Title: | Fractional Factorial Designs with 2-Level Factors |
---|---|
Description: | Regular and non-regular Fractional Factorial 2-level designs can be created. Furthermore, analysis tools for Fractional Factorial designs with 2-level factors are offered (main effects and interaction plots for all factors simultaneously, cube plot for looking at the simultaneous effects of three factors, full or half normal plot, alias structure in a more readable format than with the built-in function alias). |
Authors: | Ulrike Groemping |
Maintainer: | Ulrike Groemping <[email protected]> |
License: | GPL (>= 2) |
Version: | 2.3-3 |
Built: | 2024-11-13 06:40:32 UTC |
Source: | CRAN |
creates regular and non-regular Fractional Factorial 2-level designs. Furthermore, analysis tools for Fractional Factorial designs with 2-level factors are offered (main effects and interaction plots for all factors simultaneously, cube plot for looking at the simultaneous effects of three factors, full or half normal plot, alias structure in a more readable format than with the built-in function alias).
The package works together with packages DoE.base and DoE.wrapper.
The package is still subject to development; most key functionality is now included. Please contact me, if you have suggestions.
This package designs and analyses Fractional Factorial experiments with 2-level factors.
Regular (function FrF2
) and non-regular (function pb
) 2-level fractional factorial
designs can be generated. For regular fractional factorials, function FrF2
permits the specification of effects of interest, whose estimation is requested clear of aliasing
with other effects. The function can furthermore generate regular fractional factorials
as blocked or split-plot designs, and hard-to-change
factors can be specified in order to keep the number of level changes low.
Regular resolution V designs larger than those obtainable from function FrF2
can be created by function FrF2Large
(these are not guaranteed to be optimal).
Analysis facilities work for completely aliased designs only,
i.e. e.g. not for analysing Plackett-Burman designs with interactions.
Functions fac.design
, fractionate
or oa.design
from
Chambers and Hastie (1993) have been used as role models e.g. for
the option factor.names
or for outputting a data frame with attributes.
However, S compatibility has not been considered in devising this package. The original
above-mentioned functions are not available in R
; similar
functions have been implemented in package DoE.base
together with other general functionality for experimental designs.
In terms of analysis, package FrF2
works on linear models and enables convenient main effects and
interaction plots (functions MEPlot
and IAPlot
) similar to those
offered by Minitab software for all factors simultaneously, even though especially the
interactions are often aliased, i.e. the model is typically singular.
For the (less frequent) case of suspected three-factor-interactions, function
cubePlot
displays a cube with corners labeled with the (modeled)
means of three factors simultaneously.
Furthermore, the function DanielPlot
from package BsMD has been
modified to automatically label effects significant according to the
Lenth-criterion, to automatically distinguish between whole-plot and
split-plot effects for split-plot designs, and to provide more usage comfort
to the analyst.
Finally, the function aliases
determines the alias structure of a
Fractional Factorial 2-level design in a format more suitable for human readers
than the output from the built-in function alias
.
Ulrike Groemping
Maintainer: Ulrike Groemping <[email protected]>
Box G. E. P, Hunter, W. C. and Hunter, J. S. (2005) Statistics for Experimenters, 2nd edition. New York: Wiley.
Chambers, J.M. and Hastie, T.J. (1993). Statistical Models in S, Chapman and Hall, London.
Chen, J., Sun, D.X. and Wu, C.F.J. (1993) A catalogue of 2-level and 3-level orthogonal arrays. International Statistical Review 61, 131-145.
Daniel, C. (1959) Use of Half Normal Plots in Interpreting Two Level Experiments. Technometrics, 1, 311-340.
Groemping, U. (2014). R Package FrF2 for Creating and Analyzing Fractional Factorial 2-Level Designs. Journal of Statistical Software, 56, Issue 1, 1-56. https://www.jstatsoft.org/v56/i01/.
Hedayat, A.S., Sloane, N.J.A. and Stufken, J. (1999) Orthogonal Arrays: Theory and Applications, Springer, New York.
Lenth, R.V. (1989) Quick and easy analysis of unreplicated factorials. Technometrics, 31, 469-473.
Mee, R. (2009). A Comprehensive Guide to Factorial Two-Level Experimentation. New York: Springer.
Montgomery, D.C. (2001). Design and Analysis of Experiments (5th ed.). Wiley, New York.
Plackett, R.L.; Burman, J.P. (1946) The design of optimum multifactorial experiments. Biometrika 33, 305-325.
Ryan, K.J. and Bulutoglu, D.A. (2010). Minimum Aberration Fractional Factorial Designs With Large N. Technometrics 52, 250-255.
Sanchez, S.M. and Sanchez, P.J. (2005). Very Large Fractional Factorial and Central Composite Designs. ACM Transactions on Modeling and Computer Simulation 15, 362-377.
The key design generating functions: FrF2
, pb
, FrF2Large
S3 class design
Related packages:
DoE.base
,
DoE.wrapper
,
BsMD
;
Graphical analysis functions: MEPlot
, IAPlot
, cubePlot
,
DanielPlot
Analysis of alias structure for linear models of FrF2
designs: aliases
### for examples on design generation, cf. functions pb and FrF2 ### Injection Molding Experiment. Box et al. 1978. ## data(BM93.e3.data, package="BsMD") #from BsMD ## iMdat <- BM93.e3.data[1:16,2:10] #only original experiment ## re-create here y=c(14, 16.8, 15, 15.4, 27.6, 24, 27.4, 22.6, 22.3, 17.1, 21.5, 17.5, 15.9, 21.9, 16.7, 20.3) iMdat <- FrF2(8,7,randomize=FALSE) iMdat <- desnum(iMdat) iMdat <- rbind(cbind(iMdat,H=1),cbind(-iMdat,H=-1)) iMdat <- cbind(as.data.frame(iMdat), y=y) # make data more user-friendly colnames(iMdat) <- c("MoldTemp","Moisture","HoldPress","CavityThick","BoostPress", "CycleTime","GateSize","ScrewSpeed", "y") # linear model with all main effects and 2-factor interactions iM.lm <- lm(y ~ (.)^2, data = iMdat) # determine aliases aliases(iM.lm) # coded version aliases(iM.lm, code=TRUE) # normal plot of effects, default is autolabel with alpha=0.05 DanielPlot(iM.lm) DanielPlot(iM.lm,code=TRUE) DanielPlot(iM.lm,code=TRUE,alpha=0.5) # half normal plot of effects DanielPlot(iM.lm,code=TRUE,alpha=0.5,half=TRUE) # main effects plots MEPlot(iM.lm, las=1) # interaction plots IAPlot(iM.lm, las=1) # interaction plots with attention drawn to aliases aus <- IAPlot(iM.lm, show.alias=TRUE) # alias groups corresponding to interaction plots aliases(iM.lm)$aliases[9:15] # returned object aus # cube plot of three factors # (not very useful for this model, for demonstration only) ## per default, modeled means are shown ## this does not make a difference here, since the main effect of ## ScrewSpeed is confounded with the MoldTemp:HoldPress:BoostPress ## interaction, so that the three-factor-interaction is indirectly included ## in the modeled means cubePlot(iM.lm, "MoldTemp", "HoldPress", "BoostPress") ## modeled means without a three-factor interaction cubePlot(lm(y ~ (MoldTemp+HoldPress+BoostPress)^2, data = iMdat), "MoldTemp", "HoldPress", "BoostPress") ## modeled=FALSE reverts to showing the apparent three-factor interaction cubePlot(lm(y ~ (MoldTemp+HoldPress+BoostPress)^2, data = iMdat), "MoldTemp", "HoldPress", "BoostPress", modeled=FALSE) ## cubePlot also works on raw data cubePlot(iMdat$y, iMdat$MoldTemp, iMdat$HoldPress, iMdat$BoostPress) ## plotting functions also work directly on designs, ## if these have been generated from functions FrF2 or pb: plan <- FrF2(16, 7) plan <- add.response(plan, rnorm(16)) MEPlot(plan) IAPlot(plan) DanielPlot(plan)
### for examples on design generation, cf. functions pb and FrF2 ### Injection Molding Experiment. Box et al. 1978. ## data(BM93.e3.data, package="BsMD") #from BsMD ## iMdat <- BM93.e3.data[1:16,2:10] #only original experiment ## re-create here y=c(14, 16.8, 15, 15.4, 27.6, 24, 27.4, 22.6, 22.3, 17.1, 21.5, 17.5, 15.9, 21.9, 16.7, 20.3) iMdat <- FrF2(8,7,randomize=FALSE) iMdat <- desnum(iMdat) iMdat <- rbind(cbind(iMdat,H=1),cbind(-iMdat,H=-1)) iMdat <- cbind(as.data.frame(iMdat), y=y) # make data more user-friendly colnames(iMdat) <- c("MoldTemp","Moisture","HoldPress","CavityThick","BoostPress", "CycleTime","GateSize","ScrewSpeed", "y") # linear model with all main effects and 2-factor interactions iM.lm <- lm(y ~ (.)^2, data = iMdat) # determine aliases aliases(iM.lm) # coded version aliases(iM.lm, code=TRUE) # normal plot of effects, default is autolabel with alpha=0.05 DanielPlot(iM.lm) DanielPlot(iM.lm,code=TRUE) DanielPlot(iM.lm,code=TRUE,alpha=0.5) # half normal plot of effects DanielPlot(iM.lm,code=TRUE,alpha=0.5,half=TRUE) # main effects plots MEPlot(iM.lm, las=1) # interaction plots IAPlot(iM.lm, las=1) # interaction plots with attention drawn to aliases aus <- IAPlot(iM.lm, show.alias=TRUE) # alias groups corresponding to interaction plots aliases(iM.lm)$aliases[9:15] # returned object aus # cube plot of three factors # (not very useful for this model, for demonstration only) ## per default, modeled means are shown ## this does not make a difference here, since the main effect of ## ScrewSpeed is confounded with the MoldTemp:HoldPress:BoostPress ## interaction, so that the three-factor-interaction is indirectly included ## in the modeled means cubePlot(iM.lm, "MoldTemp", "HoldPress", "BoostPress") ## modeled means without a three-factor interaction cubePlot(lm(y ~ (MoldTemp+HoldPress+BoostPress)^2, data = iMdat), "MoldTemp", "HoldPress", "BoostPress") ## modeled=FALSE reverts to showing the apparent three-factor interaction cubePlot(lm(y ~ (MoldTemp+HoldPress+BoostPress)^2, data = iMdat), "MoldTemp", "HoldPress", "BoostPress", modeled=FALSE) ## cubePlot also works on raw data cubePlot(iMdat$y, iMdat$MoldTemp, iMdat$HoldPress, iMdat$BoostPress) ## plotting functions also work directly on designs, ## if these have been generated from functions FrF2 or pb: plan <- FrF2(16, 7) plan <- add.response(plan, rnorm(16)) MEPlot(plan) IAPlot(plan) DanielPlot(plan)
This function adds center points to a 2-level fractional factorial design. All factors must be quantitative!
add.center(design, ncenter, distribute=NULL, ...)
add.center(design, ncenter, distribute=NULL, ...)
design |
a data frame of class design that contains a 2-level fractional factorial
(regular or non-regular);
For function |
ncenter |
the number of center points to be added to each block |
distribute |
the number of positions over which to distribute the center points within each block;
note that the center points are not randomized but placed evenly throughout the
(hopefully randomdomized) design (but see also the details section);
|
... |
currently not used |
Function add.center
adds center points to 2-level fractional factorial
designs. Instead of using this function directly, center points should usually
be added directly with calls to functions FrF2
or pb
.
These make use of function add.center
for this purpose.
Center points are added to designs for three main reasons: they provide a repeated benchmark run that can alert the experimenter to unplanned changes in experimental conditions, they provide an independent estimate of experimental error, and finally they provide a possibility for checking whether a first order model is sufficient. Especially for the first purpose, package FrF2 follows the recommendation in Montgomery (2001, p.275). To distinguish them from the center points, the original fractional factorial runs are called “cube points”.
Addition of center points does not affect estimates for main effects and interactions. The difference between the averages of cube points and center points gives an indication whether quadratic terms might be needed in the model.
For blocked designs and properly replicated designs,
ncenter
center points are added to each (replication) block.
In case of repeated measurements, center points are also measured repeatedly.
Center points are distributed as evenly as possible over the distribute
selected
positions throughout each block. distribute=1
always adds all center points at the end of
each block. If distribute > 1
, (each block of) the design starts and ends
with a (group of) center point(s),
and the distribute
positions for placing center points are as evenly
placed throughout (each block of) the design as possible.
If ncenter
is not a multiple of distribute
,
some center point groups have one more center point than others. If ncenter%%distribute
is one or two only, the beginning and (for two) the end of (each block of) the design have one more center point,
otherwise the ncenter%%distribute
extra center points are randomized over the center point positions.
Function iscube
from package DoE.base
provides a logical vector that is TRUE for cube points
and FALSE for center points, which allows to use of simple functions
for “clean” 2-level fractional factorials like MEPlot
.
A data frame of class design with ncenter
center point runs per block
(or per replication block) added to the design
(and its desnum
and run.order
attributes).
The run.no.in.std.order column of run.order is “0” for the center points.
Existing response values for cube runs are preserved, and response values for the
new center point runs are NA. Note, however, that center points should be added
BEFORE running the experiment in order to benefit from all their useful properties;
this should best be done within functions pb
or FrF2
.
The design is identifiable as a design with center points by the
suffix .center
to the type
element of attribute design.info
,
and the elements ncube
and ncenter
are added
(with the updated nruns
being their sum). The element coding
is
also added to the design.info
, in order to support steepest ascent/descent
analysis from the center point.
This function is still somewhat experimental.
Ulrike Groemping
Montgomery, D.C. (2001). Design and Analysis of Experiments (5th ed.). Wiley, New York.
## purely technical example plan <- FrF2(8,5, factor.names=c("one","two","three","four","five")) add.center(plan, 6) add.center(plan, 6, distribute=1) add.center(plan, 6, distribute=6) add.center(plan, 6, distribute=4) ## very artificial analysis example plan <- FrF2(8,4, factor.names=list(one=c(0,10),two=c(1,3),three=c(25,32),four=c(3.7,4.8))) ## add some response data y <- c(2+desnum(plan)%*%c(2,3,0,0) + 1.5*apply(desnum(plan)[,c(1,2)],1,"prod") + rnorm(8)) ## the "c()" makes y into a vector rather than a 1-column matrix plan <- add.response(plan, y) ## analysing this design provides an impression MEPlot(lm(y~(.)^2, plan)) IAPlot(lm(y~(.)^2, plan)) DanielPlot(lm(y~(.)^2,plan), half=TRUE, alpha=0.2) ## tentative conclusion: factors one and two do something ## wonder whether the model with one and two and their interaction is sufficient ## look at center points (!!! SHOULD HAVE BEEN INCLUDED FROM THE START, ## but maybe better now than not at all) ## use distribute=1, because all center points are run at the end planc <- add.center(plan, 6, distribute=1) ## conduct additional runs for the center points y <- c(y, c(2+desnum(planc)[!iscube(planc),1:4]%*%c(2,3,0,0) + 1.5*apply(desnum(planc)[!iscube(planc),][,c(1,2)],1,"prod") + rnorm(6))) ## add to the design planc <- add.response(planc, y, replace=TRUE) ## sanity check: repeat previous analyses for comparison, with the help of function iscube() MEPlot(lm(y~(.)^2, planc, subset=iscube(planc))) IAPlot(lm(y~(.)^2, planc, subset=iscube(planc))) DanielPlot(lm(y~(.)^2, planc, subset=iscube(planc)), half=TRUE, alpha=0.2) ## quick check whether there a quadratic effect is needed: is the cube indicator significant ? summary(lm(y~(.)^2+iscube(planc), planc)) ## (in this unrealistic example, the quadratic effect is dominating everything else; ## with an effect that strong in practice, it is likely that ## one would either have expected a strong non-linearity before conducting the experiment, ## OR that the effect is not real but the result of some stupid mistake ## alternatively, the check can be calculated per hand (cf. e.g. Montgomery, Chapter 11): (mean(planc$y[iscube(planc)])-mean(planc$y[!iscube(planc)]))^2*8*6/(8+6)/var(y[!iscube(planc)]) ## must be compared to the F-quantile with 1 degree of freedom ## is the square of the t-value for the cube indicator in the linear model
## purely technical example plan <- FrF2(8,5, factor.names=c("one","two","three","four","five")) add.center(plan, 6) add.center(plan, 6, distribute=1) add.center(plan, 6, distribute=6) add.center(plan, 6, distribute=4) ## very artificial analysis example plan <- FrF2(8,4, factor.names=list(one=c(0,10),two=c(1,3),three=c(25,32),four=c(3.7,4.8))) ## add some response data y <- c(2+desnum(plan)%*%c(2,3,0,0) + 1.5*apply(desnum(plan)[,c(1,2)],1,"prod") + rnorm(8)) ## the "c()" makes y into a vector rather than a 1-column matrix plan <- add.response(plan, y) ## analysing this design provides an impression MEPlot(lm(y~(.)^2, plan)) IAPlot(lm(y~(.)^2, plan)) DanielPlot(lm(y~(.)^2,plan), half=TRUE, alpha=0.2) ## tentative conclusion: factors one and two do something ## wonder whether the model with one and two and their interaction is sufficient ## look at center points (!!! SHOULD HAVE BEEN INCLUDED FROM THE START, ## but maybe better now than not at all) ## use distribute=1, because all center points are run at the end planc <- add.center(plan, 6, distribute=1) ## conduct additional runs for the center points y <- c(y, c(2+desnum(planc)[!iscube(planc),1:4]%*%c(2,3,0,0) + 1.5*apply(desnum(planc)[!iscube(planc),][,c(1,2)],1,"prod") + rnorm(6))) ## add to the design planc <- add.response(planc, y, replace=TRUE) ## sanity check: repeat previous analyses for comparison, with the help of function iscube() MEPlot(lm(y~(.)^2, planc, subset=iscube(planc))) IAPlot(lm(y~(.)^2, planc, subset=iscube(planc))) DanielPlot(lm(y~(.)^2, planc, subset=iscube(planc)), half=TRUE, alpha=0.2) ## quick check whether there a quadratic effect is needed: is the cube indicator significant ? summary(lm(y~(.)^2+iscube(planc), planc)) ## (in this unrealistic example, the quadratic effect is dominating everything else; ## with an effect that strong in practice, it is likely that ## one would either have expected a strong non-linearity before conducting the experiment, ## OR that the effect is not real but the result of some stupid mistake ## alternatively, the check can be calculated per hand (cf. e.g. Montgomery, Chapter 11): (mean(planc$y[iscube(planc)])-mean(planc$y[!iscube(planc)]))^2*8*6/(8+6)/var(y[!iscube(planc)]) ## must be compared to the F-quantile with 1 degree of freedom ## is the square of the t-value for the cube indicator in the linear model
Functions to examine the alias structure of a fractional factorial 2-level design
aliases(fit, code = FALSE, condense=FALSE) aliasprint(design, ...) ## S3 method for class 'aliases' print(x, ...)
aliases(fit, code = FALSE, condense=FALSE) aliasprint(design, ...) ## S3 method for class 'aliases' print(x, ...)
fit |
a linear model object with only 2-level factors as explanatory variables; the function will return an error, if the model contains partially aliased effects (like interactions in a Plackett-Burman design for most cases) |
code |
if TRUE, requests that aliasing is given in code letters (A, B, C etc.) instead of (potentially lengthy) variable names; in this case, a legend is included in the output object. |
condense |
if TRUE, reformats the alias information to be comparable to the version calculated by internal function alias3fi; does not work with models with higher than 3-way interactions; for up to 3-way interactions, the output may be more easily readible |
design |
a data frame of class |
x |
an object of class |
... |
further arguments to function |
Function aliasprint
returns NULL and is called for its side effects only.
Per default, Function aliases
returns a list with two elements:
legend |
links the codes to variable names, if |
aliases |
is a list of vectors of aliased effects. |
If option condense
is TRUE, the function returns a list with elements legend,
main, fi2 and fi3; this may be preferrable for looking at the alias structure of larger designs.
The output object of function aliases
has class aliases
,
which is used for customized printing with the print
method.
Ulrike Groemping
Box G. E. P, Hunter, W. C. and Hunter, J. S. (2005) Statistics for Experimenters, 2nd edition. New York: Wiley.
FrF2-package
for information on the package,
alias
for the built-in R-function,
IAPlot
for effects plots
### Injection Molding Experiment. Box et al. 1978. ## data(BM93.e3.data, package="BsMD") #from BsMD ## iMdat <- BM93.e3.data[1:16,2:10] #only original experiment ## re-create here y=c(14, 16.8, 15, 15.4, 27.6, 24, 27.4, 22.6, 22.3, 17.1, 21.5, 17.5, 15.9, 21.9, 16.7, 20.3) iMdat <- FrF2(8,7,randomize=FALSE) iMdat <- desnum(iMdat) iMdat <- rbind(cbind(iMdat,H=1),cbind(-iMdat,H=-1)) iMdat <- cbind(as.data.frame(iMdat), y=y) # make data more user-friendly colnames(iMdat) <- c("MoldTemp","Moisture","HoldPress","CavityThick", "BoostPress","CycleTime","GateSize","ScrewSpeed","y") # determine aliases with all 2-factor-interactions aliases(lm(y ~ (.)^2, data = iMdat)) # coded version aliases(lm(y ~ (.)^2, data = iMdat), code=TRUE) # determine aliases with all 3-factor-interactions aliases(lm(y ~ (.)^3, data = iMdat), code=TRUE) # show condensed form aliases(lm(y ~ (.)^3, data = iMdat), code=TRUE, condense=TRUE) # determine aliases for unaliased model aliases(lm(y ~ ., data = iMdat))
### Injection Molding Experiment. Box et al. 1978. ## data(BM93.e3.data, package="BsMD") #from BsMD ## iMdat <- BM93.e3.data[1:16,2:10] #only original experiment ## re-create here y=c(14, 16.8, 15, 15.4, 27.6, 24, 27.4, 22.6, 22.3, 17.1, 21.5, 17.5, 15.9, 21.9, 16.7, 20.3) iMdat <- FrF2(8,7,randomize=FALSE) iMdat <- desnum(iMdat) iMdat <- rbind(cbind(iMdat,H=1),cbind(-iMdat,H=-1)) iMdat <- cbind(as.data.frame(iMdat), y=y) # make data more user-friendly colnames(iMdat) <- c("MoldTemp","Moisture","HoldPress","CavityThick", "BoostPress","CycleTime","GateSize","ScrewSpeed","y") # determine aliases with all 2-factor-interactions aliases(lm(y ~ (.)^2, data = iMdat)) # coded version aliases(lm(y ~ (.)^2, data = iMdat), code=TRUE) # determine aliases with all 3-factor-interactions aliases(lm(y ~ (.)^3, data = iMdat), code=TRUE) # show condensed form aliases(lm(y ~ (.)^3, data = iMdat), code=TRUE, condense=TRUE) # determine aliases for unaliased model aliases(lm(y ~ ., data = iMdat))
This help page documents the statistical and algorithmic details of blocking in FrF2
Blocking is done with the purpose to balance the design with respect to a factor that is known or strongly suspected to have an influence but is not in itself of interest, and it is usually assumed that block factors do not interact with experimental factors. Examples are batches of material that are not large enough to accomodate the complete experiment so that e.g. half the experiment is done on the first batch and the other half on the second batch (two blocks). The block factor should be orthogonal to the experimental factors, at least to their main effects. Per default, it is also requested that the block factor is orthogonal to the 2-factor interactions. This can be changed by the user, if no such design can be found.
