A package devoted to multivariate resampling-based tests. By resampling jointly on all univariate tests (e.g., sign-flip score tests by Hemerik, Goeman and Finos (2020) <doi.org/10.1111/rssb.12369>) it allows for multivariate and selective inference – i.e., weak and strong control of the familywise error rate or confidence bounds for true discovery proportion.
To install this GitHub version type (in R):
##if devtools is not installed yet:
## install.packages("devtools")
library(devtools)
install_github("livioivil/jointest")
library(jointest)Simulate data:
n=20
set.seed(123)
D=data.frame(X=rnorm(n),Z1=rnorm(n),Z2=rnorm(n))
D$Y=D$Z1+D$X+rnorm(n)
mod1=glm(Y~X+Z1+Z2,data=D)
mod2=glm(Y~X+poly(Z1,2)+Z2,data=D)
mod3=glm(Y~X+poly(Z1,2)+poly(Z2,2),data=D)
mod4=glm(Y~X+Z1+poly(Z2,2),data=D)
mods=list(mod1=mod1,mod2=mod2,mod3=mod3,mod4=mod4)Testing
res=join_flipscores(mods,n_flips = 5000, seed = 1, tested_coeffs = "X")
summary(res)
Model Coeff Estimate Score Std. Error z value Part. Cor p
1 mod1 X 1.1254046 19.78375 6.309305 3.135646 0.7605058 0.01499700
2 mod2 X 0.9552644 15.38036 5.053879 3.043278 0.7608195 0.01339732
3 mod3 X 1.0121921 14.31866 4.836669 2.960437 0.7643816 0.01279744
4 mod4 X 1.1696906 17.62979 5.875380 3.000621 0.7501552 0.00979804add adjusted
res=p.adjust(res)
summary(res)
Model Coeff Estimate Score Std. Error z value Part. Cor p
1 mod1 X 1.1254046 19.78375 6.309305 3.135646 0.7605058 0.01499700
2 mod2 X 0.9552644 15.38036 5.053879 3.043278 0.7608195 0.01339732
3 mod3 X 1.0121921 14.31866 4.836669 2.960437 0.7643816 0.01279744
4 mod4 X 1.1696906 17.62979 5.875380 3.000621 0.7501552 0.00979804
p.adj
1 0.019996
2 0.019996
3 0.019996
4 0.019996look at the combined results:
summary(combine(res))
Model Coeff Stat nMods S p
1 Overall X maxT 4 19.78375 0.00739852plot the results:
plot(res, p.values = "adjusted")Simulate data:
N=20
n=rpois(N,20)
reff=rep(rnorm(N),n)
D=data.frame(X1=rnorm(length(reff)),
X2=rep(rnorm(N),n),
Grp=factor(rep(rep(LETTERS[1:3],length.out=N),n)),
SOGG=rep(1:N,n))
D$Y=rbinom(n=nrow(D),prob=plogis( 2*D$X1 * (D$Grp=="B") + 2*D$X2+reff),size=1)Define model of interest:
formula <- Y ~ Grp * X1 + X2Define clusters structure:
cluster <- factor(D$SOGG)Define the summary statistics (here we propose the glm with firth correction from the logistf package)
summstats_within <- 'logistf::logistf(Y ~ X1, family = binomial(link = "logit"), control=logistf::logistf.control(maxit=100))'however also the classic glm function can be used:
#summstats_within <- 'glm(Y ~ X1, family = binomial(link = "logit"))'Then, compute the
res <- flip2sss(formula, D, cluster, summstats_within=summstats_within)or the family:
res <- flip2sss(formula, D, cluster, family = "binomial")Look at the results:
res <- flip2sss(formula, D, cluster, family="binomial")
summary(res)
Model Coeff Estimate Score Std. Error z value Part. Cor p
1 .Intercept. (Intercept) 4.3128538 29.0261330 24.152830 1.20176943 0.29147190 0.16116777
2 .Intercept. GrpB -0.3824975 -1.1599283 15.511347 -0.07477934 -0.01813665 0.94741052
3 .Intercept. GrpC -1.7121985 -5.5300048 16.061350 -0.34430510 -0.08350625 0.68066387
4 .Intercept. X2 9.5446012 151.9594953 51.196709 2.96814969 0.71988204 0.00579884
5 X1 X1 0.2882803 2.0179621 2.658795 0.75897611 0.17889239 0.18636273
6 X1 GrpB:X1 1.1367411 3.9785937 2.073856 1.91845180 0.45218343 0.03879224
7 X1 GrpC:X1 0.1228905 0.3970308 1.779620 0.22309867 0.05258486 0.78544291add the adjusted p.values:
res<-p.adjust(res)
summary(res)
Model Coeff Estimate Score Std. Error z value Part. Cor p p.adj
1 .Intercept. (Intercept) 4.3128538 29.0261330 24.152830 1.20176943 0.29147190 0.16116777 0.199960
2 .Intercept. GrpB -0.3824975 -1.1599283 15.511347 -0.07477934 -0.01813665 0.94741052 1.000000
3 .Intercept. GrpC -1.7121985 -5.5300048 16.061350 -0.34430510 -0.08350625 0.68066387 1.000000
4 .Intercept. X2 9.5446012 151.9594953 51.196709 2.96814969 0.71988204 0.00579884 0.009998
5 X1 X1 0.2882803 2.0179621 2.658795 0.75897611 0.17889239 0.18636273 1.000000
6 X1 GrpB:X1 1.1367411 3.9785937 2.073856 1.91845180 0.45218343 0.03879224 1.000000
7 X1 GrpC:X1 0.1228905 0.3970308 1.779620 0.22309867 0.05258486 0.78544291 1.000000plot the results:
plot(res, p.values = "adjusted")look at the combined results:
summary(combine_contrasts(res))
Model Coeff Stat nMods S p
1 .Intercept..0 (Intercept) maxT 1 29.026133 0.08198360
2 .Intercept..1 Grp maxT 2 -1.159928 0.72145571
3 .Intercept..2 X2 maxT 1 151.959495 0.00359928
4 X1.0 X1 maxT 1 2.017962 0.09078184
5 X1.1 Grp:X1 maxT 2 3.978594 0.01959608Hemerik, J., Goeman, J. J., & Finos, L. (2020). Robust testing in generalized linear models by sign flipping score contributions. Journal of the Royal Statistical Society Series B: Statistical Methodology, 82(3), 841-864. https://doi.org/10.1111/rssb.12369
De Santis, R., Goeman, J. J., Hemerik, J., Davenport, S., & Finos, L. (2022). Inference in generalized linear models with robustness to misspecified variances. arXiv preprint arXiv:2209.13918. https://arxiv.org/abs/2209.13918
Girardi, P., Vesely, A., Lakens, D., Altoè, G., Pastore, M., Calcagnì, A., & Finos, L. (2024). Post-selection Inference in Multiverse Analysis (PIMA): An Inferential Framework Based on the Sign Flipping Score Test. Psychometrika, 1-27. https://doi.org/10.1007/s11336-024-09973-6
If you encounter a bug, please file a reprex (minimal reproducible example) on github.
