[R-sig-ME] Question on semiparametric bootstrap in lme4

[Ben Bolker]

  Yes, that does look like a bug.  Thanks!  (Interesting that the
bootstrap std dev is about half the size of the parametric std error ...
qqmath(fm01) shows that the distribution of residuals is indeed thin-tailed.

  Fixed on Github (testing now ...)

  Ben Bolker

On 14-07-22 12:11 AM, Mark Lai wrote:
> Hi,
> 
> I have a question on the semiparametric bootstrap result for lme4.
> Specifically, the bootstrap standard deviation for the fixed effect is
> essentially zero. Here is an example:
> 
>> require(lme4)
>> fm01 <- lmer(Yield ~ 1|Batch, Dyestuff)
>> set.seed(1)
>> require(boot)
>> boo01_sp <- bootMer(fm01, fixef, nsim = 100, use.u = TRUE,
> +                  type = "semiparametric")
>> boo01_sp
> 
> 
> Call:
> bootMer(x = fm01, FUN = fixef, nsim = 100, use.u = TRUE, type =
> "semiparametric")
> 
> 
> Bootstrap Statistics :
>     original       bias     std. error
> t1*   1527.5 9.094947e-13 1.392467e-12
> 
> Then I took a look on the source code for the function `bootMer`, and
> found the relevant code:
> 
> if (type == "parametric") {
>         ss <- simulate(x, nsim = nsim, use.u = use.u, na.action =
> na.exclude)
>     }
>     else {
>         if (use.u) {
>             if (isGLMM(x))
>                 warning("semiparametric bootstrapping is questionable
> for GLMMs")
>             ss <- replicate(nsim, fitted(x) + sample(residuals(x,
>                 "response")), simplify = FALSE)
>         }
>         else {
>             stop("semiparametric bootstrapping with use.u=FALSE not yet
> implemented")
>         }
>     }
> 
> I notice that the semiparametric method is using sampling without
> replacement (i.e., `sample(residuals(x, "response"))`, which is
> different from what I learned about bootstrap. Should the `replace =
> TRUE` argument be added?
> 
> Mark
> 
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models