Announcement

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

  • Meta-analysis of Precalculated Effect Estimates - advice needed on coding

    Dear all, I would be very grateful for your advice. I am currently working on a meta-analysis of studies reporting RRs.

    This is the code that I'm using:

    gen lnrr = ln(rr)
    gen lnlci = ln(lci)
    gen lnuci = ln(uci)
    metan lnrr lnlci lnuci, eform effect(RR) lcols(study year)

    Is this valid for RR (or HR) summary effect estimates? Is this the fixed-effects model using the Mantel-Haenszel method?

    I read somewhere that you can only use a fixed-effects model when working with precalculated effect estimates - is this true? A random effect model seems more appropriate given the level of heterogeneity between the studies that I'm looking at.

    Many thanks in advance!

    Dearbhla

  • #2
    Anybody?

    Comment


    • #3
      Hi Dearbhla,

      The code that you propose is fine, but it will perform the fixed-effect model using the inverse-variance method, not the Mantel-Haenszel. The Mantel-Haenszel method can only be used with raw count data for each study (i.e. numbers of treated cases, non-treated cases, treated controls, non-treated controls ... or similar). Do you have access to such data?

      The statement r.e. "pre-calculated estimates" sounds confused to me, and I would advise reading through some introductory meta-analysis texts to clarify matters. The decision as to which model to use (Mantel-Haenszel, standard fixed-effect, random-effects, etc.) should ideally be pre-specified, possibly including some statement that model B or C will be used if large heterogeneity is detected, or if very few studies are eligible for inclusion, or if data is sparse, etc. and that Model A will be used otherwise.

      Does that help at all?

      Thanks,

      David.

      Comment


      • #4
        Hi David,

        Thank you very much - that is really helpful. I don't have access to raw count data for the individual studies unfortunately. There does seem to be quite a lot of heterogeneity between studies (I squared = 73% when performed using the fixed effect model above). Do you know how I could modify the above formula to get perform a random effect model?

        Best wishes and thanks again,
        Dearbhla

        Comment


        • #5
          Does anyone know how to change the following code to make it a random effects model please?

          gen lnrr = ln(rr)
          gen lnlci = ln(lci)
          gen lnuci = ln(uci)
          metan lnrr lnlci lnuci, eform effect(RR) lcols(study year)

          Comment


          • #6
            Hi Dearbhla,
            You just need to add "random" to the end of your line of code. If you type "help metan" from within Stata, you can see the list of options that you can use.
            Thanks,
            David.

            P.S. Doing this will implement the standard DerSimonian-Laird random-effects model. Other models also exist, but you'd need to read about them to decide which to use.
            Last edited by David Fisher; 03 Oct 2018, 05:11. Reason: Added P.S.

            Comment


            • #7
              Thanks very much David. Would this code be correct for a DerSimonian-Laird random-effects model?

              gen lnrr = ln(rr)
              gen lnlci = ln(lci)
              gen lnuci = ln(uci)
              metan lnrr lnlci lnuci, eform effect(RR) randomi lcols(Study Year)

              Comment


              • #8
                How does randomi know the weighting variable? I am using this code and it gives me an error..metan Fisherz3 SEz4, randomi wgt(n) lcols(Study Country) rcols(n) textsize(200) aspect (0.5) xlab (-0.3, -0.2, -0.1, 0.1, 0.2, 0.3)

                Comment

                Working...
                X