conducting GAM-GEE in gamm4 R package? - r

Question

I am trying to analyze some visual transect data of organisms to generate a habitat distribution model. Once organisms are sighted, they are followed as point data is collected at a given time interval. Because of the autocorrelation among these "follows," I wish to utilize a GAM-GEE approach similar to that of Pirotta et al. 2011, using packages 'yags' and 'splines' (http://www.int-res.com/abstracts/meps/v436/p257-272/). Their R scripts are shown here (http://www.int-res.com/articles/suppl/m436p257_supp/m436p257_supp1-code.r). I have used this code with limited success and multiple issues of models failing to converge.
Below is the structure of my data:
> str(dat2)
'data.frame': 10792 obs. of 4 variables:
$ dist_slag : num 26475 26340 25886 25400 24934 ...
$ Depth : num -10.1 -10.5 -16.6 -22.2 -29.7 ...
$ dolphin_presence: int 0 0 0 0 0 0 0 0 0 0 ...
$ block : int 1 1 1 1 1 1 1 1 1 1 ...
> head(dat2)
dist_slag Depth dolphin_presence block
1 26475.47 -10.0934 0 1
2 26340.47 -10.4870 0 1
3 25886.33 -16.5752 0 1
4 25399.88 -22.2474 0 1
5 24934.29 -29.6797 0 1
6 24519.90 -26.2370 0 1
Here is the summary of my block variable (indicating the number of groups for which autocorrelation exists within each block
> summary(dat2$block)
Min. 1st Qu. Median Mean 3rd Qu. Max.
1.00 39.00 76.00 73.52 111.00 148.00
However, I would like to use the package 'gamm4', as I am more familiar with Professor Simon Wood's packages and functions, and it appears gamm4 might be the most appropriate. It is important to note that the models have a binary response (organism presence of absence along a transect), and thus why I think gamm4 is more appropriate than gamm. In the gamm help, it provides the following example for autocorrelation within factors:
## more complicated autocorrelation example - AR errors
## only within groups defined by `fac'
e <- rnorm(n,0,sig)
for (i in 2:n) e[i] <- 0.6*e[i-1]*(fac[i-1]==fac[i]) + e[i]
y <- f + e
b <- gamm(y~s(x,k=20),correlation=corAR1(form=~1|fac))
Following this example, the following is the code I used for my dataset
b <- gamm4(dolphin_presence~s(dist_slag)+s(Depth),random=(form=~1|block), family=binomial(),data=dat)
However, by examining the output (summary(b$gam)) and specifically summary(b$mer)), I am either unsure of how to interpret the results, or do not believe that the autocorrelation within the group is being taken into consideration.
> summary(b$gam)
Family: binomial
Link function: logit
Formula:
dolphin_presence ~ s(dist_slag) + s(Depth)
Parametric coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -13.968 5.145 -2.715 0.00663 **
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Approximate significance of smooth terms:
edf Ref.df Chi.sq p-value
s(dist_slag) 4.943 4.943 70.67 6.85e-14 ***
s(Depth) 6.869 6.869 115.59 < 2e-16 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
R-sq.(adj) = 0.317glmer.ML score = 10504 Scale est. = 1 n = 10792
>
> summary(b$mer)
Generalized linear mixed model fit by the Laplace approximation
AIC BIC logLik deviance
10514 10551 -5252 10504
Random effects:
Groups Name Variance Std.Dev.
Xr s(dist_slag) 1611344 1269.39
Xr.0 s(Depth) 98622 314.04
Number of obs: 10792, groups: Xr, 8; Xr.0, 8
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
X(Intercept) -13.968 5.145 -2.715 0.00663 **
Xs(dist_slag)Fx1 -35.871 33.944 -1.057 0.29063
Xs(Depth)Fx1 3.971 3.740 1.062 0.28823
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Correlation of Fixed Effects:
X(Int) X(_)F1
Xs(dst_s)F1 0.654
Xs(Dpth)Fx1 -0.030 0.000
>
How do I ensure that autocorrelation is indeed being accounted for within each unique value of the "block" variable? What is the simplest way to interpret the output for "summary(b$mer)"?
The results do differ from a normal gam (package mgcv) using the same variables and parameters without the "correlation=..." term, indicating that something different is occurring.
However, when I use a different variable for the correlation term (season), I get the SAME output:
> dat2 <- data.frame(dist_slag = dat$dist_slag, Depth = dat$Depth, dolphin_presence = dat$dolphin_presence,
+ block = dat$block, season=dat$season)
> head(dat2)
dist_slag Depth dolphin_presence block season
1 26475.47 -10.0934 0 1 F
2 26340.47 -10.4870 0 1 F
3 25886.33 -16.5752 0 1 F
4 25399.88 -22.2474 0 1 F
5 24934.29 -29.6797 0 1 F
6 24519.90 -26.2370 0 1 F
> summary(dat2$season)
F S
3224 7568
> b <- gamm4(dolphin_presence~s(dist_slag)+s(Depth),correlation=corAR1(1, form=~1 | season), family=binomial(),data=dat2)
> summary(b$gam)
Family: binomial
Link function: logit
Formula:
dolphin_presence ~ s(dist_slag) + s(Depth)
Parametric coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -13.968 5.145 -2.715 0.00663 **
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Approximate significance of smooth terms:
edf Ref.df Chi.sq p-value
s(dist_slag) 4.943 4.943 70.67 6.85e-14 ***
s(Depth) 6.869 6.869 115.59 < 2e-16 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
R-sq.(adj) = 0.317glmer.ML score = 10504 Scale est. = 1 n = 10792
> summary(b$mer)
Generalized linear mixed model fit by the Laplace approximation
AIC BIC logLik deviance
10514 10551 -5252 10504
Random effects:
Groups Name Variance Std.Dev.
Xr s(dist_slag) 1611344 1269.39
Xr.0 s(Depth) 98622 314.04
Number of obs: 10792, groups: Xr, 8; Xr.0, 8
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
X(Intercept) -13.968 5.145 -2.715 0.00663 **
Xs(dist_slag)Fx1 -35.871 33.944 -1.057 0.29063
Xs(Depth)Fx1 3.971 3.740 1.062 0.28823
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Correlation of Fixed Effects:
X(Int) X(_)F1
Xs(dst_s)F1 0.654
Xs(Dpth)Fx1 -0.030 0.000
>
I just want to make sure it is correctly allowing for correlation within each value for the "block" variable. How do I formulate the model to say that autocorrelation can exist within each single value for block, but assume independence among blocks?
On another note, I am also receiving the following warning message after model completion for larger models (with many more variables than 2):
Warning message:
In mer_finalize(ans) : false convergence (8)

Answer 1

gamm4 is built on top of lme4, which does not allow for a correlation parameter (in contrast to the nlme, package, which underlies mgcv::gamm). mgcv::gamm does handle binary data, although it uses PQL, which is generally less accurate than Laplace/GHQ approximations as in gamm4/lme4. It is unfortunate (!!) that you're not getting a warning telling you that the correlation argument is being ignored (when I try a simple example using a correlation argument with lme4, I do get a warning, but it's possible that the extra argument is getting swallowed somewhere inside gamm4).
Your desired autocorrelation structure ("autocorrelation can exist within each single value for block, but assume independence among blocks") is exactly the way correlation structures are coded in nlme (and hence in mgcv::gamm).
I would use mcgv::gamm, and would suggest that if at all possible you try it out on some simulated data with known structure (or use the data set provided in the supplementary material above and see if you can reproduce their qualitative conclusions with your alternative approach).
StackOverflow is nice, but there is probably more mixed model expertise at r-sig-mixed-models#r-project.org