Tutorials › Causal Inference › Budget-Constrained Treatment Allocation

Causal Learning and Policy Decisions · Tutorial 970 of 972

Budget-Constrained Treatment Allocation

Derive the optimal treatment-allocation rule under a fixed capacity constraint, then apply it to subgroup and individual-level examples.

Advanced 11 min read

What You'll Learn

  • Define policy value when treatment is assigned with a probability depending on observed characteristics.
  • Distinguish an upper bound on treatment capacity from an exact quota that must be filled.
  • Derive the Lagrangian threshold rule for allocating treatment to those with the greatest conditional benefit.
  • Handle ties at the threshold using randomized allocation or a valid individual-level tie-breaker.
  • Calculate policy value and treatment capacity in hypothetical subgroup and finite-roster examples.
  • Identify why deterministic policies cannot always fill a quota when observed characteristics do not distinguish tied units.

Why Treatment Capacity Changes the Decision

When treatment is scarce, the decision is not simply whether it helps an individual. The policy must also decide who should receive the limited places. If one person is expected to benefit more than another, assigning treatment to the person with the larger expected benefit can raise the population value of the policy. The relevant comparison is therefore across individuals, not just between treatment and no treatment for one person at a time.

This tutorial considers a binary treatment and an outcome for which larger values are better. Each person has observed characteristics \(X\), and a policy may assign treatment with a probability that depends on those characteristics. The treatment capacity is a fraction \(\kappa\) of the population, where \(0\leq\kappa\leq1\). The goal is to maximize expected outcome subject to that capacity.

Definition: Let \(Y(1)\) and \(Y(0)\) be the potential outcomes under treatment and no treatment, respectively. The conditional average treatment effect is \(\tau(x)=E[Y(1)-Y(0)\mid X=x]\). A possibly randomized treatment policy is a function \(\pi(x)\in[0,1]\), interpreted as the probability of assigning treatment to a person with characteristics \(x\). Its treatment rate is \(E[\pi(X)]\), and it meets an upper capacity limit \(\kappa\) when \(E[\pi(X)]\leq\kappa\).

For this allocation problem, assume treatment is well-defined, consistency holds, and there is no interference between people’s assignments and outcomes. The conditional effects \(\tau(x)\) must also be identified or otherwise available for the decision. In observational data, a standard sufficient set of identification conditions includes conditional exchangeability, \(Y(1),Y(0)\perp D\mid X\), and positivity for treatment and control at the covariate values where the policy will be used. In randomized data, assignment supports exchangeability by design. These conditions concern learning the effects; the allocation rule below takes the effects as given.

Policy Value and the Capacity Constraint

Under policy \(\pi\), a person with characteristics \(X\) receives treatment with probability \(\pi(X)\). The policy’s value is the expected outcome that would result from applying it. By conditioning on \(X\), this value can be written as

$$ V(\pi)=E[Y(0)]+E[\pi(X)\tau(X)]. $$

The first term is the expected outcome under no treatment for everyone. The second is the population’s expected gain from the policy. Because \(E[Y(0)]\) is the same for all policies, maximizing policy value is equivalent to maximizing \(E[\pi(X)\tau(X)]\), subject to the capacity limit.

$$ \max_{0\leq \pi(X)\leq 1} E[\pi(X)\tau(X)] \quad\text{subject to}\quad E[\pi(X)]\leq\kappa. $$

This formulation treats one person receiving treatment as using one unit of capacity, regardless of their characteristics. It does not include varying treatment costs or different utilities for different people. Also distinguish an upper capacity limit from an exact quota: a limit allows the policy to leave places unused, while an exact quota requires the policy to treat a specified fraction.

Deriving the Benefit Threshold

Introduce a nonnegative multiplier \(\lambda\) for the upper capacity constraint. The Lagrangian is

$$ L(\pi,\lambda) =E[\pi(X)\tau(X)]-\lambda\{E[\pi(X)]-\kappa\} =\lambda\kappa+E[\pi(X)\{\tau(X)-\lambda\}]. $$

For a fixed \(\lambda\), the policy choice at each \(x\) maximizes \(\pi(x)\{\tau(x)-\lambda\}\) over \(0\leq\pi(x)\leq1\). If \(\tau(x)>\lambda\), this expression is largest at \(\pi(x)=1\); if \(\tau(x)<\lambda\), it is largest at \(\pi(x)=0\). If \(\tau(x)=\lambda\), every value of \(\pi(x)\) gives the same Lagrangian contribution. Thus the multiplier acts as a threshold for the conditional benefit.

