Learning Where to Look: Delaunay Matching for Discovery and Data Collection

arXiv:2607.04885 · econ.EM · Submitted 2026-07-06 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Learning Where to Look".

Warren: Here's the gist of the paper. This paper studies data-driven policy choices from a geometric perspective by deriving affordability guarantees for estimators when finite and large samples are available.

Ruth: The Decision Framework The policymaker must commit on an estimator of the target’s innovation effect that pools information across donors before the donor sample is drawn.

Warren: The PM seeks an estimator that keeps the average compensation cost below a finite budget given the collection design and for any target population.

Ruth: A key theoretical device introduced is certification, which provides affordability conditions by controlling the probability of making a mistake uniformly.

Paper summary: Ruth: So, looking at the title "Learning Where to Look: Delaunay Matching for Discovery and Data Collection," it really captures the essence of what they did. It’s not just about finding the best estimator; it’s about figuring out where to look next both among existing donors and when you need to collect new data.

Warren: I agree. What this paper shows is that when you have a finite budget, the way you choose your weights for matching estimators matters a lot, and geometric structures like the Delaunay triangulation give us a very specific, provably affordable way to handle those choices.

Iris: It means for people actually making these policy decisions that they can use this geometry to guide their data collection planning. Instead of just randomly sampling donors within the area where you already have data, they can target locations that are most informative according to this geometric rule.

Ruth: Exactly. The implication for the world, or rather for policymakers, is that designing a cost-efficient rollout strategy isn't just about getting more data; it’s about using spatial relationships between your donor and target populations to minimize the risk of making costly mistakes.

Warren: It’s a practical tool here because they show that adopting this geometric perspective can halve the cost by adding just three donor units in their empirical application, whereas random sampling takes twenty additions to reach a similar level of performance.

Iris: The main limitation they point out is that the assumptions used for these strong guarantees rely on certain regularity conditions, and they are focusing specifically on matching estimators with positive weights. So if your setup doesn't fit those exact conditions, you might need to be more careful with how you interpret the results.

Conclusion: Ruth: So we're wrapping up this look at "Learning Where to Look: Delaunay Matching for Discovery and Data Collection." Basically, these authors are showing how you can use geometry, specifically Delaunay triangulation, to figure out the best way to design data collection plans when you're trying to make policy choices.

Warren: It boils down to finding a way for a policymaker who has a limited budget to make decisions about innovation effects that stay affordable no matter what target population turns up. They’re talking about this specific technique called certification, which sets an affordability boundary based on controlling the probability of making a mistake uniformly across everyone.

Iris: And the core idea they push is that when you have large samples available, this whole problem simplifies down to a geometric one with a unique answer under standard conditions. They show that Delaunay interpolants give you the smallest budget needed to cover any target point within the area defined by your initial donor collection design.

Ruth: That's what I found really interesting—it takes something abstract about statistical decision theory and grounds it in concrete geometry. It’s not just a theoretical exercise; it directly informs where someone should look for new donor observations to reduce the risk of a bad policy choice.

Warren: And the empirical results they show are pretty compelling for me. They use an experiment based on smartcards, and this geometric plan—this Delaunay Matching approach—halved the cost by only adding three more donor units compared to a random collection strategy that would need twenty additions to get there. That's a big difference in terms of what it costs to gather the necessary information.

Iris: But we have to remember the caveats they laid out. These strong affordability guarantees depend on certain regularity conditions being met, and they’re specifically focusing on matching estimators with positive weights, so we can’t just take this exact setup and apply it blindly without checking those assumptions first.

Ruth: Right, so it's a powerful tool for designing cost-efficient data collection plans when you have these specific conditions in place. It changes how we think about the practical side of causal inference in policy.

Warren: Exactly. If you’re thinking about designing a rollout strategy, this suggests that using spatial information to guide your next data collection move is much smarter than just picking donors at random within the general area.

Iris: So what does this mean for future work? They are setting up the groundwork for using these geometric ideas in more complex scenarios, moving beyond this specific matching estimator setup.

Ruth: Right, so we've seen how geometry helps us solve a core problem in policy choice under budget constraints, and now we can look at how to apply these spatial insights elsewhere.

University of Bologna

econ.EM

Submitted: 2026-07-06

Updated: 2026-11-06

License: http://creativecommons.org/publicdomain/zero/1.0/

Importance score: 89/100

The gist: This paper studies data-driven policy choices from a geometric perspective by deriving affordability guarantees for estimators when finite and large samples are available.

