#JSM2018 Raghu Talking about using Multiple Imputation for sample design. Cites @jameswagner254 2010 Stats in Medicine paper.
#JSM2018 Raghu Have benchmark data from large national survey or admin data. Have population frame with sample units. Draw replicate sample from frame such that distribution is similar to population
#JSM2018 Raghu Impute values for survey data and admin data for unsampled units. Then use info from this to adjust selection probabilities for next replicate. Keep going until distribution in sampled data looks like population.
#JSM2018 Raghu Simulations used three different nonresponse mechanisms, and two different sampling schemes (fixed and adaptive), with different correlation structures between auxiliary variables and survey data
#JSM2018 Fixed sampling scheme gives biased estimate, whereas adaptive design gives estimate close to truth. Bias for adaptive samples reduced most when MAR on auxiliary variables has strong correlation b/w auxiliary and survey variables
#JSM2018 Raghu Real Data simulation (note all correlations much lower here than in simulation). 
#JSM2018 Raghu This provides useful framework for adaptive sampling. Potential cost reduction here for screening surveys.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