Blocking is currently implemented for regular fractional factorial designs only.
There are two principal ways to handle blocked designs, manual definition
(i.e. the user specifies exactly which columns are to be used for which purpose) and automatic
definition. Each situation has its specifics. These are detailed below. For users with
not so much mathematical/statistical background, it will often be best to use the automatic way,
specifying the treatment factors of interest via nfactors
or factor.names
and a single number for blocks
or WPs
.
Users with more mathematical background may want to use the manual definitions, perhaps
in conjunction with published catalogues of good block designs, or
after inspecting possibilities with functions blockpick
, blockpick.big
(default before version 2 for large settings) or colpick
(default since version 2 for large settings or settings with estimability requirements).
The user can start from a design with a number of factors and manually specify which factors or
interactions are to be used as block generators. If this route is chosen, blocks
can be a vector of factor names or factor letters, or of the same form as generators, except that
not only base factors but all factors can be used and single factors are permitted
(which would lead to resolution II designs if used in generators). For example, block = Letters[c(2,4,5)]
or block = list(2,4,5)
specify that the 2nd, 4th and 5th factor are to be used as block generators, while block = c("Day","Shift")
indicates that the named factors “Day” and “Shift” specified in factor.names
are to be treated as blocking factors). In this case, the number of blocks is calculated,
and a new factor with the default name “Blocks” (in general the name chosen in
option block.name
) is generated, which would for example contain as levels
the Day/Shift combinations. It is also possible to choose interaction effects rather than factors themselves
as block generators, e.g. block = c("ABCD","EFGH")
orblock = list(c(1,2,3,4),c(5,6,7,8))
.
Finally, it is also possible to specify choice of blocks using a vector of Yates column numbers,
in order to be able to use catalogued blocking structures of this form, e.g. from Sitter, Chen and Feder
(1997).
The block main effects are defined by the k.block
specified effect
and all interactions between them. The specified block effects are required to be independent from each other,
which implies that they generate 2^k.block
blocks.
CAUTION: If the user manually generates a blocked design, it is his/her responsibility to ensure a
good choice of design (e.g. by using a catalogued design from Bisgaard 1994,
Sun, Wu and Chen 1997, Sitter, Chen and Feder (1997), or Cheng and Wu 2002).
Since version 2 of package FrF2, manual blocking is also checked for confounding
of the block factor with main effects or two-factor interactions;
this implies that some earlier code will now require the additional specification
of argument alias.block.2fis=TRUE
in order to avoid errors.
If the user only specifies the number of blocks required for the experiment, function FrF2
automatically generates the blocks. For full factorial designs, function FrF2
uses function colpick
with subsequent blockgencreate
, except where the Sun, Wu and Chen (1997) catalogue of blocked designs
contains suitable block generators for a design without estimability requirements
(implemented in function blockpick
, which also calls colpick
, if that
catalogue does not offer a solution).
Otherwise, depending on the situation,
function FrF2
uses function blockpick
or function colpick
with subsequent blockgencreate
;
function blockpick
treats smaller problems (choose(nruns-1-nfactors,k.block) < 100000
)
without estimability requirements and with force.godolphin=FALSE
(the latter is per default set to TRUE
whenever alias.block.2fis=TRUE
),
other problems are treated by function colpick
.
Use of the earlier default function blockpick.big
for large cases or the earlier behavior for full factorial designs
can be requested with the argument
block.old=TRUE
; this should only be done for reproducing earlier results, as the new methodology is definitely superior.
The search for an appropriate blocked design starts with the overall best unblocked design
(in terms of aberration or MaxC2, if requested).
If this best design does not yield an adequate blocking possibility, the search continues with
the next best design and so forth (exception: with an estimability requirement, only a single design, prefiltered for the estimability requirement, is subjected to the blocking algorithm).
For the smaller problems, function blockpick
looks for k.block
independent subsets among the eligible columns of the design.
(The eligible columns are all columns of the Yates matrix that are neither occupied
by treatment main effects nor by 2fis among treatments (if alias.block.2fis=FALSE
,
which is the default), or all columns of the Yates matrix that are not occupied by treatment main effects
(if alias.block.2fis=TRUE
). Note that no effort is made to avoid aliasing with 2-factor interactions,
if alias.block.2fis=TRUE
is chosen.
For the larger problems, or blocking in combination with requiring some 2fis to be clear of aliasing,
or per default for blocking with permitting 2fis to be aliased with blocks,
function colpick
creates a X matrix for creating
blocks of size
based on the approach described by Godolphin (2021);
function
blockgencreate
creates block generators from this matrix.
This approach can be used in combination with argument estimable
,
as long as clear=TRUE
. The implementation of this approach is described in Groemping (2021).
The argument force.godolphin
of function FrF2
can enforce the Godolphin approach instead
of the default approach for small blocked designs without alias.block.2fis=TRUE
,
and the Godolphin approach can be switched off for alias.block.2fis=TRUE
applications by explicitly requesting
force.godolphin=FALSE
. Note that the Godolphin approach solely focuses on clear 2fis
of the blocked design and does not attempt to avoid confounding of the block factor with non-clear 2fis; it may
thus confound 2fis with the block factor even if this were avoidable, maintaining the same number of clear 2fis.
For the larger problems, in versions before 2.0, which can be activated in current versions with block.old=TRUE
,
function blockpick.big
permutes the k~base factors of candidate designs with nfactors + k.block
factors
in search of a design the first k.block
~factors of which can be used for block construction. Any
specification of design (via options design
or generators
) is ignored. Note that function
blockpick.big
is not guaranteed to find an existing blocked design.
Sun, Wu and Chen (1997) provided a catalogue of blocked designs
with a few quality criteria, and they stated that there is no single best design, but that the choice
depends on the situation. FrF2
always comes up with one specific solution design.
Comparisons to the catalogued designs in Sun, Wu and Chen (1997) have shown that
the designs found in FrF2
are often but not always isomorphic to the catalogued ones.
Differences do occur, especially if the base designs are resolution III, or if blockpick.big
has to be used. Expert users who want to be certain to use a “best” blocked design should manually
implement a specific catalogued design or inspect several solutions from functions blockpick
or colpick
(or, if desparate, blockpick.big
).
Please contact me with any suggestions for improvements.
Ulrike Groemping
Bisgaard, S. (1994a). Blocking generators for small designs.
J. Quality Technology 26, 288-294.
Chen, J., Sun, D.X. and Wu, C.F.J. (1993) A catalogue of 2-level and 3-level orthogonal arrays. International Statistical Review 61, 131-145.
Cheng, C.-S. and Tsai, P.-W. (2009). Optimal two-level regular fractional factorial block and split-plot designs. Biometrika 96, 83-93.
Cheng, S.W. and Wu, C.F.J. (2002). Choice of optimal blocking schemes in 2-level and 3-level designs. Technometrics 44, 269-277.
Godolphin, J. (2021). Construction of Blocked Factorial Designs to Estimate Main Effects and Selected Two-Factor Interactions. J. Royal Statistical Society B 83, 5-29. doi:10.1111/rssb.12397.
Groemping, U. (2019). An algorithm for blocking regular fractional factorial 2-level designs with clear two-factor interactions. Reports in Mathematics, Physics and Chemistry, Report 3/2019, Department II, Beuth University of Applied Sciences Berlin.
Sitter, R.R., Chen, J. and Feder, M. (1997). Fractional Resolution and Minimum Aberration in Blocked 2n-k Designs. Technometrics 39, 382–390.
Sun, D.X., Wu, C.F.J. and Chen, Y.Y. (1997).
Optimal blocking schemes for and
designs. Technometrics 39,
298-307.
See Also FrF2
for regular fractional factorials,
catlg
for the Chen, Sun, Wu catalogue of designs
and some accessor functions,
and splitplot
for the statistical aspects of split-plot designs.
########## automatic blocked designs ################### ## from a full factorial ## FrF2(8,3,blocks=2) ## with replication run.order(FrF2(8,3,blocks=2,wbreps=2)) run.order(FrF2(8,3,blocks=2,wbreps=2,repeat.only=TRUE)) run.order(FrF2(8,3,blocks=2,bbreps=2)) run.order(FrF2(8,3,blocks=2,bbreps=2,wbreps=2)) ## automatic blocked design with fractions FrF2(16,7,blocks=4,alias.block.2fis=TRUE) ## isomorphic non-catalogued design as basis FrF2(16,gen=c(7,11,14),blocks=4,alias.block.2fis=TRUE) ## FrF2 uses blockpick.big and ignores the generator FrF2(64,gen=c(7,11,14),blocks=16,alias.block.2fis=TRUE) ########## manual blocked design #################### ### example that shows why order of blocks is not randomized ### can of course be randomized by user, if appropriate FrF2(32,9,blocks=c("Day","Shift"),alias.block.2fis=TRUE, factor.names=list(Day=c("Wednesday","Thursday"), Shift=c("Morning","Afternoon"), F1="",F2="",F3="",F4="",F5="",F6="",F7=""), default.levels=c("current","new")) ########## blocked design with estimable 2fis #################### ### all interactions of last two factors to be estimable clearly ### in 64 run design with blocks of size 4 ### not possible with catalogue entry 9-3.1 FrF2(design="9-3.2", blocks=16, alias.block.2fis=TRUE, factor.names = list(C1="",C2="",C3="",C4="",C5="",C6="",C7="", N1=c("low","high"),N2=c("low","high")), default.levels = c("current","new"), estimable=compromise(9, 8:9)$requirement)
########## automatic blocked designs ################### ## from a full factorial ## FrF2(8,3,blocks=2) ## with replication run.order(FrF2(8,3,blocks=2,wbreps=2)) run.order(FrF2(8,3,blocks=2,wbreps=2,repeat.only=TRUE)) run.order(FrF2(8,3,blocks=2,bbreps=2)) run.order(FrF2(8,3,blocks=2,bbreps=2,wbreps=2)) ## automatic blocked design with fractions FrF2(16,7,blocks=4,alias.block.2fis=TRUE) ## isomorphic non-catalogued design as basis FrF2(16,gen=c(7,11,14),blocks=4,alias.block.2fis=TRUE) ## FrF2 uses blockpick.big and ignores the generator FrF2(64,gen=c(7,11,14),blocks=16,alias.block.2fis=TRUE) ########## manual blocked design #################### ### example that shows why order of blocks is not randomized ### can of course be randomized by user, if appropriate FrF2(32,9,blocks=c("Day","Shift"),alias.block.2fis=TRUE, factor.names=list(Day=c("Wednesday","Thursday"), Shift=c("Morning","Afternoon"), F1="",F2="",F3="",F4="",F5="",F6="",F7=""), default.levels=c("current","new")) ########## blocked design with estimable 2fis #################### ### all interactions of last two factors to be estimable clearly ### in 64 run design with blocks of size 4 ### not possible with catalogue entry 9-3.1 FrF2(design="9-3.2", blocks=16, alias.block.2fis=TRUE, factor.names = list(C1="",C2="",C3="",C4="",C5="",C6="",C7="", N1=c("low","high"),N2=c("low","high")), default.levels = c("current","new"), estimable=compromise(9, 8:9)$requirement)
Functions to investigate potential assignments of blocks and show alias information of resulting designs, meant for expert users
blockpick(k, gen, k.block, design = NULL, show = 10, alias.block.2fis = FALSE, select.catlg = catlg) blockpick.big(k, gen, k.block, design = NULL, show = 10, alias.block.2fis = FALSE, select.catlg = catlg)
blockpick(k, gen, k.block, design = NULL, show = 10, alias.block.2fis = FALSE, select.catlg = catlg) blockpick.big(k, gen, k.block, design = NULL, show = 10, alias.block.2fis = FALSE, select.catlg = catlg)
k |
the number of base factors (designs have |
gen |
vector of generating columns from Yates matrix;
for a full factorial, choose |
k.block |
number of base factors needed for constructing blocks;
there will be |
design |
design name (character string) of a specific design from the catalogue given in select.catlg |
show |
numeric integer indicating how many results are to be shown;
the search for possible allocations stops, once |
alias.block.2fis |
logical, indicates whether 2fis may be aliased with blocks |
select.catlg |
design catalogue of class |
Function blockpick
is used per default by function FrF2
for problems with choose(nruns-1-nfactors,k.block) < 100000
and without
estimability requirements. blockpick
will find a design, if it exists.
However, it may take a long time and/or much storage space in problems with
large numbers of runs and blocks.
In FrF2 versions before 2.0, function blockpick.big
was used for
large use cases; this can still be requested using argument block.old=TRUE
.
Since FrF2 version 2, the Godolphin (2021) based approach
is used instead, both for
large cases and for cases where blocking is combined with estimability requirements
(clear=TRUE
only); the big advantage is the ability of combining blocking with
estimability requirements, and a substantial speed gain if small blocks are needed.
All approaches investigate the potential assignment of blocks such that
main effects of treatment factors are not aliased with block main effects.
It is left to the user whether or not 2fis amoong treatment effects may be
aliased with block main effects (option alias.block.2fis
). (For the Godolphin
approach to work, one will usually need to set alias.block.2fis
to TRUE.)
Following Sun, Wu and Chen (1997), there is no single best block assignment.
blockpick
uses their catalogue for full factorials (implemented up to 256 runs).
For fractional factorials, it develops designs according to a
principle similar to that underlying the Sun Wu Chen catalogue that works also in
uncatalogued situations.
Function blockpick.big
uses a strategy similar to splitpick
and leftadjust
and often finds a solution quickly where blockpick
does not work with the
given ressources. However, it is not guaranteed to find existing solutions
or a best solution.
The function blockpick
outputs a list of entries with information on at most show
suitable
assignments. It ends with an error, if no suitable solution can be found.
gen |
generator column numbers of the base design (w.r.t. the Yates matrix) |
basics |
named vector with number of runs ( |
blockcols |
matrix with at most show rows; each row contains the |
alias.2fis.block |
list of character vectors, which contain the 2fis
aliased with block main effects for the respective rows of |
nblock.2fis |
vector with number of 2fis aliased with block main effects
for the respective rows of |
nclear.2fis |
vector with number of 2fis clear (of aliasing with block main effects
and treatment main effects or 2fis)
for the respective rows of |
clear.2fis |
list of character vectors, which contain the 2fis that are
counted in |
Ulrike Groemping
Chen, J., Sun, D.X. and Wu, C.F.J. (1993) A catalogue of 2-level and 3-level orthogonal arrays. International Statistical Review 61, 131-145.
Sun, D.X., Wu, C.F.J. and Chen, Y.Y. (1997).
Optimal blocking schemes for and
designs. Technometrics 39,
298-307.
See Also FrF2
## look at possibilities for running a 32 run design with 6 factors in 8 blocks ## running this without alias.block.2fis=TRUE throws an error: not possible ## Not run: blockpick(k=5,design="6-1.1",k.block=3) ## the 8th to 10th design have more clear 2fis than the earlier ones blockpick(k=5,design="6-1.1",k.block=3,alias.block.2fis=TRUE) ## function FrF2 can be used to manually accomodate this des32.6fac.8blocks.MaxC2 <- FrF2(32,6,blocks=c(3,12,21),alias.block.2fis=TRUE) summary(des32.6fac.8blocks.MaxC2) ## automatic block generation leads to more aliased 2fis summary(FrF2(32,6,blocks=8,alias.block.2fis=TRUE)) ## look at possibilities for blocking design 7-3.1 from Chen, Sun, Wu catalogue blockpick(4,design="7-3.1",k.block=2,alias.block.2fis=TRUE) ## big design ## running this throws an error on many machines because of too little memory ## Not run: blockpick(6,design="7-1.2",k.block=5,alias.block.2fis=TRUE) ## for obtaining a design for this scenario with blockpick.big, ## the number of factors must be increased to 7+k.block=12 ## designs 12-6.1 and 12-6.2 dont do it, 12-6.3 does bpb <- blockpick.big(6,design="12-6.3",k.block=5,alias.block.2fis=TRUE) bpb ## based on the result of blockpick.big, a blocked design can be obtained as follows: ## (not run for saving check time on CRAN) ## Not run: des64.7fac.32blocks <- FrF2(64,gen=bpb$gen[1,], blocks = as.list(1:5), alias.block.2fis=TRUE) str(des64.7fac.32blocks) ## if the seven factors are to be named A,...,G: des64.7fac.32blocks <- FrF2(64,gen=bpb$gen[1,], blocks = as.list(1:5), alias.block.2fis=TRUE, factor.names=c(paste("b",1:5,sep=""),Letters[1:7])) str(des64.7fac.32blocks) ## End(Not run)
## look at possibilities for running a 32 run design with 6 factors in 8 blocks ## running this without alias.block.2fis=TRUE throws an error: not possible ## Not run: blockpick(k=5,design="6-1.1",k.block=3) ## the 8th to 10th design have more clear 2fis than the earlier ones blockpick(k=5,design="6-1.1",k.block=3,alias.block.2fis=TRUE) ## function FrF2 can be used to manually accomodate this des32.6fac.8blocks.MaxC2 <- FrF2(32,6,blocks=c(3,12,21),alias.block.2fis=TRUE) summary(des32.6fac.8blocks.MaxC2) ## automatic block generation leads to more aliased 2fis summary(FrF2(32,6,blocks=8,alias.block.2fis=TRUE)) ## look at possibilities for blocking design 7-3.1 from Chen, Sun, Wu catalogue blockpick(4,design="7-3.1",k.block=2,alias.block.2fis=TRUE) ## big design ## running this throws an error on many machines because of too little memory ## Not run: blockpick(6,design="7-1.2",k.block=5,alias.block.2fis=TRUE) ## for obtaining a design for this scenario with blockpick.big, ## the number of factors must be increased to 7+k.block=12 ## designs 12-6.1 and 12-6.2 dont do it, 12-6.3 does bpb <- blockpick.big(6,design="12-6.3",k.block=5,alias.block.2fis=TRUE) bpb ## based on the result of blockpick.big, a blocked design can be obtained as follows: ## (not run for saving check time on CRAN) ## Not run: des64.7fac.32blocks <- FrF2(64,gen=bpb$gen[1,], blocks = as.list(1:5), alias.block.2fis=TRUE) str(des64.7fac.32blocks) ## if the seven factors are to be named A,...,G: des64.7fac.32blocks <- FrF2(64,gen=bpb$gen[1,], blocks = as.list(1:5), alias.block.2fis=TRUE, factor.names=c(paste("b",1:5,sep=""),Letters[1:7])) str(des64.7fac.32blocks) ## End(Not run)
The function calculates Bayesian posterior probabilities according to Box and Meyer (1993) for screening experiments with 2-level factors. The function is modified from function BsProb in packge BsMD with the purpose of providing usage comfort for class design objects.
BsProb.design(design, mFac = NULL, response=NULL, select=NULL, mInt = 2, p = 0.25, g = 2, ng = 1, nMod = 10)
BsProb.design(design, mFac = NULL, response=NULL, select=NULL, mInt = 2, p = 0.25, g = 2, ng = 1, nMod = 10)
design |
an experimental design of class |
response |
NULL or a character string that specifies response variable to be used,
must be an element of |
mFac |
integer. Maximum number of factors included in the models. The default is the number of factors in the design. |
select |
vector with position numbers of the factors to be included; |
mInt |
integer <= 3. Maximum order of interactions considered in the models. This can strongly impact the result. |
p |
numeric. Prior probability assigned to active factors. This can strongly impact the result. |
g |
numeric vector. Variance inflation factor(s) gamma associated to active and interaction factors; see "Details" section |
ng |
integer <=20. Number of different variance inflation factors (g) used in calculations. |
nMod |
integer <=100. Number of models to keep with the highest posterior probability. |
Factor and model posterior probabilities are computed by the Box and Meyer (1993) Bayesian procedure.
The design factors - or a selection of these given by column numbers in select
-
are considered together with the specified response or the first response of the design.
The function has been adapted from function BsProb
in package BsMD,
and a vignette in that package (../../BsMD/doc/BsMD.pdf) explains
details of the usage regarding the parameters.
If g
, the variance inflation factor (VIF) gamma, is a vector of length 1,
the same VIF is used for factor main effects and interactions.
If the length of g
is 2 and ng
is 1, g[1]
is used for factor main effects and g[2]
for the interaction effects.
If ng
greater than 1, then ng
values of VIFs between g[1]
and
g[2]
are used for calculations with the same gamma value for main effects
and interactions. The function calls the FORTRAN subroutine bm
and captures
summary results. The complete output of the FORTRAN code is save in the BsPrint.out
file in the working directory. The output is a list of class BsProb
for which print, plot and summary methods are available from package BsMD.
cf. documentation of function BsProb
This method relies on the availability of package BsMD.
Daniel Meyer, ported to R by Ernesto Barrios, port adapted to designs by Ulrike Groemping.
Barrios, E. (2013). Using the BsMD Package for Bayesian Screening and Model Discrimination. Vignette. ../../BsMD/doc/BsMD.pdf.
Box, G. E. P and R. D. Meyer (1986). An Analysis for Unreplicated Fractional Factorials. Technometrics 28, 11-18.
