predict_response() vs ggeffect()

[YyLu5]

I'm confused by the difference between these two functions. I've read that the predict_response is a "wrapper" function for ggeffect(). I used both the ggeffect(model, term) and predict_response (model, term) to compare, the results are different, they differ by small values and the SE is slightly larger in the ggeffect() function. Later when my model contains a random effect, I'm able to use ggeffect(model,term,vcov_fun="CR3"), while use predict_response(model,term,vcov_fun="CR3") caused an error "Error in vcovCR.default(obj = list(obj = list(par = c(beta = 5.34256429246318, : argument "cluster" is missing, with no default"

So I'm not sure which one is correct to use. Thank you!

[strengejacke]

The main functions are:

• ggpredict() (calls predict())
• ggeffect() (calls effects::Effect())
• ggemmeans() (calls emmeans::emmeans())
• ggaverage() (calls marginaleffects::avg_predictions()`)

For all functions, predictions are made for a reference or data grid, which is provided via the terms argument.

Thus, the above functions all return more or less similar results. Their main difference is how non-focal terms (those model predictors that are not specified in the terms argument) are handled. ggeffect() and ggemmeans() have similar approaches and usually return the same predictions. Therefore, predict_response() is a wrapper around the four above-mentioned functions, and the margin argument is responsible for which underlying function is called (see description for instance here: https://strengejacke.github.io/ggeffects/articles/ggeffects.html). Indeed, no margin option calls ggeffect() (so this function is excluded from the predict_response()-wrapper), because results are usually the same as for ggemmeans() (which is predict_response(margin = "marginalmeans")).

That is the summary of predict_response() as wrapper and the single functions. As you can see here: https://strengejacke.github.io/ggeffects/reference/ggpredict.html, ggeffect() has no vcov_* argument. This is only supported by ggpredict() or ggaverage() (or: predict_response() and predict_response(margin = "empirical")). ggeffect() ignores the vcov_fun argument and therefore throws no error.

Coming to your example that fails: The "CR3" option requires the cluster argument to be specified, then it works. See examples here: https://strengejacke.github.io/ggeffects/articles/practical_robustestimation.html

Hope that helps?

[strengejacke]

In a recent commit, I added warnings when non-supported arguments are used in function calls.

[YyLu5]

sounds great, this is very helpful, thank you!