ptcompare compares
two samples of continuous multivariate observations. It implements the
hidden Markov Pólya tree of Awaya and Ma (2024), a nonparametric
Bayesian model whose structure is a recursive partition of the data
space.
This vignette fits three comparisons: a simulation in two variables where the difference is confined to a small part of the plane, a set of 30 measurements of cell nuclei where the groups are two diagnoses, and a pair of samples drawn from the same distribution.
Both groups are standard normal in two variables,
marker_A and marker_B. They differ in that
about 15% of group 2 is drawn instead from a narrow cluster near
(1.6, 1.6).
set.seed(11)
n <- 1000L
x <- matrix(rnorm(2 * n), ncol = 2,
dimnames = list(NULL, c("marker_A", "marker_B")))
y <- matrix(rnorm(2 * n), ncol = 2,
dimnames = list(NULL, c("marker_A", "marker_B")))
extra <- runif(n) < 0.15
y[extra, ] <- cbind(rnorm(sum(extra), 1.6, 0.3),
rnorm(sum(extra), 1.6, 0.3))
sum(extra)
#> [1] 144Over most of the plane the two samples are drawn from the same distribution and lie on top of one another. The cluster occupies a small region in the upper right, and it belongs to one group only.
plot(rbind(x, y), type = "n", xlab = "marker_A", ylab = "marker_B")
points(x, pch = 0, col = "#123E7C", cex = 0.6)
points(y, pch = 2, col = "#B65D35", cex = 0.6)
legend("topleft", c("group 1", "group 2"), pch = c(0, 2),
col = c("#123E7C", "#B65D35"), bty = "n")The two samples. Group 2 carries an extra cluster near (1.6, 1.6).
compare_pt() takes the two samples; every other argument
has a default.
set.seed(12)
fit <- compare_pt(x, y)
fit
#> Bayesian two-sample comparison (ptcompare)
#> group 1: 1000; group 2: 1000; variables: 2
#> Posterior P(distributions differ): 1.000000
#> Common coordinate transformation: rank
#> Particles: 256; elapsed: 0.20 s; divisions in the representative tree: 142
#> Use summary(), regions() and plot() for details.prob_difference is the posterior probability that the
two distributions differ. That probability says only that a difference
exists somewhere, but regions() can locate it. Each row is
one division of the representative tree, the tree with the highest
posterior score among those sampled, and the rows are ordered by
prob_difference — which here is the probability that the
groups branch differently at that one division, not the global quantity
printed above.
regions(fit, prob = 0.9, min_n = 40)
#> node_id parent_id depth condition variable
#> 1 15 7 3 marker_A > 0.866 & marker_B > 0.101 marker_B
#> 2 31 15 4 marker_A > 0.866 & marker_B > 0.84 marker_A
#> 3 1 0 0 all observations marker_A
#> 4 7 3 2 marker_A > 0.866 marker_B
#> 5 3 1 1 marker_A > 0.125 marker_A
#> threshold n1 n2 prob_difference effect effect_abs lower upper
#> 1 0.841095 102 209 1.0000 1.975 1.975 1.453 2.524
#> 2 1.436290 40 172 0.9999 1.347 1.347 0.644 2.082
#> 3 0.125388 1000 1000 0.9995 0.425 0.425 0.248 0.602
#> 4 0.101036 196 304 0.9953 0.703 0.704 0.320 1.080
#> 5 0.866351 447 553 0.9708 0.434 0.434 0.000 0.698
#> prob_positive
#> 1 1.0000
#> 2 0.9999
#> 3 0.9995
#> 4 0.9950
#> 5 0.9706Take the first row. Its condition restricts attention to
the observations with marker_A > 0.866 and
marker_B > 0.101, a part of the plane holding 102
observations of group 1 and 209 of group 2. Within it the division is at
marker_B = 0.841, and the two groups fall below that
threshold in different proportions: prob_difference is
1.00, and effect, a log odds ratio of group 1 against group
2, is 1.98, so the odds of falling below the threshold are about seven
times higher in group 1. The lower and upper
columns give a 95% credible interval for that log odds ratio. The region
is the upper right part of the plane, where the simulated cluster lies,
and the sign of the effect says that group 2 is the one over-represented
above the threshold.
