From Treatment Effects to an Interpretable Rule
A treatment rule can be written as a function that maps a person's covariates to an action. A policy tree makes that function visible: a small number of questions about covariates lead to a treatment assignment. For example, a rule might ask whether a patient's baseline risk exceeds a threshold and assign treatment only on one side of that threshold.
The goal is not necessarily to explain every variation in conditional treatment effects. It is to find a useful, understandable policy: one whose estimated value is high while its sequence of decisions remains manageable. This requires distinguishing a tree built to predict treatment effects from one built to choose actions. A policy tree is optimized for the latter.
Suppose there are two actions, control and treatment, and the outcome \(Y\) is a utility measure for which larger values are preferred. Let \(Y(0)\) and \(Y(1)\) denote the potential outcomes under control and treatment. The conditional treatment effect is \(\tau(x)=E[Y(1)-Y(0)\mid X=x]\). For a policy \(\pi\), its value is the expected outcome under the action it assigns:
The second equality follows by taking conditional expectations given \(X\). The first term, \(E[Y(0)]\), is common to every policy in this comparison. Thus, when treatment has no additional cost outside the outcome scale, choosing a policy to maximize value is equivalent to choosing one to maximize its treatment gain \(E[\pi(X)\tau(X)]\). If costs are relevant, they must be incorporated into the utility or the treatment-effect quantity being optimized.
In practice the conditional effects are estimated. Write \(\widehat{\tau}_i\) for an estimated treatment effect at covariates \(X_i\) in the sample used for policy construction. The empirical objective for a tree \(T\) is commonly expressed as estimated gain,
An estimated baseline term can be added to compare estimated values, but it is the same for all policies in this expression and therefore does not change which tree maximizes estimated gain. This objective treats the estimated effects as inputs; the quality of the resulting policy still depends on how well those effects represent effects in the target population.
Choosing an Action at a Leaf
For a fixed tree structure, each leaf contains a set of observations. If the leaf assigns the same action to everyone in it, the estimated contribution to total gain from treating that leaf is the sum of its estimated effects. The best leaf action follows directly from comparing that sum with zero.
This rule is specific to choosing an action for a fixed leaf. It does not say that every person in the leaf has a positive estimated effect when the sum is positive; it says that treatment has the larger estimated aggregate gain for that leaf under the policy's shared action.
Worked Example: Choosing a Leaf Action
A hypothetical leaf contains five people with estimated treatment effects, in utility points per person, of \(1.0\), \(0.4\), \(-0.2\), \(-0.5\), and \(-0.1\). The policy must assign the same action to all five.
The sum of the estimated effects is
Treating the entire leaf therefore has an estimated total gain of \(0.6\) utility points over assigning control to everyone in the leaf, or an average gain of \(0.6/5=0.12\) utility points per person. Since the sum is positive, treatment is the estimated-value-maximizing leaf action. This does not imply that every person's estimated effect is positive: two of the five estimates are negative.
If the sum had been negative, control would maximize estimated gain for that leaf. If it were exactly zero, the two actions would tie under this objective. The calculation assumes that effects and outcomes use the same utility scale and that treatment has no omitted cost that should alter the comparison.
Splits Are Judged by Policy Gain
A split divides a node into child nodes. For each child, the tree can choose the better leaf action, so the split's estimated gain is the sum of the best gains in its children. In an unweighted sample, the best gain in a node \(I\), measured as a total rather than an average, is \(\max(0,\sum_{i\in I}\widehat{\tau}_i)\). Therefore a candidate split of node \(I\) into \(I_L\) and \(I_R\) has total estimated gain
The corresponding unsplit node has best gain \(\max(0,\sum_{i\in I}\widehat{\tau}_i)\). The split is useful to the estimated objective when its children’s total best gain exceeds the unsplit node’s best gain. This comparison evaluates how the split changes the treatment decisions and their estimated gains. A split that separates people with very different predicted outcomes but similar treatment effects may add no policy value.
Worked Example: Comparing Candidate Thresholds
Suppose a hypothetical sample has 100 people divided into four equally sized covariate bands, ordered from lowest to highest value of a risk score. The estimated treatment effects are \(0.8\), \(0.6\), \(-0.2\), and \(-0.7\) utility points per person in the four bands. Consider depth-one trees that split between adjacent bands. Each leaf assigns a single action to the bands it contains.
A split after band 1 gives the first band a total estimated treatment gain of \(25(0.8)=20\). The other leaf has total effect \(25(0.6-0.2-0.7)=-7.5\), so its best action is control and its gain is zero. Total gain is \(20\), or \(20/100=0.20\) per person.
A split after band 2 gives the lower two bands total gain \(25(0.8+0.6)=35\). The upper two bands have total effect \(25(-0.2-0.7)=-22.5\), so the best action there is control. The tree's gain is \(35/100=0.35\) utility points per person.
A split after band 3 gives the lower three bands total effect \(25(0.8+0.6-0.2)=30\), so that leaf is treated. The final band is assigned control, giving total gain \(30/100=0.30\) per person. Thus the split after band 2 has the largest estimated gain among these depth-one candidates. The depth-zero rule that treats everyone has gain \(25(0.8+0.6-0.2-0.7)/100=0.125\) per person, also below the best split. These are comparisons of estimated policy gains, not claims that the estimated effects are known without error.
