How do I specify random effects using gamm models in R?

Question

I am using gamm models in mgcv package to analyse how specific diversity measures eg. Shannon vary over time and with environmental variables eg. temperature.

I have the initial model so far to analyse time series:

modf<-gamm(y~ as.factor(year) + s(doy,bs='cc',k=kdy),method=mth,correlation=tcor,data=d,
           control=ctrl,random=NULL,gamma=1)

I want to include temperature as random effect and thought of doing something like:

modf<-gamm(y~ as.factor(year) + s(doy,bs='cc',k=kdy), + s(temp,bs="re"),method=mth,
           correlation=tcor,data=d,control=ctrl,gamma=1)

However, so far I have only seen this for gam not gamm. Does it still work this way?

An example of type of data structure:

• $ total_abundance: num 6364161 1929775 7057036 1266342 3981198 ...
• $ shannon : num 1.87 2.08 1.95 1.84 2.06 ...
• $ turnover : num 0.613 0.475 0.525 0.556 0.429 ...
• $ year : int 1985 1986 1987 1987 1987 1988 1989 1989 1991
• $ month : int 8 12 3 7 8 5 1 8 1 9 ...
• $ day : int 20 8 15 6 17 9 16 29 14 4 ...
• $ temp : num 25.5 9.87 4.8 19.72 26.03 ...
• $ doy : num 232 342 74 187 229 130 16 241 14 247 ...

where doy is 'day of year' and accounts for seasonality

Thanks

Answer 1

What you seem to be wanting to do doesn't make any sense unless you want the linear effect of temp to vary within the levels of a grouping factor.

Typically, you specify this random slope using random as

list(group = ~ x)

where group is a factor grouping variable and x is your temp .

What you are asking for with temp = ~ 1 is a random intercept for the unique values of temp , which is probably asking too much of the data.

The equivalent of s(time, bs = "re") requires you to remove the intercept from the random formula:

 list(group = ~ x - 1)

but you still need a group variable.

If you just want to control for temp , add it as a linear parametric effect ( + temp ) or a smooth effect ( + s(temp) ) in the gamm() model formula.