Tutorials › Causal Inference › Policy Value and Treatment Rules

Causal Learning and Policy Decisions · Tutorial 962 of 972

Policy Value and Treatment Rules

Learn to express a treatment rule as a covariate-based action and calculate its expected utility, including costs, for a target population.

Advanced 11 min read

What You'll Learn

  • Define a deterministic treatment rule using pretreatment covariates and available treatment actions
  • Define policy value as expected utility under the actions assigned by a rule
  • Include treatment costs in a common utility scale
  • Calculate a rule’s value by averaging across covariate groups
  • Compare the values of simple binary-treatment rules
  • Distinguish policy value from treatment-effect prediction accuracy

From Treatment Effects to Decisions

A conditional treatment effect describes how the average effect varies across people with different covariates. A policy question adds a decision: which action should be assigned to each person? To compare such decisions, we need a common measure of what a rule achieves across the population to which it will be applied. That measure is the rule’s value.

Value depends on more than treatment effects alone. The untreated outcome matters, as do treatment costs and the distribution of covariates in the target population. A rule that assigns treatment to a group with a positive treatment effect may still have low value if the cost is large. Conversely, a rule’s value can be high because untreated outcomes are favorable, even if its treatment assignments add little.

Definition: Let \(X\) denote pretreatment covariates, and let the available actions be \(A\in\{0,1\}\), where 0 is no treatment and 1 is treatment. A deterministic treatment rule is a function \(d\) that maps each covariate value \(x\) to an action: \(d(x)\in\{0,1\}\). The rule assigns action \(d(X)\) to a person with covariates \(X\).

The rule is based on covariates available when the decision is made. It does not assign different actions to otherwise identical people using information that would only become known after treatment. A rule may be simple, such as “treat everyone,” or may depend on several features, such as age group and baseline risk.

Defining Value When Treatment Has a Cost

Let \(Y(0)\) and \(Y(1)\) be the potential outcomes under no treatment and treatment. Suppose larger values of \(Y\) are better, and let \(c(X)\) be the cost of treatment, expressed in the same units as the outcome’s utility. No-treatment cost is taken to be zero. The net utility under action \(a\) is \(Y(a)-a c(X)\). Thus a treated person’s outcome is reduced by the treatment cost, while an untreated person’s outcome is not.

Definition: For a target population with covariate distribution \(P(X)\), the value of a treatment rule \(d\) is its expected net utility: $$ V(d)=E\left[Y(d(X))-d(X)c(X)\right]. $$ The expectation averages over people in the target population, including their covariates and potential outcomes. Value is measured in the same units as net utility.

This definition uses a linear cost subtraction, which is appropriate when costs can be translated into the same utility scale as outcomes. If a decision problem instead specifies a different utility function, value is the expected utility under the rule: \(V(d)=E[U(d(X),X)]\), where \(U(a,X)\) is the relevant potential utility for action \(a\) and covariates \(X\). The units and meaning of the utility must be clear before comparing rules.

For example, if \(Y\) is measured in quality-adjusted days and treatment costs have been converted to quality-adjusted days, the net utility is also in quality-adjusted days. Subtracting a dollar cost directly from a health outcome measured in days would not be meaningful without a specified conversion. Treatment harms may be included in \(Y(a)\), in the cost term, or in a broader utility function, but they must not be omitted or counted twice.

Calculating Value Across Covariate Groups

For a population divided into a finite number of covariate groups, value is a weighted average of the group-specific expected utilities. Let group \(j\) have population proportion \(p_j\), and let the rule assign action \(d_j\) in that group. Then

$$ V(d)=\sum_j p_j\left\{E[Y(d_j)\mid X\text{ in group }j]-d_j E[c(X)\mid X\text{ in group }j]\right\}. $$

When costs are constant within each group, the conditional expected cost is just that group’s cost. The group proportions must sum to 1. They weight the group utilities because the policy is being evaluated for the target population, not for a population in which every group is equally common.

Worked Example: Valuing a Rule That Treats One Group

Consider a hypothetical target population with two covariate groups. Younger patients make up 60% of the population and older patients 40%. The outcomes are measured in quality-adjusted days, and the treatment costs are converted into the same units. The expected outcomes and costs are:

GroupPopulation proportionMean outcome without treatmentMean outcome with treatmentTreatment cost
Younger0.608 days11 days2 days
Older0.407 days8 days2 days

Suppose the rule treats younger patients and does not treat older patients. In the younger group, mean net utility under treatment is \(11-2=9\) quality-adjusted days. In the older group, the rule assigns no treatment, so mean utility is 7 quality-adjusted days. Therefore,

$$ V(d)=0.60(9)+0.40(7)=5.40+2.80=8.20\text{ quality-adjusted days}. $$

For comparison, treating nobody has value \(0.60(8)+0.40(7)=7.60\) quality-adjusted days. Treating everybody has value \(0.60(11-2)+0.40(8-2)=0.60(9)+0.40(6)=7.80\) quality-adjusted days. These calculations compare the rules for this specified population and set of utilities; they do not establish that any one rule is best in another population or under another cost structure.

Connecting Value to Conditional Treatment Effects

Define the conditional average treatment effect (CATE) for the outcome \(Y\) as \(\tau(x)=E[Y(1)-Y(0)\mid X=x]\). Under the linear net-utility definition, the expected net utility for a person with covariates \(x\) is \(E[Y(0)\mid X=x]+d(x)\{\tau(x)-c(x)\}\). Averaging this expression gives a useful way to express policy value relative to the rule that treats nobody.

Result: If the treatment cost is \(c(X)\) and net utility is \(Y(a)-a c(X)\), then $$ V(d)=E[Y(0)]+E\left[d(X)\{\tau(X)-c(X)\}\right]. $$ In particular, \(V(d)-V(0)=E[d(X)\{\tau(X)-c(X)\}]\), where \(0\) denotes the rule that assigns no treatment to anyone.