Key concepts

Certification
A theoretical tool used to guarantee affordability by controlling the probability of making a mistake uniformly across all possible target groups. It sets bounds on the maximum possible compensation cost based on a chosen error rate.
Delaunay Matching Estimator
An estimator derived from computational geometry, using barycentric interpolation on the Delaunay triangulation of donor covariates. This estimator is shown to provide the smallest budget required to uniformly cover compensation costs across target populations.
Geometric Plan
A new data collection rule motivated by the optimality of Delaunay Matching. It suggests collecting new donor observations at locations where current approximations are weakest, specifically at the center of the smallest circle containing the largest simplex.

Terminology

Summary

This paper studies data-driven policy choices from a geometric perspective by deriving affordability guarantees for estimators when finite and large samples are available. p. 0

Where the paper says it

Learning Where to Look: Delaunay Matching for Policy Choice and Data Collection Giacomo Opocher∗ July 28, 2026 Abstract. This paper studies data-driven policy choices from a geometric perspective. p. 0

Where the paper says it

Learning Where to Look: Delaunay Matching for Policy Choice and Data Collection Giacomo Opocher∗ July 28, 2026 Abstract. This paper studies data-driven policy choices from a geometric perspective.

The Decision Framework

The policymaker must commit on an estimator of the target’s innovation effect that pools information across donors before the donor sample is drawn. The PM seeks an estimator that keeps the average compensation cost below a finite budget given the collection design and for any target population. p. 1

Where the paper says it

A donor sample is collected over a covariate space and a collection design is defined as the exact covariate values that realize in the donor sample. The PM seeks an estimator that keeps the average compensation cost below a finite budget given the collection design and for any target population.

p. 1
Where the paper says it

A donor sample is collected over a covariate space and a collection design is defined as the exact covariate values that realize in the donor sample. The PM seeks an estimator that keeps the average compensation cost below a finite budget given the collection design and for any target population.

A key theoretical device introduced is certification, which provides affordability conditions by controlling the probability of making a mistake uniformly. p. 2
Where the paper says it

I introduce certification as a theoretical device to provide affordability conditions. An estimator produces certified decisions if the probability of making a mistake is uniformly controlled, provided that the innovation effects are large enough in absolute value.

Lemma 1 shows that the worst-case compensation cost is bounded above by the maximum between gamma(m) and alpha taubar. p. 0
Where the paper says it

In the latter case, Delaunay interpolants, whose properties are well-known from results in computational geometry, deliver the smallest budget that covers the compensation cost uniformly over the admissible target populations, and conditional on the donor collection design. This result informs where to look for new donor observations to decrease the worst-case compensation cost the most.

Geometric Optimality

The paper connects certification constraints with well-known results in computational geometry, defining the Delaunay Matching Estimator by barycentric interpolation on the Delaunay triangulation of the donor covariates. This estimator requires the smallest lower bound on intervention effects to certify decisions at every feasible target point within the convex hull of the donor collection design. p. 3

Where the paper says it

requires the smallest lower bound on intervention effects to certify decisions at every feasible target point. Feasible target points belong to the convex hull of the donor collection design.

Theorem 3 proves that Delaunay weights solve the local geometric approximation problem over the full class of positive affine-exact weights pointwise. p. 2
Where the paper says it

Then, I study a sufficient condition for an estimator to provide certified decisions in general, and specialize this condition to the set of linearly precise matching estimators with positive weights. I derive a finite-sample certification condition for this class of estimators, and show that if we let the size of the donor and target samples grow we can study the problem from a purely geometric perspective.

Data Collection Strategies

The paper motivates a new collection rule called the geometric plan by using the optimality of Delaunay Matching together with geometric results in Waldron (1998) to guide where to look for new donor observations. The geometric plan adds donor observations at locations where the current approximation is weakest, which is the center of the smallest circle that contains the largest simplex. p. 3

Where the paper says it

The rule adds donor observations at locations where the current approximation is weakest. Waldron (1998) proves that the worst-case interpolation error is bounded by the square of the radius of the smallest ball that contains the largest simplex in the triangulation.

This collection plan is shown empirically to deliver sizeable gains relative to random collection within the donors’ convex hull. p. 3
Where the paper says it

This bound is achieved in the center of such a ball which therefore constitutes the worst target we could take a decision for and the first point to add in the donor data according to the geometric plan. This collection plan is shown empirically to deliver sizeable gains relative to random collection within the donors’ convex hull.

