Why Identification Depends on Variation
A dynamic choice model can fit observed choices well without its preference parameters being separately determined by the data. In particular, the current cost of changing an action may be hard to distinguish from the value of future consequences, which depends on the discount factor. The central identification question is therefore not simply whether choices vary, but which parts of the choice problem change across observations, and how.
Current-payoff covariates shift the immediate gain from an action. State variables such as the previous action can change the cost of switching. Other state variables, including excluded shifters, can alter the distribution of future states and therefore continuation values. These sources of variation can provide distinct information about preferences, but only under appropriate support, exclusion, and transition assumptions.
A Dynamic Choice Model and Identification
Consider a stationary model with discrete actions \(a\in\{0,1\}\) and observed state \(s=(x,l,e)\). Here \(x\) is a current-payoff covariate, \(l\) is the previous action, and \(e\) is a potential transition shifter. Let the per-period utility be
The parameter \(\theta\) governs the payoff response to \(x\), and \(\kappa\geq 0\) is a switching cost. The shocks \(\varepsilon_0,\varepsilon_1\) are independent Type I extreme-value shocks, independent over time and of the observed states. Their scale is normalized to one, fixing the utility scale. The agent discounts future utility by \(\beta\in(0,1)\). Conditional on the current state and action, the next state follows a Markov transition distribution \(F(s'|s,a)\).
The value function and choice-specific value are
Under the stated extreme-value shock and scale assumptions, the probability of choosing action 1 is logistic in the difference between the two choice-specific values:
This log-odds equation makes the identification problem visible. Current payoffs and switching costs enter directly, while \(\beta\) affects the difference in expected continuation values. To separate parameters, the data must contain variation that changes these components in distinguishable ways.
Three Sources of Identifying Variation
Payoff variation changes the immediate utility difference. If \(x\) varies over an adequate range and its coefficient is not perfectly entangled with other changing components, the response of choices to \(x\) informs \(\theta\). Repeated observations at different values of \(x\) are more useful than a large sample in which \(x\) is nearly constant.
State variation can reveal costs attached to history. For example, when the previous action \(l\) varies while current payoffs and relevant transition prospects are comparable, the action that entails switching changes. This can distinguish a switching cost from a general preference for action 0 or action 1. The comparison needs overlap: observations with different \(l\) must occur in comparable regions of the other state variables.
Transition variation changes the consequences of acting now. An excluded shifter \(e\) is a variable that affects the distribution of future states, but does not directly affect current utility after conditioning on the included payoff variables and state. If \(e\) changes the relative future prospects of the actions, it can shift the continuation-value difference. Since \(\beta\) scales future utility, this variation can help distinguish discounting from a current switching cost.
Exclusion is a substantive modeling restriction, not a fact guaranteed by observing a variable named “shifter.” If \(e\) also directly changes current utility, attributing its choice association solely to transitions can produce misleading conclusions. The transition process must also be specified or estimated appropriately, and the shifter must have meaningful variation in that process.
Worked Example: Using Lagged-Action Variation to Inform a Switching Cost
Suppose a hypothetical dataset is described by the model above, with \(\theta=1\), \(\kappa=0.4\), \(x=0.5\), and an expected continuation-value difference \(E[V(s')|a=1,s]-E[V(s')|a=0,s]=0.2\) for the observations considered. Compare otherwise similar observations with \(l=0\) and \(l=1\). Assume the transition comparison is indeed the same in both groups.
When \(l=0\), choosing action 1 incurs the switching cost, while choosing action 0 does not. Thus the log odds are
For the numerical illustration, take \(\beta=0.5\). Substitution gives \(1(0.5)-0.4+0.5(0.2)=0.2\), so \(P(a=1|s)=1/(1+\exp(-0.2))=0.5498\).
When \(l=1\), choosing action 0 instead incurs the switching cost. The log odds become \(\theta x+\kappa+\beta(0.2)\). Substitution gives \(1(0.5)+0.4+0.5(0.2)=1.0\), so \(P(a=1|s)=1/(1+\exp(-1.0))=0.7311\). The predicted log-odds difference is \(1.0-0.2=0.8=2\kappa\).
This contrast is informative about \(\kappa\) because the switching penalty moves from one alternative to the other as \(l\) changes. Its interpretation relies on the stated comparability: if the transition-value difference also changed systematically with \(l\), the contrast would combine switching-cost and continuation effects rather than isolate the former.
Excluded Shifters and Discounting
A transition shifter can be useful when it changes the consequences of today’s choice without directly changing today’s payoff. The following simple calculation isolates the source of leverage. Suppose that, after today’s action, next period’s state is either good or bad. In the next period, the good state yields an additional value \(R\) relative to the bad state, and the action-specific probability of reaching the good state depends on \(e\). The difference in expected next-period values is then \(R[p_1(e)-p_0(e)]\), where \(p_a(e)\) is the probability of a good next state following action \(a\).
For this illustration, assume that \(R\) is known, there are no other differences in continuation values, and the shifter has no direct current-payoff effect. The current log odds are
Worked Example: A Transition Shifter Provides Leverage on Discounting
Suppose \(R=2\), \(\theta=1.2\), \(\kappa=0.4\), and \(\beta=0.9\). Let \(x=0.5\) in both groups. A hypothetical excluded shifter produces \(p_1(e)-p_0(e)=0.1\) in one group and \(0.5\) in another. Assume it changes only these transition probabilities and that the stated transition differences are correctly measured.