Box, G. E. P and R. D. Meyer (1993). Finding the Active Factors in Fractionated Screening Experiments. Journal of Quality Technology 25, 94-105.
plot.BsProb
, print.BsProb
,
summary.BsProb
, BsMD
### there are several success stories and recommendations for this method ### in the simulated example here (not fabricated, ### it was the first one that came to my mind), ### the method goes wrong, at least when using mInt=2 (the default, because ### Daniel plots work quite well for pure main effects models): ### active factors are A to E (perhaps too many for the method to work), ### the method identifies F, J, and L with highest probability ### (but is quite undecided) plan <- pb(12) dn <- desnum(plan) set.seed(8655) y <- dn%*%c(2,2,2,2,3,0,0,0,0,0,0) + dn[,1]*dn[,3]*2 - dn[,5]*dn[,4] + rnorm(12)/10 plan.r <- add.response(plan, response=y) if (requireNamespace("BsMD", quiet=TRUE)){ plot(bpmInt2 <- BsProb.design(plan.r), code=FALSE) plot(bpmInt1 <- BsProb.design(plan.r, mInt=1), code=FALSE) ## much better! summary(bpmInt2) summary(bpmInt1) } ### For comparison: A Daniel plot does not show any significant effects according ### to Lenths method, but makes the right effects stick out DanielPlot(plan.r, half=TRUE, alpha=1)
### there are several success stories and recommendations for this method ### in the simulated example here (not fabricated, ### it was the first one that came to my mind), ### the method goes wrong, at least when using mInt=2 (the default, because ### Daniel plots work quite well for pure main effects models): ### active factors are A to E (perhaps too many for the method to work), ### the method identifies F, J, and L with highest probability ### (but is quite undecided) plan <- pb(12) dn <- desnum(plan) set.seed(8655) y <- dn%*%c(2,2,2,2,3,0,0,0,0,0,0) + dn[,1]*dn[,3]*2 - dn[,5]*dn[,4] + rnorm(12)/10 plan.r <- add.response(plan, response=y) if (requireNamespace("BsMD", quiet=TRUE)){ plot(bpmInt2 <- BsProb.design(plan.r), code=FALSE) plot(bpmInt1 <- BsProb.design(plan.r, mInt=1), code=FALSE) ## much better! summary(bpmInt2) summary(bpmInt1) } ### For comparison: A Daniel plot does not show any significant effects according ### to Lenths method, but makes the right effects stick out DanielPlot(plan.r, half=TRUE, alpha=1)
Functions to select elements or extract information from design catalogues of class catlg
res(catlg) ## S3 method for class 'catlg' res(catlg) ## S3 method for class 'character' res(catlg) nruns(catlg) ## S3 method for class 'catlg' nruns(catlg) ## S3 method for class 'character' nruns(catlg) nfac(catlg) ## S3 method for class 'catlg' nfac(catlg) ## S3 method for class 'character' nfac(catlg) WLP(catlg) ## S3 method for class 'catlg' WLP(catlg) ## S3 method for class 'character' WLP(catlg) nclear.2fis(catlg) ## S3 method for class 'catlg' nclear.2fis(catlg) ## S3 method for class 'character' nclear.2fis(catlg) clear.2fis(catlg) ## S3 method for class 'catlg' clear.2fis(catlg) ## S3 method for class 'character' clear.2fis(catlg) all.2fis.clear.catlg(catlg) dominating(catlg) ## S3 method for class 'catlg' dominating(catlg) ## S3 method for class 'character' dominating(catlg) catlg ## S3 method for class 'catlg' catlg[i] ## S3 method for class 'catlg' print(x, name="all", nruns="all", nfactors="all", res.min=3, MaxC2=FALSE, show=10, gen.letters=FALSE, show.alias=FALSE, ...) block.catlg
res(catlg) ## S3 method for class 'catlg' res(catlg) ## S3 method for class 'character' res(catlg) nruns(catlg) ## S3 method for class 'catlg' nruns(catlg) ## S3 method for class 'character' nruns(catlg) nfac(catlg) ## S3 method for class 'catlg' nfac(catlg) ## S3 method for class 'character' nfac(catlg) WLP(catlg) ## S3 method for class 'catlg' WLP(catlg) ## S3 method for class 'character' WLP(catlg) nclear.2fis(catlg) ## S3 method for class 'catlg' nclear.2fis(catlg) ## S3 method for class 'character' nclear.2fis(catlg) clear.2fis(catlg) ## S3 method for class 'catlg' clear.2fis(catlg) ## S3 method for class 'character' clear.2fis(catlg) all.2fis.clear.catlg(catlg) dominating(catlg) ## S3 method for class 'catlg' dominating(catlg) ## S3 method for class 'character' dominating(catlg) catlg ## S3 method for class 'catlg' catlg[i] ## S3 method for class 'catlg' print(x, name="all", nruns="all", nfactors="all", res.min=3, MaxC2=FALSE, show=10, gen.letters=FALSE, show.alias=FALSE, ...) block.catlg
catlg |
Catalogue of designs of class |
i |
vector of index positions or logical vector that can be used for indexing a |
x |
an object of class |
name |
character vector of entry names from |
nruns |
numeric integer (vector), giving the run size(s) for entries of |
nfactors |
numeric integer (vector), giving the factor number(s) for entries of |
res.min |
numeric integer giving the minimum resolution for entries of |
MaxC2 |
logical indicating whether designs are ordered by minimum aberration (default, |
show |
integer number indicating maximum number of designs to be shown; default is 10 |
gen.letters |
logical indicating whether the generators should be shown as
column numbers (default, |
show.alias |
logical indicating whether the alias structure (up to 2fis) is to be printed |
... |
further arguments to function |
block.catlg |
data frame with block generators for full factorial designs up to 256~runs, taken from Sun, Wu and Chen (1997) |
The class catlg
is a named list of design entries.
Each design entry is again a list with the following items:
resolution, numeric, i.e. 3 denotes resolution III and so forth
number of factors
number of runs
column numbers of additional factors in Yates order
word length pattern (starting with words of length 1,
i.e. the first two entries are 0 for all designs in catlg
)
number of clear 2-factor interactions (i.e. free of aliasing with main effects or other 2-factor interactions)
2xnclear.2fis
matrix of clear 2-factor interactions
(clear to be understood in the above sense); this matrix represents
each designs clear interaction graph, which can be used in automated searches for
designs that can accomodate (i.e. clearly) a certain requirement set
of 2-factor interactions; cf. also estimable.2fis
vector of factors with all 2-factor interactions clear in the above sense
logical that indicates whether the current design adds a
CIG structure that has not been seen for a design with less aberration
(cf. Wu, Mee and Tang 2012 p.196 for dominating designs);
TRUE, if so; FALSE, if current CIG is isomorphic to previous one or
has no edges (IMPORTANT: the dominance assessment refers to the current
catalogue; for designs with more than 64 runs, it is possible that
a design marked dominating in catalogue catlg
is not dominating
when considering ALL non-isomorphic designs).
This element is helpful in omitting non-promising
designs from a search for a clear design.
This element may be missing. In that case, all catalogue entries
are assumed dominating.
Reference to factors in components clear.2fis
and all.2fis.clear
is via their position number (element of (1:nfac)).
The print
function for class catlg
gives a concise overview of selected designs in any design catalogue of class catlg
.
It is possible to restrict attention to designs with certain run sizes, numbers of factors, and/or to request a minimum resolutions.
Designs are ordered in decreasing quality, where the default is aberration order, but number of clear 2fis can be requested alternatively.
The best 10 designs are displayed per default; this number can be changed by the show
option.
Options gen.letters
and show.alias
influence the style and amount of printed output.
The catalogue catlg
, which is included with package FrF2
,
is of class catlg
and is a living object, since it has to be updated with
recent research results for larger designs. In particular, new MA designs may be found,
or it may be proven that previous “good” designs are in fact of minimum aberration.
Currently, the catalogue contains
the Chen, Sun and Wu (1993) 2-level designs (complete list of 2-level fractional factorials from 4 to 32~runs, complete list of resolution IV 2-level fractional factorials with 64~runs). Note that the Chen Sun Wu paper only shows a selection of the 64~run designs, the complete catalogue has been obtained from Don Sun and is numbered according to minimum aberration (lower number = better design); numbering in the paper is not everywhere in line with this numbering.
minimum aberration (MA) resolution III designs for 33 to 63 factors in 64 runs. The first few of these have been obtained from Appendix G of Mee 2009, the designs for 38 and more factors have been constructed by combining a duplicated minimum aberration design in 32 runs and the required number of factors with columns 32 to 63 of the Yates matrix for 64 run designs. Using complementary design theory (cf. e.g. Chapter 6.2.2 in Mee 2009), it can be shown that the resulting designs are minimum aberration (because they are complementary to basically the same designs as the designs in 32 runs on which they are based). The author is grateful to Robert Mee for pointing this out.
the MA designs in 128 runs:
for up to 24 factors obtained from Xu (2009),
for 25 to 64 factors taken from Block and Mee (2005, with corrigendum 2006),
for 65 to 127 factors (resolution III): up to 69 factors coming from Appendix G in Mee, whereas the designs for 70 or more factors have been constructed according to the same principle mentioned for the 64 run designs.
various further “good” resolution IV designs in 128 runs obtained by evaluating designs from the complete catalogue by Xu (2009, catalogue on his website) w.r.t. aberration and number of clear 2fis (including also all designs that yield minimum aberration clear compromise designs according to Groemping 2010); all designs with resolution at least IV for up to 11 factors have been added with version 2.2.
the MA even designs in 128 runs, in support of blocking according to the Godolphin approach have been added with version 2.2.
Note that additional non-isomorphic resolution IV designs in 128 runs are available in package (FrF2.catlg128); since the catalogues are quite large, they are not forced upon users of this package who do not need them. Since version 1.1 of that package, the catalogues are not complete but contain high resolution fractions and even/odd fractions only (status: version 1.2-x); re-inclusion of at least selected even fractions is intended, because these may yield improved support of blocking in connection with estimable 2fis.
the best (MA) resolution IV or higher designs in
256 runs for up to 36 factors (resolution V up to 17 factors),
512 runs for up to 29 factors (resolution V for up to 23 factors).
These have been taken from Xu (2009) with additions by Ryan and Bulutoglu
(2010).
Further “good” resolution IV designs with up to 80 factors in 256 runs and up to 160 factors in 512 runs have also been implemented from Xu (2009).
the best (MA) resolution V or higher design for each number of factors or
a “good” such design (if it is not known which one is best) in
1024 runs (up to 33 factors, MA up to 28 factors, resolution VI up to 24 factors),
2048 runs (up to 47 factors, MA up to 32 factors, resolution VI up to 34 factors),
and 4096 runs (up to 65 factors, MA up to 26 factors, resolution VI up to 48 factors).
Most of the large designs in catlg
have been taken from Xu (2009),
where complete catalogues of some scenarios are provided
(cf. also his website) as well as “good” (not necessarily MA) designs for a larger
set of situations. Some of the good designs by Xu (2009) have later been shown
to be MA by Ryan and Bulutoglu (2010), who also found some additional larger MA designs,
which are also included in catlg
. Non-MA designs that
were already available before Bulutoglu (2010) are still in the catalogue with their old name.
(Note that designs that are not MA and cannot be placed in the ranking do not
have a running number in the design name; for example, the MA 2048 runs
design in 28 factors is named 28-17.1, the older previous design
that was not MA is named 28-17 (without “.1” or another placement,
because the designs position in the ranking of all designs is not known.))
There are also some non-regular 2-level fractional factorial designs of resolution V
which may be interesting, as it is possible to increase the number of factors for which
resolution V is possible (cf. Mee 2009, Chapter 8).
These are part of package DoE.base, which is automatically
loaded with this package. With versions higher than 0.9-14 of that package,
the following arrays are available: L128.2.15.8.1
, which allows 4 additional factors and blocking into up to 8 blocksL256.2.19
, which allows just 2 additonal factorsL2048.2.63
, which allows 16 additional factors.
These non-regular arrays should be fine for most purposes; the difference to the arrays generated
by function FrF2
lies in the fact that there is partial aliasing, e.g. between 3-factor interactions
and 2-factor interactions. This means that an affected 3-factor interaction is
partially aliased with several different
2-factor interactions rather than being aliased either fully or not at all.
[
selects a subset of designs based on i
, which is again a list of class catlg
, even if a single element is selected.
res
, nruns
, nfac
, nclear.2fis
and dominating
return a named vector,
the print
method does not return anything (i.e. it returns NULL
), and
the remaining functions return a list.
Ulrike Groemping
Block, R. and Mee, R. (2005) Resolution IV Designs with 128 Runs Journal of Quality Technology 37, 282-293.
Block, R. and Mee, R. (2006) Corrigenda Journal of Quality Technology 38, 196.
Chen, J., Sun, D.X. and Wu, C.F.J. (1993) A catalogue of 2-level and 3-level orthogonal arrays. Int. Statistical Review 61, 131-145.
Groemping, U. (2012). Creating clear designs: a graph-based algorithm and a catalog of clear compromise plans. IIE Transactions 44, 988-1001. doi:10.1080/0740817X.2012.654848. Early preprint at http://www1.bht-berlin.de/FB_II/reports/Report-2010-005.pdf.
Mee, R. (2009). A Comprehensive Guide to Factorial Two-Level Experimentation. New York: Springer.
Ryan, K.J. and Bulutoglu, D.A. (2010). Minimum Aberration Fractional Factorial Designs With Large N. Technometrics 52, 250-255.
Sun, D.X., Wu, C.F.J. and Chen, Y.Y. (1997).
Optimal blocking schemes for and
designs. Technometrics 39,
298-307.
Wu, H., Mee, R. and Tang, B. (2012). Fractional Factorial Designs With Admissible Sets of Clear Two-Factor Interactions. Technometrics 54, 191-197.
Xu, H. (2009) Algorithmic Construction of Efficient Fractional Factorial Designs With Large Run Sizes. Technometrics 51, 262-277.
c8 <- catlg[nruns(catlg)==8] nclear.2fis(c8) clear.2fis(c8) all.2fis.clear.catlg(c8) ## inspecting a specific catalogue element clear.2fis("9-4.2") ## usage of print function for inspecting catalogued designs ## the first 10 resolution V+ designs in catalogue catlg print(catlg, res.min=5) ## the 10 resolution V+ designs in catalogue catlg with the most factors ## (for more than one possible value of nfactors, MaxC2 does usually not make sense) print(catlg, res.min=5, MaxC2=TRUE) ## designs with 12 factors in 64 runs (minimum resolution IV because ## no resolution III designs of this size are in the catalogue) ## best 10 aberration designs print(catlg, nfactors=12, nruns=64) ## best 10 clear 2fi designs print(catlg, nfactors=12, nruns=64, MaxC2=TRUE) ## show alias structure print(catlg, nfactors=12, nruns=64, MaxC2=TRUE, show.alias=TRUE) ## show best 20 designs print(catlg, nfactors=12, nruns=64, MaxC2=TRUE, show=20) ## use vector-valued nruns print(catlg, nfactors=7, nruns=c(16,32)) ## all designs (as show=100 is larger than available number of designs) ## with 7 or 8 factors in 16 runs print(catlg, nfactors=c(7,8), nruns=16, show=100) ## the irregular resolution V arrays from package DoE.base (from version 0.9-17) ## designs can be created from them using function oa.design ## Not run: ## not run in case older version of DoE.base does not have these length3(L128.2.15.8.1) length4(L128.2.15.8.1) ## aliasing of 2fis with block factor length4(L128.2.15.8.1[,-16]) length3(L256.2.19) length4(L256.2.19) ##length3(L2048.2.63) ##length4(L2048.2.63) do not work resource wise ## but the array is also resolution V (but irregular) ## End(Not run)
c8 <- catlg[nruns(catlg)==8] nclear.2fis(c8) clear.2fis(c8) all.2fis.clear.catlg(c8) ## inspecting a specific catalogue element clear.2fis("9-4.2") ## usage of print function for inspecting catalogued designs ## the first 10 resolution V+ designs in catalogue catlg print(catlg, res.min=5) ## the 10 resolution V+ designs in catalogue catlg with the most factors ## (for more than one possible value of nfactors, MaxC2 does usually not make sense) print(catlg, res.min=5, MaxC2=TRUE) ## designs with 12 factors in 64 runs (minimum resolution IV because ## no resolution III designs of this size are in the catalogue) ## best 10 aberration designs print(catlg, nfactors=12, nruns=64) ## best 10 clear 2fi designs print(catlg, nfactors=12, nruns=64, MaxC2=TRUE) ## show alias structure print(catlg, nfactors=12, nruns=64, MaxC2=TRUE, show.alias=TRUE) ## show best 20 designs print(catlg, nfactors=12, nruns=64, MaxC2=TRUE, show=20) ## use vector-valued nruns print(catlg, nfactors=7, nruns=c(16,32)) ## all designs (as show=100 is larger than available number of designs) ## with 7 or 8 factors in 16 runs print(catlg, nfactors=c(7,8), nruns=16, show=100) ## the irregular resolution V arrays from package DoE.base (from version 0.9-17) ## designs can be created from them using function oa.design ## Not run: ## not run in case older version of DoE.base does not have these length3(L128.2.15.8.1) length4(L128.2.15.8.1) ## aliasing of 2fis with block factor length4(L128.2.15.8.1[,-16]) length3(L256.2.19) length4(L256.2.19) ##length3(L2048.2.63) ##length4(L2048.2.63) do not work resource wise ## but the array is also resolution V (but irregular) ## End(Not run)
Function CIG creates a clear interactions graph (CIG) from a catlg design (design name must be given). Function CIGstatic allows to create a static graph from a dynamically-adjusted one.
CIG(design, select.catlg = catlg, nfac = NULL, static = FALSE, layout = layout.auto, label = "num", plot = TRUE, ...) CIGstatic(graph, id, label = "num", xlim = c(-1,1), ylim = c(1,-1), ...) gen2CIG(nruns, gen)
CIG(design, select.catlg = catlg, nfac = NULL, static = FALSE, layout = layout.auto, label = "num", plot = TRUE, ...) CIGstatic(graph, id, label = "num", xlim = c(-1,1), ylim = c(1,-1), ...) gen2CIG(nruns, gen)
design |
a character string that identifies a design in the catalogue specified
by option |
select.catlg |
name of catalogue (not in quotes);
only relevant, if |
nfac |
number of factors; this is not needed for a class |
static |
logical. If |
layout |
ignored for |
label |
in effect for |
plot |
a logical that decides whether a plot is requested (default: |
... |
further arguments to be passed to function |
graph |
a graph object from package |
id |
identification number of the interactive graph to be reproduced in static form; this number can be found in the header line of the graphics window |
xlim |
horizontal axis limits |
ylim |
vertical axis limits (per default reversed in order to exactly reproduce the interactive graph) |
nruns |
number of runs of the design to be graphed |
gen |
generator (vector of Yates matrix column numbers) |
The design depicted in CIG
has to be the name (character string) of a
regular fractional factorial 2-level design present in select.catlg
.
Clear 2fis are depicted as edges in the graph. In the interactive graph, users can change the layout manually or with the menus. For example, the Reingold-Tilford layout can be chosen, with a root vertex specified; this sometimes helps in identifying groups of vertices that are not connected with each other.
Previous versions of package igraph
used to internally number the
vertices from 0 to number of vertices -1, not from 1 to number of vertices.
This has been changed
in June 2012 (FrF2 adapted to this change with version 1.5).
Function CIGstatic
serves the purpose to statically create the current
interactively modified graph;
the usual annotation possibilities for plots are available.
Function gen2CIG
returns a graph object that can be plotted or otherwise
investigated with graph-related
functionality.
For plot=FALSE
or plot=TRUE
with static=TRUE
,
function CIG
visibly (plot=FALSE
) or invisibly (plot=TRUE
)
returns a graph from package igraph
.
For plot=TRUE
with static=FALSE
,
the function returns a list with the first element graph
the element coords
with the coordinates of that graph.
Function CIGstatic
works on the list produced by function CIG
by plotting the graph statically using the positioning from the
current interactive picture.
Function gen2CIG
returns a clear interactions graph that can e.g. be
plotted with functions plot
(plot.igraph
) or
tkplot
.
Ulrike Groemping
Groemping, U. (2012). Creating clear designs: a graph-based algorithm and a catalog of clear compromise plans. IIE Transactions 44, 988-1001. doi:10.1080/0740817X.2012.654848. Early preprint at http://www1.bht-berlin.de/FB_II/reports/Report-2010-005.pdf.
plot.igraph
, tkplot
,
plot.common
## Not run: ex.CIG <- CIG("9-4.2", vertex.color="white", vertex.label.color="darkred") ## play around with the dynamic graph until it looks right ## look up its id number in the title bar of the graph window and use it for id par(xpd=TRUE) CIGstatic(ex.CIG, id=1) ## End(Not run) graph1 <- CIG("9-4.2", plot=FALSE) ### create graph object from design name ### calculate graph properties require(igraph) degree(graph1) clique.number(graph1) independence.number(graph1) largest.cliques(graph1) graph2 <- gen2CIG(32, c(7,11,14,29)) ### create graph object from generator columns ### check isomorphism to graph1 graph.isomorphic(graph1, graph2) ## Not run: ## use a CIG for manual design search ## requirement set: estim <- compromise(9, 8:9)$requirement ## all interactions of factors 8 and 9 (H, J) ## graph the requirement set CIG CIG(estim, vertex.color="white", vertex.label.color="darkred") ## a human can easily see that columns 1, 8 and 9 are worth a try for factors P, Q and R CIG("9-4.1", vertex.color="white", vertex.label.color="darkred") ## obviously, 9-4.1 cannot accommodate the requirement set CIG("9-4.2", vertex.color="white", vertex.label.color="darkred") ## 9-4.2 can, by assigning factors H and J to columns 5 and 9 ## function FrF2 automatically does such matchings ## End(Not run)
## Not run: ex.CIG <- CIG("9-4.2", vertex.color="white", vertex.label.color="darkred") ## play around with the dynamic graph until it looks right ## look up its id number in the title bar of the graph window and use it for id par(xpd=TRUE) CIGstatic(ex.CIG, id=1) ## End(Not run) graph1 <- CIG("9-4.2", plot=FALSE) ### create graph object from design name ### calculate graph properties require(igraph) degree(graph1) clique.number(graph1) independence.number(graph1) largest.cliques(graph1) graph2 <- gen2CIG(32, c(7,11,14,29)) ### create graph object from generator columns ### check isomorphism to graph1 graph.isomorphic(graph1, graph2) ## Not run: ## use a CIG for manual design search ## requirement set: estim <- compromise(9, 8:9)$requirement ## all interactions of factors 8 and 9 (H, J) ## graph the requirement set CIG CIG(estim, vertex.color="white", vertex.label.color="darkred") ## a human can easily see that columns 1, 8 and 9 are worth a try for factors P, Q and R CIG("9-4.1", vertex.color="white", vertex.label.color="darkred") ## obviously, 9-4.1 cannot accommodate the requirement set CIG("9-4.2", vertex.color="white", vertex.label.color="darkred") ## 9-4.2 can, by assigning factors H and J to columns 5 and 9 ## function FrF2 automatically does such matchings ## End(Not run)
Addelman (1962) and Ke and Wu (2005) discuss compromise plans of different types. Their creation is supported by the function compromise.
compromise(nfactors, G1, class=3, msg=TRUE)
compromise(nfactors, G1, class=3, msg=TRUE)
nfactors |
overall number of factors |
G1 |
vector with indices of factors in group G1 (cf. details) |
class |
class of compromise designs that is to be generated; 1, 2, 3, or 4, cf. details below |
msg |
logical stating whether the |
For compromise plans, the factors are decomposed into a group G1 and a group G2.
