We use the IFIH1 genotype data in Table 5 to illustrate differences between GEE and ESS. Using the corstr=“exchangeable” option in the gee subroutine in R statistical package (VJ Carey, T Lumley, and B Ripley, “The gee package”, version 4.13–12, Feb 2007), the averaged within-family correlation coefficient for the allele count variable was estimated as r =0.4349. This r value is slightly smaller than that for sibpairs (r =0.5), but close, reflecting the fact that this dataset is dominated by sibpairs. In the Results section, we have shown that using r = 0.5, the sample size reduction for the dataset in Table 5 is equal to 0.649. If we use the within-group correlation coefficient r =0.4349 estimated by GEE, the sample size reduction is 0.678. The GEE and ESS results are more or less the same, though GEE does not seem to correct the correlation enough. A similar observation that GEE tends to underestimate variance for smaller sample sizes was made in (Trégouët et al., 1997).
