| Title: | Clustered Covariate Regression |
|---|---|
| Description: | Clustered covariate regression enables estimation and inference in both linear and non-linear models with linear predictor functions even when the design matrix is column rank deficient. Routines in this package implement algorithms in Soale and Tsyawo (2019) <doi:10.13140/RG.2.2.32355.81441>. |
| Authors: | Emmanuel S Tsyawo [aut, cre], Abdul-Nasah Soale [aut] |
| Maintainer: | Emmanuel S Tsyawo <[email protected]> |
| License: | GPL-2 |
| Version: | 1.1.0 |
| Built: | 2024-11-20 06:40:03 UTC |
| Source: | CRAN |
A function for constructing functions for concrete classes of models for the chmod() family of
of functions.
c_chmod(Y, X, modclass = "lm")c_chmod(Y, X, modclass = "lm")
Y |
vector of the outcome variable |
X |
matrix of covariates; excepting intercepts 1's |
modclass |
the class of model. Currently, "lm" for linear regression, "logit" (logit model), "qreg" (quantile regression), "probit" (probit model), "gammainverse" (gamma with inverse link), "gammalog" (gamma with log link), "poissonlog" (poisson model with log link), "poissonidentity" (poisson with identity link), "poissonsqrt" (poisson with sqrt link), "negbin" (negative binomial) are supported. |
object an object list with class attribute modclass.
CCRls runs regressions with potentially more covariates than observations.
See c_chmod() for the list of models supported.
CCRls(Y, X, kap = 0.1, modclass = "lm", tol = 1e-06, reltol = TRUE, rndcov = NULL, report = NULL, ...)CCRls(Y, X, kap = 0.1, modclass = "lm", tol = 1e-06, reltol = TRUE, rndcov = NULL, report = NULL, ...)
Y |
vector of dependent variable Y |
X |
design matrix (without intercept) |
kap |
maximum number of parameters to estimate in each active sequential step,
as a fraction of the less of total number of observations n or number of covariates p.
i.e. |
modclass |
a string denoting the desired the class of model. See c_chmod for details. |
tol |
level of tolerance for convergence; default |
reltol |
a logical for relative tolerance instead of level. Defaults at TRUE |
rndcov |
seed for randomising assignment of covariates to partitions; default |
report |
number of iterations after which to report progress; default |
... |
additional arguments to be passed to the model |
betas parameter estimates (intercept first),
iter number of iterations,
dev increment in the objective function value at convergence
fval objective function value at convergence
set.seed(14) #Generate data N = 1000; (bets = rep(-2:2,4)); p = length(bets); X = matrix(rnorm(N*p),N,p) Y = cbind(1,X)%*%matrix(c(0.5,bets),ncol = 1) CCRls(Y,X,kap=0.1,modclass="lm",tol=1e-6,reltol=TRUE,rndcov=NULL,report=8)set.seed(14) #Generate data N = 1000; (bets = rep(-2:2,4)); p = length(bets); X = matrix(rnorm(N*p),N,p) Y = cbind(1,X)%*%matrix(c(0.5,bets),ncol = 1) CCRls(Y,X,kap=0.1,modclass="lm",tol=1e-6,reltol=TRUE,rndcov=NULL,report=8)
This function is a wrapper for linrclus. It requires less input.
CCRls.coord(Y, X, k, nC = 1, ...)CCRls.coord(Y, X, k, nC = 1, ...)
Y |
vector of outcome variable |
X |
matrix of covariates. Should not include 1's for the intercept |
k |
number of clusters |
nC |
first nC-1 covariates in X not to cluster. Must be at least 1 for the intercept |
... |
additional parameters to be passed to lm |
mobj the low dimension lm regression object
clus cluster assignments of covariates (excluding the first nC
covariates - including the intercept 1)
set.seed(14) #Generate data N = 1000; (bets = rep(-2:2,4)); p = length(bets); X = matrix(rnorm(N*p),N,p) Y = cbind(1,X)%*%matrix(c(0.5,bets),ncol = 1) CCRls.coord(Y,X,k=5,nC=1)set.seed(14) #Generate data N = 1000; (bets = rep(-2:2,4)); p = length(bets); X = matrix(rnorm(N*p),N,p) Y = cbind(1,X)%*%matrix(c(0.5,bets),ncol = 1) CCRls.coord(Y,X,k=5,nC=1)
CCRseqk runs regressions with potentially more covariates than observations with
k clusters. See c_chmod() for the list of models supported.
