[R-sig-ME] glmmADMB definition of levels for random component

[Ben Bolker]

Juan José Santos Blanco <juan.santos at ...> writes:

>  The trouble I have regarding the option (a) is that I'm assuming
> that sampled trips come from the overall trip population (T), and
> from my point of view this is not true; Following the sampling
> scheme, trip "t" is randomly drawn from the trip population "T_i"
> from the year "i" .  Many factors (such as shift in fishing
> behaviour and techniques, changes in fish populations, yearly
> quotas...),makes me think that trip population can not be assumed to
> be the same across years. I expect that using the (a) approach
> produce an over-inflation of the random effect sigmas.

> >option (d):
> 
> >  ndiscarded~year+ [other fixed effects] + (1|year:trip)+(1|year:trip:haul)
> 
> >which expands the trip/haul nesting (which would otherwise expand
> >to trip+trip:haul) to include an interaction with year.  If this gives
> >you trouble you could also do this even more explicitly by adding
> >a 'yeartrip' variable to your data set which was defined as 
> >interaction(year,trip) and similarly for 'yeartriphaul' ...


  Actually, I believe that you will get exactly the same answer from
these two approaches.  What you're assuming in either case is that
the distribution *within* years of trips, or the distribution within
trips of hauls, is the same across years (or hauls).  Specifically,
you're assuming something like

  y_ijk ~ beta_i + eps_{ij} + gamma_{ijk}

where i=year, j=trip, k=haul ...

  You could try out the two different approaches (possibly on
 a subset of your data, or a simpler/made-up example) and
confirm for yourself that they do indeed give the same answers ...

   Ben Bolker