Tutorials › Causal Inference › Deriving the Optimal Treatment Rule

Causal Learning and Policy Decisions · Tutorial 963 of 972

Deriving the Optimal Treatment Rule

Learn to derive the treatment choice that maximizes conditional expected utility and apply it when treatment costs are constant or vary with covariates.

Advanced 10 min read

What You'll Learn

  • Define conditional net utility as treatment benefit minus incremental treatment cost.
  • Derive the pointwise utility-maximizing rule for two treatment options.
  • Distinguish a strictly better treatment from a treatment that is tied for best.
  • Apply a constant cost threshold to subgroup-specific treatment effects.
  • Solve a rule with covariate-dependent effects and costs.
  • Identify the conditions under which pointwise maximization yields the best overall policy.

Why the Treatment Rule Is a Utility Comparison

A conditional treatment effect describes how expected outcomes differ between treatment and control for people with covariates \(X=x\). A policy decision requires one more step: the treatment benefit must be compared with what treatment costs. A positive effect on the outcome alone does not guarantee that treatment is the preferred action if its incremental cost is larger than that benefit.

The goal here is to derive the decision rule for two actions, treatment and control, when each person’s choice can be made separately using their covariates. The rule is an oracle rule: it uses the true conditional expectations and therefore describes which action maximizes expected utility, not how those expectations should be estimated from data.

Definition: Let \(A\in\{0,1\}\) denote control and treatment, respectively, and let \(X\) be the covariates available for the decision. Let \(Y(a)\) be the benefit outcome under action \(a\), and let \(C(a)\) be its cost. Define the conditional treatment effect as \(\tau(x)=E[Y(1)-Y(0)\mid X=x]\), and the conditional incremental treatment cost as \(c(x)=E[C(1)-C(0)\mid X=x]\). The conditional net utility advantage of treatment is \(\Delta(x)=\tau(x)-c(x)\). The outcome and cost must be expressed on a common utility scale for this subtraction to be meaningful.

A treatment rule is a function \(d(X)\in\{0,1\}\): it assigns control when \(d(X)=0\) and treatment when \(d(X)=1\). For a rule \(d\), define its expected policy value as \(V(d)=E[Y(d(X))-C(d(X))]\). The rule is unconstrained when there is no limit on the number of people treated and no other requirement linking one person’s decision to another’s. These conditions let us compare the two actions separately at each covariate value.

The Pointwise Optimal Rule

At a fixed covariate value \(x\), compare conditional expected utility under the two actions. The difference between treatment and control is

$$ E[Y(1)-C(1)\mid X=x]-E[Y(0)-C(0)\mid X=x] =\tau(x)-c(x)=\Delta(x). $$

If \(\Delta(x)\) is positive, treatment has greater conditional expected utility. If it is negative, control has greater conditional expected utility. If it is zero, the two actions have equal conditional expected utility. Thus, the rule can be written with an explicit convention for ties:

$$ d^*(x)= \begin{cases} 1, & \Delta(x)>0,\\ 0, & \Delta(x)<0,\\ 0\text{ or }1, & \Delta(x)=0. \end{cases} $$
Key result: Treatment is strictly better than control at \(x\) exactly when \(\tau(x)>c(x)\). Treatment is a maximizing action when \(\tau(x)\ge c(x)\). At equality, either treatment or control maximizes conditional expected utility; neither is strictly better.

When costs are constant, \(c(x)=c\), the treatment threshold is the same for everyone: treatment is strictly better where \(\tau(x)>c\), and either action is optimal where \(\tau(x)=c\). A varying cost \(c(x)\) changes the threshold across covariate values. The relevant comparison is always the net advantage \(\Delta(x)\), not the treatment effect considered in isolation.

Why Pointwise Choices Maximize Overall Value

The pointwise comparison gives an overall optimum under the unconstrained setup, not merely a locally sensible recommendation. To see why, define \(\Delta(X)=\tau(X)-c(X)\). For any rule \(d\), its value can be expressed relative to the value of assigning control to everyone:

$$ V(d)=E[Y(0)-C(0)]+E[d(X)\Delta(X)]. $$

The first term is the same for every rule. The second term is the expected incremental utility from assigning treatment where \(d(X)=1\). Let \(d^*\) choose treatment when \(\Delta(X)>0\), control when \(\Delta(X)<0\), and either action when \(\Delta(X)=0\). For any other rule \(d\),

$$ V(d^*)-V(d)=E[(d^*(X)-d(X))\Delta(X)]\ge 0. $$

The inequality follows pointwise. When \(\Delta(X)>0\), \(d^*(X)=1\), so \(d^*(X)-d(X)\ge 0\), and their product is nonnegative. When \(\Delta(X)<0\), \(d^*(X)=0\), so \(d^*(X)-d(X)\le 0\), and its product with the negative \(\Delta(X)\) is again nonnegative. When \(\Delta(X)=0\), the product is zero for either action. Taking expectations preserves the inequality. Therefore no other unconstrained rule has greater expected value. Any rule that makes maximizing choices at all covariate values also attains the maximum.

Worked Example: A Constant Treatment Cost

Suppose a hypothetical program has a constant incremental cost of \(5\) utility units per person. The conditional benefit effects and proportions in three covariate groups are as follows. The outcome and cost have already been converted to the same utility scale.

