Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Cox regression assumption

    I am doing a research to examine the association between sleep and all-cause mortality. The average follow-up time is only 5 years for both men and women.

    I was thinking to run cox regression since my outcome is death. However, after checking the assumption, unfortunately, I have evidence of non-proportional hazards with almost all covariates.
    Please any suggestion of what to do next or should I run different regression type?


    Thank you

  • #2
    Talla:
    you can explore the -strata- option for -stcox- (see Example 8, -stcox- entry, Stata .pdf manual).
    As an aside, if your outcome is -death-, in addition to -stcox-, it would be perfectly legal to use a non-parametric model (such as Kaplan-Meier curves) or a parametric model (see -help streg-).
    Kind regards,
    Carlo
    (Stata 19.0)

    Comment


    • #3
      Thank you Carlo for your reply. I am just wondering if I can use logistic regression since I can use a binary variable (died-live) or I can not?

      Comment


      • #4
        Talla:
        I would stick with survival analysis.
        Logistic regression does not take time into account: time to a given event can be investigated via survival analysis only.
        It is also worth remembering that non-parametric and parametric regeression models for survival analysis give back hazard ratios, which differ from probabilities and odds ratios.
        Last edited by Carlo Lazzaro; 01 Jul 2019, 11:38.
        Kind regards,
        Carlo
        (Stata 19.0)

        Comment


        • #5
          Thank you Carlo for your explanation and clarification I appriciated it.

          Comment


          • #6
            What timescale are you using? Attained age is probably the most appropriate timescale. If you are using time since entry as the timescale and individuals have different ages at entry, then this may induce non-proportional hazards.

            As Carlo said, stratifying on sex is probably advisable.

            If your primary interest is in studying the effect of sleep on all-cause mortality then non-proportional hazards for confounders is probably not a big problem; just model them as time-varying.

            Interpretation is easier if the effect of your primary exposure (sleep) is proportional, but if it's not then that is what you have. You could fit a different model (additive hazards or AFT) but standard is to report the time-varying HR.

            Comment

            Working...
            X