MM-ML Strategy Classification

mmmlrh4002This post is derived from “Chapter 5, Multiple Measure Strategy Classification-Outcomes, Decision Times, and Confidence Ratings” authored by Andreas Glockner from Foundations for Tracing Intuition– Challenges and Methods, edited by Andreas Glockner and Cilia Witteman 2010 Psychology Press NY. It shines a little light on how intuition experiments seeking to answer if a person is using a Take the Best strategy or a Parallel Constraint Satisfaction strategy, etc are actually done. It is written more understandably than a typical paper for a journal. It will hopefully give more meaning to the letters MM-ML.

One of the core features of intuition is that at least parts of the information integration processes remain opaque to the decider. Intuition researchers cannot use introspection based methods or self-report measures to get a first idea about the processes and to derive and later test hypotheses. Properties of the information integration process underlying intuition have to be derived from readily apparent behavior, physiological reaction or by using neuronal activation measures. Glockner outlines a general methodological approach and a statistical method that permits investigating intuitive as well as deliberate processes based on simple behavioral measures only.

The multiple measure maximum likelihood MM-ML strategy classification method estimates the total maximum likelihood for the observed choice behavior including choices, decision times, and confidence ratings of a participant given the application of a certain intuitive or deliberative strategy. According to Glockner, the resulting likelihoods for different strategies can be compared and people should be classified as users of the strategy with the highest likelihood. Glockner sets up an example to explain how it works.

Experimental task and research question

In Glockner’s example, participants take the role of a purchaser of oranges and have to decide repeatedly from which of the two orange producers to buy. They are provided with predictions of four testers which predict whether the oranges will be good or bad and are explicitly informed about their cue validities. The researcher has to decide which decision strategies are to be compared. A simple deliberate strategy would be the so-called take the best strategy TTB, which suggests that individuals consider the prediction of the most valid cue only. If this cue differentiates the favored option is instantly selected. If this is not the case, the second cue is considered, and so on. Another simple strategy would be to ignore cue validities and to select the option which has the highest sum of cue values without weighting them by validity, a strategy which is called equal weight EQW. The application of a weighted additive strategy WADD in which individual calculate a weighted sum of cue values and select the option with the higher weighted sum would require more effort. One possible intuitive strategy is specified by the Parallel Constraint Satisfaction PCS model, which predicts quick weighted compensatory information integration based on automatic processes of perception.

————————-Oranges A Oranges B
Tester 1 80% valid      +                   –
Tester 2 70% valid      –                   +
Tester 3 60% valid      –                   +
Tester 4 55% valid     +                   –

Based on the table above, TTB would predict choices for option A because the most valid cue speaks for A, and EQW would predict a random selection between A and B because there are two positive cues for each option. The application of a WADD strategy can be assumed to entail a mediation step in which the explicitly provided objective validities are transformed into subjective weights. The simplest transformation would be to use the validities as the weights w=v. According to such a so called ignorant WADD strategy, the weighted sum for options A and B would be .05 and -0.05 and option A should be selected. Note, that the strategy is dubbed ignorant because it does not correct for the fact that in choices between two options cues with a validity of .5 have no predictive power at all since they are equivalent to a coin flip. Adding several non-predictive cues with w=.50 which all speak for option A would overrule all other information and lead to choices for option A only. This problem could be accounted for by correcting validity by subtracting the chance probability .50 (w=v-.50). A third possibility would be to transform cue validities using a non-linear transformation. This could be done using a log-transformed odds-ratio of the validity according to w=ln(v/(l-v))(with 0<v<1). Under the assumption that cues are independent and that the validities represent prior probabilities according to Bayes’ theorem, the application of these log-transformed decision weights in a WADD would lead to the mathematically optimal solution (normative WADD).

Decision time predictions. Since TTB, EQW and WADD are conceptualized as deliberate decision strategies, the number of computational steps (elementary information processes EIP) which are necessary to come to a decision determines decision time. According to Glockner, in the Orange example, TTB should take 4 EIPs (read cue1 info1, read cue1 info2, compare, select A), while EQW would take 16 EIPs and WADD 32 EIPs. Thus, WADD should show the longest decision time.

Confidence predictions. Confidence predictions could also be derived from the strategies. For TTB, confidence in a choice should be equal to the validity of the cue which differentiates between the options. For EQW and WADD confidence should be proportional to the difference between the unweighted or weighted sum of cue values.

Predictions of intuitive strategies
Now the deliberate strategies need to be contrasted with intuitive ones. For example, PCS models postulate that individuals form a mental representation of the decision task which can be modeled by interactive activation networks. The PCS processes operate towards maximizing consistency in the network by changing the activation of nodes. Initial advantages of the favored option are accentuated by highlighting supportive information (cues that speak for the favored option) and by devaluing contrary information (cues supporting the non-favored option). PCS models basically predict weighted compensatory information of cue validities and cue values and thus make essentially the same choice predictions as WADD.  However, for PCS, decision time should decrease with increasing superiority of one option over the other which makes consistency maximizing easier and decision times should be generally short because decisions are based on automatic processing. Furthermore, confidence should increase with increasing superiority of one option over the other.

Selection of diagnostic tasks for the considered set of strategies
To determine which strategy participants are most likely to use, tasks have to be selected for which the strategies make different predictions concerning at least one of the dependent variables: choices, decision time, and/or confidence. There is no formal optimal solution to construct the most appropriate tasks, but basically types of tasks should be selected that allow pitting the specific properties of the strategies against each other (e.g. ignoring cue validities for EQW, ignoring less valid information for TTB, increasing decision time for low consistency for PCS, increasing confidence with increasing weighted difference between options for WADD).

In the end MM-ML uses statistics, that I will ignore here, to allow for differentiating between strategies which make similar choice predictions and within which all dependent variables are considered simultaneously and in a compensatory manner.