Finding and Converting the Delta R Squared

[matt starcke]

Hi all - long time reader, first time poster.

I'm working on duplicating findings from a report I was given and would love some assistance. They began with a regression - easy enough to duplicate: reg DV IV1 IV2 IV3 IV4 IV5. I get lost on the next step, however. They generated an impact score (on a 7 point scale) by converting the delta r squared such that impact(IV1) = 7 - abs(log(delta r squared)), etc...

My first thought was to use hireg DV (IV1) (IV2) (IV3) (IV4) (IV5), but the delta r squared values didn't match those provided, and even if they did I'm not sure how I'd proceed to store/log those values.

At any rate, any assistance would be greatly appreciated. Thanks.

[Joseph Luchman]

Matt,

I assume hireg is the one available from SSC (you are asked to identify the source of non-built in modules on Statalist). The hireg program assumes adding variables sequentially starting from the left moving right. Thus, in your syntax, you would have 5 regressions: 1) DV IV1, 2) DV IV1 IV2, 3) DV IV1 IV2 IV3, 4) DV IV1 IV2 IV3 IV4, and 5) DV IV1 IV2 IV3 IV4 IV5. they used.

The delta R² will stem from those models. It's not clear that hireg saves the delta R² values in memory though that are accessible post-estimation.

I think the real question is to discern what the base model the delta R² is being differenced from. My suspicion is that it is the full model less the model with 4 of the 5 predictors (i.e., less the focal predictor.

If this turns out to be true, those values will be equivalent to the conditional dominance values from the package domin (SSC) in the final (i.e., 5th) column of the e(cdldom) matrix. You can then operate on those values using Stata or Mata's matrix commands.

One suggestion is (using 3 predictors and the auto data):

sysuse auto

domin price mpg trunk turn

mata: st_matrix("values", 7:-abs(log(st_matrix("e(cdldom)")[., 3])))

matrix list values

Each row represents each variables "impact" as you define above. Again, this assumes that the difference from the full model less the focal predictor is what they used.

- joe