Empirical Validation

In a semi-synthetic application based on the Smartcards experiment, the Delaunay Matching Estimator takes a decision for 21 out of 25 target cells and certifies 16 cells asymptotically. The geometric plan achieves the target level with only three additional donor points, whereas a standard collection plan that picks new donors at random within the convex hull of the collection design fails to reach it within twenty steps. p. 4

Where the paper says it

I then consider a policymaker who wants to halve the current worst-case compensation cost by collecting additional donor data. The geometric plan achieves the target level with only three additional donor points, whereas a standard collection plan that picks new donors at random within the convex hull of the collection design fails to reach it within twenty steps.

This finding suggests sizeable gains from adopting a geometric perspective in designing cost-efficient collection plans for policy choice problems. p. 4
Where the paper says it

The geometric plan achieves the target level with only three additional donor points, whereas a standard collection plan that picks new donors at random within the convex hull of the collection design fails to reach it within twenty steps. This finding suggests sizeable gains from adopting a geometric perspective in designing cost-efficient collection plans for policy choice problems.

Conclusion

The paper concludes that Delaunay Matching provides the best asymptotic affordability guarantee in the set of matching estimators with positive weights. It also demonstrates how geometric collections deliver substantial gains relative to random collection. p. 3

Where the paper says it

This bound is achieved in the center of such a ball which therefore constitutes the worst target we could take a decision for and the first point to add in the donor data according to the geometric plan. This collection plan is shown empirically to deliver sizeable gains relative to random collection within the donors’ convex hull.

This work contributes to the broader literatures on statistical decision theory and causal inference highlighting the value of studying the performance of data-driven decisions from a geometric perspective. p. 6
Where the paper says it

and overlap, (ii) they maximize empirical welfare over the policy space while holding the matching procedure fixed, and (iii) their guarantees are not conditional on the collection design. Overall, this paper contributes to the broader literatures on statistical decision theory and causal inference highlighting the value of studying the performance of data-driven decisions from a geometric perspective.

The rest of the paper is organized as follows. p. 6
Where the paper says it

Overall, this paper contributes to the broader literatures on statistical decision theory and causal inference highlighting the value of studying the performance of data-driven decisions from a geometric perspective. The rest of the paper is organized as follows.

">

--- Page 1 ---

Learning Where to Look: Delaunay Matching for Policy Choice and Data Collection

Giacomo Opocher∗ July 28, 2026 Abstract. This paper studies data-driven policy choices from a geometric perspective. A sample from a donor population fully exposed to an innovation informs a policymaker (PM) on whether to innovate groups in a target population where the innovation was not introduced. p. 0

Where the paper says it

This paper studies data-driven policy choices from a geometric perspective. A sample from a donor population fully exposed to an innovation informs a policymaker (PM) on whether to innovate groups in a target population where the innovation was not introduced.

Any wrong decision must be compensated from a finite budget and the PM seeks an estimator of the innovation’s effect that guarantees an affordable compensation cost. p. 0
Where the paper says it

A sample from a donor population fully exposed to an innovation informs a policymaker (PM) on whether to innovate groups in a target population where the innovation was not introduced. Any wrong decision must be compensated from a finite budget and the PM seeks an estimator of the innovation’s effect that guarantees an affordable compensation cost.

I focus on matching estimators with positive weights and derive affordability guarantees when finite and large samples of the populations of interest are available. p. 0
Where the paper says it

Any wrong decision must be compensated from a finite budget and the PM seeks an estimator of the innovation’s effect that guarantees an affordable compensation cost. I focus on matching estimators with positive weights and derive affordability guarantees when finite and large samples of the populations of interest are available.

In the latter case, Delaunay interpolants, whose properties are well-known from results in computational geometry, deliver the smallest budget that covers the compensation cost uniformly over the admissible target populations, and conditional on the donor collection design. p. 0
Where the paper says it

I focus on matching estimators with positive weights and derive affordability guarantees when finite and large samples of the populations of interest are available. In the latter case, Delaunay interpolants, whose properties are well-known from results in computational geometry, deliver the smallest budget that covers the compensation cost uniformly over the admissible target populations, and conditional on the donor collection design.

This result informs where to look for new donor observations to decrease the worst-case compensation cost the most. p. 0
Where the paper says it

In the latter case, Delaunay interpolants, whose properties are well-known from results in computational geometry, deliver the smallest budget that covers the compensation cost uniformly over the admissible target populations, and conditional on the donor collection design. This result informs where to look for new donor observations to decrease the worst-case compensation cost the most.

In an empirical application, I show that such collection plans halve the cost by adding three donor units, while random sampling fails to reach the same target within twenty additions. p. 0
Where the paper says it

This result informs where to look for new donor observations to decrease the worst-case compensation cost the most. In an empirical application, I show that such collection plans halve the cost by adding three donor units, while random sampling fails to reach the same target within twenty additions.

