tests/testthat/test-modules-smd.R
In kaz-yos/tableone: Create 'Table 1' to Describe Baseline Characteristics with or without Propensity Score Weights
################################################################################
### Unit tests for the standardized mean difference modules
## Reference: http://adv-r.had.co.nz/Testing.html
## Created on:
## Author: Kazuki Yoshida
################################################################################
### Structure
## expectations within tests within context
###
### References
################################################################################
## A unified approach to measuring the effect size between two groups using SAS
## http://support.sas.com/resources/papers/proceedings12/335-2012.pdf
###
### Prepare environment
################################################################################
library(testthat)
library(survey)
library(Matrix)
###
### Actual tests
################################################################################
context("Tests for functions for standardized mean differences")
### Provide data
data(nhanes)
nhanesSvy <- svydesign(ids = ~ SDMVPSU, strata = ~ SDMVSTRA, weights = ~ WTMEC2YR,
nest = TRUE, data = nhanes)
## ‘SDMVPSU’ Primary sampling units
## ‘SDMVSTRA’ Sampling strata
## ‘WTMEC2YR’ Sampling weights
## ‘HI_CHOL’ Numeric vector: 1 for total cholesterol over 240mg/dl, 0
## under 240mg/dl
## ‘race’ 1=Hispanic, 2=non-Hispanic white, 3=non-Hispanic black,
## 4=other
## ‘agecat’ Age group ‘(0,19]’ ‘(19,39]’ ‘(39,59]’ ‘(59,Inf]’
## ‘RIAGENDR’ Gender: 1=male, 2=female
## Use HI_CHOL for continuous, RIAGENDR for binary, race and agecat as categorical
### Old functions explicitly for 3 groups (Keep; used for testing)
## Standardize differences
StdDiffOld <- function(variable, group) {
## For each group
means <- tapply(variable, group, mean)
vars <- tapply(variable, group, var)
## Calculate for all three possible pairs
g1g2 <- abs(means[1] - means[2]) / sqrt((vars[1] + vars[2]) / 2)
g2g3 <- abs(means[2] - means[3]) / sqrt((vars[2] + vars[3]) / 2)
g3g1 <- abs(means[3] - means[1]) / sqrt((vars[3] + vars[1]) / 2)
## 2vs1, 3vs1, then 3vs2 to be consistent with regression
out <- c(g1g2, g3g1, g2g3)
names(out) <- c("3vs2","3vs1","2vs1")
out
}
svyStdDiffOld <- function(varName, groupName, design) {
varFormula <- as.formula(paste("~", varName))
groupFormula <- as.formula(paste("~", groupName))
means <- svyby(formula = varFormula, by = groupFormula, FUN = svymean, design = design)[,2]
vars <- svyby(formula = varFormula, by = groupFormula, FUN = svyvar, design = design)[,2]
## Calculate for all three possible pairs
g1g2 <- abs(means[1] - means[2]) / sqrt((vars[1] + vars[2]) / 2)
g2g3 <- abs(means[2] - means[3]) / sqrt((vars[2] + vars[3]) / 2)
g3g1 <- abs(means[3] - means[1]) / sqrt((vars[3] + vars[1]) / 2)
## 2vs1, 3vs1, then 3vs2 to be consistent with regression
out <- c(g1g2, g3g1, g2g3)
names(out) <- c("3vs2","3vs1","2vs1")
out
}
test_that("continuous standardized difference is correct (nhanes unweighted)", {
## Two group with only one contrast 1 vs 2
means1 <- tapply(nhanes$HI_CHOL, nhanes$RIAGENDR, mean, na.rm = TRUE)
vars1 <- tapply(nhanes$HI_CHOL, nhanes$RIAGENDR, var, na.rm = TRUE)
meanDiffs1 <- (means1[1] - means1[2]) / sqrt((vars1[1] + vars1[2]) / 2)
names(meanDiffs1) <- NULL
expect_equal(StdDiff(nhanes$HI_CHOL, nhanes$RIAGENDR), abs(meanDiffs1))
## Four groups with 6 contrasts
means2 <- tapply(nhanes$HI_CHOL, nhanes$race, mean, na.rm = TRUE)
vars2 <- tapply(nhanes$HI_CHOL, nhanes$race, var, na.rm = TRUE)
meanDiffs2 <-
c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
names(meanDiffs2) <- NULL
## Individual numbers
expect_equal(StdDiff(nhanes$HI_CHOL, nhanes$race), abs(meanDiffs2))
## Average across
expect_equal(mean(StdDiff(nhanes$HI_CHOL, nhanes$race)),
mean(abs(meanDiffs2)))
})
test_that("continuous standardized difference is correct (nhanes weighted)", {
## Two group
means1 <- svyby( ~ HI_CHOL, by = ~ RIAGENDR, design = nhanesSvy, FUN = svymean, na.rm = TRUE)[,2]
vars1 <- svyby( ~ HI_CHOL, by = ~ RIAGENDR, design = nhanesSvy, FUN = svyvar, na.rm = TRUE)[,2]
meanDiffs1 <- (means1[1] - means1[2]) / sqrt((vars1[1] + vars1[2]) / 2)
expect_equal(svyStdDiff("HI_CHOL", "RIAGENDR", design = nhanesSvy),
abs(meanDiffs1))
## Four groups
means2 <- svyby( ~ HI_CHOL, by = ~ race, design = nhanesSvy, FUN = svymean, na.rm = TRUE)[,2]
vars2 <- svyby( ~ HI_CHOL, by = ~ race, design = nhanesSvy, FUN = svyvar, na.rm = TRUE)[,2]
meanDiffs2 <-
c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
names(meanDiffs2) <- NULL
## Individual numbers
expect_equal(svyStdDiff("HI_CHOL", "race", design = nhanesSvy),
abs(meanDiffs2))
## Average across
expect_equal(mean(svyStdDiff("HI_CHOL", "race", design = nhanesSvy)),
mean(abs(meanDiffs2)))
})
test_that("binary standardized difference is correct (nhanes unweighted)", {
## Two group with only one contrast 1 vs 2
means1 <- tapply(nhanes$HI_CHOL, nhanes$RIAGENDR, mean, na.rm = TRUE)
vars1 <- means1 * (1 - means1)
meanDiffs1 <- (means1[1] - means1[2]) / sqrt((vars1[1] + vars1[2]) / 2)
names(meanDiffs1) <- NULL
expect_equal(StdDiff(nhanes$HI_CHOL, nhanes$RIAGENDR, binary = TRUE),
abs(meanDiffs1))
## Four groups with 6 contrasts
means2 <- tapply(nhanes$HI_CHOL, nhanes$race, mean, na.rm = TRUE)
vars2 <- means2 * (1 - means2)
meanDiffs2 <-
c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
names(meanDiffs2) <- NULL
## Individual numbers
expect_equal(StdDiff(nhanes$HI_CHOL, nhanes$race, binary = TRUE),
abs(meanDiffs2))
## Average across
expect_equal(mean(StdDiff(nhanes$HI_CHOL, nhanes$race, binary = TRUE)),
mean(abs(meanDiffs2)))
})
test_that("binary standardized difference is correct (nhanes weighted)", {
## Two group
means1 <- svyby( ~ HI_CHOL, by = ~ RIAGENDR, design = nhanesSvy, FUN = svymean, na.rm = TRUE)[,2]
vars1 <- means1 * (1 - means1)
meanDiffs1 <- (means1[1] - means1[2]) / sqrt((vars1[1] + vars1[2]) / 2)
expect_equal(svyStdDiff("HI_CHOL", "RIAGENDR", design = nhanesSvy, binary = TRUE),
abs(meanDiffs1))
## Four groups
means2 <- svyby( ~ HI_CHOL, by = ~ race, design = nhanesSvy, FUN = svymean, na.rm = TRUE)[,2]
vars2 <- means2 * (1 - means2)
meanDiffs2 <-
c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
names(meanDiffs2) <- NULL
## Individual numbers
expect_equal(svyStdDiff("HI_CHOL", "race", design = nhanesSvy, binary = TRUE),
abs(meanDiffs2))
## Average across
expect_equal(mean(svyStdDiff("HI_CHOL", "race", design = nhanesSvy, binary = TRUE)),
mean(abs(meanDiffs2)))
})
test_that("multinomal helpers are correct", {
means <- c(0.310336708264657, 0.224393689663292, 0.234283023310572, 0.230986578761479)
covs <- MultinomialVar(means)
## Expectations
## Diagonal p(1-p)
expect_equal(diag(covs), means * (1 - means))
## Off diagonal -1 * p_i * p_j
expect_equal(covs[1,2], -1 * means[1] * means[2])
expect_equal(covs[1,3], -1 * means[1] * means[3])
expect_equal(covs[1,4], -1 * means[1] * means[4])
expect_equal(covs[2,3], -1 * means[2] * means[3])
expect_equal(covs[2,4], -1 * means[2] * means[4])
expect_equal(covs[3,4], -1 * means[3] * means[4])
})
test_that("categorical standardized difference is correct (nhanes unweighted)", {
## Four groups with 6 contrasts
means2 <- tapply(nhanes$HI_CHOL, nhanes$race, mean, na.rm = TRUE)
vars2 <- means2 * (1 - means2)
meanDiffs2 <-
c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
names(meanDiffs2) <- NULL
## Individual numbers
expect_equal(StdDiffMulti(nhanes$HI_CHOL, nhanes$race),
abs(meanDiffs2))
## Average across
expect_equal(mean(StdDiffMulti(nhanes$HI_CHOL, nhanes$race)),
mean(abs(meanDiffs2)))
## Two group with only one contrast 1 vs 2
strataByLevels1 <- xtabs( ~ RIAGENDR + agecat, data = nhanes)
## drop first column after calculating proportion
propTable1 <- prop.table(strataByLevels1, margin = 1)[, -1, drop = FALSE]
meanDiffs1 <- propTable1[1,] - propTable1[2,]
covMean1 <- (MultinomialVar(propTable1[1,]) + MultinomialVar(propTable1[2,])) / 2
## R is not strict about transposition
expect_equal(meanDiffs1 %*% MASS::ginv(covMean1) %*% meanDiffs1,
t(meanDiffs1) %*% MASS::ginv(covMean1) %*% t(t(meanDiffs1)))
## Test actual functions
expect_equal(StdDiffMulti(nhanes$agecat, nhanes$RIAGENDR),
sqrt(drop(t(meanDiffs1) %*% MASS::ginv(covMean1) %*% t(t(meanDiffs1)))))
## Four groups with 6 contrasts
strataByLevels2 <- xtabs( ~ race + agecat, data = nhanes)
propTable2 <- prop.table(strataByLevels2, margin = 1)[, -1, drop = FALSE]
meanDiffs2 <- list(propTable2[1,] - propTable2[2,],
propTable2[1,] - propTable2[3,],
propTable2[1,] - propTable2[4,],
propTable2[2,] - propTable2[3,],
propTable2[2,] - propTable2[4,],
propTable2[3,] - propTable2[4,])
covMean2 <-
list((MultinomialVar(propTable2[1,]) + MultinomialVar(propTable2[2,])) / 2,
(MultinomialVar(propTable2[1,]) + MultinomialVar(propTable2[3,])) / 2,
(MultinomialVar(propTable2[1,]) + MultinomialVar(propTable2[4,])) / 2,
(MultinomialVar(propTable2[2,]) + MultinomialVar(propTable2[3,])) / 2,
(MultinomialVar(propTable2[2,]) + MultinomialVar(propTable2[4,])) / 2,
(MultinomialVar(propTable2[3,]) + MultinomialVar(propTable2[4,])) / 2)