The example illustrates a useful distinction: splits are not selected merely because the estimated effects differ across the resulting groups. They matter to the policy objective when that heterogeneity changes which actions are worthwhile and by enough to improve total gain.
Optimizing Under a Depth Constraint
A depth constraint limits how many sequential decisions the policy requires. To state the optimization problem, let \(\mathcal{T}_D\) be the set of allowable trees with depth at most \(D\), under a specified set of candidate split variables and thresholds. The empirical policy-learning problem is
The constraint matters because a tree with many splits can create very small leaves, depend on fragile distinctions, and be difficult to communicate or implement. Searching over a larger set of trees can also find a high objective by chance, especially when estimated treatment effects are noisy. A shallow tree limits complexity, but it may combine people whose treatment effects differ and therefore sacrifice some estimated value.
The depth limit alone does not specify the full optimization problem. One must also define the available covariates, candidate split points, minimum leaf sizes if required, and any restrictions on which actions can appear in leaves. These choices determine the set \(\mathcal{T}_D\), and hence what “best tree” means.
Worked Example: Estimated Value Versus Simplicity
Consider a hypothetical sample of 120 people. The estimated sum of their control outcomes is 1,090 utility points. A binary covariate \(X\) identifies 50 people with an estimated treatment effect of \(2.6\) points each; the other 70 have an estimated effect of \(-40/70\), approximately \(-0.5714\) points each. Suppose the policy can either use no split or split once on \(X\).
With depth zero, the tree must assign one action to everyone. Treating everyone has total estimated gain
Because this total is positive, the better depth-zero action is to treat everyone, giving estimated value \((1090+90)/120=1180/120\approx9.833\) utility points per person. Assigning control to everyone would instead give \(1090/120\approx9.083\).
With a depth-one split on \(X\), the tree treats the 50 people with positive estimated effects and assigns control to the other 70. Its total estimated gain is \(50(2.6)+70(0)=130\). Its estimated value, calculated before rounding, is
The split tree improves estimated value over the depth-zero rule by \((130-90)/120=0.3333\) utility points per person. It is also somewhat more complex because it requires one covariate check. If the implementation cost or interpretability requirement permits only depth zero, the simpler rule is the best allowable policy; if depth one is acceptable, the split has higher estimated value among these candidates. The comparison depends on the chosen complexity constraint, not on a universal rule that more complex trees are always preferable.
How to Search and What the Result Means
For a small set of allowable splits, one can enumerate candidate trees up to the depth limit, calculate the best leaf actions and estimated gain for each, and choose the maximizer. For larger searches, specialized optimization or tree-growing algorithms may be used. A greedy algorithm chooses the locally best available split at each step; it is computationally convenient, but it need not find the global maximizer over all trees satisfying the depth constraint. For example, a split that looks modest at the root could enable much better decisions in its descendants, while a locally attractive split might prevent those later distinctions.
Choose the maximum depth, candidate covariates and split points, and any leaf-size or action restrictions before comparing trees.
For each leaf, compare the total estimated effect of treating with zero; assign treatment when the total is positive and control when it is negative.
Sum the estimated gains across leaves and select the tree with the largest objective among the trees allowed by the constraint.
Report the covariate questions, leaf actions, depth, and estimated value, while making clear that the value is based on estimated effects.
The interpretation of the selected tree is concrete: it is a compact rule for assigning actions, together with an estimated objective under the data and estimates used to construct it. It is not proof that the rule is optimal in the population. In particular, searching over many candidate splits can make the winning estimated gain optimistic, since random estimation errors may help some candidates appear unusually favorable. Restricting depth improves interpretability and can reduce opportunities for such searching, but it does not by itself eliminate estimation error or guarantee good performance in new data.
Common Mistakes
- Confusing treatment-effect prediction with policy optimization. A tree need not reproduce every fluctuation in estimated effects. Its splits are valuable only insofar as they improve the action assignments and the objective.
- Including negative effects for people assigned control. A policy's treatment gain sums estimated effects only for people assigned treatment. Effects for people assigned control contribute zero treatment gain, not their estimated treatment effects.
- Calling a positive leaf sum proof of individual benefit. The sum supports treatment for the leaf under the shared-action objective; it does not establish positive effects for every person in that leaf.
- Assuming the best greedy split is the best constrained tree. A locally chosen split may not yield the best complete tree. Global optimality requires comparison or optimization over the full allowable class.
- Reporting the training objective as established future value. It is estimated from the effects and data used in construction. It should not be interpreted as a guaranteed population value.
Check Your Understanding
Try these questions using the definitions and objective above.
- A leaf has estimated effects \(-0.3\), \(0.2\), and \(0.4\). Which action maximizes estimated gain for the leaf, and what is the average gain per person?
- Explain why a split that improves prediction of outcomes but does not change the preferred action need not improve a policy tree's estimated gain.
- For 80 people, a tree treats 30 people with mean estimated effect \(0.5\) and assigns control to the other 50. What is its estimated treatment gain per person?
- What does a depth-one constraint allow, and what does it rule out?
- Why can a tree's maximum estimated gain be optimistic when many candidate splits are searched?