Sensitivity analysis

[Emma_at_SAS]

In my research, I asked the gender of the teenagers from the kids and their parents. In about 99.5% of the cases, the gender reported by kids and parents match. When I look at the association between gender and risky behavior, I want to use one of these gender variables. If I want to use the gender reported by the kids and do a sensitivity analysis that the results would still be almost the same if I was using the gender reported by parents, how would I do that? I am at the bivariate analysis stage. Thanks

[sbxkoenk]

Hello,

So you want to find out how the results change if

predictor1 (gender_repd_by_kids)

is replaced by the very highly correlated (almost equal) 

predictor2 (gender_repd_by_parents).

That's not my definition of a sensitivity analysis, but I see what you mean.

I would first check to what extent the confidence bounds overlap in analysis 1 versus analysis 2 (for whatever parameter you try to estimate). But others may have better ideas.

Cheers,

Koen

[Emma_at_SAS]

Hello, and thank you, Koen! I very much appreciate your insight. 

[SteveDenham · Accepted solution]

Try looking at log likelihoods or information criteria for these four models:

Model 1: predictor1 (gender_repd_by_kids)

Model 2: predictor2 (gender_repd_by_parents).

Model 3: both predictors in the model

Model 4: Null model (no predictors)

You can either do likelihood ratio tests, or use the AIC to calculate the relative amount of information retained.  Comparisons of models 1 and 2 to model 4 tell you something about the effect of each of the predictors, while comparison of model 3 to models 1 and 2 would tell you something about how much additional information is in the added variable (for 3 vs 1, it tells you how much gain there is in predictor 2, and vice versa for 2 vs. 1).

SteveDenham

[Emma_at_SAS]

Thank you, SteveDenham! It was very helpful.

Does this mean that I need to check them in a regression model, even for bivariate analysis?

Thanks

[SteveDenham]

The approach I suggested does require fitting a model, even for bivariate examination, as I don't see an easy way to get the necessary values for the null model otherwise.

SteveDenham

[Emma_at_SAS]

Thank you, SteveDenham. I appreciate it. 

Thanks