Conditional Average Treatment Effects (CATE) | Carlos Mendez

Causal Machine Learning and the Resource Curse with Stata 19

Wed, 06 May 2026 00:00:00 +0000

1. Overview

Imagine discovering that the very thing that should make a country rich — abundant natural resources — actually makes it poorer. This is the resource curse hypothesis, first documented by Sachs and Warner (1995): countries rich in oil, minerals, or other extractive resources often experience slower growth, weaker institutions, and more conflict than resource-poor nations.

But the story is more nuanced than “resources are bad.” Mehlum, Moene, and Torvik (2006) argued that institutional quality determines whether resource wealth becomes a blessing or a curse. Countries with strong rule of law and quality governance channel resource revenues productively, while weak institutions allow rent-seeking and conflict.

This tutorial is inspired by Hodler, Lechner & Raschky (2023), who brought causal machine learning to this debate. Using a Modified Causal Forest on sub-national mining districts across Sub-Saharan Africa, they uncovered three key findings:

Mining increases development and conflict — Districts that begin mining experience higher nighttime lights (a proxy for economic activity) and more conflict events.
Price effects are non-linear — The effect of mineral prices on outcomes is small at moderate prices but jumps sharply at high prices.
Institutions moderate mining but NOT prices — Institutional quality amplifies the development benefits of mining (upward-sloping GATEs), but does not moderate the effect of global price shocks (flat GATEs).

This tutorial uses Stata 19’s cate command to replicate all three findings on a simulated dataset with known ground-truth causal effects (3,000 observations = 300 districts $\times$ 10 years). Because the data-generating process is known, we can directly compare our estimates against the true parameter values. The cate command provides native access to generalized random forests, doubly robust estimation, and formal hypothesis tests — all without external packages.

Prerequisite. This post requires Stata 19 or later. The cate command does not exist in Stata 18. The companion do-file aborts on startup if it detects an older Stata.

Runtime. The full analysis takes approximately 20–30 minutes on a modern machine. Each cate estimation takes 60–90 seconds with 5-fold cross-fitting.

For a deeper introduction to the CATE framework and the cate command on a binary-treatment dataset, see the companion tutorial Conditional Average Treatment Effects (CATE) with Stata 19.

1.1 Learning objectives

By the end of this tutorial you should be able to:

Understand the resource curse hypothesis and why treatment effects may vary with institutional quality.
Apply Stata 19’s cate command to a multi-valued treatment via binary pairwise comparisons.
Distinguish PO and AIPW estimators and when each is preferred.
Estimate ATEs, GATEs, and IATEs for multiple treatment contrasts.
Interpret GATE patterns to identify institutional moderation of treatment effects.
Diagnose treatment-effect heterogeneity with formal hypothesis tests (estat heterogeneity, estat gatetest).
Visualize individualized treatment effects using categraph postestimation tools.
Connect statistical results to substantive findings from published research.

1.2 Analytical roadmap

The diagram below shows the five stages of this tutorial, from data exploration through advanced diagnostics.

flowchart LR
A["<b>Data &<br/>Descriptives</b><br/><i>Sections 3--4</i>"]:::data
B["<b>Naive vs<br/>Ground Truth</b><br/><i>Section 5</i>"]:::naive
C["<b>ATE<br/>Estimation</b><br/><i>Sections 6--7</i>"]:::ate
D["<b>GATE<br/>Heterogeneity</b><br/><i>Section 8</i>"]:::gate
E["<b>Advanced<br/>Diagnostics</b><br/><i>Section 9</i>"]:::diag
A --> B --> C --> D --> E
classDef data fill:#6a9bcc,stroke:#141413,color:#fff
classDef naive fill:#d97757,stroke:#141413,color:#fff
classDef ate fill:#00d4c8,stroke:#141413,color:#141413
classDef gate fill:#d97757,stroke:#141413,color:#fff
classDef diag fill:#141413,stroke:#d97757,color:#fff

1.3 Key concepts at a glance

The post leans on a small vocabulary repeatedly. The rest of the tutorial assumes you can move between these terms quickly. Each concept below has three parts. The definition is always visible. The example and analogy sit behind clickable cards: open them when you need them, leave them collapsed for a quick scan. If a later section mentions “PO vs AIPW” or “honest splitting” and the term feels slippery, this is the section to re-read.

1. Potential outcomes $Y_i(t)$. The outcome unit $i$ would take under treatment value $t$. Each unit has one potential outcome per treatment level. We observe only one of them: the one matching the treatment actually received. The rest are counterfactual. They live in worlds we never see.

Example

Take district 47 in 2008. Four potential NTL outcomes exist for it: $Y_{47,2008}(0)$, $Y_{47,2008}(1)$, $Y_{47,2008}(2)$, and $Y_{47,2008}(3)$. They correspond to no mining, low prices, medium prices, and high prices. Only one is in the dataset. It is the one matching whatever treatment value that district-year actually had. The other three are forever invisible.

Analogy

Every life decision is a fork in the road. You took one fork. The parallel-universe versions of yourself took the other forks. Their lives are real conceptual objects. You just cannot directly observe them. Causal inference reconstructs those parallel universes. It does so by looking at people who did take the other forks.

2. CATE — Conditional Average Treatment Effect, $\tau(\mathbf{x})$. The average treatment effect for units with covariate profile $\mathbf{x}$. The CATE is a function of $\mathbf{x}$, not a single number. Where the CATE bends with $\mathbf{x}$, the treatment helps some units more than others. Stata’s cate command estimates exactly this function.

Example

Take a well-governed district profile in our data: exec_constraints = 6, quality_of_govt = 0.7, and so on. For that profile the CATE is $\tau(\mathbf{x}) \approx 0.26$. Mining lifts log-NTL by about 0.26 for that profile. Now move to the weakest-institutions case: exec_constraints = 1. The same function gives only $\tau(\mathbf{x}) \approx 0.18$. The CATE is what makes this comparison possible.

Analogy

A drug’s “average effect” might be a 5-point reduction in blood pressure. But a doctor cares about a specific patient. Maybe a 65-year-old male with diabetes. The CATE is that personalized effect. It takes a patient profile in. It returns the expected effect for someone like them.

3. GATE — Group Average Treatment Effect. The CATE averaged over a pre-specified subgroup. The subgroup is defined by some variable. GATEs test targeted moderation hypotheses. A typical question: “does institutional quality moderate the effect of mining?”

Example

Sort districts by exec_constraints (1–6). Average the per-observation CATEs inside each level. At level 1 we get $\widehat{\mathrm{GATE}} \approx 0.18$. The number climbs to $\approx 0.26$ at level 6. That climb is the moderation pattern §8 visualizes. It is exactly what estat gateplot reports after the cate command.

Analogy

A nationwide marketing campaign might lift sales by 5% on average. Before scaling it up, the company asks a simple question: did it work better in cities than in rural towns? The GATE answers exactly that. It reports the campaign’s effect inside each store type. It surfaces heterogeneity that the headline ATE hides.

4. ATE — Average Treatment Effect, $E[\tau(\mathbf{X})]$. The CATE averaged over the entire sample. The headline policy number. It answers a single question: if we turned the treatment on for everyone, what average effect would we see?

Example

Take our 3,000 district-years. The PO ATE for the 1-vs-0 mining contrast is 0.194 (SE = 0.010). AIPW gives a more conservative 0.149 (SE = 0.011). Both are reported in §7. They are two estimates of the same population-level number.

Analogy

“This drug lowers cholesterol by 12 points on average.” That is an ATE statement. A single number, suitable for a press release. It says nothing about whether the drug works better in some patients than others. That question belongs to GATEs and CATEs.

5. Nuisance functions $g_0, m_0$. Two conditional means. $g_0(\mathbf{x}, \mathbf{w}) = E[Y \mid \mathbf{X}, \mathbf{W}]$ predicts the outcome from covariates. $m_0(\mathbf{x}, \mathbf{w}) = E[T \mid \mathbf{X}, \mathbf{W}]$ predicts the treatment from covariates. We call them nuisance because we do not care about their values directly. We estimate them only to strip out the part of $Y$ and $T$ that is predictable from $(\mathbf{X}, \mathbf{W})$. What remains is the variation that identifies the causal effect.

Example

In this post Stata’s cate fits both $g_0$ and $m_0$ as random forests behind the scenes. We never see them. We never tune them directly. They are intermediate machinery the command consumes and discards on its way to the CATE.

Analogy

Two surveyors map two different layers of the same terrain. One maps elevation. The other maps soil type. Neither map is the goal. The goal is to subtract them from a third map and see what is left. That residue is what we actually care about.

6. Cross-fitting and honest splitting. Cross-fitting splits the sample into $K$ folds. Nuisance models are fit on $K-1$ folds and applied to the held-out fold. The roles rotate. No observation is ever scored by a model that saw it during training. Honest splitting goes one step further inside each tree. It uses one subsample to choose where to split. It uses a separate subsample to estimate the leaf values. Both tricks remove the over-fitting bias that would otherwise contaminate the CATE.

Example

Stata’s cate does this internally. We pass xfolds(5) and the rest is automatic. We never call separate train/test commands. The 5 folds rotate behind the scenes; the user sees only the final estimates.

Analogy

Two-pass exam grading. One TA writes the rubric without seeing your paper. A different TA applies the rubric without writing it. The separation is what makes the grade defensible. Mixing the two roles is exactly the over-fitting bias these tricks remove.

7. PO vs AIPW estimators. Two ways to map nuisance estimates to a CATE. PO (Partialing Out) residualizes both $Y$ and $T$ against the covariates, then regresses one residual on the other. Simple, transparent, sensitive to extreme propensity scores. AIPW (Augmented Inverse-Probability Weighting) reweights observations by inverse propensity and adds a regression correction. More complex, but doubly robust: it stays consistent if either $g_0$ or $m_0$ is right.

Example

This post fits both. They disagree by about 0.045 on the 1-vs-0 contrast (PO 0.194, AIPW 0.149). That gap is the model-disagreement diagnostic. When PO and AIPW disagree, the overlap is suspect or one of the nuisance models is mis-specified.

Analogy

Two judges hear the same case. They follow slightly different reasoning paths. When their verdicts agree, you trust the case. When they disagree, you re-read the evidence. The disagreement is the signal, not noise.

8. Heterogeneity test. A formal test that $\tau(\mathbf{x})$ varies with $\mathbf{x}$. The null hypothesis is constant treatment effects: every unit gets the same effect. Rejection licenses CATE and GATE interpretation. Failing to reject does not mean effects are constant. It means the test could not detect variation at this sample size.

Example

After cate, run estat heterogeneity in §9. It returns a $\chi^2$ statistic and a $p$-value. A small $p$-value is the green light to inspect GATEs and CATEs. A large $p$-value is a caution: the heterogeneity story may not be in the data.

Analogy

A metal detector for hidden moderation. It does not tell you where in the field the metal is buried. It only tells you whether to keep digging.

2. The CATE framework

2.1 From ATE to CATE

The Average Treatment Effect (ATE) summarizes causal effects as a single number for the entire population. But when effects are heterogeneous — varying across subgroups — the ATE can mask important patterns. The Conditional Average Treatment Effect (CATE) captures this heterogeneity:

$$\tau(\mathbf{x}) = E\{y_i(1) - y_i(0) \mid \mathbf{x}_i = \mathbf{x}\}$$

where $y_i(1)$ and $y_i(0)$ are potential outcomes under treatment and control, and $\mathbf{x}$ is a vector of characteristics that may moderate the treatment effect. If $\tau(\mathbf{x})$ is constant across all $\mathbf{x}$, we are back at the ATE. Whenever it varies, the ATE is an average of these subgroup effects weighted by how common each $\mathbf{x}$ is in the data.

2.2 The partial linear model

Stata 19’s cate estimates CATEs within a partial linear framework:

$$y = d \cdot \tau(\mathbf{x}) + g(\mathbf{x}, \mathbf{w}) + \epsilon, \qquad d = f(\mathbf{x}, \mathbf{w}) + u$$

where $\tau(\mathbf{x})$ is the heterogeneous treatment effect function, $g(\cdot)$ and $f(\cdot)$ are flexible nuisance functions estimated by machine learning, $\mathbf{x}$ are CATE covariates (potential moderators), and $\mathbf{w}$ are additional controls.

Think of the nuisance functions as background noise that must be cleaned away before the treatment effect signal becomes visible. The cate command uses cross-fitting to prevent the nuisance models from overfitting: data are split into $K$ folds, and each fold’s nuisance predictions are made using models trained on the other $K-1$ folds.

2.3 Two estimators

Stata 19 provides two estimators for the CATE:

Partialing-Out (PO). Think of PO like cleaning two messy signals before comparing them. It residualizes both the outcome and treatment against $\mathbf{x}$ and $\mathbf{w}$, then estimates $\tau(\mathbf{x})$ from the residuals using a generalized random forest (Nie & Wager, 2021). PO is robust when propensity scores get close to 0 or 1.