CCRseqk(Y, X, k, nC = 1, kap = 0.1, modclass = "lm", tol = 1e-06, reltol = TRUE, rndcov = NULL, report = NULL, ...)CCRseqk(Y, X, k, nC = 1, kap = 0.1, modclass = "lm", tol = 1e-06, reltol = TRUE, rndcov = NULL, report = NULL, ...)
Y |
vector of dependent variable Y |
X |
design matrix (without intercept) |
k |
number of clusters |
nC |
first |
kap |
maximum number of parameters to estimate in each active sequential step,
as a fraction of the less of total number of observations n or number of covariates p.
i.e. |
modclass |
a string denoting the desired the class of model. See c_chmod for details. |
tol |
level of tolerance for convergence; default |
reltol |
a logical for relative tolerance instead of level. Defaults at TRUE |
rndcov |
seed for randomising assignment of covariates to partitions; default |
report |
number of iterations after which to report progress; default |
... |
additional arguments to be passed to the model |
a list of objects
mobj low dimensional model object of class lm, glm, or rq (depending on modclass)
clus cluster assignments of covariates
iter number of iterations
dev decrease in the function value at convergence
set.seed(14) #Generate data N = 1000; (bets = rep(-2:2,4)/2); p = length(bets); X = matrix(rnorm(N*p),N,p) Y = cbind(1,X)%*%matrix(c(0.5,bets),ncol = 1); nC=1 zg=CCRseqk(Y,X,k=5,nC=nC,kap=0.1,modclass="lm",tol=1e-6,reltol=TRUE,rndcov=NULL,report=8) (del=zg$mobj$coefficients) # delta (bets = c(del[1:nC],(del[-c(1:nC)])[zg$clus])) #construct betaset.seed(14) #Generate data N = 1000; (bets = rep(-2:2,4)/2); p = length(bets); X = matrix(rnorm(N*p),N,p) Y = cbind(1,X)%*%matrix(c(0.5,bets),ncol = 1); nC=1 zg=CCRseqk(Y,X,k=5,nC=nC,kap=0.1,modclass="lm",tol=1e-6,reltol=TRUE,rndcov=NULL,report=8) (del=zg$mobj$coefficients) # delta (bets = c(del[1:nC],(del[-c(1:nC)])[zg$clus])) #construct beta
A generic S3 function as wrapper for internal R routines for classes of models implemented in this package. See details c_chmod for the list of classes supported.
chmod(object, ...)chmod(object, ...)
object |
the object to be passed to the concrete class constructor |
... |
additional paramters to be passed to the internal routine |
A gamma regression implementation for the "gammainverse" class. It uses glm
with the Gamma link function set to "inverse"
## S3 method for class 'gammainverse' chmod(object, ...)## S3 method for class 'gammainverse' chmod(object, ...)
object |
a list of Y - outcome variable and X - design matrix of class "probit" |
... |
additional parameters to be passed to |
fitted model object
chmod(c_chmod(Y=women$height,X=women$weight,modclass="gammainverse"))chmod(c_chmod(Y=women$height,X=women$weight,modclass="gammainverse"))
A gamma regression implementation for the "gammalog" class. It uses glm
with the Gamma link function set to "log"
## S3 method for class 'gammalog' chmod(object, ...)## S3 method for class 'gammalog' chmod(object, ...)
object |
a list of Y - outcome variable and X - design matrix of class "probit" |
... |
additional parameters to be passed to |
fitted model object
chmod(c_chmod(Y=women$height,X=women$weight,modclass="gammalog"))chmod(c_chmod(Y=women$height,X=women$weight,modclass="gammalog"))
A linear regression implementation for the "lm" class. It uses lm
## S3 method for class 'lm' chmod(object, ...)## S3 method for class 'lm' chmod(object, ...)
object |
a list of Y - outcome variable and X - design matrix of class "lm" |
... |
additional parameters to be passed to |
fitted model object
chmod(c_chmod(Y=women$height,X=women$weight,modclass="lm"))chmod(c_chmod(Y=women$height,X=women$weight,modclass="lm"))
A logit regression implementation for the "logit" class. It uses glm
with the binomial link function set to "logit"
