Quantile Treatment Effects in R

What are Quantile Treatment Effects?

The Average Treatment Effect (ATE) or Average Treatment Effect on the Treated (ATT) summarizes the causal effect of a treatment by a single number: the mean difference in outcomes between treated and untreated units. This is often a natural quantity, but it can miss important heterogeneity. If a job training program substantially raises earnings at the bottom of the distribution while having little effect at the top, the ATE obscures this. If a minimum wage increase compresses the wage distribution, an average effect conceals the compression.

The Quantile Treatment Effect at level τ is QTE(τ) = QY(1)(τ) − QY(0)(τ), where QY(d)(τ) is the τ-th quantile of the potential outcome distribution under treatment status d. Mapping QTE(τ) across τ ∈ (0, 1) traces how the treatment shifts the entire outcome distribution. The Quantile Treatment Effect on the Treated (QTT) conditions on the treated group.

This vignette demonstrates unc_qte(), the qte package’s estimator for cross-sectional settings (no panel data required). For DiD-based estimators and staggered treatment adoption, see vignette("panel-estimators") and vignette("staggered-adoption").

library(qte)
library(ggplot2)
set.seed(42)
data(lalonde)
xf <- ~ age + I(age^2) + education + black + hispanic + married + nodegree

Random assignment

Under random assignment, no covariate adjustment is needed: the unconditional quantiles of the treated and control outcome distributions identify the QTE directly. unc_qte() implements the semiparametrically efficient estimator of Firpo (2007). We use the experimental Lalonde data (lalonde.exp), where treatment was randomly assigned.

res_exp <- unc_qte(
  yname  = "re78",
  dname  = "treat",
  data   = lalonde.exp,
  target = "qte",
  probs  = seq(0.1, 0.9, 0.1),
  biters = 100
)
summary(res_exp)
#> 
#> Overall ATE:  
#>       ATE    Std. Error     [ 95%  Conf. Int.]  
#>  1794.343      566.6378   683.7534    2904.933 *
#> 
#> 
#> QTE:
#>  Tau      QTE Std. Error [ 95% Simult.  Conf. Band] 
#>  0.1    0.000     0.0000         0.0000      0.0000 
#>  0.2    0.000   141.1888      -438.1941    438.1941 
#>  0.3  929.884   311.5726       -37.1140   1896.8819 
#>  0.4 1176.944   937.1751     -1731.6769   4085.5641 
#>  0.5 1091.468   841.2708     -1519.5036   3702.4394 
#>  0.6 1466.690   717.0160      -758.6433   3692.0243 
#>  0.7 1810.696   943.5590     -1117.7378   4739.1293 
#>  0.8 2300.454  1011.9796      -840.3301   5441.2377 
#>  0.9 2856.068  2239.1618     -4093.4035   9805.5402 
#> ---
#> Signif. codes: `*' confidence band does not cover 0
#> Note: the outcome looks discrete (308 distinct values; 31% at 0); quantile effects may be unreliable.

Earnings are zero for 31% of this sample, so unc_qte() warns that the outcome looks discrete, and summary() repeats the note. Quantiles in the mass point are zero in both groups, so the QTE there is zero by construction; interpret the low quantiles with care.

autoplot(res_exp, ylab = "QTE (earnings, 1978)")

QTE curve under random assignment

The point estimates are zero at the lowest quantiles, inside the mass point, and then increase with τ, from about $900 at the 30th percentile to about $2,900 at the 90th. With 445 participants, though, the uniform band covers zero at every quantile, even though the ATE of about $1,800 is statistically significant: a band that holds simultaneously over nine quantiles is much wider than a single interval.

Unconfoundedness

When treatment is not randomly assigned, the unconfoundedness assumption (conditional independence of potential outcomes given covariates) provides identification (Firpo 2007). We use the observational Lalonde data (lalonde.psid), which combines the treated group from the experiment with a comparison group from the PSID.

unc_qte() supports three estimation methods:

  • "ipw" — inverse probability weighting (propensity score re-weighting)
  • "reg" — outcome regression (conditional quantile regression)
  • "aipw" — augmented IPW (doubly robust; recommended)
xf <- ~ age + I(age^2) + education + black + hispanic + married + nodegree

