I am trying to run a multiple linear regression using weighted cases (there were more women than men in my sample), and I am trying to look at the Cook's D for any outliers that may be affecting my model, but it only calculated it for men. I think this may be the reason why my model is significant, but the individual IVs come up as not statistically significant.

Thanks!

Julia

I think you are using fractional weights. All the cases will be used in the regression estimation, but certain output variables, including Cook's distance, df betas, and certain adjusted variables do not include cases where the weight is less than 1.

BTW, be sure that the sum of the weights equals the number of cases.

Hi Jon,

Thanks for the response. In my class we have only been taught how to do weighted variables using fractional weights. This is what I did:

% population/ % sample.

1. Females: 49.9/87/5 = 0.570
2. Males: 50.1/12.5 = 4.008.
3. Compute variables - If - Sex -1(females) = 0.570
   1. If - Sex-2(males) = 4.008
4. Data - Weight Cases - reweigh sex with the new weight variable.
5. Then ran my regression as normal.

Should I be doing something else?

So you normalized so the weighted N equals the unweighted N. That's good, but those weights are pretty extreme at almost 8 to 1, so this could have a big effect on your results. There are several different kinds of weights - too much to detail here. The ideal strategy could be simple weighting (which you can also do in REGRESSION as a weighted least squares weight), as a probability of selection weight using the Complex Samples procedures, and, as an alternative, unweighted but including the weight or it determinates as a regressor. Or you could run separate regressions for M and F.

It's often a good idea to try several alternatives to see how robust your results are.

Although you can't get Cook's distance from the Regression procedure in this case, you can get it for all the cases if you use GENLIN (Analyze > Generalized Linear Models < Generalized Linear Models with the same equation.