Augmented Inverse-Probability Weighting (AIPW). AIPW is like having a backup GPS — if one route fails, the other still gets you there. It constructs doubly robust scores that combine outcome modeling and propensity score weighting. Even if one model is misspecified, the estimator remains consistent (Knaus, 2022; Kennedy, 2023).

2.4 Three levels of treatment effects

flowchart LR
A["<b>Panel Data</b><br/>3,000 obs<br/>300 districts x 10 years"]:::data
A --> B["<b>cate po / aipw</b><br/>Binary pairwise<br/>comparisons"]:::main
B --> C["<b>IATEs</b><br/>Per-observation<br/>effects tau(x_i)"]:::iate
B --> D["<b>GATEs</b><br/>Group averages<br/>by institutions"]:::gate
B --> E["<b>ATE</b><br/>Overall population<br/>average"]:::ate
C --> F["categraph histogram<br/>categraph iateplot"]:::post
D --> G["categraph gateplot<br/>estat gatetest"]:::post
E --> H["estat heterogeneity<br/>estat ate"]:::post
classDef data fill:#6a9bcc,stroke:#141413,color:#fff
classDef main fill:#141413,stroke:#141413,color:#fff
classDef iate fill:#00d4c8,stroke:#141413,color:#141413
classDef gate fill:#d97757,stroke:#141413,color:#fff
classDef ate fill:#6a9bcc,stroke:#141413,color:#fff
classDef post fill:#f5f5f5,stroke:#141413,color:#141413

IATE (Individualized Average Treatment Effects): One effect per observation, $\tau(\mathbf{x}_i)$
GATE (Group Average Treatment Effects): Average effect within prespecified groups, $\tau(g) = E\{\tau(\mathbf{x}) \mid G = g\}$
ATE: Overall population average, $\text{ATE} = E\{\tau(\mathbf{x})\}$

3. Data preparation

We use a simulated dataset of 3,000 observations (300 districts $\times$ 10 years) across 8 fictional countries. The data mirror the structure of Hodler et al. (2023) but with known ground-truth causal effects, enabling direct validation of our estimates.

* Import the simulated resource curse dataset
* GitHub: import delimited using "https://github.com/quarcs-lab/data-open/raw/master/stata19/sim_resource_curse.csv", clear
import delimited using "sim_resource_curse.csv", clear
* Label variables
label variable district_id "District ID (1-300)"
label variable country_id "Country ID (1-8)"
label variable year "Year (2003-2012)"
label variable treatment "Treatment group (0=none, 1=low, 2=med, 3=high)"
label variable mining "Mining district (binary)"
label variable price_index "Mineral price index"
label variable exec_constraints "Constraints on Executive (1-6)"
label variable quality_of_govt "Quality of Government (0.22-0.70)"
label variable gdp_pc "GDP per capita"
label variable elevation "Elevation (meters)"
label variable temperature "Mean temperature (Celsius)"
label variable ruggedness "Terrain ruggedness"
label variable distance_capital "Distance to capital (meters)"
label variable agri_suitability "Agricultural suitability (0-1)"
label variable population "Population"
label variable ethnic_frac "Ethnic fractionalization (0-1)"
label variable ntl_log "Log nighttime lights"
label variable conflict "Conflict event (binary)"
* Create integer version of exec_constraints for group()
gen int exec_con = round(exec_constraints)
label variable exec_con "Executive Constraints (integer 1-6)"
* Save as .dta
save "sim_resource_curse.dta", replace
* Report dataset dimensions
describe, short

Contains data from sim_resource_curse.dta
Observations: 3,000
Variables: 19 6 May 2026
Sorted by:

The dataset contains 3,000 observations organized as a balanced panel: 300 districts observed over 10 years (2003–2012) in 8 fictional countries. The key variables are:

Variable	Description	Type
`treatment`	Treatment group (0=none, 1=low, 2=med, 3=high price)	Categorical
`ntl_log`	Log nighttime lights (development proxy)	Continuous outcome
`conflict`	Conflict event indicator	Binary outcome
`exec_constraints`	Constraints on executive (1–6 scale)	Institutional moderator
`quality_of_govt`	Quality of government (0.22–0.70)	Institutional moderator
`gdp_pc`, `elevation`, `temperature`, …	Economic and geographic covariates	Controls

4. Descriptive statistics

* Summary statistics for key variables
tabstat ntl_log conflict exec_constraints quality_of_govt gdp_pc ///
elevation temperature ruggedness distance_capital ///
agri_suitability population ethnic_frac, ///
statistics(mean sd min max) columns(statistics) format(%9.3f)

 Variable | Mean SD Min Max
-------------+----------------------------------------
ntl_log | -1.096 0.435 -2.503 0.265
conflict | 0.123 0.328 0.000 1.000
exec_const~s | 3.680 1.489 1.000 6.000
quality_of~t | 0.440 0.152 0.220 0.700
gdp_pc | 2198.000 1469.937 500.000 5000.000
elevation | 499.083 302.031 0.000 1357.232
temperature | 23.913 3.920 13.993 35.000
ruggedness | 24.423 17.803 0.000 76.953
distance_c~l | 2.68e+05 1.44e+05 10813.747 4.97e+05
agri_suita~y | 0.395 0.197 0.000 0.983
population | 82028.426 85186.961 4134.682 5.97e+05
ethnic_frac | 0.550 0.202 0.201 0.899
------------------------------------------------------

* Treatment distribution
tab treatment, missing
* Mining share
count if treatment > 0
* Outcomes by treatment group
table treatment, statistic(mean ntl_log) statistic(mean conflict) ///
statistic(count ntl_log) nformat(%9.3f)

 Treatment |
group |
(0=none, |
1=low, |
2=med, |
3=high) | Freq. Percent Cum.
------------+-----------------------------------
0 | 2,550 85.00 85.00
1 | 150 5.00 90.00
2 | 150 5.00 95.00
3 | 150 5.00 100.00
------------+-----------------------------------
Total | 3,000 100.00
Mining share: 15.0%
--------------------------------------------------------------
| Mean
| ntl_log conflict
-----------------------------------------------+--------------------
Treatment group (0=none, 1=low, 2=med, 3=high) |
0 | -1.137 0.107
1 | -1.028 0.180
2 | -0.930 0.180
3 | -0.615 0.280
Total | -1.096 0.123
--------------------------------------------------------------

Treatment imbalance. The treatment distribution is highly imbalanced: approximately 85% of observations are in the control group (no mining), while each treated group contains only about 5% of observations. This mirrors real-world mining data where few districts have active mines. Stata’s cate handles this via honest random forests with appropriate sample-splitting.

The descriptive statistics reveal important patterns. Mean ntl_log varies across treatment groups, but these raw differences mix the causal effect with confounding — mining districts differ systematically from non-mining districts in geography, institutions, and economic conditions. The next section demonstrates this directly.

5. Naive comparison vs ground truth

Before applying any causal method, we compute raw mean differences and compare them to the known ground-truth ATEs from the data-generating process.

* Naive difference-in-means (biased by confounders)
display as text _newline "=== Naive Difference-in-Means (biased) ==="
display as text "Comparison" _col(20) "NTL diff" _col(35) "Ground Truth"
display as text "{hline 50}"
* 1-0: mining vs no mining
quietly summarize ntl_log if treatment == 1
local m1 = r(mean)
quietly summarize ntl_log if treatment == 0
local m0 = r(mean)
display as result "1 vs 0" _col(20) %7.4f (`m1' - `m0') _col(35) "0.25"
* 3-1: high vs low prices
quietly summarize ntl_log if treatment == 3
local m3 = r(mean)
display as result "3 vs 1" _col(20) %7.4f (`m3' - `m1') _col(35) "0.30"
* 2-1: medium vs low prices
quietly summarize ntl_log if treatment == 2
local m2 = r(mean)
display as result "2 vs 1" _col(20) %7.4f (`m2' - `m1') _col(35) "0.05"

=== Naive Difference-in-Means (biased) ===
Comparison NTL diff Ground Truth
--------------------------------------------------
1 vs 0 0.1092 0.25
2 vs 0 0.2077 0.30
3 vs 0 0.5227 0.55
2 vs 1 0.0985 0.05
3 vs 1 0.4135 0.30
3 vs 2 0.3150 0.25
--------------------------------------------------

The ground-truth ATEs for all six pairwise comparisons are:

Contrast	Ground Truth	Interpretation
1-0	0.25	Mining effect at mean institutions
2-0	0.30	Mining + medium price premium
3-0	0.55	Mining + high price premium
2-1	0.05	Medium price premium (small)
3-1	0.30	High price premium (large)
3-2	0.25	High vs medium step

The naive comparisons are biased because mining districts differ systematically from non-mining districts in geography, institutions, and economic conditions. Some confounders push the raw difference above the truth, others below. This motivates the use of causal machine learning methods that adjust for these confounders.

6. Estimation strategy

6.1 Binary pairwise comparisons

Stata 19’s cate command requires a binary treatment variable. Since our treatment has 4 levels (0, 1, 2, 3), we run separate estimations for each pairwise comparison, subsetting the data to the two relevant groups each time. This yields 6 binary comparisons that map directly to the three key findings:

graph TD
T["<b>4-Level Treatment</b><br/>0: No mining<br/>1: Low price<br/>2: Medium price<br/>3: High price"]:::data
subgraph F1["Finding 1: Mining Effect"]
C10["1 vs 0"]:::f1
C20["2 vs 0"]:::f1
C30["3 vs 0"]:::f1
end
subgraph F2["Finding 2: Price Non-linearity"]
C21["2 vs 1<br/><i>small ~ 0.05</i>"]:::f2
C31["3 vs 1<br/><i>large ~ 0.30</i>"]:::f2
C32["3 vs 2"]:::f2
end
T --> F1
T --> F2
classDef data fill:#6a9bcc,stroke:#141413,color:#fff
classDef f1 fill:#00d4c8,stroke:#141413,color:#141413
classDef f2 fill:#d97757,stroke:#141413,color:#fff

Contrast	Comparison	Finding	Ground Truth
1-0	Mining (any price) vs No mining	Finding 1	0.25
2-0	Mining (medium price) vs No mining	Finding 1	0.30
3-0	Mining (high price) vs No mining	Finding 1	0.55
2-1	Medium vs Low prices (within mining)	Finding 2 (small)	0.05
3-1	High vs Low prices (within mining)	Finding 2 (large)	0.30
3-2	High vs Medium prices (within mining)	Finding 2	0.25

6.2 Variable specification

We separate variables into two groups following the cate framework:

* CATE variables (x): potential drivers of treatment-effect heterogeneity
global catevars exec_constraints quality_of_govt gdp_pc ///
elevation temperature ruggedness distance_capital ///
agri_suitability population ethnic_frac
* Controls (w): nuisance variables for background adjustment only
global controls i.country_id i.year

The catevarlist ($\mathbf{x}$) contains the 10 covariates that may drive heterogeneity — institutional, economic, and geographic variables. The controls ($\mathbf{w}$) contain country and year fixed effects to absorb panel-level confounding without overcomplicating the CATE function.

We use xfolds(5) rather than the default 10 to ensure adequate sample sizes per fold, especially for the within-mining comparisons. With rseed(12345), all results are reproducible.

Small sample warning. The mining-vs-no-mining comparisons (1-0, 2-0, 3-0) use approximately 2,700 observations — adequate for the causal forest. However, the within-mining price comparisons (2-1, 3-1, 3-2) use only about 300 observations (two treated groups of ~150 each). With xfolds(5), each fold has only ~60 observations. Expect wider confidence intervals for these comparisons.

7. Average treatment effects

We estimate ATEs for all 6 NTL contrasts and key conflict contrasts. For the two most important comparisons (NTL 1-0 and NTL 3-1), we show both PO and AIPW estimators. Remaining comparisons use AIPW only.

Runtime. Each cate estimation takes approximately 60–90 seconds with 5-fold cross-fitting. The full section runs in approximately 15–20 minutes.

7.1 NTL: Mining effect (1-0) — PO vs AIPW

This is the most important contrast: does mining increase nighttime lights? The ground truth is 0.25.

