0

I am running into an error when I try to compute the predicted values from a model averaged object using the MuMIn package's predict.averaging. I have been assured that when the full argument is set to FALSE the function should return predicted values based off the conditional average coefficients. However, it returns an error. See example below using the cars dataset. It is very similar to my actual set up.

library(MuMIn)
options(na.action = "na.fail")
global.model <- glm(mpg ~ hp + drat + wt, 
data=mtcars)
dr <- dredge(global.model)
mod.avg <- model.avg(dr, subset = delta < 2, fit = T)
summary(mod.avg)
predict(mod.avg, se.fit = TRUE, full = FALSE)

The error indicates that full is ignored, meaning that the full model coefficients are used for the predicted values (not what I want). I have confirmed this by some simple manual checking of values. It is also evident my examining predict() output. Notice how the values jump, suggesting that a coefficient is set to zero or something. It has also been suggested that changing glm to lm will fix the issue but it does not, at least for me.

Thanks!

wraymond
  • 295
  • 1
  • 6
  • 17

1 Answers1

2

Comparing predictions from the component models to the averaged ones you can see that the "full averaged" predictions fall within the component predictions (which is as it should be).

On the other hand, the "subset averaged" coefficients produce predictions that are quite biased. This is because the effects are augmented due to ignoring the zero coefficients when calculating the mean.

# Full/subset averaged predictions 
pyfa <- predict(mod.avg, full = TRUE)
pysa <- predict(mod.avg, full = FALSE)
# Note: full=FALSE works only with se.fit=FALSE

# Predictions from component models
pycm <- do.call("cbind", lapply(get.models(mod.avg, TRUE), predict))

n <- ncol(pycm)
k <- rep(1:3, c(n, 1, 1))
lty <- c(2,1,1); lwd <- c(1,2,2); col <- c(3,1,2)
matplot(cbind(pycm, pyfa, pysa), type = "l",
        lty = lty[k], lwd = lwd[k], col = col[k],
        ylab = "predicted")
legend("topleft", legend = c("component", "full average", "subset average"),
       lty = lty, lwd = lwd, col = col)

comparison of model predictions

Kamil Bartoń
  • 1,482
  • 9
  • 10
  • I suppose I might be asking the wrong question. From what I gather from the literature, its seems that sometimes the conditional average is more appropriate to interpret over the full average. If you go that route it just means that you have to compute predicted values manually, especially if you want standard errors. Correct? – wraymond Nov 01 '18 at 18:40
  • 1
    The "conditional" average is, according to some authors, more appropriate in terms of parameter averaging, but it is clearly not suitable for prediction. Still you can get it from `predict.averaging` but without SE. – Kamil Bartoń Nov 02 '18 at 15:09