GroupShare of population\(\tau(x)\)\(c(x)\)\(\Delta(x)=\tau(x)-c(x)\)
A0.301257
B0.40550
C0.30-45-9

For group A, the calculation is \(\Delta(A)=12-5=7\). Treatment is strictly better, so the rule treats this group. For group B, \(\Delta(B)=5-5=0\); treatment and control tie, so either action maximizes expected utility for this group. For group C, \(\Delta(C)=-4-5=-9\), so control is strictly better.

The expected incremental value relative to assigning control to everyone is \(0.30(7)+0.40(0)+0.30(0)=2.1\) utility units per person. Treating group B as well would leave that value unchanged: its treatment adds exactly as much cost as benefit. This is a tie, not a strict preference for control.

When Costs Vary Across People

If incremental cost depends on covariates, it is not enough to compare every person’s effect with one common cutoff. The decision compares that person’s conditional effect with their own conditional incremental cost. A group with a larger effect may still be less attractive for treatment if its cost is sufficiently larger.

Worked Example: Subgroup-Specific Costs and a Tie

Consider a hypothetical service with three subgroups. The table gives each subgroup’s conditional treatment effect, conditional incremental cost, and share of the population; all utility values use the same scale.

GroupShare of population\(\tau(x)\)\(c(x)\)\(\Delta(x)\)
10.50633
20.30880
30.2024-2

For group 1, \(6-3=3>0\), so treatment is strictly better. For group 2, \(8-8=0\), so either action maximizes expected utility. For group 3, \(2-4=-2<0\), so control is strictly better. Thus a maximizing rule treats group 1, assigns either action to group 2, and assigns control to group 3.

Relative to control for everyone, treating group 1 adds \(0.50(3)=1.5\) expected utility units per person in the population. Treating group 2 as well adds \(0.30(0)=0\), and treating group 3 would reduce value by \(0.20(2)=0.4\). The rule therefore gains \(1.5\) units per person over universal control. This comparison shows why a positive treatment effect is not sufficient: group 3’s benefit is positive, but its cost is larger.

A Continuously Varying Threshold

With continuous covariates, the same rule can be solved by algebra. The boundary between actions is where \(\tau(x)=c(x)\), provided such a boundary exists. On one side treatment has positive net advantage; on the other, it has negative net advantage. The boundary itself is a tie.

Worked Example: A Covariate-Dependent Cost

Suppose, hypothetically, that \(x\) is a risk score measured in score points and lies between 0 and 6. On a common utility scale, the conditional effect and incremental cost are

$$ \tau(x)=14-2x,\qquad c(x)=2+x. $$

Subtracting the cost gives the conditional net advantage:

$$ \Delta(x)=(14-2x)-(2+x)=12-3x. $$

Treatment is strictly better when \(12-3x>0\), or \(x<4\). At \(x=4\), the effect and cost are both \(6\) utility units, so either action is optimal. For \(x>4\), \(\Delta(x)<0\) and control is strictly better. For example, at \(x=2\), \(\tau(2)=10\), \(c(2)=4\), and \(\Delta(2)=6\), so treatment is preferred. At \(x=5\), \(\tau(5)=4\), \(c(5)=7\), and \(\Delta(5)=-3\), so control is preferred.

Conditions and Common Pitfalls

The derivation depends on a specific decision problem. It considers two available actions, a conditional expected benefit and incremental cost on a common utility scale, and no constraint that links one person’s assignment to another’s. The quantities \(\tau(x)\) and \(c(x)\) must refer to the same covariates used by the rule. If costs are random, \(c(x)\) is their conditional expected incremental cost, consistent with maximizing expected utility. If a cost is measured in money while the outcome is measured in a clinical unit, they cannot be subtracted directly without a specified conversion or utility function.

The rule also assumes that the conditional quantities used in the comparison are the target quantities of interest. In an application, causal identification conditions are needed to interpret \(\tau(x)\) as a treatment effect from data, and estimates of \(\tau(x)\) may be uncertain. The rule derived here describes the decision given the true conditional expectations; replacing them with estimates does not make the estimated rule automatically optimal in reality.

  • Ignoring cost. The claim “treat whenever \(\tau(x)>0\)” is correct only when incremental cost is zero on the utility scale. In general, compare \(\tau(x)\) with \(c(x)\).
  • Calling a tie a strict preference. If \(\tau(x)=c(x)\), treatment is not strictly better, but it is still a maximizing action. Control also maximizes at that value.
  • Mixing incompatible units. Subtracting dollars from a clinical outcome measured in, for example, symptom points has no defined meaning until both are represented on a common utility scale.
  • Assuming choices are always separable. A limit on total treatment capacity or another shared constraint can make one person’s choice affect feasible choices for others. The unconstrained pointwise proof does not cover that decision problem.

The key comparison is therefore \(\Delta(x)=\tau(x)-c(x)\). A positive value makes treatment strictly better, a negative value makes control strictly better, and zero means both actions attain the same conditional expected utility. Keeping those three cases distinct gives a precise treatment rule, including at its threshold.

Check Your Understanding

Use the conditional net utility comparison to answer each question.

  1. If \(\tau(x)=9\) utility units and \(c(x)=6\), which action is strictly better, and by how much?
  2. If \(\tau(x)=c(x)\), is treatment a maximizing action? Is it strictly better than control?
  3. A group has a positive conditional treatment effect of 3 units and an incremental cost of 5 units. What does the pointwise rule recommend, and why?
  4. If \(\tau(x)=16-2x\) and \(c(x)=4+x\), solve for the covariate value at which the actions tie and state the optimal action on either side.
  5. Name one kind of constraint that would prevent the unconstrained pointwise proof from applying directly.
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