used, a negative bias of order 1/T is introduced in the coefficient on the lagged dependent
variable (Nickell, 1981). OLS estimation of the model in first differences partially
corrects the bias, but does not entirely alleviate the endogeneity of the lagged dependent
variable. Arellano and Bover (1995) and Blundell and Bond (1998) derive a generalized
methods of moments estimator (known as system GMM) that simultaneously estimates
the model in levels and first differences. Blundell and Bond (1998) perform Monte Carlo
simulations that demonstrate the system GMM estimator is superior to both the OLS
fixed effects and GMM estimations using first differences only. Further lagged values of
the levels and first difference of the dependent variables are used as instruments for the
lagged dependent variable.22 The system GMM estimator is appropriate when the
coefficient on the lagged dependent variable is 0.8 or greater. For this estimator to be
valid, the lagged dependent variable must have a constant correlation with the state
effects and be uncorrelated with present and past values of the error term. Further, it is
assumed that the error terms have a mean of zero and are not serially correlated. Robust
standard errors that are consistent in the presence of heteroskedasticity and
autocorrelation within states are used in calculating t-statistics.
Data Used
Summary statistics of the data used are provided in Table 2-1. The sample is
comprised of quarterly data. The date in which a state enters the sample is determined by
the date the utility commission in that state and the corresponding leader state ordered its
initial UNE rates. That date ranges from April 1997 to October 1997. Data for all states
22 In the estimations that follow, the previous four quarters of the level of the lagged dependent variable are
used as instruments in the first differences equation, while one lag of the first difference of the lagged
dependent variable is used as an instrument in the levels equation.