Keywords: Statistical Decision Theory, Matching, Delaunay Triangulation. p. 0
Where the paper says it

In an empirical application, I show that such collection plans halve the cost by adding three donor units, while random sampling fails to reach the same target within twenty additions. Keywords: Statistical Decision Theory, Matching, Delaunay Triangulation.

--- Page 2 ---

1 Introduction Policymakers (PMs) often need to decide whether to introduce an innovation across groups in a target population where no unit has ever received it before. Evidence may instead come from a random sample of a distinct donor population where all units were exposed to the innovation. p. 1

Where the paper says it

1 Introduction Policymakers (PMs) often need to decide whether to introduce an innovation across groups in a target population where no unit has ever received it before. Evidence may instead come from a random sample of a distinct donor population where all units were exposed to the innovation.

Non-random staggered roll-out policies fit this description. p. 1
Where the paper says it

Evidence may instead come from a random sample of a distinct donor population where all units were exposed to the innovation. Non-random staggered roll-out policies fit this description.

The PM decides who should be innovated first following fairness, equality, or any non-data driven and non-random concerns; once these groups are innovated, she wants to learn how to proceed further with the roll-out. p. 1
Where the paper says it

Non-random staggered roll-out policies fit this description. The PM decides who should be innovated first following fairness, equality, or any non-data driven and non-random concerns; once these groups are innovated, she wants to learn how to procede further with the roll-out.

In such settings, the policymaker must determine both which donors are informative about each target group and, when existing evidence is insufficient, where additional donor data should be collected. p. 1
Where the paper says it

The PM decides who should be innovated first following fairness, equality, or any non-data driven and non-random concerns; once these groups are innovated, she wants to learn how to procede further with the roll-out. In such settings, the policymaker must determine both which donors are informative about each target group and, when existing evidence is insufficient, where additional donor data should be collected.

Matching estimators constitute a natural option for solving this problem. p. 1
Where the paper says it

In such settings, the policymaker must determine both which donors are informative about each target group and, when existing evidence is insufficient, where additional donor data should be collected. Matching estimators constitute a natural option for solving this problem.

Yet, that same flexibility can leave the choice of weights ambiguous. p. 1
Where the paper says it

Abadie and Imbens, 2006; Abadie, 2021). Yet, that same flexibility can leave the choice of weights ambiguous.

Several donor combinations, regardless of the distance they cover around the target unit, can fit it equally well in the matching space. p. 1
Where the paper says it

Yet, that same flexibility can leave the choice of weights ambiguous. Several donor combinations, regardless of the distance they cover around the target unit, can fit it equally well in the matching space.

Therefore, weighting schemes that are equally justified ex-ante can lead to dramatic differences in performance. p. 1
Where the paper says it

Several donor combinations, regardless of the distance they cover around the target unit, can fit it equally well in the matching space. Therefore, weighting schemes that are equally justified ex-ante can lead to dramatic differences in performance (see Abadie and L’Hour, 2021).

--- Page 3 ---

1 Introduction This paper studies this research design problem from a new (statistical) decision theoretic angle. The PM must commit on an estimator of the target’s innovation effect that pools information across donors before the donor sample is drawn. p. 1

Where the paper says it

This paper studies this research design problem from a new (statistical) decision theoretic angle. The PM must commit on an estimator of the target’s innovation effect that pools information across donors before the donor sample is drawn.

A donor sample is collected over a covariate space and a collection design is defined as the exact covariate values that realize in the donor sample. p. 1
Where the paper says it

The PM must commit on an estimator of the target’s innovation effect that pools information across donors before the donor sample is drawn. A donor sample is collected over a covariate space and a collection design is defined as the exact covariate values that realize in the donor sample.

The PM seeks an estimator that keeps the average compensation cost below a finite budget given the collection design and for any target population. This goal disciplines the choice of weights and, when large samples are available, reduces it to a purely geometric problem with a unique solution under standard regularity conditions. p. 1
Where the paper says it

The PM seeks an estimator that keeps the average compensation cost below a finite budget given the collection design and for any target population. This goal disciplines the choice of weights and, when large samples are available, reduces it to a purely geometric problem with a unique solution under standard regularity conditions.

Not only does this perspective provide analytical guarantees for the resulting choice of weights, but it also guides the design of data collection plans that empirically outperform random sampling in reducing worst-case compensation cost. p. 1
Where the paper says it

This goal disciplines the choice of weights and, when large samples are available, reduces it to a purely geometric problem with a unique solution under standard regularity conditions. Not only does this perspective provide analytical guarantees for the resulting choice of weights, but it also guides the design of data collection plans that empirically outperform random sampling in reducing worst-case compensation cost.

