Skip to content

priors for multivariate mixed model in MCMCglmm with random intercepts and slopes

1 message · John Morrongiello

#
Hi Malcolm thanks for the suggestions regarding the priors. Changing the 7
to an 8 for the random effect priors worked a treat, as did your
specification for uninformative priors. The model runs and results are
sensible.

In regards to your other questions about the data and model specification:
I lied a bit in outlining the level of coverage. Some individuals have the
7 observations across each of 124 time periods, others less. This is partly
because of different aged individuals and also because I ensured that I'm
only analysing data from a given time point for an individual when all 7
response variables are present. I allow Age to randomly vary across months
because the month term captures most of the extrinsic sources of variation
in the model without specifying what this is (obviously this can be
attributed when environmental terms are added to the model). It is highly
likely thathow an individual responds to a change in the environment is age
dependent, hence the random age slope on month (which just so happens to
greatly improve model performance).

Finally, I model growth as opposed to size at age as we only have
information from the otolith on increment width which is a proxy for
growth. I could do some form of back-calculation to convert increment
measurements into fish size, but this process is not perfect and can
introduce bias into the data. Also, I find model interpretation easier when
I talk about growth (e.g. conditions in year x were good for growth, warm
temperatures favour higher growth) rather than changes in fish size, which
is very much conditional on the size of fish at the previous time step, and
the range of fish that contribute to a given section of the chronology.

Cheers
John



On Sat, Jan 24, 2015 at 5:49 AM, Malcolm Fairbrother <
M.Fairbrother at bristol.ac.uk> wrote: