INSTRUMENT STRENGTH AND PERFORMANCE OF A PARAMETRIC LATENT INDEX MODEL IN ESTIMATING AVERAGE TREATMENT EFFECT (ATE)
Author(s)
Wu EQ1, Johnson KA2, Nichol MB2, 1Analysis Group/Economics, Boston, MA, USA; 2University of Southern California, Los Angeles, CA, USA
Presentation Documents
OBJECTIVES: Testing the sensitivity of switching regression model's estimator of ATE to the joint normality assumption of error term distribution when there is selection bias; assessing the relationship between the sensitivity and the instrument strength. METHODS: We used the Monte Carlo method to generate data with selection bias under four different distributions: joint normal distribution, joint t distribution (df=3), a joint right skewed distribution, and a joint symmetric two-peak distribution. The parameters in the simulation were calibrated using an asthma patient population selected from Medi-Cal eligibles between January 1995 and December 2000. The simulated outcome variable can be considered as log-cost or any continuous outcome variable in healthcare. We varied the strength of the simulated instrumental variable. For each instrument strength, 500 data sets were generated each with 1500 observations. We measured the sensitivity of switching regression model's estimator of ATE to the joint normality assumption in terms of root mean square relative error (RMSRE) and relative bias. RESULTS: For data simulated with error terms from all of the four distributions, RMSRE of the ATE estimates converge and decrease as instrument strength increases. When the instrument strength equals 85%, the ATE estimate for data simulated from the joint t error term distribution has the highest RMSRE at 48%, which is close to the smallest RMSRE value at 40% for data simulated from the joint normal distribution. The relative bias also converges as the instrument strength increases. CONCLUSIONS: With a strong instrument, the parametric switching regression model was robust. This result justifies the use of the parametric switching regression model to adjust for selection bias with a strong instrument. The sensitivity of the switching regression model with only weak instrument(s) results mostly from the lack of information to identify the treatment effects, rather than from the erroneous assumption of error term distribution.
Conference/Value in Health Info
2003-05, ISPOR 2003, Arlington, VA, USA
Value in Health, Vol. 6, No. 3 (May/June 2003)
Code
PMD3
Topic
Methodological & Statistical Research
Topic Subcategory
Modeling and simulation
Disease
Multiple Diseases