As a general principle, when comparing different models, e.g. progressive adjustment for confounders, one should always restrict the data to the same set of participants. Otherwise the models aren´t comparable.

In terms of your actual question, you can get an idea of which is happening by: 1) running your model without the confounder but restricted to people who are not missing the confounder, and then comparing to 2) running the model with the confounder. That way both models will be in exactly the same people. Look at the ORs and 95% CIs to decide whether there is any material/important change. You could also do a loglikelihood ratio test of the two models to see whether they are significantly different. Since you have so many missing observations for the confounder, consider including the confounder only as part of a sensitivity analysis.

Hope this helps.

As a general principle, when comparing different models, e.g. progressive adjustment for confounders, one should always restrict the data to the same set of participants. Otherwise the models aren´t comparable.

In terms of your actual question, you can get an idea of which is happening by: 1) running your model without the confounder but restricted to people who are not missing the confounder, and then comparing to 2) running the model with the confounder. That way both models will be in exactly the same people. Look at the ORs and 95% CIs to decide whether there is any material/important change. You could also do a loglikelihood ratio test of the two models to see whether they are significantly different. Since you have so many missing observations for the confounder, consider including the confounder only as part of a sensitivity analysis.

Hope this helps.

Subhojit,

In addition to endorsing Philip's general principles, I'll add one:

I wouldn't use an ad hoc missing data strategy such as listwise deletion. With current tools, there's rarely a solid reason to not use either a raw-data estimator or multiple imputation, and plenty of reasons to avoid listwise (especially with such a large number of missing).

Pat

Subhojit, I largely agree with Patrick. I almost completely avoid listwise deletion unless it results in very trivial loss of data. In response to your original question, there is no way to say whether the data loss or the case deletion causes the change in results. You can in almost any stats software work it out so that you are dropping the same cases for both models to at least satisfy that questions. But you are dropping 2000 cases to do that, and I don't know how many cases you are already dropping to run the first model without the confounder in it. Despite the hesitation many people have about imputation (assuming a valid approach is used), it's been demonstrated that listwise deletion is more biasing under most circumstances. When it's not biasing, under what's called missing completely at random (MCAR) nothing else really is either. Under most circumstances, multiple imputation yields unbiased results. Other approaches, such as models estimated under full information maximum likelihood also work well. Bob

Thank you guys and special thanks to Philip. I did what you suggested - made the models comparable and bingo! NO confounding. The change in values was indeed caused due to the loss of data. I guess I have no choice but to drop this variable due to which I am loosing 2000+ cases in the model.

Adding to Robert's comment: Even in the case of MCAR, listwise deletion is one of the worst strategies in terms of hurting power.

And Subhojit, you *do* have other choices, as Robert and I have been discussing. It is, of course, your call whether they're best for your situation.

Pat