1

I don't know a lot about statistics and survival analysis so this might be a dull question.

I'm performing a survival analysis where I want to study the possible risk or protective factors that could lead to the incidence of a proctological pathology after the surgery of another proctological patology. I would like to obtain the Kaplan Meier curves and compare them using the logrank test. My question is if I can obtain a univariant HZ with the logrank test (since I've been reading that de Cox regression analysis is multivariant). After that,I would like to introduce to a multivariant analysis the variables that are statistally significative in the univariant analysis. In this case, should I use Cox?

kjetil b halvorsen
  • 77,844
Arthurr
  • 73

1 Answers1

1

There's no reason to do log-rank tests comparing Kaplan-Meier curves. For categorical predictors the score test of a Cox regression is equivalent to the log-rank test. For continuous predictors you would need to categorize the values, which is generally a bad idea; a Cox model can handle continuous predictors readily and model them flexibly as with regression splines.

I would like to introduce to a multivariant analysis the variables that are statistally significative in the univariant analysis.

I interpret "mutltivariant" and "univariate" here to mean "multiple-predictor" and "single-predictor" regressions. What you propose, although often done, isn't good practice. Whenever you use the outcomes to choose predictors of a model you risk optimistic overfitting; your model might fit your current data sample but won't extend well to new data samples.

There are principled ways to proceed. Frank Harrell's online Regression Modeling Strategies describes general aspects of regression modeling (Chapter 2), strategies for developing multiple-predictor models without peeking at the outcomes (Chapter 4), and specific issues in survival analysis (Chapters 17-21).

EdM
  • 92,183
  • 10
  • 92
  • 267