Which estimand better fits the matched pair design within units? ATT or ATE (Conditional or Local ATE?)

Question

As an Amazon Associate, I earn from qualifying purchases. For more practice problems like this, see Schaum’s Outline of Statistics, 6th Edition.

1. What the student is asking – restated in plain language

The student has an observational study in which each site contains two plots:

one plot that is “treated’’ (adjacent to land‑use type 1)
one plot that is “control’’ (adjacent to land‑use type 2).

The sites were chosen because they have both land‑use types, not because they were only treated or only untreated.

The student wants to know which causal estimand is appropriate for the analysis:

Option	Rough meaning
ATT (Average Treatment effect on the Treated)	Effect that would be observed only for the units that actually received the treatment (here, the treated plots).
ATE (Average Treatment Effect)	Effect averaged over all units in a well‑defined population (here, all plots that could possibly appear, treated or not).
Conditional (or “local’’) ATE	Effect averaged over the sub‑population that has the same pre‑treatment covariates as the observed sample (e.g., the set of sites that satisfy the geographic and climatic constraints used for sampling).

Because the design pairs a treated and a control plot within the same site, the student wonders whether the estimand is automatically an ATT, a (sample) ATE, or a conditional ATE that only applies to sites with the same geoclimatic characteristics.

2. Step‑by‑step reasoning

Step 1 – Write down the potential‑outcome (Rubin) framework

For each plot (i) we denote

(Y_i(1)) – outcome we would observe if the plot were treated (adjacent to land‑use 1)
(Y_i(0)) – outcome we would observe if the plot were control (adjacent to land‑use 2)

The observed outcome is

[ Y_i = Z_i Y_i(1) + (1-Z_i) Y_i(0), ]

where (Z_i\in{0,1}) indicates treatment (1 = treated).

Because plots are nested inside sites, let (s(i)) denote the site that contains plot (i).
Within each site we have exactly one treated plot ((Z_i=1)) and one control plot ((Z_i=0)).

Step 2 – Define the three candidate estimands

Estimand	Formal definition (finite‑sample version)	Population it refers to
ATT	(\displaystyle \tau_{\text{ATT}} = \frac{1}{N_T}\sum_{i:Z_i=1}\bigl{Y_i(1)-Y_i(0)\bigr})	The treated plots only (the “treated’’ side of each pair).
ATE	(\displaystyle \tau_{\text{ATE}} = \frac{1}{N}\sum_{i=1}^{N}\bigl{Y_i(1)-Y_i(0)\bigr})	All plots that could belong to the target population (treated + control).
Conditional (local) ATE	(\displaystyle \tau_{\text{CATE}}(x) = \mathbb{E}!\big[\,Y(1)-Y(0)\mid X=x\,\big]) where (X) are the pre‑treatment covariates used for sampling (geoclimatic, landscape, etc.). In a finite sample it is the average treatment effect within the sub‑sample that shares the observed covariate pattern.	The sub‑population that satisfies the same site‑selection criteria as the observed sample.

Step 3 – Examine how the sample was generated

Selection rule: Choose sites that contain both land‑use types.
Within each selected site, we observe both a treated and a control plot.
The selection rule does not depend on whether a site is “treated’’ or “control’’ (both exist in every selected site).
The rule does depend on geoclimatic and landscape characteristics (e.g., elevation, precipitation). Those characteristics are the same for the treated and control plot inside a site, and they are the only source of variation across sites.

Thus the sample is conditional on a set of covariates (X). The sample does not preferentially pick treated plots; it picks pairs of plots that satisfy the covariate constraints.

Step 4 – Determine which estimand the data naturally identify

Because every selected site contributes one treated and one control plot, the pairwise difference

[ D_s = Y_{s,T} - Y_{s,C} ]

is an unbiased estimator of the site‑specific treatment effect

[ \tau_s = Y_{s,T}(1) - Y_{s,C}(0), ]

provided the stable unit treatment value assumption (SUTVA) holds within a site (no interference between the two plots) and the conditional ignorability assumption holds given the covariates that are constant within a site (i.e., after conditioning on the site‑level covariates, treatment assignment is as‑if random).

Taking the average of the pairwise differences over all sampled sites gives

[ \widehat{\tau}{\text{pair}} = \frac{1}{S}\sum{s=1}^{S} D_s . ]

What does (\widehat{\tau}_{\text{pair}}) estimate?

It is not an ATT, because ATT would weight each treated plot equally and would ignore the control‑only plots. In our data every treated plot has a matched control, so the ATT is exactly the same numerical average if we average only over treated plots and use the same control outcomes that belong to the same sites. However, the ATT definition does not condition on the covariate restrictions that generated the sample; it implicitly averages over all treated plots in the broader population (including sites that do not satisfy the selection rule).
It is a conditional (local) ATE: the average effect over the sub‑population of sites that meet the geoclimatic/landscape constraints. Formally,

[ \tau_{\text{CATE}} = \mathbb{E}!\big[\,Y(1)-Y(0)\mid X\in\mathcal{X}_\text{sample}\,\big], ]

where (\mathcal{X}_\text{sample}) denotes the region of covariate space used for sampling.

