All analyses were conducted in SAS version 9.3, Stata version 14, or SUDAAN version 11, and utilized procedures capable of incorporating survey weights and design variables into variance estimation algorithms. Full population prevalence estimates were calculated for each administration of each survey and analyzed graphically. To characterize time-trends for both the full population and for demographic subpopulations within each survey, we used log-binomial regression to model prevalence as a function of year. This yields an estimate of the relative risk (RR) associated with year, which is related to the average annual percentage change in prevalence, calculated as 100*(RR-1). For example, an RR of 1.02 indicates that the relative increase in prevalence is 2% per year. To summarize overall and demographic trend analyses graphically, we utilized forest plots of the estimated trend regression coefficients (β) from each survey. The regression coefficient is the natural logarithm of the RR; for small values of β (e.g., −.09 to 0.09), the RR can be approximated by 1+β with accuracy to the second decimal place and the average annual percentage change is 100*β. We sometimes
