[R-sig-ME] Post hoc on glmer for specific hypotheses

[Ben Bolker]

  That's fine.  (You could always extract the values and exponentiate
them yourself ...)
 I still can't tell whether there are any remaining issues, i.e. if
type="response" worked as expected would you be happy? (Noting that
the significance tests of the contrasts would still be based on the
estimates and SDs from the log-odds scale ...)

On Wed, Mar 9, 2022 at 2:21 PM Timothy MacKenzie <fswfswt using gmail.com> wrote:
>
> Ben, type = "response" exponentiates the link-scale results. It just
> seems that it doesn't work inside the pairs(...) call. Russ can better
> speak to that.
>
> emmeans provides the capability to convert the results from the link
> scale into the odds scale (which for comparisons would be odds
> ratios).
>
> On Wed, Mar 9, 2022 at 1:14 PM Ben Bolker <bbolker using gmail.com> wrote:
> >
> >      Hmm.  Why would you want to test on the response scale?  It
> > almost always makes sense to test contrasts on the link scale, where
> > the sampling distributions of the parameters are more likely to be
> > approximately Gaussian.
> >      I haven't gone to look at your data/problem in detail, but I'm a
> > little bit surprised that `type="response"` doesn't exponentiate, i.e.
> > convert the estimated values from the log-odds to the odds scale?
> > (Converting to the probability scale is not really feasible because of
> > the way the math works out ...)
> >
> > On Wed, Mar 9, 2022 at 1:46 PM Timothy MacKenzie <fswfswt using gmail.com> wrote:
> > >
> > > Thanks, Ben! Apparently, the 7th coefficient in the glmer() output
> > > tests the first hypothesis that I'm looking for.
> > >
> > > For the second hypothesis, I was able to use emmeans::emmeans() to get
> > > what I want by running:
> > >
> > > ems <- emmeans(m2, pairwise ~ time*item_type, type = "response",infer
> > > = c(FALSE, TRUE), adjust = "tukey")[[2]]
> > > vv <- pairs(ems, simple = "each", infer = c(FALSE, TRUE), type =
> > > "response") # on log-odds scale
> > >
> > > In this set up, vv[288, ] tests the second hypothesis. But on the Log
> > > odds scale. Is there any way to convert it back to the response scale?
> > >
> > > Thanks,
> > > Tim M
> > >
> > > On Wed, Mar 9, 2022 at 12:22 PM Ben Bolker <bbolker using gmail.com> wrote:
> > > >
> > > >   One way to do this would be via `multcomp::glht`, which allows you
> > > > to specify any linear combination(s) of parameters to test.  You do
> > > > have to figure out the association between the parameters you have and
> > > > the contrasts you want. I wrote some (possibly scrutable) stuff about
> > > > that problem here:
> > > > http://bbolker.github.io/bbmisc/mgreen_contrasts.html
> > > >
> > > >   by hand:
> > > >
> > > >   assuming factor levels are {baseline, post} and {MC, prod}
> > > >
> > > > baseline_MC = intercept
> > > > post_MC = intercept + time_post
> > > > baseline_prod = intercept + gram_prod
> > > > post_prod = intercept + time_post + gram_prod + interax
> > > >
> > > > so your first contrast would be {(intercept - (intercept + time_post))
> > > > - (intercept + gram_prod - (intercept + time_post + gram_prod +
> > > > interax))
> > > >    = (-time_post -time_post - interax) = (-2*time_post - interax)
> > > >
> > > > so the contrast would be (0, -2, 0, -1) assuming that the parameters
> > > > are ordered (intercept, time_post, gram_prod, interax)
> > > >
> > > >    (1) You should definitely check my algebra; (2) there may be a
> > > > quicker (if less transparent) way to do this by setting up appropriate
> > > > contrasts from the beginning; (3) I noticed that I left the "vocab"
> > > > levels out of the analysis, but I don't think that changes anything
> > > > important.
> > > >
> > > >
> > > > On Wed, Mar 9, 2022 at 12:59 PM Timothy MacKenzie <fswfswt using gmail.com> wrote:
> > > > >
> > > > > Hello All,
> > > > >
> > > > > My glmer model below analyzes the performance of a single group of
> > > > > subjects on a test at two time points. The test has 4 item types.
> > > > >
> > > > > Data and code are below.
> > > > >
> > > > > Is there a way to test only the following two hypotheses?
> > > > >
> > > > > 1- ((Baseline multiple-choice_grammar) - (Post-test
> > > > > multiple-choice_grammar)) - (Baseline production_grammar - (Post-test
> > > > > production_grammar))
> > > > >
> > > > > 2- ((Baseline multiple-choice_vocabulary) - (Post-test
> > > > > multiple-choice_vocabulary)) - (Baseline production_vocabulary -
> > > > > (Post-test production_vocabulary))
> > > > >
> > > > > dat <- read.csv("https://raw.githubusercontent.com/fpqq/w/main/d.csv")
> > > > >
> > > > > form2 <- y ~ item_type*time + (1 | user_id)
> > > > >
> > > > > m2 <- glmer(form2, family = binomial, data = dat,
> > > > >             control =
> > > > >               glmerControl(optimizer = "bobyqa"))
> > > > >
> > > > > Sincerely,
> > > > > Tim M
> > > > >
> > > > > _______________________________________________
> > > > > R-sig-mixed-models using r-project.org mailing list
> > > > > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models