Hello,
We’re exploring temporal change in summer range selection in radio-collared
elk during a 3 decade period. Elk are marked on the wintering grounds and
tracked back to one of 5 summer ranges. There are 180 individuals; most (n
= 121) individuals were monitored >=2 years, resulting in a total of 546
summer range observations from the 180 elk. We’ve run multinomial
regression models with year as a predictor in packages ‘mlogit’ and ‘nnet’
using only the last summer range choice for each individual (n = 180):
m2.year<-multinom(summercategory~n.year, data=elk, Hess=T)
# weights: 15 (8 variable)
initial value 289.698824
iter 10 value 237.592576
final value 237.060236
converged
> summary(m2.year)
Call:
multinom(formula = summercategory ~ n.year, data = elk, Hess = T)
Coefficients:
(Intercept) n.year
GTNPCEB 1.1306086 0.008673583
GTNPS -2.7164924 0.128711359
TW -0.4452291 0.007482690
YNP 0.8210785 -0.042043091
Std. Errors:
(Intercept) n.year
GTNPCEB 0.3936527 0.01869625
GTNPS 0.9689580 0.03476211
TW 0.5450637 0.02550461
YNP 0.4229157 0.02325437
Residual Deviance: 474.1205
AIC: 490.1205
We’ve also run a mixed-effect model in nnet using the individual as a
random effect to account for the longitudinal data of repeated observations
through time for some individuals with all 546 observations.
> mn.year<-multinom(summercategory~n.year + (1|elkid), data=elk, Hess=T)
# weights: 20 (12 variable)
initial value 907.722983
iter 10 value 734.516684
final value 733.300162
converged
> summary(mn.year)
Call:
multinom(formula = summercategory ~ n.year + (1 | elkid), data = elk,
Hess = T)
Coefficients:
(Intercept) n.year 1 | elkidTRUE
GTNPCEB 0.6748545 0.005700504 0.6748545
GTNPS -1.0455490 0.115835869 -1.0455490
TW -0.1948539 -0.002288967 -0.1948539
YNP 0.4134208 -0.037711158 0.4134208
Std. Errors:
(Intercept) n.year 1 | elkidTRUE
GTNPCEB 0.1338889 0.01291928 0.1338889
GTNPS 0.2444979 0.01964817 0.2444979
TW 0.1889981 0.01843526 0.1889981
YNP 0.1478354 0.01548316 0.1478354
Residual Deviance: 1466.6
AIC: 1482.6
We get comparable results for the effect of year from all three models
we’ve run. What we’re concerned about is the fact that 59 individuals only
have a single observation, and 43 only have 2. It seems like the estimation
of weights wouldn’t work because of this and we may be asking too much
trying to use a mixed-effect model with this data?
Thanks,
Jeff
[[alternative HTML version deleted]]