Package {drlate}

drlate: Doubly Robust Estimation of Local Average Treatment Effects

Description

Estimates the local average treatment effect (LATE) and the local average treatment effect on the treated (LATT) using observational data with a binary instrument, implementing the complete estimator suite of Sloczynski, Uysal, and Wooldridge: the doubly robust estimators of Sloczynski, Uysal, and Wooldridge (2022) doi:10.48550/arXiv.2208.01300 – inverse probability weighted regression adjustment (IPWRA), inverse probability weighting (IPW), augmented inverse probability weighting (AIPW), and regression adjustment (RA) – and the Abadie-kappa weighting estimators of Sloczynski, Uysal, and Wooldridge (2025) doi:10.1080/07350015.2024.2332763. Supports linear, logistic, probit, Poisson, and fractional (fractional-logit and fractional-probit) outcome and treatment models, and instrument propensity scores estimated by maximum likelihood, covariate balancing (CBPS), or inverse probability tilting (IPT). Standard errors are computed jointly for all estimation stages by stacking the moment conditions of every model into a single M-estimation system; weak-instrument-robust Fieller confidence sets, cluster-aware bootstrap inference, design diagnostics, and a doubly robust Hausman-type test of unconfoundedness are included. Estimates and standard errors are validated against the authors' Stata commands 'drlate' (Statistical Software Components S459708) and 'kappalate' (S459257).

Author(s)

Maintainer: Kailas Venkitasubramanian kailasv@gmail.com

Other contributors:

• S. Derya Uysal (Author of the original Stata package 'drlate') [contributor, copyright holder]

• Tymon Sloczynski (Author of the original Stata package 'drlate') [contributor, copyright holder]

• Jeffrey M. Wooldridge (Author of the original Stata package 'drlate') [contributor, copyright holder]

See Also

Useful links:

• https://github.com/kvenkita/drlate

• https://kvenkita.github.io/drlate/

• Report bugs at https://github.com/kvenkita/drlate/issues

Covariate balance across instrument arms

Description

Computes standardized mean differences (SMDs) of the model covariates between the two instrument arms, before and after weighting by the inverse of the estimated instrument propensity score. Well-balanced weighted covariates (conventionally, absolute SMD below 0.1) indicate that the propensity score model is doing its job.

Usage

balance(object, ...)

## S3 method for class 'drlate'
balance(object, detail = FALSE, ...)

Arguments

Details

The covariate set is the union of the columns of the instrument, outcome, and treatment model matrices (the intercept is dropped). The SMD denominator is the unweighted pooled standard deviation \sqrt{(s_1^2 + s_0^2)/2} in both columns, so the two columns are directly comparable. Weighted arm means are Hájek means using the inverse-propensity weights implied by the fit (for estimand = "latt", the Z=0 arm uses the ATT odds weights p/(1-p), matching the estimator).

Value

A data frame with one row per covariate and columns variable, smd_unweighted, and smd_weighted; with detail = TRUE, the four additional columns described above.

See Also

plot.drlate() with type = "balance" for the love plot.

Imai-Ratkovic covariate-balance test

Description

Tests whether the estimated instrument propensity score balances the covariates, using the overidentification test of Imai and Ratkovic (2014). The propensity-score MLE score equations identify the coefficients; the covariate-balancing (CBPS) moments are the overidentifying restrictions. A large statistic is evidence that the propensity-score model does not balance the covariates — a misspecification diagnostic. This is the Stata latebalance overid postestimation feature.

Usage

balance_test(object)

Arguments

Value

An object of class drlate_balance_test: a list with statistic (Hansen's J), df, p.value, ivmodel, and n, with a print method.

References

Imai, K. and Ratkovic, M. (2014). Covariate Balancing Propensity Score. Journal of the Royal Statistical Society B 76(1), 243–263.

See Also

balance() for the standardized-mean-difference diagnostics.

Examples

fit <- drlate(lwage ~ age + educ, nvstat ~ age + educ,
              rsncode ~ age + educ, data = drlate_sim)
balance_test(fit)

Complier covariate means

Description

Compares the average of each covariate in the full estimation sample with its average in the complier subpopulation, the latter computed with the normalized Abadie kappa weights of kappa_weights(). Because the local average treatment effect is a causal effect for compliers, knowing how compliers differ from the population aids interpretation. This is the Stata estat compliers postestimation feature.

Usage

complier_means(object, vars = NULL)

Arguments

Details

Covariate values are reported on their original scale.

Value

A data frame with one row per covariate and columns variable, population_mean, complier_mean, and difference (complier_mean - population_mean).

