Question

I’m trying to compute Wald tests for interaction terms in a generalized additive model (gam in mgcv package) using R. The documentation for regTermTest says that the function takes a model with coef and vcov methods, but putting a gam model in doesn’t seem to work. How can I get this method to work? If it can’t be done, is there another simple way to compute the same values in R?

lm model works

model <- lm(sysbp~pmper10*se_race4,data=mesa)
regTermTest(model,"pmper10:se_race4")
# Wald test for pmper10:se_race4
# in lm(formula = sysbp ~ pmper10 * se_race4, data = mesa)
# F =  3.940545  on  3  and  43621  df: p= 0.0080253 

but gam model does not

model <- gam(sysbp~pmper10*se_race4,data=mesa)
regTermTest(model,"pmper10:se_race4")
# Error in solve.default(V) : 'a' is 0-diml

Answer 1

This will work if you use the library gam but not mgcv.

The regTermTest function relies on extracting the 'assign' attribute from the model matrix for a given model.


aa <- attr(model.matrix(model), "assign")[okbeta]


This object aa is used to construct the indices when creating V

A gam object from the package mgcv does not have this attribute. Using the example from survey::regTermTest

data(esoph)
 model1 <- glm(cbind(ncases, ncontrols) ~ agegp + tobgp * 
   alcgp, data = esoph, family = binomial())


regTermTest(model1,"tobgp")

If we use the package gam

library(gam)
 model2 <- gam(cbind(ncases, ncontrols) ~ agegp + tobgp * 
   alcgp, data = esoph, family = binomial())

 regTermTest(model2,"tobgp")

Gives the output

Wald test for tobgp
   in gam(formula = cbind(ncases, ncontrols) ~ agegp + tobgp * alcgp, 
     family = binomial(), data = esoph)
  F =  3.961947  on  3  and  67  df: p= 0.011609 

While

library(mgcv)

model3 <- gam(cbind(ncases, ncontrols) ~ agegp + tobgp * 
   alcgp, data = esoph, family = binomial())


 regTermTest(model3,"tobgp")

Gives the error reported by the OP.