Methods

Tracing participants
Response rates for the baseline surveys were generally high, ranging from 74 to 98%. For the follow-up, which was pre-planned at baseline, we have collected the names, addresses and telephone numbers of three non co-resident local relatives, friends or neighbours who can be contacted to help trace the participant if they are not at their original address. The population in each of the settings where we are working is stable. Apart from a likely 6% per annum attrition from mortality, we anticipate a high follow-up rate.

Protocol
The interviews for the incidence phase largely involve a repeat of the one phase dementia diagnostic protocol previously described, with the aim of identifying incident dementia. We will modify the DSMIV algorithm to incorporate cognitive test and informant data from baseline into the direct determination of cognitive and functional decline. We will also reassess risk exposures that may have changed over the three years since baseline. All participants will receive the same comprehensive assessment, very similar to that administered in the baseline survey, regardless of their baseline status. Those with dementia at baseline will have been at risk for stroke, and we will repeat dementia assessments to assess course and outcome and to validate baseline diagnoses. We will also review care arrangements, dependency and indices of caregiver strain.

New elements in the incidence phase protocol
Mortality
For those who died between baseline and follow-up (information ascertained on follow-up visit) we will complete a verbal autopsy interview with a co-resident, relative or other person well-placed to know the circumstances of death, using methods developed and validated by the 'Million Deaths' project for use in India (1,2) to identify underlying cause according to ICD10 criteria. The interview takes 30-45 minutes using structured questions and an open-ended narrative, with a symptom list to assist attribution. Cause of death is allocated by the consensus judgement of two physicians. We will also attempt to ascertain whether the participant had an onset of dementia before death, using the standard 10/66 informant assessment.

Stroke
We also include a two phase clinical protocol designed to identify incident stroke. All those claiming (by participant or informant report) to have experienced a stroke in the interval between assessments, and all those with suggestive neurological signs (asymmetric long tract signs, dysphasia, marked gait disturbance) not apparent at baseline will be offered physician assessment including physical examination, clinical history, and examination of clinical notes and investigations where available ('cold pursuit'). We will then seek consensus diagnosis from two local independent experts. Stroke diagnosis (ARIC criteria (3)) require

1. evidence of sudden or rapid onset of neurological symptoms,
2. lasting for more than 24 hours or leading to death,
3. in the absence of evidence for a nonstroke cause (brain trauma, neoplasm, coma attributable to metabolic disorders or disorders of fluid or electrolyte balance, CNS vasculitis or infection, and peripheral neuropathy).

In our baseline data set 96% of those who reported a stroke had a clinical diagnosis, 81% from a specialist. However, neuroimaging will not routinely be available, and we cannot be confident that we will have enough information to attribute stroke sub-type with confidence. Silent brain infarctions (picked up by neuroimaging without clinical symptoms and signs) will not be included.

New retrospective exposure data (not collected at baseline)

• Number of remaining teeth, use and quality of dentures
• Use of lipid-lowering agents
• Calf circumference
• Weight (using digital scales)
• Interviewer assessment of mobility (bedbound/chairbound/housebound/limited mobility outside/unrestricted mobility

New biochemical assays
Fresh frozen serum from baseline collections are available in Cuba, DR, Venezuela, Peru and Argentina. We will use these to investigate associations between micronutrient deficiency and dementia/ AD. We will use a nested incident case control design whereby for each incident case, we shall select three controls at random from among those matching for age, gender and education, who were free of dementia at the time of onset for the case. Assuming 219 incident cases in these six centres (see below) this will imply 876 samples to be assayed for Vitamin B12, folate, TSH and T4.

Analyses
Description of incidence
Person-years at risk will be calculated as the interval between baseline and follow-up assessment, or the estimated time of onset of dementia, or the time of death, whichever occurs sooner. Age-specific incidence (with Poisson standard errors and 95% confidence intervals) will be estimated for each country, by gender and age in 5-year bands by dividing number of cases by number of person-years contributed in each age band. Dementia onset is assumed to be the midpoint between the last date when known to be dementia free and the first date of dementia diagnosis (either by survey ascertainment or clinical information). A similar procedure will be used to calculate stroke attack rates.

Hypothesis testing
We will use Cox's proportional hazards regression throughout to estimate risk associations. Informative censoring may occur through the competing risks of participants dying and becoming lost to follow-up in the interval between assessments. This problem is only rarely addressed in dementia cohort studies (4). We will do so firstly by verbal autopsy interviews with key informants for all deceased persons, and secondly (in a sensitivity analysis) by using proportional subdistribution hazards regression (5) to account for informative censoring through explicit analysis of competing risks.

The ADMIXMAP program (6) will be used to model genetic admixture. In two-step analysis, estimates of individual admixture generated by ADMIXMAP are simply plugged into standard programs for statistical analysis. One-step analysis fits regression models for the effect of individual admixture and other risk factors (genetic or environmental) on disease risk. This allows for the uncertainty in estimation of individual admixture from the marker data, eliminating residual confounding by population stratification that may occur in a two-stage analysis using a small number of markers, and allows us to model haplotypes given unphased genotype data. For this study, ADMIXMAP will be extended to support Cox regression (regression algorithms are based on a GLM approach). We shall test for

1. effects of individual admixture on dementia, AD and other outcomes
2. effects of individual admixture on the slope of relationship of APOE to dementia
3. effects of other environmental factors or genetic polymorphisms on dementia, again controlling for population stratification.

1. Jha P, Gajalakshmi V, Gupta PC, et al. Prospective Study of One Million Deaths in India: Rationale, Design, and Validation Results. PLoS Med 2005 Dec 20;3:e18.
2. Gajalakshmi V, Peto R. Verbal autopsy of 80,000 adult deaths in Tamilnadu, South India. BMC Public Health 2004 Oct 15;4:47.
3. Toole JF, Lefkowitz DS, Chambless LE, Wijnberg L, Paton CC, Heiss G. Self-reported transient ischemic attack and stroke symptoms: methods and baseline prevalence. The ARIC Study, 1987-1989. Am J Epidemiol 1996 Nov 1;144:849-856.
4. Bursi F, Rocca WA, Killian JM, et al. Heart Disease and Dementia: A Population-based Study. Am J Epidemiol 2006 Jan 15;163:135-141.
5. Fine JP, Gray RJ. A proportional hazards model for the subdistribution of a competing risk., 94 ed 1999:496-509.
6. The ADMIXMAP program http://www.ucd.ie/genepi/software 2007.