Survival analysis

[Paolo Costa]

Hello, I've got a question for anyone so kind to answer me:

When speaking about survival analysis, KM curves are usually not so right in description of survival functions in case of non RCTs since confounders may play a role in these cases.
Therefore, Cox regression could be the solution for adjusting for such covariates, and therefore have the right graph with HR adjusted.

My question is: can I perform a PSM on the two cohorts and they're just use KM curves? I mean, in this situation adjusting doesn't have sense and matching just eliminate the need of complicated regression models, isn't it?

Thanks,

[Carlo Lazzaro]

Paolo:
yes, you can.
The issue is that, presenting the results of the (semiparametric) Cox regression, the reader can have an idea of the HRs, that are more informative than the KM log-rank.

[Paolo Costa]

Thanks Carlo.
The fact is that since I would have to adjust for lots of covariates, matching seems to be easier to me. Moreover, it would be difficult to give a mean value to each variable to have the real survival curve function, since I cannot represent KM curves adjusted.

Paolo

[Carlo Lazzaro]

Paolo:
most depends on the tribal rules in your research field.
Adjusting for (too) many covariates may overfit your results.

[Paul Dickman]

I cannot represent KM curves adjusted.

Why not? What's preventing you from fitting a model and calculating marginal survival curves?

[Jeff Wooldridge]

I agree with Carlo and Paul. It’s not really easier to include covariates in propensity score estimation than the hazard or survival function. You can estimate separate ones for control and treatment and then average out the covariates. And then you can do proper inference rather than conditioning on the PS and matched sample.