* --- PO estimator ---
preserve
keep if treatment == 1 | treatment == 0
gen byte treat_1v0 = (treatment == 1)
cate po (ntl_log $catevars) (treat_1v0), ///
controls($controls) ///
rseed(12345) xfolds(5) ///
omethod(rforest) tmethod(rforest)
estimates store po_ntl_1v0
restore

Conditional average treatment effects Number of observations = 2,700
Estimator: Partialing out Number of folds in cross-fit = 5
Outcome model: Random forest Number of outcome controls = 28
Treatment model: Random forest Number of treatment controls = 28
CATE model: Random forest Number of CATE variables = 10
------------------------------------------------------------------------------
| Robust
ntl_log | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
treat_1v0 |
(Mining ..) |
vs |
No mining) | .1936814 .0097428 19.88 0.000 .1745858 .212777
-------------+----------------------------------------------------------------
POmean |
treat_1v0 |
No mining | -1.142413 .0079236 -144.18 0.000 -1.157943 -1.126883
------------------------------------------------------------------------------

* --- AIPW estimator ---
preserve
keep if treatment == 1 | treatment == 0
gen byte treat_1v0 = (treatment == 1)
cate aipw (ntl_log $catevars) (treat_1v0), ///
controls($controls) ///
rseed(12345) xfolds(5) ///
omethod(rforest) tmethod(rforest)
estimates store aipw_ntl_1v0
restore

Conditional average treatment effects Number of observations = 2,700
Estimator: Augmented IPW Number of folds in cross-fit = 5
Outcome model: Random forest Number of outcome controls = 28
Treatment model: Random forest Number of treatment controls = 28
CATE model: Random forest Number of CATE variables = 10
------------------------------------------------------------------------------
| Robust
ntl_log | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
treat_1v0 |
(Mining ..) |
vs |
No mining) | .1489842 .0105686 14.10 0.000 .1282701 .1696983
-------------+----------------------------------------------------------------
POmean |
treat_1v0 |
No mining | -1.142416 .0079187 -144.27 0.000 -1.157936 -1.126896
------------------------------------------------------------------------------

The PO ATE is 0.194 (SE = 0.010) and the AIPW ATE is 0.149 (SE = 0.011) — both positive and significant, confirming that mining increases nighttime lights. The estimates differ somewhat from each other and from the ground truth (0.25), which is expected with 5-fold cross-fitting on a moderately sized sample. The PO estimate is closer to the truth here, while AIPW is more conservative. Both confirm the directional finding.

7.2 NTL: High vs low prices (3-1) — PO vs AIPW

The price effect comparison tests Finding 2. The ground truth is 0.30 — a large jump from low to high prices. Note that this comparison uses only mining districts (~300 observations), so estimates will be noisier.

* --- PO estimator ---
preserve
keep if treatment == 3 | treatment == 1
gen byte treat_3v1 = (treatment == 3)
display as text "N = " _N " observations (mining districts only)"
cate po (ntl_log $catevars) (treat_3v1), ///
controls($controls) ///
rseed(12345) xfolds(5) ///
omethod(rforest) tmethod(rforest)
estimates store po_ntl_3v1
restore

N = 300 observations (mining districts only)
Conditional average treatment effects Number of observations = 300
Estimator: Partialing out Number of folds in cross-fit = 5
Outcome model: Random forest Number of outcome controls = 28
Treatment model: Random forest Number of treatment controls = 28
CATE model: Random forest Number of CATE variables = 10
------------------------------------------------------------------------------
| Robust
ntl_log | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
treat_3v1 |
(High price |
vs |
Low price) | .5945629 .0313138 18.99 0.000 .5331891 .6559368
-------------+----------------------------------------------------------------
POmean |
treat_3v1 |
Low price | -1.12839 .0280085 -40.29 0.000 -1.183285 -1.073494
------------------------------------------------------------------------------

* --- AIPW estimator ---
preserve
keep if treatment == 3 | treatment == 1
gen byte treat_3v1 = (treatment == 3)
cate aipw (ntl_log $catevars) (treat_3v1), ///
controls($controls) ///
rseed(12345) xfolds(5) ///
omethod(rforest) tmethod(rforest)
estimates store aipw_ntl_3v1
restore

Conditional average treatment effects Number of observations = 300
Estimator: Augmented IPW Number of folds in cross-fit = 5
Outcome model: Random forest Number of outcome controls = 28
Treatment model: Random forest Number of treatment controls = 28
CATE model: Random forest Number of CATE variables = 10
------------------------------------------------------------------------------
| Robust
ntl_log | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
treat_3v1 |
(High price |
vs |
Low price) | .4052631 .0254935 15.90 0.000 .3552968 .4552293
-------------+----------------------------------------------------------------
POmean |
treat_3v1 |
Low price | -1.029871 .0240718 -42.78 0.000 -1.077051 -.9826917
------------------------------------------------------------------------------

The PO ATE is 0.595 (SE = 0.031) and the AIPW ATE is 0.405 (SE = 0.025) — both large and highly significant, confirming that the price premium from low to high is substantial (ground truth = 0.30). With only 300 observations and 5-fold cross-fitting, estimates are noisier than the 1-0 contrast, and both overshoot the ground truth, but the directional finding is robust.

7.3 NTL: Remaining comparisons (AIPW only)

For the remaining four NTL contrasts we use AIPW with default lasso methods (faster than random forests on smaller subsamples).

* --- NTL: 2 vs 0 (medium mining vs no mining) ---
preserve
keep if treatment == 2 | treatment == 0
gen byte treat_2v0 = (treatment == 2)
cate aipw (ntl_log $catevars) (treat_2v0), ///
controls($controls) rseed(12345) xfolds(5)
estimates store aipw_ntl_2v0
restore

Conditional average treatment effects Number of observations = 2,700
Estimator: Augmented IPW Number of folds in cross-fit = 5
Outcome model: Linear lasso Number of outcome controls = 28
Treatment model: Logit lasso Number of treatment controls = 28
CATE model: Random forest Number of CATE variables = 10
------------------------------------------------------------------------------
| Robust
ntl_log | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
treat_2v0 |
(1 vs 0) | .2891968 .0250557 11.54 0.000 .2400886 .3383049
------------------------------------------------------------------------------

* --- NTL: 3 vs 0 (high mining vs no mining) ---
preserve
keep if treatment == 3 | treatment == 0
gen byte treat_3v0 = (treatment == 3)
cate aipw (ntl_log $catevars) (treat_3v0), ///
controls($controls) rseed(12345) xfolds(5)
estimates store aipw_ntl_3v0
restore

Conditional average treatment effects Number of observations = 2,700
Estimator: Augmented IPW Number of folds in cross-fit = 5
Outcome model: Linear lasso Number of outcome controls = 28
Treatment model: Logit lasso Number of treatment controls = 28
CATE model: Random forest Number of CATE variables = 10
------------------------------------------------------------------------------
| Robust
ntl_log | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
treat_3v0 |
(1 vs 0) | .6111885 .0250606 24.39 0.000 .5620707 .6603063
------------------------------------------------------------------------------

* --- NTL: 2 vs 1 (medium vs low prices, within mining) ---
preserve
keep if treatment == 2 | treatment == 1
gen byte treat_2v1 = (treatment == 2)
cate aipw (ntl_log $catevars) (treat_2v1), ///
controls($controls) rseed(12345) xfolds(5)
estimates store aipw_ntl_2v1
restore

Conditional average treatment effects Number of observations = 300
Estimator: Augmented IPW Number of folds in cross-fit = 5
Outcome model: Linear lasso Number of outcome controls = 28
Treatment model: Logit lasso Number of treatment controls = 28
CATE model: Random forest Number of CATE variables = 10
------------------------------------------------------------------------------
| Robust
ntl_log | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
treat_2v1 |
(1 vs 0) | -.0112177 .0883033 -0.13 0.899 -.1842889 .1618535
------------------------------------------------------------------------------

* --- NTL: 3 vs 2 (high vs medium prices, within mining) ---
* Note: AIPW fails on this tiny subsample (N=300) due to propensity
* score overlap violations. PO with relaxed tolerance handles this,
* but the estimate is unreliable due to the extreme small sample.
preserve
keep if treatment == 3 | treatment == 2
gen byte treat_3v2 = (treatment == 3)
cate po (ntl_log $catevars) (treat_3v2), ///
controls($controls) rseed(12345) xfolds(5) ///
pstolerance(1e-8)
estimates store aipw_ntl_3v2
restore

Overlap failure. The 3-2 comparison (high vs medium prices) has only 300 observations split roughly evenly between two treated groups. The AIPW estimator fails entirely due to propensity scores near zero. Even the PO estimator with pstolerance(1e-8) — which relaxes the minimum acceptable propensity score from the default 1e-5 to 1e-8 — produces an unreliable ATE of –43,825 (SE = 43,752, p = 0.317). This comparison is excluded from the summary table below. The remaining five contrasts are well-identified.

7.4 Conflict: Mining effect (1-0) — PO vs AIPW

Does mining increase conflict? We show both estimators for the key contrast. Unlike NTL, the conflict ground truths are not specified in the DGP, so we interpret directionally.

* --- Conflict: 1 vs 0 (PO estimator) ---
preserve
keep if treatment == 1 | treatment == 0
gen byte treat_1v0 = (treatment == 1)
cate po (conflict $catevars) (treat_1v0), ///
controls($controls) ///
rseed(12345) xfolds(5) ///
omethod(rforest) tmethod(rforest)
estimates store po_conf_1v0
restore

Conditional average treatment effects Number of observations = 2,700
Estimator: Partialing out Number of folds in cross-fit = 5
Outcome model: Random forest Number of outcome controls = 28
Treatment model: Random forest Number of treatment controls = 28
CATE model: Random forest Number of CATE variables = 10
------------------------------------------------------------------------------
| Robust
conflict | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
treat_1v0 |
(1 vs 0) | .0630853 .0130031 4.85 0.000 .0375997 .0885709
------------------------------------------------------------------------------

* --- Conflict: 1 vs 0 (AIPW estimator) ---
preserve
keep if treatment == 1 | treatment == 0
gen byte treat_1v0 = (treatment == 1)
cate aipw (conflict $catevars) (treat_1v0), ///
controls($controls) ///
rseed(12345) xfolds(5) ///
omethod(rforest) tmethod(rforest)
estimates store aipw_conf_1v0
restore

Conditional average treatment effects Number of observations = 2,700
Estimator: Augmented IPW Number of folds in cross-fit = 5
Outcome model: Random forest Number of outcome controls = 28
Treatment model: Random forest Number of treatment controls = 28
CATE model: Random forest Number of CATE variables = 10
------------------------------------------------------------------------------
| Robust
conflict | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
treat_1v0 |
(1 vs 0) | .0659767 .0122036 5.41 0.000 .042058 .0898954
------------------------------------------------------------------------------

Both estimators produce positive and significant ATEs (PO = 0.063, AIPW = 0.066, both p < 0.001), confirming Finding 1: mining increases both nighttime lights and conflict. The baseline conflict probability for non-mining districts is approximately 10.7%, and mining increases it by about 6.5 percentage points.

7.5 Conflict: Remaining comparisons (AIPW only)

* Loop over remaining conflict comparisons
local comparisons "2_0 3_0 2_1 3_1 3_2"
foreach comp of local comparisons {
local t_hi = substr("`comp'", 1, 1)
local t_lo = substr("`comp'", 3, 1)
preserve
keep if treatment == `t_hi' | treatment == `t_lo'
gen byte treat_bin = (treatment == `t_hi')
quietly cate aipw (conflict $catevars) (treat_bin), ///
controls($controls) rseed(12345) xfolds(5)
matrix b = e(b)
matrix V = e(V)
display as result "Conflict `t_hi' vs `t_lo': ATE = " %7.4f b[1,1] ///
" SE = " %7.4f sqrt(V[1,1])
estimates store aipw_conf_`t_hi'v`t_lo'
restore
}

=== Conflict: Treatment 2 vs 0 ===
N = 2700
ATE = 0.0728 SE = 0.0330
=== Conflict: Treatment 3 vs 0 ===
N = 2700
ATE = 0.1586 SE = 0.0380
=== Conflict: Treatment 2 vs 1 ===
N = 300
ATE = -0.0677 SE = 0.0497
=== Conflict: Treatment 3 vs 1 ===
N = 300
ATE = 0.1126 SE = 0.0293
=== Conflict: Treatment 3 vs 2 ===
N = 300
ATE = 3.5e+04 SE = 3.5e+04 (overlap failure -- unreliable)

7.6 ATE summary

* Compile NTL AIPW ATEs into a comparison table
display as text "{hline 70}"
display as text "SUMMARY: Average Treatment Effects (NTL Outcome)"
display as text "{hline 70}"
display as text "Contrast" _col(15) "AIPW ATE" _col(30) "SE" _col(42) "Ground Truth"
display as text "{hline 70}"
local comps "1v0 2v0 3v0 2v1 3v1 3v2"
local gts "0.25 0.30 0.55 0.05 0.30 0.25"
local i = 1
foreach comp of local comps {
local gt : word `i' of `gts'
quietly estimates restore aipw_ntl_`comp'
matrix b = e(b)
matrix V = e(V)
local ate = b[1,1]
local se = sqrt(V[1,1])
display as result "`comp'" _col(15) %7.4f `ate' _col(30) %7.4f `se' _col(42) "`gt'"
local ++i
}
display as text "{hline 70}"

----------------------------------------------------------------------
SUMMARY: Average Treatment Effects (NTL Outcome)
----------------------------------------------------------------------
Contrast ATE SE Ground Truth
----------------------------------------------------------------------
1v0 0.1490 0.0106 0.25
2v0 0.2892 0.0251 0.30
3v0 0.6112 0.0251 0.55
2v1 -0.0112 0.0883 0.05
3v1 0.4053 0.0255 0.30
3v2 (overlap failure) 0.25
----------------------------------------------------------------------

Several estimates deviate from the ground truth (e.g., 1v0 = 0.149 vs truth 0.25; 3v1 = 0.405 vs truth 0.30). These deviations reflect finite-sample variability, the particular random seed, and the challenge of estimating effects with only 150 treated observations per group. The directional patterns are robust: all mining effects are positive, and the price non-linearity is clear. With larger samples or different seeds, estimates would converge closer to the DGP values.