Two further columns are reported for each division:
effect_abs, the posterior mean of the absolute log odds
ratio, which is not in general the absolute value of
effect, and prob_positive, the posterior
probability that the log odds ratio is positive.
The same regions can be drawn with the data. min_n is
passed again so that the panels match the rows of the table above.
The four most probable regions. The box is the region, the dashed line its threshold.
Each panel is one region. The box is the region, the dashed line is the threshold that divides it, and the observations inside are drawn as blue squares and brown triangles, with the rest of the data in gray for context. The first panel is the region discussed above; the brown triangles concentrated inside it are the simulated cluster.
The tree shows how the regions are nested and which variable each
division uses. Here max_depth limits how much of the tree
is drawn; it does not change the fit.
The representative tree, drawn to depth 3.
With two variables the regions at each depth can also be shaded by their probability of a difference, which shows how the comparison narrows as the partition becomes finer.
Regions by depth, shaded by the probability of a difference.
At depth 0 the only region is the whole plane, so the shading reports no more than that the samples differ somewhere. The shaded area contracts as the tree deepens, and by depth 3 it has settled on the upper right corner.
The wdbc data in the mclust package are
the Wisconsin Diagnostic Breast Cancer data of Street, Wolberg and
Mangasarian (1993). For each of 569 patients, a fine needle aspirate was
taken from a breast mass and photographed, and the boundary of each cell
nucleus in the image was traced. Ten geometric features were computed
for every nucleus: radius, texture, perimeter, area, smoothness,
compactness, concavity, concave points, symmetry and fractal dimension.
Each feature was then summarized over the nuclei of an image by its
mean, its standard error and an extreme value, which is how the 30
variable names below are composed. The mass was separately diagnosed as
benign or malignant.
data("wdbc", package = "mclust")
features <- setdiff(names(wdbc)[sapply(wdbc, is.numeric)], "ID")
features
#> [1] "Radius_mean" "Texture_mean" "Perimeter_mean"
#> [4] "Area_mean" "Smoothness_mean" "Compactness_mean"
#> [7] "Concavity_mean" "Nconcave_mean" "Symmetry_mean"
#> [10] "Fractaldim_mean" "Radius_se" "Texture_se"
#> [13] "Perimeter_se" "Area_se" "Smoothness_se"
#> [16] "Compactness_se" "Concavity_se" "Nconcave_se"
#> [19] "Symmetry_se" "Fractaldim_se" "Radius_extreme"
#> [22] "Texture_extreme" "Perimeter_extreme" "Area_extreme"
#> [25] "Smoothness_extreme" "Compactness_extreme" "Concavity_extreme"
#> [28] "Nconcave_extreme" "Symmetry_extreme" "Fractaldim_extreme"
table(wdbc$Diagnosis)
#>
#> B M
#> 357 212The question is how nuclear morphology differs between the two
diagnoses. Here the groups are a label on a single data set rather than
two matrices, so they are given through group. The levels
of that factor fix the order, so group 1 is B, the benign
masses, and group 2 is M.
set.seed(2)
fit_wdbc <- compare_pt(wdbc[, features], group = wdbc$Diagnosis)
summary(fit_wdbc, top = 3)
#> Bayesian two-sample comparison
#> N = 357 / 212; variables = 30
#> P(equal) = 2.08552e-106; P(different) = 1
#> Common coordinate transformation: rank
#> Final ESS = 181.5; resampling steps = 4; elapsed = 0.04 s
#> Regions that reached max_depth: 3 (86.1% of observations)
#> Divisions with the largest prob_difference (see ?regions):
#> node_id variable threshold n1 n2 prob_difference effect lower upper
#> 1 Perimeter_extreme 102.05000 357 212 1.000 5.021 4.333 5.797
#> 23 Nconcave_mean 0.05302 41 182 1.000 3.478 2.638 4.396
#> 27 Concavity_mean 0.07865 10 166 0.997 3.976 2.175 5.950
#> effect > 0: group 1 is more likely to satisfy variable <= threshold within the region.The two diagnoses differ, and summary() also reports the
computation: the effective number of particles at the end of the
sequential Monte Carlo, how often the particles were resampled, and the
running time.
