I would like to compare survival among patients stratified by insurance status, where insurance status has 3 potential values (private insurance, Medicaid, uninsured). I would like to treat private insurance as the reference group and compare survival among Medicaid patients and uninsured patients to patients with private insurance; I am not interested in comparing survival between Medicaid and uninsured patients.
If I were to make this comparison using Kaplan-Meier curves, the generally accepted way appears to be to first perform a log-rank test to see if there is a significant association between insurance and survival. If that log-rank test is significant at p < 0.05, then I could perform post hoc tests where I treat private insurance as the reference group and compare Medicaid survival to private insurance and uninsured survival to private insurance using a p-value adjustment for multiple comparisons. In SAS, a reasonable adjustment method in PROC LIFETEST appears to be ADJUST = DUNNETT, which allows you to treat one group as reference rather than perform all pairwise comparisons.
If I were to make this comparison using a Cox proportional hazards regression model, the generally accepted way appears to be to treat private insurance as the reference group in the model and compute hazard ratios comparing Medicaid mortality to private insurance and uninsured mortality to private insurance. For those 2 comparisons, significance is then evaluated using 95% confidence intervals.
My first question is why are p-value/confidence level adjustments encouraged for Kaplan-Meier curves but not hazard ratios from the Cox model in this scenario (assuming my understanding of generally accepted practices is correct)? Shouldn’t the confidence intervals for the hazard ratios be made wider to account for multiple comparisons? This never appears to be done in the literature, though.
My second question is since the comparisons to the reference group are of primary interest rather than the overall significance of the independent variable, when using Kaplan-Meier curves, is it ever acceptable to treat the comparisons to reference as planned comparisons and not perform the overall log-rank test? In a Cox model, the Type III p-value appears to function as a test for the overall association between the independent variable and mortality, but I don’t think I’ve ever seen it reported in the literature, but log-rank values for the overall association between the independent variable and survival are regularly reported.
Because p-value adjustments are based on multiple tests within 1 data set. The curve situation has more than 1 from your description. these references may help:
https://b-ok.cc/book/1059549/e822e0
Best, D. Booth
Dr. David Eugene Booth, thank you so much for your response and references. I agree with your assessment that for the Kaplan-Meier curves, there is multiple testing when comparing survival among private insurance (the reference group) to Medicaid and uninsured patients. In regard to Cox models, is the confidence level not adjusted for multiple comparisons because comparing mortality among private insurance to Medicaid and uninsured patients is done simultaneously within the model?
Not quite it just depends on the circumstances If you are making multiple conclusions about related things then it can be. If the hypotheses are independent no. I believe the ETH document discusses that. If not, google multiple comparison tests or look at this attachment: https://pdfs.semanticscholar.org/31d3/dfb112f8e4c94c1a4077a06e82bd80be845d.pdf my Mentor believed that it is more important to use one than exactly which one. That's what the ETH people do but carefully. Best, David Booth
Matthew C Simpson Perhaps a more clear explanation than mine is given here:
https://en.wikipedia.org/wiki/Multiple_comparisons_problem
Best wishes, D. Booth
- Badges
- Science topic
- Similar topics
- Cognitive Science
- Thinking