[R-sig-ME] glmmTMB: testing for temporal variation in effect of fixed predictor on response variable

[Brenna Levine]

Dr. Bolker,

I'm hoping that I might be able to bother you with a quick question. I am
trying to *test whether there is significant temporal variation in the
effect of a fixed predictor on my response variable *with a model in which
year is specified as a random effect (I have 20 years of data). Currently,
I am doing this by fitting an interaction between the fixed effect and the
random effect of year as* (1|year:fixed)* (per a recommendation that I saw
on RSeek.org at some point and some tips that I have read on this
list-serve), and am testing the significance of this random interaction
with a LRT (i.e., with a model lacking this interaction).

Could you tell me if (a) (1|year:fixed) is the correct way to specify this,
and (b) if not, do you have a recommendation for how I should specify this
interaction to test for temporal variation in the effect of a fixed
predictor on my response variable?

Thanks.

[[alternative HTML version deleted]]

[Ben Bolker]

Yes, to amplify slightly: suppose you have categorical fixed effects
f1, f2, f3 and continuous fixed effect x1.

The most complete random-effects model would be (1+f1+f2+f3+x1|year)
(assuming that all of the fixed effects vary among years and so it
even makes sense to estimate year-by-effect variation), but this is
very likely to be too complex to fit, especially if your categorical
predictors have more than 2 levels.

(1|year) + (1|f1:year) + (1|f2:year) + (1|f3:year) + (0+x1|year)

would be a reasonable simplification (this only fits 5 variance
parameters), but does assume that the effects vary independently.

Also note that likelihood ratio tests of variance components are
generally conservative (see details at
http://bbolker.github.io/mixedmodels-misc/glmmFAQ.html#testing-significance-of-random-effects
)
On Wed, Nov 7, 2018 at 10:53 AM Thierry Onkelinx via

Dear Brenna,
Please keep the mailing list in cc.
(1 + fixed|year) fits a random intercept and a random slope along "fixed"
for every "year". Keep in mind that you need enough data to support such a
model. See e.g.
https://www.muscardinus.be/2018/02/highly-correlated-random-effects/
Best regards,
ir. Thierry Onkelinx
Statisticus / Statistician
Vlaamse Overheid / Government of Flanders
INSTITUUT VOOR NATUUR- EN BOSONDERZOEK / RESEARCH INSTITUTE FOR NATURE AND
FOREST
Team Biometrie & Kwaliteitszorg / Team Biometrics & Quality Assurance
Havenlaan 88 bus 73, 1000 Brussel
www.inbo.be
///////////////////////////////////////////////////////////////////////////////////////////
To call in the statistician after the experiment is done may be no more
than asking him to perform a post-mortem examination: he may be able to say
what the experiment died of. ~ Sir Ronald Aylmer Fisher
The plural of anecdote is not data. ~ Roger Brinner
The combination of some data and an aching desire for an answer does not
ensure that a reasonable answer can be extracted from a given body of data.
~ John Tukey
///////////////////////////////////////////////////////////////////////////////////////////
<https://www.inbo.be>

Hi Thierry,
Thanks for your detailed response. I have one fixed effect that is not
discrete. How should I fit this interaction in this case?
Thanks!
Brenna

Dear Brenna,
Adding a random effect (1|year:fixed) makes sense, assuming that both
year and fixed are discrete. Note that adding this allows for a very
liberal temporal variantion by the fixed effect. Each level of the
interaction is independent from all other levels.
Best regards,
ir. Thierry Onkelinx
Statisticus / Statistician
Vlaamse Overheid / Government of Flanders
INSTITUUT VOOR NATUUR- EN BOSONDERZOEK / RESEARCH INSTITUTE FOR NATURE
AND FOREST
Team Biometrie & Kwaliteitszorg / Team Biometrics & Quality Assurance
Havenlaan 88 bus 73, 1000 Brussel
www.inbo.be
///////////////////////////////////////////////////////////////////////////////////////////
To call in the statistician after the experiment is done may be no more
than asking him to perform a post-mortem examination: he may be able to say
what the experiment died of. ~ Sir Ronald Aylmer Fisher
The plural of anecdote is not data. ~ Roger Brinner
The combination of some data and an aching desire for an answer does not
ensure that a reasonable answer can be extracted from a given body of data.
~ John Tukey
///////////////////////////////////////////////////////////////////////////////////////////
<https://www.inbo.be>
Op wo 7 nov. 2018 om 00:16 schreef Brenna Levine <

Post by Brenna Levine
Dr. Bolker,
I'm hoping that I might be able to bother you with a quick question. I am
trying to *test whether there is significant temporal variation in the
effect of a fixed predictor on my response variable *with a model in which
year is specified as a random effect (I have 20 years of data). Currently,
I am doing this by fitting an interaction between the fixed effect and the
random effect of year as* (1|year:fixed)* (per a recommendation that I saw
on RSeek.org at some point and some tips that I have read on this
list-serve), and am testing the significance of this random interaction
with a LRT (i.e., with a model lacking this interaction).
Could you tell me if (a) (1|year:fixed) is the correct way to specify this,
and (b) if not, do you have a recommendation for how I should specify this
interaction to test for temporal variation in the effect of a fixed
predictor on my response variable?
Thanks.
[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models

[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models