Microeconomic simulation output analysis using a two-way random effects metamodel.

Microeconomic simulation models are widely used by policy analysis agencies to explore the impact of social welfare and tax programs. The National Research Council's report strongly recommended including information on the sources and levels of uncertainty in estimates based on model results. Those concerns are directly addressed by this work, which develops statistical methodology for estimating the expected value of the output variable from a random replication of the simulation model. An example model of annual earnings of married women is used throughout to demonstrate the application of the methodology. This methodology is useful in all aspects of simulation model output analysis: model validation, sensitivity analysis, and policy exploration. A microsimulation model contains a sample of microeconomic decision units with heterogeneous demographic characteristics; the value of each decision unit's response variable is stochastically determined as a function of the macroeconomic environment. Multiple, independent, Monte Carlo replications of the simulation model are made from the same initial state. The output from a simulation experiment consists of a matrix of observed values for the dependent variable; each column in the matrix represents a replication of the experiment and each row represents a decision unit. The two-way random effects model is a useful metamodel for microsimulation output analysis since it facilitates identifying, separating and evaluating the sources of uncertainty inherent in the output; decision unit sampling error is represented by the row effects variance, replication sampling error is represented by the column effects variance, and modeling error is represented by the disturbance term variance. This work develops a posterior distribution for the mean of the observations from a randomly selected replication of the simulation model. The likelihood function is derived from the two-way random effects model. Conjugate, informative prior distributions are used. Analytic-numeric techniques, based on LaPlace's method for integral approximation, are developed to evaluate the intractable integrals over the variance parameters. Comparisons between the Bayesian and sampling theory analyses, using output from the example model, demonstrate the differences in inferential results which may occur under the two approaches.PhDBusiness AdministrationUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/103218/1/9308319.pdfDescription of 9308319.pdf : Restricted to UM users only