The different classes of compromise plans require estimability of different subsets
of 2fis in addition to main effects:
Class 1: all 2fis within group G1 are estimable
Class 2: all 2fis within group G1 are estimable,
as well as all 2fis within group G2
Class 3: all 2fis within group G1 are estimable,
as well as all 2fis between groups G1 and G2
Class 4: all 2fis between groups G1 and G2 are estimable
The function returns a list of four components (cf. section “Value”).
They can be used as input for the function FrF2
, if compromise
plans are to be created. Both distinct designs (Addelman 1962) and clear designs
(Ke, Tang and Wu 2005) can be constructed,
depending on the settings of option clear
in function
FrF2
. More explanations on specifying estimability requirements
for 2fis in general are provided under estimable.2fis
.
Value is a list of the four components perms.full
, requirement
,
class
, and minnrun.clear
. The last two components are purely imformative,
while the first two provide input parameters for function FrF2
.requirement
can be used for specifying the required 2fis in the estimable
option,
both with clear=FALSE
and clear=TRUE
.
For clear=FALSE
, perms.full
can be used in the perms
option
for speeding up the search into a hopefully realistic time frame.minnrun.clear
indicates the minimum number of runs needed for a clear design.
Note that the catalogue catlg
contains all designs needed for
accomodating existing clear compromise designs in up to 128 runs (even minimum aberration
among all existing clear compromise designs; for a catalogue of these, cf. Gr\"omping 2010).
Ulrike Groemping
Addelman, S. (1962). Symmetrical and asymmetrical fractional factorial plans. Technometrics 4, 47-58.
Groemping, U. (2012). Creating clear designs: a graph-based algorithm and a catalog of clear compromise plans. IIE Transactions 44, 988-1001. doi:10.1080/0740817X.2012.654848. Early preprint at http://www1.bht-berlin.de/FB_II/reports/Report-2010-005.pdf.
Ke, W., Tang, B. and Wu, H. (2005). Compromise plans with clear two-factor interactions. Statistica Sinica 15, 709-715.
See Also FrF2
for creation of regular fractional factorial designs
as well as estimable.2fis
for statistical and algorithmic information on estimability of 2-factor interactions
## seven factors two of which are in group G1 C1 <- compromise(7, c(2,4), class=1) C1$perms.full ## the same for all classes C1$requirement C2 <- compromise(7, c(2,4), class=2) C2$requirement C3 <- compromise(7, c(2,4), class=3) C3$requirement C4 <- compromise(7, c(2,4), class=4) C4$requirement ## Not run: ########## usage of estimable ########################### ## design with with BD clear in 16 runs FrF2(16,7,estimable = C1$requirement) ## design with BD estimable on a distinct column in 16 runs (any design will do, ## if resolution IV!!! FrF2(16,7,estimable = C1$requirement, clear=FALSE, perms=C1$perms.full) ## all four classes, mostly clear, for 32 runs FrF2(32,7,estimable = C1$requirement) FrF2(32,7,estimable = C2$requirement) ## requires resolution V ## as clear class 2 compromise designs do not exist due to Ke et al. 2005 FrF2(32,7,estimable = C2$requirement, clear=FALSE, perms=C2$perms.full) FrF2(32,7,estimable = C3$requirement) FrF2(32,7,estimable = C4$requirement) ## two additional factors H and J that do not show up in the requirement set FrF2(32,9,estimable = C3$requirement) ## two additional factors H and J that do not show up in the requirement set FrF2(32,9,estimable = C3$requirement, clear=FALSE) ## note that this is not possible for distinct designs in case perms is needed, ## because perms must have nfactors columns ## End(Not run)
## seven factors two of which are in group G1 C1 <- compromise(7, c(2,4), class=1) C1$perms.full ## the same for all classes C1$requirement C2 <- compromise(7, c(2,4), class=2) C2$requirement C3 <- compromise(7, c(2,4), class=3) C3$requirement C4 <- compromise(7, c(2,4), class=4) C4$requirement ## Not run: ########## usage of estimable ########################### ## design with with BD clear in 16 runs FrF2(16,7,estimable = C1$requirement) ## design with BD estimable on a distinct column in 16 runs (any design will do, ## if resolution IV!!! FrF2(16,7,estimable = C1$requirement, clear=FALSE, perms=C1$perms.full) ## all four classes, mostly clear, for 32 runs FrF2(32,7,estimable = C1$requirement) FrF2(32,7,estimable = C2$requirement) ## requires resolution V ## as clear class 2 compromise designs do not exist due to Ke et al. 2005 FrF2(32,7,estimable = C2$requirement, clear=FALSE, perms=C2$perms.full) FrF2(32,7,estimable = C3$requirement) FrF2(32,7,estimable = C4$requirement) ## two additional factors H and J that do not show up in the requirement set FrF2(32,9,estimable = C3$requirement) ## two additional factors H and J that do not show up in the requirement set FrF2(32,9,estimable = C3$requirement, clear=FALSE) ## note that this is not possible for distinct designs in case perms is needed, ## because perms must have nfactors columns ## End(Not run)
A cube plot for the combined effect of three factors is produced (function cubePlot). Functions cubedraw, cubecorners, cubelabel and myscatterplot3d are not intended for users.
cubePlot(obj, eff1, eff2, eff3, main=paste("Cube plot for",respnam), cex.title=1.5,cex.lab=par("cex.lab"), cex.ax=par("cex.axis"), cex.clab=1.2, size=0.3, round=NULL, abbrev=4,y.margin.add=-0.2, modeled=TRUE)
cubePlot(obj, eff1, eff2, eff3, main=paste("Cube plot for",respnam), cex.title=1.5,cex.lab=par("cex.lab"), cex.ax=par("cex.axis"), cex.clab=1.2, size=0.3, round=NULL, abbrev=4,y.margin.add=-0.2, modeled=TRUE)
obj |
a vector of response values to be analyzed OR a linear model object with 2-level factors or numerical 2-level variables (CAUTION: numerical x-variable have to be coded as -1 and +1 only!); the structure must be such that effects are either fully aliased or orthogonal, like in a fractional factorial 2-level design |
eff1 |
cf. |
eff2 |
cf. |
eff3 |
effects to be included in the cube plot (x-, y-, z-direction),
EITHER vectors of equal length (two-level factors or numerical
with the two values -1 and 1)
OR variable names of main effects within the |
main |
title for the plot, |
cex.title |
multiplier for size of overall title
( |
cex.ax |
size of axis tick marks, defaults to |
cex.lab |
size of axis labels |
cex.clab |
size of corner labels |
size |
size of cube corners |
round |
optional rounding of corner labels (digits argument for function
|
abbrev |
number of characters shown for factor levels |
y.margin.add |
adjustment parameter for placement of y-axis labeling |
modeled |
TRUE (default: show modeled means; FALSE: show averages NOTE: Even when showing modeled means, there also appears to be a three-factor-interaction, if the model contains an effect that is aliased with this interaction! |
cubePlot
produces a cube plot of the modeled means or averages of
all combinations for three factors. The other functions are internal
and are called by cubePlot.
myscatterplot3d
is a modified version of scatterplot3d,
made more suitable for this situation.
cubePlot
is used for its side effects only.
Ulrike Groemping
Box G. E. P, Hunter, W. C. and Hunter, J. S. (2005) Statistics for Experimenters, 2nd edition. New York: Wiley.
FrF2-package
for examples
The function is modified from the same-name function in packge BsMD with the purpose of providing more usage comfort (correct effect sizes in case of factors, automatic annotation, automatic labelling of the most significant factors only).
DanielPlot(fit, ...) ## S3 method for class 'design' DanielPlot(fit, ..., response = NULL) ## Default S3 method: DanielPlot(fit, code = FALSE, autolab = TRUE, alpha = 0.05, faclab = NULL, block = FALSE, datax = TRUE, half = FALSE, pch = "*", cex.fac = par("cex.lab"), cex.lab = par("cex.lab"), cex.pch = par("cex"), cex.legend = par("cex.lab"), main = NULL, subtitle=NULL, ...)
DanielPlot(fit, ...) ## S3 method for class 'design' DanielPlot(fit, ..., response = NULL) ## Default S3 method: DanielPlot(fit, code = FALSE, autolab = TRUE, alpha = 0.05, faclab = NULL, block = FALSE, datax = TRUE, half = FALSE, pch = "*", cex.fac = par("cex.lab"), cex.lab = par("cex.lab"), cex.pch = par("cex"), cex.legend = par("cex.lab"), main = NULL, subtitle=NULL, ...)
fit |
an experimental design of class |
... |
further arguments to be passed to the default function,
or graphical parameters to be passed to |
response |
NULL or a character string that specifies response variable to be used,
must be an element of |
code |
logical. If |
autolab |
If TRUE, only the significant factors according to the Lenth method
(significance level given by |
alpha |
significanc level for the Lenth method |
faclab |
NULL or list.
If |
block |
logical. If |
datax |
logical. If |
half |
logical. If |
pch |
numeric or character. Points character. |
cex.fac |
numeric. Factor label character size. |
cex.lab |
numeric. Labels character size. |
cex.pch |
numeric. Points character size. |
cex.legend |
numeric. Legend size in case of codes. |
main |
NULL or character. Title of plot. If NULL, automatic title is generated. |
subtitle |
NULL or character. Sub title of plot. Should not be used for split-plot designs, because automatic subtitle is generated for these. |
The design underlying fit
has to be a (regular or non-regular) fractional factorial 2-level design.
Effects (except for the intercept) are displayed in a normal or half-normal
plot with the effects in the x-axis by default.
If fit
is a design with at least one response variable
rather than a linear model fit,
the lm
-method for class design
is applied to it with
degree high enough that at least one effect is assigned to each column of the Yates matrix,
and the default method for DanielPlot
is afterwards applied to the
resulting linear model.
For split-plot designs, whole plot effects are shown as different plotting characters, because they are potentially subject to larger variability, and one should not be too impressed, if they look impressively large, as this may well be indication of plot-to-plot variability rather than a true effect.
The function invisibly returns a data frame with columns: x
, y
,
no
, effect
, coded
(if coded plot was requested)
and pchs
, for the coordinates, the position numbers,
the effect names, the coded effect names, and the plotting characters
for plotted points.
The plotting characters are particularly useful for split-plot designs and can be used for subsequent separate plotting of whole-plot and split-plot effects, if necessary.
If you load package BsMD after package FrF2,
a mere call to function DanielPlot
will use the function from package BsMD
rather than the one from package FrF2. You can explicitly request
usage of the FrF2 function by FrF2::DanielPlot
.
Ernesto Barrios, modified by Ulrike Groemping.
Box G. E. P, Hunter, W. C. and Hunter, J. S. (2005) Statistics for Experimenters, 2nd edition. New York: Wiley.
Daniel, C. (1959) Use of Half Normal Plots in Interpreting Two Level Experiments. Technometrics 1, 311–340.
Daniel, C. (1976) Application of Statistics to Industrial Experimentation. New York: Wiley.
Lenth, R.V. (1989) Quick and easy analysis of unreplicated factorials. Technometrics 31, 469–473.
Lenth, R.V. (2006) Lenth s Method for the Analysis of Unreplicated Experiments. To appear in Encyclopedia of Statistics in Quality and Reliability, Wiley, New York. Downloadable at http://www.wiley.com/legacy/wileychi/eqr/docs/sample_1.pdf.
qqnorm
, halfnormal
,
LenthPlot
, BsMD-package
This help page documents the statistical and algorithmic details of requesting 2-factor interactions to be estimable in FrF2
The option estimable
allows to specify 2-factor interactions (2fis) that
have to be estimable in the model. Whenever a resolution V or higher design is available,
this option is unnecessary, because all 2fis are estimable in the sense that they are
not aliased with any main effect or any other 2fi. If resolution V or higher is not affordable,
the option estimable
can ensure that certain 2fis can nevertheless be estimated.
Per default, it is assumed that a resolution IV design is required,
as it is normally not reasonable to allow main effects to be
aliased with other 2-factor interactions in this situation. There are two types of
estimability that are distinguished by the setting of option clear
in
function FrF2
(cf. Groemping 2010).
Let us first consider designs of at least resolution IV.
With option clear=TRUE
, FrF2
searches for
a model for which all main effects and all 2fis given in estimable
are
clear of aliasing with any other 2fis. This is a weaker requirement than resolution V,
because 2fis outside those specified in estimable
may be aliased with
each other. But it is much stronger than what is done in case of clear=FALSE
:
For the latter, FrF2
searches for a design that has a distinct column in
the model matrix for each main effect and each interaction requested
in estimable
.
Users can explicitly permit that resolution III designs are included in the
search of designs for which the specified 2fis are estimable (by the res3=TRUE
option).
In case of clear=TRUE
, this leads to the somewhat
strange situation that main effects can be aliased with 2fis from outside
estimable
while 2fis from inside estimable
are not aliased with
any main effects or 2fis.
With clear=TRUE
, the algorithm compares the requirement set to
catalogued sets of clear 2fis by a graph isomorphism algorithm from R-package
igraph. For details of this algorithm,
cf. Groemping (2012). With the catalogue catlg
available in this package,
the best (minimum aberration) existing clear designs are guaranteed to be found
for up to 64 runs and have a good chance to be found for 128 runs. For 128 runs,
it is possible to load an additional large catalogue (package FrF2.catlg128)
in order to also guarantee that the best clear design is found. For 256 and 512 runs,
only one or two resolution IV designs of each size are catalogued so that
option estimable
can try to influence allocation of factors to columns,
but may fail although an appropriate clear design would exist outside the catalogued
designs.
The search for a clear design
is often fast. If it isn't, option sort
of function FrF2
can help. For the occasional situation where
this doesn't help either, a manual search may help, see CIG
for an example of how to proceed.
Since version 2 of package FrF2, requesting 2fis to be clear is compatible
with blocking a design. The algorithm behind that functionality is based on
Godolphin (2021) and is described in Groemping (2021).
The default implementation strives for a guartanteed and best possible result.
Arguments firsthit
and useV
to function FrF2
can be used for trying to obtain a possibly not best result (firsthit
)
faster or to use a (sometimes) faster algorithm that is not guaranteed
to deliver a result even though it might exist for resolution IV situations
(useV=FALSE
).
With clear=FALSE
, the algorithm loops through the eligible designs from
catlg.select
from good to worse (in terms of MA) and, for each design, loops
through all eligible permutations of the experiment factors from perms
.
If perms
is omitted, the permutations are looped through in lexicographic
order starting from 1:nfac or perm.start
. Especially in this case,
run times of the search algorithm can be very long.
The max.time
option allows to limit this run time.
If the time limit is reached, the final situation (catalogued design and
current permutation of experiment factors) is printed so that the user can
decide to proceed later with this starting point (indicated by catlg.select
for the catalogued design(s) to be used and perm.start
for the current
permutation of experiment factors).
With clear=TRUE
, the algorithm loops through the eligible designs from
catlg.select
from good to worse (in terms of MA) and, for each design,
uses a subgraph isomorphism check from package igraph
. There are two such
algorithms, VF2 (the default, Cordella et al. 2001) and LAD (introduced with
version 1.7 of package FrF2, Solnon 2010),
which can be chosen with the method
option.
Run times of the subgraph isomorphism search are often fast,
but can also be very very slow in unlucky situations.
Where the VF2 algorithm is particularly slow, the LAD algorithm is often fast
(see Groemping 2014b).
Especially for the VF2 algorithm, run times may strongly depend on the ordering
of factors, which can be influenced by the option sort
.
As the slowness of the process is intrinsic to the subgraph isomorphism
search problem (which is NP-complete), a max.time
option analogous to
the clear=FALSE
situation would be of very limited use only and is
therefore not available. Instead, it is possible to have a look at the
number of the design that was in the process of being searched when the
process was interrupted (with the command FrF2.currentlychecked()
).
Note that - according to the structure of the catalogued designs and the lexicographic
order of checking permutations - the initial order of the factors has a strong influence
on the run time for larger or unlucky problems. For example, consider
an experiment in 32~runs and 11~factors, for six of which the pairwise interactions are to be estimable
(Example 1 in Wu and Chen 1992). estimable
for this model can be specified as formula("~(F+G+H+J+K+L)^2")
OR formula("~(A+B+C+D+E+F)^2")
.
The former runs a lot faster than the latter (I have not yet seen the latter finish
the first catalogued design, if perms
is not specified).
The reason is that the latter needs more permutations of the experiment factors than
the former, since the factors with high positions
change place faster and more often than those with low positions.
For this particular design, it is very advisable to constrain the permutations of the experiment factors to the different subset selections of six factors from eleven, since permutations within the sets do not change the possibility of accomodating a design. The required permutations for the second version of this example can be obtained e.g. by the following code:
perms.6 <- combn(11,6)
perms.full <- matrix(NA,ncol(perms.6),11)
for (i in 1:ncol(perms.6))
perms.full[i,] <- c(perms.6[,i],setdiff(1:11,perms.6[,i]))
Handing perms.full to the procedure using the perms
option makes the second version of the
requested interaction terms fast as well, since up to almost 40 Mio permutations of experiment
factors are reduced to at most 462. Thus, whenever possible,
one should try to limit the permutations necessary in case of clear=FALSE
.
In order to support relatively comfortable creation of distinct designs of some frequently-used types
of required interaction patterns, the function compromise
has been
divised: it supports creation of the so-called compromise plans of classes 1 to 4 (cf.
e.g. Addelman 1962; Ke, Tang and Wu 2005; Groemping 2012).
The list it returns also contains a component perms.full
that can be used as input
for the perms
option.
Please contact me with any suggestions for improvements.
Ulrike Groemping
Addelman, S. (1962). Symmetrical and asymmetrical fractional factorial plans. Technometrics 4, 47-58.
Chen, J., Sun, D.X. and Wu, C.F.J. (1993). A catalogue of 2-level and 3-level orthogonal arrays. International Statistical Review 61, 131-145.
Cordella, L.P., Foggia, P., Sansone, C. and Vento, M. (2001). An improved algorithm for matching large graphs. Proc. of the 3rd IAPR TC-15 Workshop on Graphbased Representations in Pattern Recognition, 149–159.
Godolphin, J. (2021). Construction of Blocked Factorial Designs to Estimate Main Effects and Selected Two-Factor Interactions. J. Royal Statistical Society B 83, 5-29. doi:10.1111/rssb.12397.
Groemping, U. (2010). “Clear” and “Distinct”: two approaches for regular fractional factorial designs with estimability requirements. Reports in Mathematics, Physics and Chemistry, report 02/2010, Department II, Beuth University of Applied Sciences Berlin. http://www1.bht-berlin.de/FB_II/reports/Report-2010-002.pdf.
Groemping, U. (2012). Creating clear designs: a graph-based algorithm and a catalogue of clear compromise plans. IIE Transactions 44, 988–1001. Early preprint available at http://www1.bht-berlin.de/FB_II/reports/Report-2010-005.pdf.
Groemping, U. (2014a). R Package FrF2 for Creating and Analyzing Fractional Factorial 2-Level Designs. Journal of Statistical Software, 56, Issue 1, 1-56. https://www.jstatsoft.org/v56/i01/.
Groemping, U. (2014b). A Note on Dominating Fractional Factorial Two-Level Designs With Clear Two-Factor Interactions. Technometrics 56, 42–45.
Groemping, U. (2021). An algorithm for blocking regular fractional factorial 2-level designs with clear two-factor interactions. Computational Statistics and Data Analysis 153, 1-18. doi:10.1016/j.csda.2020.107059. Preprint at Report 3/2019.
Ke, W., Tang, B. and Wu, H. (2005). Compromise plans with clear two-factor interactions. Statistica Sinica 15, 709-715.
Solnon, C. (2010). AllDifferent-based Filtering for Subgraph Isomorphism. Artificial Intelligence 174, 850–864.
Wu, C.F.J. and Chen, Y. (1992) A graph-aided method for planning two-level experiments when certain interactions are important. Technometrics 34, 162-175.
