0
$\begingroup$

I followed two groups of people for one year. Both were subjected to the same conditions, just in two different periods. One group was followed throughout 2019 and another was followed throughout 2020. My intention is to make a Kaplan Meier estimate for the two groups. My question is: can I combine them into one sample and make a single Kaplan Meier estimate or should I carry out my analysis separately, stratifying it for the two years? Thank you. I'm only on my second question here, so forgive me if my request for an explanation sounds a bit naive.

$\endgroup$

1 Answer 1

0
$\begingroup$

Look at the data first, and let the data be your guide.

If there were no differences between 2019 and 2020 with respect to the phenomenon of interest, it might make sense to combine the calendar-year groups. But blindly combining the groups without examining the data might lead you to miss an important distinction between 2019 and 2020. Given the 2020 pandemic, one might expect some substantial differences between those 2 calendar years.

If it's just a matter of following the same type of individuals in each of 2019 and 2020 without any treatment interventions or evaluation of covariate effects, you could do a log-rank test on separate Kaplan-Meier curves for the 2 years (equivalent in that case to the score-test result of a Cox regression with calendar year as a categorical predictor) to see if the calendar years differ by a greater amount than would be expected by chance. If not, combining the 2 years might make sense.

Maybe you have a treatment/control comparison performed in each of the calendar years. With 1 treatment group and 1 control group for each calendar year, that's only 4 Kaplan-Meier curves, quite easy to visualize. A Cox model with an interaction between treatment and calendar year can both test whether the treatment effect differs between the years and evaluate whether the survival under control conditions differs between the 2 years.

$\endgroup$
3
  • $\begingroup$ I am interested in determining the effect of a categorical variable on the event of interest. There are no treatment/control comparisons. Just observation of individuals with different peculiarities. I performed a cox regression with the calendar year as univariate and found a significant difference at the log-rank test. I ran the same regression adding my variable of interest as a cohort and the year lost statistical significance. Which result should I refer to in order to make the choice? I believe the one from the second cox regression but I am not so sure. Thanks anyway for your reply. $\endgroup$
    – n_cer
    Apr 11, 2021 at 16:32
  • $\begingroup$ @n_cer did your second Cox model include an interaction between the categorical variable and the calendar year? Or did you just examine additive effects of those predictors? A different balance between categories or a different association of category with outcome from year to year might be responsible for the log-rank test difference when you just examined calendar year as predictor. A model with the interaction will help distinguish those possibilities most clearly and provide the most information about any possible association of calendar year with outcome. $\endgroup$
    – EdM
    Apr 11, 2021 at 16:45
  • $\begingroup$ The model with the interaction showed an effect of the calendar year on the outcome (survival probability was lower in 2020). Therefore, I chose to present the results for each individual year. Thanks again. $\endgroup$
    – n_cer
    Apr 11, 2021 at 17:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.