## These should match in length.
expect_equal(length(meanDiffs2), length(covMean2))
smds <- unlist(lapply(seq_along(meanDiffs2), function(i) {
sqrt(drop(t(meanDiffs2[[i]]) %*% MASS::ginv(covMean2[[i]]) %*% t(t(meanDiffs2[[i]]))))
}))
## Individual numbers
expect_equal(StdDiffMulti(nhanes$agecat, nhanes$race),
smds)
## Average across
expect_equal(mean(StdDiffMulti(nhanes$agecat, nhanes$race)),
mean(smds))
})
test_that("categorical standardized difference is correct (nhanes weighted)", {
## Binary four groups
means2 <- svyby( ~ HI_CHOL, by = ~ race, design = nhanesSvy, FUN = svymean, na.rm = TRUE)[,2]
vars2 <- means2 * (1 - means2)
meanDiffs2 <-
c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
names(meanDiffs2) <- NULL
## Individual numbers
expect_equal(svyStdDiffMulti("HI_CHOL", "race", design = nhanesSvy),
abs(meanDiffs2))
## Average across
expect_equal(mean(svyStdDiffMulti("HI_CHOL", "race", design = nhanesSvy)),
mean(abs(meanDiffs2)))
## Two group with only one contrast 1 vs 2
strataByLevels1 <- svytable( ~ RIAGENDR + agecat, design = nhanesSvy)
propTable1 <- prop.table(strataByLevels1, margin = 1)[, -1, drop = FALSE]
meanDiffs1 <- propTable1[1,] - propTable1[2,]
covMean1 <- (MultinomialVar(propTable1[1,]) + MultinomialVar(propTable1[2,])) / 2
## Test actual functions
expect_equal(svyStdDiffMulti("agecat", "RIAGENDR", design = nhanesSvy),
sqrt(drop(t(meanDiffs1) %*% MASS::ginv(covMean1) %*% t(t(meanDiffs1)))))
## Four groups with 6 contrasts
strataByLevels2 <- svytable( ~ race + agecat, design = nhanesSvy)
propTable2 <- prop.table(strataByLevels2, margin = 1)[, -1, drop = FALSE]
meanDiffs2 <- list(propTable2[1,] - propTable2[2,],
propTable2[1,] - propTable2[3,],
propTable2[1,] - propTable2[4,],
propTable2[2,] - propTable2[3,],
propTable2[2,] - propTable2[4,],
propTable2[3,] - propTable2[4,])
covMean2 <-
list((MultinomialVar(propTable2[1,]) + MultinomialVar(propTable2[2,])) / 2,
(MultinomialVar(propTable2[1,]) + MultinomialVar(propTable2[3,])) / 2,
(MultinomialVar(propTable2[1,]) + MultinomialVar(propTable2[4,])) / 2,
(MultinomialVar(propTable2[2,]) + MultinomialVar(propTable2[3,])) / 2,
(MultinomialVar(propTable2[2,]) + MultinomialVar(propTable2[4,])) / 2,
(MultinomialVar(propTable2[3,]) + MultinomialVar(propTable2[4,])) / 2)
## These should match in length.
expect_equal(length(meanDiffs2), length(covMean2))
smds <- unlist(lapply(seq_along(meanDiffs2), function(i) {
sqrt(drop(t(meanDiffs2[[i]]) %*% MASS::ginv(covMean2[[i]]) %*% t(t(meanDiffs2[[i]]))))
}))
## Individual numbers
expect_equal(svyStdDiffMulti("agecat", "race", design = nhanesSvy),
smds)
## Average across
expect_equal(mean(svyStdDiffMulti("agecat", "race", design = nhanesSvy)),
mean(smds))
})
### Test on anomalous data
################################################################################
test_that("decent results are returned for anomalous/difficult data", {
data(nhanes)
nhanes$onlyOne <- 1
nhanes$onlyNa <- as.numeric(NA)
nhanes$logi <- nhanes$RIAGENDR == 1
nhanes$race2 <- nhanes$race
nhanesSvy <- svydesign(ids = ~ SDMVPSU, strata = ~ SDMVSTRA,
weights = ~ WTMEC2YR, nest = TRUE,
data = nhanes)
## Note: 2015-08-01
## Categorical SMD: When both of two group have at one or more zero
## categories, both contribute singular matrices to the pooled
## variance covariance matrix S in Yang and Dalton 2012.
##
## When both groups have a single category that is probability 1 and
## other cells that are all zeros, both contribute zero matrices to
## the pooled variance covariance matrix S. solve() dies on this
## singular matrix. MASS::ginv() returns generalized inverse matrix,
## which is the zero matrix of the same dimension, instead.
##
## This results in SMD of zero regardless of the mean vector difference
## (T-C in Yang and Dalton 2012).
##
## This behavior looks odd, thus, categorical SMD functions were
## modified to return NaN when S is a zero matrix.
## Logical
## 0 due to [0]^- = 0
table(nhanes$logi, nhanes$RIAGENDR)
expect_equal(StdDiffMulti(nhanes$logi, group = nhanes$RIAGENDR), NaN)
## Matches with result from original variable
expect_equal(StdDiffMulti(nhanes$logi, group = nhanes$race),
StdDiffMulti(nhanes$RIAGENDR, group = nhanes$race))
## Multiple group where T-C is not a zero vector.
## All groups have single group that is probability 1, thus,
## all contrasts have zero matrix S.
table(nhanes$race2, nhanes$race)
expect_equal(StdDiffMulti(nhanes$race2, group = nhanes$race), rep(NaN, 6))
expect_equal(svyStdDiffMulti("race2", group = "race", design = nhanesSvy), rep(NaN, 6))
## Identical constant comparison
## T-C vector is a zero vector, but divisor is also zero
## NaN due to division by zero variance -> Redefined as zero
by(nhanes$onlyOne, nhanes$RIAGENDR, summary)
expect_equal(StdDiff(nhanes$onlyOne, group = nhanes$RIAGENDR), 0)
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined NaN in (svy)StdDiffMulti
table(nhanes$onlyOne, nhanes$RIAGENDR)
expect_equal(StdDiffMulti(nhanes$onlyOne, group = nhanes$RIAGENDR), 0)
## When weighted problematic
means1 <- svyby(~ onlyOne, by = ~ RIAGENDR, nhanesSvy, FUN = svymean)[,2]
vars1 <- svyby(~ onlyOne, by = ~ RIAGENDR, nhanesSvy, FUN = svyvar)[,2]
## Very small difference is inflated by even smaller variance
## on sparc-sun-solaris sign was opposite; abs() solves this issue
## as svyStdDiff() uses abs() internally
expect_equal(svyStdDiff("onlyOne", "RIAGENDR", nhanesSvy),
abs((means1[1] - means1[2]) / sqrt(sum(vars1) / 2)))
## NaN should be the case, but it's not, but it's consistent with survey
## expect_equal(svyStdDiff("onlyOne", "RIAGENDR", nhanesSvy), NaN)
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined NaN in (svy)StdDiffMulti
## No error even with a single level variable (constant) as redundant
## level drop from table occurs only when 2+ levels are present.
## If any group has more than 2 levels, then strata-by-level table
## is correctly created, which is not the case here. Redefined as 0.
expect_equal(svyStdDiffMulti("onlyOne", "RIAGENDR", nhanesSvy), 0)
## Four groups (six contrasts)
## NaN due to division by zero variance -> Redefined as 0
by(nhanes$onlyOne, nhanes$race, summary)
expect_equal(StdDiff(nhanes$onlyOne, group = nhanes$race), rep(0, 6))
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined NaN in (svy)StdDiffMulti
expect_equal(StdDiffMulti(nhanes$onlyOne, group = nhanes$race), rep(0, 6))
## When weighted problematic; not in this case??
means2 <- svyby(~ onlyOne, by = ~ race, nhanesSvy, FUN = svymean)[,2]
vars2 <- svyby(~ onlyOne, by = ~ race, nhanesSvy, FUN = svyvar)[,2]
meanDiffs2 <- c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
## Redefined as 0 as T-C is a zero vector
## Fails if no long double is available in R.
##
## noLD: Tests without long double (tests on x86_64 Linux with R-devel configured --without-long-double)
## ── 1. Failure: decent results are returned for anomalous/difficult data (@test-m
## svyStdDiff("onlyOne", "race", nhanesSvy) not equal to rep(0, 6).
## 6/6 mismatches (average diff: 1.33)
## [1] 0.117 - 0 == 0.117
## [2] 1.987 - 0 == 1.987
## [3] 1.129 - 0 == 1.129
## [4] 1.997 - 0 == 1.997
## [5] 1.162 - 0 == 1.162
## [6] 1.608 - 0 == 1.608
##
## base::.Machine$sizeof.longdouble: the number of bytes in a C long double type. Will be zero if there is no such type (or its use was disabled when R was built), otherwise possibly 12 (most 32-bit builds) or 16 (most 64-bit builds).
if (base::.Machine$sizeof.longdouble >= 16) {
## Test only when long double is available in 64bit system.
expect_equal(svyStdDiff("onlyOne", "race", nhanesSvy), rep(0,6))
}
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined NaN in (svy)StdDiffMulti
expect_equal(svyStdDiffMulti("onlyOne", "race", nhanesSvy), rep(0, 6))
## onlyNa
## NA as na.rm is turned off
expect_warning(expect_equal(StdDiff(nhanes$onlyNa, group = nhanes$RIAGENDR),
as.numeric(NA)),
"Variable has only NA's in at least one stratum. na.rm turned off.")
## defined 0 as NA is a level
expect_warning(expect_equal(StdDiffMulti(nhanes$onlyNa, group = nhanes$RIAGENDR), 0),
"Variable has only NA's in all strata. Regarding NA as a level")
## When weighted problematic
means1 <- svyby(~ onlyNa, by = ~ RIAGENDR, nhanesSvy, FUN = svymean)[,2]
vars1 <- svyby(~ onlyNa, by = ~ RIAGENDR, nhanesSvy, FUN = svyvar)[,2]
## Very small difference is inflated by even smaller variance
expect_warning(svyStdDiff("onlyNa", "RIAGENDR", nhanesSvy),
"onlyNa has only NA's in at least one stratum. na.rm turned off.")
expect_warning(expect_equal(svyStdDiff("onlyNa", "RIAGENDR", nhanesSvy),
as.numeric(NA)),
"onlyNa has only NA's in at least one stratum. na.rm turned off.")
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined NaN in (svy)StdDiffMulti
expect_warning(expect_equal(svyStdDiffMulti("onlyNa", "RIAGENDR", nhanesSvy), 0),
"onlyNa has only NA's in all strata. Regarding NA as a level.")
## Four groups (six contrasts)
## NaN due to division by zero variance
expect_warning(expect_equal(StdDiff(nhanes$onlyNa, group = nhanes$race), rep(NaN, 6)),
"Variable has only NA's in at least one stratum. na.rm turned off.")
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined NaN in (svy)StdDiffMulti
expect_warning(expect_equal(StdDiffMulti(nhanes$onlyNa, group = nhanes$race), rep(0, 6)),
"Variable has only NA's in all strata. Regarding NA as a level.")
## When weighted problematic; not in this case??
means2 <- svyby(~ onlyNa, by = ~ race, nhanesSvy, FUN = svymean)[,2]
vars2 <- svyby(~ onlyNa, by = ~ race, nhanesSvy, FUN = svyvar)[,2]
meanDiffs2 <- c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
expect_warning(expect_equal(svyStdDiff("onlyNa", "race", nhanesSvy), meanDiffs2),
"onlyNa has only NA's in at least one stratum. na.rm turned off.")
expect_warning(expect_equal(svyStdDiff("onlyNa", "race", nhanesSvy), rep(NaN, 6)),
"onlyNa has only NA's in at least one stratum. na.rm turned off.")
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined 0 as NA is a level
expect_warning(expect_equal(svyStdDiffMulti("onlyNa", "race", nhanesSvy), rep(0, 6)),
"onlyNa has only NA's in all strata. Regarding NA as a level.")