Two findings emerge from the ATE summary:

Finding 1: Mining increases nighttime lights. All three mining-vs-no-mining comparisons (1-0, 2-0, 3-0) show positive and significant ATEs (0.149, 0.289, 0.611), with magnitudes increasing as the mineral price level rises. The 3-0 contrast (high-price mining vs no mining) is the largest at 0.611 — the combined effect of mining itself plus the high price premium.

Finding 2: Price effects are non-linear. The within-mining price contrasts confirm non-linearity: 2-1 (medium vs low prices) is essentially zero (–0.011, p = 0.90), while 3-1 (high vs low prices) is large and significant (0.405, p < 0.001). Price effects are not a smooth dose-response — they “jump” sharply only at high prices. The step from low to medium prices does nothing; the step from low to high prices does a lot.

8. Treatment effect heterogeneity (GATEs)

The key innovation of causal machine learning is detecting how treatment effects vary across subgroups. We compute GATEs (Group Average Treatment Effects) by institutional variables to test Finding 3: institutions moderate mining effects but NOT price effects.

8.1 GATEs by executive constraints: Mining effect (1-0)

preserve
keep if treatment == 1 | treatment == 0
gen byte treat_1v0 = (treatment == 1)
cate aipw (ntl_log $catevars) (treat_1v0), ///
controls($controls) ///
group(exec_con) ///
rseed(12345) xfolds(5) ///
omethod(rforest) tmethod(rforest)
categraph gateplot
estat gatetest
estimates store gate_ntl_1v0_exec
restore

Conditional average treatment effects Number of observations = 2,700
Estimator: Augmented IPW Number of folds in cross-fit = 5
------------------------------------------------------------------------------
| Robust
ntl_log | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
GATE |
exec_con |
1 | .2748407 .0403765 6.81 0.000 .1957042 .3539772
2 | .3155337 .0204714 15.41 0.000 .2754106 .3556569
3 | .1674459 .020837 8.04 0.000 .1266061 .2082857
4 | .1131603 .0263687 4.29 0.000 .0614785 .164842
5 | .0998745 .0296118 3.37 0.001 .0418364 .1579127
6 | .0508165 .0220009 2.31 0.021 .0076955 .0939374
-------------+----------------------------------------------------------------
ATE |
treat_1v0 |
(1 vs 0) | .1517508 .0111077 13.66 0.000 .12998 .1735216
------------------------------------------------------------------------------
Group treatment-effects heterogeneity test
H0: Group average treatment effects are homogeneous
chi2(5) = 96.90
Prob > chi2 = 0.0000

The GATE plot reveals a downward slope: districts with weaker executive constraints (lower values on the x-axis) experience larger mining effects on nighttime lights (GATE = 0.275 at exec_con = 1 vs 0.051 at exec_con = 6). The estat gatetest strongly rejects GATE equality (chi2(5) = 96.90, p < 0.0001), confirming that institutional quality moderates mining effects.

This pattern — weaker institutions, larger mining benefit — differs from the sign that Hodler et al. (2023) found in real Sub-Saharan African data. In the full paper, stronger institutions amplified the development benefits of mining. In our simulated data, the DGP produces the opposite sign: mining has a larger positive effect on NTL in weakly-governed districts, perhaps because these districts start from a lower baseline and have more room for growth when mining begins. The key takeaway is that institutional moderation exists (the GATEs are clearly heterogeneous), even though the direction differs from the full paper’s parametrization.

8.2 GATEs by executive constraints: Price effect (3-1)

preserve
keep if treatment == 3 | treatment == 1
gen byte treat_3v1 = (treatment == 3)
cate aipw (ntl_log $catevars) (treat_3v1), ///
controls($controls) ///
group(exec_con) ///
rseed(12345) xfolds(5) ///
omethod(rforest) tmethod(rforest)
categraph gateplot
estat gatetest
estimates store gate_ntl_3v1_exec
restore

Conditional average treatment effects Number of observations = 300
Estimator: Augmented IPW Number of folds in cross-fit = 5
------------------------------------------------------------------------------
| Robust
ntl_log | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
GATE |
exec_con |
1 | .3211384 .0790347 4.06 0.000 .1662332 .4760437
2 | .2868582 .0726244 3.95 0.000 .1445171 .4291993
3 | .3729897 .0413647 9.02 0.000 .2919164 .4540629
4 | .5891193 .0542141 10.87 0.000 .4828616 .6953771
5 | .4870458 .0590345 8.25 0.000 .3713403 .6027514
6 | .3400699 .0596378 5.70 0.000 .2231819 .4569579
-------------+----------------------------------------------------------------
ATE |
treat_3v1 |
(1 vs 0) | .4062996 .0252193 16.11 0.000 .3568707 .4557284
------------------------------------------------------------------------------
Group treatment-effects heterogeneity test
H0: Group average treatment effects are homogeneous
chi2(5) = 18.92
Prob > chi2 = 0.0020

The GATE plot for the price effect (3-1) shows a non-monotone pattern: GATEs range from 0.29 to 0.59 across executive constraint levels with no clear directional trend. While the estat gatetest rejects equality (chi2(5) = 18.92, p = 0.002), the pattern lacks the clear monotone slope seen in the mining effect. The price effect is positive everywhere and does not systematically vary with institutional quality in the same way.

The key contrast (Finding 3). Compare the two GATE plots above. Mining effect (1-0): clear monotone downward slope — a strong, systematic relationship between institutional quality and the mining effect (chi2 = 96.90). Price effect (3-1): no monotone pattern — while some variation exists, there is no clear directional relationship between institutions and the price premium. This asymmetry supports the paper’s core insight: institutional quality systematically moderates mining effects, but does not systematically shape how global commodity price shocks affect local economic activity.

8.3 GATEs by quality of government: Mining effect (1-0)

We repeat the analysis using an alternative institutional measure — quality of government — discretized into quartiles.

preserve
keep if treatment == 1 | treatment == 0
gen byte treat_1v0 = (treatment == 1)
egen qog_cat = cut(quality_of_govt), group(4) label
cate aipw (ntl_log $catevars) (treat_1v0), ///
controls($controls) ///
group(qog_cat) ///
rseed(12345) xfolds(5) ///
omethod(rforest) tmethod(rforest)
categraph gateplot
estat gatetest
estimates store gate_ntl_1v0_qog
restore

GATE |
qog_cat |
.22- | .2978846 .0215258 13.84 0.000 .2556947 .3400745
.32- | .168479 .0205681 8.19 0.000 .1281663 .2087917
.42- | .1080724 .0242792 4.45 0.000 .060486 .1556589
.58- | .0728521 .0179392 4.06 0.000 .037692 .1080123
-------------+----------------------------------------------------------------
ATE |
(1 vs 0) | .1504898 .0107088 14.05 0.000 .1295009 .1714786
------------------------------------------------------------------------------
Group treatment-effects heterogeneity test
H0: Group average treatment effects are homogeneous
chi2(3) = 69.19
Prob > chi2 = 0.0000

The quality-of-government GATE plot confirms the same downward pattern seen with executive constraints: districts in the lowest QoG quartile (0.22–0.32) show a GATE of 0.298, while the highest quartile (0.58+) shows only 0.073. The estat gatetest strongly rejects equality (chi2(3) = 69.19, p < 0.0001). Both institutional measures tell the same story: weaker governance, larger mining effect.

8.4 GATEs by quality of government: Price effect (3-1)

preserve
keep if treatment == 3 | treatment == 1
gen byte treat_3v1 = (treatment == 3)
egen qog_cat = cut(quality_of_govt), group(4) label
cate aipw (ntl_log $catevars) (treat_3v1), ///
controls($controls) ///
group(qog_cat) ///
rseed(12345) xfolds(5) ///
omethod(rforest) tmethod(rforest)
categraph gateplot
estat gatetest
estimates store gate_ntl_3v1_qog
restore

GATE |
qog_cat |
.22- | .3224114 .0784869 4.11 0.000 .1685799 .476243
.28- | .3510705 .0408896 8.59 0.000 .2709285 .4312126
.38- | .4843956 .0567094 8.54 0.000 .3732473 .5955438
.48- | .4447202 .039344 11.30 0.000 .3676074 .5218329
-------------+----------------------------------------------------------------
ATE |
(1 vs 0) | .4057735 .0253689 15.99 0.000 .3560514 .4554957
------------------------------------------------------------------------------
Group treatment-effects heterogeneity test
H0: Group average treatment effects are homogeneous
chi2(3) = 5.81
Prob > chi2 = 0.1211

The price effect GATEs by QoG range from 0.322 to 0.484 without a clear monotone pattern, and the estat gatetest fails to reject equality (chi2(3) = 5.81, p = 0.121). This confirms the asymmetry: institutional quality does not systematically moderate the price premium, unlike the mining effect where the relationship is strong and monotone.

8.5 GATEs for conflict: Mining effect (1-0)

preserve
keep if treatment == 1 | treatment == 0
gen byte treat_1v0 = (treatment == 1)
cate aipw (conflict $catevars) (treat_1v0), ///
controls($controls) ///
group(exec_con) ///
rseed(12345) xfolds(5) ///
omethod(rforest) tmethod(rforest)
categraph gateplot
estat gatetest
estimates store gate_conf_1v0_exec
restore

GATE |
exec_con |
1 | .0924768 .0440987 2.10 0.036 .0060449 .1789087
2 | .032707 .0348295 0.94 0.348 -.0355576 .1009715
3 | .0600398 .0292415 2.05 0.040 .0027275 .1173522
4 | .0486042 .0273151 1.78 0.075 -.0049326 .1021409
5 | .0643314 .0205048 3.14 0.002 .0241427 .1045202
6 | .1057752 .021425 4.94 0.000 .0637831 .1477674
-------------+----------------------------------------------------------------
ATE |
(1 vs 0) | .0648653 .0122278 5.30 0.000 .0408994 .0888313
------------------------------------------------------------------------------
Group treatment-effects heterogeneity test
H0: Group average treatment effects are homogeneous
chi2(5) = 5.00
Prob > chi2 = 0.4162

For conflict, the GATEs range from 0.033 to 0.106 across executive constraint levels with no clear monotone pattern. The estat gatetest fails to reject equality (chi2(5) = 5.00, p = 0.416), indicating that institutional quality does not significantly moderate the conflict effect of mining in this simulated dataset. All groups show a positive conflict effect, but without systematic variation.

9. Advanced diagnostics

Stata 19’s cate suite provides several postestimation tools that go beyond group-level summaries. This section demonstrates IATE distributions, formal heterogeneity tests, subpopulation ATEs, linear projections, and IATE function plots.

9.1 IATE distribution and heterogeneity test

We re-estimate the NTL mining effect (1-0) with i.exec_con in the catevarlist to enable reestimate group(exec_con) later.

preserve
keep if treatment == 1 | treatment == 0
gen byte treat_1v0 = (treatment == 1)
cate aipw (ntl_log exec_constraints quality_of_govt gdp_pc ///
elevation temperature ruggedness distance_capital ///
agri_suitability population ethnic_frac ///
i.exec_con) (treat_1v0), ///
controls($controls) ///
rseed(12345) xfolds(5) ///
omethod(rforest) tmethod(rforest)
* Distribution of individual effects
categraph histogram

Conditional average treatment effects Number of observations = 2,700
Estimator: Augmented IPW Number of folds in cross-fit = 5
Outcome model: Random forest Number of outcome controls = 34
Treatment model: Random forest Number of treatment controls = 34
CATE model: Random forest Number of CATE variables = 16
------------------------------------------------------------------------------
| Robust
ntl_log | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
treat_1v0 |
(1 vs 0) | .1517508 .0111077 13.66 0.000 .12998 .1735216
------------------------------------------------------------------------------

The histogram shows the full distribution of estimated individual treatment effects $\hat{\tau}(\mathbf{x}_i)$ across all districts. A wide spread indicates substantial heterogeneity; a spike at one value would indicate near-homogeneity. The distribution is centered around the ATE of approximately 0.15, with meaningful spread reflecting how institutional quality, geography, and economic conditions create different mining effects across districts.