Theorem: For an upper capacity limit, an optimal policy treats units with \(\tau(x)>\lambda\), does not treat units with \(\tau(x)<\lambda\), and may treat units with \(\tau(x)=\lambda\) with any probability between 0 and 1. The multiplier \(\lambda\geq0\) is chosen so that the capacity constraint is met when binding. If the constraint is slack, complementary slackness requires \(\lambda=0\). When the threshold group must be divided to meet capacity, fractional treatment can be implemented by randomization or by a tie-breaker that distinguishes individuals.

A short inequality verifies why this rule is optimal. For any feasible policy and any \(\lambda\geq0\), its incremental value satisfies

$$ E[\pi(X)\tau(X)] =E[\pi(X)\{\tau(X)-\lambda\}]+\lambda E[\pi(X)] \leq E[(\tau(X)-\lambda)_+]+\lambda\kappa, $$

where \(a_+=\max(a,0)\). The inequality follows because \(0\leq\pi\leq1\), so \(\pi(\tau-\lambda)\leq(\tau-\lambda)_+\), and because \(E[\pi]\leq\kappa\). The threshold policy attains this bound when it treats above the threshold, does not treat below it, and uses the available threshold ties as needed. This proves the allocation rule.

If treatment effects are positive for more than the available capacity, the capacity binds and \(\lambda\) is a cutoff among the most beneficial units. If fewer than \(\kappa\) of the population have positive effects, an upper-limit policy can leave capacity unused: it treats units with positive effects, not units expected to be harmed. An exact quota is different. It may require treatment even when effects are negative, and its threshold can then be below zero.

Worked Example: Allocating Capacity Across Three Groups

Suppose a hypothetical population has three groups. Group A is 20% of the population and has a conditional treatment effect of 8 outcome units; Group B is 30% and has an effect of 4 units; Group C is 50% and has an effect of 1 unit. Assume these effects are identified and the policy has an upper capacity limit of 40% of the population. What allocation maximizes expected outcome?

The effects rank the groups A, B, then C. Treat all of A, which uses 20 percentage points of the population and therefore half of the 40% capacity. There are 20 percentage points of capacity remaining. Group B is 30% of the population, so treating two-thirds of B uses exactly the remaining capacity:

$$ 0.20+0.30\left(\frac{2}{3}\right)=0.20+0.20=0.40. $$

The threshold is \(\lambda=4\) outcome units: all of A is above it, all of C is below it, and B is tied at it. A policy that observes only the three group labels cannot deterministically treat two-thirds of group B. It can instead randomize treatment with probability \(2/3\) for each member of B, or use an available individual-level tie-breaker to select that fraction. Randomization gives the required treatment rate in expectation.

The incremental expected gain per person in the population is

$$ 0.20(8)+0.30\left(\frac{2}{3}\right)(4) =1.60+0.80=2.40\text{ outcome units}. $$

Among those treated, the mean expected gain is \(2.40/0.40=6.00\) outcome units per treated person. If the population’s mean outcome under no treatment is hypothetically 50 units, policy value is \(50+2.40=52.40\) units. The baseline value is not needed to rank allocations, but it converts the incremental gain into total expected outcome.

Capacity Limits Versus Exact Quotas

The distinction between a limit and an exact quota matters when the capacity exceeds the share of people with positive effects. With an upper limit, the best policy need not fill every place. With an exact quota, the policy must fill the quota even if the marginal treatment effect is negative. The following example makes the difference explicit.

Worked Example: Leaving Capacity Unused When Effects Are Negative

Suppose a hypothetical population has 25% of people with an effect of 6 units, 25% with an effect of 2 units, and 50% with an effect of \(-1\) unit. The program can treat up to 70% of the population. Assume the effects are known for this example.