})
### Define a function for pooled S matrix (used in two tests)
pooledSMat <- function(strataByLevels) {
lstMeans <- LstMeansFromFullTable(strataByLevels)
S1 <- MultinomialVar(lstMeans[[1]])
S2 <- MultinomialVar(lstMeans[[2]])
S <- (S1 + S2) / 2
## first group element, second group element, pooled
list(S1 = S1, S2 = S2, S = S)
}
test_that("zero-cell vectors are decently handled by categorical SMD (artificial data)", {
set.seed(2015080200)
## Simple artificial data
wt <- runif(n = 100)
dat <- data.frame(group = rep(c(1,2), each = 100),
var1 = factor(c(rep(1:4, c(0,10,20,70)), rep(1:4, c( 0,10,20,70))), level = 1:4),
var2 = factor(c(rep(1:4, c(0,10,20,70)), rep(1:4, c( 0,20,30,50))), level = 1:4),
var3 = factor(c(rep(1:4, c(0,10,20,70)), rep(1:4, c(20, 0,30,50))), level = 1:4),
var4 = factor(c(rep(1:4, c(0,10,20,70)), rep(1:4, c(10,20,30,40))), level = 1:4),
wt = rep(wt, 2))
datSvy <- svydesign(ids = ~ 1, data = dat, weights = ~wt)
## Expectations
## First case
var1S <- pooledSMat(table(dat$group, dat$var1))
expect_true(rankMatrix(var1S[[1]]) < 3)
expect_true(rankMatrix(var1S[[2]]) < 3)
expect_true(rankMatrix(var1S[[3]]) < 3)
## Using ginv zero due to numerator (OK)
expect_equal(StdDiffMulti(dat$var1, dat$group), 0)
## Second case
var2S <- pooledSMat(table(dat$group, dat$var2))
expect_true(rankMatrix(var2S[[1]]) < 3)
expect_true(rankMatrix(var2S[[2]]) < 3)
expect_true(rankMatrix(var2S[[3]]) < 3)
## Using ginv() non-zero
expect_true(StdDiffMulti(dat$var2, dat$group) > 0)
## Third case
var3S <- pooledSMat(table(dat$group, dat$var3))
expect_true(rankMatrix(var3S[[1]]) < 3)
expect_true(rankMatrix(var3S[[2]]) < 3)
## pooled can be full rank if zero cell positions are different
expect_true(rankMatrix(var3S[[3]]) == 3)
## Using true inverse from ginv, non-zero
expect_true(StdDiffMulti(dat$var3, dat$group) > 0)
## Third case
var4S <- pooledSMat(table(dat$group, dat$var4))
expect_true(rankMatrix(var4S[[1]]) < 3)
## All cells populated and full rank
expect_true(rankMatrix(var4S[[2]]) == 3)
## pooled can be full rank if one group is full rank
expect_true(rankMatrix(var4S[[3]]) == 3)
expect_true(StdDiffMulti(dat$var4, dat$group) > 0)
## Weighted
## First case
var1S <- pooledSMat(svytable( ~ group + var1, datSvy))
expect_true(rankMatrix(var1S[[1]]) < 3)
expect_true(rankMatrix(var1S[[2]]) < 3)
expect_true(rankMatrix(var1S[[3]]) < 3)
## Using ginv zero due to numerator (OK)
expect_equal(svyStdDiffMulti("var1", "group", datSvy), 0)
## Second case
var2S <- pooledSMat(svytable( ~ group + var2, datSvy))
expect_true(rankMatrix(var2S[[1]]) < 3)
expect_true(rankMatrix(var2S[[2]]) < 3)
expect_true(rankMatrix(var2S[[3]]) < 3)
## Using ginv() non-zero
expect_true(svyStdDiffMulti("var2", "group", datSvy) > 0)
## Third case
var3S <- pooledSMat(svytable( ~ group + var3, datSvy))
expect_true(rankMatrix(var3S[[1]]) < 3)
expect_true(rankMatrix(var3S[[2]]) < 3)
## pooled can be full rank if zero cell positions are different
expect_true(rankMatrix(var3S[[3]]) == 3)
## Using true inverse from ginv, non-zero
expect_true(svyStdDiffMulti("var3", "group", datSvy) > 0)
## Third case
var4S <- pooledSMat(svytable( ~ group + var4, datSvy))
expect_true(rankMatrix(var4S[[1]]) < 3)
## All cells populated and full rank
expect_true(rankMatrix(var4S[[2]]) == 3)
## pooled can be full rank if one group is full rank
expect_true(rankMatrix(var4S[[3]]) == 3)
expect_true(svyStdDiffMulti("var4", "group", datSvy) > 0)
})
test_that("zero-cell vectors are decently handled by categorical SMD (real data)", {
data(nhanes)
set.seed(2015080200)
## grouping is by a binary nhanes$RIAGENDR
## (0,0.1,0.2,0.7)^T for both groups: same distribution
nhanes$var1 <- NA
nhanes$var1[nhanes$RIAGENDR == 1] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 1)), size = 1, prob = c(0,0.1,0.2,0.7)))
nhanes$var1[nhanes$RIAGENDR == 2] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 2)), size = 1, prob = c(0,0.1,0.2,0.7)))
nhanes$var1 <- factor(nhanes$var1, levels = 1:4)
## (0,0.1,0.2,0.7)^T vs (0,0.2,0.3,0.5)^T: different distribution, but same zero cell
nhanes$var2 <- NA
nhanes$var2[nhanes$RIAGENDR == 1] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 1)), size = 1, prob = c(0,0.1,0.2,0.7)))
nhanes$var2[nhanes$RIAGENDR == 2] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 2)), size = 1, prob = c(0,0.2,0.3,0.5)))
nhanes$var2 <- factor(nhanes$var2, levels = 1:4)
## (0,0.1,0.2,0.7)^T vs (0.2,0,0.3,0.5)^T: different distribution and zero cell
nhanes$var3 <- NA
nhanes$var3[nhanes$RIAGENDR == 1] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 1)), size = 1, prob = c(0,0.1,0.2,0.7)))
nhanes$var3[nhanes$RIAGENDR == 2] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 2)), size = 1, prob = c(0.2,0,0.3,0.5)))
nhanes$var3 <- factor(nhanes$var3, levels = 1:4)
## (0,0.1,0.2,0.7)^T vs (0.1,0.2,0.3,0.5)^T: different distribution and one is ok
nhanes$var4 <- NA
nhanes$var4[nhanes$RIAGENDR == 1] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 1)), size = 1, prob = c(0,0.1,0.2,0.7)))
nhanes$var4[nhanes$RIAGENDR == 2] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 2)), size = 1, prob = c(0.1,0.2,0.3,0.4)))
nhanes$var4 <- factor(nhanes$var4, levels = 1:4)
## svydesign
nhanesSvy <- svydesign(ids = ~ SDMVPSU, strata = ~ SDMVSTRA,
weights = ~ WTMEC2YR, nest = TRUE,
data = nhanes)
## Expectations
## First case
var1S <- pooledSMat(table(nhanes$RIAGENDR, nhanes$var1))
expect_true(rankMatrix(var1S[[1]]) < 3)
expect_true(rankMatrix(var1S[[2]]) < 3)
expect_true(rankMatrix(var1S[[3]]) < 3)
## Using ginv, near zero due to near-zero numerator
expect_true(StdDiffMulti(nhanes$var1, nhanes$RIAGENDR) < 0.1)
## Second case
var2S <- pooledSMat(table(nhanes$RIAGENDR, nhanes$var2))
expect_true(rankMatrix(var2S[[1]]) < 3)
expect_true(rankMatrix(var2S[[2]]) < 3)
expect_true(rankMatrix(var2S[[3]]) < 3)
## Using ginv() non-zero
expect_true(StdDiffMulti(nhanes$var2, nhanes$RIAGENDR) > 0)
## Third case
var3S <- pooledSMat(table(nhanes$RIAGENDR, nhanes$var3))
expect_true(rankMatrix(var3S[[1]]) < 3)
expect_true(rankMatrix(var3S[[2]]) < 3)
## pooled can be full rank if zero cell positions are different
expect_true(rankMatrix(var3S[[3]]) == 3)
## Using true inverse from ginv, non-zero
expect_true(StdDiffMulti(nhanes$var3, nhanes$RIAGENDR) > 0)
## Fourth case
var4S <- pooledSMat(table(nhanes$RIAGENDR, nhanes$var4))
expect_true(rankMatrix(var4S[[1]]) < 3)
## All cells populated and full rank
expect_true(rankMatrix(var4S[[2]]) == 3)
## pooled can be full rank if one group is full rank
expect_true(rankMatrix(var4S[[3]]) == 3)
expect_true(StdDiffMulti(nhanes$var4, nhanes$RIAGENDR) > 0)
## Weighted
## First case
var1S <- pooledSMat(svytable( ~ RIAGENDR + var1, nhanesSvy))
expect_true(rankMatrix(var1S[[1]]) < 3)
expect_true(rankMatrix(var1S[[2]]) < 3)
expect_true(rankMatrix(var1S[[3]]) < 3)
## Using ginv, near zero due to small numerator vector
expect_true(svyStdDiffMulti("var1", "RIAGENDR", nhanesSvy) < 0.1)
## Second case
var2S <- pooledSMat(svytable( ~ RIAGENDR + var2, nhanesSvy))
expect_true(rankMatrix(var2S[[1]]) < 3)
expect_true(rankMatrix(var2S[[2]]) < 3)
expect_true(rankMatrix(var2S[[3]]) < 3)
## Using ginv() non-zero
expect_true(svyStdDiffMulti("var2", "RIAGENDR", nhanesSvy) > 0)
## Third case
var3S <- pooledSMat(svytable( ~ RIAGENDR + var3, nhanesSvy))
expect_true(rankMatrix(var3S[[1]]) < 3)
expect_true(rankMatrix(var3S[[2]]) < 3)
## pooled can be full rank if zero cell positions are different
expect_true(rankMatrix(var3S[[3]]) == 3)
## Using true inverse from ginv, non-zero
expect_true(svyStdDiffMulti("var3", "RIAGENDR", nhanesSvy) > 0)
## Third case
var4S <- pooledSMat(svytable( ~ RIAGENDR + var4, nhanesSvy))
expect_true(rankMatrix(var4S[[1]]) < 3)
## All cells populated and full rank
expect_true(rankMatrix(var4S[[2]]) == 3)
## pooled can be full rank if one RIAGENDR is full rank
expect_true(rankMatrix(var4S[[3]]) == 3)
expect_true(svyStdDiffMulti("var4", "RIAGENDR", nhanesSvy) > 0)
})
### Tests working on multiple variables
################################################################################
test_that("multiple variables can be looped", {
contVars <- c("WTMEC2YR", "HI_CHOL", "race", "RIAGENDR")
catVars <- c("HI_CHOL", "race", "agecat", "RIAGENDR")
## Two groups
expect_equal(as.numeric(sapply(contVars, function(var) {
StdDiff(variable = nhanes[,var], group = nhanes[,"RIAGENDR"])
})),
c(StdDiff(variable = nhanes[,contVars[1]], group = nhanes[,"RIAGENDR"]),
StdDiff(variable = nhanes[,contVars[2]], group = nhanes[,"RIAGENDR"]),
StdDiff(variable = nhanes[,contVars[3]], group = nhanes[,"RIAGENDR"]),
StdDiff(variable = nhanes[,contVars[4]], group = nhanes[,"RIAGENDR"])))
expect_equal(as.numeric(sapply(catVars, function(var) {
StdDiffMulti(variable = nhanes[,var], group = nhanes[,"RIAGENDR"])
})),
c(StdDiffMulti(variable = nhanes[,catVars[1]], group = nhanes[,"RIAGENDR"]),
StdDiffMulti(variable = nhanes[,catVars[2]], group = nhanes[,"RIAGENDR"]),
StdDiffMulti(variable = nhanes[,catVars[3]], group = nhanes[,"RIAGENDR"]),
StdDiffMulti(variable = nhanes[,catVars[4]], group = nhanes[,"RIAGENDR"])))
## Four groups
expect_equal(lapply(contVars, function(var) {
StdDiff(variable = nhanes[,var], group = nhanes[,"race"])
}),
list(StdDiff(variable = nhanes[,contVars[1]], group = nhanes[,"race"]),
StdDiff(variable = nhanes[,contVars[2]], group = nhanes[,"race"]),
StdDiff(variable = nhanes[,contVars[3]], group = nhanes[,"race"]),
StdDiff(variable = nhanes[,contVars[4]], group = nhanes[,"race"])))
expect_equal(lapply(catVars, function(var) {
StdDiffMulti(variable = nhanes[,var], group = nhanes[,"race"])