## S3 method for class 'logit' chmod(object, ...)## S3 method for class 'logit' chmod(object, ...)
object |
a list of Y - outcome variable and X - design matrix of class "logit" |
... |
additional parameters to be passed to |
fitted model object
chmod(c_chmod(Y=women$height<=50,X=women$weight,modclass="logit"))chmod(c_chmod(Y=women$height<=50,X=women$weight,modclass="logit"))
A negative binomial regression implementation for the "negbin" class. It uses glm.nb
## S3 method for class 'negbin' chmod(object, ...)## S3 method for class 'negbin' chmod(object, ...)
object |
a list of Y - outcome variable and X - design matrix of class "negbin" |
... |
additional parameters to be passed to |
fitted model object
A poisson regression implementation for the "poissonidentity" class. It uses glm
with the poisson link function set to "identity"
## S3 method for class 'poissonidentity' chmod(object, ...)## S3 method for class 'poissonidentity' chmod(object, ...)
object |
a list of Y - outcome variable and X - design matrix of class "poissonidentity" |
... |
additional parameters to be passed to |
fitted model object
chmod(c_chmod(Y=women$height,X=women$weight,modclass="poissonidentity"))chmod(c_chmod(Y=women$height,X=women$weight,modclass="poissonidentity"))
A poisson regression implementation for the "poissonlog" class. It uses glm
with the poisson link function set to "log"
## S3 method for class 'poissonlog' chmod(object, ...)## S3 method for class 'poissonlog' chmod(object, ...)
object |
a list of Y - outcome variable and X - design matrix of class "poissonlog" |
... |
additional parameters to be passed to |
fitted model object
chmod(c_chmod(Y=women$height,X=women$weight,modclass="poissonlog"))chmod(c_chmod(Y=women$height,X=women$weight,modclass="poissonlog"))
A poisson regression implementation for the "poissonsqrt" class. It uses glm
with the poisson link function set to "sqrt"
## S3 method for class 'poissonsqrt' chmod(object, ...)## S3 method for class 'poissonsqrt' chmod(object, ...)
object |
a list of Y - outcome variable and X - design matrix of class "poissonsqrt" |
... |
additional parameters to be passed to |
fitted model object
chmod(c_chmod(Y=women$height,X=women$weight,modclass="poissonsqrt"))chmod(c_chmod(Y=women$height,X=women$weight,modclass="poissonsqrt"))
A probit regression implementation for the "probit" class. It uses glm
with the binomial link set to "probit"
## S3 method for class 'probit' chmod(object, ...)## S3 method for class 'probit' chmod(object, ...)
object |
a list of Y - outcome variable and X - design matrix of class "probit" |
... |
additional parameters to be passed to |
fitted model object
chmod(c_chmod(Y=women$height<=50,X=women$weight,modclass="probit"))chmod(c_chmod(Y=women$height<=50,X=women$weight,modclass="probit"))
A quantile regression implementation for the "qreg" class. It uses rq
## S3 method for class 'qreg' chmod(object, ...)## S3 method for class 'qreg' chmod(object, ...)
object |
a list of Y - outcome variable and X - design matrix of class "qreg" |
... |
additional parameters to be passed to |
fitted model object
chmod(c_chmod(Y=women$height,X=women$weight,modclass="qreg"),tau=0.45)chmod(c_chmod(Y=women$height,X=women$weight,modclass="qreg"),tau=0.45)
A deterministic clustering device of vector elements into k clusters
dcluspar(k, vec)dcluspar(k, vec)
k |
number of clusters |
vec |
the vector of real valued elements |
clus integer assignment of corresponding elements in vec in up to k clusters
set.seed(2); (v=c(rnorm(4,0,0.5),rnorm(3,3,0.5))[sample(1:7)]) dcluspar(k=2,vec = v)set.seed(2); (v=c(rnorm(4,0,0.5),rnorm(3,3,0.5))[sample(1:7)]) dcluspar(k=2,vec = v)
Minimising a continuous univariate function using the golden section search algorithm.
goldensearch(fn, interval, tol = 1)goldensearch(fn, interval, tol = 1)
fn |
the function; should be scalar valued |
interval |
a vector containing the lower and upper bounds of search |
tol |
tolerance level for convergence |
a list of objects
k: minimiser
value: mimimum value
iter: number of iterations before convergence
iterfn: number of function evaluations
fn = function(x) (x-1)^2; goldensearch(fn=fn,interval=c(-2,3),tol=1)fn = function(x) (x-1)^2; goldensearch(fn=fn,interval=c(-2,3),tol=1)
This function conducts an integer golden search minimisation of a univariate function.
