Welfare Measures from Dynamic Choice Models

Why Dynamic Welfare Requires a Lifetime Comparison

A policy can change more than the payoff received today. It may also change which action is optimal, which states are reached, and the value of choices available later. A welfare measure based only on the policy’s immediate utility effect can therefore miss important consequences. In a dynamic choice model, welfare comparisons should be made using expected value functions that include both current utility and discounted continuation value.

A compensating variation expresses a policy’s welfare effect in monetary units. The amount depends on what kind of monetary adjustment is being considered: a one-time payment today and a recurring income change have different effects on lifetime value. The adjustment must also be defined with a sign convention, so that a positive value has an unambiguous interpretation.

Definition: Let policy \(r\in\{0,1\}\) denote the baseline and alternative policy, and let \(\bar V^r_t(s)\) be the expected value at date \(t\), before the current choice shocks are realized, for an agent in state \(s\). The compensating variation is the monetary adjustment under the alternative policy that makes its expected value equal to the baseline value. A positive compensating variation, under the convention used here, is money that could be taken away under the alternative policy while leaving the agent as well off as under the baseline.

Expected Value Functions and Welfare

Write \(a\) for an available action, \(\epsilon_a\) for its choice shock, \(u^r_t(a,s)\) for its current utility under policy \(r\), and \(\beta\), with \(0\leq\beta<1\), for the discount factor. Let \(F^r(s'\mid s,a)\) denote the distribution of the next state under policy \(r\). The expected value function satisfies the Bellman equation

\bar V^r_t(s) = \mathbb{E}_{\epsilon} \left[ \max_a \left\{ u^r_t(a,s)+\epsilon_a + \beta\mathbb{E}_{s'\mid s,a,r} \left[\bar V^r_{t+1}(s')\right] \right\} \right].

The expectation over next states is conditional on the current state, action, and policy. In a finite-horizon model, the recursion ends at a specified terminal value. In a stationary infinite-horizon model, the value function satisfies the corresponding time-invariant Bellman equation. The expectation over choice shocks is part of the value calculation; the value is not simply the payoff from one selected action.

For a fixed initial state \(s\), the policy’s welfare effect in utility units is \(\Delta V(s)=\bar V^1_0(s)-\bar V^0_0(s)\). A positive \(\Delta V(s)\) means the alternative policy raises expected value for an agent starting in that state. It does not, by itself, give a monetary amount: that conversion requires specifying how money enters utility and when the monetary adjustment occurs.

Convention: A recurring compensating variation \(c\) is a cash reduction of \(c\) in every period under policy 1, including the current period. It is defined by \(\bar V^1_0(s;y-c)=\bar V^0_0(s;y)\), where \(y\) denotes the relevant cash income. Thus \(c>0\) means the alternative policy creates enough welfare to permit a recurring cash reduction of \(c\) per period. A one-time compensating variation instead changes cash only at the current date.

If money enters every period’s utility linearly, with marginal utility \(\lambda>0\) utility units per dollar, and a cash reduction does not otherwise change choices or transitions, recurring cash has a simple discounted value. Over \(H\) remaining periods, its utility cost is \(\lambda c\) times the discounted annuity factor. In an infinite-horizon stationary model, that factor is \(1/(1-\beta)\). These shortcuts require the stated linearity and timing assumptions; otherwise compensating variation generally must be found by solving the value equality.

Converting a Value Difference into Money

Result: Suppose the policy’s value difference is \(\Delta V(s)\), and reducing cash by \(c\) lowers utility by \(\lambda c\) in each of \(H\) periods, without changing the choice or transition structure. Then the recurring compensating variation is \(c=\Delta V(s)/\{\lambda\sum_{j=0}^{H-1}\beta^j\}\). With infinitely many periods, it is \(c=(1-\beta)\Delta V(s)/\lambda\).

