Preference Estimation

In general datasets Prest estimates which deterministic model/heuristic in its toolkit is the best match for each agent.

It does so by identifying “how far” each model/heuristic is from explaining that person’s choices.

In this process, Prest also recovers the agent’s preferences conditional on this best-matching model/heuristic.

A model or heuristic’s proximity to an agent’s data is captured by its distance score.

This is the number of observations that need to be removed from an agent’s data in order for the remaining observations to be fully compatible with the model in question.

Prest also provides information about the compatible instances of every model that is optimal in the sense of having a minimum distance score. [1]

Note

Since the release of Prest 1.1.0 users can also visualize this preference-estimation output and save it in .png format. This service builds on GraphViz and requires GraphViz to be installed and its ‘dot’ binary file to be available in Prest’s root directory. Users can then view and save the directed preference graph that corresponds to some instance of a model used in estimation by right-clicking and viewing the relevant model-estimation dataset in the workspace, locating the subject, model and instance of interest from the relevant list, and clicking on the question mark icon that appears next to the code label of that instance.

These estimation features allow users to test for the proximity of choice behaviour not only with the textbook model of rational choice/utility maximization but also with several models/heuristics that explain well-documented behavioural phenomena such as context-dependent choices, cyclic choices, status-quo biased choices, choice deferral and choice overload.

Note

By default, Prest’s core program is designed to utilise all available computing power by simultaneously engaging all CPU cores. To change that, check the “Disable parallelism” box at the bottom of the “Model estimation” window.

Note

If your dataset includes observations where the deferral/outside option was chosen by some agent(s) and you wish to ignore these observations, you can do so by checking the “Disregard deferrals” box at the bottom of the “Model estimation” window.

Note

If your dataset includes observations where multiple items were chosen by some agent(s) in some menu(s), you can do model and preference estimation either by selecting the “Houtman-Maks” (default selection) or the “Jaccard-Houtman-Maks” method from the “Distance score” scroll-down menu on the bottom left of the “Model estimation” window. The two methods coincide when at most one alternative is chosen at every menu in the dataset. When this is not the case, the second method is less punitive in its computation of the distance score by accounting for the Jaccard [2] (dis)similarity between the agent’s actual choices at a menu and what would have been the model-optimal choices at that menu.

Footnotes

[1]

The concept of a distance score as described above was introduced by Houtman and Maks (1985) for the model of rational choice / utility maximization with budgetary choice datasets. It was extended in Costa-Gomes, Cueva, Gerasimou, and Tejiščák (2022) to other models, on general choice datasets.

[2]

The Jaccard metric (see, for example, Levandowsky and Winter, 1971) identifies the dissimilarity between two finite sets by the elements they have in common relative to their total number of unique elements: \(J(A,B) = 1 - \dfrac{|A \cap B|}{|A \cup B|}\).