Both positive-effect groups together make up 50% of the population, less than the 70% limit. The upper-limit policy treats the two positive-effect groups and leaves the negative-effect group untreated. Its treatment rate is

$$ 0.25+0.25=0.50\leq0.70. $$

The incremental value is

$$ 0.25(6)+0.25(2)=1.50+0.50=2.00\text{ outcome units per person}. $$

Here the constraint is slack and the benefit cutoff is zero: effects above zero are treated, while effects below zero are not. If instead the policy had an exact quota of 70%, it would have to treat 20 percentage points of the population with effect \(-1\). Its incremental value would be \(0.25(6)+0.25(2)+0.20(-1)=1.80\) units per person. Filling that exact quota lowers expected value compared with using an upper limit.

Handling Ties in a Finite Allocation

The threshold rule may leave a group tied at the cutoff. A fractional treatment probability is a valid way to describe a randomized policy, but it should not be confused with a deterministic rule based only on characteristics that fail to distinguish people in that group. In a finite roster with an exact number of places, one can select the required number from the tied units using a randomized lottery or an additional individual-level tie-breaker.

Worked Example: Choosing Three Places From a Six-Person Roster

Consider a hypothetical roster of six people whose conditional treatment effects, in outcome units, are 9, 7, 7, 4, 1, and \(-2\). Exactly three of the six can be treated, so the quota is \(3/6=0.50\). Suppose the policy observes each person’s effect category but has no other feature that distinguishes the two people tied at 7.

Treat the person with effect 9 and one of the two people with effect 7, then use a fair lottery to choose which of the tied pair receives the third place. The expected incremental gain for the roster is

$$ 9+\left(\frac{1}{2}\right)7+\left(\frac{1}{2}\right)7 =9+3.5+3.5=16\text{ outcome units}. $$

The expected gain per person in the six-person roster is \(16/6=2.667\) units, and the expected gain per treated person is \(16/3=5.333\) units. Because the lottery selects exactly one person from the tied pair, the realized treatment count is always three. A policy that independently treated each tied person with probability \(1/2\) would have the same expected treatment count within the pair, but could realize zero, one, or two treatments from that pair; it would not guarantee exactly three treatments on every roster.

If a valid individual-level characteristic distinguished the two tied people, a deterministic policy could use it to select exactly one. Without such a distinction, claiming that a deterministic policy based only on the shared effect category treats exactly one of them is not justified.

Practical Interpretation and Common Pitfalls

The threshold \(\lambda\) is the expected benefit of the marginal treatment place under the capacity-constrained policy. Raising capacity generally permits the policy to include people with smaller conditional effects; reducing capacity raises the cutoff. If the effect estimates are noisy, a ranking can be unstable, especially near the threshold. The allocation calculation describes the rule given the conditional effects; it does not make uncertain effect estimates known or eliminate the need for appropriate estimation and evaluation.

  • Confusing percentage points with a share of capacity. Treating 20% of the population under a 40% capacity limit uses half the capacity, not 20% of it.
  • Assuming every capacity limit is a quota. Under an upper limit, negative-effect units need not be treated merely to fill places. An exact quota can require those treatments.
  • Ignoring ties. A deterministic rule based only on \(X\) cannot split people who have identical observed \(X\) values. Use randomized assignment or an available tie-breaker and state which one.
  • Ranking by treatment effects without defining the outcome scale. The threshold and gains are expressed in the units of \(Y\); the interpretation must name those units.
  • Calling an estimated ranking causal by itself. The allocation rule uses conditional causal effects, so those effects must be identified under appropriate assumptions before interpreting the resulting policy as optimizing causal value.
Key takeaway: With equal treatment capacity per person, allocate scarce treatment to the greatest conditional benefits. The optimal rule is a benefit threshold, with careful handling of tied units; an upper capacity limit may be left partly unused when the remaining effects are negative.

Check Your Understanding

Use the threshold rule and distinguish an upper limit from an exact quota.

  1. A group makes up 15% of a population and receives treatment under a policy with a 30% capacity limit. What fraction of the available capacity does that group use?
  2. Why can a policy based only on three group labels fail to treat exactly two-thirds of one group deterministically?
  3. If 40% of a population has positive conditional effects and the upper capacity limit is 60%, must the policy treat 60%? Explain.
  4. In a finite roster with an exact number of places, name two valid ways to allocate a place among units tied at the threshold.
  5. For an exact quota, why might the benefit threshold be negative?
Need help applying this method to your own data?

Our team supports biostatistics, statistical programming and pharmacometric analyses. Send us a request and we will reply by email.

Request an analysis
Clinical Biostats