To justify the finite-horizon formula, the cash reduction changes current utility by \(-\lambda c\). At the next date it changes utility by another \(-\lambda c\), discounted by \(\beta\); at date \(j\), the discounted change is \(-\beta^j\lambda c\). Summing the changes gives

\bar V^1_0(s;y-c)-\bar V^1_0(s;y) = -\lambda c\sum_{j=0}^{H-1}\beta^j.

The compensating variation sets the adjusted policy-1 value equal to the baseline value. Since \(\bar V^1_0(s;y)-\bar V^0_0(s;y)=\Delta V(s)\), equality requires \(\Delta V(s)-\lambda c\sum_{j=0}^{H-1}\beta^j=0\). Solving yields the stated formula. For an infinite horizon, the geometric sum is \(1/(1-\beta)\). The argument relies on the cash adjustment lowering every action’s utility by the same amount, so it does not alter the choice ranking or transitions.

Worked Example: Two Periods of Policy Benefits

Consider a hypothetical two-period model with \(\beta=0.90\). At every date and for every available action, the alternative policy raises current utility by \(0.80\) utility units and leaves transitions unchanged. Assume utility falls by \(\lambda=0.020\) utility units per dollar of cash reduction, and that this cash adjustment affects both periods. Calculate the recurring compensating variation. Also calculate the one-time cash equivalent if money is reduced only today.

Because the policy adds the same utility to every action, it does not change the choice ranking. Its effect on value is the current gain plus the discounted gain in the final period:

\Delta V = 0.80+0.90(0.80) = 1.52\text{ utility units}.

There are two periods in which the recurring cash reduction applies, so the discounted annuity factor is \(1+0.90=1.90\). Substituting into the formula gives

c = \frac{1.52}{0.020(1.90)} = \frac{1.52}{0.038} = 40.0\text{ dollars per period}.

Thus, the policy’s welfare gain is equivalent to a recurring cash increase of \(40.0\) dollars per period, or, with the chosen sign convention, it permits a cash reduction of \(40.0\) dollars per period while preserving baseline value. If the adjustment is made only today, its utility cost is \(0.020c\), not \(0.020c(1.90)\). The one-time equivalent is therefore \(c=1.52/0.020=76.0\) dollars today. The two amounts differ because one is a recurring flow and the other is a single payment.

Infinite-Horizon Welfare and the Timing of Money

Worked Example: A Constant Gain in a Stationary Model

Suppose a hypothetical stationary model has one state that returns to itself each period. Under the baseline policy, expected utility is \(1.00\) per period. The alternative policy adds \(0.40\) utility units per period to every action and does not change transitions. Let \(\beta=0.95\) and let \(\lambda=0.040\) utility units per dollar. Find the recurring compensating variation and compare it with the one-period cash equivalent of the current utility gain.

The baseline value is \(1.00/(1-0.95)=20.0\) utility units. Under the alternative it is \(1.40/(1-0.95)=28.0\). Hence the lifetime value difference is \(28.0-20.0=8.00\) utility units. Equivalently, the repeated policy gain is \(0.40/(1-0.95)=8.00\).

The discounted annuity factor is \(1/(1-0.95)=20.0\). A recurring cash reduction \(c\) therefore lowers value by \(0.040c(20.0)\). Setting that equal to the policy gain gives

c = \frac{8.00}{0.040(20.0)} = \frac{8.00}{0.800} = 10.0\text{ dollars per period}.

The current utility gain alone is equivalent to \(0.40/0.040=10.0\) dollars for one period. In this example, that is also the recurring compensating variation because the same gain and the same cash adjustment repeat every period. It would be incorrect to call \(8.00/0.040=200\) dollars the recurring compensating variation: that calculation divides a lifetime utility sum by a one-period marginal utility of money without accounting for the discounted stream of cash adjustments.

When transitions or action values differ across states, an overall welfare measure must also specify who is being evaluated. For a known initial state, compare values at that state. For an initial-state distribution, one may compute expected value differences over that distribution. If compensating variations are calculated state by state, their average need not equal a single compensating variation for the average value difference, especially when the monetary mapping is nonlinear.

