I have a significant R2 prediction model with a large effect size (R2=.34). I want to comment on the unique contribution of a couple of the variables but I'm unsure of how to describe the size of the unique contributions since sr2 is somewhat affected by the relationships between variables. Are there any known guidelines for commenting on the size of sr2? For example, if sr2 indicates a variable only uniquely accounts for 3% or 5% of the variance, do I note this as just a small contribution? Any ideas? I can't seem to find any conventions on small, medium, or large for this statistic.
One idea is to perform a linear regression between each independent and the dependent variable to determine the specific R2 effect between each VI.
I think the semipartial correlation (or squared) is a good measure to assess the individual contribution of each variable. But I also think that you cant find any useful benchmarks, which are universal. Even the famous benchmarks by Cohen are only rules of thumb (what he already stated himself in his publication on effect sizes...), which should not be generalized to all fields of social sciences (what it is usually done). It is more important to look at your very specific research field and your specific variables. What have other researchers IN YOUR FIELD already found to be of interest/practical significance? What do YOU think is a small, medium or large effect?
If semipartial correlations are too abstract, please look at the regression coefficients and interpret them!! How much does an increase of 1 unit in your predictor increase/decrease in your dependent variable? If you have two scales (IV and DV) from 1 to 10 and the b-coefficient is 2, then this might be a huge effect, if you increase your IV by 1 unit, your DV changes by 2 units. But if your DV scale is from 1 to 100 (instead of 1 to 10), then a b-weight of 2 might be considered as small effect. So, get the intepretations of your variables clear to answer the question, there cant be an overall correct answer to your question!
I am going to suggest lasso or gradient boosting. -Look at what I did in this paper to find variable importance with these two methods. If you have any thoughts or questions, please let me know. Best wishes, D. Booth
- Badges
- Science method
- Similar topics
- Teaching