0

does anyone know how to bootstrap confidence intervals for p-values of an lmer object (lmer and glmer as well) from lme4? Maybe via bootMer? I am doing planned/ one sided contrasts but all I can find are intervals for pairwise contrasts.

Am glad about any help. Thanks!

StupidWolf
  • 45,075
  • 17
  • 40
  • 72
a.henrietty
  • 55
  • 1
  • 1
  • 8
  • There are two things that confuse me in your question: First, do you litterally want to get a confidence interval for a _p_-value? That would be highly exceptional and I doubt that it is useful. Or did you mean a confidence interval for a specific significance level (alpha, **not** _p_)? If so, try `confint(YourLmerModel, method="boot")`. Second, when you say pairwise, do you mean _two-sided_ in the sense that the coefficient of interest may be positive or negative under the H1? If yes, for a one-sided test, just double your 1-alpha, i.e. set `level=.90` instead of the default .95. – benimwolfspelz Jun 29 '20 at 07:29
  • Thanks for your answer. I will explain my procedure to you: First, I compared models in a likelihood ratio test. In the second step, I took the final model, set contrasts (0 vs. 1) and compared the parameter levels via summary(). With glmer() I just proceeded, with lmer() i used satterthwaites approximation. By one sided, I thus meant: directed comparison, so not like tukey`s HSD test. Now, I would like to check for the validity of the comparison via bootstrap confidence intervals (So i guess I am comparing beta-values?). I am kind of confused in this field, sorry. So, alpha * 2 is needed? – a.henrietty Jun 30 '20 at 09:10
  • So you want the bootstrap confidence interval for a specific regression parameter _b_ and use it to test a one-sided / directional hypothesis (e.g. that _b_ > 0 or _b_ < 0), right? If so, `confint` should be good, see `help(confint.merMod)`. Make sure to use much more than the default 500 bootstrap-samples, rather 5.000 or so. This may take some hours to run an average laptop. Now, whether or not you can do .90 or have to use .95 depends on your field. While 90%-CIs are formally correct to test one-sided hypotheses with alpha = .95, e.g. in psychology, 90%-CIs are generally not accepted. – benimwolfspelz Jun 30 '20 at 10:30

0 Answers0