Model selection and multiple research goals: The case of rational addiction

Abstract

A comparison of rational addiction and time inconsistency models of addiction highlights the complexities of model selection when researchers have goals in addition to empirical fit. Although currently the two models of addiction are underdetermined by data, each offers a different understanding of addiction; moreover, the two models offer starkly different policy implications. When the goals of understanding and policy usefulness are added to the goal of empirical fit, a more complex account of model selection is needed. First, the principle of parsimony loses some of its force when researchers also value understanding and policy usefulness. Second, when economists value understanding as well as pure prediction, a broader justification of the realism of assumptions becomes possible. Third, because radically different policy advice flows from these empirically equivalent models, this literature underscores the difficulty of separating the seemingly positive analysis of consumer behavior from normative analysis.

Keywords:

• rational addiction
• time inconsistency
• model selection
• parsimony
• realism of assumptions

Notes

1. Because Quine (1953) asserts that theories are always underdetermined by data, he asserts that model selection is always made on ‘pragmatic’ grounds. Hausman (1992) lists ‘aesthetic appeal’, ‘heuristic power’ and ‘normative force’ as pragmatic considerations relevant to model selection.

2. Although the ‘official’ aim of economics is predictionist, Hausman (1992) and McCloskey (1994) both assert that economists are not really predictionists, interested only in ‘empirical fit’, in practice. Hausman traces the heritage of Millian deductive method in economics. Economics reasons from a set of accepted lawlike assertions which are inexact, to construct a compact if imperfect understanding of the economy. McCloskey (1994) criticizes predictionist method as unworkable and unpracticed, and alleges that economists use a much more eclectic method to ascertain truth about the economy. Coase (1994) notes that one of the virtues of models is that they help economists to understand the social world, and Robbins (1963) acknowledges that most economists want to produce work that informs policy.

3. Musgrave also speculates that empirical testing affects modeling assumptions. The encounter with data can force an economist to transform a negligibility assumption (‘transaction costs are zero’) into a domain assumption (‘the model only applies where transaction costs are zero’) into a heuristic assumption (‘for now, assume that transaction costs are zero; we will complicate the model with transaction costs later’).

4. Mäki's discussion of the role of abstraction highlights the fact that parsimony can promote understanding as well as empirical fit, although the evaluation of its contributions to understanding requires more judgment about the trade‐off between abstraction and true description.

5. Gruber and Köszegi (2004) summarize the arguments in favor of time inconsistency; Goldfarb et al. (2001) argue that survey evidence, empirical tests and evidence from health studies do not clearly favor any one model.

6. Understood in this way, the assumption of time consistency is an example of negligibility assumption, as defined by Musgrave (1981).

7. A full account of model selection with multiple research goals is difficult to formulate methodologically: there is no pre‐existing set of disciplinary preferences on which to evaluate trade‐offs among goals, and disagreements will not always be resolvable by appeals to data. The intellectual faculty brought into play by this complex context goes by the name of ‘prudence’, or ‘practical reason’. Yuengert (2004) uses prudence as an organizing principle for his account of economic research.