The end of follow-up for analysis was May 24, 2010. Follow-up time for each participant was calculated from date of in-home visit to date of first stroke, death, or last telephone follow-up. Stroke incidence rates were calculated as number of incident events (WHO or clinical or probable stroke) divided by person-years at risk within each age-race-sex-region stratum, with 95% CIs calculated assuming a Poisson distribution. Adjusted rates were standardized to the 2000 U.S. population. Incidence rate ratios (IRRs) were calculated for stroke belt and stroke buckle relative to the rest of the country, and for blacks relative to whites within each age stratum.
To maximize the number of events included, we selected a recent date for end of follow-up that was close to the time of analysis. Thus, not all potential events were fully processed. Additionally, we were unable to retrieve records for some suspected events. Exclusion of these incompletely processed potential events would reduce estimates of event rates and could bias results. The outcome of these potential events for which the adjudication process was not completed are simply missing data, in these the use of multiple imputation has been encouraged as an approach to reduce potential biases and improve precision.18 (link) We applied multiple imputation techniques to classify potential stroke events still in process, using a logistic function predicting the likelihood that an attempted record retrieval would result in an adjudicated stroke.19 This model included as predictors race, region, and reason prompting record retrieval attempt; no other available factors were associated with likelihood of adjudication as stroke. Ten datasets including imputed outcomes were generated, with estimates of incidence rates and IRRs produced by the SAS procedure MIANALYZE. Details of this approach are available elsewhere.19 Analyses were carried out in SAS version 9.02 (SAS Institute, Cary, NC).
To maximize the number of events included, we selected a recent date for end of follow-up that was close to the time of analysis. Thus, not all potential events were fully processed. Additionally, we were unable to retrieve records for some suspected events. Exclusion of these incompletely processed potential events would reduce estimates of event rates and could bias results. The outcome of these potential events for which the adjudication process was not completed are simply missing data, in these the use of multiple imputation has been encouraged as an approach to reduce potential biases and improve precision.18 (link) We applied multiple imputation techniques to classify potential stroke events still in process, using a logistic function predicting the likelihood that an attempted record retrieval would result in an adjudicated stroke.19 This model included as predictors race, region, and reason prompting record retrieval attempt; no other available factors were associated with likelihood of adjudication as stroke. Ten datasets including imputed outcomes were generated, with estimates of incidence rates and IRRs produced by the SAS procedure MIANALYZE. Details of this approach are available elsewhere.19 Analyses were carried out in SAS version 9.02 (SAS Institute, Cary, NC).
