I want to compare the importance or 'predictive power' of the same feature/covariate in 2 different datasets. Specifically let $[\bf{y}_1,\bf{V}_1]$ be my output & design matrix of dataset 1 & $[\bf{y}_2,\bf{V}_2]$ be the same for dataset 2.

I want to see if predictor X (which is a column of V) is more important in dataset 1 or 2.

To complicate matters:

1) X is expanded on a set of 10 basis functions, so $\bf{X}$ makes up columns 1-10 of $\bf{V}$.

2) y has other predictors besides X. Thus $\bf{V}$ has 22 columns, of which only the first 10 are $\bf{X}$

3) the output variable y is count data so I am doing Poisson regression (GLM with log link)

So far my thought is to compare the mean partial residuals (Pearson or Deviance) of $\bf{r=y-Xc}$ where c are the estimated coefficients of the columns of X. Then, I would conclude that the dataset with the lower mean residuals is the dataset where X is more important. Is this correct? One issue I see is if the mean of $y_1$ and $y_2$ are different (I would need a Poisson version of standardized coefficients?).


This is a topic which has been addressed in the context of logistic regression using dominance analysis, a useful relative importance approach.

Given that the only difference between Poisson and logit regression is the link and disribution, the results from the logit model should apply in a straightforward way to Poisson.

The only software I am familiar with to estimate such a model is in Stata currently. There is an R version which could accommodate, but I have yet to try this package out myself.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.