Working Paper: NBER ID: w26849
Authors: Jason Abaluck; Giovanni Compiani
Abstract: We state a sufficient condition under which choice data alone suffices to identify consumer preferences when choices are not fully informed. Suppose that: (i) the data generating process is a search model in which the attribute hidden to consumers is observed by the econometrician; (ii) if a consumer searches good j, she also searches goods which are better than j in terms of the non-hidden component of utility; and (iii) consumers choose the good that maximizes overall utility among searched goods. Canonical models will be biased: the value of the hidden attribute will be understated because consumers will be unresponsive to variation in the attribute for goods that they do not search. Under the conditions above and additional mild restrictions, an alternative method of recovering preferences using cross derivatives of choice probabilities succeeds regardless of the search protocol and is thus robust to whether consumers are informed. The approach nests several standard models, including full information. Our methods suggest natural tests for full information and can be used to forecast how consumers will respond to additional information. We verify in a lab experiment that our approach succeeds in recovering preferences when consumers engage in costly search.
Keywords: Discrete Choice Models; Consumer Search; Preference Identification; Search Costs
JEL Codes: C5; C8; C9; D0; D6; D8
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Higher visible utility (D11) | Greater likelihood of searching and subsequently choosing a good (C52) |
| Conditions of the model met (C52) | Recover preferences from choice data alone (D11) |
| Identification approach (C52) | Isolate consumers who behave as if fully informed (D83) |
| Method can be used to simulate counterfactual scenarios (C59) | Assess how consumers respond to changes in search costs or additional information (D83) |
| Method's robustness (C59) | Yields robust estimates of consumer preferences (D11) |