See also FrF2
for regular fractional factorials,
catlg
for the Chen, Sun, Wu (1993) and larger catalogues of designs
and some accessor functions, and function compromise
for a convenience
function to handle estimability requests for compromise plans
########## usage of estimable ########################### ## design with all 2fis of factor A estimable on distinct columns in 16 runs FrF2(16, nfactors=6, estimable = rbind(rep(1,5),2:6), clear=FALSE) FrF2(16, nfactors=6, estimable = c("AB","AC","AD","AE","AF"), clear=FALSE) FrF2(16, nfactors=6, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=FALSE) ## formula would also accept self-defined factor names ## from factor.names instead of letters A, B, C, ... ## estimable does not need any other input FrF2(estimable=formula("~(A+B+C)^2+D+E")) ## estimable with factor names ## resolution three must be permitted, as FrF2 first determines that 8 runs ## would be sufficient degrees of freedom to estimate all effects ## and then tries to accomodate the 2fis from the model clear of aliasing in 8 runs FrF2(estimable=formula("~one+two+three+four+two:three+two:four"), factor.names=c("one","two","three","four"), res3=TRUE) ## clear=FALSE allows to allocate all effects on distinct columns in the ## 8 run MA resolution IV design FrF2(estimable=formula("~one+two+three+four+two:three+two:four"), factor.names=c("one","two","three","four"), clear=FALSE) ## 7 factors instead of 6, but no requirements for factor G FrF2(16, nfactors=7, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=FALSE) ## larger design for handling this with all required effects clear FrF2(32, nfactors=7, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=TRUE) ## 16 run design for handling this with required 2fis clear, but main effects aliased ## (does not usually make sense) FrF2(16, nfactors=7, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=TRUE, res3=TRUE) ## example for necessity of perms for the clear=FALSE case ## based on Wu and Chen Example 1 ## Not run: ## runs per default about max.time=60 seconds, before throwing error with ## interim results ## results could be used in select.catlg and perm.start for restarting with ## calculation of further possibilities FrF2(32, nfactors=11, estimable = formula("~(A+B+C+D+E+F)^2"), clear=FALSE) ## would run for a long long time (I have not yet been patient enough) FrF2(32, nfactors=11, estimable = formula("~(A+B+C+D+E+F)^2"), clear=FALSE, max.time=Inf) ## End(Not run) ## can be easily done with perms, ## as only different subsets of six factors are non-isomorphic perms.6 <- combn(11,6) perms.full <- matrix(NA,ncol(perms.6),11) for (i in 1:ncol(perms.6)) perms.full[i,] <- c(perms.6[,i],setdiff(1:11,perms.6[,i])) ## function compromise will calculate the necessary perms entries automatically compromise(11,1:6)$perms.full FrF2(32, nfactors=11, estimable = formula("~(A+B+C+D+E+F)^2"), clear=FALSE, perms = perms.full )
########## usage of estimable ########################### ## design with all 2fis of factor A estimable on distinct columns in 16 runs FrF2(16, nfactors=6, estimable = rbind(rep(1,5),2:6), clear=FALSE) FrF2(16, nfactors=6, estimable = c("AB","AC","AD","AE","AF"), clear=FALSE) FrF2(16, nfactors=6, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=FALSE) ## formula would also accept self-defined factor names ## from factor.names instead of letters A, B, C, ... ## estimable does not need any other input FrF2(estimable=formula("~(A+B+C)^2+D+E")) ## estimable with factor names ## resolution three must be permitted, as FrF2 first determines that 8 runs ## would be sufficient degrees of freedom to estimate all effects ## and then tries to accomodate the 2fis from the model clear of aliasing in 8 runs FrF2(estimable=formula("~one+two+three+four+two:three+two:four"), factor.names=c("one","two","three","four"), res3=TRUE) ## clear=FALSE allows to allocate all effects on distinct columns in the ## 8 run MA resolution IV design FrF2(estimable=formula("~one+two+three+four+two:three+two:four"), factor.names=c("one","two","three","four"), clear=FALSE) ## 7 factors instead of 6, but no requirements for factor G FrF2(16, nfactors=7, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=FALSE) ## larger design for handling this with all required effects clear FrF2(32, nfactors=7, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=TRUE) ## 16 run design for handling this with required 2fis clear, but main effects aliased ## (does not usually make sense) FrF2(16, nfactors=7, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=TRUE, res3=TRUE) ## example for necessity of perms for the clear=FALSE case ## based on Wu and Chen Example 1 ## Not run: ## runs per default about max.time=60 seconds, before throwing error with ## interim results ## results could be used in select.catlg and perm.start for restarting with ## calculation of further possibilities FrF2(32, nfactors=11, estimable = formula("~(A+B+C+D+E+F)^2"), clear=FALSE) ## would run for a long long time (I have not yet been patient enough) FrF2(32, nfactors=11, estimable = formula("~(A+B+C+D+E+F)^2"), clear=FALSE, max.time=Inf) ## End(Not run) ## can be easily done with perms, ## as only different subsets of six factors are non-isomorphic perms.6 <- combn(11,6) perms.full <- matrix(NA,ncol(perms.6),11) for (i in 1:ncol(perms.6)) perms.full[i,] <- c(perms.6[,i],setdiff(1:11,perms.6[,i])) ## function compromise will calculate the necessary perms entries automatically compromise(11,1:6)$perms.full FrF2(32, nfactors=11, estimable = formula("~(A+B+C+D+E+F)^2"), clear=FALSE, perms = perms.full )
This function creates a foldover design for a 2-level fractional factorial. The purpose is to dealias (some) effects. Per default, all factors are folded upon, which makes the resulting design at least resolution IV. Different foldover versions can be requested.
fold.design(design, columns = "full", ...)
fold.design(design, columns = "full", ...)
design |
a data frame of class design that contains a 2-level fractional factorial;
currently, |
columns |
indicates which columns to fold on; the default “full” folds on all columns,
i.e. swaps levels for all columns. |
... |
currently not used |
Foldover is a method to dealias effects in relatively small 2-level fractional factorial designs. The folded design has twice the number of runs from the original design, and an additional column “fold” that distinguishes the original runs from the mirror runs. This column should be used in analyses, since it captures a block effect on time (often the mirror runs are conducted substantially later than the original experiment).
Like most other software, this function conducts a full foldover per default,
i.e. the mirror portion reverses the levels of all factors. In terms of the
convenient -1/1 notation for factor levels, this can be written as
a multiplication with “-1” for the mirror portion of all factors.
Thus, all confounding relations involving
an odd number of factors (e.g. A=BC) are resolved, because the odd side of the
equation involves a minus for the mirror runs, and the even side does not
(since the minuses cancel each other). (These
confounding relations are replaced by even ones
for which the odd side of the equation is multiplied with minus the new mirror factor fold
.)
There are many situations, for which the default full foldover is not the best possible foldover fraction, cf. e.g. Li and Mee (2002). It is therefore possible to choose an arbitrary foldover fraction. For example, folding on one particular factor alone dealiases all confounding relations for that factor, folding on two particular factors dealiases all confounding relations of these two with others but not of these two together with others and so on.
Folding Plackett-Burman designs also removes the (partial) aliasing with 2-factor interactions for all main effects that are mirrored.
A data frame of class design with twice as many rows as design
and
the additional factor fold
(added as the last factor for folded pb
designs, as the first factor for splitplot designs,
and as the last base factor for other folded regular fractional
factorial designs).
Existing response values are of course preserved, and response values for the new mirror runs are NA.
The type in attribute design.info
is suffixed with “.folded”, and
nruns
(and, if applicable, nWPs
) is doubled,
nfactors
(and, if applicable, nfac.WP
)
is increased by one (for the factor fold, which
is a block factor and can also be treated as such, but will currently be treated as a fixed
(whole plot) factor by any automated analysis routine). The creator element receives a list entry for the fold columns.
For regular fractional factorials (design type starting with FrF2
), the generator element is adjusted
(the generators for all generated fold factors now involve the folding factor), and an existing
catlg.entry element is replaced by a new generators element. The aliased
element is
adapted to the new alias structure. Note that the fold factor enters as a new base factor and therefore
is added to the factor matrix after the first log2(nruns) factors. This implies that all factor
letters previously used for the generated factors are changed - for avoiding confusion it is always recommended to
work with factor names that are meaningful in a subject-matter sense.
Furthermore, for the regular fractional factorial designs,
the column run.no.in.std.order in attribute run.order
for the mirror portion of the design is
populated such that the base factors remain in the conventional order when ordered by
run.no.in.std.order (regardless whether or not they are included in the fold;
it is always possible to reorder runs such that the original base factors
together with the folding factor form the new base in standard order).
This function is still somewhat experimental.
Ulrike Groemping
Li, H. and Mee, R. (2002). Better foldover fractions for resolution III 2^(k-p) designs. Technometrics 44, 278–283. New York: Springer.
Mee, R. (2009). A Comprehensive Guide to Factorial Two-Level Experimentation. New York: Springer.
Montgomery, D.C. (2001). Design and Analysis of Experiments (5th ed.). Wiley, New York.
## create resolution III design plan <- FrF2(8,5, factor.names=c("one","two","three","four","five")) ## add some resonse data y <- c(2+desnum(plan)%*%c(2,3,0,0,0) + 1.5*apply(desnum(plan)[,c(1,2)],1,"prod") + rnorm(8)) ## the "c()" makes y into a vector rather than a 1-column matrix plan <- add.response(plan, y) DanielPlot(lm(y~(.)^2,plan), alpha=0.2, half=TRUE) ## alias information design.info(plan) ## full foldover for dealiasing all main effects plan <- fold.design(plan) design.info(plan) ## further data, shifted by -2 y <- c(y, desnum(plan)[9:16,1:5]%*%c(2,3,0,0,0) + 1.5*apply(desnum(plan)[9:16,c(1,2)],1,"prod") + rnorm(8)) plan <- add.response(plan, y, replace=TRUE) linmod <- lm(y~(.)^2,plan) DanielPlot(linmod, alpha=0.2, half=TRUE) MEPlot(linmod) IAPlot(linmod) ## fold on factor a only (also removes main effect aliasing here) plan <- FrF2(8,5, factor.names=c("one","two","three","four","five")) aliasprint(plan) plan <- fold.design(plan, columns=1) aliasprint(plan) ## fold a Plackett-Burman design with 11 factors plan <- pb(12) fold.design(plan)
## create resolution III design plan <- FrF2(8,5, factor.names=c("one","two","three","four","five")) ## add some resonse data y <- c(2+desnum(plan)%*%c(2,3,0,0,0) + 1.5*apply(desnum(plan)[,c(1,2)],1,"prod") + rnorm(8)) ## the "c()" makes y into a vector rather than a 1-column matrix plan <- add.response(plan, y) DanielPlot(lm(y~(.)^2,plan), alpha=0.2, half=TRUE) ## alias information design.info(plan) ## full foldover for dealiasing all main effects plan <- fold.design(plan) design.info(plan) ## further data, shifted by -2 y <- c(y, desnum(plan)[9:16,1:5]%*%c(2,3,0,0,0) + 1.5*apply(desnum(plan)[9:16,c(1,2)],1,"prod") + rnorm(8)) plan <- add.response(plan, y, replace=TRUE) linmod <- lm(y~(.)^2,plan) DanielPlot(linmod, alpha=0.2, half=TRUE) MEPlot(linmod) IAPlot(linmod) ## fold on factor a only (also removes main effect aliasing here) plan <- FrF2(8,5, factor.names=c("one","two","three","four","five")) aliasprint(plan) plan <- fold.design(plan, columns=1) aliasprint(plan) ## fold a Plackett-Burman design with 11 factors plan <- pb(12) fold.design(plan)
Regular fractional factorial 2-level designs are provided. Apart from obtaining the usual minimum aberration designs in a fixed number of runs, it is possible to request highest number of free 2-factor interactions instead of minimum aberration or to request the smallest design that fulfills certain requirements (e.g. resolution V with 8 factors).
FrF2(nruns = NULL, nfactors = NULL, factor.names = if (!is.null(nfactors)) { if (nfactors <= 50) Letters[1:nfactors] else paste("F", 1:nfactors, sep = "")} else NULL, default.levels = c(-1, 1), ncenter=0, center.distribute=NULL, generators = NULL, design = NULL, resolution = NULL, select.catlg=catlg, estimable = NULL, clear = TRUE, method="VF2", sort="natural", ignore.dom = !isTRUE(all.equal(blocks,1)), useV = TRUE, firsthit=FALSE, res3 = FALSE, max.time = 60, perm.start=NULL, perms = NULL, MaxC2 = FALSE, replications = 1, repeat.only = FALSE, randomize = TRUE, seed = NULL, alias.info = 2, blocks = 1, block.name = "Blocks", block.old=FALSE, force.godolphin=alias.block.2fis, bbreps=replications, wbreps=1, alias.block.2fis = FALSE, hard = NULL, check.hard=10, WPs=1,nfac.WP=0, WPfacs=NULL, check.WPs = 10, ...) FrF2.currentlychecked()
FrF2(nruns = NULL, nfactors = NULL, factor.names = if (!is.null(nfactors)) { if (nfactors <= 50) Letters[1:nfactors] else paste("F", 1:nfactors, sep = "")} else NULL, default.levels = c(-1, 1), ncenter=0, center.distribute=NULL, generators = NULL, design = NULL, resolution = NULL, select.catlg=catlg, estimable = NULL, clear = TRUE, method="VF2", sort="natural", ignore.dom = !isTRUE(all.equal(blocks,1)), useV = TRUE, firsthit=FALSE, res3 = FALSE, max.time = 60, perm.start=NULL, perms = NULL, MaxC2 = FALSE, replications = 1, repeat.only = FALSE, randomize = TRUE, seed = NULL, alias.info = 2, blocks = 1, block.name = "Blocks", block.old=FALSE, force.godolphin=alias.block.2fis, bbreps=replications, wbreps=1, alias.block.2fis = FALSE, hard = NULL, check.hard=10, WPs=1,nfac.WP=0, WPfacs=NULL, check.WPs = 10, ...) FrF2.currentlychecked()
nruns |
Number of runs, must be a power of 2 (4 to 4096), if given. The number of runs can also be omitted. In that case,
if If If estimable is specified and |
nfactors |
is the number of 2-level factors to be investigated.
It can be omitted, if it is obvious from If For blocked designs, block generator columns are not included in
For automatically-generated split-plot designs (cf. details section),
|
factor.names |
a character vector of |
default.levels |
default levels (vector of length 2) for all factors for which no specific levels are given |
ncenter |
number of center points per block; |
center.distribute |
the number of positions over which the center points
are to be distributed for each block; if NULL (default), center points are
distributed over end, beginning, and middle (in that order, if there are fewer than three center points)
for randomized designs, and appended to the end for non-randomized designs.
for more detail, see function |
generators |
There are
a list of vectors with position numbers of base factors (e.g. c(1,3,4) stands for the interaction between first, third and fourth base factor) a vector of character representations of these interactions, e.g. “ACD” stands for the same interaction as above a vector of columns numbers in Yates order (e.g. 13 stands for ACD).
Note that the columns 1, 2, 4, 8, etc., i.e. all powers of 2, are reserved
for the base factors and cannot be used for assigning additional factors,
because the design would become a resolution II design. For looking up
which column number stands for which interaction, type e.g.
In all cases, preceding the respective entry with a minus sign
(e.g. -c(1,3,4), “-ACD”, -13) implies that the levels
of the respective column are reversed. |
design |
is a character string specifying the name of a design listed
in the catalogue specified as |
resolution |
is the arabic numeral for the requested resolution of the design.
|
select.catlg |
specifies a catalogue of class If a specific different catalogue of designs is available,
this can be specified here. Names of catalogues from package FrF2.catlg128
can be given here without prior loading of that package; loading of the package
and the selected catalogue will then happen automatically, provided the
package is installed (for version >=1.2 of package FrF2.catlg128;
for earlier versions, the suitable catalogue has to be manually loaded
using the |
estimable |
indicates the 2-factor interactions (2fis) that are to be estimable in
the design. Consult the specific help file (
|
clear |
logical, indicating how estimable is to be used. See |
method |
character string indicating which subgraph isomorphism search routine
of package igraph is used ( |
sort |
character string indicating how the estimability requirement and the
candidate design clear 2fis are handed to
the subgraph isomorphism search routine of package igraph.
The default This option is relevant for |
ignore.dom |
logical, default FALSE for unblocked designs, TRUE for blocked designs;
if TRUE, |
useV |
NULL or logical; relevant for designs with |
firsthit |
logical; relevant for designs with |
res3 |
logical; if TRUE, |
max.time |
maximum time for design search as requested by |
perm.start |
used with |
perms |
used with It is planned to automatically generate perms for certain structures like compromise designs in the (not so near) future. |
MaxC2 |
is a logical and defaults to FALSE. If TRUE,
maximizing the number of clear 2-factor interactions takes precedence
over minimizing aberration. Resolution is always considered first.
|
replications |
positive integer number. Default 1 (i.e. each row just once).
If larger, each design run is executed replication times.
If Otherwise (default), the full experiment is first carried out once, then for the second replication and so forth. In case of randomization, each such blocks is randomized separately. In this case, replication variance is more likely suitable for usage as error variance (unless e.g. the same parts are used for replication runs although build variation is important). |
repeat.only |
logical, relevant only if replications > 1. If TRUE,
replications of each run are grouped together
(repeated measurement rather than true replication). The default is
|
randomize |
logical. If TRUE, the design is randomized. This is the default.
In case of replications, the nature of randomization depends on the setting of
option |
seed |
optional seed for the randomization process |
alias.info |
can be 2 or 3, gives the order of interaction effects for which
alias information is to be included in the |
blocks |
is EITHER
If If the experiment is randomized, randomization happens within blocks.
In case of many blocks, units should also be randomized to blocks wherever possible! For the statistical and algorithmic background of blocked designs, see |
block.name |
name of the block factor, default “Blocks” |
block.old |
logical; if TRUE, blocking behavior of FrF2 version 1.7.2 is activated |
force.godolphin |
logical; if TRUE, blocking is forced to be done with the
Godolphin method (using function |
bbreps |
between block replications; these are always taken as genuine replications,
not repeat runs; default: equal to |
wbreps |
within block replications; whether or not these are taken as genuine replications
depends on the setting of |
alias.block.2fis |
logical indicating whether blocks may be aliased
with 2fis (default: |
hard |
gives the number of hard to change factors. These must be the
first factors in |
check.hard |
is the number of candidate designs from the catalogue specified
in |
WPs |
is the number of whole plots and must be a power of 2. For statistical and algorithmic information on treatment of split-plot designs
see the separate help file |
nfac.WP |
is the number of whole plot factors and must be smaller than The If a design is provided and whole plot factors are manually provided
(design or generators option together with If If |
WPfacs |
is per default NULL. In this case, the first
|
check.WPs |
is the number of potential split-plot designs that are
compared by function |
... |
currently not used |
Per default, the function picks the best design from the
default design catalogue catlg
(a list object of class catlg
).
Alternatively, the user can explicitly specify a design through accessing
a specific catalogued design using the design
option or specifying non-catalogued
generators via the generators
option.
Apart from generation of simple fractional factorial designs based on catalogued
or non-catalogued generators, function FrF2
allows specification of blocked designs
and split-plot designs, as well as specification of a set of 2fis that are required to be estimable.
The implementation of these possibilities is explained in the separate help files block
,
splitplot
and estimable.2fis
. If you consider to use
option hard
, it may also be worth while to look at the splitplot
help file.
Function FrF2
is still under development, although most features are
now included, and the principle structure of inputs and outputs should not change
much any more. Please contact me with any suggestions for improvements.
Function FrF2.currentlychecked
is meant as a diagnostic tool,
when searching for designs with option estimable
and clear=TRUE
. If the search takes very long, it can be interrupted
(CAUTION: in some igraph versions, interrupting the search may crash R).
After a successful interruption, and FrF2.currentlychecked()
returns a character string with the name of the design that was checked at the
time of interruption.
Function FrF2
returns a data frame of S3 class design
that has attached attributes that can be accessed
by functions desnum
,
run.order
and
design.info
.
The data frame itself contains the design with levels coded as requested.
If no center points have been requested, the design columns are factors with
contrasts -1
and +1
(cf. also contr.FrF2
); in case
of center points, the design columns are numeric.
The following attributes are attached to it:
desnum |
Design matrix in -1/1 coding |
run.order |
a three column data frame; |
design.info |
list with the entries
|
Since R version 3.6.0, the behavior of function sample
has changed
(correction of a biased previous behavior that should not be relevant for the randomization of designs).
For reproducing a randomized design that was produced with an earlier R version,
please follow the steps described with the argument seed
.
Ulrike Groemping
Bingham, D.R., Schoen, E.D. and Sitter, R.R. (2004). Designing Fractional Factorial Split-Plot Experiments with Few Whole-Plot Factors. Applied Statistics 53, 325-339.
Bingham, D. and Sitter, R.R. (2003). Fractional Factorial Split-Plot Designs for Robust Parameter Experiments. Technometrics 45, 80-89.
Bisgaard, S. (1994a). Blocking generators for small designs.
J. Quality Technology 26, 288-294.
Chen, J., Sun, D.X. and Wu, C.F.J. (1993) A catalogue of 2-level and 3-level orthogonal arrays. International Statistical Review 61, 131-145.
Cheng, C.-S., Martin, R.J., and Tang, B. (1998). Two-level factorial designs with extreme numbers of level changes. Annals of Statistics 26, 1522-1539.
Cheng, C.-S. and Tsai, P.-W. (2009). Optimal two-level regular fractional factorial block and split-plot designs. Biometrika 96, 83-93.
Cheng, S.W. and Wu, C.F.J. (2002). Choice of optimal blocking schemes in 2-level and 3-level designs. Technometrics 44, 269-277.
Cordella, L.P., Foggia, P., Sansone, C. and Vento, M. (2001). An improved algorithm for matching large graphs. Proc. of the 3rd IAPR TC-15 Workshop on Graphbased Representations in Pattern Recognition, 149–159.
Godolphin, J. (2021). Construction of Blocked Factorial Designs to Estimate Main Effects and Selected Two-Factor Interactions. J. Royal Statistical Society B 83, 5-29. doi:10.1111/rssb.12397.
Groemping, U. (2012). Creating clear designs: a graph-based algorithm and a catalogue of clear compromise plans. IIE Transactions 44, 988–1001. Early preprint available at http://www1.bht-berlin.de/FB_II/reports/Report-2010-005.pdf.
Groemping, U. (2014a). R Package FrF2 for Creating and Analyzing Fractional Factorial 2-Level Designs. Journal of Statistical Software, 56, Issue 1, 1-56. https://www.jstatsoft.org/v56/i01/.
Groemping, U. (2014b). A Note on Dominating Fractional Factorial Two-Level Designs With Clear Two-Factor Interactions. Technometrics 56, 42–45.
Groemping, U. (2021). An algorithm for blocking regular fractional factorial 2-level designs with clear two-factor interactions. Computational Statistics and Data Analysis 153, 1-18. doi:10.1016/j.csda.2020.107059. Preprint at Report 3/2019.
Huang, P., Chen, D. and Voelkel, J.O. (1998). Minimum-Aberration Two-Level Split-Plot Designs. Technometrics 40, 314-326.
Mee, R. (2009). A Comprehensive Guide to Factorial Two-Level Experimentation. New York: Springer.
Solnon, C. (2010). AllDifferent-based Filtering for Subgraph Isomorphism. Artificial Intelligence 174, 850–864.
Sun, D.X., Wu, C.F.J. and Chen, Y.Y. (1997).
Optimal blocking schemes for and
designs. Technometrics 39,
298-307.
Wu, C.F.J. and Chen, Y. (1992) A graph-aided method for planning two-level experiments when certain interactions are important. Technometrics 34, 162-175.
See also FrF2Large
for regular fractional factorial designs
with more than 4096 runs (these are not supported by a design catalogue, except for
a few resolution V designs which have not been checked for any optimality among the resolution V designs), pb
for non-regular fractional factorials according
to Plackett-Burman, catlg
for the underlying design catalogue and some accessor functions,
and block
, splitplot
or estimable.2fis
for statistical and algorithmic information on the respective topic.