To justify the result, condition on \(X=x\). If \(d(x)=0\), expected net utility is \(E[Y(0)\mid X=x]\). If \(d(x)=1\), it is \(E[Y(1)\mid X=x]-c(x)\). The identity \(E[Y(1)\mid X=x]=E[Y(0)\mid X=x]+\tau(x)\) makes these two cases one expression:

$$ E[Y(d(x))-d(x)c(x)\mid X=x] =E[Y(0)\mid X=x]+d(x)\{\tau(x)-c(x)\}. $$

Taking expectation over \(X\) yields the stated formula. This also explains why CATE estimates alone do not generally give a rule’s absolute value: the formula includes \(E[Y(0)]\), the population’s expected outcome without treatment. For comparing a rule with treating nobody, that baseline term cancels, leaving the rule’s average incremental net utility.

Worked Example: Comparing Rules Using Group-Specific Effects

In a hypothetical population, 50% of people are in group A and 50% are in group B. Mean outcomes and treatment costs, all expressed in utility units, are as follows:

GroupProportionMean outcome without treatmentMean outcome with treatmentCost
A0.50471
B0.50672

The treatment effects are \(7-4=3\) units in group A and \(7-6=1\) unit in group B. The incremental net utilities from treatment are therefore \(3-1=2\) units in A and \(1-2=-1\) unit in B. Consider the rule that treats A and not B. Its value is

$$ V(d)=0.50(7-1)+0.50(6)=3.00+3.00=6.00\text{ utility units}. $$

Treating nobody has value \(0.50(4)+0.50(6)=5.00\) units, so the rule’s incremental value relative to that rule is \(6.00-5.00=1.00\) unit. The same difference follows from the CATE expression: \(0.50(2)+0.50(0)=1.00\). A rule that treats both groups has value \(0.50(6)+0.50(5)=5.50\) units; its incremental value relative to treating nobody is \(0.50(2)+0.50(-1)=0.50\) units.

The group proportions matter: the incremental contribution from treating a group is its population proportion multiplied by that group’s incremental net utility. A large treatment effect in a very small group need not contribute more to population value than a modest effect in a common group.

Target Population and Causal Interpretation

A policy value is always tied to a target population. The same rule can have different values in populations with different covariate distributions, even when the conditional potential-outcome means are the same. For example, a rule that treats only group A contributes more incremental value in a population where group A is more common, provided its within-group utility is unchanged.

The definition is causal because it refers to the outcomes that would result under the action the rule assigns. In a hypothetical calculation where the conditional potential-outcome means are supplied, the arithmetic directly computes the value for the specified population. In an observational data analysis, those means are not automatically observed for every person under both actions. Causal identification requires conditions such as consistency, no interference, conditional exchangeability given the measured covariates, and positivity for the actions considered. In brief: observed treatment and outcome must correspond to the defined intervention; one person’s action must not change another’s outcome; measured covariates must suffice to remove treatment-outcome confounding; and both actions must be possible at relevant covariate values. Without appropriate identification, a numerical estimate of value may reflect confounding or unsupported extrapolation rather than the policy’s causal value.

Policy value is also distinct from prediction accuracy. A model can predict outcomes well overall but produce poor treatment decisions if it misestimates effects in groups where decisions change. Conversely, a rule’s value is an average utility quantity; it does not by itself show how precisely that value has been estimated or how much it might vary in a new sample.

Worked Example: Why Population Weights Change Value

Suppose a hypothetical rule treats only group A. Treatment adds 2 utility units net of cost in group A, while the rule assigns no treatment to group B. If group A is 20% of the target population, the rule’s incremental value over treating nobody is

$$ 0.20(2)+0.80(0)=0.40\text{ utility units per person}. $$

If instead group A makes up 70% of the target population, with the same within-group utilities, the incremental value is

$$ 0.70(2)+0.30(0)=1.40\text{ utility units per person}. $$

The group-specific treatment contribution has not changed; only the target population composition has changed. The two values answer different population-level questions and should not be treated as contradictory.

Common Mistakes

  • Reporting an effect instead of a value. A treatment effect describes a contrast between actions; policy value averages the utility of the actions actually assigned, including the untreated baseline and any costs.
  • Ignoring treatment costs. A positive outcome effect does not guarantee positive incremental net utility. Subtract the cost on the same utility scale before calculating the contribution of treatment.
  • Averaging groups equally. Unless groups are equally common in the target population, an unweighted average does not represent that population’s policy value.
  • Using the wrong population distribution. Value calculated for one population need not carry over to a population with different covariate composition.
  • Calling a calculated value universally optimal. A value comparison applies to the specified rules, utilities, costs, and target population. It does not establish that the preferred rule remains preferable under different conditions.
Key takeaway: A treatment rule maps pretreatment covariates to actions; its value is the target-population average utility produced by those actions. For binary treatment with a cost subtracted from outcome utility, calculate each assigned action’s expected net utility and average using the target population’s covariate distribution.

Check Your Understanding

Use the definitions and calculations above to answer these questions.

  1. In your own words, what makes a function from covariates to actions a treatment rule?
  2. A rule treats a group comprising 30% of a population. Treatment has an incremental net utility of 4 units in that group, and the rule leaves everyone else untreated. What is its incremental value over treating nobody?
  3. Why can a CATE estimate alone be insufficient to calculate a rule’s absolute value?
  4. Two populations have identical group-specific potential-outcome means but different proportions in each group. Must a rule have the same value in both populations? Explain.
  5. Name two conditions needed to interpret a value calculated from observational data as a causal policy value.
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