Skip to content
Prev 16487 / 20628 Next

random slopes model specification

On Wed, Apr 18, 2018 at 3:04 PM, Karista Hudelson <karistaeh at gmail.com> wrote:
Notice that there is not much variability  in your coefficients
across watersheds
(zero variability in FishLength effects because you left it out of your random
effects on purpose, about 10% variability in Spring_MST and Summer_Rain).
Normally it's not recommended to fit a model with three correlated RE terms
(intercept, spring_MST, summer_rain) to only a grouping factor with
only 5 levels -- this is
equivalent to trying to estimate a 3x3 variance-covariance matrix with
5 observations
(I'm a little surprised this worked at all).  Are all the values in
getME(fitted_model,"theta") 'reasonable',
i.e. larger than (say) 1e-4?

   There is a huge, ongoing debate about what to do in this case,
where the RE model
you want is theoretically identifiable (you do have multiple
measurements of spring_MST
and summer_RAIN in each watershed), but practically very poorly
constrained, suggestions ranging from

  "keep it maximal" (Barr et al): fit the full model, reduce when the
model is singular
  try to constrain complexity _a priori_, reduce by model selection
(AIC or p-value based)
  use Bayesian methods to add priors/regularize the problem
  substitute fixed effects instead.
Fitting separate models will be almost equivalent to fitting a single

Hg~(FishLength+Spring_MST+Summer_Rain):watershed

model -- the only difference will be whether the residual variance is
pooled/assumed to be
the same across watersheds.