JMASM27: An Algorithm for Implementing Gibbs Sampling for ...
Developing a dynamic sampling algorithm for cohort studies
description
Transcript of Developing a dynamic sampling algorithm for cohort studies
Developing a dynamic sampling algorithm for cohort studies
M.H.P. HofA.C.J. RavelliM.B. SnijderK. StronksA.H. Zwinderman
Setting Increasing number of non-Dutch inhabitants
Welfare, health, and illness varies between different ethnic groups Why?
Unclear whether current healthcare and treatment (mainly based on the Dutch Caucasian population) guidelines can be used
Source: O+S Amsterdam blabla
Setting
HELIUS(HEalthy Life in an Urban Setting) Study Large multi-ethnic cohort study among
Moroccan, Surinamese (-Creole and –Hindustani) Turkish, West-African Dutch/Caucasian
Group size ± 10,000 individuals Participants will undergo extensive interviews, medical
investigations, and biomaterial will be collected. Recruitment period: ± 1 year
Problem Definition
High generalizability Representativeness Sample size
Recruitment period of great importance Sampling Design
Current Sampling Designs
(Restricted) randomized sampling Double stage sampling
Stage 1: Sample a large group and obtain distributions of characteristics
Stage 2: Use stratified randomization with stage 1 results
Current Sampling Designs
Problems: Expensive Non-response
differences in subgroups undetected
Limited number of strata possible
Results are very depended on pre-assumptions
Stepwise Sampling Algorithm
Development of stepwise sampling algorithm Actively invite participants with certain characteristics
Minimize difference population and sample
HELIUS study focusses on representativeness on 4 categorized variables Known for each individual
x1 = Age (4 categories) x2 = Gender (2 categories)
Unknown for each individual x3 = Household situation (7 categories) x4 = Income (5 categories)
Stepwise Sampling Algorithm
Problems of active selection Joint distribution of population
composition f(x1, x2, x3, x4) unavailable Estimation of population composition
Prior knowledge: f(x1 * x2) f(x3) * f(x4) Without Prior knowledge Updated with sample composition
f(x1, x2, x3, x4)
Individuals could only be selected on x1 and x2
x1 = Age x2 = Genderx3 = Household situationx4 = Income
Stepwise Sampling Algorithm
Recruitment period has n iterations Each iteration:
Individuals were invited with optimal characteristics f(x1 , x2) and estimated f(x3) and f(x4) Minimizing differences between
sample- and estimated population-composition
Weighted for response and participation chance
Population Estimation was updated with f(x1, x2, x3, x4) from the sample
x1 = Age x2 = Genderx3 = Household situationx4 = Income
Stepwise Sampling Algorithm
Hypothesis:
Random Sampling Stepwise Sampling
Simulation Setting Stepwise Sampling Algorithm versus Random sampling
(With prior knowledge)(Without prior knowledge)
Recruitment period consists of 50 iterations and a sample size of 10,000 per ethnic group is desired
Population O+S Research and Statistics Amsterdam Data from 2009 Five ethnic groups
Dutch (Largest) Surinamese Moroccan Turkish Antillean (Smallest) .
Response rates varying between all characteristics Invited persons responded and participated one iteration later Non-responders were sent a reminder once Performance measured by
Representativeness and Compared to Sample Size
Stepwise Sampling Algorithm Characteristics
Results Simulation
Discussion
Stepwise Sampling Algorithm Strengths
Non-response adjustment Better representativeness and sample size Large number of characteristics representative Less depended on prior knowledge
Weakness High burden of registration during recruitment No increase in representativeness of individually
unknown characteristics
Conclusion
The Stepwise Sampling Algorithm outperforms Random Sampling on representativeness
Questions