res_psid <- unc_qte(
  yname      = "re78",
  dname      = "treat",
  data       = lalonde.psid,
  xformula   = xf,
  est_method = "aipw",
  target     = "qte",
  probs      = seq(0.1, 0.9, 0.1),
  biters     = 100
)
summary(res_psid)
#> 
#> Overall ATE:  
#>        ATE    Std. Error     [ 95%  Conf. Int.]  
#>  -12537.37      1740.177  -15948.05   -9126.687 *
#> 
#> 
#> QTE:
#>  Tau        QTE Std. Error [ 95% Simult.  Conf. Band]  
#>  0.1      0.000   449.2543      -2737.819   2737.8192  
#>  0.2  -6404.679  2113.9444     -19287.354   6477.9962  
#>  0.3  -8523.773  2135.0722     -21535.204   4487.6582  
#>  0.4 -11114.085  2405.7382     -25774.993   3546.8225  
#>  0.5 -12684.516  3728.2369     -35404.917  10035.8858  
#>  0.6 -13532.837  3330.9643     -33832.203   6766.5296  
#>  0.7 -14181.889  1556.9702     -23670.285  -4693.4931 *
#>  0.8 -18285.313   806.9748     -23203.131 -13367.4950 *
#>  0.9 -22580.785  3752.0629     -45446.385    284.8154  
#> ---
#> Signif. codes: `*' confidence band does not cover 0
#> Note: the outcome looks discrete (740 distinct values; 12% at 0); quantile effects may be unreliable.
autoplot(res_psid, ylab = "QTE (earnings, 1978)")

QTE curve under unconfoundedness

These estimates are large and negative, far from the experimental benchmark above. The QTE is about the whole population, which here is mostly PSID households with much higher earnings than the program’s participants, and these covariates (which leave out earnings before the program) do not make the two groups comparable (LaLonde 1986; Dehejia and Wahba 1999).

QTT under unconfoundedness

Setting target = "qtt" estimates the Quantile Treatment Effect on the Treated — the distributional effect for the subpopulation that actually received treatment.

res_qtt <- unc_qte(
  yname      = "re78",
  dname      = "treat",
  data       = lalonde.psid,
  xformula   = xf,
  est_method = "aipw",
  target     = "qtt",
  probs      = seq(0.1, 0.9, 0.1),
  biters     = 100
)
summary(res_qtt)
#> 
#> Overall ATT:  
#>        ATT    Std. Error     [ 95%  Conf. Int.]  
#>  -4543.922      816.4717  -6144.177   -2943.667 *
#> 
#> 
#> QTT:
#>  Tau         QTT Std. Error [ 95% Simult.  Conf. Band]  
#>  0.1      0.0001   143.1839      -824.0687    824.0688  
#>  0.2  -1002.7420   262.6508     -2514.3807    508.8967  
#>  0.3  -3400.5673  2264.5917    -16434.0115   9632.8768  
#>  0.4  -5009.2491  1082.4184    -11238.9106   1220.4125  
#>  0.5  -4602.4652   643.5540     -8306.3228   -898.6077 *
#>  0.6  -5229.1454  1173.0966    -11980.6888   1522.3980  
#>  0.7  -5507.4720  1243.2416    -12662.7219   1647.7779  
#>  0.8  -6885.7529  1635.1317    -16296.4547   2524.9489  
#>  0.9 -10517.0625  2494.8751    -24875.8609   3841.7358  
#> ---
#> Signif. codes: `*' confidence band does not cover 0
#> Note: the outcome looks discrete (740 distinct values; 12% at 0); quantile effects may be unreliable.
autoplot(res_qtt, ylab = "QTT (earnings, 1978)")

QTT curve under unconfoundedness

The QTT only needs a counterfactual for the participants, so it is much smaller in magnitude than the QTE, but it is still negative and grows with τ, and the band excludes zero at the median. It remains well below the experimental estimates, a reminder that unconfoundedness given these covariates is a strong assumption for this comparison group.

Reading the output

summary() prints the overall ATT, the QTE/QTT at each quantile, and standard errors with confidence intervals. autoplot() returns a ggplot object so you can add layers:

autoplot(res_qtt) +
  ggplot2::labs(title = "QTT — Lalonde (observational)",
                subtitle = "AIPW with pre-treatment covariates")

By default autoplot() shows a uniform confidence band (cband = TRUE), which provides simultaneous coverage over all quantile levels — a stronger guarantee than pointwise intervals. Pass cband = FALSE for pointwise intervals instead.

Standard errors are computed via the empirical bootstrap. The biters argument controls the number of iterations (default 100). Parallel computation is available via the cl argument.

References

Dehejia, Rajeev, and Sadek Wahba. 1999. “Causal Effects in Nonexperimental Studies: Reevaluating the Evaluation of Training Programs.” Journal of the American Statistical Association 94 (448): 1053–62.
Firpo, Sergio. 2007. “Efficient Semiparametric Estimation of Quantile Treatment Effects.” Econometrica 75 (1): 259–76.
LaLonde, Robert. 1986. “Evaluating the Econometric Evaluations of Training Programs with Experimental Data.” American Economic Review, 604–20.