* Formal test: are treatment effects heterogeneous?
estat heterogeneity

Treatment-effects heterogeneity test
H0: Treatment effects are homogeneous
chi2(1) = 53.05
Prob > chi2 = 0.0000

Interpreting the heterogeneity test. The estat heterogeneity test uses the method of Chernozhukov et al. (2006). A significant result (p < 0.05) provides statistical evidence that treatment effects vary across observations — they are not constant. This justifies the use of CATE methods rather than a simple ATE.

9.2 GATE equality test with reestimate

The reestimate option recycles the IATE function from the previous estimation. We do NOT refit the (slow) causal forest — we just recompute group means. This makes it fast to explore different grouping variables.

* Recompute GATEs by Executive Constraints from existing IATEs
cate, reestimate group(exec_con)
* Test H0: GATEs are equal across executive constraint levels
estat gatetest

GATE |
exec_con |
1 | .2748407 .0403765 6.81 0.000 .1957042 .3539772
2 | .3155337 .0204714 15.41 0.000 .2754106 .3556569
3 | .1674459 .020837 8.04 0.000 .1266061 .2082857
4 | .1131603 .0263687 4.29 0.000 .0614785 .164842
5 | .0998745 .0296118 3.37 0.001 .0418364 .1579127
6 | .0508165 .0220009 2.31 0.021 .0076955 .0939374
Group treatment-effects heterogeneity test
H0: Group average treatment effects are homogeneous
chi2(5) = 96.90
Prob > chi2 = 0.0000

9.3 ATE for subpopulations

We can estimate ATEs for specific subsets of the data using estat ate. This answers the policy question: “What is the average effect of mining specifically for districts with strong (or weak) institutions?”

* ATE for districts with strong institutions
estat ate if exec_constraints >= 4
* ATE for districts with weak institutions
estat ate if exec_constraints <= 2

--- ATE for districts with exec_constraints >= 4 ---
Treatment-effects estimation Number of obs = 1,526
------------------------------------------------------------------------------
| Robust
| Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
treat_1v0 |
(1 vs 0) | .0922992 .0156897 5.88 0.000 .0615479 .1230506
------------------------------------------------------------------------------
--- ATE for districts with exec_constraints <= 2 ---
Treatment-effects estimation Number of obs = 558
------------------------------------------------------------------------------
| Robust
| Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
treat_1v0 |
(1 vs 0) | .2970104 .0215156 13.80 0.000 .2548407 .3391801
------------------------------------------------------------------------------

The subpopulation ATEs reveal a stark difference: districts with weak institutions (exec_constraints $\leq$ 2) have an ATE of 0.297 (SE = 0.022), while districts with strong institutions (exec_constraints $\geq$ 4) have an ATE of only 0.092 (SE = 0.016). The mining effect is more than three times larger in weakly-governed districts. This confirms the GATE pattern and demonstrates that institutions systematically moderate the magnitude of mining’s developmental impact.

9.4 Linear projection of IATEs

The estat projection command regresses the estimated $\hat{\tau}_i$ on covariates linearly. This provides an interpretable summary of which variables drive heterogeneity — think of it as “an OLS view of the function $\tau(\mathbf{x})$.”

estat projection exec_constraints quality_of_govt gdp_pc elevation temperature

Treatment-effects linear projection Number of obs = 2,700
F(5, 2694) = 13.95
Prob > F = 0.0000
R-squared = 0.0235
------------------------------------------------------------------------------
| Robust
| Coefficient std. err. t P>|t| [95% conf. interval]
-------------+----------------------------------------------------------------
exec_const~s | -.026133 .0470456 -0.56 0.579 -.1183822 .0661162
quality_of~t | -.862502 .6669174 -1.29 0.196 -2.170224 .4452196
gdp_pc | .000067 .0000388 1.73 0.084 -9.06e-06 .000143
elevation | -.0001258 .0000341 -3.69 0.000 -.0001928 -.0000589
temperature | .0045969 .0022266 2.06 0.039 .0002309 .0089629
_cons | .4351266 .1043275 4.17 0.000 .2305565 .6396967
------------------------------------------------------------------------------

The linear projection (R-squared = 0.024) reveals that elevation (coeff = –0.0001, p < 0.001) and temperature (coeff = 0.005, p = 0.039) are the strongest linear predictors of the individual treatment effect. Institutional variables (exec_constraints and quality_of_govt) have negative coefficients but are not individually significant in this linear summary, despite driving the GATE heterogeneity. This is expected — the relationship between institutions and the treatment effect is nonlinear (as the GATE plots show), and a linear projection cannot capture the full pattern that the random forest identifies.

9.5 IATE function plots

The categraph iateplot command shows how the IATE function varies with one covariate at a time, holding all others at their reference values. This is the most intuitive visualization of heterogeneity.

categraph iateplot exec_constraints

categraph iateplot quality_of_govt
restore

The IATE function plots show downward trends: as institutional quality increases (whether measured by executive constraints or quality of government), the predicted treatment effect of mining on nighttime lights decreases. This provides visual confirmation of the GATE findings and complements the bar charts with a continuous view of the relationship. The downward slope is consistent with the subpopulation ATEs: mining has a larger developmental effect in weakly-governed districts.

10. Conclusion

10.1 Mapping results to paper findings

Finding	Paper Result	Tutorial Evidence	Stata Command
1. Mining increases NTL	Positive ATEs (1-0, 2-0, 3-0)	Confirmed: ATEs 0.15–0.61	`cate aipw ... (treat_1v0)`
2. Non-linear prices	ATE(2-1) « ATE(3-1)	Confirmed: -0.01 vs 0.41	`estimates restore`
3. Institutions moderate mining	Monotone GATE slope for 1-0	Confirmed (downward): chi2 = 96.9	`cate ... group(exec_con)`
3. NOT prices	No monotone GATE for 3-1	Confirmed: chi2 = 5.81, p = 0.12 (QoG)	`cate ... group(qog_cat)`

Note on direction. The paper found that stronger institutions amplify mining benefits (upward slope). Our simulated data shows the opposite sign — weaker institutions yield larger mining effects — but the key structural finding (systematic institutional moderation of mining, not of prices) is reproduced. The direction difference reflects DGP parametrization, not a methodological failure.

10.2 What differs from the full paper

Aspect	Tutorial	Full Paper
Data	3,000 simulated obs	60,121 real obs
Districts	300	3,800
Countries	8 (fictional)	42 (Sub-Saharan Africa)
Covariates	10	60+
Treatment	4-level, simulated	29 minerals, real prices
Inference	5-fold cross-fitting	1,000 bootstrap
Outcomes	2 (NTL, Conflict)	2 (same)
Method	Stata `cate` (GRF)	MCF (Lechner, 2019)

10.3 Glossary

Term	Definition
ATE	Average Treatment Effect — the mean effect across the entire population
CATE	Conditional Average Treatment Effect — the ATE conditional on characteristics $\mathbf{x}$
IATE	Individualized ATE — one effect per observation, $\tau(\mathbf{x}_i)$
GATE	Group ATE — average effect within prespecified groups
GATES	Sorted Group ATE — groups formed by quantiles of IATE estimates (data-driven)
PO	Partialing-Out estimator — residualizes outcome and treatment, then estimates $\tau(\mathbf{x})$ via GRF
AIPW	Augmented IPW — doubly robust estimator combining outcome model and propensity score
GRF	Generalized Random Forest — the nonparametric method underlying Stata’s `cate` (Athey et al., 2019)
Cross-fitting	Sample-splitting procedure to prevent overfitting of nuisance models
Honest tree	Tree that uses separate subsamples for splitting and leaf estimation (enables valid inference)
catevarlist	Variables in Stata’s `cate` that drive treatment-effect heterogeneity ($\mathbf{x}$)
controls()	Additional variables for nuisance models only ($\mathbf{w}$), not for heterogeneity

10.4 Key advantages of Stata 19’s `cate`

No external packages — everything is built into Stata 19.

Formal hypothesis tests — estat heterogeneity and estat gatetest provide rigorous inference that specialized Python packages do not offer natively.

Publication-ready visualization — categraph produces polished plots directly.

Integrated workflow — seamlessly combines with Stata’s ecosystem (estimates store, preserve/restore, margins).

Doubly robust — AIPW estimator is consistent even if one nuisance model is wrong.

10.5 Exercises

Change the estimator: Re-run the NTL 1-0 comparison using omethod(lasso) tmethod(lasso) instead of random forest. Do the ATE and GATEs change substantially?
Vary cross-fitting folds: Try xfolds(10) instead of 5 for the 1-0 comparison. Is there a precision gain? Does the ATE shift?
Data-driven groups: Replace group(exec_con) with group(4) to let the data discover heterogeneity groups via IATE quantiles (GATES). Do the data-driven groups align with institutional quality?
Subpopulation policy analysis: Use estat ate if gdp_pc > 6000 to estimate the ATE for richer districts. How does the mining effect compare to the full-sample ATE?
Additional heterogeneity: Investigate whether GDP per capita moderates treatment effects using categraph iateplot gdp_pc. Is there a clear pattern?

10.6 References

Hodler, R., Lechner, M., & Raschky, P. A. (2023). Institutions and the resource curse: New insights from causal machine learning. PLoS ONE, 18(5), e0284968.
Athey, S., Tibshirani, J., & Wager, S. (2019). Generalized random forests. Annals of Statistics, 47(2), 1148–1178.
Nie, X., & Wager, S. (2021). Quasi-oracle estimation of heterogeneous treatment effects. Biometrika, 108(2), 299–319.
Knaus, M. C. (2022). Double machine learning-based programme evaluation under unconfoundedness. Econometrics Journal, 25(3), 602–627.
Kennedy, E. H. (2023). Towards optimal doubly robust estimation of heterogeneous causal effects. Electronic Journal of Statistics, 17(2), 3008–3049.
Sachs, J. D., & Warner, A. M. (1995). Natural resource abundance and economic growth. NBER Working Paper No. 5398.
Mehlum, H., Moene, K., & Torvik, R. (2006). Institutions and the resource curse. The Economic Journal, 116(508), 1–20.
StataCorp. (2025). Stata 19 Treatment-Effects Reference Manual: cate. College Station, TX: Stata Press.

Conditional Average Treatment Effects (CATE) with Stata 19

Fri, 01 May 2026 00:00:00 +0000

1. Overview

The textbook causal-inference workflow ends with a single number — the Average Treatment Effect (ATE). But policy makers, doctors, and managers rarely care only about the average. They want to know for whom the program works best, for whom it does little, and whether the gains are worth the cost in any particular subgroup. This question — how the treatment effect varies across the covariates — is captured by the Conditional Average Treatment Effect (CATE), also written $\tau(x) = E\{y(1) - y(0) \mid x = x\}$.

Until very recently, estimating CATE in Stata required hand-rolled forvalues loops, careful interactions, and uncomfortably-ad-hoc inference. Stata 19 changed that with the new cate command, which builds on the doubly robust scores of Athey, Tibshirani & Wager (2019) and the partialing-out workflow of Chernozhukov et al. (2018). With one command, Stata 19 now runs cross-fitted lasso for the nuisance functions, a generalized random forest for the individual-effect function $\tau(x)$, and an honest-tree bootstrap for confidence intervals. Postestimation tools — estat heterogeneity, estat projection, categraph gateplot, estat classification, estat series — turn the resulting object into pictures that beginners can read directly.

This tutorial walks through the full cate workflow on the canonical 401(k) eligibility study (webuse assets3, 9,913 households). We start with a single ATE, show that it hides a wide fan of household-level effects, and then peel back the heterogeneity in five complementary ways: a histogram of individual effects, an IATE-by-covariate plot, a GATE on prespecified income groups, GATES on data-driven quartiles, and a smooth nonparametric series fit. The result is a complete picture of who benefits — and a reusable template you can drop into your own observational data.

Prerequisite. This post requires Stata 19 or later. The cate command does not exist in Stata 18. The do-file aborts on startup if it detects an older Stata.

1.1 Learning objectives

By the end of this tutorial you should be able to:

Understand why the ATE alone can mislead and how the CATE function $\tau(x)$ describes treatment effect heterogeneity.
Implement Stata 19’s cate command using both the partialing-out (PO) and the augmented inverse-probability weighting (AIPW) estimators on observational data.
Estimate group-level effects (GATE) on prespecified groups and data-driven quartiles (GATES) of the predicted effect.
Diagnose treatment-effect heterogeneity with estat heterogeneity, summarize who responds with estat projection and estat classification, and visualize the dose-response with estat series.
Compare doubly robust ML estimates (PO, AIPW) to a parametric teffects aipw benchmark and judge whether the average is hiding important variation.

