Abstract
Consider an agent who faces choice problems and learns information about an objective state of the world through a technology of sequential experiments. We consider two cases of learning costs. In the first, the agent discounts future payoffs geometrically. In the second, she incurs a constant flow cost of time. If the observable data consist only of the joint distributions over chosen actions and decision times, an analyst can uniquely identify the discount factor in the first case and the flow cost of time in the second case. Moreover, we show how an analyst can recover the agent’s ex ante welfare in both cases, besides identifying her prior belief. Our approach does not rely on any knowledge about the underlying sequential experiment.
We’re sorry, something doesn't seem to be working properly.
Please try refreshing the page. If that doesn't work, please contact support so we can address the problem.
Notes
See Proposition 3.G.4 in Mas-Collel et al. (1995) for the classical Roy’s identity.
It follows that with discounting, the agent’s behavior satisfies a property of homotheticity; whereas with additive costs, her behavior satisfies a weak notion of linearity. We address these behavioral differences in Sect. 5 in more detail.
We take this prize space for ease in exposition. Results can be proven for general metric space Z.
In the case of the flow cost c, the decision time in our model may be interpreted differently in specific situations. For example, assume the agent has access to a fixed experiment and makes repetitive draws from it (i.i.d. draws from the same experiment). Then, the ‘decision time’ t is interpreted as the number of experiments she chooses to perform.
Random choice as an observable may be interpreted in two distinct ways. In the single-agent interpretation, the analyst observes the limiting frequency of choices, as well as of decision times of a single agent, coming from many repetitions of the same decision problem. In the population interpretation, the analyst observes choices from menus, as well as of decision times from a population of homogeneous individuals with the same taste, prior and costs of information having access to the same sequential experiment. Our results hold for both interpretations, but we pick the former for ease of exposition.
This proof was suggested by the co-editor and uses only elementary methods. Previous versions of the paper contained a proof which uses the Envelope Theorem from Milgrom and Segal (2002).
The addendum to the proof in the appendix contains a derivation of the first statement of this fact. The second statement is straightforward given the first one.
The addendum to the proof in the appendix contains more details about this step.
The approach of recovering the indirect utility of a menu through ex post random choice first appears in Lu (2016). In his static model, private information is exogenous and menu independent.
In a static rational inattention model where private information is optimally acquired, Lin (2019) shows through an envelope theorem argument that random choice recovers the ex ante valuation of a menu. His recoverability result applies directly to our SeSa-LC model because the latter can be reduced to a static rational inattention model.
\(\mathcal {A}\) is equipped with the standard Hausdorff topology.
The equation (5) is by no means the only procedure to identify the prior belief, but it is the one we found most elegant for exposition.
Given u identified, an act \(f\in \mathbb {F}\) corresponds to the utility act \(\tilde{f}\) defined as \(\tilde{f}(s) = u\circ f(s)\, \forall s\in S\).
Readers can verify that in Example 1, no learning is optimal for all menus \(\{f,r\}\) if and only if \(\delta \le \frac{2}{3}\).
In Table 1 and all the following tables, a pair (g, t) in the upper row with g an act and \(t\in \mathcal {T}\) is the argument of the RCDT under consideration. Readers can verify all the tables in our examples given the specified parameters.
This is unless we assume the analyst is in the idealized situation where he has some algorithm to check if there is at least one pair \((\tau (A',\cdot ),\mathcal {P}(A',\cdot ))\) that can explain the observable behavior \(P_{A'}\). We think this may be too strong a requirement.
References
Anscombe, F. J., & Aumann, R. J. (1963). A definition of subjective probability. The Annals of Mathematical Statistics, 34(1), 199–205.
Arrow, K. J., Blackwell, D., & Girshick, M. A. (1949). Bayes and minimax solutions of sequential decision problems. Econometrica, 17(3/4), 213–244.
Caplin, A., & Dean, M. (2015). Revealed preference, rational inattention, and costly information acquisition. American Economic Review, 105(7), 2183–2203.
Che, Y. K., & Mierendorff, K. (2019). Optimal dynamic allocation of attention. American Economic Review, 109(8), 2993–3029.
de Oliveira, H., Denti, T., Mihm, M., & Ozbek, K. (2017). Rationally inattentive preferences and hidden information costs. Theoretical Economics, 12(2), 621–654.