In this sense, this work studies two related questions: where to look among existing donors to inform policy choices, and where to look next when collecting new donor observations. p. 2
Where the paper says it

related questions: where to look among existing donors to inform policy choices, and where to look next when collecting new donor observations. The setting of interest is defined in further detail as follows.

--- Page 4 ---

The setting of interest is defined in further detail as follows. Consider two counterfactual states: the status quo and one innovation. p. 2

Where the paper says it

The setting of interest is defined in further detail as follows. Consider two counterfactual states: the status quo and one innovation.

A policymaker (PM) decides whether to innovate groups in a target population. p. 2
Where the paper says it

Consider two counterfactual states: the status quo and one innovation. A policymaker (PM) decides whether to innovate groups in a target population.

The donor and target populations are distinct. p. 2
Where the paper says it

A policymaker (PM) decides whether to innovate groups in a target population. The donor and target populations are distinct.

The PM commits to an estimator of the innovation effect, observes a sample drawn from an innovated donor population according to a given collection design, and then observes the target units. p. 2
Where the paper says it

The donor and target populations are distinct. The PM commits to an estimator of the innovation effect, observes a sample drawn from an innovated donor population according to a given collection design, and then observes the target units.

Decisions are made by a threshold rule that takes the value one if the estimated innovation effect for units in the target is strictly positive. p. 2
Where the paper says it

The PM commits to an estimator of the innovation effect, observes a sample drawn from an innovated donor population according to a given collection design, and then observes the target units. Decisions are made by a threshold rule that takes the value one if the estimated innovation effect for units in the target is strictly positive.

The PM must compensate any wrong decision out of a finite compensation budget. p. 2
Where the paper says it

Decisions are made by a threshold rule that takes the value one if the estimated innovation effect for units in the target is strictly positive. The PM must compensate any wrong decision out of a finite compensation budget.

For simplicity, compensation equals the innovation effect that would (not) have occurred had the PM chosen (not) to innovate. p. 2
Where the paper says it

The PM must compensate any wrong decision out of a finite compensation budget. For simplicity, compensation equals the innovation effect that would (not) have occurred had the PM chosen (not) to innovate.

Once the target population realizes, decisions are made and the compensation cost must be paid out. p. 2
Where the paper says it

For simplicity, compensation equals the innovation effect that would (not) have occurred had the PM chosen (not) to innovate. Once the target population realizes, decisions are made and the compensation cost must be paid out.

--- Page 5 ---

2 Formal Description of the PM’s Objective Denote by (Yi(0), Yi(1)) the potential outcomes of unit i under the status quo and the innovation respectively. Denote by (Xi, Ui(0), Ui(1)) a random vector containing a set of continuous covariates Xi and idiosyncratic factors (Ui(0), Ui(1)) that depend on the counterfactual state. p. 6

Where the paper says it

2 Formal Description of the PM’s Objective Denote by (Yi(0), Yi(1)) the potential outcomes of unit i under the status quo and the innovation respectively. Denote by (Xi , Ui(0), Ui(1)) a random vector containing a set of continous covariates Xi and idiosyncratic factors (Ui(0), Ui(1)) that depend on the counterfactual state.

Let d ∈ [0, 1] denote the state. p. 6
Where the paper says it

Denote by (Xi , Ui(0), Ui(1)) a random vector containing a set of continous covariates Xi and idiosyncratic factors (Ui(0), Ui(1)) that depend on the counterfactual state. Let d ∈ {0, 1} denote the state.

Assumption 1 (Potential Outcomes Function) Assume the potential outcomes functions have the following form: Yi(d) = fd(Xi) + Ui(d) (1) where fd belongs to the admissible functions set Fc defined as f0, f1 ∈ Fc:= (f: X → Y: sup x∈X ∇2 f(x) ≤ c < ∞) for any x ∈ X, and Yi(1) − Yi(0) ≤ taubar <sup

--- Page 6 ---

c < ∞. Assumption 2 (Distributions Family) Define P = over X × U2: ER[Ui(d)Xi] = 0, Ui(d) ≤ Ubar d < ∞, for d ∈ [0, 1]. p. 7

Where the paper says it

(4) The probability laws in the donor and target population, P and Q, are distinct but belong to the same family P. Assumption 2 (Distributions Family) Define P = {R over X × U2 : ER[Ui(d)|Xi ] = 0, |Ui(d)| ≤ U¯ d < ∞, for d ∈ {0, 1}}.

