My answer to What is the best data mining method to predict dementia?
Answer by Connie b. Dellobuono:
Health data on Diabetes and Alzheimer’s
Diabetes Mellitus and Risk of Alzheimer's Disease and Dementia with Stroke in a Multiethnic Cohort
Data analysis
Prevalences of diabetes and other covariates were compared between subjects with and without Alzheimer's disease and between subjects with and without stroke-associated dementia. Continuous variables were compared by analysis of variance, and categorical variables were compared by χ test. Cox proportional hazards modeling was used for multivariate analyses. The time-to-event variable was age at onset of dementia; the models were stratified by year of entry into the cohort in order to control for period effects, as recommended for longitudinal studies. There was one model for each outcome mentioned above. All covariates were treated as time-constant covariates using the baseline values. In 26 of the 255 subjects with diabetes, the diagnosis was made after baseline, but these persons were treated as having had diabetes at baseline. An additional analysis was carried out treating diabetes as a time-dependent covariate taking into account the date of reporting of the diabetes diagnosis; this analysis was conducted to examine how the definition of diabetes (baseline vs. follow-up) affected the analysis. A similar analysis was carried out treating all variables as time-dependent covariates with the beginning of exposure used as the beginning of observation (or later for the 26 subjects diagnosed with diabetes after baseline), to compare the results with the time-constant covariate model. Subjects without the outcome were censored at the time of the last follow-up visit. Subjects with a type of dementia different than the one considered in the specific model were censored at the time of onset of dementia. For example, when Alzheimer's disease was examined as the outcome, persons with stroke-associated dementia were censored at the time of dementia onset. Additional analyses were performed using nondementia cognitive impairment without stroke and nondementia cognitive impairment with stroke as the outcomes; persons with nondementia cognitive impairment at baseline were excluded.
The population attributable risk (PAR) for diabetes in relation to dementia was calculated for each ethnic group using the formula PAR = Pr(HR − 1)/1 + Pr(HR − 1), where HR is the adjusted hazard ratio obtained from the multivariate models and Pr is the prevalence of diabetes in each ethnic group in the cohort; 95 percent confidence intervals were calculated for the population attributable risk using methods described for prospective studies. SAS for Windows, version 7 (SAS Institute, Inc., Cary, North Carolina), was used for all analyses.