}),
list(StdDiffMulti(variable = nhanes[,catVars[1]], group = nhanes[,"race"]),
StdDiffMulti(variable = nhanes[,catVars[2]], group = nhanes[,"race"]),
StdDiffMulti(variable = nhanes[,catVars[3]], group = nhanes[,"race"]),
StdDiffMulti(variable = nhanes[,catVars[4]], group = nhanes[,"race"])))
## Two groups
expect_equal(as.numeric(sapply(contVars, function(var) {
svyStdDiff(var, "RIAGENDR", nhanesSvy)
})),
c(svyStdDiff(contVars[1], "RIAGENDR", nhanesSvy),
svyStdDiff(contVars[2], "RIAGENDR", nhanesSvy),
svyStdDiff(contVars[3], "RIAGENDR", nhanesSvy),
svyStdDiff(contVars[4], "RIAGENDR", nhanesSvy)))
expect_equal(as.numeric(sapply(catVars, function(var) {
svyStdDiffMulti(var, "RIAGENDR", nhanesSvy)
})),
c(svyStdDiffMulti(catVars[1], "RIAGENDR", nhanesSvy),
svyStdDiffMulti(catVars[2], "RIAGENDR", nhanesSvy),
svyStdDiffMulti(catVars[3], "RIAGENDR", nhanesSvy),
svyStdDiffMulti(catVars[4], "RIAGENDR", nhanesSvy)))
## Four groups
expect_equal(lapply(contVars, function(var) {
svyStdDiff(var, "race", nhanesSvy)
}),
list(svyStdDiff(contVars[1], "race", nhanesSvy),
svyStdDiff(contVars[2], "race", nhanesSvy),
svyStdDiff(contVars[3], "race", nhanesSvy),
svyStdDiff(contVars[4], "race", nhanesSvy)))
expect_equal(lapply(catVars, function(var) {
svyStdDiffMulti(var, "race", nhanesSvy)
}),
list(svyStdDiffMulti(catVars[1], "race", nhanesSvy),
svyStdDiffMulti(catVars[2], "race", nhanesSvy),
svyStdDiffMulti(catVars[3], "race", nhanesSvy),
svyStdDiffMulti(catVars[4], "race", nhanesSvy)))
})
### Older tests with simpler data
################################################################################
test_that("binary standardized difference is correct", {
## Prepare data
variable <- c(1,0,0,0,0, 1,1,0,0,0, 1,1,1,0,0)
group <- rep(1:3, each = 5)
data1 <- data.frame(variable = variable, group = group,
weight = rep(0.5, 15))
data1Svy <- svydesign(ids = ~ 1, data = data1, weights = ~ weight)
## Gold standard
props <- c(0.2, 0.4, 0.6)
vars <- c(0.2*0.8, 0.4*0.6, 0.6*0.4)
g1g2 <- abs(props[1] - props[2]) / sqrt((vars[1] + vars[2]) / 2)
g2g3 <- abs(props[2] - props[3]) / sqrt((vars[2] + vars[3]) / 2)
g3g1 <- abs(props[3] - props[1]) / sqrt((vars[3] + vars[1]) / 2)
## Expectations
## Binary case
expect_equal(StdDiff(variable, group, binary = TRUE),
c(g1g2, g3g1, g2g3))
## Continuous case
expect_equal(StdDiff(variable, group, binary = FALSE),
as.vector(StdDiffOld(variable, group)))
## Weighted
## Binary case
expect_equal(svyStdDiff("variable","group", design = data1Svy, binary = TRUE),
c(g1g2, g3g1, g2g3))
## Continuous case
expect_equal(svyStdDiff("variable","group", design = data1Svy, binary = FALSE),
as.vector(svyStdDiffOld("variable","group", design = data1Svy)))
## Multiple variables
expect_equal(StdDiffs(data = data1, vars = c("variable","variable"),
groupVar = "group", binaryVars = "variable"),
list(c(g1g2, g3g1, g2g3), c(g1g2, g3g1, g2g3)))
expect_equal(StdDiffs(data = data1, vars = c("variable","variable"),
groupVar = "group", binaryVars = NA),
list(as.vector(StdDiffOld(variable, group)),
as.vector(StdDiffOld(variable, group))))
expect_equal(svyStdDiffs(data = data1Svy, vars = c("variable","variable"),
groupVar = "group", binaryVars = "variable"),
list(c(g1g2, g3g1, g2g3), c(g1g2, g3g1, g2g3)))
expect_equal(svyStdDiffs(data = data1Svy, vars = c("variable","variable"),
groupVar = "group", binaryVars = NA),
list(as.vector(svyStdDiffOld("variable","group", design = data1Svy)),
as.vector(svyStdDiffOld("variable","group", design = data1Svy))))
})
test_that("Multinomial SMD is correct", {
create_dummy <- function(x) {
df <- data.frame(x = factor(x))
model.matrix( ~ -1 + x, data = df)
}
set.seed(102)
## 4-category variable
probs <- c(0.1,0.3,0.5,0.2)
X <- rmultinom(n = 100, size = 1, prob = probs)
X <- as.vector(seq_along(probs) %*% X)
## Three groups
group <- c(rep(1,20), rep(2,30), rep(3,50))
## Drop first column to avoid dependent column
## The result does not change if generalized inverse is used
dummyX <- create_dummy(X)[, -1, drop = FALSE]
lstDummyX <- split(x = as.data.frame(dummyX), f = group, drop = FALSE)
## Means for each indicator
means <- lapply(lstDummyX, MultinomialMeans)
## cov mat for each indicators
covs <- lapply(means, MultinomialVar)
meanDiff1 <- means[[1]] - means[[2]]
meanDiff2 <- means[[1]] - means[[3]]
meanDiff3 <- means[[2]] - means[[3]]
covMean1 <- (covs[[1]] + covs[[2]]) / 2
covMean2 <- (covs[[1]] + covs[[3]]) / 2
covMean3 <- (covs[[2]] + covs[[3]]) / 2
smd1 <- drop(sqrt(t(meanDiff1) %*% MASS::ginv(covMean1) %*% t(t(meanDiff1))))
smd2 <- drop(sqrt(t(meanDiff2) %*% MASS::ginv(covMean2) %*% t(t(meanDiff2))))
smd3 <- drop(sqrt(t(meanDiff3) %*% MASS::ginv(covMean3) %*% t(t(meanDiff3))))
## Calculate using multi-category to SMD function
outSmd <- StdDiffMulti(variable = X, group = group)
expect_equal(length(outSmd), 3)
expect_equal(outSmd, c(smd1,smd2,smd3))
##
## Repeat for dummy variable matrix with all columns (k-1 independent)
dummyX <- create_dummy(X)
lstDummyX <- split(x = as.data.frame(dummyX), f = group, drop = FALSE)
## Means for each indicator
means <- lapply(lstDummyX, MultinomialMeans)
## cov mat for each indicators
covs <- lapply(means, MultinomialVar)
meanDiff1 <- means[[1]] - means[[2]]
meanDiff2 <- means[[1]] - means[[3]]
meanDiff3 <- means[[2]] - means[[3]]
covMean1 <- (covs[[1]] + covs[[2]]) / 2
covMean2 <- (covs[[1]] + covs[[3]]) / 2
covMean3 <- (covs[[2]] + covs[[3]]) / 2
smd1 <- drop(sqrt(t(meanDiff1) %*% MASS::ginv(covMean1) %*% t(t(meanDiff1))))
smd2 <- drop(sqrt(t(meanDiff2) %*% MASS::ginv(covMean2) %*% t(t(meanDiff2))))
smd3 <- drop(sqrt(t(meanDiff3) %*% MASS::ginv(covMean3) %*% t(t(meanDiff3))))
expect_equal(outSmd, c(smd1,smd2,smd3))
## Check behavior with a binary variable
set.seed(102)
## Binary variable
X <- rbinom(n = 100, size = 1, prob = 0.2)
group <- rep(c(0,1), c(20,80))
dummyX <- create_dummy(X)[, -1, drop = FALSE]
lstDummyX <- split(x = as.data.frame(dummyX), f = group, drop = FALSE)
## Means for each indicator
means <- lapply(lstDummyX, MultinomialMeans)
## cov mat for each indicators
covs <- lapply(means, MultinomialVar)
##
meanDiff1 <- means[[1]] - means[[2]]
covMean1 <- (covs[[1]] + covs[[2]]) / 2
smd1 <- drop(sqrt(t(meanDiff1) %*% MASS::ginv(covMean1) %*% t(t(meanDiff1))))
## Comparison to manual calculation
expect_equal(StdDiffMulti(variable = X, group = group), smd1)
## Comparison to continuous version with binary support
expect_equal(StdDiffMulti(variable = X, group = group),
StdDiff(variable = X, group = group, binary = TRUE))
## Check missing value handling (binary)
X[13] <- NA
## within group values
means <- tapply(X, group, mean, na.rm = TRUE)
covs <- means * (1 - means)
## Diff and cov mean
meanDiff1 <- means[[1]] - means[[2]]
covMean1 <- (covs[[1]] + covs[[2]]) / 2
smd1 <- drop(sqrt(t(meanDiff1) %*% MASS::ginv(covMean1) %*% t(t(meanDiff1))))
smd2 <- meanDiff1 / sqrt(covMean1)
expect_equal(smd1, smd2)
expect_equal(StdDiff(variable = X, group = group, binary = TRUE), smd1)
expect_equal(StdDiff(variable = X, group = group, binary = TRUE), smd2)
expect_equal(StdDiffMulti(variable = X, group = group),
StdDiff(variable = X, group = group, binary = TRUE))
expect_equal(StdDiffMulti(variable = X, group = group),
StdDiff(variable = X, group = group, binary = TRUE))
})
test_that("mutlinomial SMD for survey data is correct", {
set.seed(102)
## 4-category variable
probs <- c(0.1,0.3,0.5,0.2)
X <- rmultinom(n = 100, size = 1, prob = probs)
X <- as.vector(seq_along(probs) %*% X)
## Three groups
group <- c(rep(1,20), rep(2,30), rep(3,50))
dat <- data.frame(X = X, group = group, weight = 1)
datSvy <- svydesign(ids = ~ 1, data = dat, weights = ~ weight)
## Check tables match with weight of 1 for everyone
tab1svytable <- svytable(~ group + X, design = datSvy)
tab1xtabs <- xtabs(~ group + X, data = dat)
expect_true(all(tab1svytable == tab1xtabs))
## Check equality for equal weight case
expect_equal(svyStdDiffMulti("X", "group", datSvy),
StdDiffMulti(dat$X, group = dat$group))
## Missing value is introduced
dat$X[13] <- NA
## Recreate
datSvy <- svydesign(ids = ~ 1, data = dat, weights = ~ weight)
## Tables again
tab1svytable <- svytable(~ group + X, design = datSvy)
tab1xtabs <- xtabs(~ group + X, data = dat)
expect_true(all(tab1svytable == tab1xtabs))
expect_equal(svyStdDiffMulti("X", "group", datSvy),
StdDiffMulti(dat$X, group = dat$group))
})
###
### User level functionality check
################################################################################
test_that("SMDs are correctly shown in print()", {
vars <- c("race","agecat")
strata <- c("HI_CHOL","RIAGENDR")
## Create an interaction variable
nhanes$strataVar <- interaction(nhanes$HI_CHOL, nhanes$RIAGENDR, sep = ":")
## Create an weighted data
nhanesSvy <- svydesign(ids = ~ SDMVPSU, strata = ~ SDMVSTRA, weights = ~ WTMEC2YR,
nest = TRUE, data = nhanes)
## Unweighed version
## Create a table
tab1 <- CreateTableOne(vars = vars, strata = strata, data = nhanes)
## Show summary including
summary(tab1)
expect_output(summary(tab1), "Standardize mean differences")
## First variable changes fastest, and its consistent with by()'s order
expect_equal(levels(nhanes$strataVar),
c("0:1", "1:1", "0:2", "1:2"))
expect_equal(levels(nhanes$strataVar),
colnames(print(tab1$ContTable))[1:4])
## Check ordering is correct (continuous)
expect_equal(as.vector(attr(tab1$ContTable, "smd"))[-1],
StdDiff(nhanes$race, nhanes$strataVar))
## Check ordering is correct (categorical)
expect_equal(as.vector(attr(tab1$CatTable, "smd"))[-1],
StdDiffMulti(nhanes$agecat, nhanes$strataVar))
out1 <- print(tab1, smd = TRUE)
