Postscript : Rejoinder to Brandstätter , Gigerenzer , and Hertwig ( 2008 ) Eric

vector 0 and variance terms xy. Because the dependent measures, time and frequency, are skewed, they were log-transformed to make them more symmetrically distributed. We note that because of the discrete nature of the frequency data, a linear regression might not be an appropriate model. We therefore also estimated a nonlinear Poisson regression model. However, because using a Poisson specification yielded similar results to the ones obtained under the linear model of the log-transformed data, we do not discuss the results of this analysis in detail. For transitions, we modeled the relevant transitions between the eight boxes using a similar model, which has 56 observations per gamble per participant due to two factors: a transition identifier (28 levels) and phase. Again, we used an intercept, phase and number of reasons as random factors. In addition, as described in the text, we estimated a model that allowed the impact of kind of transition to vary across respondents, which resulted in improved fit and substantively interesting interpretations. We used two-tailed tests throughout, both because this analysis has substantial power and because prior data suggests results that conflict with the predictions of the priority heuristic.