See Also

kappa_weights()

Examples

fit <- drlate(lwage ~ age + educ, nvstat ~ age + educ,
              rsncode ~ age + educ, data = drlate_sim)
complier_means(fit)

Confidence intervals for drlate fits

Description

Confidence intervals for drlate fits

Usage

## S3 method for class 'drlate'
confint(object, parm, level = 0.95, method = c("default", "fieller"), ...)

Arguments

Value

For method = "default", a numeric matrix with one row per requested coefficient (parm) and two columns holding the lower and upper confidence limits. The columns are labelled with the corresponding percentiles (for the default 95% level, "2.5 %" and "97.5 %"). The limits are Wald intervals from the joint sandwich covariance, or percentile intervals from the resampling draws when the fit was computed with vcov = "bootstrap".

For method = "fieller", an object of class "drlate_fieller": a list describing the weak-instrument-robust confidence set for the LATE/LATT ratio (its endpoints and shape, the estimand name, and the confidence level), with its own print method. Because a Fieller set need not be a bounded interval, it is returned in this form rather than as a matrix of endpoints.

Doubly robust Hausman test of unconfoundedness

Description

Tests whether the treatment is unconfounded given the covariates, using the comparison proposed by Słoczyński, Uysal, and Wooldridge (2022, Section 5), building on Donald, Hsu, and Lieli (2014). Under one-sided noncompliance (nobody takes the treatment without the instrument: \Pr(D = 1 \mid Z = 0) = 0), the LATT identified through the instrument equals the ATT identified through unconfoundedness of the treatment — so a significant difference between the doubly robust LATT estimate (which uses the instrument) and the doubly robust ATT estimate (which does not) is evidence against unconfoundedness. Unlike the textbook OLS-vs-IV Hausman test, this comparison is robust to treatment effect heterogeneity.

Usage

dr_hausman(
  outcome,
  treatment,
  instrument,
  data,
  omodel = c("linear", "logit", "poisson"),
  tmodel = c("logit", "linear", "poisson"),
  ivmodel = c("logit", "ipt"),
  weights = NULL,
  cluster = NULL,
  pstolerance = 1e-05,
  subset = NULL
)

Arguments

Details

The DR ATT estimator follows the paper's equation (33): a treatment propensity score \Pr(D = 1 \mid X) is fitted by logit QMLE on the treatment-equation covariates; the outcome model is fitted on the untreated sample weighted by the odds \hat p/(1-\hat p); and \hat\tau_{ATT} is the treated-sample mean outcome minus the mean imputed counterfactual. The standard error of the difference comes from stacking the moment conditions of both estimators (and the difference) into one M-estimation system, so the covariance between them is accounted for analytically — the analytic option suggested in the paper.

Note that the two halves adjust on their respective formulas: the LATT half's propensity score uses the instrument-equation covariates, while the ATT half's uses the treatment-equation covariates (both share the outcome model). Supply the same covariate set to all three formulas unless you intend them to differ.

Value

An object of class "htest" with the z statistic, p-value, and the DR LATT, DR ATT, and difference estimates.

References

Słoczyński, T., S. D. Uysal, and J. M. Wooldridge (2022). "Doubly Robust Estimation of Local Average Treatment Effects Using Inverse Probability Weighted Regression Adjustment." doi:10.48550/arXiv.2208.01300

Donald, S. G., Y.-C. Hsu, and R. P. Lieli (2014). "Testing the Unconfoundedness Assumption via Inverse Probability Weighted Estimators of (L)ATT." Journal of Business & Economic Statistics 32(3), 395-415.

Examples

d <- drlate_sim
d$nvstat[d$rsncode == 0] <- 0L   # impose one-sided noncompliance
dr_hausman(lwage ~ age + educ, nvstat ~ age + educ,
           rsncode ~ age + educ, data = d)

Doubly robust estimation of the LATE and LATT

Description

Estimates the local average treatment effect (LATE) or the local average treatment effect on the treated (LATT) with a binary instrument, following Słoczyński, Uysal, and Wooldridge (2022). A faithful R port of the Stata package drlate (SSC S459708): point estimates come from sequential weighted regressions, and standard errors are computed jointly for the instrument propensity score, the outcome regression, the treatment regression, and the causal estimand by stacking all moment conditions into a single M-estimation system.

Usage

drlate(
  outcome,
  treatment,
  instrument,
  data,
  omodel = c("linear", "logit", "probit", "poisson", "flogit", "fprobit"),
  tmodel = c("logit", "probit", "linear", "poisson"),
  ivmodel = c("logit", "cbps", "ipt", "probit"),
  method = c("ipwra", "ipw", "aipw", "ra", "kappa", "kappa0", "kappa10"),
  estimand = c("late", "latt"),
  normalized = TRUE,
  weights = NULL,
  cluster = NULL,
  pstolerance = 1e-05,
  osample = FALSE,
  subset = NULL,
  keep_data = TRUE,
  vcov = c("analytic", "bootstrap"),
  boot_reps = 999L,
  boot_seed = NULL,
  cores = 1L
)