1.2 Methodological overview

The diagram below shows the two routes through the cate command and the postestimation tools that probe the resulting CATE object.

flowchart TB
A["assets3 dataset<br/>9,913 households<br/>e401k -> assets"]:::data
A --> B{cate command}:::main
B -->|"<b>cate po</b><br/>Partial-linear model<br/>Robust to small propensities"| C["PO estimator<br/>Cross-fit lasso + causal forest"]:::po
B -->|"<b>cate aipw</b><br/>Fully interactive model<br/>Doubly robust, more efficient"| D["AIPW estimator<br/>Cross-fit lasso + causal forest"]:::aipw
C --> E["IATE function<br/>tau-hat(x_i) per household"]:::iate
D --> E
E --> F1["categraph histogram<br/>distribution of effects"]:::post
E --> F2["categraph iateplot<br/>tau vs covariate"]:::post
E --> F3["estat heterogeneity<br/>H0: tau(x) constant"]:::post
E --> F4["estat projection<br/>linear summary of who"]:::post
E --> F5["GATE / GATES<br/>group-level effects"]:::post
E --> F6["estat classification<br/>top vs bottom profile"]:::post
E --> F7["estat series<br/>smooth derivative"]:::post
classDef data fill:#6a9bcc,stroke:#141413,color:#fff
classDef main fill:#141413,stroke:#141413,color:#fff
classDef po fill:#6a9bcc,stroke:#141413,color:#fff
classDef aipw fill:#d97757,stroke:#141413,color:#fff
classDef iate fill:#00d4c8,stroke:#141413,color:#141413
classDef post fill:#f5f5f5,stroke:#141413,color:#141413

The two branches (PO and AIPW) make different model assumptions but produce the same kind of object: a function $\hat{\tau}(x_i)$ that returns a predicted treatment effect for every household. Postestimation commands then summarize that function in different ways — as a distribution (histogram), a function of one covariate (iateplot), a test (estat heterogeneity), a regression summary (estat projection), or a group-level table (GATE / GATES). All seven postestimation views answer slightly different questions, and the last three sections of this post show why a beginner should look at all of them rather than picking one favorite.

1.3 Key concepts at a glance

The post leans on a small vocabulary repeatedly. The rest of the tutorial assumes you can move between these terms quickly. Each concept below has three parts. The definition is always visible. The example and analogy sit behind clickable cards: open them when you need them, leave them collapsed for a quick scan. If a later section mentions “GATE vs GATES” or “doubly robust” and the term feels slippery, this is the section to re-read.

1. Potential outcomes $Y_i(t)$. The outcome unit $i$ would take under treatment value $t$. Each household has two potential outcomes here: assets if eligible for a 401(k), assets if not. We observe only one. The other is counterfactual. It lives in a world we never see.

Example

Take household 1234 with e401k = 1 (eligible). We observe its assets under eligibility. Its potential outcome under non-eligibility, $Y_{1234}(0)$, is forever invisible. Causal inference is the art of imputing that missing potential outcome from comparable ineligible households.

Analogy

Every life decision is a fork in the road. You took one fork. The parallel-universe versions of yourself took the other. Their lives are real conceptual objects, but you cannot directly observe them.

2. CATE — Conditional Average Treatment Effect, $\tau(\mathbf{x})$. The average treatment effect for households with covariate profile $\mathbf{x}$. The CATE is a function of $\mathbf{x}$, not a single number. Where it bends with $\mathbf{x}$, eligibility helps some households more than others.

Example

For a high-income household (income in the top quintile), the CATE is roughly \$20,511. For a low-income household, it is closer to \$4,087. Same e401k = 1, very different effects on assets.

Analogy

A drug’s “average effect” is a 5-point reduction in blood pressure. But a doctor cares about a specific patient. Maybe a 65-year-old male with diabetes. The CATE is that personalized effect.

3. ATE — Average Treatment Effect, $E[\tau(\mathbf{X})]$. The CATE averaged across the entire sample. The headline policy number. It answers a single question: if we made everyone eligible, what would the average bump in assets be?

Example

AIPW gives an ATE of \$8,120 (95% CI [\$5,846, \$10,395]) on our 9,913 households. PO gives \$7,937 (95% CI [\$5,677, \$10,197]). The two estimates are within \$200. Their joint message: eligibility raises mean assets by about \$8,000.

Analogy

“This drug lowers cholesterol by 12 points on average.” Single number. Suitable for a press release. Says nothing about who responds best.

4. GATE — Group Average Treatment Effect. The CATE averaged inside a pre-specified subgroup. The subgroup is fixed before estimation. GATEs test moderation hypotheses you formulated in advance: “do high-income households benefit more than low-income ones?”

Example

Sort households by incomecat (lowest to highest income quintile). Average CATEs inside each level. The lowest quintile gets \$4,087. The highest gets \$20,511. The pattern is monotone and steep.

Analogy

A nationwide marketing campaign lifts sales 5% on average. Before scaling up, you ask: did it work better in cities than rural towns? Same data, broken down by a subgroup you defined in advance.

5. GATES — Group Average Treatment Effects via predicted effect quartiles. A data-driven version of GATE. Sort households by their estimated CATE $\hat{\tau}_i$, slice into quartiles Q1–Q4, then average the actual response in each quartile. The contrast Q4-vs-Q1 is the strongest moderation signal a beginner can find without naming the moderator.

Example

GATES Q1 (lowest predicted effect) = \$17,279. GATES Q4 (highest predicted effect) = \$2,919. The top-to-bottom ratio is 5.9×. Note that GATES is sorted by predicted effect, so the labels feel inverted: we let the model tell us who responds.

Analogy

Letting the data sort the patients for you. You do not need to know in advance whether age, gender, or kidney function matters. You ask: “based on the model, who is in the top 25% of predicted responders?” Then you check whether they actually respond more.

6. PO vs AIPW estimators. Two ways to map nuisance estimates into a CATE. PO (Partialing Out, partial-linear model) residualizes both assets and e401k against the covariates, then regresses one residual on the other. Simple, transparent, sensitive to extreme propensity scores. AIPW (Augmented Inverse-Probability Weighting) reweights by inverse propensity and adds a regression correction. More machinery, but doubly robust.

Example

This post fits both. They land within \$200 (PO \$7,937 vs AIPW \$8,120). When PO and AIPW are close, the model-disagreement diagnostic is green. When they diverge, the overlap is suspect or one of the nuisance models is mis-specified.

Analogy

Two judges hear the same case via different reasoning. When their verdicts agree, you trust the case. When they disagree, you re-read the evidence.

7. Heterogeneity test. A formal $\chi^2$ test that $\tau(\mathbf{x})$ varies with $\mathbf{x}$. The null is constant treatment effects: every household responds the same way. Rejection licenses the CATE / GATE / GATES interpretation.

Example

After cate, estat heterogeneity returns χ²(1) = 4.11 (p = 0.043) for PO and χ²(1) = 5.54 (p = 0.019) for AIPW. Both reject the constant-effect null at conventional levels. The post’s heterogeneity story has formal backing.

Analogy

A metal detector for hidden moderation. It does not tell you where in the field the metal is buried. It only tells you whether to keep digging.

8. Doubly robust property. A property of AIPW (and other DR estimators). The estimator stays consistent for the ATE if either the outcome model is correctly specified or the propensity model is correctly specified. Both right is gravy. Only one right is enough. Both wrong is the only failure mode.

Example

This is why AIPW (\$8,120) is given more weight in our discussion than IPW alone would be. Even if our random forests under-fit either nuisance, AIPW still recovers the truth.

Analogy

Belt and suspenders. If the belt fails, the suspenders hold. If the suspenders fail, the belt holds. Two failures simultaneously? Time to buy new pants.

2. The dataset: 401(k) eligibility and household assets

We use assets3, an excerpt from Chernozhukov & Hansen (2004) shipped with Stata 19. Each row is one household. The outcome is total net financial assets in dollars; the treatment is whether the household head’s employer offers a 401(k) plan (i.e. eligibility, not actual participation). The economic question is whether eligibility on its own — independent of contribution choices — increases retirement wealth, and the standard concern is that eligible workers differ systematically from ineligible workers (they earn more, are older, work for larger employers).

We load the data, declare which variables describe the heterogeneity we care about, and inspect the basic descriptive stats:

webuse assets3, clear
* Define the heterogeneity-of-interest covariates and (for this tutorial)
* the same set as nuisance controls.
global catecovars age educ i.incomecat i.pension i.married i.twoearn i.ira i.ownhome
global controls age educ i.incomecat i.pension i.married i.twoearn i.ira i.ownhome
global rseed 12345671
describe asset e401k age educ income incomecat pension married twoearn ira ownhome
summarize asset e401k age educ income, detail
tab e401k, missing

Variable Storage Display Value
name type format label Variable label
-------------------------------------------------------------
assets float %9.0g Net total financial assets
e401k byte %12.0g lbe401 401(k) eligibility
age byte %9.0g Age
educ byte %9.0g Years of education
income float %9.0g Household income
incomecat byte %9.0g Income category
pension byte %16.0g lbpen Pension benefits
married byte %11.0g lbmar Marital status
twoearn byte %9.0g lbyes Two-earner household
ira byte %9.0g lbyes IRA participation
ownhome byte %9.0g lbyes Homeowner
401(k) |
eligibility | Freq. Percent Cum.
-------------+-----------------------------------
Not eligible | 6,231 62.86 62.86
Eligible | 3,682 37.14 100.00
Total | 9,913 100.00

The dataset contains 9,913 households, of which 3,682 (37.1%) are eligible for a 401(k) and 6,231 (62.9%) are not. The asset distribution is extraordinarily right-skewed — mean \$18,054 against a median of just \$1,499, with a maximum of \$1.5 million and a minimum of −\$502,302 (households with negative net worth). Income, age, and education show much milder skew. Four key features matter for what follows: the treatment is roughly balanced (37% vs 63%, plenty of overlap on average), the outcome has heavy tails (so the treatment effect almost certainly varies across the distribution), and we have a rich set of demographic covariates to condition on.

3. The naive view (and why it fails)

Before reaching for any causal estimator it is healthy to look at the raw mean difference. If e401k were randomly assigned the comparison would be the ATE. It isn’t — eligibility is a function of who chooses what employer — so the raw difference is biased. Showing this gap explicitly motivates everything that follows.

* Raw means by eligibility
tabstat asset, by(e401k) statistics(mean sd n)

Summary for variables: assets
Group variable: e401k (401(k) eligibility)
e401k | Mean SD N
-------------+------------------------------
Not eligible | 10789.9 54527.02 6231
Eligible | 30347.39 74800.21 3682
-------------+------------------------------
Total | 18054.17 63528.63 9913

Eligible households hold an average of \$30,347 in net financial assets versus \$10,790 for ineligible ones — a raw gap of \$19,557. If we believed in random assignment we would call that the average effect of eligibility. But eligible workers are systematically different: they tend to be older, more educated, and earn substantially more. Some of that \$19,557 is causal, but a meaningful share is just selection. The next section pins down how much of the gap is causal once we adjust for those covariates.

4. A first ATE: parametric `teffects aipw`

Stata’s mature teffects suite already supports doubly robust ATE estimation with parametric models. We use it here as a familiar, fast benchmark before introducing the new cate command. The estimand is

$$\text{ATE} = E\{y(1) - y(0)\}$$

In words, this is the average treatment effect across all households in the population. The augmented inverse-probability weighting (AIPW) estimator is doubly robust: it returns the right ATE if either the outcome model or the propensity score model is correctly specified — we don’t need both.

teffects aipw ///
(asset c.age c.educ i.incomecat i.pension i.married i.twoearn i.ira i.ownhome) ///
(e401k c.age c.educ i.incomecat i.pension i.married i.twoearn i.ira i.ownhome)

Treatment-effects estimation Number of obs = 9,913
Estimator : augmented IPW
Outcome model : linear by ML
Treatment model: logit
------------------------------------------------------------------------------
ATE |
e401k |
(Eligible |
vs |
Not elig..) | 8019.463 1152.038 6.96 0.000 5761.51 10277.42
-------------+----------------------------------------------------------------
POmean |
e401k |
Not eligi.. | 13930.46 817.613 17.04 0.000 12327.97 15532.96
------------------------------------------------------------------------------

The doubly robust ATE is \$8,019 with a 95% confidence interval of [\$5,762, \$10,277] — about 58% above the average baseline assets of ineligible households (\$13,930). The naive raw gap (\$19,557) was therefore inflated by a factor of 2.4: roughly 60% of the observed asset gap between eligible and ineligible households is selection — they would have held more assets even without the program — and only 40% is the causal effect of eligibility. That said, \$8,019 is still just one number. The cross-tabulation of mean assets by income category and eligibility (which we computed but suppressed for length here, see analysis.log) shows differences ranging from \$5,011 in the lowest income category to \$20,949 in the highest — a 4× spread that the ATE flattens out. That spread is the CATE we now estimate properly.

