Hey! I want to know the hazard ratio for male vs female using keplan-meier survival command, however so far in SPSS they can only compute survival plot and statistic significance without telling me the Hazard Ratio, is there any way to compute it? (Using log rank, it only provides significance)
You may find some relevant info (not specific to SPSS) here:
- https://stats.stackexchange.com/questions/136472/how-to-calculate-hazard-ratio-from-kaplan-meier-curve
Q. Is there some reason why you do not want to use Cox regression?
Bruce Weaver Hello, thanks for your response!
To my knowledge, Kaplan Meier is the first thing we need to do before proceeding into the next step, this is because Kaplan Meier (with Log Rank) provides p-value that shows statistical significance for my variables in regards to the capability of surviving (variable selection in bivariate analysis before proceeding into multivariate stage). If the log rank value shows p-value < 0.25, I'd put them into the multivariate model and otherwise, I'd exclude them from the model.
Please correct me if I am wrong!
Hello Fadhaa Aditya Kautsar Murti. The approach you describe tends to produce overfitted models, IMO. See the "univariate pretesting" section in this 2004 article by Mike Babyak:
- https://people.duke.edu/~mababyak/papers/babyakregression.pdf
PS- Babyak uses the term univariate pretesting where I would say univariable pretesting. See the following article for more info on the distinctions between univariate/univariable and multivariate/multivariable.
Article Multivariate or Multivariable Regression?
Dear Bruce Weaver thanks for responding! I've read the article you referred. I've summarized three possible ways to avoid overfitting:
- Gather more data
- Add more covariates/predictors
- Penalization (not familiar with this one)
But however, I have questions regarding the first and the second one: to what extent that we need more data and predictor? I believe that there are points where more data lead to unnecessary findings (i.e, the higher sample size the more likely you get significant statistic result in chi-square, thus big sample size could lead to bias result)
Hello Fadhaa Aditya Kautsar Murti. I don't know where you got "Add more covariates/predictors" in Babyak's article. The second suggestion in that section was to combine predictors if there is a sensible way to do so. And doing that would reduce the number of degrees of freedom eaten up by predictors. Indeed, noticed what the first sentence of that section says:
"One obvious approach to preserve degrees of freedom is to
reduce the number of predictors in the model."
You asked about hazard rates, which implies that you have a survival model. For survival models, the limiting sample size is the number of events. Babyak's recommendation is that you have at least 10-15 events per explanatory variable degree of freedom. Nowadays, Frank Harrell recommends at least 20 EPV.
https://discourse.datamethods.org/t/author-checklist/3407#arrow_rightoverfitting-and-lack-of-model-validation-26
I hope this helps.
- Badges
- Science topic