Table 1 Summary of the pros and cons of GEE and QIF

✔ GEE parameter estimates are efficient provided the true correlation structure is closely approximated. Parameter estimates are optimal in this case²;

✔ Modules for GEE are widely available in many statistical software applications;

✔ GEE parameter estimates are consistent irrespective of the covariance structure chosen, as long as the linear predictor and link function are correctly specified^1,2

✔ Has all the pros of GEE highlighted in the adjacent column³;

✔ Parameter estimates are efficient irrespective of correlation structure specified³;

✔ Includes a "chi-squared inference function" for testing goodness-of-fit and regression misspecification. The function follows a chi-squared distribution irrespective of the specified correlation structure. P-values less than 0.5 suggests that the specified model may be inadequate to describe the observed data ³;

✔ The goodness-of-fit test is analogous to the LRT, thus model selection criteria such as AIC (Akaike Information Criterion) and BIC (Bayes Information Criterion) are natural extensions³;

✔ Gives robust parameter estimates in the presence of outliers/contaminated clusters, by using an "automatic down-weighting strategy" through a weighting matrix³. This property is illustrated in Qu and Song ⁹;

✔ Existence of a lower bound is guaranteed since the function has a lower bound of 0, thus solving the problem of multiple roots associated with GEE ³;

✔ QIF gives similar results as the GEE when the independent correlation structure is assumed³.

✔ GEE assumes that the chosen model is correctly specified. It is often difficult to assess the goodness-of-fit of models built using GEE due to lack of an inference function like the likelihood ratio test (LRT) ⁷. The likelihood function for marginal models using GEE is often difficult to evaluate and intractable, especially for data that is not normally distributed ²;

✔ GEE parameter estimates are sensitive to the presence of outliers as illustrated in Diggle et al ² (page 165) and Qu et al³;

✔ GEE parameter estimates are not efficient if the correlation structure is misspecified. Inefficient estimates may lead to faulty inferences from hypotheses tests³;

✔ Non-convergence of results due to lack of an objective function, and the "multiple roots" problem associated with estimating functions like the quasi-likelihood function ²⁹.

✔ No software implementation available, but SAS macro available for download²⁸;

✔ Is only presently applicable to three working covariance structures: Independent, Exchangeable and AR(1) ²⁸;