Arguments

Value

An object of class "drlate", a list with components including coefficients (the causal estimate, the numerator effect of Z on Y, and the denominator effect of Z on D), vcov3 (their variance matrix, diagonal by construction, as in the Stata package), vcov_full (the joint variance matrix of all stacked parameters), theta (all stacked parameter estimates), N, dmeanz1, dmeanz0, and the call. For method = "kappa10" only the causal estimate is reported (the estimator is a difference of two ratios, so no single numerator/denominator pair exists). For "kappa" and "kappa0" the third coefficient is the mean of the corresponding kappa weight: under the LATE assumptions it estimates the same complier share as the IPW first-stage contrast (the population ATE of Z on D), but it is a different sample statistic and the two can diverge under propensity score misspecification.

References

Słoczyński, T., S. D. Uysal, and J. M. Wooldridge (2022). "Doubly Robust Estimation of Local Average Treatment Effects Using Inverse Probability Weighted Regression Adjustment." doi:10.48550/arXiv.2208.01300

Słoczyński, T., S. D. Uysal, and J. M. Wooldridge (2025). "Abadie's Kappa and Weighting Estimators of the Local Average Treatment Effect." Journal of Business & Economic Statistics 43(1), 164–177. doi:10.1080/07350015.2024.2332763

Examples

data(drlate_sim)
fit <- drlate(lwage ~ age + educ, nvstat ~ age + educ,
              rsncode ~ age + educ, data = drlate_sim)
summary(fit)

Compare drlate estimators in one call

Description

Runs several estimators on the same specification and collects the causal estimates with their confidence intervals — the sensitivity comparison applied papers routinely report. Formula restrictions are handled automatically: method = "ipw" drops the outcome/treatment covariates and method = "ra" drops the instrument covariates (each with a message), matching the requirements of those estimators.

Usage

drlate_compare(
  outcome,
  treatment,
  instrument,
  data,
  methods = c("ipwra", "ipw", "aipw", "ra"),
  both_norms = FALSE,
  ...
)

Arguments

Details

Because IPW carries no outcome/treatment regressions and RA carries no instrument propensity score, the automatic formula adjustment means the rows do not share a single adjustment specification: differences between the IPW or RA row and the doubly robust rows reflect both the estimator and the reduced specification. Read the comparison as a robustness display, not as a test that isolates estimator choice; the doubly robust rows (IPWRA, AIPW) are the like-for-like pair.

Value

An object of class "drlate_compare": a data frame with columns method, normalized, estimate, se, ci_lo, ci_hi, with a print method and a dot-whisker plot method.

Examples

cmp <- drlate_compare(lwage ~ age + educ, nvstat ~ age + educ,
                      rsncode ~ age + educ, data = drlate_sim)
cmp

Simulated example data for drlate

Description

A simulated dataset with a binary instrument, a binary treatment with two-sided noncompliance, and continuous, positive, and binary outcome variables, designed to exercise every model family supported by drlate(). The complier average treatment effect (LATE) used in the data-generating process is 0.5. The treatment is genuinely endogenous (compliance type shifts the baseline outcome, so naive OLS is biased upward) and the instrument is only conditionally valid (its propensity depends on age and educ, so the raw Wald ratio is biased too).

Usage

drlate_sim

Format

A data frame with 2,000 rows and 7 variables:

lwage

continuous outcome

kwage

positive outcome (for Poisson models), exp(lwage / 2)

hijob

binary outcome (for logit models)

nvstat

binary treatment

rsncode

binary instrument

age

continuous covariate

educ

factor covariate with levels hs, college, graduate

Source

Simulated; see data-raw/drlate_sim.R in the package sources.

Abadie's kappa weights

Description

Returns the per-observation Abadie kappa weight implied by a fitted drlate() object,

\kappa = 1 - \frac{D(1 - Z)}{1 - p(X)} - \frac{(1 - D) Z}{p(X)},

where p(X) is the estimated instrument propensity score. The kappa weights identify the complier subpopulation: for any function g of the data, E[g \mid \mathrm{complier}] = E[\kappa g] / E[\kappa] (Abadie 2003). They are the weights used by complier_means() and are the Stata estat compliers, genkappa() object.

Usage

kappa_weights(object, normalize = TRUE)

Arguments

Value

A numeric vector with one entry per estimation-sample observation.

See Also

complier_means()

Examples

fit <- drlate(lwage ~ age + educ, nvstat ~ age + educ,
              rsncode ~ age + educ, data = drlate_sim)
head(kappa_weights(fit))

Diagnostic plots for drlate fits

Description

Diagnostic plots for drlate fits

Usage

## S3 method for class 'drlate'
plot(
  x,
  type = c("overlap", "balance", "balance_density", "weights"),
  bins = 30,
  geom = c("histogram", "density"),
  var = NULL,
  ...
)

Arguments

Value

A ggplot object.