Advice on observational study data analysis

[Omar Wur]

Hello everyone,

I am doing a fairly large population study to estimate the treatment effect of a drug in a particular patient group. More specifically, I would like to compare the average age at death for patients that do use this drug vs patients that do not.

One approach I am considering is using propensity score matching to match both groups with their healthy counterparts (which do not use the drug) based on several variables to see what the treatment effect is in both groups. Another approach I have seen is a Kaplan–Meier analysis with a log rank test. Some advice on the best approach would be highly appreciated.

Also, one issue that I am stumbling upon is that most of the subjects are still alive and currently the average age at death is slightly below the average age of the population. Would this be a problem?

[Clyde Schechter]

It is unclear from your description whether you have three groups (sick patients who take the drug, sick patients who do not take the drug, and healthy people [who do not take the drug]) or just the first two.

In any case, in observational data analysis we always worry about the confounding effects of other differences between the groups being compared. A Kaplan-Meier analysis with log-rank test is not able to make any adjustments for other variables, so it is generally not suitable for analyzing observational studies. That is, while one might present the Kaplan-Meier analysis as a preliminary take, its result cannot be considered sufficient: additional analyses making adjustments are needed.

The Cox proportional hazards model is probably the most popular approach to this kind of problem, and it is implemented in Stata as the -stcox- command.

There are other, parametric, approaches to survival analysis, but these are less popular and generally rely on more unverifiable assumptions than the Cox model does.

As for propensity matching, it isn't clear what it will buy you if your data sample is as large as you seem to imply. It doesn't really offer any benefit you can't achieve by just including the same variables you would use in your propensity model as covariates in the survival analysis itself. The one situation where a propensity approach is beneficial is when the number of covariates needed is too large to include in the survival analysis (usuaully due to sample size limitations). But with a very large sample and a reasonable number of potential confounding variables, this advantage evaporates.

As for most of the subjects' being stlil alive, these are known as censored observations. All of the commonly used survival analysis techniques (including Cox proportional hazards and Kaplan-Meier) rely on the assumption that the censoring is uninformative. This assumption can be questionable, particularly in studies of drug effects, because patients may withdraw from observation (and, hence, be alive at the end of their participation in the study) as a result of poor response to the drug or other causes of deteriorating health. In this situation the censoring can be related to the drug effect itself. (It can also go the other way: participants may withdraw due to side effects, and side effects may sometimes imply that the drug is working very well in terms of treating the underlying fatal illness.) So there are a lot of issues to think about and discuss when you present your results.

[Omar Wur]

Thank you for the advice, Clyde! It is indeed three groups. I will take a look at stcox in the coming days.

[Barc]

It is unclear from your description whether you have three groups (sick patients who take the drug, sick patients who do not take the drug, and healthy people [who do not take the drug]) or just the first two.

In any case, in observational data analysis we always worry about the confounding effects of other differences between the groups being compared. A Kaplan-Meier analysis with log-rank test is not able to make any adjustments for other variables, so it is generally not suitable for analyzing observational studies. That is, while one might present the Kaplan-Meier analysis as a preliminary take, its result cannot be considered sufficient: additional analyses making adjustments are needed.

The Cox proportional hazards model is probably the most popular approach to this kind of problem, and it is implemented in Stata as the -stcox- command.

There are other, parametric, approaches to survival analysis, but these are less popular and generally rely on more unverifiable assumptions than the Cox model does.

As for propensity matching, it isn't clear what it will buy you if your data sample is as large as you seem to imply. It doesn't really offer any benefit you can't achieve by just including the same variables you would use in your propensity model as covariates in the survival analysis itself. The one situation where a propensity approach is beneficial is when the number of covariates needed is too large to include in the survival analysis (usuaully due to sample size limitations). But with a very large sample and a reasonable number of potential confounding variables, this advantage evaporates.

As for most of the subjects' being stlil alive, these are known as censored observations. All of the commonly used survival analysis techniques (including Cox proportional hazards and Kaplan-Meier) rely on the assumption that the censoring is uninformative. This assumption can be questionable, particularly in studies of drug effects, because patients may withdraw from observation (and, hence, be alive at the end of their participation in the study) as a result of poor response to the drug or other causes of deteriorating health. In this situation the censoring can be related to the drug effect itself. (It can also go the other way: participants may withdraw due to side effects, and side effects may sometimes imply that the drug is working very well in terms of treating the underlying fatal illness.) So there are a lot of issues to think about and discuss when you present your results.

Thanks for the advice.