Econometrics of nonstationary panel data applied to CEO compensation analysis.

摘要

The first article of this dissertation examines estimation of and inference for the average long run relation parameter proposed by Phillips and Moon (1999a) in unit root nonstationary panels when the cross-sectional dimension n is possibly much larger than the longitudinal dimension T. A new estimator based on cross-sectional pooling of kernel long run covariance matrix estimators of the time series of the panel is proposed. The estimator is consistent and asymptotically normal at rate $n$ in spurious and heterogeneously cointegrated panels and at rate $nT/K$ in homogeneously and near-homogeneously cointegrated panels both under sequential and joint limits, where K is a kernel bandwidth. Biases in the asymptotic distributions can be eliminated if $n$/T → 0. A relatively wide bandwidth is usually needed for the validity of the asymptotic approximation but this seems to have little or no adverse effect on the estimator variance.; The second article uses these tools to evaluate the connection between CEO compensation and company performance. Unit root tests indicate that standard variables in CEO pay-performance sensitivity regressions possess unit roots. Estimated sensitivities of total compensation, which takes into account changes in executive stock and stock option values, to company market value are significantly lower than in the literature. The sensitivity of total compensation to firm accounting performance is of the same order of magnitude as to market performance. Sensitivities of narrower compensation measures to company market and accounting performance are economically negligible. Little support for relative performance evaluation is found and some econometric problems in its estimation are demonstrated. Statistical tests cannot detect inter-industry differences in pay-performance sensitivities.; Phillips and Moon (1999a) discussed linear regressions in unit root nonstationary panels and derived probability limits and limit distributions for estimators and assumed n/T → 0 in joint limit distribution theory. The third article derives joint limit distributions for the pooled least squares estimator in spurious and cointegrated panels when n/T → k ∈ [0, ∞) and shows that although the limit distributions are still normal, there generally appears a bias in the distributions. Allowing n/T → k does not affect the variances of the limit distributions but merely shifts them by finite nonrandom factors that are proportional to $k$.