## maximum resolution minimum aberration design with 4 factors in 8 runs FrF2(8,4) ## the design with changed default level codes FrF2(8,4, default.level=c("current","new")) ## the design with number of factors specified via factor names ## (standard level codes) FrF2(8,factor.names=list(temp="",press="",material="",state="")) ## the design with changed factor names and factor-specific level codes FrF2(8,4, factor.names=list(temp=c("min","max"),press=c("low","normal"), material=c("current","new"),state=c("new","aged"))) ## a full factorial FrF2(8,3, factor.names=list(temp=c("min","max"),press=c("low","normal"), material=c("current","new"))) ## a replicated full factorial (implicit by low number of factors) FrF2(16,3, factor.names=list(temp=c("min","max"),press=c("low","normal"), material=c("current","new"))) ## three ways for custom specification of the same design FrF2(8, generators = "ABC") FrF2(8, generators = 7) FrF2(8, generators = list(c(1,2,3))) ## more than one generator FrF2(8, generators = c("ABC","BC")) FrF2(8, generators = c(7,6)) FrF2(8, generators = list(c(1,2,3),c(2,3))) ## alias structure for three generators that differ only by sign design.info(FrF2(16,generators=c(7,13,15),randomize=FALSE))$aliased design.info(FrF2(16,generators=c(7,-13,15),randomize=FALSE))$aliased design.info(FrF2(16,generators=c(-7,-13,-15),randomize=FALSE))$aliased ## finding smallest design with resolution 5 in 7 factors FrF2(nfactors=7, resolution=5) ## same design, but with 12 center points in 6 positions FrF2(nfactors=7, resolution=5, ncenter=12, center.distribute=6) ## maximum resolution minimum aberration design with 9 factors in 32 runs ## show design information instead of design itself design.info(FrF2(32,9)) ## maximum number of free 2-factor interactions instead of minimum aberration ## show design information instead of design itself design.info(FrF2(32,9,MaxC2=TRUE)) ## usage of replication ## shows run order instead of design itself run.order(FrF2(8,4,replication=2,randomize=FALSE)) run.order(FrF2(8,4,replication=2,repeat.only=TRUE,randomize=FALSE)) run.order(FrF2(8,4,replication=2)) run.order(FrF2(8,4,replication=2,repeat.only=TRUE)) ## Not run: ## examples below do work, but are repeated in the ## respective method's separate help file and are therefore prevented ## from running twice ########## automatic blocked designs ################### ## from a full factorial ## FrF2(8,3,blocks=2) ## with replication run.order(FrF2(8,3,blocks=2,wbreps=2)) run.order(FrF2(8,3,blocks=2,wbreps=2,repeat.only=TRUE)) run.order(FrF2(8,3,blocks=2,bbreps=2)) run.order(FrF2(8,3,blocks=2,bbreps=2,wbreps=2)) ## automatic blocked design with fractions FrF2(16,7,blocks=4,alias.block.2fis=TRUE,factor.names=c("MotorSpeed", "FeedMode","FeedSizing","MaterialType","Gain","ScreenAngle","ScreenVibLevel")) ## isomorphic non-catalogued design as basis, using Godolphin approach FrF2(16,gen=c(7,11,14),blocks=4,alias.block.2fis=TRUE) ## isomorphic non-catalogued design as basis, not using Godolphin approach ## (different design of comparable quality in this case) FrF2(16,gen=c(7,11,14),blocks=4,alias.block.2fis=TRUE, force.godolphin=FALSE) ## FrF2 uses blockpick.big and ignores the generator FrF2(64,gen=c(7,11,14),blocks=16,alias.block.2fis=TRUE, block.old=TRUE) ## FrF2 uses Godolphin approach, regardless of force.godolphin argument ## because the setting is large FrF2(64,gen=c(7,11,14),blocks=16,alias.block.2fis=TRUE) ########## manual blocked design #################### ### example that shows why order of blocks is not randomized ### can of course be randomized by user, if appropriate FrF2(32,9,blocks=c("Day","Shift"),alias.block.2fis=TRUE, factor.names=list(Day=c("Wednesday","Thursday"), Shift=c("Morning","Afternoon"), F1="",F2="",F3="",F4="",F5="",F6="",F7=""), default.levels=c("current","new")) ########## blocked design with estimable 2fis #################### ### all interactions of last two factors to be estimable clearly ### in 64 run design with blocks of size 4 ### not possible with catalogue entry 9-3.1 FrF2(64, 6, blocks=16, factor.names=Letters[15:20], estimable=compromise(6,3)$requirement, alias.block.2fis=TRUE, randomize=FALSE) FrF2(design="9-3.2", blocks=16, alias.block.2fis=TRUE, factor.names = list(C1="",C2="",C3="",C4="",C5="",C6="",C7="", N1=c("low","high"),N2=c("low","high")), default.levels = c("current","new"), estimable=compromise(9, 8:9)$requirement) FrF2(256, 13, blocks=64, alias.block.2fis=TRUE, factor.names = list(C1="",C2="",C3="",C4="",C5="",C6="",C7="",C8="", N1=c("low","high")), default.levels = c("current","new"), estimable=compromise(13, 1)$requirement) ########## hard to change factors #################### ## example from Bingham and Sitter Technometrics 19999 ## MotorSpeed, FeedMode,FeedSizing,MaterialType are hard to change BS.ex <- FrF2(16,7,hard=4, factor.names=c("MotorSpeed", "FeedMode","FeedSizing","MaterialType", "Gain","ScreenAngle","ScreenVibLevel"), default.levels=c("-","+"),randomize=FALSE) design.info(BS.ex) BS.ex ## NOTE: the design has 8 whole plots. ## If randomize=FALSE is used like here, the first hard-to-change factors ## do not always change between whole plots. ## A conscious and honest decision is required whether this is ## acceptable for the situation at hand! ## randomize=TRUE would cause more changes in the first four factors. ########## automatic generation for split plot ########## ## 3 control factors, 5 noise factors, control factors are whole plot factors ## 8 plots desired in a total of 32 runs ## Bingham Sitter 2003 BS.ex2a <- FrF2(32, 8, WPs=8, nfac.WP=3, factor.names=c(paste("C",1:3,sep=""), paste("N",1:5,sep="")),randomize=TRUE) ## manual generation of this same design BS.ex2m <- FrF2(32, 8, generators=c("ABD","ACD","BCDE"),WPs=8, WPfacs=c("C1","C2","C3"), nfac.WP=3, factor.names=c(paste("C",1:3,sep=""),paste("N",1:5,sep="")),randomize=TRUE) ## design with few whole plot factors ## 2 whole plot factors, 7 split plot factors ## 8 whole plots, i.e. one extra WP factor needed BSS.cheese.exa <- FrF2(32, 9, WPs=8, nfac.WP=2, factor.names=c("A","B","p","q","r","s","t","u","v")) design.info(BSS.cheese.exa) ## manual generation of the design used by Bingham, Schoen and Sitter ## note that the generators include a generator for the 10th spplitting factor ## s= ABq, t = Apq, u = ABpr and v = Aqr, splitting factor rho=Apqr BSS.cheese.exm <- FrF2(32, gen=list(c(1,2,4),c(1,3,4),c(1,2,3,5),c(1,4,5),c(1,3,4,5)), WPs=8, nfac.WP=3, WPfacs=c(1,2,10), factor.names=c("A","B","p","q","r","s","t","u","v","rho")) design.info(BSS.cheese.exm) ########## usage of estimable ########################### ## design with all 2fis of factor A estimable on distinct columns in 16 runs FrF2(16, nfactors=6, estimable = rbind(rep(1,5),2:6), clear=FALSE) FrF2(16, nfactors=6, estimable = c("AB","AC","AD","AE","AF"), clear=FALSE) FrF2(16, nfactors=6, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=FALSE) ## formula would also accept self-defined factor names ## from factor.names instead of letters A, B, C, ... ## estimable does not need any other input FrF2(estimable=formula("~(A+B+C)^2+D+E")) ## estimable with factor names ## resolution three must be permitted, as FrF2 first determines that 8 runs ## would be sufficient degrees of freedom to estimate all effects ## and then tries to accomodate the 2fis from the model clear of aliasing in 8 runs FrF2(estimable=formula("~one+two+three+four+two:three+two:four"), factor.names=c("one","two","three","four"), res3=TRUE) ## clear=FALSE allows to allocate all effects on distinct columns in the ## 8 run MA resolution IV design FrF2(estimable=formula("~one+two+three+four+two:three+two:four"), factor.names=c("one","two","three","four"), clear=FALSE) ## 7 factors instead of 6, but no requirements for factor G FrF2(16, nfactors=7, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=FALSE) ## larger design for handling this with all required effects clear FrF2(32, nfactors=7, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=TRUE) ## 16 run design for handling this with required 2fis clear, but main effects aliased ## (does not usually make sense) FrF2(16, nfactors=7, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=TRUE, res3=TRUE) ## End(Not run) ## example for the sort option added with version 1.6-1 ## and for usage of a catalogue from package FrF2.catlg128 (simplified with version 1.6-5) ## Not run: estim <- compromise(17,15:17)$requirement ## all interactions of factors 15 to 17 (P,Q,R) ## VF2 algorithm without pre-sorting of vertices ### CAUTION: in some igraph versions, the following may crash R ### FrF2(128, 17, estimable=estim, select.catlg=catlg128.17) ## very slow, interrupt with ESC key after a short while ## !!! save all important work before, in case R crashes FrF2.currentlychecked() ## displays the design that was currently checked ## should be 17-10.2407, if the interrupt was successful ## VF2 algorithm with pre-sorting of vertices FrF2(128, 17, estimable=estim, sort="high", select.catlg=catlg128.17) ## very fast FrF2(128, 17, estimable=estim, sort="low", select.catlg=catlg128.17) ## very fast ## LAD algorithm FrF2(128, 17, estimable=estim, method="LAD", select.catlg=catlg128.17) ## very fast ## guaranteed to be MA clear design ## only works, if package FrF2.catlg128 is installed ## End(Not run) ## example for necessity of perms, and uses of select.catlg and perm.start ## based on Wu and Chen Example 1 ## Not run: ## runs per default about max.time=60 seconds, before throwing error with ## interim results ## results could be used in select.catlg and perm.start for restarting with ## calculation of further possibilities FrF2(32, nfactors=11, estimable = formula("~(A+B+C+D+E+F)^2"), clear=FALSE) ## would run for a long long time (I have not yet been patient enough) FrF2(32, nfactors=11, estimable = formula("~(A+B+C+D+E+F)^2"), clear=FALSE, max.time=Inf) ## can be easily done with perms, ## as only different subsets of six factors are non-isomorphic perms.6 <- combn(11,6) perms.full <- matrix(NA,ncol(perms.6),11) for (i in 1:ncol(perms.6)) perms.full[i,] <- c(perms.6[,i],setdiff(1:11,perms.6[,i])) FrF2(32, nfactors=11, estimable = formula("~(A+B+C+D+E+F)^2"), clear=FALSE, perms = perms.full ) ## End(Not run)
## maximum resolution minimum aberration design with 4 factors in 8 runs FrF2(8,4) ## the design with changed default level codes FrF2(8,4, default.level=c("current","new")) ## the design with number of factors specified via factor names ## (standard level codes) FrF2(8,factor.names=list(temp="",press="",material="",state="")) ## the design with changed factor names and factor-specific level codes FrF2(8,4, factor.names=list(temp=c("min","max"),press=c("low","normal"), material=c("current","new"),state=c("new","aged"))) ## a full factorial FrF2(8,3, factor.names=list(temp=c("min","max"),press=c("low","normal"), material=c("current","new"))) ## a replicated full factorial (implicit by low number of factors) FrF2(16,3, factor.names=list(temp=c("min","max"),press=c("low","normal"), material=c("current","new"))) ## three ways for custom specification of the same design FrF2(8, generators = "ABC") FrF2(8, generators = 7) FrF2(8, generators = list(c(1,2,3))) ## more than one generator FrF2(8, generators = c("ABC","BC")) FrF2(8, generators = c(7,6)) FrF2(8, generators = list(c(1,2,3),c(2,3))) ## alias structure for three generators that differ only by sign design.info(FrF2(16,generators=c(7,13,15),randomize=FALSE))$aliased design.info(FrF2(16,generators=c(7,-13,15),randomize=FALSE))$aliased design.info(FrF2(16,generators=c(-7,-13,-15),randomize=FALSE))$aliased ## finding smallest design with resolution 5 in 7 factors FrF2(nfactors=7, resolution=5) ## same design, but with 12 center points in 6 positions FrF2(nfactors=7, resolution=5, ncenter=12, center.distribute=6) ## maximum resolution minimum aberration design with 9 factors in 32 runs ## show design information instead of design itself design.info(FrF2(32,9)) ## maximum number of free 2-factor interactions instead of minimum aberration ## show design information instead of design itself design.info(FrF2(32,9,MaxC2=TRUE)) ## usage of replication ## shows run order instead of design itself run.order(FrF2(8,4,replication=2,randomize=FALSE)) run.order(FrF2(8,4,replication=2,repeat.only=TRUE,randomize=FALSE)) run.order(FrF2(8,4,replication=2)) run.order(FrF2(8,4,replication=2,repeat.only=TRUE)) ## Not run: ## examples below do work, but are repeated in the ## respective method's separate help file and are therefore prevented ## from running twice ########## automatic blocked designs ################### ## from a full factorial ## FrF2(8,3,blocks=2) ## with replication run.order(FrF2(8,3,blocks=2,wbreps=2)) run.order(FrF2(8,3,blocks=2,wbreps=2,repeat.only=TRUE)) run.order(FrF2(8,3,blocks=2,bbreps=2)) run.order(FrF2(8,3,blocks=2,bbreps=2,wbreps=2)) ## automatic blocked design with fractions FrF2(16,7,blocks=4,alias.block.2fis=TRUE,factor.names=c("MotorSpeed", "FeedMode","FeedSizing","MaterialType","Gain","ScreenAngle","ScreenVibLevel")) ## isomorphic non-catalogued design as basis, using Godolphin approach FrF2(16,gen=c(7,11,14),blocks=4,alias.block.2fis=TRUE) ## isomorphic non-catalogued design as basis, not using Godolphin approach ## (different design of comparable quality in this case) FrF2(16,gen=c(7,11,14),blocks=4,alias.block.2fis=TRUE, force.godolphin=FALSE) ## FrF2 uses blockpick.big and ignores the generator FrF2(64,gen=c(7,11,14),blocks=16,alias.block.2fis=TRUE, block.old=TRUE) ## FrF2 uses Godolphin approach, regardless of force.godolphin argument ## because the setting is large FrF2(64,gen=c(7,11,14),blocks=16,alias.block.2fis=TRUE) ########## manual blocked design #################### ### example that shows why order of blocks is not randomized ### can of course be randomized by user, if appropriate FrF2(32,9,blocks=c("Day","Shift"),alias.block.2fis=TRUE, factor.names=list(Day=c("Wednesday","Thursday"), Shift=c("Morning","Afternoon"), F1="",F2="",F3="",F4="",F5="",F6="",F7=""), default.levels=c("current","new")) ########## blocked design with estimable 2fis #################### ### all interactions of last two factors to be estimable clearly ### in 64 run design with blocks of size 4 ### not possible with catalogue entry 9-3.1 FrF2(64, 6, blocks=16, factor.names=Letters[15:20], estimable=compromise(6,3)$requirement, alias.block.2fis=TRUE, randomize=FALSE) FrF2(design="9-3.2", blocks=16, alias.block.2fis=TRUE, factor.names = list(C1="",C2="",C3="",C4="",C5="",C6="",C7="", N1=c("low","high"),N2=c("low","high")), default.levels = c("current","new"), estimable=compromise(9, 8:9)$requirement) FrF2(256, 13, blocks=64, alias.block.2fis=TRUE, factor.names = list(C1="",C2="",C3="",C4="",C5="",C6="",C7="",C8="", N1=c("low","high")), default.levels = c("current","new"), estimable=compromise(13, 1)$requirement) ########## hard to change factors #################### ## example from Bingham and Sitter Technometrics 19999 ## MotorSpeed, FeedMode,FeedSizing,MaterialType are hard to change BS.ex <- FrF2(16,7,hard=4, factor.names=c("MotorSpeed", "FeedMode","FeedSizing","MaterialType", "Gain","ScreenAngle","ScreenVibLevel"), default.levels=c("-","+"),randomize=FALSE) design.info(BS.ex) BS.ex ## NOTE: the design has 8 whole plots. ## If randomize=FALSE is used like here, the first hard-to-change factors ## do not always change between whole plots. ## A conscious and honest decision is required whether this is ## acceptable for the situation at hand! ## randomize=TRUE would cause more changes in the first four factors. ########## automatic generation for split plot ########## ## 3 control factors, 5 noise factors, control factors are whole plot factors ## 8 plots desired in a total of 32 runs ## Bingham Sitter 2003 BS.ex2a <- FrF2(32, 8, WPs=8, nfac.WP=3, factor.names=c(paste("C",1:3,sep=""), paste("N",1:5,sep="")),randomize=TRUE) ## manual generation of this same design BS.ex2m <- FrF2(32, 8, generators=c("ABD","ACD","BCDE"),WPs=8, WPfacs=c("C1","C2","C3"), nfac.WP=3, factor.names=c(paste("C",1:3,sep=""),paste("N",1:5,sep="")),randomize=TRUE) ## design with few whole plot factors ## 2 whole plot factors, 7 split plot factors ## 8 whole plots, i.e. one extra WP factor needed BSS.cheese.exa <- FrF2(32, 9, WPs=8, nfac.WP=2, factor.names=c("A","B","p","q","r","s","t","u","v")) design.info(BSS.cheese.exa) ## manual generation of the design used by Bingham, Schoen and Sitter ## note that the generators include a generator for the 10th spplitting factor ## s= ABq, t = Apq, u = ABpr and v = Aqr, splitting factor rho=Apqr BSS.cheese.exm <- FrF2(32, gen=list(c(1,2,4),c(1,3,4),c(1,2,3,5),c(1,4,5),c(1,3,4,5)), WPs=8, nfac.WP=3, WPfacs=c(1,2,10), factor.names=c("A","B","p","q","r","s","t","u","v","rho")) design.info(BSS.cheese.exm) ########## usage of estimable ########################### ## design with all 2fis of factor A estimable on distinct columns in 16 runs FrF2(16, nfactors=6, estimable = rbind(rep(1,5),2:6), clear=FALSE) FrF2(16, nfactors=6, estimable = c("AB","AC","AD","AE","AF"), clear=FALSE) FrF2(16, nfactors=6, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=FALSE) ## formula would also accept self-defined factor names ## from factor.names instead of letters A, B, C, ... ## estimable does not need any other input FrF2(estimable=formula("~(A+B+C)^2+D+E")) ## estimable with factor names ## resolution three must be permitted, as FrF2 first determines that 8 runs ## would be sufficient degrees of freedom to estimate all effects ## and then tries to accomodate the 2fis from the model clear of aliasing in 8 runs FrF2(estimable=formula("~one+two+three+four+two:three+two:four"), factor.names=c("one","two","three","four"), res3=TRUE) ## clear=FALSE allows to allocate all effects on distinct columns in the ## 8 run MA resolution IV design FrF2(estimable=formula("~one+two+three+four+two:three+two:four"), factor.names=c("one","two","three","four"), clear=FALSE) ## 7 factors instead of 6, but no requirements for factor G FrF2(16, nfactors=7, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=FALSE) ## larger design for handling this with all required effects clear FrF2(32, nfactors=7, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=TRUE) ## 16 run design for handling this with required 2fis clear, but main effects aliased ## (does not usually make sense) FrF2(16, nfactors=7, estimable = formula("~A+B+C+D+E+F+A:(B+C+D+E+F)"), clear=TRUE, res3=TRUE) ## End(Not run) ## example for the sort option added with version 1.6-1 ## and for usage of a catalogue from package FrF2.catlg128 (simplified with version 1.6-5) ## Not run: estim <- compromise(17,15:17)$requirement ## all interactions of factors 15 to 17 (P,Q,R) ## VF2 algorithm without pre-sorting of vertices ### CAUTION: in some igraph versions, the following may crash R ### FrF2(128, 17, estimable=estim, select.catlg=catlg128.17) ## very slow, interrupt with ESC key after a short while ## !!! save all important work before, in case R crashes FrF2.currentlychecked() ## displays the design that was currently checked ## should be 17-10.2407, if the interrupt was successful ## VF2 algorithm with pre-sorting of vertices FrF2(128, 17, estimable=estim, sort="high", select.catlg=catlg128.17) ## very fast FrF2(128, 17, estimable=estim, sort="low", select.catlg=catlg128.17) ## very fast ## LAD algorithm FrF2(128, 17, estimable=estim, method="LAD", select.catlg=catlg128.17) ## very fast ## guaranteed to be MA clear design ## only works, if package FrF2.catlg128 is installed ## End(Not run) ## example for necessity of perms, and uses of select.catlg and perm.start ## based on Wu and Chen Example 1 ## Not run: ## runs per default about max.time=60 seconds, before throwing error with ## interim results ## results could be used in select.catlg and perm.start for restarting with ## calculation of further possibilities FrF2(32, nfactors=11, estimable = formula("~(A+B+C+D+E+F)^2"), clear=FALSE) ## would run for a long long time (I have not yet been patient enough) FrF2(32, nfactors=11, estimable = formula("~(A+B+C+D+E+F)^2"), clear=FALSE, max.time=Inf) ## can be easily done with perms, ## as only different subsets of six factors are non-isomorphic perms.6 <- combn(11,6) perms.full <- matrix(NA,ncol(perms.6),11) for (i in 1:ncol(perms.6)) perms.full[i,] <- c(perms.6[,i],setdiff(1:11,perms.6[,i])) FrF2(32, nfactors=11, estimable = formula("~(A+B+C+D+E+F)^2"), clear=FALSE, perms = perms.full ) ## End(Not run)
Large regular fractional factorial 2-level designs in 8192 or more runs are provided: Resolution V designs in 8096 to 32768 runs with up to 120 factors according to the suggestion by Sanchez and Sanchez 2005 are automatically created (these are not necessarily optimal). Furthermore, manual generation of large regular fractional factorial designs via specification of generators is possible.
FrF2Large(nruns, nfactors = NULL, factor.names = if (!is.null(nfactors)){ if (nfactors <= 50) Letters[1:nfactors] else paste("F", 1:nfactors, sep = "") } else NULL, default.levels = c(-1, 1), ncenter = 0, center.distribute = NULL, generators = NULL, replications = 1, repeat.only = FALSE, randomize = TRUE, seed = NULL, alias.info = 2, ...) nrunsV(nfactors)
FrF2Large(nruns, nfactors = NULL, factor.names = if (!is.null(nfactors)){ if (nfactors <= 50) Letters[1:nfactors] else paste("F", 1:nfactors, sep = "") } else NULL, default.levels = c(-1, 1), ncenter = 0, center.distribute = NULL, generators = NULL, replications = 1, repeat.only = FALSE, randomize = TRUE, seed = NULL, alias.info = 2, ...) nrunsV(nfactors)
nruns |
Number of runs, must be a power of 2 (8192 to 32768). The number of runs must match the number of factors. Function |
nfactors |
is the number of 2-level factors to be investigated.
It can be omitted, if it is obvious from options |
factor.names |
a character vector of |
default.levels |
default levels (vector of length 2) for all factors for which no specific levels are given |
ncenter |
number of center points per block; |
center.distribute |
the number of positions over which the center points
are to be distributed for each block; if NULL (default), center points are
distributed over end, beginning, and middle (in that order, if there are fewer than three center points)
for randomized designs, and appended to the end for non-randomized designs.
for more detail, see function |
generators |
There are
a list of vectors with position numbers of base factors (e.g. c(1,3,4) stands for the interaction between first, third and fourth base factor) a vector of character representations of these interactions, e.g. “ACD” stands for the same interaction as above a vector of columns numbers in Yates order (e.g. 13 stands for ACD).
Note that the columns 1, 2, 4, 8, etc., i.e. all powers of 2, are reserved
for the base factors and cannot be used for assigning additional factors,
because the design would become a resolution II design. For looking up
which column number stands for which interaction, type e.g.
WARNING: Contrary to function |
replications |
positive integer number. Default 1 (i.e. each row just once).
If larger, each design run is executed replication times.
