Pustaka Ilmiah Universitas Padjadjaran

Abstrak

Bootstrapping For A Structural Equation Model With A Nearly Non-positive Definite Fitted Covaraince Matrix

Structural equation modeling (SEM) has been widely adopted for measuring causal relationship. Instead of fitting individual observations as in regression analysis, SEM works in a different way by fitting a sample covariance matrix to a model implied covariance matrix and producing a fitted covariance matrix. One problem emerges when a fitted covariance matrix is nearly non-positive definite. The estimated maximum likelihood standard errors in the model become implausibly large which in turn jeopardize inferences. In this paper, we show that bootstrapping can overcome this problem. We compare standard errors from two equivalent SEM models, i.e. a discrete and a continuous time hedonic price autoregression panel model. We find out that the bootstrap standard errors obtained from the continuous time model , which suffers from nearly non-positive definite fitted covariance matrix problem, are comparable to the maximum likelihood standard errors obtained from the discrete time model which is free from the problem.