After having read David Hardman's comments on my previous entry
(see here), some things were left dingling in the air with little (to no) explanation.
Regarding the 95% accuracy, he is absolutely right pointing out that it does not refer to "correct" decisions, but to the fact that the Matching Heuristics model predicted accurately the decision bail/no bail on a 95% of the cases. Of course, on a decision like this there is no "correct" or "incorrect" decision, since it is, firstly, impossible to predict what would have happened if the decision was the opposite. Secondly, there is no "right" or "wrong" answer, since there is no established evaluation process. It can, however, be "popular" or "impopular", as in how it is regarded from the public's point of view.
Also, David mentions that I have not defined what the Matching Heuristic model actually is, which is the (main) point being adressed in this post.
The Matching Heuristic model is a non-compensatory model, which means that not all cues being considered (or available) to undertake the decision weight the same (i.e. the colour of the new car you want to buy is less important than the make, for example, whilst still keeping the price range the top priority (the one that weights more) on deciding which car to buy).
The objective of a non-compensatory model is to be able to base a decision on a minimal, small subset of possible cues, whilst still being able to provide a reasoned, based explanation.
For example, on Dhami & Ayton (1999), the matching heuristic model used to predict bail/no bail decisions (or punitive/non-punitive, respectively) used 2 cues, which were subdivided into two sets of questions:
1: Does the 1st cue give a reason for being punitive? If so, predict a punitive decision. Otherwise, go to question 2.
2: Does the 2nd cue give a reason for being punitive? If so, predict a punitive decision, otherwise, predict a non-punitive decision.
Fair enough, it uses plenty of weasel words, like "cue", without defining them. Specifically, they suggest this model for k=2 different cues taken into account. On their main experiment, they found that, for people using the matching heuristics model, the maximum amount of cues taken was k=3, so the more specific questions came out:
1: Did prosecution request conditional bail or oppose bail? If yes, then predict punitive decision. Otherwise (or no information), go to q2.
2: Did previous court impose conditions or remand in custody? If yes, predict punitive decision. Otherwise (or no information again), go to q3.
3: Did police impose conditions or remand in custody? If yes, punitive, otherwise (or no info), predict non punitive decision.
It is interesting how this model ressembles those "find your own adventure" books, in which you had to keep turning pages based on x or y decision at the end of each page. Has the prosecution requested bail? Then, punitive decision! This model is the ONE mentioned on the earliest post which provides 95% accuracy PREDICTING (I got it right this time!) the magistrates' decisions.
Some examples can be made for effect.
1st time offender, non violent, petty crime. Passes question 1 OK, question 3 OK (both for non violent crime, + petty crime), AND it's the first offence (so no chance of previous court), thus passing question 2 as well (order would still be 1-2-3). So, this offender goes on bail very nicely.
1st time offender, petty crime, violent to the police when arrested (aggressive drunk behaviour, for example?)
Question 1 goes as "pass", or maybe "no info". First time offender (might have been caught before but no previous conviction or first time through court), so question 2 goes as "ok". Question 3, however, might impose some differences, since the offender is likely to drink again, become very violent again, and probably offend again, thus police asks for his/her remand in custody. Question 3 fails, thus magistrate decides to give punitive action.
Seems reasonable enough, I'd say?
