Can I use Cox regression in a nested case-control study to get adjusted risk ratios?

I have 70 patients who died suddenly and I've drawn 140 controls for these patients (2:1 matching on age, gender, coronary artery disease and smoking). When looking at the baseline characteristics of these patients, the groups differ statistically significantly only in terms of one factor (arterial hypertension) and one factor is of borderline significance (left ventricular hypertrophy).

Now, I have three new markers (binary) that I'd like to study. Normally I would just do Cox regression univariately for each variable and then do multivariable Cox regression analysis and adjust the risk ratios for those factors that had significance of P<0.100 and as well as for few other factors which are usually always included, like body-mass-index and heart rate, previous myocardial infarction (even if they're not univariately predictive).

I've done this and get pretty reasonable risk ratios: e.g. univariate/multivariate: 5.8/6.8 (for some reason the multivariate RR is higher than that of the univariate).

Then I read somewhere, that I couldn't use Cox regression in nested case-control study. Instead I should do conditional logistic regression. I did this in SPSS by "tricking" the Cox regression as instructed at

The thing is that I got risk ratios that don't seem right. Univariate/multivariate: 18.4/34.6. I've never seen risk ratios this high ever before.

So, my question is:

1) Can I use Cox regression in this nested case-control study to account for the clinical factors not used in the matching?

2) If not, is logistic regression the way to go or is there some better way to account for the group differences?

All help is welcome. Thanks,


PS. all variables entered into the model are binary and there aren't any missing values...
Re: Can I use Cox regression in a nested case-control study to get adjusted risk rati

Is the data cross sectional? For cross sectional data I would use logistic regression (I'm not very familiar with cox regression, perhaps it's appropriate here but I thought that longitudinal data was needed). I would advise against choosing covariates based on P values. The differences between univariate and multivariate estimates are possible, this reflects confounding by one or more of the covariates. I also suggest you construct tables and simply look at the raw numbers and percentages first, before doing regression. This might shed some light on the effects of covariates on the risk estimates.