Kaplan-Meier curves for retrospective studies While doing a review i noticed that Kaplan-Meier curves are often used for retrospective studies in the medical field.
However, is it correct to do so even if the registry of data is not prospective? I suppose that the risk of bias is way higher than prospective registries. Wouldn't it be more precise to use other tipe of analysis, such as normal regressions models, even if more assumptions are made in such models (and they actually test for different questions)?
Thanks for the answers
 A: What most people would call "normal regression models" don't handle the censored survival times typical of such studies. The Kaplan-Meier display of survival data has the advantage of requiring minimal assumptions while handling such data. As with any data analysis method, it's only as good as the underlying data allow.
Whether retrospective or prospective, all observational cohort studies (as opposed to randomized prospective trials) run a risk that the apparent association between the primary exposure of interest and outcome might be due to other, confounding variables. In both situations you need some way to try to account for confounders; in both situations, there's a danger that you won't have data on important unknown confounders.
Kaplan-Meier curves by themselves don't provide an efficient way to account for confounders. That's why such studies usually go beyond Kaplan-Meier plots and perform some type of regression analysis appropriate to censored survival data, typically a Cox proportional hazards regression, to try to account for the measured potential confounders.
There are two major differences between retrospective and prospective cohort studies in terms of potential bias. First, a prospective study can be designed to ensure more complete data on potential confounders. Second, a prospective study might allow study-specific follow up, beyond what's available in standard clinical records, on those enrolled. That can minimize "informative" loss to follow up/censoring, a potential source of substantial bias.
In a setting with reasonably complete data on confounders and patients, like hospital records on patients who come back regularly for clinical visits, the risk of bias in a retrospective study need not be "way higher" than in a prospective study. This review on "Cohort Studies: Prospective versus Retrospective" by Euser et al., Nephron Clin Pract 2009;113:c214–c217 is a brief outline of the issues.