## With default test = TRUE, missing = FALSE, last column should be SMD.
expect_equal(as.vector(out1[,ncol(out1)][2:3]),
c(sprintf(" %.3f", attr(tab1$ContTable, "smd")[1,1]),
sprintf(" %.3f", attr(tab1$CatTable, "smd")[1,1])))
})
kaz-yos/tableone documentation built on April 20, 2022, 12:16 p.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
################################################################################
### Unit tests for the standardized mean difference modules
## Reference: http://adv-r.had.co.nz/Testing.html
## Created on:
## Author: Kazuki Yoshida
################################################################################
### Structure
## expectations within tests within context
###
### References
################################################################################
## A unified approach to measuring the effect size between two groups using SAS
## http://support.sas.com/resources/papers/proceedings12/335-2012.pdf
###
### Prepare environment
################################################################################
library(testthat)
library(survey)
library(Matrix)
###
### Actual tests
################################################################################
context("Tests for functions for standardized mean differences")
### Provide data
data(nhanes)
nhanesSvy <- svydesign(ids = ~ SDMVPSU, strata = ~ SDMVSTRA, weights = ~ WTMEC2YR,
nest = TRUE, data = nhanes)
## ‘SDMVPSU’ Primary sampling units
## ‘SDMVSTRA’ Sampling strata
## ‘WTMEC2YR’ Sampling weights
## ‘HI_CHOL’ Numeric vector: 1 for total cholesterol over 240mg/dl, 0
## under 240mg/dl
## ‘race’ 1=Hispanic, 2=non-Hispanic white, 3=non-Hispanic black,
## 4=other
## ‘agecat’ Age group ‘(0,19]’ ‘(19,39]’ ‘(39,59]’ ‘(59,Inf]’
## ‘RIAGENDR’ Gender: 1=male, 2=female
## Use HI_CHOL for continuous, RIAGENDR for binary, race and agecat as categorical
### Old functions explicitly for 3 groups (Keep; used for testing)
## Standardize differences
StdDiffOld <- function(variable, group) {
## For each group
means <- tapply(variable, group, mean)
vars <- tapply(variable, group, var)
## Calculate for all three possible pairs
g1g2 <- abs(means[1] - means[2]) / sqrt((vars[1] + vars[2]) / 2)
g2g3 <- abs(means[2] - means[3]) / sqrt((vars[2] + vars[3]) / 2)
g3g1 <- abs(means[3] - means[1]) / sqrt((vars[3] + vars[1]) / 2)
## 2vs1, 3vs1, then 3vs2 to be consistent with regression
out <- c(g1g2, g3g1, g2g3)
names(out) <- c("3vs2","3vs1","2vs1")
out
}
svyStdDiffOld <- function(varName, groupName, design) {
varFormula <- as.formula(paste("~", varName))
groupFormula <- as.formula(paste("~", groupName))
means <- svyby(formula = varFormula, by = groupFormula, FUN = svymean, design = design)[,2]
vars <- svyby(formula = varFormula, by = groupFormula, FUN = svyvar, design = design)[,2]
## Calculate for all three possible pairs
g1g2 <- abs(means[1] - means[2]) / sqrt((vars[1] + vars[2]) / 2)
g2g3 <- abs(means[2] - means[3]) / sqrt((vars[2] + vars[3]) / 2)
g3g1 <- abs(means[3] - means[1]) / sqrt((vars[3] + vars[1]) / 2)
## 2vs1, 3vs1, then 3vs2 to be consistent with regression
out <- c(g1g2, g3g1, g2g3)
names(out) <- c("3vs2","3vs1","2vs1")
out
}
test_that("continuous standardized difference is correct (nhanes unweighted)", {
## Two group with only one contrast 1 vs 2
means1 <- tapply(nhanes$HI_CHOL, nhanes$RIAGENDR, mean, na.rm = TRUE)
vars1 <- tapply(nhanes$HI_CHOL, nhanes$RIAGENDR, var, na.rm = TRUE)
meanDiffs1 <- (means1[1] - means1[2]) / sqrt((vars1[1] + vars1[2]) / 2)
names(meanDiffs1) <- NULL
expect_equal(StdDiff(nhanes$HI_CHOL, nhanes$RIAGENDR), abs(meanDiffs1))
## Four groups with 6 contrasts
means2 <- tapply(nhanes$HI_CHOL, nhanes$race, mean, na.rm = TRUE)
vars2 <- tapply(nhanes$HI_CHOL, nhanes$race, var, na.rm = TRUE)
meanDiffs2 <-
c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
names(meanDiffs2) <- NULL
## Individual numbers
expect_equal(StdDiff(nhanes$HI_CHOL, nhanes$race), abs(meanDiffs2))
## Average across
expect_equal(mean(StdDiff(nhanes$HI_CHOL, nhanes$race)),
mean(abs(meanDiffs2)))
})
test_that("continuous standardized difference is correct (nhanes weighted)", {
## Two group
means1 <- svyby( ~ HI_CHOL, by = ~ RIAGENDR, design = nhanesSvy, FUN = svymean, na.rm = TRUE)[,2]
vars1 <- svyby( ~ HI_CHOL, by = ~ RIAGENDR, design = nhanesSvy, FUN = svyvar, na.rm = TRUE)[,2]
meanDiffs1 <- (means1[1] - means1[2]) / sqrt((vars1[1] + vars1[2]) / 2)
expect_equal(svyStdDiff("HI_CHOL", "RIAGENDR", design = nhanesSvy),
abs(meanDiffs1))
## Four groups
means2 <- svyby( ~ HI_CHOL, by = ~ race, design = nhanesSvy, FUN = svymean, na.rm = TRUE)[,2]
vars2 <- svyby( ~ HI_CHOL, by = ~ race, design = nhanesSvy, FUN = svyvar, na.rm = TRUE)[,2]
meanDiffs2 <-
c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
names(meanDiffs2) <- NULL
## Individual numbers
expect_equal(svyStdDiff("HI_CHOL", "race", design = nhanesSvy),
abs(meanDiffs2))
## Average across
expect_equal(mean(svyStdDiff("HI_CHOL", "race", design = nhanesSvy)),
mean(abs(meanDiffs2)))
})
test_that("binary standardized difference is correct (nhanes unweighted)", {
## Two group with only one contrast 1 vs 2
means1 <- tapply(nhanes$HI_CHOL, nhanes$RIAGENDR, mean, na.rm = TRUE)
vars1 <- means1 * (1 - means1)
meanDiffs1 <- (means1[1] - means1[2]) / sqrt((vars1[1] + vars1[2]) / 2)
names(meanDiffs1) <- NULL
expect_equal(StdDiff(nhanes$HI_CHOL, nhanes$RIAGENDR, binary = TRUE),
abs(meanDiffs1))
## Four groups with 6 contrasts
means2 <- tapply(nhanes$HI_CHOL, nhanes$race, mean, na.rm = TRUE)
vars2 <- means2 * (1 - means2)
meanDiffs2 <-
c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
names(meanDiffs2) <- NULL
## Individual numbers
expect_equal(StdDiff(nhanes$HI_CHOL, nhanes$race, binary = TRUE),
abs(meanDiffs2))
## Average across
expect_equal(mean(StdDiff(nhanes$HI_CHOL, nhanes$race, binary = TRUE)),
mean(abs(meanDiffs2)))
})
test_that("binary standardized difference is correct (nhanes weighted)", {
## Two group
means1 <- svyby( ~ HI_CHOL, by = ~ RIAGENDR, design = nhanesSvy, FUN = svymean, na.rm = TRUE)[,2]
vars1 <- means1 * (1 - means1)
meanDiffs1 <- (means1[1] - means1[2]) / sqrt((vars1[1] + vars1[2]) / 2)
expect_equal(svyStdDiff("HI_CHOL", "RIAGENDR", design = nhanesSvy, binary = TRUE),
abs(meanDiffs1))
## Four groups
means2 <- svyby( ~ HI_CHOL, by = ~ race, design = nhanesSvy, FUN = svymean, na.rm = TRUE)[,2]
vars2 <- means2 * (1 - means2)
meanDiffs2 <-
c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
names(meanDiffs2) <- NULL
## Individual numbers
expect_equal(svyStdDiff("HI_CHOL", "race", design = nhanesSvy, binary = TRUE),
abs(meanDiffs2))
## Average across
expect_equal(mean(svyStdDiff("HI_CHOL", "race", design = nhanesSvy, binary = TRUE)),
mean(abs(meanDiffs2)))
})
test_that("multinomal helpers are correct", {
means <- c(0.310336708264657, 0.224393689663292, 0.234283023310572, 0.230986578761479)
covs <- MultinomialVar(means)
## Expectations
## Diagonal p(1-p)
expect_equal(diag(covs), means * (1 - means))
## Off diagonal -1 * p_i * p_j
expect_equal(covs[1,2], -1 * means[1] * means[2])
expect_equal(covs[1,3], -1 * means[1] * means[3])
expect_equal(covs[1,4], -1 * means[1] * means[4])
expect_equal(covs[2,3], -1 * means[2] * means[3])
expect_equal(covs[2,4], -1 * means[2] * means[4])
expect_equal(covs[3,4], -1 * means[3] * means[4])
})
test_that("categorical standardized difference is correct (nhanes unweighted)", {
## Four groups with 6 contrasts
means2 <- tapply(nhanes$HI_CHOL, nhanes$race, mean, na.rm = TRUE)
vars2 <- means2 * (1 - means2)
meanDiffs2 <-
c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
names(meanDiffs2) <- NULL
## Individual numbers
expect_equal(StdDiffMulti(nhanes$HI_CHOL, nhanes$race),
abs(meanDiffs2))
## Average across
expect_equal(mean(StdDiffMulti(nhanes$HI_CHOL, nhanes$race)),
mean(abs(meanDiffs2)))
## Two group with only one contrast 1 vs 2
strataByLevels1 <- xtabs( ~ RIAGENDR + agecat, data = nhanes)
## drop first column after calculating proportion
propTable1 <- prop.table(strataByLevels1, margin = 1)[, -1, drop = FALSE]
meanDiffs1 <- propTable1[1,] - propTable1[2,]
covMean1 <- (MultinomialVar(propTable1[1,]) + MultinomialVar(propTable1[2,])) / 2
## R is not strict about transposition
expect_equal(meanDiffs1 %*% MASS::ginv(covMean1) %*% meanDiffs1,
t(meanDiffs1) %*% MASS::ginv(covMean1) %*% t(t(meanDiffs1)))
## Test actual functions
expect_equal(StdDiffMulti(nhanes$agecat, nhanes$RIAGENDR),
sqrt(drop(t(meanDiffs1) %*% MASS::ginv(covMean1) %*% t(t(meanDiffs1)))))
## Four groups with 6 contrasts
strataByLevels2 <- xtabs( ~ race + agecat, data = nhanes)
propTable2 <- prop.table(strataByLevels2, margin = 1)[, -1, drop = FALSE]
meanDiffs2 <- list(propTable2[1,] - propTable2[2,],
propTable2[1,] - propTable2[3,],
propTable2[1,] - propTable2[4,],
propTable2[2,] - propTable2[3,],
propTable2[2,] - propTable2[4,],
propTable2[3,] - propTable2[4,])
covMean2 <-
list((MultinomialVar(propTable2[1,]) + MultinomialVar(propTable2[2,])) / 2,
(MultinomialVar(propTable2[1,]) + MultinomialVar(propTable2[3,])) / 2,
(MultinomialVar(propTable2[1,]) + MultinomialVar(propTable2[4,])) / 2,
(MultinomialVar(propTable2[2,]) + MultinomialVar(propTable2[3,])) / 2,
(MultinomialVar(propTable2[2,]) + MultinomialVar(propTable2[4,])) / 2,
(MultinomialVar(propTable2[3,]) + MultinomialVar(propTable2[4,])) / 2)
