[Simulation on design-based and model-based methods in descriptive analysis of complex samples].

Area of Science:

• Epidemiology
• Biostatistics
• Survey Methodology

Background:

• Complex survey data often presents challenges in descriptive analysis due to potential biases in sample structure.
• Evaluating statistical methods is crucial for accurate health surveillance and risk factor assessment.

Purpose of the Study:

• To compare the effectiveness of design-based versus model-based methods for descriptive analysis of complex sample data.
• To assess the performance of routine, multi-level model, and design-based methods in estimating population parameters and constructing confidence intervals.

Main Methods:

• Utilized data from the 2010 China chronic disease and risk factors surveillance, involving 1,000 samples with a multistage random sampling design.
• Simulated systematic age structure deviations in samples to evaluate method robustness.
• Calculated Mean Squared Error (MSE) and 95% Confidence Interval (CI) coverage probability for systolic blood pressure (SBP) and raised blood pressure prevalence using design-based, routine, and multi-level model methods.

Main Results:

• The design-based method yielded the lowest Mean Squared Error (MSE) for both SBP (1.38) and prevalence (4.80) estimators.
• Design-based analysis achieved the highest 95% Confidence Interval (CI) coverage probability for both mean SBP (97.5%) and prevalence (97.3%).
• Routine and multi-level models showed higher MSE and lower CI coverage, indicating potential for increased Type I error.

Conclusions:

• The design-based method significantly outperformed routine and multi-level models in both point estimation and confidence interval construction for complex samples.
• Design-based methods are the preferred approach for descriptive statistical analysis of complex survey data, particularly when sample structures are systematically biased.