Multiple linear regression analysis methods for complex random sampled data and their application / 军事医学

Objective To study the significance and function of the comprehensive weight in multiple linear regression analysis of complex random sampled data .Methods Based on the concept of Monte Carlo random simulation , two different multiple linear regression analysis procedures in SAS-REG and SURVEYREG were used to perform regression modeling for the same batch of complex random sampled data ( n=6756 ) at different random sampling proportions .The results were compared.Results In the results of the fitting multiple linear regression model when observation weight and sampling weight were considered or not , it was found that the size of the partial regression coefficient , standard error and P value of independent variables varied .Conclusion In complex random sampled data based on different proportions ,especially in regression modeling of stratified random sampling survey information , the multiple linear regression model makes it possible to more accurately and sensitively perform parameter estimates of regression coefficients and statistical prediction of outcome variables if the comprehensive weight of the survey data is incorporated into the statistical analysis .

Multiple Indicator Cluster Survey 2003 in Afghanistan: Outdated Sampling Frame and the Effect of Sampling Weights on Estimates of Maternal and Child Health Coverage.

Due to an urgent need for information on the coverage of health service for women and children after the fall of Taliban regime in Afghanistan, a multiple indicator cluster survey (MICS) was conducted in 2003 using the outdated 1979 census as the sampling frame. When 2004 pre-census data became available, population- sampling weights were generated based on the survey-sampling scheme. Using these weights, the population estimates for seven maternal and child healthcare-coverage indicators were generated and compared with the unweighted MICS 2003 estimates. The use of sample weights provided unbiased estimates of population parameters. Results of the comparison of weighted and unweighted estimates showed some wide differences for individual provincial estimates and confidence intervals. However, the mean, median and absolute mean of the differences between weighted and unweighted estimates and their confidence intervals were close to zero for all indicators at the national level. Ranking of the five highest and the five lowest provinces on weighted and unweighted estimates also yielded similar results. The general consistency of results suggests that outdated sampling frames can be appropriate for use in similar situations to obtain initial estimates from household surveys to guide policy and programming directions. However, the power to detect change from these estimates is lower than originally planned, requiring a greater tolerance for error when the data are used as a baseline for evaluation. The generalizability of using outdated sampling frames in similar settings is qualified by the specific characteristics of the MICS 2003—low replacement rate of clusters and zero probability of inclusion of clusters created after the 1979 census.