## These should match in length.
expect_equal(length(meanDiffs2), length(covMean2))
smds <- unlist(lapply(seq_along(meanDiffs2), function(i) {
sqrt(drop(t(meanDiffs2[[i]]) %*% MASS::ginv(covMean2[[i]]) %*% t(t(meanDiffs2[[i]]))))
}))
## Individual numbers
expect_equal(StdDiffMulti(nhanes$agecat, nhanes$race),
smds)
## Average across
expect_equal(mean(StdDiffMulti(nhanes$agecat, nhanes$race)),
mean(smds))
})
test_that("categorical standardized difference is correct (nhanes weighted)", {
## Binary four groups
means2 <- svyby( ~ HI_CHOL, by = ~ race, design = nhanesSvy, FUN = svymean, na.rm = TRUE)[,2]
vars2 <- means2 * (1 - means2)
meanDiffs2 <-
c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
names(meanDiffs2) <- NULL
## Individual numbers
expect_equal(svyStdDiffMulti("HI_CHOL", "race", design = nhanesSvy),
abs(meanDiffs2))
## Average across
expect_equal(mean(svyStdDiffMulti("HI_CHOL", "race", design = nhanesSvy)),
mean(abs(meanDiffs2)))
## Two group with only one contrast 1 vs 2
strataByLevels1 <- svytable( ~ RIAGENDR + agecat, design = nhanesSvy)
propTable1 <- prop.table(strataByLevels1, margin = 1)[, -1, drop = FALSE]
meanDiffs1 <- propTable1[1,] - propTable1[2,]
covMean1 <- (MultinomialVar(propTable1[1,]) + MultinomialVar(propTable1[2,])) / 2
## Test actual functions
expect_equal(svyStdDiffMulti("agecat", "RIAGENDR", design = nhanesSvy),
sqrt(drop(t(meanDiffs1) %*% MASS::ginv(covMean1) %*% t(t(meanDiffs1)))))
## Four groups with 6 contrasts
strataByLevels2 <- svytable( ~ race + agecat, design = nhanesSvy)
propTable2 <- prop.table(strataByLevels2, margin = 1)[, -1, drop = FALSE]
meanDiffs2 <- list(propTable2[1,] - propTable2[2,],
propTable2[1,] - propTable2[3,],
propTable2[1,] - propTable2[4,],
propTable2[2,] - propTable2[3,],
propTable2[2,] - propTable2[4,],
propTable2[3,] - propTable2[4,])
covMean2 <-
list((MultinomialVar(propTable2[1,]) + MultinomialVar(propTable2[2,])) / 2,
(MultinomialVar(propTable2[1,]) + MultinomialVar(propTable2[3,])) / 2,
(MultinomialVar(propTable2[1,]) + MultinomialVar(propTable2[4,])) / 2,
(MultinomialVar(propTable2[2,]) + MultinomialVar(propTable2[3,])) / 2,
(MultinomialVar(propTable2[2,]) + MultinomialVar(propTable2[4,])) / 2,
(MultinomialVar(propTable2[3,]) + MultinomialVar(propTable2[4,])) / 2)
## These should match in length.
expect_equal(length(meanDiffs2), length(covMean2))
smds <- unlist(lapply(seq_along(meanDiffs2), function(i) {
sqrt(drop(t(meanDiffs2[[i]]) %*% MASS::ginv(covMean2[[i]]) %*% t(t(meanDiffs2[[i]]))))
}))
## Individual numbers
expect_equal(svyStdDiffMulti("agecat", "race", design = nhanesSvy),
smds)
## Average across
expect_equal(mean(svyStdDiffMulti("agecat", "race", design = nhanesSvy)),
mean(smds))
})
### Test on anomalous data
################################################################################
test_that("decent results are returned for anomalous/difficult data", {
data(nhanes)
nhanes$onlyOne <- 1
nhanes$onlyNa <- as.numeric(NA)
nhanes$logi <- nhanes$RIAGENDR == 1
nhanes$race2 <- nhanes$race
nhanesSvy <- svydesign(ids = ~ SDMVPSU, strata = ~ SDMVSTRA,
weights = ~ WTMEC2YR, nest = TRUE,
data = nhanes)
## Note: 2015-08-01
## Categorical SMD: When both of two group have at one or more zero
## categories, both contribute singular matrices to the pooled
## variance covariance matrix S in Yang and Dalton 2012.
##
## When both groups have a single category that is probability 1 and
## other cells that are all zeros, both contribute zero matrices to
## the pooled variance covariance matrix S. solve() dies on this
## singular matrix. MASS::ginv() returns generalized inverse matrix,
## which is the zero matrix of the same dimension, instead.
##
## This results in SMD of zero regardless of the mean vector difference
## (T-C in Yang and Dalton 2012).
##
## This behavior looks odd, thus, categorical SMD functions were
## modified to return NaN when S is a zero matrix.
## Logical
## 0 due to [0]^- = 0
table(nhanes$logi, nhanes$RIAGENDR)
expect_equal(StdDiffMulti(nhanes$logi, group = nhanes$RIAGENDR), NaN)
## Matches with result from original variable
expect_equal(StdDiffMulti(nhanes$logi, group = nhanes$race),
StdDiffMulti(nhanes$RIAGENDR, group = nhanes$race))
## Multiple group where T-C is not a zero vector.
## All groups have single group that is probability 1, thus,
## all contrasts have zero matrix S.
table(nhanes$race2, nhanes$race)
expect_equal(StdDiffMulti(nhanes$race2, group = nhanes$race), rep(NaN, 6))
expect_equal(svyStdDiffMulti("race2", group = "race", design = nhanesSvy), rep(NaN, 6))
## Identical constant comparison
## T-C vector is a zero vector, but divisor is also zero
## NaN due to division by zero variance -> Redefined as zero
by(nhanes$onlyOne, nhanes$RIAGENDR, summary)
expect_equal(StdDiff(nhanes$onlyOne, group = nhanes$RIAGENDR), 0)
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined NaN in (svy)StdDiffMulti
table(nhanes$onlyOne, nhanes$RIAGENDR)
expect_equal(StdDiffMulti(nhanes$onlyOne, group = nhanes$RIAGENDR), 0)
## When weighted problematic
means1 <- svyby(~ onlyOne, by = ~ RIAGENDR, nhanesSvy, FUN = svymean)[,2]
vars1 <- svyby(~ onlyOne, by = ~ RIAGENDR, nhanesSvy, FUN = svyvar)[,2]
## Very small difference is inflated by even smaller variance
## on sparc-sun-solaris sign was opposite; abs() solves this issue
## as svyStdDiff() uses abs() internally
expect_equal(svyStdDiff("onlyOne", "RIAGENDR", nhanesSvy),
abs((means1[1] - means1[2]) / sqrt(sum(vars1) / 2)))
## NaN should be the case, but it's not, but it's consistent with survey
## expect_equal(svyStdDiff("onlyOne", "RIAGENDR", nhanesSvy), NaN)
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined NaN in (svy)StdDiffMulti
## No error even with a single level variable (constant) as redundant
## level drop from table occurs only when 2+ levels are present.
## If any group has more than 2 levels, then strata-by-level table
## is correctly created, which is not the case here. Redefined as 0.
expect_equal(svyStdDiffMulti("onlyOne", "RIAGENDR", nhanesSvy), 0)
## Four groups (six contrasts)
## NaN due to division by zero variance -> Redefined as 0
by(nhanes$onlyOne, nhanes$race, summary)
expect_equal(StdDiff(nhanes$onlyOne, group = nhanes$race), rep(0, 6))
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined NaN in (svy)StdDiffMulti
expect_equal(StdDiffMulti(nhanes$onlyOne, group = nhanes$race), rep(0, 6))
## When weighted problematic; not in this case??
means2 <- svyby(~ onlyOne, by = ~ race, nhanesSvy, FUN = svymean)[,2]
vars2 <- svyby(~ onlyOne, by = ~ race, nhanesSvy, FUN = svyvar)[,2]
meanDiffs2 <- c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
## Redefined as 0 as T-C is a zero vector
## Fails if no long double is available in R.
##
## noLD: Tests without long double (tests on x86_64 Linux with R-devel configured --without-long-double)
## ── 1. Failure: decent results are returned for anomalous/difficult data (@test-m
## svyStdDiff("onlyOne", "race", nhanesSvy) not equal to rep(0, 6).
## 6/6 mismatches (average diff: 1.33)
## [1] 0.117 - 0 == 0.117
## [2] 1.987 - 0 == 1.987
## [3] 1.129 - 0 == 1.129
## [4] 1.997 - 0 == 1.997
## [5] 1.162 - 0 == 1.162
## [6] 1.608 - 0 == 1.608
##
## base::.Machine$sizeof.longdouble: the number of bytes in a C long double type. Will be zero if there is no such type (or its use was disabled when R was built), otherwise possibly 12 (most 32-bit builds) or 16 (most 64-bit builds).
if (base::.Machine$sizeof.longdouble >= 16) {
## Test only when long double is available in 64bit system.
expect_equal(svyStdDiff("onlyOne", "race", nhanesSvy), rep(0,6))
}
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined NaN in (svy)StdDiffMulti
expect_equal(svyStdDiffMulti("onlyOne", "race", nhanesSvy), rep(0, 6))
## onlyNa
## NA as na.rm is turned off
expect_warning(expect_equal(StdDiff(nhanes$onlyNa, group = nhanes$RIAGENDR),
as.numeric(NA)),
"Variable has only NA's in at least one stratum. na.rm turned off.")
## defined 0 as NA is a level
expect_warning(expect_equal(StdDiffMulti(nhanes$onlyNa, group = nhanes$RIAGENDR), 0),
"Variable has only NA's in all strata. Regarding NA as a level")
## When weighted problematic
means1 <- svyby(~ onlyNa, by = ~ RIAGENDR, nhanesSvy, FUN = svymean)[,2]
vars1 <- svyby(~ onlyNa, by = ~ RIAGENDR, nhanesSvy, FUN = svyvar)[,2]
## Very small difference is inflated by even smaller variance
expect_warning(svyStdDiff("onlyNa", "RIAGENDR", nhanesSvy),
"onlyNa has only NA's in at least one stratum. na.rm turned off.")
expect_warning(expect_equal(svyStdDiff("onlyNa", "RIAGENDR", nhanesSvy),
as.numeric(NA)),
"onlyNa has only NA's in at least one stratum. na.rm turned off.")
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined NaN in (svy)StdDiffMulti
expect_warning(expect_equal(svyStdDiffMulti("onlyNa", "RIAGENDR", nhanesSvy), 0),
"onlyNa has only NA's in all strata. Regarding NA as a level.")
## Four groups (six contrasts)
## NaN due to division by zero variance
expect_warning(expect_equal(StdDiff(nhanes$onlyNa, group = nhanes$race), rep(NaN, 6)),
"Variable has only NA's in at least one stratum. na.rm turned off.")
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined NaN in (svy)StdDiffMulti
expect_warning(expect_equal(StdDiffMulti(nhanes$onlyNa, group = nhanes$race), rep(0, 6)),
"Variable has only NA's in all strata. Regarding NA as a level.")
## When weighted problematic; not in this case??
means2 <- svyby(~ onlyNa, by = ~ race, nhanesSvy, FUN = svymean)[,2]
vars2 <- svyby(~ onlyNa, by = ~ race, nhanesSvy, FUN = svyvar)[,2]
meanDiffs2 <- c((means2[1] - means2[2]) / sqrt((vars2[1] + vars2[2]) / 2),
(means2[1] - means2[3]) / sqrt((vars2[1] + vars2[3]) / 2),
(means2[1] - means2[4]) / sqrt((vars2[1] + vars2[4]) / 2),
(means2[2] - means2[3]) / sqrt((vars2[2] + vars2[3]) / 2),
(means2[2] - means2[4]) / sqrt((vars2[2] + vars2[4]) / 2),
(means2[3] - means2[4]) / sqrt((vars2[3] + vars2[4]) / 2))
expect_warning(expect_equal(svyStdDiff("onlyNa", "race", nhanesSvy), meanDiffs2),
"onlyNa has only NA's in at least one stratum. na.rm turned off.")
expect_warning(expect_equal(svyStdDiff("onlyNa", "race", nhanesSvy), rep(NaN, 6)),
"onlyNa has only NA's in at least one stratum. na.rm turned off.")
## 0 because [0]^- = 0, and [1]^T [0]^-1 [1] = 0; defined 0 as NA is a level
expect_warning(expect_equal(svyStdDiffMulti("onlyNa", "race", nhanesSvy), rep(0, 6)),
"onlyNa has only NA's in all strata. Regarding NA as a level.")