Assumption 2 imposes that any distribution belonging to P, whose typical element is indexed by R, gives rise to bounded idiosyncratic factors with zero conditional mean. p. 7
Where the paper says it

Assumption 2 (Distributions Family) Define P = {R over X × U2 : ER[Ui(d)|Xi ] = 0, |Ui(d)| ≤ U¯ d < ∞, for d ∈ {0, 1}}. Assumption 2 imposes that any distribution belonging to P, whose typical element is indexed by R, gives rise to bounded idiosyncratic factors with zero conditional mean.

Taken together, Assumptions 1 and 2 allow for P and Q to have arbitrarily different covariate distributions and different idiosyncratic term distributions as long as Assumption 2 is satisfied. p. 7
Where the paper says it

Assumption 2 imposes that any distribution belonging to P, whose typical element is indexed by R, gives rise to bounded idiosyncratic factors with zero conditional mean. Taken together, Assumptions 1 and 2 allow for P and Q to have arbitrarily different covariate distributions and different idiosyncratic term distributions as long as Assumption 2 is satisfied.

Given these two, the distributions of potential outcomes are determined by the causal law in Assumption 1 which is common in P. p. 7
Where the paper says it

Taken together, Assumptions 1 and 2 allow for P and Q to have arbitrarily different covariate distributions and different idiosyncratic term distributions as long as Assumption 2 is satisfied. Given these two, the distributions of potential outcomes are determined by the causal law in Assumption 1 which is common in P.

Define the individual innovation effect and conditional average innovation effect for a unit in the target as: τj:= Yj (1) − Yj (0), τ(x):= EQ[τj Xj = x]. p. 7
Where the paper says it

Given these two, the distributions of potential outcomes are determined by the causal law in Assumption 1 which is common in P. Define the individual innovation effect and conditional average innovation effect for a unit in the target as: τj := Yj (1) − Yj (0), τ (x) := EQ[τj | Xj = x].

--- Page 7 ---

Definition 1 Innovation Effect Estimator Given a donor sample SN, define τˆj = m(Xj, Yj (0), SN) (6) Therefore, an estimator is a function that takes as input the donor sample SN and one target draw j with random values (Xj, Yj (0)), and gives as output an estimate of the innovation effect. Let Nx:= XJ j=1 1if Xj = x, and any x such that Nx ≥ 1, define the innovation effect estimate τˆm(x):= 1/Nx XJ j=1 m(Xj, Yj (0), SN) 1if Xj = x. p. 8

Where the paper says it

(6) Therefore, an estimator is a function that takes as input the donor sample S N and one target draw j with random values (Xj , Yj (0)), and gives as output an estimate of the innovation effect. Let Nx := X J j=1 1{Xj = x}, (7) and any x such that Nx ≥ 1, define the innovation effect estimate τˆm(x) := 1 Nx X J j=1 m(Xj , Yj (0), SN )1{Xj = x}.

Given the estimate ˆτm(x), the policy rule is defined as ˆdm(x):= 1if τˆm(x) ≥ 0. p. 8
Where the paper says it

Let Nx := X J j=1 1{Xj = x}, (7) and any x such that Nx ≥ 1, define the innovation effect estimate τˆm(x) := 1 Nx X J j=1 m(Xj , Yj (0), SN )1{Xj = x}. (8) Given the estimate ˆτm(x), the policy rule is defined as ˆdm(x) := 1{τˆm(x) ≥ 0}.

Define the oracle decision rule as d∗(x):= 1if τ (x) ≥ 0. p. 8
Where the paper says it

(8) Given the estimate ˆτm(x), the policy rule is defined as ˆdm(x) := 1{τˆm(x) ≥ 0}. (9) Define the oracle decision rule as d ∗ (x) := 1{τ (x) ≥ 0}.

--- Page 8 ---

Definition 2 PM’s Objective Assume the PM needs to compensate for any wrong decision. In particular, define the compensation cost at group x as: l(x):= τ (x) · 1if d∗(x)! p. 8

Where the paper says it

Definition 2 (PM’s Objective) Assume the PM needs to compensate for any wrong decision. In particular, define the compensation cost at group x as: ℓ(x) := |τ (x)| · 1{d ∗ (x) ̸= ˆdm(x)}.

= ˆdm(x). Given a compensation budget B0, the PM aims at setting ex-ante an estimator m(·) such that: sup Q∈P sup f0,f1∈Fc EQJ×P N [l(Xj) XN] ≤ B0. p. 8
Where the paper says it

In particular, define the compensation cost at group x as: ℓ(x) := |τ (x)| · 1{d ∗ (x) ̸= ˆdm(x)}. (11) Given a compensation budget B0, the PM aims at setting ex-ante an estimator m(·) such that: sup Q∈P sup f0,f1∈Fc EQJ×P N [ℓ(Xj ) | XN ] ≤ B0.