If Otherwise (default), the full experiment is first carried out once, then for the second replication and so forth. In case of randomization, each such blocks is randomized separately. In this case, replication variance is more likely suitable for usage as error variance (unless e.g. the same parts are used for replication runs although build variation is important). |
repeat.only |
logical, relevant only if replications > 1. If TRUE,
replications of each run are grouped together
(repeated measurement rather than true replication). The default is
|
randomize |
logical. If TRUE, the design is randomized. This is the default.
In case of replications, the nature of randomization depends on the setting of
option |
seed |
optional seed for the randomization process |
alias.info |
can be 2 or 3, gives the order of interaction effects for which
alias information is to be included in the |
... |
currently not used |
If generators are not explicitly specified, function FrF2Large
creates a
resolution V design according to the
rules by Sanchez and Sanchez (2005) for the specified number of factors in
the specified number of runs. The Sanchez and Sanchez article offers designs with
at least 1024 runs for 25 to 29 factors (1024 up to 33 factors with FrF2
),
at least 2048 runs for 30 to 38 factors (2048 up to 47 factors with FrF2
),
at least 4096 runs for 39 to 52 factors (4096 up to 65 factors with FrF2
),
at least 8192 runs for 53 to 69 factors (up to 65 factors in half the run size with FrF2
),
at least 16384 runs for 70 to 92 factors, (
at least 32768 runs for 93 to 120 factors.
For designs with up to 4096 runs, function FrF2
creates better automatic designs.
Therefore, function FrF2Large
is restricted to usage for larger designs.
Users can explicitly specify a design through specifying
generators via the generators
option. For up to 4096 runs, this is also possible
with function FrF2
, even with more flexibility. Therefore, manual design generation
with function FrF2Large
is also restricted to designs of at least 8192 runs.
Manual generation of large designs with the option generators
is limited by
computer memory only. nruns
must be at least large enough to accomodate the
rightmost generator column; for example, if generators contains an element ABEP
,
P
is the 15th base factor (15th letter in Letters
),
i.e. nruns
must be at least 2^15
=32768;
if the largest generator column number in Yates column notation is 4201,
nruns
must be at least 2^ceiling(log2(4201))
=8192.
Function nrunsV
invisibly returns the number of runs requested and
prints a message with the number of runs and the appropriate function.
Function FrF2Large
returns a data frame of S3 class
design
and has attached attributes that can be accessed
by functions desnum
,
run.order
and
design.info
.
The data frame itself contains the design with levels coded as requested.
If no center points have been requested, the design columns are factors with
contrasts -1
and +1
(cf. also contr.FrF2
); in case
of center points, the design columns are numeric.
The following attributes are attached to it:
desnum |
Design matrix in -1/1 coding |
run.order |
three column data frame, first column contains the run number in standard order, second column the run number as randomized, third column the run number with replication number as postfix; useful for switching back and forth between actual and standard run number |
design.info |
list with the entries
|
Since R version 3.6.0, the behavior of function sample
has changed
(correction of a biased previous behavior that might be relevant for the randomization
of very large designs).
For reproducing a randomized design that was produced with an earlier R version,
please follow the steps described with the argument seed
.
Ulrike Groemping
Mee, R. (2009). A Comprehensive Guide to Factorial Two-Level Experimentation. New York: Springer.
Sanchez, S.M. and Sanchez, P.J. (2005). Very Large Fractional Factorial and Central Composite Designs. ACM Transactions on Modeling and Computer Simulation 15, 362-377.
See also FrF2
for smaller regular fractional factorials and
oacat
for two non-regular resolution V fractional factorials (reported e.g. by
Mee 2009) for up to 19 factors in 256 runs or up to 63 factors in 2048 runs
## numbers of runs needed for resolution V designs in different numbers of factors nrunsV(8) nrunsV(18) needed <- nrunsV(27) needed nrunsV(65) nrunsV(71) ## Not run: plan <- FrF2Large(nrunsV(75),75) summary(plan) ## End(Not run)
## numbers of runs needed for resolution V designs in different numbers of factors nrunsV(8) nrunsV(18) needed <- nrunsV(27) needed nrunsV(65) nrunsV(71) ## Not run: plan <- FrF2Large(nrunsV(75),75) summary(plan) ## End(Not run)
Function colpick handles the creation of X matrices for blocking, function FF_from_X blocks a full factorial, function X_from_profile creates an X matrix from a profile, function phimax calculations the maximum number of clear 2fis from Godolphin's approach. Further helper functions support the use of the method. The functions are meant for expert users only.
colpick(design, q, all = FALSE, select.catlg = catlg, estimable = NULL, method = "VF2", sort = "natural", res3 = FALSE, all0 = FALSE, quiet = FALSE, firsthit = is.numeric(design)) FF_from_X(X, randomize = TRUE, seed = NULL, alias.info=2) X_from_profile(n, q, profile = NULL) clear2fis_from_profile(n, q, profile = NULL) X_from_parts(n, q, parts) phimax(n, q, profile = NULL) blockgencreate(X, p = 0) Xcalc(XI, gen) blockgengroup(X, p = 0, num = FALSE) colpickIV(design, q, all = FALSE, select.catlg = catlg, estimable = NULL, method = "VF2", sort = "natural", res3 = FALSE, all0 = FALSE, quiet = FALSE, firsthit = is.numeric(design))
colpick(design, q, all = FALSE, select.catlg = catlg, estimable = NULL, method = "VF2", sort = "natural", res3 = FALSE, all0 = FALSE, quiet = FALSE, firsthit = is.numeric(design)) FF_from_X(X, randomize = TRUE, seed = NULL, alias.info=2) X_from_profile(n, q, profile = NULL) clear2fis_from_profile(n, q, profile = NULL) X_from_parts(n, q, parts) phimax(n, q, profile = NULL) blockgencreate(X, p = 0) Xcalc(XI, gen) blockgengroup(X, p = 0, num = FALSE) colpickIV(design, q, all = FALSE, select.catlg = catlg, estimable = NULL, method = "VF2", sort = "natural", res3 = FALSE, all0 = FALSE, quiet = FALSE, firsthit = is.numeric(design))
design |
a character string that identifies a design in the cataloge specified
by option |
q |
the requested block size is |
all |
if TRUE (default FALSE), all possible X matrices are returned;
otherwise, |
select.catlg |
name of catalogue (not in quotes);
only relevant, if |
estimable |
a specification of 2fis to be kept clear in the blocked design, either as a character vector of pairs of factor letters (using the first elements of 'Letters') or as a two-row matrix of pairs of factor numbers occurring in 2fis) |
method |
character string identifying a subgraph isomorphism method
(VF2 or LAD), see |
sort |
character string specifying a presort strategy for subgraph
isomorphism search, see |
res3 |
relevant only if |
all0 |
per default ( |
quiet |
if TRUE, the message about failure is suppressed (for using the function
inside other functions, like |
firsthit |
if TRUE, the function does not attempt to optimize the
number of clear 2fis but accepts the first acceptable blocking
(relevant for non-null |
X |
a |
randomize |
logical. If TRUE, the design is randomized. This is the default.
Randomization is implemented using function
|
seed |
optional seed for the randomization process |
alias.info |
degree of effects aliased with blocks to be included in the
|
profile |
profile to use for calculation
(NULL or integer vector of up to |
parts |
list that provides factor partitions; list entries
must either be all integers from 1 to |
n |
number of factors |
p |
the number of generated factors (among the n factors);
|
XI |
a |
gen |
generators for extending |
num |
if TRUE (default FALSE), Yates column numbers are returned instead of their character representations |
These are the functions for the Godolphin (2021) approach to blocking; most of them are user-visible. This approach and its implementation are described in Groemping (2021). Direct use of this functions is intended for expert use only.
Function colpick
is
the main workhorse function for blocking larger situations
in function FrF2
(since version 2 of the package, it replaces the earlier approach
with function blockpick.big
); it makes use of function
blockgencreate
, and of the internal function blockgengroup
.
Function FF_from_X
creates a class design
object.
Design size is limited by computer memory and run time.
The function can use an X matrix that was produced by function
colpick
; but note that it is quite easy to hand-craft
an X matrix for a full factorial, even with estimability requirements.
The light-weight function does not have arguments for customization; it can
be post-processed, however, e.g. using function
factor.names<-
.
Function X_from_profile
creates an X matrix that corresponds to
the specified profile.
Function phimax
returns the maximum number of 2fis that can
be kept clear when blocking a full factorial design in factors into
blocks of size
, given the specified profile.
Function blockgencreate
creates block generators from an X matrix
for blocking a design in runs into blocks of size
,
where
and
are derived from
X
as the number of columns and rows,
respectively. The generators are returned as a character vector that consists of
strings of base factor letters.
Function Xcalc
extends a
matrix X_I by
columns
(X_II in Godolphin notatation) based
on the generators provided in
gen
.
Function blockgengroup
is internal only,
as are functions colpickIV
and clear2fis_from_profile
.
Function colpick
returns a list of at least two elements:
if all
is FALSE, the list consists of the matrix X
,
the character vector
clear.2fis
and possibly the integer vector map
,
otherwise of list-valued elements X_matrices
, clearlist
and
profiles
and maplist
.
Function FF_from_X
returns a class design object of type FrF2.blocked
.
Function phimax
returns a real number.
Function blockgencreate
returns a character vector of generators
in terms of Letters
combinations of the first $n-p$ factors.
Function Xcalc
returns a
matrix (in case of a single generator) or a list of such matrices
(if
gen
is a class catlg
object with more than one element).
The internal function blockgengroup
returns a character vector of all effects
(denoted as base column letter combinations) aliased with the block main effect,
or corresponding Yates column numbers.
The internal function colpickIV
returns almost the same type of results as colpick
.
The difference:
if all
is TRUE, there is an integer vector map
instead of the
maplist
element, because the map
does not depend on the choice of X-matrix
(separate subgraph isomorphism checking is skipped with this function).
Ulrike Groemping
Groemping, U. (2012). Creating clear designs: a graph-based algorithm and a catalogue of clear compromise plans. IIE Transactions 44, 988–1001. Early preprint available at http://www1.bht-berlin.de/FB_II/reports/Report-2010-005.pdf.
Godolphin, J. (2021). Construction of Blocked Factorial Designs to Estimate Main Effects and Selected Two-Factor Interactions. J. Royal Statistical Society B 83, 5-29. doi:10.1111/rssb.12397.
Groemping, U. (2021). An algorithm for blocking regular fractional factorial 2-level designs with clear two-factor interactions. Computational Statistics and Data Analysis 153, 1-18. doi:10.1016/j.csda.2020.107059. Preprint at Report 3/2019.
plot.igraph
, tkplot
,
plot.common
phimax(7, 2) ## 16 2fis can be clear, if 128 run full factorial is blocked ## into 32 blocks of size 2^2=4 ## X matrices for blocking full factorials ## do not care about which factors have which role X_from_profile(7, 2, c(3,2,2)) # X_from_profile(7, 2, c(2,2,3)) returns same matrix ## ensure specific partition, i.e. specific requirement CIG to be accommodated X <- X_from_parts(7, 2, parts=list(c("A","D","F"), c("B","G"), c("C","E"))) ## blocked full factorial summary(FF_from_X(X)) ## using colpick ## estimable in standard letters requ <- c("BA", "BC", "BD", "BE", "BF", "BG", "BH", "BJ") ## estimability requirement in factor names fn <- Letters[15:23] ## P to X requfn <- requ requfn <- sapply(1:8, function(obj) gsub(Letters[obj], fn[obj], requfn[obj])) ## obtain X matrix for accommodating estimability requirement in 9-4.2 (aus <- colpick("9-4.2", 2, estimable=requ)) ## obtain the same matrix manually with Xcalc XI <- aus$X[,1:5] ## obtain the same matrix manually with Xcalc all(Xcalc(XI, catlg["9-4.2"])==aus$X) ## inspect X matrices generated from XI Xcalc(XI, catlg[nruns(catlg)==32 & nfac(catlg)==9 & res(catlg)>=4]) ## factor permutation needed aus$map ## calculate block generators blockgencreate(aus$X, p=4) ## automatic creation from the design 9-4.2 uses these block generators summary(FrF2(32, 9, blocks=8, estimable=requ, factor.names=fn, alias.block.2fis = TRUE, select.catlg = catlg["9-4.2"]), brief=TRUE) ## can also be reproduced manually (internal function invperm does the permuting) summary(FrF2(design="9-4.2", blocks=blockgencreate(aus$X, p=4), factor.names=fn[FrF2:::invperm(aus$map)], alias.block.2fis = TRUE), brief=TRUE)
phimax(7, 2) ## 16 2fis can be clear, if 128 run full factorial is blocked ## into 32 blocks of size 2^2=4 ## X matrices for blocking full factorials ## do not care about which factors have which role X_from_profile(7, 2, c(3,2,2)) # X_from_profile(7, 2, c(2,2,3)) returns same matrix ## ensure specific partition, i.e. specific requirement CIG to be accommodated X <- X_from_parts(7, 2, parts=list(c("A","D","F"), c("B","G"), c("C","E"))) ## blocked full factorial summary(FF_from_X(X)) ## using colpick ## estimable in standard letters requ <- c("BA", "BC", "BD", "BE", "BF", "BG", "BH", "BJ") ## estimability requirement in factor names fn <- Letters[15:23] ## P to X requfn <- requ requfn <- sapply(1:8, function(obj) gsub(Letters[obj], fn[obj], requfn[obj])) ## obtain X matrix for accommodating estimability requirement in 9-4.2 (aus <- colpick("9-4.2", 2, estimable=requ)) ## obtain the same matrix manually with Xcalc XI <- aus$X[,1:5] ## obtain the same matrix manually with Xcalc all(Xcalc(XI, catlg["9-4.2"])==aus$X) ## inspect X matrices generated from XI Xcalc(XI, catlg[nruns(catlg)==32 & nfac(catlg)==9 & res(catlg)>=4]) ## factor permutation needed aus$map ## calculate block generators blockgencreate(aus$X, p=4) ## automatic creation from the design 9-4.2 uses these block generators summary(FrF2(32, 9, blocks=8, estimable=requ, factor.names=fn, alias.block.2fis = TRUE, select.catlg = catlg["9-4.2"]), brief=TRUE) ## can also be reproduced manually (internal function invperm does the permuting) summary(FrF2(design="9-4.2", blocks=blockgencreate(aus$X, p=4), factor.names=fn[FrF2:::invperm(aus$map)], alias.block.2fis = TRUE), brief=TRUE)
Main effects plots and interaction plots are produced. The other documented functions are not intended for users.
MEPlot(obj, ...) ## S3 method for class 'design' MEPlot(obj, ..., response = NULL) ## Default S3 method: MEPlot(obj, main = paste("Main effects plot for", respnam), pch = 15, cex.xax = par("cex.axis"), cex.yax = cex.xax, mgp.ylab = 4, cex.title = 1.5, cex.main = par("cex.main"), lwd = par("lwd"), las=par("las"), abbrev = 3, select = NULL, ...) IAPlot(obj, ...) ## S3 method for class 'design' IAPlot(obj, ..., response = NULL) ## Default S3 method: IAPlot(obj, main = paste("Interaction plot matrix for", respnam), pch = c(15, 17), cex.lab = par("cex.lab"), cex = par("cex"), cex.xax = par("cex.axis"), cex.yax = cex.xax, cex.title = 1.5, lwd = par("lwd"), las=par("las"), abbrev = 4, select = NULL, show.alias = FALSE, ...) intfind(i, j, mat) check(obj) remodel(obj)
MEPlot(obj, ...) ## S3 method for class 'design' MEPlot(obj, ..., response = NULL) ## Default S3 method: MEPlot(obj, main = paste("Main effects plot for", respnam), pch = 15, cex.xax = par("cex.axis"), cex.yax = cex.xax, mgp.ylab = 4, cex.title = 1.5, cex.main = par("cex.main"), lwd = par("lwd"), las=par("las"), abbrev = 3, select = NULL, ...) IAPlot(obj, ...) ## S3 method for class 'design' IAPlot(obj, ..., response = NULL) ## Default S3 method: IAPlot(obj, main = paste("Interaction plot matrix for", respnam), pch = c(15, 17), cex.lab = par("cex.lab"), cex = par("cex"), cex.xax = par("cex.axis"), cex.yax = cex.xax, cex.title = 1.5, lwd = par("lwd"), las=par("las"), abbrev = 4, select = NULL, show.alias = FALSE, ...) intfind(i, j, mat) check(obj) remodel(obj)
obj |
an experimental design of class |
... |
further arguments to be passed to the default function; |
response |
character string that specifies response variable to be used,
must be an element of |
main |
overall title for the plot assembly |
pch |
Plot symbol number |
cex.xax |
size of x-axis annotation, defaults to |
cex.yax |
size of y-axis annotation, defaults to cex.xax |
mgp.ylab |
horizontal placement of label of vertical axis in |
cex.title |
multiplier for size of overall title (cex.main is multiplied with this factor) |
cex.main |
size of individual plot titles in |
cex.lab |
Size of variable names in diagonal panels of interaction plots
produced by |
cex |
size of plot symbols in interaction plots |
lwd |
line width for plot lines and axes |
las |
orientation for tick mark labels ( |
abbrev |
number of characters shown for factor levels |
select |
vector with position numbers of the main effects to be displayed; |
show.alias |
if TRUE, the interaction plot shows the number
of the list entry from aliases(obj) (cf. |
i |
integer, for internal use only |
j |
integer, for internal use only |
mat |
matrix, for internal use only |
For functions MEPlot
or IAPlot
,
if obj
is a design with at least one response variable
rather than a linear model fit,
the lm
-method for class design
is applied to it with the
required degree (1 or 2),
and the default method for the respective function is afterwards applied to the
resulting linear model.
If the design contains a block factor, the plot functions show non-block effects only.
produces plots of all treatment main effects in the model,
or selected ones if select
is specified
produces plots of all treatment interaction effects in the model,
or selected ones if select
is specified
is an internal function not directly useful for users
is an internal function for checking whether the model complies
with assumptions (fractional factorial of 2-level factors
with full or no aliasing, not partial aliasing;
this implies that Plackett-Burman designs with partial aliasing
of 2-factor interactions give an OK (=TRUE) in check
for
pure main effects models only.)
is an internal function that redoes factor values into -1 and 1 coding, regardless of the contrasts that have been used for the original factors; numerical data are transformed by subtracting the mean and dividing by half the range (max-min), which also transforms them to -1 and 1 coding in the 2-level case (and leads to an error otherwise)
MEPlot
and IAPlot
invisibly return the plotted effects (two-row
matrix or four-row matrix, respectively). If show.alias=TRUE
,
the matrix returned by IAPlot has as the attribute aliasgroups
,
which contains all alias groups (list element number corresponds to
number in the graphics tableau).
The internal function check
is used within other functions for checking
whether the model is a fractional factorial with 2-level factors and
no partial aliasing, as requested for the package to work.
It is applied to remodeled objects only and returns a logical.
If the returned value is FALSE, the calling function fails.
The internal function intfind
returns an integer (length 1 or 0).
It is not useful for users.
The internal function remodel
is applied to a linear model object and
returns a list of two components:
model |
is the redone model with x-variables recoded to numeric -1 and 1 notation and aov objects made into “pure” lm objects |
labs |
is a list preserving the level information from original factors (levels are minus and plus for numerical variables) |
Ulrike Groemping
Box G. E. P, Hunter, W. C. and Hunter, J. S. (2005) Statistics for Experimenters, 2nd edition. New York: Wiley.
FrF2-package
for examples
creates a class catlg catalogue with a single element for use in functions colpick or FrF2
makecatlg(k, gen)
makecatlg(k, gen)
k |
number of base factors spanning a full factorial with the desired number of runs |
gen |
generators as a numeric vector of Yates column numbers |
If generators are available in a different format, they must be transformed to Yates column numbers.
For a character vector genc
with elements like ABC
, ADE
, etc.,
a code for obtaining Yates columns with order preserved is
sapply(1:length(genc), function(obj) which(names(Yates)==genc[obj]))
(a solution with which
applied to the entire vector at once does not preserve the order).
Yet different formats like 123
, 145
, etc., can e.g.
be preprocessed by picking the suitable elements from Letters
, e.g.
paste(Letters[as.numeric(unlist(strsplit("123","")))],collapse="")
.
The function returns a list of class catlg
with a single element.
This package is still under development, but does already provide useful and well-tested results.
Ulrike Groemping
See also FrF2
## Xu's fraction 13-5.2 genXu <- c(127, 143, 179, 85, 150) catXu <- makecatlg(k=8, genXu) colpick(catXu, q=2) ## Godolphin blocking into blocks of size 4 yields 56 clear 2fis FrF2(256, 13, blocks=64, alias.block.2fis=TRUE, select.catlg=catXu)
## Xu's fraction 13-5.2 genXu <- c(127, 143, 179, 85, 150) catXu <- makecatlg(k=8, genXu) colpick(catXu, q=2) ## Godolphin blocking into blocks of size 4 yields 56 clear 2fis FrF2(256, 13, blocks=64, alias.block.2fis=TRUE, select.catlg=catXu)
The function generates Plackett-Burman designs and in some cases other screening designs in run numbers that are a multiple of 4. These designs are particularly suitable for screening a large number of factors, since interactions are not fully aliased with one main effect each but partially aliased. (The design in 8 runs is an exception from this rule.)
pb(nruns, nfactors = nruns - 1, factor.names = if (nfactors <= 50) Letters[1:nfactors] else paste("F", 1:nfactors, sep = ""), default.levels = c(-1, 1), ncenter=0, center.distribute=NULL, boxtyssedal = TRUE, n12.taguchi = FALSE, replications = 1, repeat.only = FALSE, randomize = TRUE, seed = NULL, oldver = FALSE, ...) pb.list
pb(nruns, nfactors = nruns - 1, factor.names = if (nfactors <= 50) Letters[1:nfactors] else paste("F", 1:nfactors, sep = ""), default.levels = c(-1, 1), ncenter=0, center.distribute=NULL, boxtyssedal = TRUE, n12.taguchi = FALSE, replications = 1, repeat.only = FALSE, randomize = TRUE, seed = NULL, oldver = FALSE, ...) pb.list
nruns |
number of runs, must be a multiple of 4 |
nfactors |
number of factors, default is nruns - 1,
and it is recommended to retain this default. |
factor.names |
a character vector of factor names (length up to nfactors)
or a list with |
default.levels |
default levels (vector of length 2) for all factors for which no specific levels are given |
ncenter |
number of center points; |
center.distribute |
the number of positions over which the center points
are to be distributed ; if NULL (default), center points are
distributed over end, beginning, and middle (in that order, if there are fewer than three center points)
for randomized designs, and appended to the end for non-randomized designs.
for more detail, see function |
boxtyssedal |
logical, relevant only for nruns=16. If FALSE, the geometric (=standard) 16 run plan is used. If TRUE, the proposal by Box and Tyssedal is used instead, which has the advantage (for screening) of aliasing each interaction with several main effects, like the other Plackett-Burman designs. |
n12.taguchi |
logical, relevant only for nruns=12. If TRUE, the 12 run design is given in Taguchi order. |
replications |
positive integer number. Default 1 (i.e. each row just once).