Denti T (2020) Posterior separable cost of information. working paper
Duraj, J. (2018). Dynamic random subjective expected utility. arXiv:1808.00296.
Duraj, J., & Lin, Y.H. (2021). Costly information and random choice. Econ Theory. https://doi.org/10.1007/s00199-021-01361-w.
Echenique, F., & Saito, K. (2017). Response time and utility. Journal of Economic Behavior & Organization, 139, 49–59.
Epstein, L. G. (1983). Stationary cardinal utility and optimal growth under uncertainty. Journal of Economic Theory, 31(1), 133–152.
Fishburn, P. C., & Rubinstein, A. (1982). Time preference. International Economic Review, 23(3), 677–694.
Frick, M., Iijima, R., & Strzalecki, T. (2019). Dynamic random utility. Econometrica, 87(6), 1941–2002.
Fudenberg, D., Strack, P., & Strzalecki, T. (2018). Speed, accuracy, and the optimal timing of choices. American Economic Review, 108(12), 3651–3684.
Hébert, B., & Woodford, M. (2019). Rational inattention when decisions take time. NBER Working Paper No. w26415, Available at SSRN: https://ssrn.com/abstract=3476495.
Koopmans, T. C. (1960). Stationary ordinal utility and impatience. Econometrica, 28(2), 287–309.
Lin, YH. (2019). Stochastic choice and rational inattention. working paper.
Lu, J. (2016). Random choice and private information. Econometrica, 84(6), 1983–2027.
Magnac, T., & Thesmar, D. (2002). Identifying dynamic discrete decision processes. Econometrica, 70(2), 801–816.
Mas-Collel, A., Whinston, M. D., & Green, J. R. (1995). Microeconomic Theory. New York: Oxford University Press.
Matějka, F., & McKay, A. (2015). Rational inattention to discrete choices: A new foundation for the multinomial logit model. American Economic Review, 105(1), 272–298.
Milgrom, P., & Segal, I. (2002). Envelope theorems for arbitrary choice sets. Econometrica, 70(2), 583–601.
Wald, A. (1947). Foundations of a general theory of sequential decision functions. Econometrica, 15(4), 279–313.
Zhong, W. (2019). Optimal dynamic information acquisition. working paper.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
We are thankful to Larry Epstein, Drew Fudenberg and Tomasz Strzalecki for their continuous encouragement and support in this project. We also thank Jerry Green, Kevin He, and Jay Lu for their insightful comments. Finally, we thank participants in the conferences RUD 2019 and NASMES 2019 for questions and feedback. Any errors are ours.
Appendices
Addendum to the proof of theorem 1 for SeSa-GD
1.1 Proof of fact 3.
Given a history \(e^1\), a menu A and an act \(f\in A\), let \(P_A^{e^1}(f,t)\) be the probability that the agent chooses f from A at period \(t\ge 1\) after history \(e^1\) occurs. Note that \(W_1(A\cup f^u_r)=\mathbb {E}_{e^1}[\tilde{W}_1(A\cup f^u_r,e^1)]\). By the induction hypothesis,
So we have
Hence \(\frac{d}{dr}\delta W_1(A\cup \{f^u_r\})=\delta \mathbb {E}_{e^1}[\frac{d}{dr}\tilde{W}_1(A\cup f^u_r,e^1)]<1\).
1.2 Proof that \(P_{A\cup \{f_r^u\}}(f_r^u,0) = 1\) for \(r>\bar{r}\), and \(P_{A\cup \{f^u_r\}}(f^u_r,0) = 0\) for \(r<\bar{r}\).
When \(r>\bar{r}\ge \underline{r}\) it holds \(P_{A\cup \{f_r^u\}}(f_r^u,0) = 1\), because of Fact 2 and
When \(\underline{r}<r<\bar{r}\), we have
and thus \(P_{A\cup \{f^u_r\}}(A\cup \{f^u_r\},0) = 0\) (this ensures \(\mathbb {E}_{e^1\sim \mu (\pi _0)}[P^{e^1}_{A\cup \{f^u_r\}}(f^u_r,t)]=P_{A\cup \{f^u_r\}}(f^u_r,t)\)). When \(r<\underline{r}\le \bar{r}\), we have \(r<\delta W_1(A\cup \{f_r^u\})< V(A,u,\pi _0)\), implying
Proof of Theorem 1 for SeSa-LC
Just as in the case of SeSa-GD, the result is trivially true if the menu A contains only a single constant act which yields only the worst prize. Hence, we exclude this case in the following.
