Announcement

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

  • Question: Power, sample size, effect size analysis for GEE?

    Hello,

    I am planning on a longitudinal analysis of repeated cross-sectional measures using logistic regression with GEE and an independent correlation structure. The response variable is binary and the predictor an ordinal categorical variable (i.e. increasing levels of exposure so I know the sample size at each level of the predictor).

    When I've used cross-sectional data in the past I've used the minimum detectable odds ratio as a standardized effect size. Initially, I calculated the minimum detectable OR using Pearson’s chi-squared tests to compare two independent proportions specifying a .05 significance level and an 80% power level and used the local prevalence of the outcome to be the prevalence in the reference group. Is this an appropriate approach for repeated measures data? Are there alternative power analyses would you recommend for this in STATA?

    I would be grateful for any guidance or insight on the matter!

    Sincerely,
    Bill

  • #2
    What is it that you're powering your GEE model to detect?

    Comment


    • #3
      The minimum detectable odds ratio of the outcome occurring at each level of the exposure variable. The underlying population prevalence is 17%. I'm not sure that answers your question.

      Comment


      • #4
        Well, "exposure" to me is the product of intensity (your ordered-categorical predictor?) and time. Your longitudinal model includes both, and perhaps their interaction. So, what is your null and alternate statistical hypothesis pair—expressed in terms of a linear combination of regression coefficients from the GEE model—for which you have set rates of Type I and II error at 5 and 20%?

        Starting from that, you could compute sample size or minimum detectable effect by simulation, using the GEE model that you intend to use, and computing the test statistic (again, expressed in terms of a linear combination of regression coefficients) that you intend to use, over a range of assumptions as to parameters for correlation structure, sample size (if estimating minimum detectable difference) and effect size (if estimating sample size).

        Comment

        Working...
        X