Appropriate statistical methods to account for similarities in binary outcomes between fellow eyes

Many ocular measurements are more alike between fellow eyes than between eyes from different individuals. To make appropriate inferences using data from both eyes rather than the best or worst eye, statistical methods that account for the association between fellow eyes must be used.

Marginal and conditional regression models account for the association between fellow eyes in different ways. The authors compare and contrast these methods using data from a series of patients with retinitis pigmentosa in whom the primary object is to identify risk factors, some of which are subject specific and some of which are eye specific, for visual acuity loss (as a binary outcome) among affected subjects.

Odds ratios for age, gender, presence of posterior subcapsular cataract, and genetic type of retinitis pigmentosa obtained from the marginal model were all larger than those from the conditional model. Familial aggregation of visual acuity loss was statistically significant in the marginal, but not in the conditional, model.

The estimates and interpretation of the association between an ocular outcome and risk factors can differ significantly between these two approaches. The choice of model depends on the scientific questions of interest rather than on statistical considerations. Computer programs are available for implementing both models.