Announcement

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

  • How man observations are needed for a robust Cox proportional hazards estimation?

    Dear list,

    I'd like to use survival analysis to analyze the construction of wind turbine projects. However, my number of observations is low. I have between 13 and 22 wind farm projects (in various areas), which I know is low for OLS. What is the minimum number of observations necessary for a robust Cox estimation?

    PS: I'm working my way through An Intro to Survivaly Analysis Using Stata (Cleves, et al) and I've searched on the Stata Forum and other online resources, but I don't seem to find any advice about this.

    Thanks!

    -nick

  • #2

    Nick-

    For Cox sample size calculations, the key quantity is the number of events ("failures"), not the number of observations. But the calculation depends on the goal of the analysis (hypothesis testing, estimation, prediction). A quick Google search on "sample size Cox regression" turns up many references. See especially Vitinghoff and McCulloch, who say that 5-9 events per variable may be satisfactory for some purposes. (http://www.medcalc.org/manual/cox_pr...al_hazards.php).

    There are also online calculators, such as that at http://www.medcalc.org/manual/cox_pr...al_hazards.php. In Stata, stpower cox will give the sample size for testing the effect of a single covariate; stpower logrank will compare two curves whose hazard functions are proportional.

    Note the strong preference on the List for usernames that show first and last names. (FAQ Section 6). You are not hard to identify, so why not make it official? You can re-register via the Contact Us button at the bottom right of the page.
    Last edited by Steve Samuels; 28 Jan 2015, 19:53.
    Steve Samuels
    Statistical Consulting
    [email protected]

    Stata 14.2

    Comment


    • #3
      Dear Steve,

      Thanks for your post and the clear explanation -- this is very helpful. Since, in my dataset, there is no right censoring observations = events. However, I have only 3-4 variables so it seems like a relatively small data set should be effective.

      A related question: When evaluating Cox estimation outputs, can I treat the Chi2 p-value as I would treat an F-test p-value in OLS?

      Thanks again for all the help!

      -nick

      PS: Per your request, I have contacted the mods.

      Comment


      • #4
        Thanks for re-registering, Nick. To answer your question: yes, you can treat the Chi-square test as you would the F-test in OLS.

        Although you have not many units in any given area, you could fit a combined model, with a separate intercept for each area, Since you have no censoring, you can also fit non-survival models, for example, ordinary regress for mean duration or log duration, but with robust standard errors. You could even fit quantile regression qreg, or the very robust mmregress package, by Verardi and Croux s("robust" in its sense of insensitivity to outliers and high leverage points (findit mmregress).
        Last edited by Steve Samuels; 29 Jan 2015, 16:43.
        Steve Samuels
        Statistical Consulting
        [email protected]

        Stata 14.2

        Comment

        Working...
        X