Partial effect plot: Which options do I have?


I have a multiple regression model, and I would like to do an effect plot with confidence bands regarding one of the predictors. One solution I see is to create a fake-dataset, where all other predictors are constant. However, this produces pretty large cofindece bands, since they are based on these specific covariate values and not on the entire dataset.

I also used a modified regression model with only the predictor of interest and plotted this, but this does not make much sense, since the neglection of these covariates produces bias.

Do I have any other options? What do the people usually do? Should't it be possible to use the entire dataset to calculate confidence intervals, since this actually happens if significance values for regression coefficients are calculated?



Less is more. Stay pure. Stay poor.
Yeah, this definitely seems plausible. Can't say I have done it. Yes, you are going to want to control for covariates. I guess I usually just report the partial effect without it.

It may help others if you mention what programs you are using or willing to use.


Less is more. Stay pure. Stay poor.
What do you want the graph to look like in particular. I guess, going back and looking at your description I am unsure.
Hi hlsmith, I try to explain: I have an outcome Y and a predictor I am intersted in, let's say YEAR. Now I want to plot Y~YEAR with confidence bands. However, I have some covariates in order to prevent bias. E.g., I have the categorical covariate "site", since in different years the intensity of data collection differs between the sites, and I want to correct for this.

If I now perform a prediction plot using predict() in R, I can create a fake dataset, but I have to choose a certain value for the covariate. Thus, I can select site = X, but now, confidence intervals are only based on these data from site X and not on the entire dataset, and thus are relatively large compared to the "overall standard error" connected to the predictor YEAR, which is calculated using all levels of "site". Thus, I want to calculate confidence intervals based on all collected data. I am aware of the fact that in this case, I can't anylonger plot Y on an absolute scale (since Y depends on the covariate "site"), but on a relative scale (e.g. starting at 1).

I hope it is clearer now.