})
### Define a function for pooled S matrix (used in two tests)
pooledSMat <- function(strataByLevels) {
lstMeans <- LstMeansFromFullTable(strataByLevels)
S1 <- MultinomialVar(lstMeans[[1]])
S2 <- MultinomialVar(lstMeans[[2]])
S <- (S1 + S2) / 2
## first group element, second group element, pooled
list(S1 = S1, S2 = S2, S = S)
}
test_that("zero-cell vectors are decently handled by categorical SMD (artificial data)", {
set.seed(2015080200)
## Simple artificial data
wt <- runif(n = 100)
dat <- data.frame(group = rep(c(1,2), each = 100),
var1 = factor(c(rep(1:4, c(0,10,20,70)), rep(1:4, c( 0,10,20,70))), level = 1:4),
var2 = factor(c(rep(1:4, c(0,10,20,70)), rep(1:4, c( 0,20,30,50))), level = 1:4),
var3 = factor(c(rep(1:4, c(0,10,20,70)), rep(1:4, c(20, 0,30,50))), level = 1:4),
var4 = factor(c(rep(1:4, c(0,10,20,70)), rep(1:4, c(10,20,30,40))), level = 1:4),
wt = rep(wt, 2))
datSvy <- svydesign(ids = ~ 1, data = dat, weights = ~wt)
## Expectations
## First case
var1S <- pooledSMat(table(dat$group, dat$var1))
expect_true(rankMatrix(var1S[[1]]) < 3)
expect_true(rankMatrix(var1S[[2]]) < 3)
expect_true(rankMatrix(var1S[[3]]) < 3)
## Using ginv zero due to numerator (OK)
expect_equal(StdDiffMulti(dat$var1, dat$group), 0)
## Second case
var2S <- pooledSMat(table(dat$group, dat$var2))
expect_true(rankMatrix(var2S[[1]]) < 3)
expect_true(rankMatrix(var2S[[2]]) < 3)
expect_true(rankMatrix(var2S[[3]]) < 3)
## Using ginv() non-zero
expect_true(StdDiffMulti(dat$var2, dat$group) > 0)
## Third case
var3S <- pooledSMat(table(dat$group, dat$var3))
expect_true(rankMatrix(var3S[[1]]) < 3)
expect_true(rankMatrix(var3S[[2]]) < 3)
## pooled can be full rank if zero cell positions are different
expect_true(rankMatrix(var3S[[3]]) == 3)
## Using true inverse from ginv, non-zero
expect_true(StdDiffMulti(dat$var3, dat$group) > 0)
## Third case
var4S <- pooledSMat(table(dat$group, dat$var4))
expect_true(rankMatrix(var4S[[1]]) < 3)
## All cells populated and full rank
expect_true(rankMatrix(var4S[[2]]) == 3)
## pooled can be full rank if one group is full rank
expect_true(rankMatrix(var4S[[3]]) == 3)
expect_true(StdDiffMulti(dat$var4, dat$group) > 0)
## Weighted
## First case
var1S <- pooledSMat(svytable( ~ group + var1, datSvy))
expect_true(rankMatrix(var1S[[1]]) < 3)
expect_true(rankMatrix(var1S[[2]]) < 3)
expect_true(rankMatrix(var1S[[3]]) < 3)
## Using ginv zero due to numerator (OK)
expect_equal(svyStdDiffMulti("var1", "group", datSvy), 0)
## Second case
var2S <- pooledSMat(svytable( ~ group + var2, datSvy))
expect_true(rankMatrix(var2S[[1]]) < 3)
expect_true(rankMatrix(var2S[[2]]) < 3)
expect_true(rankMatrix(var2S[[3]]) < 3)
## Using ginv() non-zero
expect_true(svyStdDiffMulti("var2", "group", datSvy) > 0)
## Third case
var3S <- pooledSMat(svytable( ~ group + var3, datSvy))
expect_true(rankMatrix(var3S[[1]]) < 3)
expect_true(rankMatrix(var3S[[2]]) < 3)
## pooled can be full rank if zero cell positions are different
expect_true(rankMatrix(var3S[[3]]) == 3)
## Using true inverse from ginv, non-zero
expect_true(svyStdDiffMulti("var3", "group", datSvy) > 0)
## Third case
var4S <- pooledSMat(svytable( ~ group + var4, datSvy))
expect_true(rankMatrix(var4S[[1]]) < 3)
## All cells populated and full rank
expect_true(rankMatrix(var4S[[2]]) == 3)
## pooled can be full rank if one group is full rank
expect_true(rankMatrix(var4S[[3]]) == 3)
expect_true(svyStdDiffMulti("var4", "group", datSvy) > 0)
})
test_that("zero-cell vectors are decently handled by categorical SMD (real data)", {
data(nhanes)
set.seed(2015080200)
## grouping is by a binary nhanes$RIAGENDR
## (0,0.1,0.2,0.7)^T for both groups: same distribution
nhanes$var1 <- NA
nhanes$var1[nhanes$RIAGENDR == 1] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 1)), size = 1, prob = c(0,0.1,0.2,0.7)))
nhanes$var1[nhanes$RIAGENDR == 2] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 2)), size = 1, prob = c(0,0.1,0.2,0.7)))
nhanes$var1 <- factor(nhanes$var1, levels = 1:4)
## (0,0.1,0.2,0.7)^T vs (0,0.2,0.3,0.5)^T: different distribution, but same zero cell
nhanes$var2 <- NA
nhanes$var2[nhanes$RIAGENDR == 1] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 1)), size = 1, prob = c(0,0.1,0.2,0.7)))
nhanes$var2[nhanes$RIAGENDR == 2] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 2)), size = 1, prob = c(0,0.2,0.3,0.5)))
nhanes$var2 <- factor(nhanes$var2, levels = 1:4)
## (0,0.1,0.2,0.7)^T vs (0.2,0,0.3,0.5)^T: different distribution and zero cell
nhanes$var3 <- NA
nhanes$var3[nhanes$RIAGENDR == 1] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 1)), size = 1, prob = c(0,0.1,0.2,0.7)))
nhanes$var3[nhanes$RIAGENDR == 2] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 2)), size = 1, prob = c(0.2,0,0.3,0.5)))
nhanes$var3 <- factor(nhanes$var3, levels = 1:4)
## (0,0.1,0.2,0.7)^T vs (0.1,0.2,0.3,0.5)^T: different distribution and one is ok
nhanes$var4 <- NA
nhanes$var4[nhanes$RIAGENDR == 1] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 1)), size = 1, prob = c(0,0.1,0.2,0.7)))
nhanes$var4[nhanes$RIAGENDR == 2] <-
as.vector(t(1:4) %*% rmultinom(n = nrow(subset(nhanes, RIAGENDR == 2)), size = 1, prob = c(0.1,0.2,0.3,0.4)))
nhanes$var4 <- factor(nhanes$var4, levels = 1:4)
## svydesign
nhanesSvy <- svydesign(ids = ~ SDMVPSU, strata = ~ SDMVSTRA,
weights = ~ WTMEC2YR, nest = TRUE,
data = nhanes)
## Expectations
## First case
var1S <- pooledSMat(table(nhanes$RIAGENDR, nhanes$var1))
expect_true(rankMatrix(var1S[[1]]) < 3)
expect_true(rankMatrix(var1S[[2]]) < 3)
expect_true(rankMatrix(var1S[[3]]) < 3)
## Using ginv, near zero due to near-zero numerator
expect_true(StdDiffMulti(nhanes$var1, nhanes$RIAGENDR) < 0.1)
## Second case
var2S <- pooledSMat(table(nhanes$RIAGENDR, nhanes$var2))
expect_true(rankMatrix(var2S[[1]]) < 3)
expect_true(rankMatrix(var2S[[2]]) < 3)
expect_true(rankMatrix(var2S[[3]]) < 3)
## Using ginv() non-zero
expect_true(StdDiffMulti(nhanes$var2, nhanes$RIAGENDR) > 0)
## Third case
var3S <- pooledSMat(table(nhanes$RIAGENDR, nhanes$var3))
expect_true(rankMatrix(var3S[[1]]) < 3)
expect_true(rankMatrix(var3S[[2]]) < 3)
## pooled can be full rank if zero cell positions are different
expect_true(rankMatrix(var3S[[3]]) == 3)
## Using true inverse from ginv, non-zero
expect_true(StdDiffMulti(nhanes$var3, nhanes$RIAGENDR) > 0)
## Fourth case
var4S <- pooledSMat(table(nhanes$RIAGENDR, nhanes$var4))
expect_true(rankMatrix(var4S[[1]]) < 3)
## All cells populated and full rank
expect_true(rankMatrix(var4S[[2]]) == 3)
## pooled can be full rank if one group is full rank
expect_true(rankMatrix(var4S[[3]]) == 3)
expect_true(StdDiffMulti(nhanes$var4, nhanes$RIAGENDR) > 0)
## Weighted
## First case
var1S <- pooledSMat(svytable( ~ RIAGENDR + var1, nhanesSvy))
expect_true(rankMatrix(var1S[[1]]) < 3)
expect_true(rankMatrix(var1S[[2]]) < 3)
expect_true(rankMatrix(var1S[[3]]) < 3)
## Using ginv, near zero due to small numerator vector
expect_true(svyStdDiffMulti("var1", "RIAGENDR", nhanesSvy) < 0.1)
## Second case
var2S <- pooledSMat(svytable( ~ RIAGENDR + var2, nhanesSvy))
expect_true(rankMatrix(var2S[[1]]) < 3)
expect_true(rankMatrix(var2S[[2]]) < 3)
expect_true(rankMatrix(var2S[[3]]) < 3)
## Using ginv() non-zero
expect_true(svyStdDiffMulti("var2", "RIAGENDR", nhanesSvy) > 0)
## Third case
var3S <- pooledSMat(svytable( ~ RIAGENDR + var3, nhanesSvy))
expect_true(rankMatrix(var3S[[1]]) < 3)
expect_true(rankMatrix(var3S[[2]]) < 3)
## pooled can be full rank if zero cell positions are different
expect_true(rankMatrix(var3S[[3]]) == 3)
## Using true inverse from ginv, non-zero
expect_true(svyStdDiffMulti("var3", "RIAGENDR", nhanesSvy) > 0)
## Third case
var4S <- pooledSMat(svytable( ~ RIAGENDR + var4, nhanesSvy))
expect_true(rankMatrix(var4S[[1]]) < 3)
## All cells populated and full rank
expect_true(rankMatrix(var4S[[2]]) == 3)
## pooled can be full rank if one RIAGENDR is full rank
expect_true(rankMatrix(var4S[[3]]) == 3)
expect_true(svyStdDiffMulti("var4", "RIAGENDR", nhanesSvy) > 0)
})
### Tests working on multiple variables
################################################################################
test_that("multiple variables can be looped", {
contVars <- c("WTMEC2YR", "HI_CHOL", "race", "RIAGENDR")
catVars <- c("HI_CHOL", "race", "agecat", "RIAGENDR")
## Two groups
expect_equal(as.numeric(sapply(contVars, function(var) {
StdDiff(variable = nhanes[,var], group = nhanes[,"RIAGENDR"])
})),
c(StdDiff(variable = nhanes[,contVars[1]], group = nhanes[,"RIAGENDR"]),
StdDiff(variable = nhanes[,contVars[2]], group = nhanes[,"RIAGENDR"]),
StdDiff(variable = nhanes[,contVars[3]], group = nhanes[,"RIAGENDR"]),
StdDiff(variable = nhanes[,contVars[4]], group = nhanes[,"RIAGENDR"])))
expect_equal(as.numeric(sapply(catVars, function(var) {
StdDiffMulti(variable = nhanes[,var], group = nhanes[,"RIAGENDR"])
})),
c(StdDiffMulti(variable = nhanes[,catVars[1]], group = nhanes[,"RIAGENDR"]),
StdDiffMulti(variable = nhanes[,catVars[2]], group = nhanes[,"RIAGENDR"]),
StdDiffMulti(variable = nhanes[,catVars[3]], group = nhanes[,"RIAGENDR"]),
StdDiffMulti(variable = nhanes[,catVars[4]], group = nhanes[,"RIAGENDR"])))
## Four groups
expect_equal(lapply(contVars, function(var) {
StdDiff(variable = nhanes[,var], group = nhanes[,"race"])
}),
list(StdDiff(variable = nhanes[,contVars[1]], group = nhanes[,"race"]),
StdDiff(variable = nhanes[,contVars[2]], group = nhanes[,"race"]),
StdDiff(variable = nhanes[,contVars[3]], group = nhanes[,"race"]),
StdDiff(variable = nhanes[,contVars[4]], group = nhanes[,"race"])))
expect_equal(lapply(catVars, function(var) {
StdDiffMulti(variable = nhanes[,var], group = nhanes[,"race"])
}),
list(StdDiffMulti(variable = nhanes[,catVars[1]], group = nhanes[,"race"]),
StdDiffMulti(variable = nhanes[,catVars[2]], group = nhanes[,"race"]),