That is, the PM wants to commit to an estimator that provides affordable decisions conditional on the donor points collected. p. 9
Where the paper says it

That is, the PM wants to commit to an estimator that provides affordable decisions conditional on the donor points collected. According to Definition 2, the PM aims at committing on an estimator m(·) that guarantees the average compensation cost over the target units to be below a certain budget B0.

--- Page 9 ---

Remark 1 The loss in (11) is the same as regret (as defined in Manski, 2004) conditional on the collection design XN. Under Assumptions 1 and 2, τ(x) = f1(x) − f0(x). p. 5

Where the paper says it

Their problem differs from the one considered here along several dimensions. The three main differences are that (i) they study a single observational population under unconfoundedness 1For references on these assumptions see e.

Therefore, R(m):= EQJ×P N h fd∗(Xj)(Xj) - fdˆm(Xj) (Xj) XN i = 1 = EQJ×P N h τ (Xj) 1if d∗(Xj)! p. 9
Where the paper says it

Under Assumptions 1 and 2, τ (x) = f1(x) − f0(x). (13) Therefore, R(m) := EQJ×P N h fd ∗(Xj )(Xj ) − fdˆm(Xj ) (Xj ) | XN i (14) = EQJ×P N h τ (Xj ) d ∗ (Xj ) − ˆdm(Xj ) | XN i (15) = EQJ×P N h |τ (Xj )|1{d ∗ (Xj ) ̸= ˆdm(Xj )} | XN i (16) = EQJ×P N [ℓ(Xj ) | XN ].

= ˆdm(Xj) XN i = EQJ×P N [l(Xj) XN]. p. 8
Where the paper says it

In particular, define the compensation cost at group x as: ℓ(x) := |τ (x)| · 1{d ∗ (x) ̸= ˆdm(x)}. (11) Given a compensation budget B0, the PM aims at setting ex-ante an estimator m(·) such that: sup Q∈P sup f0,f1∈Fc EQJ×P N [ℓ(Xj ) | XN ] ≤ B0.

--- Page 10 ---

Solving the problem defined in Definition 2 is challenging because the donor and target population do not coincide, and the assignment mechanism is deterministic. As a result, the assumptions of unconfoundedness and strict overlap, which are common in the policy learning literature do not hold. p. 9

Where the paper says it

As a result, the assumptions of unconfoundedness and strict overlap, which are common in the policy learning literature (see, e. Manski, 2004; Kitagawa and Tetenov, 2018; Athey and Wager, 2021; Mbakop and Tabord-Meehan, 2021; Li and Yan, 2025), do not hold.

Moreover, the guarantee need to hold conditional on the collection design, and not for the average collection. p. 9
Where the paper says it

Manski, 2004; Kitagawa and Tetenov, 2018; Athey and Wager, 2021; Mbakop and Tabord-Meehan, 2021; Li and Yan, 2025), do not hold. Moreover, the guarantee need to hold conditional on the collection design, and not for the average collection.

--- Page 11 ---

I introduce certified decisions as an alternative theoretical device to provide affordability conditions. Definition 3 Fix an error rate alpha ∈ (0, 1) and an estimator m ∈ M. p. 7

Where the paper says it

Define the individual innovation effect and conditional average innovation effect for a unit in the target as: τj := Yj (1) − Yj (0), τ (x) := EQ[τj | Xj = x]. (5) Definition 1 (Innovation Effect Estimator) Let M denote a class of estimators of τj .

A nonnegative scalar gamma(m) is an α-certification boundary for m if defining Cα(m):= x: τ (x) > gamma(m), it satisfies sup Q∈P sup f0,f1∈Fc PQJ×P N d∗(Xj)! p. 6
Where the paper says it

Let d ∈ {0, 1} denote the state. Assumption 1 (Potential Outcomes Function) Assume the potential outcomes functions have the following form: Yi(d) = fd(Xi) + Ui(d) (1) where fd belongs to the admissible functions set Fc defined as: f0, f1 ∈ Fc := ( f : X → Y : sup x∈X ||∇2 f(x)|| ≤ c < ∞ ) (2) for any x ∈ X , and |Yi(1) − Yi(0)| ≤ τ¯.

= ˆdm(Xj) Xj ∈ Cα(m), XN ≤ α. p. 8
Where the paper says it

Let Nx := X J j=1 1{Xj = x}, (7) and any x such that Nx ≥ 1, define the innovation effect estimate τˆm(x) := 1 Nx X J j=1 m(Xj , Yj (0), SN )1{Xj = x}. (8) Given the estimate ˆτm(x), the policy rule is defined as ˆdm(x) := 1{τˆm(x) ≥ 0}.