5. The CATE: definition, model, and the `cate` command

The Conditional Average Treatment Effect at covariate value $x$ is defined as

$$\tau(\mathbf{x}) = E\{y(1) - y(0) \mid \mathbf{x} = \mathbf{x}\}$$

In words, this says: among all households whose covariates are $\mathbf{x}$, what is their average treatment effect? The CATE is a function of covariates, not a single number. If $\tau(\mathbf{x})$ happened to be constant, we’d be back at the ATE. Whenever it varies, the ATE is an average of these subgroup effects weighted by how common each $\mathbf{x}$ is in the data.

To estimate $\tau(\mathbf{x})$ Stata 19 offers two model specifications:

Partial-linear (PO) model. Assumes the outcome can be written as

$$y = d \cdot \tau(\mathbf{x}) + g(\mathbf{x}, \mathbf{w}) + \epsilon, \qquad d = f(\mathbf{x}, \mathbf{w}) + u$$

In words, the outcome is the treatment $d$ times the per-household effect $\tau(\mathbf{x})$, plus a flexible function $g$ of all covariates, plus noise; and the treatment itself is a flexible function $f$ of those covariates plus its own noise. PO partials out $g$ and $f$ using out-of-sample predictions (cross-fitting), then fits a generalized random forest on the residuals to recover $\tau(\mathbf{x})$. PO is the more robust choice when propensity scores can get close to 0 or 1.

Fully interactive (AIPW) model. Assumes $y(1) = g_1(\mathbf{x}, \mathbf{w}) + \epsilon_1$ and $y(0) = g_0(\mathbf{x}, \mathbf{w}) + \epsilon_0$ — separate outcome models for treated and untreated households — and combines them with the propensity score to form the doubly-robust AIPW score (Section 9). AIPW is more efficient (narrower CIs) when both models are well-specified, but more sensitive to extreme propensities.

We start with PO. The variables in $catecovars are the inputs to $\tau(\mathbf{x})$ — the dimensions on which we want to see heterogeneity — and the controls (left at the default, which equals catecovars) are passed to the nuisance models $g$ and $f$. The rseed() option fixes the cross-fitting and random-forest internals so the run is reproducible.

cate po (asset $catecovars) (e401k), rseed($rseed)

Conditional average treatment effects Number of observations = 9,913
Estimator: Partialing out Number of folds in cross-fit = 10
Outcome model: Linear lasso Number of outcome controls = 17
Treatment model: Logit lasso Number of treatment controls = 17
CATE model: Random forest Number of CATE variables = 17
------------------------------------------------------------------------------
| Robust
assets | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
e401k |
(Eligible |
vs |
Not elig..) | 7937.182 1153.017 6.88 0.000 5677.309 10197.05
------------------------------------------------------------------------------

The PO ATE — averaged over the estimated $\hat{\tau}(\mathbf{x}_i)$ across the sample — is \$7,937 with a 95% CI of [\$5,677, \$10,197]. That’s within \$80 of the parametric teffects aipw ATE in the previous section, even though cate po is doing something fundamentally different under the hood (cross-fit lasso for the nuisance models, causal forest for the IATE). When two very different estimators agree on the average, you can trust that average — and you can move on to looking at the heterogeneity.

5.1 Is there heterogeneity at all? `estat heterogeneity`

Before exploring how $\tau(\mathbf{x})$ varies, it is worth asking whether it varies. The estat heterogeneity command tests the null hypothesis that $\tau(\mathbf{x})$ is constant — that there is, in fact, no heterogeneity to study.

estat heterogeneity

Treatment-effects heterogeneity test
H0: Treatment effects are homogeneous
chi2(1) = 4.11
Prob > chi2 = 0.0427

The test rejects homogeneity at the 5% level: χ²(1) = 4.11, p = 0.043. In plain English: the data have enough information to distinguish the estimated CATE function $\hat{\tau}(\mathbf{x})$ from a constant. The rest of this tutorial is therefore not a hunt for noise — there is real heterogeneity, and the next sections describe what shape it takes.

5.2 Who responds most? `estat projection`

A causal forest fits $\hat{\tau}(\mathbf{x})$ flexibly, but a flexible function is hard to summarize in a paragraph. estat projection regresses $\hat{\tau}_i$ on the covariates linearly. The coefficients are not causal (they’re a projection of an already-estimated nonlinear function onto a linear basis), but they answer the practical question “which variables shift the predicted effect, and by how much?”.

estat projection $catecovars

Treatment-effects linear projection Number of obs = 9,913
F(11, 9901) = 4.90
Prob > F = 0.0000
age | 205.12 117.98 1.74 0.082 -26.15 436.39
educ | -442.46 488.47 -0.91 0.365 -1399.96 515.05
incomecat 1 | -2439.22 2013.52 -1.21 0.226 -6386.14 1507.69
incomecat 2 | 1874.82 2295.16 0.82 0.414 -2624.15 6373.79
incomecat 3 | 5707.69 3298.34 1.73 0.084 -757.73 12173.11
incomecat 4 | 18194.60 5398.39 3.37 0.001 7612.65 28776.54
pension Y | 3817.36 2454.44 1.56 0.120 -993.84 8628.55
ownhome Y | 3162.65 1669.59 1.89 0.058 -110.08 6435.38

The single dominant signal is income. Relative to households in the lowest income category, those in the highest income category have a predicted effect that is \$18,195 higher (p = 0.001) — the only coefficient significant at the 1% level. Homeownership lifts the predicted effect by another \$3,163 (p = 0.058) and each additional year of age adds \$205 (p = 0.082); both are borderline. Education, marriage, two-earner status, and IRA participation are essentially flat. The R² of 0.0045 is not a critique — it tells us most of the heterogeneity is genuinely nonlinear (curvature that the random forest captures and a linear projection cannot). The rest of the post zooms into where that nonlinearity lives.

6. The shape of individual-level heterogeneity

Before slicing the CATE by groups, it helps to look at the distribution of household-level effects. categraph histogram plots the predicted $\hat{\tau}_i$ for every household in the sample.

categraph histogram, ///
title("Distribution of individual treatment effects (PO)") ///
xtitle("Estimated tau_hat_i (dollars)") ///
note("Source: assets3, Stata 19 cate po")
graph export "stata_cate_iate_histogram_po.png", replace width(1200)

The distribution is strongly right-skewed. Most households cluster around a modest positive effect (the bulk of the mass sits near \$5,000–\$10,000), but a long right tail extends to \$80,000 and beyond. A small left tail dips into negative territory: a meaningful minority of households are estimated to gain little or nothing from 401(k) eligibility. This is the visual answer to “is the average hiding something?” — the average of \$7,937 is genuinely close to the median, but the spread on either side is huge. The next two views — IATE plots and GATE — describe who sits in the right tail.

6.1 How does the effect vary with one covariate? IATE plots

The categraph iateplot command holds all covariates except one fixed at sample-mean (continuous) or base (factor) values, and varies the one covariate of interest. The result is a slice through the multi-dimensional CATE function with confidence bands.

categraph iateplot age, ///
title("Estimated CATE by age") ///
ytitle("tau_hat (dollars)") xtitle("Age (years)")
graph export "stata_cate_iateplot_age.png", replace width(1200)

The age slice is broadly increasing. Younger workers (mid-20s to early 30s) have small or even slightly negative predicted effects; the line crosses into clearly positive territory around age 35–40 and continues climbing through the 50s. The intuition is straightforward: 401(k) eligibility is most valuable to workers with the financial slack and the planning horizon to take advantage of tax-deferred saving. Confidence bands narrow in the middle of the age range where most of the data lives and widen at the extremes.

The same exercise with education looks rather different:

categraph iateplot educ, ///
title("Estimated CATE by years of education") ///
ytitle("tau_hat (dollars)") xtitle("Education (years)")
graph export "stata_cate_iateplot_educ.png", replace width(1200)

The education slice is broadly flat at around \$1,000–\$3,000 across the entire range from 8 to 18 years of schooling. This is consistent with the linear projection (where the education coefficient was small and not significant). It is also a useful negative finding — once you condition on income, education adds little to the predicted effect.

7. Group-level effects: GATE on prespecified groups

Individual-level $\hat{\tau}_i$ is informative but noisy. A common practice is to summarize them by groups — either prespecified (income category, region, education tier) or data-driven (top vs bottom quartile of predicted effect). Stata’s GATE and GATES estimators are the formal versions of these two strategies.

The Group ATE (GATE) on a prespecified group $g$ is

$$\tau(g) = E\{\Gamma_i \mid G_i = g\}$$

In words, this says: the average AIPW orthogonal score $\Gamma_i$ within group $g$ — i.e., the doubly robust per-household effect score, averaged over households assigned to that group. We compute it on the income categories incomecat. The clever bit is reestimate: after running cate po once, we tell Stata to recycle the fitted IATE function and just recompute group means, saving a slow second causal-forest fit.

cate, group(incomecat) reestimate
estat gatetest

GATE | Coefficient Std. err. z P>|z| [95% conf. interval]
incomecat |
0 | 4087.014 987.7124 4.14 0.000 2151.13 6022.90
1 | 1399.398 1663.193 0.84 0.400 -1860.40 4659.20
2 | 5154.329 1349.842 3.82 0.000 2508.69 7799.97
3 | 8532.238 2287.664 3.73 0.000 4048.50 13015.98
4 | 20510.94 4723.741 4.34 0.000 11252.58 29769.30
Group treatment-effects heterogeneity test
H0: Group average treatment effects are homogeneous
chi2(4) = 18.44
Prob > chi2 = 0.0010

categraph gateplot, ///
title("GATE by income category") ///
ytitle("tau_hat (dollars)") xtitle("Income category (1 = low, 5 = high)")
graph export "stata_cate_gate_incomecat.png", replace width(1200)

The five group-level effects span an order of magnitude: \$4,087 (lowest income), \$1,399 (income category 1, not significant at p = 0.40), \$5,154, \$8,532, and \$20,511 in the highest income category — roughly five times the average. The joint test of equality (estat gatetest) rejects strongly: χ²(4) = 18.44, p = 0.001. There is one mild departure from monotonicity at category 1, which is interesting but lies just within sampling variability (its CI overlaps zero). Two important policy facts emerge: the marginal household in the top income category gains an average of about \$20,500 from 401(k) eligibility, and the marginal household at the bottom of the distribution gains about \$4,000 — but the middle-low (category 1) gains effectively nothing.

8. Data-driven groups: GATES on quartiles of $\hat{\tau}$

GATE on prespecified groups is principled but presupposes that the analyst already knows which groups matter. GATES (“Group Average Treatment Effects Sorted”) flips this around: it lets the data sort households by their predicted effect, bins them into quantiles, and reports the mean effect within each bin. Cross-fitting protects against p-hacking — each unit’s bin is determined by an out-of-sample prediction, so observations cannot leak their own outcomes into their bin assignment.

cate po (asset $catecovars) (e401k), rseed($rseed) group(4)

GATES | Coefficient Std. err. z P>|z| [95% conf. interval]
rank |
1 | 17278.94 3440.125 5.02 0.000 10536.42 24021.46
2 | 8121.04 1691.008 4.80 0.000 4806.73 11435.35
3 | 3443.83 1437.640 2.40 0.017 626.11 6261.56
4 | 2919.20 2110.320 1.38 0.167 -1216.96 7055.35
ATE | 7938.21 1152.994 6.88 0.000 5678.38 10198.04

categraph gateplot, ///
title("GATES by data-driven quartile of estimated effect") ///
ytitle("tau_hat (dollars)") xtitle("Quartile (1 = highest tau_hat, 4 = lowest)")
graph export "stata_cate_gates_quartiles.png", replace width(1200)

The data-driven ladder is clean and monotonic: the top quartile gains an average of \$17,279 (CI \$10,536–\$24,021), the second \$8,121, the third \$3,444, and the bottom \$2,919 — and the bottom quartile is not statistically distinguishable from zero (p = 0.167). The top-to-bottom ratio is 5.9×. This is the single most informative summary of heterogeneity in the dataset because the bins are constructed by the data itself rather than by a researcher choice. Roughly one in four households in the sample appears to gain little or nothing from 401(k) eligibility, while another quarter gains over twice the average effect.

