Announcement

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

  • Meta-analysis of incidence rate ratios - problem with weights

    Dear Statalisters, (sorry posted in the wrong area initially)

    I have a data set of four studies that I would like to meta-analyse using a random effects model. In the papers, the data is presented as incidence, as per study 2 the incidence is 66.8 per 100,000 person years compared to the general population of 5.7 per 100,000 person years. They do not supply total person years from their study. The data from the studies is in Table 1 below:

    Table 1
    study id personyear1 events1 personyear0 events0 patientswithtb samplesizeofcases followuptimeyears selogirr logirr _LCI _UCI _WT
    1 9672 13 100000 11.9 13 2806 6 0.401194 2.424346 5.144972 24.79574 10.78166
    2 100000 66.8 100000 5.7 37 14506 6 0.436358 2.461237 4.982791 27.56326 9.113976
    3 21620 105 100000 69.8 105 4131 12 0.154436 1.939878 5.140697 9.41747 72.76086
    4 9381 7 100000 10.7 7 9381 2 0.486123 1.94215 2.689533 18.08229 7.343501
    gen selogirr=sqrt( 1/events1 + 1/events0)
    gen logirr=log( (events1/ personmonth1)/( events0/ personmonth0))
    I get the effect sizes (ES) and their pooled ES by using :
    metan logirr selogirr,random eform

    but the problem is, that I need to get the correct weight of the random effects model from the sample size of cases (variable name samplesizeofcases), and not the computed ES-- personyear1 event1 personyear0 event0. I only have the data in Table 1 to work with.

    The weight assigned (Table 1 WT variable) for study 2 with the biggest sample size (n=14506) was only 9.1%, as opposed to study 4 with 72% of the weight but only 9381 cases total. This is wrong.

    Any suggestions on how I should proceed?

    Many thanks in advance, Carole

    Reference: Bagos PG, Nikolopoulos GK. Mixed-effects Poisson regression models for meta-analysis of follow-up studies with constant or varying durations. 2009, The International Journal of Biostatistics, 5(1), Article 21 [PDF] http://www.compgen.org/tools/poisson-meta-analysis

  • #2
    Carol:

    As you saw when you registered, Statalist etiquette is to use full real names names. Although this is not mandatory, my own policy is not to engage with posters whose names I don't know. For any follow-up discussion, therefore, please change your user name to your full name-just click the "contact us" button on the bottom right of the page.

    The standard error formula requires the actual number of events, which is a whole number. You have assumed that the incidence rate of 66.8 per 100,000 for Group 1, Study 2 meant that there were 66.8 observed events, a fractional number, and 100,000 PY. Presumably the correct number of events is 37, the "patientswithtb" for this group. (This would make the actual PY (37/66.8)x100,000 = 55,389). Likewise "events0" is not the actual number of events in the general population control groups. To approximate the correct numbers, estimate the actual person-years from external data (population x years), then multiply by the published incidence rates in column 4 (divided by 100,000). The resulting numbers may turn out to be so large that they can be ignored in the SE calculation.
    Last edited by Steve Samuels; 28 Jul 2014, 07:42.
    Steve Samuels
    Statistical Consulting
    [email protected]

    Stata 14.2

    Comment


    • #3
      Dear Steve,
      Your reply to my post is much appreciated, especially given I had violated forum protocol in my ignorance.

      In the fields for personyear1, event1, personyear0, event0 I give the numbers that the papers report. Study 2 states (or reports) a rate ratio of 11.7 in the paper.

      Study 2 : the incidence is 5.7/100,000/year in the general population (n=19,855,283). The study gives the number of events in the general population as 6276.
      The events in the patient group is 37, with an incidence of 66.8/100,000.The study reports a rate ratio of 11.7.

      So then according to your suggested calculation: (6276*5.7)*100,000= 3,577,320,000 person years in general population.

      I would then plug these into my rate ratio calculation: irr= (events1/ personyear1) / ( events0/ personyear0)
      irr= (37/55,389) / (6276/3,577,320,000) = 380.76

      So the rate ratio here would be 380.76. But the study reports a rate ratio of 11.7.

      You continued help with this is greatly appreciated, Carole


      Comment


      • #4
        Hi Carol/any other Statalisters,
        How did you handle the issue of incorrect weigths assigned with the metan logirr selogirr,random eform command?

        Thank you,
        Jackson

        Comment

        Working...
        X