for systematic heterogeneity in response to changes in compliance costs improves the
fit of the model.26
Random parameter logit results
Several different specifications of the RPL model were tested. The best results
were obtained when all cost coefficients are allows to vary randomly. In the RPL
model presented in Table 3, the estimated standard deviations of all but one of the
random coefficients are all highly significant, indicating that these parameters do
vary across managers, even after allowing for observed, systematic variation across
electricity market types and plant vintages. The results of a nested likelihood ratio
test imply that allowing for response heterogeneity dramatically improves the fit of
the model. These RPL estimation results are robust to various optimization routines
and initial starting values.
In the RPL model, unobserved variation is decomposed into an extreme value
stochastic term and variance of the random parameters. In the CL models, all unob-
served variation in anticipated costs is captured by the extreme value stochastic term.
Consequently, normalizing coefficients by the variance of the extreme value compo-
nent of the disturbance term will make RPL parameters larger in absolute value. The
significant increase in the magnitude of the cost coefficient estimates suggests that
the variation in random parameters constitutes a significant portion of the variance
in (unobserved) perceived compliance costs. Conversely, the technology specific fixed
effects get smaller in absolute value, and some cease to be significant. This suggests
that the statistical significance of these fixed effects in the CL specifications was partly
due to random response heterogeneity to variations in costs.
All of the cost coefficients are assumed to be normally distributed.27 The means
of both the variable cost coefficient and the variable cost∕restructured market inter-
action term are negative and significant at the 1% level. The estimated standard
23