The proof proceeds by induction on the number of periods T.
For \(T=0\) the agent’s choice follows Subjective Expected Utility, up to tie-breaking. In particular, Theorem 2 from Lu (2016) applies, and immediately gives the statement.
Suppose the statement is true for \(T=n\), and consider a model with \(T=n+1\). Fix a menu \(A\in \mathcal {A}\) which is not the constant menu containing only the worst prize. Define \(\underline{r} = \inf \{r\in \mathbb {R}_+:W_1(A\cup \{f_r^u\})-c\ge V(A,u,\pi _0)\}\) and \(\bar{r} = \sup \{r\in \mathbb {R}_+: W_1(A\cup \{f_r^u\})-c\ge r\}\).
The following easily proven Facts are true for \(\underline{r},\bar{r}\).
-
Fact 1:
\(\bar{r}<\infty\), and \(\underline{r} = 0\) if and only if \(W_1(A)-c\ge V(A,u,\pi _0)\).
-
Fact 2:
For every \(r>\bar{r}\), it holds \(\tau (A\cup \{f_r^u\}, e^0)=1\). The same is true for any \(r\in (0,\underline{r})\) by definition of \(\underline{r}\).
-
Fact 3:
\(r\mapsto \delta W_1(A\cup \{f_r^u\})\) is continuous, non-decreasing with derivative almost everywhere strictly smaller than 1. For \(r<\bar{r}\) it holds \(\delta W_1(A\cup \{f_r^u\})> r\) and for \(r>\bar{r}\) it holds \(\delta W_1(A\cup \{f_r^u\})< r\).
The first statement of Fact 3 can be proven as in the case of SeSa-GD.
There are two cases to consider for the proof.
Case 1: Menu A has \(\underline{r}>\bar{r}\). Since \(\underline{r}\ge \bar{r}>0\) we have \(V(A,u,\pi _0)>W_1(A)-c\), so that \(W_c(A) = V(A,u,\pi _0)\). From Fact 2 it follows overall \(\tau (A\cup \{f_r^u\}, e^0)=1\) for every \(r\in \mathbb {R}_+\). Conditioning on the value of r, it follows that \(P_{A\cup \{f_r^u\}}(f_r^u,0) = 0\) for \(r<V(A,u,\pi _0)\) and \(P_{A\cup \{f_r^u\}}(f_r^u,0) = 1\) for \(r>V(A,u,\pi _0)\). This implies for such a menu A
Case 2: Menu A has \(\underline{r}\le \bar{r}\). Note that in this case \(\bar{r} = 0\) cannot happen, due to Fact 1. Continuity implies that \(\bar{r} = W_1(A\cup \{f_{\bar{r}}^u\})-c\). Given a history \(e^1\), a menu A and an act \(f\in A\), let \(P_A^{e^1}(f,t)\) be the probability that the agent chooses f from A at period \(t\ge 1\) after history \(e^1\) occurs. By the induction hypothesis, as well as similar arguments to the discounting case, we have
Overall,
Analogously to the discounting case, it holds \(W_c(A) = W_1(A\cup \{f_{\underline{r}}^u\}) -c\). The induction hypothesis and calculations analogous to the derivation of (10) deliver
Here, in the sixth equality, we have used (10), and in the seventh, we have interchanged the expectation with the integral. In the last equality we have used that for \(r>\bar{r}\) it holds \(V(A,u,\pi _0)<W_1(A\cup \{f_r^u\})-c <r\), and for \(r<\underline{r}\) it holds \(r\le W_1(A\cup \{f_r^u\})-c<V(A,u,\pi _0)\). These last statements can be proven by the same type of arguments as for the case of SeSa-GD.
Rights and permissions
About this article
Cite this article
Duraj, J., Lin, YH. Identification and welfare evaluation in sequential sampling models. Theory Decis 92, 407–431 (2022). https://doi.org/10.1007/s11238-021-09826-z
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11238-021-09826-z