Chinese Journal of Sociology ›› 2017, Vol. 37 ›› Issue (1): 186-210.

Previous Articles     Next Articles

Uncertainty of Statistical Models and Propensity Score Methods

HU Anning   

  1. Department of Sociology, Fudan University
  • Online:2017-01-20 Published:2017-01-20
  • Supported by:

    This study was supported by the National Social Science Foundation (15CSH030), the Innovation Program of Shanghai Municipal Education Commission (15ZS001), and the Outstanding Scholar (Zhuoxue） Grant from Fudan University

PDF

919

Abstract:

Quantitative sociological research has always employed certain specific statistical models. Over the past several decades, the focus on causal relationship in sociological studies has led to a wide spread application of propensity score methods.Using an explicit estimation of the probability of being subject to a specific treatment or intervention, sociologists are able to mimic random experiments to predict causal effects. In practice, propensity score methods require an estimation from two models:one predicts propensity scores and the other estimates causal effects. However, the model structure and coefficient of both contain considerable uncertainty. This study offers a systematic review of the model structure and coefficient uncertainty in propensity score methods as well ascertain strategies to tackle the issue. By analyzing Monte Carlo's simulated data along with empirical survey statistics, the paper demonstrates how researchers can use Bayesian Model Averaging to select multiple backup models and deal with possible model-coefficient uncertainty with the joint maximum likelihood estimation in propensity score methods. The paper also finds that after taking into account of various sources of uncertainty,the estimated causal effects display a narrower confidence interval but a higher level of statistical efficiency.

Key words: Propensity Score Method, Statistical Efficiency, Model Coefficient Uncertainty, Model Form Uncertainty, Bayesian Averaging