I am conducting an individual patient level data meta-analysis using the survey package in R (I tried Stata 13 but I get stuck with an error).

As per meta-analytic practice, I would prefer to conduct both fixed and random effect analyses, but it appears no such option is available with the survey packages in either Stata or R. Briefly, study-level meta-analysis are typically based on a fixed-effect approach (eg Peto) when there is limited statistical inconsistency/heterogeneity. Conversely, a random-effect approach (eg DerSimonian-Laird) is used when inconsistency/heterogeneity is significant or when a more robust (but potentially less sensitive) analysis is sought (see for instance Kelley and Kelley for a brief tutorial on study-level meta-analysis).

The approach to patient-level meta-analysis is less established. A common strategy is to conduct analysis in two phases (first analyzing each study as if it was alone), and then combining with study-level meta-analytic methods the study-specific results. This approach, called two-stage, is easily conducted and reasonably robust.

More recently, one-stage approaches have been advocated, in which patient-level modelling takes into account study-level clustering. Several methods have been proposed for this analytical strategy, from conditional regression to generalized linear models or generalized estimating equations (see for instance Burke et al).

I have recognized that the survey packages in Stata and R can perform clustered analysis easily, but I am not sure whether I can specify fixed vs random effect in such models, nor whether I can quantify statistical inconsistency/heterogeneity between studies within such model.

My assumption though is that the survey package provides analyses which are equivalent to those of a generalized linear model or generalizing estimating equations with fixed effects with explicit correlations acknowledge in the model.

  • 4
    $\begingroup$ I don't think they are. You can use the lme4 package in R or xtmixed in stata to do random effects models. $\endgroup$ – Jeremy Miles Oct 3 '17 at 13:43
  • 3
    $\begingroup$ A reproducible example (including the one that produces an error) would help. For the benefit of me... since nobody looked at your question for a month ;)... can you explain the advice regarding both the fixed and random effect approaches to meta-analysis? Keeping in mind that different people mean different things re: fixed/random effects (stats.stackexchange.com/a/27693/5739). Please edit your question above, do not reply below. $\endgroup$ – StasK Oct 25 '17 at 4:02
  • 3
    $\begingroup$ Generally, the functionality of GEE vs. survey is complementary: GEE (attempt to) model correlations that survey just clusters on. But survey can incorporate stratification and other bells and whistles. So I agree with @JeremyMiles: there is no way to say that a particular GEE exactly coincides with a given survey-aware estimation command. Unstructured covariance sorts of goes in that direction, but it requires some sort of balance in the data, which survey estimation does not care about. $\endgroup$ – StasK Oct 25 '17 at 4:04
  • $\begingroup$ @StasK Thanks for the comment. I have substantially revised the question $\endgroup$ – Joe_74 Oct 25 '17 at 9:05
  • 2
    $\begingroup$ I would use lme4 to do a random effects regression in a one stage approach. Why do you need the survey package? $\endgroup$ – Jeremy Miles Oct 30 '17 at 18:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.