Statistical Analyses
Dichotomous and categorical variables were summarized by percentages and continuous
variables were summarized by means (standard deviations) for all variables. At baseline,
9.5% of participants had one or more missing values for covariates. For the multivariable
analyses, missing covariate values were replaced with those generated using the multiple
imputation (MI) procedure in SAS® software (Cary, NC). In MI, missing values of variables
are simultaneously predicted using existing values of variables by modeling the joint
distribution of all the covariates plus selected other predictors. Conditional on the
nonmissing values for each individual, the missing values have a distribution from which
several joint random samples are drawn. Each imputation dataset is analyzed separately as if
there were no missing values, then the results are combined in a manner that reflects the
uncertainty due to missing values. The results from five imputation datasets were combined
to obtain regression coefficient estimates and confidence intervals (CIs). Bivariable logistic
regression was conducted to examine any associations between current (new and continuing
combined) and past medication use and incontinence. Multivariable logistic regression analyses using a backward selection approach (alpha=0.15) identified those demographic,
health behaviors and health status factors to be added as covariates along with the seven
drug use independent variables in the final model. This final model calculated adjusted odds
ratios and 95% confidence intervals relating urinary incontinence (yes/no) in the previous 12
months before year 4 (three years from baseline) with specific classes of medication use at
year 3 (two years from baseline).21 All other variables were fixed at baseline values except
bladder antispasmodics, number of prescription drugs and persistent lower extremity
limitation which reflect data from year 3. We also tested all two way interactions between
race and the various forms of current medication use with potential urological activity.
Underlying statistical assumptions were evaluated and verified. 21 All statistical analyses
were conducted using SAS® Version 9.1.