8.1 Who is in the top vs the bottom quartile? `estat classification`

The data sorted itself; now we can ask what makes the top quartile different. estat classification runs a two-sample t-test for one variable at a time, comparing its mean in the top-effect rank group against its mean in the bottom-effect rank group.

estat classification age
estat classification educ
estat classification income

Variable	Top quartile (n=2,480)	Bottom quartile (n=2,471)	Difference	t-statistic
Age (years)	45.15	34.98	10.17	35.67
Education (years)	14.02	12.65	1.37	18.62
Income (\$)	62,739	26,861	35,878	56.22

The high-effect quartile is sharply different from the low-effect quartile on every dimension: about 10 years older on average (45.1 vs 35.0), with 1.4 more years of education (14.0 vs 12.7), and \$35,878 higher household income (\$62,739 vs \$26,861). All three differences are huge in t-statistic terms (19, 36, 56). Income is the dominant marker — exactly what the linear projection and the GATE-by-income picture already suggested. The story behind the numbers: 401(k) eligibility helps people who already have the financial slack and time-horizon to actually use it, and a substantial minority of the population has neither.

9. AIPW: a doubly-robust contrast

So far we have used the partialing-out estimator. The fully interactive AIPW estimator fits separate outcome models for treated and untreated households and combines them with the propensity score via the AIPW orthogonal score:

$$\Gamma_i = \left[\hat{y}(1)_i + \frac{d_i \, \{y_i - \hat{y}(1)_i\}}{\hat{f}_i}\right] - \left[\hat{y}(0)_i + \frac{(1-d_i) \, \{y_i - \hat{y}(0)_i\}}{1-\hat{f}_i}\right]$$

In words, this says: the doubly robust per-household effect score is the predicted treated outcome minus the predicted untreated outcome, each corrected by an inverse-propensity weighted residual. It is “doubly robust” because it stays consistent if either the outcome models OR the propensity-score model is correct — you only need to get one of them right. The cost is sensitivity to extreme propensities: if some households have $\hat{f}_i$ close to 0 or 1 the inverse weights blow up.

cate aipw (asset $catecovars) (e401k), rseed($rseed)
estat heterogeneity

ATE |
e401k |
(Eligible |
vs |
Not elig..) | 8120.264 1160.538 7.00 0.000 5845.652 10394.88
Treatment-effects heterogeneity test
H0: Treatment effects are homogeneous
chi2(1) = 5.54
Prob > chi2 = 0.0186

categraph histogram, ///
title("Distribution of individual treatment effects (AIPW)") ///
xtitle("Estimated tau_hat_i (dollars)") ///
note("Source: assets3, Stata 19 cate aipw")
graph export "stata_cate_iate_histogram_aipw.png", replace width(1200)

categraph iateplot educ, ///
title("Estimated CATE by education (AIPW)") ///
ytitle("tau_hat (dollars)") xtitle("Education (years)")
graph export "stata_cate_iateplot_educ_aipw.png", replace width(1200)

The AIPW ATE is \$8,120 — within \$200 of both the parametric teffects aipw ATE (\$8,019) and the PO ATE (\$7,937). The heterogeneity test now rejects more strongly (χ²(1) = 5.54, p = 0.019) than under PO (p = 0.043), consistent with AIPW’s higher efficiency when both nuisance models are well-specified. The AIPW IATE histogram (Figure 6) has the same right-skewed shape as the PO histogram but a slightly wider support — AIPW puts more mass in the tails because of the inverse-propensity correction, which is the visual signature of the overlap-sensitivity warning above. The AIPW education slice (Figure 7) is essentially identical in shape to the PO version: a broadly flat profile around the average. Across estimators, the substantive story does not change.

10. The smooth income gradient: `estat series`

categraph iateplot showed the CATE as a function of one variable with the others fixed at reference values. estat series is a complementary view — it fits a flexible smoother (cubic B-spline by default) of the predicted effect against one continuous covariate, marginalizing over the joint distribution of the others. For continuous variables like income this gives the cleanest “dose-response” picture.

estat series income if income <= 150000, graph knots(5)
graph export "stata_cate_series_income.png", replace width(1200)

Nonparametric series regression for IATE
Cubic B-spline estimation Number of obs = 9,884
Number of knots = 5
------------------------------------------------------------------------------
| Robust
| Effect std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
income | .2131162 .0502993 4.24 0.000 .1145313 .311701
------------------------------------------------------------------------------
Note: Effect estimates are averages of derivatives.

The reported “Effect” is the average derivative of the predicted treatment effect with respect to income: 0.213 (SE 0.050, p < 0.001, 95% CI [0.115, 0.312]). Translated into dollars: each additional \$1,000 of household income raises the predicted 401(k) treatment effect by about \$213 on average. The B-spline fit (Figure 8) reveals that this derivative is not constant — the slope is steepest in the middle of the income distribution and flatter at both ends — which is why a single linear-projection coefficient (\$18,195 for the highest income category) only partially captured the gradient. The series view smooths over the binning entirely.

11. Putting it all together: comparison table

The four causal estimators we ran agree closely on the average and disagree only marginally on the heterogeneity p-value:

Estimator	ATE	95% CI	Heterogeneity test	Notes
Naive raw difference	19,557	n/a	n/a	Raw mean gap; mostly selection
`teffects aipw` (parametric)	8,019	[5,762, 10,277]	—	Mature, fast benchmark
`cate po` (lasso + causal forest)	7,937	[5,677, 10,197]	χ²(1) = 4.11, p = 0.043	Robust to extreme propensities
`cate aipw` (lasso + causal forest, doubly robust)	8,120	[5,846, 10,395]	χ²(1) = 5.54, p = 0.019	Most efficient; uses AIPW score

Three independent ML and parametric estimators bracket the true ATE within a \$183 spread. Both ML estimators reject homogeneity at the 5% level. The naive raw difference of \$19,557 was inflated by a factor of 2.4 — about \$11,500 of it was selection.

12. Discussion: answering the question

We opened with the question, for whom does 401(k) eligibility increase financial assets? The eight figures and four estimators in this post answer it concretely:

The average household gains about \$8,000. Across three estimators of the ATE that span very different model assumptions, the answer is \$7,937 to \$8,120 with a narrow range. The naive raw gap of \$19,557 overstated the causal effect by 2.4×.
But the average hides substantial heterogeneity. Both estat heterogeneity tests reject a constant CATE at the 5% level; the GATE joint test rejects equality across income groups at p = 0.001; and the GATES quartile ladder spans \$2,919 (bottom quartile, not significant) to \$17,279 (top quartile) — a factor of 5.9.
Income is the dominant moderator. The linear projection coefficient on the highest income category is \$18,195 (p = 0.001). The smooth B-spline says each extra \$1,000 of income raises the effect by \$213 on average. The classification analysis says households in the top-effect quartile earn \$35,878 more on average than households in the bottom-effect quartile.
About a quarter of households gain little or nothing. The bottom GATES quartile cannot reject zero (p = 0.167), and a small left tail in both IATE histograms shows households with predicted effects close to zero or even slightly negative.
Age and homeownership matter at the margin. Older workers and homeowners gain more, but the effects are smaller and more uncertain than the income effect. Education and marital status are essentially flat once income is controlled for.

The “so what?” for policy: a 401(k) eligibility expansion targeted at low-income workers will have a much smaller per-capita asset effect than one targeted at high-income workers — but the lowest-income households still gain a real, statistically significant \$4,000 on average, suggesting the program is not pointless for them. A blanket expansion that ignores heterogeneity would systematically underestimate the gains to high earners and overestimate the gains to households in the second income decile.

13. Summary and next steps

Method takeaways.

Stata 19’s cate command unifies cross-fit ML nuisance estimation, doubly robust scores, causal-forest IATE estimation, and honest-tree inference into a single workflow. Two estimators (PO and AIPW) and seven postestimation views cover almost all practical heterogeneity questions.
PO (partialing-out) is more robust to extreme propensity scores; AIPW is more efficient when both nuisance models are well-specified. They agree on the ATE in this dataset (\$7,937 vs \$8,120, a difference of 2.3%), which is the strongest possible robustness check.
The four heterogeneity views — estat heterogeneity, estat projection, GATE/GATES, and estat series — answer different questions. A beginner should look at all of them rather than picking a favorite.

Data takeaways.

The 401(k) eligibility ATE on the assets3 sample is \$8,019 ± \$1,150.
The CATE varies from \$1,399 (income category 1) to \$20,511 (highest income category) — a 15× spread.
One in four households shows essentially no effect; one in four shows over twice the average.

Limitations.

The CATE is identified under unconfoundedness (no unmeasured confounders) given the rich set of demographic covariates. If, for instance, employer match rates differ systematically across the income distribution and we don’t observe match rates, that would bias the income gradient.
The bootstrap-of-little-bags inference behind the IATE confidence bands assumes honest random forests. With the default xfolds(10) and the default forest settings, runtime is ≈9 minutes on Stata SE 19; StataNow MP cuts this by roughly 3×.
We did not formally check propensity overlap in this post. As a follow-up, run teffects overlap after the parametric AIPW or check estat osample after the cate command.

Next steps. Try cate aipw ..., omethod(rforest) tmethod(rforest) oob for a fully nonparametric specification with out-of-bag inference (faster and more flexible than the lasso default). Or move to the lung dataset shipped with Stata 19 and explore estat policyeval to compare expected outcomes under hypothetical assignment policies (e.g., “treat only households with predicted positive effect”).

14. Exercises

Compare specifications. Re-run cate po with omethod(rforest) tmethod(rforest) (random-forest nuisance instead of lasso). How much do the GATE-by-income estimates change? Use oob to speed up the run.
Build a custom group. Create a “high-effect candidate” indicator that is 1 if age > 40 & income > 50000 & ownhome == 1, 0 otherwise. Run cate, group(high_eff_candidate) reestimate and compare the two GATEs to the GATES top vs bottom quartile in this post.
Explore another moderator. Use categraph iateplot to plot the predicted CATE against pension, married, twoearn, and ira. Which one shows the biggest difference between its categories?

Conditional Average Treatment Effects (CATE) | Carlos Mendez

Causal Machine Learning and the Resource Curse with Stata 19

1. Overview

1.1 Learning objectives

1.2 Analytical roadmap

1.3 Key concepts at a glance

2. The CATE framework

2.1 From ATE to CATE

2.2 The partial linear model

2.3 Two estimators

2.4 Three levels of treatment effects

3. Data preparation

4. Descriptive statistics

5. Naive comparison vs ground truth

6. Estimation strategy

6.1 Binary pairwise comparisons

6.2 Variable specification

7. Average treatment effects

7.1 NTL: Mining effect (1-0) — PO vs AIPW

7.2 NTL: High vs low prices (3-1) — PO vs AIPW

7.3 NTL: Remaining comparisons (AIPW only)

7.4 Conflict: Mining effect (1-0) — PO vs AIPW

7.5 Conflict: Remaining comparisons (AIPW only)

7.6 ATE summary

8. Treatment effect heterogeneity (GATEs)

8.1 GATEs by executive constraints: Mining effect (1-0)

8.2 GATEs by executive constraints: Price effect (3-1)

8.3 GATEs by quality of government: Mining effect (1-0)

8.4 GATEs by quality of government: Price effect (3-1)

8.5 GATEs for conflict: Mining effect (1-0)

9. Advanced diagnostics

9.1 IATE distribution and heterogeneity test

9.2 GATE equality test with reestimate

9.3 ATE for subpopulations

9.4 Linear projection of IATEs

9.5 IATE function plots

10. Conclusion

10.1 Mapping results to paper findings

10.2 What differs from the full paper

10.3 Glossary

10.4 Key advantages of Stata 19’s cate

10.5 Exercises

10.6 References

Conditional Average Treatment Effects (CATE) with Stata 19

1. Overview

1.1 Learning objectives

1.2 Methodological overview

1.3 Key concepts at a glance

2. The dataset: 401(k) eligibility and household assets

3. The naive view (and why it fails)

4. A first ATE: parametric teffects aipw

5. The CATE: definition, model, and the cate command

5.1 Is there heterogeneity at all? estat heterogeneity

5.2 Who responds most? estat projection

6. The shape of individual-level heterogeneity

6.1 How does the effect vary with one covariate? IATE plots

7. Group-level effects: GATE on prespecified groups

8. Data-driven groups: GATES on quartiles of $\hat{\tau}$

8.1 Who is in the top vs the bottom quartile? estat classification

9. AIPW: a doubly-robust contrast

10. The smooth income gradient: estat series

11. Putting it all together: comparison table

12. Discussion: answering the question

13. Summary and next steps

14. Exercises

15. References

10.4 Key advantages of Stata 19’s `cate`

4. A first ATE: parametric `teffects aipw`

5. The CATE: definition, model, and the `cate` command

5.1 Is there heterogeneity at all? `estat heterogeneity`

5.2 Who responds most? `estat projection`

8.1 Who is in the top vs the bottom quartile? `estat classification`

10. The smooth income gradient: `estat series`