[The Conundrum of Relationships in the Multivariate Dataset of nvAMD Treatment]

Background The evaluation of relationships in clinical case series is complicated by statistical pitfalls, one of which is spurious correlation. In this case, one or more potential factors are actually not related to a target measure, yet it may be wrongly inferred that they are. This may be due to either coincidence or the presence of a one or more confounding unregarded factors. In randomised clinical trials, it is attempted to eliminate such confounding factors or to control them by randomisation. However, this approach is not applicable in retrospective case series, and other statistical methods must be applied, such as multivariate analyses, taking into account all known variables. Patients/Material and Methods Statistical evaluation of the impact of different variables on the change in visual acuity after year 2 in a cohort of 1300 cases of nvAMD. The investigated variables are: visual acuity (VA) at baseline, central retinal thickness at baseline, age at baseline, mean interval between visits, number of injections, initial gain in VA after upload, and a change of the anti-VEGF compound. The target outcome measure was the change in VA as letter score at the end of year 2. Results All investigated variables have a marked impact on VA outcome when only single regression analysis is performed. Initial analysis of a switch in therapy gives a seemingly large effect on VA. However, multivariate analysis to avoid spurious correlations shows that only the following variables influence VA outcome with statistical significance: VA at baseline, initial gain in VA after upload, number of injections, mean interval between visits, and central retinal thickness at baseline. A switch in therapy had no statistically significant effect on VA outcome. Conclusion A switch in anti-VEGF therapy is not causally related to better VA outcome. In the statistical evaluation of non-randomised, retrospective case series, special attention must be given to potential confounding variables. When subgroups under investigation are not well balanced for confounding factors, misleading spurious correlations can result, leading to incorrect inferences.