If larger, each design run is executed replication times.
If Otherwise (default), the full experiment is first carried out once, then for the second replication and so forth. In case of randomization, each such blocks is randomized separately. In this case, replication variance is more likely suitable for usage as error variance (unless e.g. the same parts are used for replication runs although build variation is important). |
repeat.only |
logical, relevant only if replications > 1. If TRUE,
replications of each run are grouped together
(repeated measurement rather than true replication). The default is
|
randomize |
logical. If TRUE, the design is randomized. This is the default. |
seed |
optional seed for the randomization process |
oldver |
logical. If |
... |
currently not used |
pb
stands for Plackett-Burman. Plackett-Burman designs (Plackett and Burman 1946) are generally
used for screening many variables in relatively few runs, when interest is in
main effects only, at least initially. Different from the regular
fractional factorial designs created by function FrF2
, they
do not perfectly confound interaction terms with main effects but distribute
interaction effects over several main effects. The designs with number of runs
a power of 2 are an exception to this rule: they are just the resolution III
regular fractional factorial designs and are as such not very suitable for
screening because of a high risk of very biased estimates for the main effects
of the factors. Where possible, these are therefore replaced by different designs (cf. below).
For most run numbers, function pb
uses Plackett-Burman designs,
and simply fills columns from left to right.
The generating rows for these designs can be found in the list pb.list
(a 0 entry indicates that the design is constructed by a different method, e.g. doubling).
For 12 runs, the isomorphic design by Taguchi can be requested. For 16 runs, the default is to use the designs suggested by Box and Tyssedal (2001), which up to 14 factors do not suffer from perfect aliasing. For 32 runs, a cyclic design with generating row given in Samset and Tyssedal (1999) is used. For 64 runs, the 32 run design is doubled. For 92 runs, a design is constructed according to the Williamson construction with matrices A, B, C and D from Hedayat and Stufken (1999), p. 160.
Designs up to 100~runs are covered.
Usage of the 8 run design for more than 4 factors is discouraged, as it completely aliases main effects with individual two-factor interactions. It is recommended to use at least the 12 run design instead for screening more than 4 factors.
Value is a data frame of S3 class design
and has attached attributes that can be accessed
by functions desnum
,
run.order
and
design.info
.
The data frame itself contains the design with levels coded as requested.
If no center points have been requested, the design columns are factors with
contrasts -1
and +1
(cf. also contr.FrF2
); in case
of center points, the design columns are numeric.
The following attributes are attached to it:
desnum |
Design matrix in -1/1 coding |
run.order |
three column data frame, first column contains the run number in standard order, second column the run number as randomized, third column the run number with replication number as postfix; useful for switching back and forth between actual and standard run number |
design.info |
list with entries
|
With version 1.0-5 of package FrF2,
design generation for the designs based on doubling has changed (internal function
double.des
). This affected designs for 40,56,64,88,96 runs.
With version 1.3 of package FrF2, this and further behaviors (52, 76) has changed
again, in the interest of improving generalized resolution of desigs produced by function pb
.
For the affected run sizes, package versions from 1.0-5 onwards
cannot exactly reproduce pb designs that have been created with a version before
1.0-5. Package versions from 1.3 onwards reproduce the behavior of versions 1.0-5 to 1.2-10
through option oldver
.
Since R version 3.6.0, the behavior of function sample
has changed
(correction of a biased previous behavior that should not be relevant for the randomization of designs).
For reproducing a randomized design that was produced with an earlier R version,
please follow the steps described with the argument seed
.
Ulrike Groemping
Box, G.E.P. and Tyssedal, J. (2001) Sixteen Run Designs of High Projectivity for Factor Screening. Communications in Statistics - Simulation and Computation 30, 217-228.
Hedayat, A.S., Sloane, N.J.A. and Stufken, J. (1999) Orthogonal Arrays: Theory and Applications, Springer, New York.
Groemping, U. (2014). R Package FrF2 for Creating and Analyzing Fractional Factorial 2-Level Designs. Journal of Statistical Software, 56, Issue 1, 1-56. https://www.jstatsoft.org/v56/i01/.
Mee, R. (2009). A Comprehensive Guide to Factorial Two-Level Experimentation. New York: Springer.
Plackett, R.L.; Burman, J.P. (1946) The design of optimum multifactorial experiments. Biometrika 33, 305-325.
Samset, O.; Tyssedal, J. (1999) Two-level designs with good projection properties. Technical Report 12, Department of Mathematical Sciences, The Norwegian University of Science and Technology, Norway.
Williamson, J. (1946) Determinants whose elements are 0 and 1. American Mathematical Monthly 53, 427-434.
See also FrF2
for regular fractional factorial designs,
generalized.word.length
for functions length3
and length4
used in examples
pb(12,randomize=FALSE) pb(12,randomize=FALSE,n12.taguchi=TRUE) pb(20,seed=29869) pb(16,factor.names=list(A="",B="",C="",D=c("min","max"), E="",F="",G="",H="",J=c("new","old"))) pb(8,default.levels=c("current","new")) test <- pb(40) ## design created by doubling the 20 run design pb(12, ncenter=6) ## 6 center points with default placement ## Not run: ## note: designs in 40, 56, 64, 88, and 96 runs are resolution IV, ## if the number of factors is up to nruns/2 - 1, e.g.: plan1 <- pb(40, 19) length3(plan1) ## 0 generalized words of length 3 length4(plan1) ## 228 generalized words of length 4 ## they can be made resolution IV by oldver=TRUE for ## nfactors=nruns/2, e.g.: plan2 <- pb(40, 20) plan3 <- pb(40, 20, oldver=TRUE) length3(plan2) ## 9 generalized words of length 3 length3(plan3) ## 0 generalized words of length 3 length4(plan3) ## 285 generalized words of length 4 ## note: designs in 52, 76, and 100 runs are almost resolution IV, ## if the number of factors is up to nruns/2 - 1, e.g.: plan4 <- pb(52, 25) GR(plan4) ## generalized resolution 3.92 ## note: versions >1.3 avoid complete and heavy aliasing of triples of factors ## for up to nruns-2 factors for 40, 52, 56, 64, 76, 88, 92 and 96 runs ## (the same for 100 runs, which were not implemented before version 1.3) plan5 <- pb(40, 38) plan6 <- pb(40, 38, oldver=TRUE) GR(plan5) ## generalized resolution 3.4 GR(plan6) ## generalized resolution 3 plan7 <- pb(52, 50) plan8 <- pb(52, 50, oldver=TRUE) GR(plan7) ## generalized resolution 3.62 GR(plan8) ## generalized resolution 3.15 ## End(Not run)
pb(12,randomize=FALSE) pb(12,randomize=FALSE,n12.taguchi=TRUE) pb(20,seed=29869) pb(16,factor.names=list(A="",B="",C="",D=c("min","max"), E="",F="",G="",H="",J=c("new","old"))) pb(8,default.levels=c("current","new")) test <- pb(40) ## design created by doubling the 20 run design pb(12, ncenter=6) ## 6 center points with default placement ## Not run: ## note: designs in 40, 56, 64, 88, and 96 runs are resolution IV, ## if the number of factors is up to nruns/2 - 1, e.g.: plan1 <- pb(40, 19) length3(plan1) ## 0 generalized words of length 3 length4(plan1) ## 228 generalized words of length 4 ## they can be made resolution IV by oldver=TRUE for ## nfactors=nruns/2, e.g.: plan2 <- pb(40, 20) plan3 <- pb(40, 20, oldver=TRUE) length3(plan2) ## 9 generalized words of length 3 length3(plan3) ## 0 generalized words of length 3 length4(plan3) ## 285 generalized words of length 4 ## note: designs in 52, 76, and 100 runs are almost resolution IV, ## if the number of factors is up to nruns/2 - 1, e.g.: plan4 <- pb(52, 25) GR(plan4) ## generalized resolution 3.92 ## note: versions >1.3 avoid complete and heavy aliasing of triples of factors ## for up to nruns-2 factors for 40, 52, 56, 64, 76, 88, 92 and 96 runs ## (the same for 100 runs, which were not implemented before version 1.3) plan5 <- pb(40, 38) plan6 <- pb(40, 38, oldver=TRUE) GR(plan5) ## generalized resolution 3.4 GR(plan6) ## generalized resolution 3 plan7 <- pb(52, 50) plan8 <- pb(52, 50, oldver=TRUE) GR(plan7) ## generalized resolution 3.62 GR(plan8) ## generalized resolution 3.15 ## End(Not run)
This help page documents the statistical and algorithmic details of split-plot designs in FrF2
A split-plot design is similar to a block
ed design, with the difference that
there are also factors of interest that can be only changed on block level (so-called whole
plot factors). The blocks are called “plots” in the context of split-plot designs.
The factors that can (and should!) be varied within a plot are called split-plot factors.
Note that the experiment provides more information on split-plot factors than on whole-plot factors.
Warning: In terms of analysis, split-plot designs would have to be treated by advanced random effects models, but often are not. At the very least, the user must be aware that all whole-plot effects (i.e. effects on columns that only change between plots) are (likely to be) more variable than split-plot effects so that e.g. it does not necessarily mean anything if they stick out in a normal or half-normal effects plot.
Designs for hard-to-change factors are also treated by the split-plot approach in function FrF2
,
although they are not quite split-plot designs: The are non-randomized split-plot designs arranged in an order
such that the first whole-plot factors have as few as possible changes. This gives very poor information on
these first whole-plot factors (which in the extreme are only changed once or twice),
if there is variability involved with setting the factor levels.
If hard-to-change factors can be implemented as true whole-plot factors with randomization, this is by far preferrable from a statistical
point of view (but may nevertheless be rejected from a feasibility point of view, as the necessary changes may seem unaffordable).
For design generation, there are two principal ways to handle split-plot designs, manual definition
(i.e. the user specifies exactly which columns are to be used for which purpose) and automatic
definition. Each situation has its specifics. These are detailed below. For users with
not so much mathematical/statistical background, it will often be best to use the automatic way,
specifying the treatement factors of interest via nfactors
or factor.names
and a single number for WPs
.
Users with more mathematical background may want to use the manual definitions, perhaps
in conjunction with published catalogues of good split-plot designs, or
after inspecting possibilities with function splitpick
.
The user can specify a design with the design
or the generators
option
and specify manually with the WPfacs
option, which factors are whole plot factors
(i.e. factors that do not change within a plot).
The other factors become split-plot factors (i.e. factors that do change within a plot).
If the user chooses this route, WPfacs
must be character vectors of factor names, factor letters,
factor numbers preceded by capital F, or a vector or list of factor position numbers (NOT: Yates column numbers).
Caution: It is the users responsibility to ensure a
good choice of split-plot design (e.g. by using a catalogued design from Huang, Chen and Voelkel 1998,
Bingham and Sitter 2003, or Bingham Schoen and Sitter 2004).
In case of a user-mistake such that the resulting design is not a split-plot design with
the alleged number of whole plots, an error is thrown.
As mentioned above, split-plot designs differ from block designs by the fact that the block main
effects are purely nuisance parameters which are assumed (based on prior knowledge)
to be relevant but are not of interest, while the plots are structured by
nfac.WP
whole plot factors, which are of interest.
The user has to decide on a number of whole plots (WPs
) as well as
the number of whole plot factors nfac.WP
.
If log2(WPs) <= nfac.WP <= WPs-1
, it is obviously in principle possible to accomodate the
desired number of whole plot factors in the desired number of whole plots. If nfac.WP > WPs/2
,
the base design for the split-plot structure has to be of resolution III. Sometimes,
subject matter considerations limit whole plot sizes, and there are only few interesting
whole plot factors, i.e. nfac.WP < log2(WPs)
.
In this case, it is of course nevertheless necessary to
have a total of log2(WPs)
whole plot construction factors;
the missing log2(WPs) - nfac.WP
factors are added
to the design (names starting with WP
), and nfactors
is increased accordingly.
In all cases, the first nfac.WPs
user-specified factors are treated as whole plot factors, the
remaining factors as split-plot factors.
From there, function FrF2
proceeds like in the blocked situation by starting
with the best design and working its way down to worse designs, if the best design cannot
accomodate the desired split-plot structure. For each design, function FrF2
calls function splitpick
, which permutes base factors until
the requested whole plot / split-plot structure is achieved, or until impossibility for
this design with these base factors has been ascertained. In the latter case, function FrF2
proceeds to the next best design and so forth.
If several competing split-plot designs based on the same base design are found,
the best possible resolution among the first check.WPs
such designs is chosen.
No further criteria are automatically implemented, and no more than check.WPs
designs are checked. If not satisfied with the structure of the whole plot portion of the experiment,
increasing check.WPs
vs. the default 10 may help.
Expert users may want to inspect possibilities,
using function splitpick
directly.
Note that the algorithm does not necessarily find an existing split-plot design. It has been checked out which catalogued designs it can find: designs for all catalogued situations from Bingham and Sitter (2003) have been found, as well as for most catalogued situations from Huang, Chen and Voelkel (1998). Occasionally, a better design than catalogued has been found, e.g. for 4 whole plot and 10 split plot factors in 32 runs with 16 whole plots, the design found by the algorithm is resolution IV, while Huang, Chen and Voelkel propose a resolution III design. The algorithm has the largest difficulties with extreme designs in the sense that a large number of whole plots with a small number of whole plot factors are to be accomodated; thus it does not find designs for the more extreme situations in Bingham, Schoen and Sitter (2004).
Please contact me with any suggestions for improvements.
Ulrike Groemping
Bingham, D.R., Schoen, E.D. and Sitter, R.R. (2004). Designing Fractional Factorial Split-Plot Experiments with Few Whole-Plot Factors. Applied Statistics 53, 325-339.
Bingham, D. and Sitter, R.R. (2003). Fractional Factorial Split-Plot Designs for Robust Parameter Experiments. Technometrics 45, 80-89.
Chen, J., Sun, D.X. and Wu, C.F.J. (1993) A catalogue of 2-level and 3-level orthogonal arrays. International Statistical Review 61, 131-145.
Cheng, C.-S. and Tsai, P.-W. (2009). Optimal two-level regular fractional factorial block and split-plot designs. Biometrika 96, 83-93.
Huang, P., Chen, D. and Voelkel, J.O. (1998). Minimum-Aberration Two-Level Split-Plot Designs. Technometrics 40, 314-326.
See Also FrF2
for regular fractional factorials,
catlg
for the Chen, Sun, Wu catalogue of designs
and some accessor functions,
and block
for the statistical aspects of blocked designs.
########## hard to change factors #################### ## example from Bingham and Sitter Technometrics 19999 ## MotorSpeed, FeedMode,FeedSizing,MaterialType are hard to change BS.ex <- FrF2(16,7,hard=4, factor.names=c("MotorSpeed", "FeedMode","FeedSizing","MaterialType", "Gain","ScreenAngle","ScreenVibLevel"), default.levels=c("-","+")) design.info(BS.ex) BS.ex ## NOTE: the design has 8 whole plots. ## The first hard-to-change factors have very few changes only ## between whole plots. ## A conscious and honest decision is required whether it is ## acceptable for the situation at hand not to reset them! ## A proper split-plot design with resetting all whole plot factors ## for each whole plot would be strongly preferred from a ## statistical point of view. ########## automatic generation for split plot ########## ## 3 control factors, 5 noise factors, control factors are whole plot factors ## 8 plots desired in a total of 32 runs ## Bingham Sitter 2003 BS.ex2a <- FrF2(32, 8, WPs=8, nfac.WP=3, factor.names=c(paste("C",1:3,sep=""), paste("N",1:5,sep="")),randomize=TRUE) ## manual generation of this same design BS.ex2m <- FrF2(32, 8, generators=c("ABD","ACD","BCDE"),WPs=8, WPfacs=c("C1","C2","C3"), nfac.WP=3, factor.names=c(paste("C",1:3,sep=""),paste("N",1:5,sep="")),randomize=TRUE) ## design with few whole plot factors ## 2 whole plot factors, 7 split plot factors ## 8 whole plots, i.e. one extra WP factor needed BSS.cheese.exa <- FrF2(32, 9, WPs=8, nfac.WP=2, factor.names=c("A","B","p","q","r","s","t","u","v")) design.info(BSS.cheese.exa) ## manual generation of the design used by Bingham, Schoen and Sitter ## note that the generators include a generator for the 10th spplitting factor ## s= ABq, t = Apq, u = ABpr and v = Aqr, splitting factor rho=Apqr BSS.cheese.exm <- FrF2(32, gen=list(c(1,2,4),c(1,3,4),c(1,2,3,5),c(1,4,5),c(1,3,4,5)), WPs=8, nfac.WP=3, WPfacs=c(1,2,10), factor.names=c("A","B","p","q","r","s","t","u","v","rho")) design.info(BSS.cheese.exm)
########## hard to change factors #################### ## example from Bingham and Sitter Technometrics 19999 ## MotorSpeed, FeedMode,FeedSizing,MaterialType are hard to change BS.ex <- FrF2(16,7,hard=4, factor.names=c("MotorSpeed", "FeedMode","FeedSizing","MaterialType", "Gain","ScreenAngle","ScreenVibLevel"), default.levels=c("-","+")) design.info(BS.ex) BS.ex ## NOTE: the design has 8 whole plots. ## The first hard-to-change factors have very few changes only ## between whole plots. ## A conscious and honest decision is required whether it is ## acceptable for the situation at hand not to reset them! ## A proper split-plot design with resetting all whole plot factors ## for each whole plot would be strongly preferred from a ## statistical point of view. ########## automatic generation for split plot ########## ## 3 control factors, 5 noise factors, control factors are whole plot factors ## 8 plots desired in a total of 32 runs ## Bingham Sitter 2003 BS.ex2a <- FrF2(32, 8, WPs=8, nfac.WP=3, factor.names=c(paste("C",1:3,sep=""), paste("N",1:5,sep="")),randomize=TRUE) ## manual generation of this same design BS.ex2m <- FrF2(32, 8, generators=c("ABD","ACD","BCDE"),WPs=8, WPfacs=c("C1","C2","C3"), nfac.WP=3, factor.names=c(paste("C",1:3,sep=""),paste("N",1:5,sep="")),randomize=TRUE) ## design with few whole plot factors ## 2 whole plot factors, 7 split plot factors ## 8 whole plots, i.e. one extra WP factor needed BSS.cheese.exa <- FrF2(32, 9, WPs=8, nfac.WP=2, factor.names=c("A","B","p","q","r","s","t","u","v")) design.info(BSS.cheese.exa) ## manual generation of the design used by Bingham, Schoen and Sitter ## note that the generators include a generator for the 10th spplitting factor ## s= ABq, t = Apq, u = ABpr and v = Aqr, splitting factor rho=Apqr BSS.cheese.exm <- FrF2(32, gen=list(c(1,2,4),c(1,3,4),c(1,2,3,5),c(1,4,5),c(1,3,4,5)), WPs=8, nfac.WP=3, WPfacs=c(1,2,10), factor.names=c("A","B","p","q","r","s","t","u","v","rho")) design.info(BSS.cheese.exm)
Functions to restructure a fractional factorial by permuting the base factors such that the leftmost base factors have a suitable alias structure for the problem at hand; meant for expert users
splitpick(k, gen, k.WP, nfac.WP, show=10) leftadjust(k, gen, early=NULL, show=10)
splitpick(k, gen, k.WP, nfac.WP, show=10) leftadjust(k, gen, early=NULL, show=10)
k |
the number of base factors (designs have |
gen |
vector of generating columns from Yates matrix |
k.WP |
integer number of base factors used for whole plot generation;
there will be |
nfac.WP |
integer number of whole plot factors, must not be smaller than |
show |
numeric integer indicating how many results are to be shown;
for function |
early |
number that indicates how many “leftmost” factors are
needed in the design; used by |
These functions exploit the fact that a factorial design can be arranged such
that the 2^k.WP-1
leftmost columns have exactly 2^k.WP
different patterns. They can thus accomodate whole plot effects if 2^k.WP
plots are available; also, with a specially rearranged version of the Yates matrix,
the leftmost columns can have particularly few or particularly many level changes,
cf. e.g. Cheng, Martin and Tang 1998.
By permuting the k
base factors , the functions try to find 2^k.WP
ones that accomodate the current needs, if taken as the first base factors. They are
used by function FrF2
, if a user requests an automatically-generated
split-plot design or a design with some factors declared hard-to-change.
There may be a possibility to better accomodate the experimenters needs within
a given design by trying different sets of base factors. This is not done
in these functions. Also, custom user needs may be better fulfilled, if an expert
user directly uses one of these functions for inspecting the possibilities, rather
than relying on the automatic selection routine in function FrF2
.
Both functions output a list of entries with information on at most show
suitable
permutations. splitpick
ends with an error, if no suitable
solution can be found.
orig |
original generator column numbers |
basics |
named vector with the following entries: |
perms |
matrix with rows containing permutations of base factors |
res.WP |
for |
maxpos |
for |
k.early |
for |
gen |
matrix the rows of which contain the generator columns for the respective rows of perms |
Ulrike Groemping
Cheng, C.-S., Martin, R.J., and Tang, B. (1998). Two-level factorial designs with extreme numbers of level changes. Annals of Statistics 26, 1522-1539.
See Also FrF2
## leftadjusting MA design from table 6.22 in BHH2, 9 factors, 32 runs ## NOTE: nevertheless not as well left-adjusted as the isomorphic design 9-4.1 from catlg leftadjust(5,c(30,29,27,23)) ## with option early=4 (i.e. 4 columns as early as possible are requested) leftadjust(5,c(30,29,27,23),early=4) leftadjust(5,catlg$'9-4.1'$gen,early=4) ## look for a split plot design in 32 runs with 7 factors, ## 3 of which are whole plot factors, ## and 8 plots splitpick(5,catlg$'7-2.1'$gen,nfac.WP=3,k.WP=3)
## leftadjusting MA design from table 6.22 in BHH2, 9 factors, 32 runs ## NOTE: nevertheless not as well left-adjusted as the isomorphic design 9-4.1 from catlg leftadjust(5,c(30,29,27,23)) ## with option early=4 (i.e. 4 columns as early as possible are requested) leftadjust(5,c(30,29,27,23),early=4) leftadjust(5,catlg$'9-4.1'$gen,early=4) ## look for a split plot design in 32 runs with 7 factors, ## 3 of which are whole plot factors, ## and 8 plots splitpick(5,catlg$'7-2.1'$gen,nfac.WP=3,k.WP=3)