goldopt(fn, interval, tol = 1)goldopt(fn, interval, tol = 1)
fn |
function to be minimised. fn should return a list, with fval as the function value. |
interval |
a vector of length two containing the minimum and maximum interger values within which to search for the minimiser. |
tol |
the tolerance level. Defaults at 1 |
k minimiser of fn()
crit the minimum
iter total number of iterations
iterfn total number of function evaluations of fn()
fobj an object of the function minimisation
key a logical for warning if fobj may not correspond to k
set.seed(14) #Generate data N = 1000; (bets = rep(-2:2,4)); p = length(bets); X = matrix(rnorm(N*p),N,p) Y = cbind(1,X)%*%matrix(c(0.5,bets),ncol = 1) fn=function(k){du=CCRls.coord(Y,X,k=k,nC=1) return(list(fval=BIC(du$mobj),obj=du))} goldopt(fn=fn,interval=c(2,7),tol=1)set.seed(14) #Generate data N = 1000; (bets = rep(-2:2,4)); p = length(bets); X = matrix(rnorm(N*p),N,p) Y = cbind(1,X)%*%matrix(c(0.5,bets),ncol = 1) fn=function(k){du=CCRls.coord(Y,X,k=k,nC=1) return(list(fval=BIC(du$mobj),obj=du))} goldopt(fn=fn,interval=c(2,7),tol=1)
Covariate assignment to k clusters using the coordinate descent algorithm. This
function is a wrapper for the C function linreg_coord_clus
linrclus(Y, X, k, coefs, clus, clusmns, nC = 1, x = FALSE)linrclus(Y, X, k, coefs, clus, clusmns, nC = 1, x = FALSE)
Y |
vector of outcome variable |
X |
matrix of covariates. Should not include 1's for the intercept |
k |
number of clusters |
coefs |
vector of coefficients as starting values. Should not include the intercept. |
clus |
vector of covariate cluster assignments as starting values |
clusmns |
vector k cluster parameter centers |
nC |
first nC-1 covariates in X not to cluster. Must be at least 1 for the intercept |
x |
a logical for returning the design matrix |
clus cluster assignments
coefs vector of coefficients as starting values
clusmns vector of cluster means
set.seed(14) #Generate data N = 1000; (bets = rep(-2:2,4)); p = length(bets); X = matrix(rnorm(N*p),N,p) Y = cbind(1,X)%*%matrix(c(0.5,bets),ncol = 1) begin_v<- rep(NA,p) for (j in 1:p) { begin_v[j] = stats::coef(lm(Y~X[,j]))[2] } set.seed(12); klus_obj<- kmeans(begin_v,centers = 5) linrclus(Y,X,k=5,coefs=c(0,begin_v),clus=klus_obj$cluster,clusmns=klus_obj$centers)set.seed(14) #Generate data N = 1000; (bets = rep(-2:2,4)); p = length(bets); X = matrix(rnorm(N*p),N,p) Y = cbind(1,X)%*%matrix(c(0.5,bets),ncol = 1) begin_v<- rep(NA,p) for (j in 1:p) { begin_v[j] = stats::coef(lm(Y~X[,j]))[2] } set.seed(12); klus_obj<- kmeans(begin_v,centers = 5) linrclus(Y,X,k=5,coefs=c(0,begin_v),clus=klus_obj$cluster,clusmns=klus_obj$centers)
This function creates the design matrix for a latent network structure using a balanced panel
netdat(datf, Y, X, Wi, W = NULL, panvar, tvar, factors, scaling = TRUE, unicons = TRUE)netdat(datf, Y, X, Wi, W = NULL, panvar, tvar, factors, scaling = TRUE, unicons = TRUE)
datf |
the entire data frame of balanced panel with NT rows of unit-time observations |
Y |
dependent variable in the data frame datf |
X |
the covariate(s) generating spillovers |
Wi |
other unit-varying (can be time-invariant) control variables |
W |
global variables. these are only time varying but are common to all units. eg. GDP for individual/state-level data. Note that W has to be a vector of length T so cannot be in the data frame datf |
panvar |
the panel variable eg. unique person/firm identifiers |
tvar |
time variable, eg. years |
factors |
a vector of characters of factors in the data |
scaling |
a logical indicating whether non-discrete covariates should be scaled by their standard deviations |
unicons |
a logical indicating whether to include unit-specific constant term |
Y vector of dependent variables
X a block matrix of spillover matrix ( x )
Wm a matrix corresponding to covariate Wi
Wf a matrix of dummies corresponding to factors