For the first group, the log odds equal \(1.2(0.5)-0.4+0.9(2)(0.1)=0.38\). The probability of action 1 is \(1/(1+\exp(-0.38))=0.5939\). For the second group, the log odds equal \(1.2(0.5)-0.4+0.9(2)(0.5)=1.10\), giving probability \(1/(1+\exp(-1.10))=0.7503\).
The log-odds difference is \(1.10-0.38=0.72\). Under this simplified model, the difference also equals \(\beta R(0.5-0.1)\). If \(R=2\) is known, the calculation gives \(\beta=0.72/[2(0.4)]=0.9000\). This illustrates how shifter-induced changes in future prospects can inform discounting, separately from a constant switching cost.
The conclusion depends on the assumptions. If \(R\) is unknown, this contrast identifies the product \(\beta R\), not necessarily \(\beta\) by itself. If the shifter also changes current utility, its choice effect cannot be assigned to continuation values without modeling that direct effect.
Why Weak Transition Variation Confounds Parameters
A useful local diagnostic comes from the same log-odds equation. Write \(\Delta(e)=R[p_1(e)-p_0(e)]\), and suppose \(x\), the transition process, and the scale normalization are otherwise handled as specified. The log odds are \(\theta x-\kappa+\beta\Delta(e)\). Their sensitivities to \(\kappa\) and \(\beta\) are \(-1\) and \(\Delta(e)\). If \(\Delta(e)\) is constant across the observed data, these two sensitivity patterns cannot distinguish a change in \(\kappa\) from a compensating change in \(\beta\).
More explicitly, for any small change \(d\beta\), setting \(d\kappa=\Delta\,d\beta\) leaves the log odds unchanged to first order: \(-d\kappa+\Delta\,d\beta=0\). If \(\Delta\) varies meaningfully across observations, this compensation cannot generally keep every log odds unchanged. This is a local rank argument: variation can help separate the parameters, but it does not by itself prove global identification. Other model restrictions and the support of the observed states still matter.
Worked Example: Nearly Constant Transition Differences Leave a Near-Confounding
Consider the simplified log-odds model with \(R=2\), \(x=0\), and two transition differences \(p_1(e)-p_0(e)=0.20\) and \(0.21\). With \((\kappa,\beta)=(0.50,0.75)\), the log odds in the two groups are
Now use \((\kappa,\beta)=(0.60,1.00)\). The corresponding values are \(-0.60+1.00(2)(0.20)=-0.200\) and \(-0.60+1.00(2)(0.21)=-0.180\). At the first transition difference, the predictions coincide exactly; at the second, they differ by only \(0.005\). The probabilities in the first group are both \(1/(1+\exp(0.200))=0.4502\). In the second group they are approximately \(0.4539\) and \(0.4551\), respectively.
With exact population probabilities, the slight second-group difference can in principle distinguish the parameter pairs in this simplified model. In a finite sample, however, that small difference may be difficult to detect. If the two transition differences were exactly equal, the log odds would depend on \(\kappa\) and \(\beta\) only through \(-\kappa+2\beta(0.20)\), so these data would not separately identify them. Weak variation therefore creates weak identification: estimates may be highly sensitive to small changes in the data or specification even when the model can technically distinguish parameters.
A Practical Identification Check
Before interpreting estimated dynamic preferences, inspect what varies in the observed state space and what the model says that variation changes. The following sequence is a useful diagnostic, not a substitute for the model’s formal assumptions.
Express the difference in choice-specific values and identify the terms involving current payoffs, switching costs, and continuation values.
Check whether payoff covariates, lagged actions, and transition shifters vary enough—and separately enough—to move those terms in distinguishable ways.
Justify why a proposed shifter affects transitions but not current utility, and check that comparisons are supported by observations with comparable other states.
Confirm that the transition model and shock normalization are explicit. A coefficient’s scale and a continuation-value effect cannot be interpreted independently of these conventions.
Ask both whether distinct parameter values imply distinct choice distributions and whether the observed variation is large enough to estimate those distinctions precisely.
Common Pitfall
A common mistake is to argue that a large sample automatically separates a switching cost from discounting. More observations reduce sampling uncertainty only when they provide information about the parameters. If every observation has the same transition-value difference, increasing the sample does not break the exact confounding in the simplified model. If differences are merely small, a larger sample may help, but the resulting estimates can still be imprecise or sensitive.
The correction is to examine the structure of the choice-value differences and the support of the relevant states. A credible identification argument explains which variation shifts which component, why the exclusion restriction is plausible, and which parameters remain linked if that variation is absent. Numerical convergence of a likelihood optimizer is not evidence that the underlying parameters are identified.
Check Your Understanding
Answer each question using the choice-value logic developed above.
- In the binary-choice model, what changes in the current utility difference when the previous action changes from 0 to 1?
- Why does an excluded shifter need to affect transitions, and what additional restriction must it satisfy?
- If the continuation-value difference is constant across all observations, explain why switching costs and discounting may not be separately identified.
- In the shifter example, what parameter combination is informed by the shifter contrast if \(R\) is unknown?
- How can weak but nonzero transition variation differ from exact nonidentification?