1

I am running a mixed model in R and would like to perform forward selection using the step function. However, when I set the direction to forward

model = lmer(y ~ .)
step(model,direction='forward',scope=biggest)

In the output, it states and prints the tables for "backward reduced fixed-effect table" and "backward reduced random-effect table". Why isn't forward selection working?

rtob
  • 31
  • 1
  • 7
  • 1
    voting to close as off-topic (about programming, not statistical aspects). I'm guessing you must be using this method: `?step.lmerModLmerTest`, which only does backward stepwise regression (and silently ignores unused arguments ...) – Ben Bolker Jun 25 '19 at 12:35

1 Answers1

1

The most important point here is that forward stepwise selection doesn't work well at all. See this page, among many others on this site, for why this is a poor strategy. In some circumstances backward stepwise could be considered, but even then the coefficient estimates will be biased and p-values will be unreliable.

In any event, the step() function uses the AIC to compare among models. The default model fitting method for lmer() is REML, which cannot be used for comparisons among mixed models with different fixed effects, as explained here and here. So I'm somewhat surprised that you even got results for a "backward reduced fixed-effect table" and it probably shouldn't be trusted.

In this case, however, this problem with step might have saved you from the potentially serious statistical problems that would come from relying on forward stepwise selection.

EdM
  • 57,766
  • 7
  • 66
  • 187