Abstract
This paper analyzes the small sample properties of GMM estimator under the condition that errors are serially correlated. Some of the interesting results are as follows: 1) The standard deviation of parameter estimates tends to decrease as the number of lags used for instruments increases. 2) Tauchen (1986) insisted that there is a variance/bias trade-off regarding the number of lags used for instruments. However, we found that this result does not generally hold true. It depends on the law of motion of exogenous variables. 3) The test of overidentifying restrictions is biased toward accepting the specification of the model. The magnitude of this bias increases as the number of lags used for forming instruments increases. The overidentifying restriction test is not affected by the difference of the law of motion. Thus, it can be said that, since the desirable choice of instruments depends on the specification of exogenous variables, it is generally recommended to check the law of motion of the exogenous variables in advance in applied works.
