Statistical Analysis

1

Appendix e-1

Statistical Analysis

First, we analyzed differences between rCELs and nCELs in summary measurements such as mean, maximum, starting and total lesion volume. All summary lesion volumes were transformed using a natural log function because volume distributions of each summary lesion were not normally distributed (i.e., skewed to the right). Secondly, the likelihood of a CEL to convert into a cBH was assessed.

For the first two sets of analyses, generalized estimating equation methods were used. We assumed that distributions of log transformed summary volumes follow normal distributions and occurrence of cBHs follows binomial distribution respectively. A generalized estimating equation model with intercept only was used to investigate the proportion of rCELs.

For all generalized estimating equation methods, covariance structure for measurements per subject was modelled as intraclass covariance matrix in order to take into account the correlations of multiple measurements on the same individual. Estimates of parameters were produced with empirical standard errors estimates, and z-statistics were computed to assess significance. Summary descriptions of lesion volume were obtained by back transforming the summary statistics of log transformed lesion volume. Back transformation yields geometric means that are estimates of median summary lesion volume. The distributionsof the summary lesion volume for two lesion types are described using 90% confidence interval (CI), obtained through back transformation, but raw data were used to create graphs.

Length of persistence of cBHs originating from rCELs and nCELs were estimated using Kaplan-Meier curves for each type of lesion and differences between lesion types were compared by the log-rank test using a proportional hazard model. Duration of enhancement was used as a covariate in the proportional hazard model fitted by marginal model using the option of COVSANDWISH,e1 in PROC PHREG in SAS.

Lastly, we investigated the association between the percentage of rCELs with respect to total CELs and either the number of relapses or the change in EDSS score during the study period by using Pearson’s correlation coefficient.

All statistical tests were two-sided and conducted with an alpha of 0.05. Generalized estimating equation was implemented in the GENMOD procedure of SAS version 9.1.3 (SAS Institutes Inc., Cary, NC).

References

e1. Wei LJ, Lin DY, Weisfeld L. Regression analysis of multivariate incomplete failure time data by using the marginal distributions. Journal of American Statistical Association 1989;84:1065-1073.