Sample size estimation based on GEE requires several
parameters, all of which were unknown at the design stage
of the study. These included the correlation of physicians’
recommendations for TKA, specifying the structure of the
correlation matrix, and a scale factor allowing for extra
variation (or dispersion) in the response beyond the
assumed variance [12]. For these reasons, although we
planned to analyze our data using GEE, we based our sample
size of h 5 58 on what was considered to be a conservative estimate using the exact McNemar test. Did
we commit a type III error?