Timeline for Why are there large discrepancies between Wald and bootstrapped confidence intervals for parameters of a lmer model in R?
Current License: CC BY-SA 4.0
9 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jan 17, 2020 at 20:14 | comment | added | Oliver | Note that the above mentioned references are not really what i'd called simple material. I do however agree that one should be careful with the standard wald-test, as it may become very misleading indeed, even in situations where one would feel it shouldn't be. The approximations quoted might be good alternatives to bootstrapping. | |
| Jan 17, 2020 at 20:11 | comment | added | Oliver | @AdamO various references are available, with alternative. 1, 2 and 3 would probably be some often cited origins. The latter is only concerned with single parameter test, while the first two concerns multivariate test of fixed effects. These are implemented in the lmerTest package. Also bootstrapping becomes overly slow if the number of observations becomes large. | |
| Jan 16, 2020 at 19:29 | comment | added | EdM |
@TKraft I think what you are looking for is the confint.merMod() function in the lme4 package.
|
|
| Jan 16, 2020 at 19:04 | comment | added | AdamO | @BenBolker hmm, can you link a reference on denominator dfs? Curious to patch up any misunderstanding on my part. | |
| Jan 16, 2020 at 18:59 | history | edited | AdamO | CC BY-SA 4.0 |
added 2391 characters in body
|
| Jan 16, 2020 at 18:31 | comment | added | TKraft | In relation to this last comment: is there an obvious way to get bootstrapped or profile SEs rather than Wald SEs reported in summary? It's not exactly clear to me why Doug Bates did not want the Wald SEs omitted for the same reason as the p-values. | |
| Jan 16, 2020 at 1:57 | comment | added | Ben Bolker |
And ... summary.merMod omits p-values because denominator dfs are hard, not because Wald intervals are bad. The standard errors quoted in summary are exactly Wald-type estimates ...
|
|
| Jan 16, 2020 at 1:55 | comment | added | Ben Bolker | I don't really agree with this. It's true that Wald statistics are inferior to profile and parametric bootstrap statistics, but they're orders of magnitude faster, and in many cases they're good enough. (Typically better for fixed effects than for random-effects parameters [variances & covariances]; for REs they're better on log than on linear scales ... all other things being equal, they're better for larger data sets, which is exactly when you need speed) The trick is knowing when. | |
| Jan 15, 2020 at 21:25 | history | answered | AdamO | CC BY-SA 4.0 |