Asking for the regions with a probability of at least 0.9 and at least 25 observations of each group returns two. The third division in the summary above is dropped by the second condition, not the first: it holds only 10 benign cases.
found <- regions(fit_wdbc, prob = 0.9, min_n = 25)
found[, c("depth", "variable", "n1", "n2", "prob_difference", "effect")]
#> depth variable n1 n2 prob_difference effect
#> 1 0 Perimeter_extreme 357 212 1 5.021
#> 2 6 Nconcave_mean 41 182 1 3.478Left: the division at the root. Right: a division within the large nuclei.
The first region is at depth 0, which is to say it is the whole data
set, and it reproduces the familiar result: the nuclei of malignant
masses are larger. The division is on Perimeter_extreme at
102, with an effect of about 5 on the log odds scale. This is a
difference between two marginal distributions, and needs no partition to
find.
The second region is at depth 6, and is the kind of statement that does.
local <- found[found$depth > 0, ][1, ]
local$condition
#> [1] "Radius_mean > 13.77 & Perimeter_mean > 88.66 & Area_mean > 584.4 & Radius_extreme > 15.5 & Perimeter_extreme > 102 & Area_extreme > 740"Six conditions, all of them lower bounds on measurements of size,
restrict attention to the images whose nuclei are large: 41 benign
masses and 182 malignant ones. Within that region the method reports a
further division, and it falls on a measure of shape rather than size,
Nconcave_mean, which summarizes the concave points on the
nucleus boundary.
region_data() returns the observations of a region, with
a column giving the group, so the division can be examined directly.
cells <- region_data(fit_wdbc, local$node_id)
tapply(cells$Nconcave_mean, cells$group, median)
#> B M
#> 0.03890 0.08813
tapply(cells$Nconcave_mean <= local$threshold, cells$group, mean)
#> B M
#> 0.75609756 0.08791209Among these large nuclei the median benign image sits at 0.039 against 0.088 for the malignant ones, and 76% of the benign images fall below the threshold against 9% of the malignant ones. The two diagnoses are therefore separated twice over by different features: once by size across the whole data set, and again by the irregularity of the nucleus boundary among the masses whose nuclei are large. The second statement is about a conditional distribution inside a subset of the data, and it is the kind of result the partition is there to produce.
The last comparison uses the same 30 variables, but splits the benign cases into two halves at random and compares those with each other. The two samples come from the same distribution by construction.
benign <- wdbc[wdbc$Diagnosis == "B", features]
set.seed(3)
half <- sample(c(TRUE, FALSE), nrow(benign), replace = TRUE)
set.seed(4)
fit_null <- compare_pt(benign[half, ], benign[!half, ])
fit_null
#> Bayesian two-sample comparison (ptcompare)
#> group 1: 173; group 2: 184; variables: 30
#> Posterior P(distributions differ): 0.072867
#> Common coordinate transformation: rank
#> Particles: 256; elapsed: 0.03 s; divisions in the representative tree: 17
#> Use summary(), regions() and plot() for details.
nrow(regions(fit_null, prob = 0.9))
#> [1] 0The posterior probability of a difference is 0.07, and no division reaches a probability of 0.9.
The regions and effect sizes are reported for the representative tree, the one with the highest posterior score among the trees that were sampled. Regions selected by a probability threshold can be nested inside one another, and they carry no frequentist multiple-testing guarantee.
?compare_pt for the arguments, including
transform, which sets the common scale the variables are
put on before the space is divided.?pt_prior and ?pt_control for the model
and the computational budget, and for the defaults used above.?regions for the meaning of each column, and for
rank_by = "effect", which orders by the size of a
difference rather than its probability.?region_rows and ?region_data for
extracting the observations of a region.?plot.ptcompare_fit for the plots used here. With three
or more variables a panel draws two of them, so the box is the region
seen in those two variables only, and the symbols rather than the box
indicate membership. This applies to both panels of the 30-variable
figure above.system.file("STATISTICAL-SPECIFICATION.md", package = "ptcompare")
for the model, the target distribution of the sequential Monte Carlo and
the definition of the effect sizes.Awaya, N. and Ma, L. (2024). Hidden Markov Pólya Trees for High-Dimensional Distributions. Journal of the American Statistical Association, 119(545), 189–201. https://doi.org/10.1080/01621459.2022.2105223
Street, W. N., Wolberg, W. H. and Mangasarian, O. L. (1993). Nuclear feature extraction for breast tumor diagnosis. IS&T/SPIE International Symposium on Electronic Imaging: Science and Technology, 1905, 861–870.