Stata/SE 10 and 12 (StataCorp, College Station, Tex.) were used
for data analyses. Chi-square or Fisher's exact tests were performed as appropriate for crosstabulated categorical data. Continuous variables were compared with t tests or between-group analysis of variance. Given low outcome base rates and small sample sizes, logistic models were examined using a penalized log likelihood adjustment for complete or partial separation to correct bias (20, 21). Likelihood ratio chi-square tests were performed to explore differences between nested groups of variables in penalized logistic models, after examining patterns of missing data.