20
respondent did not report his or her household income, and both the recoded PCAPPINC
and INCMISS are included in the right-hand side of the model.13
• Other individual characteristics of the respondent, such as education (measured by
a dummy, COLLEGE, denoting whether the respondent has a University degree or
better), age, and marital status (the dummy MARRIED).14
• A dummy denoting whether the respondent finds the site to be in “very good”
condition (GOODSTATE).
Finally, it is important to tackle the issue of substitute sites. Ideally, if substitute sites
exist, the price per trip to a substitute site should be included in the model. Failure to do
so results in a biased estimate of the coefficient on price per trip, the severity of the bias
depending on the correlation between the two price variables. In practice, we do not have
information about which sites, if any, would be considered reasonable substitute for the
study sites. This forces us to omit this variable altogether from the regression model.15
6. Results
A. Actual Trips
Our first order of business is to fit the Poisson equation corrected for the on-site
nature of sample using only the actual trips taken by the respondents. In other words, the
sample is restricted to j=1 and results in a total sample size of 468 (32 observations are
13 The coefficient on INCMISS captures any systematic differences in the number of trips among those
respondents who did and did not report income. The coefficient on PCAPPINC should be interpreted as the
marginal effect of income on trips, conditional on information on income being available.
14 In our initial runs, we experimented with including age squared, household size, and other variables, but
the models were poorly behaved, so we decided to exclude these regressors from the specifications reported
in this document.
15 Similar reasons drove Forrest et al. (2000) and Poor and Smith (2004) to omit the travel cost to a
substitute site in their applications of the travel cost method.