--- Page 12 ---

Lemma 1 (Affordability of Certified Decisions) Under Assumption 1, and for a choice of m(·) that satisfies (19), sup Q∈P,f0,f1∈Fc EQJ×P N [l(Xj) XN] ≤ maxgamma(m), alpha taubar. As a consequence, m(·) achieves the PM’s objective if: maxgamma(m), alpha taubar ≤ B0. p. 3

Where the paper says it

This result is proved via a standard lifting argument. As a consequence, Delaunay Matching also delivers the smallest worst-case asymptotic compensation cost, thereby providing the tightest affordability guarantee.

--- Page 13 ---

Lemma 2 (Sufficient Condition for Certification) Let Cα(m):= x: τ (x) > gamma(m). Any estimator m ∈ M that satisfies sup Q∈P sup f0,f1∈Fc PQJ×P N (τˆm(Xj) − τ (Xj) ≥ gamma(m) Xj ∈ Cα(m), XN) ≤ α also satisfies sup Q∈P sup f0,f1∈Fc PQJ×P N (d∗(Xj)! p. 8

Where the paper says it

In particular, define the compensation cost at group x as: ℓ(x) := |τ (x)| · 1{d ∗ (x) ̸= ˆdm(x)}. (11) Given a compensation budget B0, the PM aims at setting ex-ante an estimator m(·) such that: sup Q∈P sup f0,f1∈Fc EQJ×P N [ℓ(Xj ) | XN ] ≤ B0.

= ˆdm(Xj) Xj ∈ Cα(m), XN) ≤ α. p. 8
Where the paper says it

Let Nx := X J j=1 1{Xj = x}, (7) and any x such that Nx ≥ 1, define the innovation effect estimate τˆm(x) := 1 Nx X J j=1 m(Xj , Yj (0), SN )1{Xj = x}. (8) Given the estimate ˆτm(x), the policy rule is defined as ˆdm(x) := 1{τˆm(x) ≥ 0}.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Geometric Data Collection Planning: The improved system can design donor data collection plans that empirically outperform random sampling by adding donor observations at locations where current approximations are weakest, as described in Definition 7 (Geometric Plan). This reduces the worst-case compensation cost from a finite budget by achieving significant gains relative to random collection.

  2. Optimal Estimator Selection: The system can select a matching estimator, specifically the Delaunay Matching Estimator (DME), which is proven to require the smallest lower bound on intervention effects to certify decisions at every feasible target point within the class of matching estimators with positive weights. This provides a tighter affordability guarantee than other weighting schemes.

  3. Certification-based Decision Making: The system can employ certified decisions as a theoretical device to provide affordability conditions, linking uniform control of mistake probabilities away from the decision frontier to worst-case loss. This allows for making decisions that are guaranteed to stay within a finite budget B0 under certain conditions derived in Lemma 1.

  4. Adaptive Data Collection Rule: The system can leverage the optimality of Delaunay Matching together with geometric results from Waldron (1998) to motivate a new collection rule, which is called the geometric plan. This rule systematically reduces worst-case compensation cost by adding donor observations at locations where the current approximation is weakest.

  5. Asymptotic Performance Guarantees: For large samples, the system can ensure that certification radii become asymptotically equivalent to their realized geometric components, meaning the guarantee retains its design-conditional interpretation even as data volume increases. This ensures that as more data is collected, the worst-case compensation cost converges to a bound dictated by the geometry of the donor collection design.

Abstract

A researcher observes a population in the status quo and wants to learn from a small, possibly selected sample exposed to an intervention its effect on the rest of the population. How should they combine these sources of information? And who should receive the intervention next to improve on estimation the most? I focus on convex matching estimators and answer both questions by adopting a statistical decision theoretic perspective and drawing new connections with results from computational geometry. First, I show that interpolation over the Delaunay simplex containing a target point minimizes worst-case bias. Second, I provide robust confidence intervals and show that inference-optimal weighting rules lay on a frontier that connects the proposed estimator to Synthetic Controls. Finally, I characterize the minimax refinement problem and show that the Delaunay triangulation of the existing data provides sufficient information to solve it numerically and restricts the space where to look for optimal additions. In a semi-synthetic application in development economics, I show that Delaunay matching achieves a lower worst-case bias compared to commonly used estimators in practice (e.g. around 7 times smaller than synthetic controls) and that the minimax collection plan decreases it by around 90% within 20 additions.

Related papers