Closed drbenvincent closed 6 years ago
Updated position of red line in lower right plot to reflect new control model in #185. Here the observed frequency of choosing delayed option was ~40%. In terms of predicting responses, the control model would then predict all responses to be immediate, as this is under the 50% decision threshold. This can then explain 100-40% =60% of all responses, and the red line now reflects this.
We also have the response data and posterior predictive checks (discount functions) plotted so we can do a good job of model checking just by looking at this figure alone.
So we now calculate the (distribution of) Log Loss goodness of fit metric. The point estimates are also exported in the .csv
file of parameter estimates.
There was some unfinished business from #185.
The "percent predicted" subplot
Fix posterior predictive check warnings
Need to update the flagging of problematic experiments in
ResultsExporter.any_percent_predicted_warnings()
.PosteriorPrediction
.Add discount function subplot to the posterior prediction plots
Replace "goodness of fit" score with "log loss"
Rather than use my ad hoc goodness of fit score, calculate and report the log loss score. This is a cross entropy measure appropriate for binary outcome variables. See more about log loss here. Note that this is just a goodness of fit metric, and is not complexity penalised, so it not appropriate for model comparison between models with varying numbers of parameters.
Lower values of log loss correspond to better fits, ie more correct classifications of participant responses.