anova doesnt flag different sample sizes and instead returns p value

[ngalanter]

The anova function for regress models doesn't report an error if there are different sample sizes between the full and reduced models, which can occur if a variable in the full and not the reduced model has missingness. Instead a (very low) p-value is returned. This is different than the behavior of the anova function and lmtest::lrtest for glm models, which both report errors.

Here is an example:

library(rigr)
library(lmtest)

set.seed(236)

x1 <- as.factor(rep(1:4,25))

x1[1:10] <- NA

x2 <- rnorm(100)

y <- rbinom(size = 1, n = 100, p = 1/(1+exp(0.5*x2)))

dat <- data.frame(x1=x1, x2=x2, y=y)

mod_full_regress <- regress("odds", y ~ x1+x2,data = dat)
mod_reduced_regress <- regress("odds", y ~ x2, data = dat)

mod_full_regress

anova(mod_reduced_regress,mod_full_regress,test = "Wald")
anova(mod_reduced_regress,mod_full_regress,test = "LRT")

mod_full_glm <- glm(y ~ x1+x2, data = dat, family = binomial)
mod_reduced_glm <- glm(y ~ x2, data = dat, family = binomial)

anova(mod_reduced_glm,mod_full_glm,test = "LRT")

lrtest(mod_reduced_glm,mod_full_glm)

[adw96]

Thanks, @ngalanter . If you have a pull request that addresses this issue, I would can review it, but I won't be able to address it myself for some time. Thanks for understanding.