Worked Example: A Policy Changes Future States

Consider a hypothetical one-decision, two-period model with no choice shocks. At the first date, an agent chooses maintain or invest. Under the baseline policy, maintain gives current utility \(5.00\) and leads to terminal value \(0\). Invest gives current utility \(3.00\) and expected terminal value \(4.00\). The discount factor is \(0.90\). Under the alternative policy, current utility rises by \(0.50\) for either action; in addition, the transition following investment raises its expected terminal value from \(4.00\) to \(5.50\). The transition after maintain is unchanged. Calculate the welfare effect and its one-time cash equivalent when current utility has marginal cash value \(\lambda=0.050\) utility units per dollar.

Under the baseline, the action values are

Q^0(\text{maintain})=5.00+0.90(0)=5.00, \qquad Q^0(\text{invest})=3.00+0.90(4.00)=6.60.

The agent chooses invest, so the baseline value is \(\bar V^0=6.60\). Under the alternative policy,

Q^1(\text{maintain})=5.50+0.90(0)=5.50, \qquad Q^1(\text{invest})=3.50+0.90(5.50)=8.45.

The agent still chooses invest, so \(\bar V^1=8.45\) and \(\Delta V=8.45-6.60=1.85\) utility units. This difference comprises the \(0.50\)-unit current gain on the chosen action and the discounted increase in expected terminal value, \(0.90(5.50-4.00)=1.35\). Their sum is \(0.50+1.35=1.85\).

A one-time cash reduction today lowers utility by \(0.050c\) and leaves terminal value unchanged. Its compensating variation is \(c=1.85/0.050=37.0\) dollars today. By contrast, converting only the current \(0.50\)-unit payoff gain gives \(0.50/0.050=10.0\) dollars. That smaller amount omits the welfare effect of the policy-induced transition to a more valuable terminal state.

When a Simple Conversion Is Not Valid

A value difference can be converted by division only under a specific monetary mapping. If utility is nonlinear in cash, a fixed dollar adjustment may have different effects across states or income levels. If cash affects the state transition, the adjustment may change future opportunities. If it changes the ranking of actions, the relevant value function must be recomputed after the adjustment. In these cases, define the compensating variation by the value equality and solve for it, often numerically:

\bar V^1_0(s;y-c)=\bar V^0_0(s;y).

This definition is still meaningful even when no closed-form formula exists. The solution should be reported with its timing and units—for example, “a one-time payment of \(c\) dollars today” or “a recurring reduction of \(c\) dollars per period”—rather than as an unspecified monetary welfare effect.

Common Pitfalls

Using only the current payoff change. This misses discounted future utility and changes in the state distribution. Compare expected value functions, including continuation values.
Mixing up a flow and a lump sum. A recurring cash adjustment affects multiple periods; a one-time adjustment affects only the specified date. State which one is being calculated.
Ignoring the money-to-utility conversion. Utility units are not dollars. Specify the marginal utility of money or solve the value equality using the model’s utility function.
Reporting an average without defining the population calculation. Welfare can depend on the initial state. State whether the result is conditional on a particular state or averaged over an initial-state distribution.

Key takeaway: Dynamic compensating variation equates expected values under two policies after a clearly timed monetary adjustment. Its magnitude reflects discounted lifetime consequences, not merely a current-period payoff difference.

Check Your Understanding

Answer each question using the value comparison and monetary timing specified.

In a two-period model with \(\beta=0.8\), a policy adds \(1.00\) utility unit in each period. If \(\lambda=0.05\) utility units per dollar, what recurring compensating variation makes the agent indifferent?
Why can a one-time cash equivalent differ from a recurring compensating variation for the same policy?
A policy raises current payoff by \(0.4\) utility units but also changes transitions. What additional model quantity is needed to calculate its full dynamic welfare effect?
Under what conditions can a value difference be divided by \(\lambda\) to obtain a one-time cash equivalent?
Explain why a compensating variation should specify an initial state or initial-state distribution.