StdDiffMulti(variable = nhanes[,catVars[3]], group = nhanes[,"race"]),
StdDiffMulti(variable = nhanes[,catVars[4]], group = nhanes[,"race"])))
## Two groups
expect_equal(as.numeric(sapply(contVars, function(var) {
svyStdDiff(var, "RIAGENDR", nhanesSvy)
})),
c(svyStdDiff(contVars[1], "RIAGENDR", nhanesSvy),
svyStdDiff(contVars[2], "RIAGENDR", nhanesSvy),
svyStdDiff(contVars[3], "RIAGENDR", nhanesSvy),
svyStdDiff(contVars[4], "RIAGENDR", nhanesSvy)))
expect_equal(as.numeric(sapply(catVars, function(var) {
svyStdDiffMulti(var, "RIAGENDR", nhanesSvy)
})),
c(svyStdDiffMulti(catVars[1], "RIAGENDR", nhanesSvy),
svyStdDiffMulti(catVars[2], "RIAGENDR", nhanesSvy),
svyStdDiffMulti(catVars[3], "RIAGENDR", nhanesSvy),
svyStdDiffMulti(catVars[4], "RIAGENDR", nhanesSvy)))
## Four groups
expect_equal(lapply(contVars, function(var) {
svyStdDiff(var, "race", nhanesSvy)
}),
list(svyStdDiff(contVars[1], "race", nhanesSvy),
svyStdDiff(contVars[2], "race", nhanesSvy),
svyStdDiff(contVars[3], "race", nhanesSvy),
svyStdDiff(contVars[4], "race", nhanesSvy)))
expect_equal(lapply(catVars, function(var) {
svyStdDiffMulti(var, "race", nhanesSvy)
}),
list(svyStdDiffMulti(catVars[1], "race", nhanesSvy),
svyStdDiffMulti(catVars[2], "race", nhanesSvy),
svyStdDiffMulti(catVars[3], "race", nhanesSvy),
svyStdDiffMulti(catVars[4], "race", nhanesSvy)))
})
### Older tests with simpler data
################################################################################
test_that("binary standardized difference is correct", {
## Prepare data
variable <- c(1,0,0,0,0, 1,1,0,0,0, 1,1,1,0,0)
group <- rep(1:3, each = 5)
data1 <- data.frame(variable = variable, group = group,
weight = rep(0.5, 15))
data1Svy <- svydesign(ids = ~ 1, data = data1, weights = ~ weight)
## Gold standard
props <- c(0.2, 0.4, 0.6)
vars <- c(0.2*0.8, 0.4*0.6, 0.6*0.4)
g1g2 <- abs(props[1] - props[2]) / sqrt((vars[1] + vars[2]) / 2)
g2g3 <- abs(props[2] - props[3]) / sqrt((vars[2] + vars[3]) / 2)
g3g1 <- abs(props[3] - props[1]) / sqrt((vars[3] + vars[1]) / 2)
## Expectations
## Binary case
expect_equal(StdDiff(variable, group, binary = TRUE),
c(g1g2, g3g1, g2g3))
## Continuous case
expect_equal(StdDiff(variable, group, binary = FALSE),
as.vector(StdDiffOld(variable, group)))
## Weighted
## Binary case
expect_equal(svyStdDiff("variable","group", design = data1Svy, binary = TRUE),
c(g1g2, g3g1, g2g3))
## Continuous case
expect_equal(svyStdDiff("variable","group", design = data1Svy, binary = FALSE),
as.vector(svyStdDiffOld("variable","group", design = data1Svy)))
## Multiple variables
expect_equal(StdDiffs(data = data1, vars = c("variable","variable"),
groupVar = "group", binaryVars = "variable"),
list(c(g1g2, g3g1, g2g3), c(g1g2, g3g1, g2g3)))
expect_equal(StdDiffs(data = data1, vars = c("variable","variable"),
groupVar = "group", binaryVars = NA),
list(as.vector(StdDiffOld(variable, group)),
as.vector(StdDiffOld(variable, group))))
expect_equal(svyStdDiffs(data = data1Svy, vars = c("variable","variable"),
groupVar = "group", binaryVars = "variable"),
list(c(g1g2, g3g1, g2g3), c(g1g2, g3g1, g2g3)))
expect_equal(svyStdDiffs(data = data1Svy, vars = c("variable","variable"),
groupVar = "group", binaryVars = NA),
list(as.vector(svyStdDiffOld("variable","group", design = data1Svy)),
as.vector(svyStdDiffOld("variable","group", design = data1Svy))))
})
test_that("Multinomial SMD is correct", {
create_dummy <- function(x) {
df <- data.frame(x = factor(x))
model.matrix( ~ -1 + x, data = df)
}
set.seed(102)
## 4-category variable
probs <- c(0.1,0.3,0.5,0.2)
X <- rmultinom(n = 100, size = 1, prob = probs)
X <- as.vector(seq_along(probs) %*% X)
## Three groups
group <- c(rep(1,20), rep(2,30), rep(3,50))
## Drop first column to avoid dependent column
## The result does not change if generalized inverse is used
dummyX <- create_dummy(X)[, -1, drop = FALSE]
lstDummyX <- split(x = as.data.frame(dummyX), f = group, drop = FALSE)
## Means for each indicator
means <- lapply(lstDummyX, MultinomialMeans)
## cov mat for each indicators
covs <- lapply(means, MultinomialVar)
meanDiff1 <- means[[1]] - means[[2]]
meanDiff2 <- means[[1]] - means[[3]]
meanDiff3 <- means[[2]] - means[[3]]
covMean1 <- (covs[[1]] + covs[[2]]) / 2
covMean2 <- (covs[[1]] + covs[[3]]) / 2
covMean3 <- (covs[[2]] + covs[[3]]) / 2
smd1 <- drop(sqrt(t(meanDiff1) %*% MASS::ginv(covMean1) %*% t(t(meanDiff1))))
smd2 <- drop(sqrt(t(meanDiff2) %*% MASS::ginv(covMean2) %*% t(t(meanDiff2))))
smd3 <- drop(sqrt(t(meanDiff3) %*% MASS::ginv(covMean3) %*% t(t(meanDiff3))))
## Calculate using multi-category to SMD function
outSmd <- StdDiffMulti(variable = X, group = group)
expect_equal(length(outSmd), 3)
expect_equal(outSmd, c(smd1,smd2,smd3))
##
## Repeat for dummy variable matrix with all columns (k-1 independent)
dummyX <- create_dummy(X)
lstDummyX <- split(x = as.data.frame(dummyX), f = group, drop = FALSE)
## Means for each indicator
means <- lapply(lstDummyX, MultinomialMeans)
## cov mat for each indicators
covs <- lapply(means, MultinomialVar)
meanDiff1 <- means[[1]] - means[[2]]
meanDiff2 <- means[[1]] - means[[3]]
meanDiff3 <- means[[2]] - means[[3]]
covMean1 <- (covs[[1]] + covs[[2]]) / 2
covMean2 <- (covs[[1]] + covs[[3]]) / 2
covMean3 <- (covs[[2]] + covs[[3]]) / 2
smd1 <- drop(sqrt(t(meanDiff1) %*% MASS::ginv(covMean1) %*% t(t(meanDiff1))))
smd2 <- drop(sqrt(t(meanDiff2) %*% MASS::ginv(covMean2) %*% t(t(meanDiff2))))
smd3 <- drop(sqrt(t(meanDiff3) %*% MASS::ginv(covMean3) %*% t(t(meanDiff3))))
expect_equal(outSmd, c(smd1,smd2,smd3))
## Check behavior with a binary variable
set.seed(102)
## Binary variable
X <- rbinom(n = 100, size = 1, prob = 0.2)
group <- rep(c(0,1), c(20,80))
dummyX <- create_dummy(X)[, -1, drop = FALSE]
lstDummyX <- split(x = as.data.frame(dummyX), f = group, drop = FALSE)
## Means for each indicator
means <- lapply(lstDummyX, MultinomialMeans)
## cov mat for each indicators
covs <- lapply(means, MultinomialVar)
##
meanDiff1 <- means[[1]] - means[[2]]
covMean1 <- (covs[[1]] + covs[[2]]) / 2
smd1 <- drop(sqrt(t(meanDiff1) %*% MASS::ginv(covMean1) %*% t(t(meanDiff1))))
## Comparison to manual calculation
expect_equal(StdDiffMulti(variable = X, group = group), smd1)
## Comparison to continuous version with binary support
expect_equal(StdDiffMulti(variable = X, group = group),
StdDiff(variable = X, group = group, binary = TRUE))
## Check missing value handling (binary)
X[13] <- NA
## within group values
means <- tapply(X, group, mean, na.rm = TRUE)
covs <- means * (1 - means)
## Diff and cov mean
meanDiff1 <- means[[1]] - means[[2]]
covMean1 <- (covs[[1]] + covs[[2]]) / 2
smd1 <- drop(sqrt(t(meanDiff1) %*% MASS::ginv(covMean1) %*% t(t(meanDiff1))))
smd2 <- meanDiff1 / sqrt(covMean1)
expect_equal(smd1, smd2)
expect_equal(StdDiff(variable = X, group = group, binary = TRUE), smd1)
expect_equal(StdDiff(variable = X, group = group, binary = TRUE), smd2)
expect_equal(StdDiffMulti(variable = X, group = group),
StdDiff(variable = X, group = group, binary = TRUE))
expect_equal(StdDiffMulti(variable = X, group = group),
StdDiff(variable = X, group = group, binary = TRUE))
})
test_that("mutlinomial SMD for survey data is correct", {
set.seed(102)
## 4-category variable
probs <- c(0.1,0.3,0.5,0.2)
X <- rmultinom(n = 100, size = 1, prob = probs)
X <- as.vector(seq_along(probs) %*% X)
## Three groups
group <- c(rep(1,20), rep(2,30), rep(3,50))
dat <- data.frame(X = X, group = group, weight = 1)
datSvy <- svydesign(ids = ~ 1, data = dat, weights = ~ weight)
## Check tables match with weight of 1 for everyone
tab1svytable <- svytable(~ group + X, design = datSvy)
tab1xtabs <- xtabs(~ group + X, data = dat)
expect_true(all(tab1svytable == tab1xtabs))
## Check equality for equal weight case
expect_equal(svyStdDiffMulti("X", "group", datSvy),
StdDiffMulti(dat$X, group = dat$group))
## Missing value is introduced
dat$X[13] <- NA
## Recreate
datSvy <- svydesign(ids = ~ 1, data = dat, weights = ~ weight)
## Tables again
tab1svytable <- svytable(~ group + X, design = datSvy)
tab1xtabs <- xtabs(~ group + X, data = dat)
expect_true(all(tab1svytable == tab1xtabs))
expect_equal(svyStdDiffMulti("X", "group", datSvy),
StdDiffMulti(dat$X, group = dat$group))
})
###
### User level functionality check
################################################################################
test_that("SMDs are correctly shown in print()", {
vars <- c("race","agecat")
strata <- c("HI_CHOL","RIAGENDR")
## Create an interaction variable
nhanes$strataVar <- interaction(nhanes$HI_CHOL, nhanes$RIAGENDR, sep = ":")
## Create an weighted data
nhanesSvy <- svydesign(ids = ~ SDMVPSU, strata = ~ SDMVSTRA, weights = ~ WTMEC2YR,
nest = TRUE, data = nhanes)
## Unweighed version
## Create a table
tab1 <- CreateTableOne(vars = vars, strata = strata, data = nhanes)
## Show summary including
summary(tab1)
expect_output(summary(tab1), "Standardize mean differences")
## First variable changes fastest, and its consistent with by()'s order
expect_equal(levels(nhanes$strataVar),
c("0:1", "1:1", "0:2", "1:2"))
expect_equal(levels(nhanes$strataVar),
colnames(print(tab1$ContTable))[1:4])
## Check ordering is correct (continuous)
expect_equal(as.vector(attr(tab1$ContTable, "smd"))[-1],
StdDiff(nhanes$race, nhanes$strataVar))
## Check ordering is correct (categorical)
expect_equal(as.vector(attr(tab1$CatTable, "smd"))[-1],
StdDiffMulti(nhanes$agecat, nhanes$strataVar))
out1 <- print(tab1, smd = TRUE)
## With default test = TRUE, missing = FALSE, last column should be SMD.
expect_equal(as.vector(out1[,ncol(out1)][2:3]),
c(sprintf(" %.3f", attr(tab1$ContTable, "smd")[1,1]),
sprintf(" %.3f", attr(tab1$CatTable, "smd")[1,1])))
})
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.