How can I run mice with a GAMM?

[kjetil-uis]

I try to impute a GAMM model but can not figure out how.
If I use predictorMatrix I can set the class variable for random effect, but no splines or variable crossing.
If I use `formulas can set splines and variable crossings but I am not able to set the random/class variable.

I tried this:

imp  <- mice(nhanes,
                       formulas = list( 
                            #bmi = bmi ~ s(chl) + (1 + age | hyp)
                            bmi = bmi ~ poly(chl) + (1 + age | hyp),
                            chl = chl ~ age + bmi + hyp,
                            hyp = hyp ~ bmi + chl*poly(age)
                       ), 
                       #method = "norm.predict", # runs when s() is not present 
                       #method = "2l.norm",         # Error: No class variable 
                       #method = "2l.pan",           # Error: No class variable 
                       #method = "2l.lmer",          # Error: lmer does not run 
                       #method = '2l.pmm',          # Error in str2lang(x) : <text>:1:27: unexpected ')'
                       #method = "2l.glm.norm",   # Error in factor(x[, clust], labels = 1:n.class) :invalid 'labels'; length 2 should be 1 or 0
                       print = FALSE
                 )

When using method = ´norm.predict´ it runs with poly() instead of s() - but does it account for the random effect?
If I use method = '2l.method' it does not run (Error is commented in the code above).

What is wrong?

[stefvanbuuren]

You are trying to specify imputation models that are more complex than mice() can handle. Generalised additive mixed models (GAMM) are not supported. The best advice that I can offer at the moment is to simplify your model in one of two ways:

1. run a flat model (no class variable) with data transforms. This may be sensible if the ICC is low (so clustering is not that important) and data transforms are critical;
2. run a linear multilevel model (no transforms) if ICC is high, and use an imputation method that is relatively robust against violations of linearity, e.g., 2l.pmm().

Your choice would depend on the data at hand.

[kjetil-uis]

Thank you for a quick response.
I have longitudinal data with missing in only the dependent variable, so I think alternative 2 would be most appropriate.

To include non-linearity is not critical, but I would like to have random slope. Is it possible? By using formulas notation like (1 + age | hyp) or other ways?

[stefvanbuuren]

You should be able to find that in the documentation.

[stefvanbuuren]

Addition: This is work in progress.

See f5d5c99 line 624:

The random slope version fml <- list(ReadDis ~ ReadAchiev + (1 + ReadAchieve|ID)) does not yet work due to improper formula processing by mice()

But having a random slope and intercept is possible by predictorMatrix route.

[kjetil-uis]

It will be great when you can use formulas - notation.

I found that for 2l.panyou can use 4 in the predictorMatrix to get a random slope.

Also I was also able to run formulas = list( y ~ poly(x,2) + (1 + x | ID)) with method = "jomoImpute". However I was not able to plot the convergence -plot and it took 10min vs 0.1 min when I used mitml:jomoImpute directly on the same data - so I am a bit uncertain how much trust I should put in that result.

Thank you again and keep up the good work.