Spatial patterns in data have become an important feature to consider in the statistical analysis in medical esearch. The analysis of data from observational studies requires methods such as the propensity scores to create a quasi- andomized design. Considering first the case of no variates, we evaluated the ordinary and generalized least squares estimators, and corresponding variance estimators, or an effect of the exposure/treatment. In order to provide robust inference for spatially correlated data, a number of new methods for variance estimation were proposed and compared, including parametric spatial bootstrap, jackknife, and empirical variogram approaches. Next, the use of propensity scores based on spatially and non-patially correlated covariates was studied. Several propensity score and outcome models were evaluated along with the difference approaches to assessing balance and exposure effects. Performance of the methods was evaluated using simulation studies. In addition, the methods were applied to discharge data from California hospitals in2004 taken from the Healthcare Cost and Utilization Project (HCUP) State Inpatient Databases (SID).
According to the simulation results, the ordinary least squares estimator with the empirical variance estimator was recommended in the absence of prior knowledge about underlying variance-covariance structure. Also, the empirical balance score was recommended for assessing the balance of a variable. Moreover, the simulation results showed that the exposure/treatment effect of the outcome model including all observed covariates and propensity scores did reduce the bias. Here, both outcome and propensity score models including all observed covariates were recommended. Applying these models to a real dataset, we found that the mean length of stay and the number of patients who died were lower for high-quality hospitals. In contrast, the median of total charges for low-quality hospitals was smaller.