It can also be described as the sample ATE (the finite‑sample average treatment effect) within that restricted sample. In finite‑sample notation:

[ \tau_{\text{sample}} = \frac{1}{2S}\sum_{s=1}^{S}\bigl{Y_{s,T}(1)-Y_{s,T}(0) + Y_{s,C}(1)-Y_{s,C}(0)\bigr} = \frac{1}{S}\sum_{s=1}^{S} D_s, ]

because the control plot’s potential outcome under treatment, (Y_{s,C}(1)), is never observed but cancels out under the pairing design.

Hence the identifiable estimand is the average treatment effect conditional on the covariates used for site selection.

Step 5 – When can we reinterpret it as an ATT or a broader ATE?

If the investigator is willing to assume that the conditional effect does not vary with the covariates (i.e., homogeneous treatment effect), then

[ \tau_{\text{CATE}} = \tau_{\text{ATE}} = \tau_{\text{ATT}} . ]

In that special case the distinction is moot, but the assumption is strong and must be justified.

If external data are available that describe the distribution of the covariates in the full target population (all sites, not only those with both land‑use types), one could re‑weight the conditional ATE to obtain an estimate of the population ATE (or ATT). The re‑weighting formula is the standard IPW or g‑formula:

[ \tau_{\text{ATE}} = \frac{\displaystyle\sum_{s} w_s D_s}{\displaystyle\sum_{s} w_s}, \qquad w_s = \frac{P_{\text{pop}}(X_s)}{P_{\text{sample}}(X_s)} . ]

Absent such re‑weighting, the most honest description is conditional ATE.

Step 6 – Summarise the answer to the original question

Question	Answer
Is the estimand an ATT?	No, because the sample was not drawn by selecting treated sites only; it was drawn by selecting sites that satisfy certain covariate constraints and contain both a treated and a control plot.
Is it a (sample) ATE?	Yes, but only within the restricted covariate space used for sampling. This is often called the conditional or local ATE.
Does it generalise to other sites with the same geoclimatic/landscape characteristics?	Yes, under the assumption that the treatment effect is exchangeable for all sites sharing those covariates. The estimand is the average effect among sites that fall in the same covariate region as the observed sample.
Can it be interpreted as the usual population ATE or ATT?	Only with additional assumptions (effect homogeneity) or with re‑weighting using external information about the full population’s covariate distribution.

3. Final answer

The paired‑plot design yields an estimator that is the average treatment effect conditional on the site‑level covariates used to select the sample (the “conditional ATE’’ or “local ATE’’). It is not an ATT because the sampling scheme does not preferentially select treated sites; it is also not the unrestricted population ATE unless one imposes a homogeneity assumption or applies external re‑weighting. Consequently, the causal claim that can be made is:

“Among sites that satisfy the same geoclimatic and landscape constraints as those in the study, the average effect of changing the adjacent land‑use from type 2 to type 1 is (\widehat{\tau}_{\text{pair}}).”

4. Common mistakes for this type of problem

Mistake	Why it is wrong	How to avoid it
Calling the estimand ATT because the data contain treated plots.	ATT refers to all treated units in the target population, not just the treated units that happen to appear in a sample selected on other criteria.	Verify the sampling frame: if units are selected independently of treatment status, ATT is not the natural estimand.
Equating “sample ATE’’ with the population ATE without justification.	The sample may be a non‑random slice of the population (e.g., restricted to a narrow climatic zone).	Explicitly state the conditioning set (\mathcal{X}_\text{sample}) and discuss whether you can assume transportability.
Ignoring the matching/pairing structure and averaging treated and control outcomes separately.	The paired design creates a difference estimator that directly targets the average effect within each pair. Averaging separately can re‑introduce bias if the pairwise balance is not respected.	Use the within‑pair difference (or a matched‑pair regression) as the primary estimator.
Assuming SUTVA automatically holds because plots are in the same site.	There can be spill‑over (e.g., water flow, species movement) from the treated to the control plot.	Examine the scientific context; if interference is plausible, consider hierarchical or spatial models that relax SUTVA.
Treating covariates that are constant within a site as “plot‑level’’ variables.	Such covariates cannot explain within‑site variation and therefore do not help with ignorability at the plot level.	Recognise they are site‑level covariates and include them in the conditioning set for the conditional ATE.
Believing that matching guarantees unbiasedness without checking the ignorability assumption.	Matching balances observed covariates but cannot address unobserved confounding.	State the conditional ignorability assumption explicitly and, if possible, conduct sensitivity analysis.

By keeping these points in mind, you can correctly specify the estimand for a matched‑pair design that is sampled conditional on site‑level characteristics.

Original question: Which estimand better fits the matched pair design within units? ATT or ATE (Conditional or Local ATE?) on Cross Validated (Stats Stack Exchange), licensed CC BY-SA.

1. What the student is asking – restated in plain language

2. Step‑by‑step reasoning