Skip to contents

Honest sensitivity analysis for parallel-trends violations

Source: R/honest_did.R
honest_sensitivity.Rd

Robust inference and sensitivity analysis for event-study / DiD designs following Rambachan and Roth (2023). Instead of assuming parallel trends holds exactly, it asks how large a violation of parallel trends would have to be before the causal conclusion changes, returning confidence sets for a post-treatment effect under a sequence of restrictions on the possible difference in trends.

Two restriction families are supported:

  • "relative_magnitude" (\(\Delta^{RM}(\bar M)\)): post-treatment violations are at most \(\bar M\) times the largest pre-treatment violation (Rambachan & Roth 2023, Section 2.4.1).

  • "smoothness" (\(\Delta^{SD}(M)\)): the difference in trends deviates from linearity by at most \(M\) per period (Section 2.4.3).

Inference uses the Andrews-Roth-Pakes (ARP) conditional moment-inequality test (Section 3.2.1), which is uniformly valid and recommended for general restriction sets.

Usage

honest_sensitivity(
  object = NULL,
  type = c("relative_magnitude", "smoothness"),
  Mvec = NULL,
  l_vec = NULL,
  alpha = 0.05,
  gridPoints = 1000L,
  betahat = NULL,
  sigma = NULL,
  numPrePeriods = NULL,
  numPostPeriods = NULL
)

Arguments

object

An es_result from run_es() (with estimator = "twfe", i.e. classic or method = "sunab", which carry the event-study coefficient covariance). Alternatively, supply betahat and sigma directly (see below) for any estimator.

type

Restriction family: "relative_magnitude" (default) or "smoothness".

Mvec

Numeric vector of restriction parameters. For "relative_magnitude" these are \(\bar M\) values (default seq(0, 2, by = 0.5)); for "smoothness" these are \(M\) values (default a data-driven sequence from 0 to the largest pre-period SD).

l_vec

Numeric weight vector over post-treatment periods defining the target \(\theta = l'\tau_{post}\). Defaults to the first post-treatment period.

alpha

Significance level (default 0.05 for 95% confidence sets).

gridPoints

Number of grid points for test inversion (default 1000).

betahat

Optional event-study coefficient vector (pre then post, excluding the reference period), ordered by relative time. Required when object does not carry an es_vcov attribute.

sigma

Optional covariance matrix of betahat.

numPrePeriods, numPostPeriods

Optional integer counts; inferred from object or from betahat/l_vec when omitted.

Value

A data.frame of class "honest_result" with one row per restriction value plus the original (parallel-trends) confidence interval, with columns M, lb, ub, method, and type. The breakdown value (largest restriction at which the robust CI still excludes 0) is stored in attr(., "breakdown").

References

Rambachan, A. and Roth, J. (2023). A More Credible Approach to Parallel Trends. Review of Economic Studies, 90(5), 2555-2591.

See also

run_es(), plot_honest()