Skip to content

Robust inference in difference-in-differences and event study designs

License

Unknown, MIT licenses found

Licenses found

Unknown
LICENSE
MIT
LICENSE.md
Notifications You must be signed in to change notification settings

asheshrambachan/HonestDiD

Repository files navigation

HonestDiD

The HonestDiD R package implements the tools for robust inference and sensitivity analysis for differences-in-differences and event study designs developed in Rambachan and Roth (2022). There is also an HonestDiD Stata package, and a Shiny app developed by Chengcheng Fang.

Background

The robust inference approach in Rambachan and Roth formalizes the intuition that pre-trends are informative about violations of parallel trends. They provide a few different ways of formalizing what this means.

Bounds on relative magnitudes. One way of formalizing this idea is to say that the violations of parallel trends in the post-treatment period cannot be much bigger than those in the pre-treatment period. This can be formalized by imposing that the post-treatment violation of parallel trends is no more than some constant larger than the maximum violation of parallel trends in the pre-treatment period. The value of = 1, for instance, imposes that the post-treatment violation of parallel trends is no longer than the worst pre-treatment violation of parallel trends (between consecutive periods). Likewise, setting = 2 implies that the post-treatment violation of parallel trends is no more than twice that in the pre-treatment period.

Smoothness restrictions. A second way of formalizing this is to say that the post-treatment violations of parallel trends cannot deviate too much from a linear extrapolation of the pre-trend. In particular, we can impose that the slope of the pre-trend can change by no more than M across consecutive periods, as shown in the figure below for an example with three periods.

diagram-smoothness-restriction

diagram-smoothness-restriction

Thus, imposing a smoothness restriction with M = 0 implies that the counterfactual difference in trends is exactly linear, whereas larger values of M allow for more non-linearity.

Other restrictions. The Rambachan and Roth framework allows for a variety of other restrictions on the differences in trends as well. For example, the smoothness restrictions and relative magnitudes ideas can be combined to impose that the non-linearity in the post-treatment period is no more than times larger than that in the pre-treatment periods. The researcher can also impose monotonicity or sign restrictions on the differences in trends as well.

Robust confidence intervals. Given restrictions of the type described above, Rambachan and Roth provide methods for creating robust confidence intervals that are guaranteed to include the true parameter at least 95% of the time when the imposed restrictions on satisfied. These confidence intervals account for the fact that there is estimation error both in the treatment effects estimates and our estimates of the pre-trends.

Sensitivity analysis. The approach described above naturally lends itself to sensitivity analysis. That is, the researcher can report confidence intervals under different assumptions about how bad the post-treatment violation of parallel trends can be (e.g., different values of or M.) They can also report the “breakdown value” of (or M) for a particular conclusion – e.g. the largest value of for which the effect is still significant.

Package installation

The package may be installed by using the function install_github() from the remotes package:

## Installation

# Install remotes package if not installed
install.packages("remotes")

# Turn off warning-error-conversion, because the tiniest warning stops installation
Sys.setenv("R_REMOTES_NO_ERRORS_FROM_WARNINGS" = "true")

# install from github
remotes::install_github("asheshrambachan/HonestDiD")

Example usage – Medicaid expansions

As an illustration of the package, we will examine the effects of Medicaid expansions on insurance coverage using publicly-available data derived from the ACS. We first load the data and packages relevant for the analysis.

#Install here, dplyr, did, haven, ggplot2, fixest packages from CRAN if not yet installed
#install.packages(c("here", "dplyr", "did", "haven", "ggplot2", "fixest"))

library(here)
library(dplyr)
library(did)
library(haven)
library(ggplot2)
library(fixest)
library(HonestDiD)

df <- read_dta("https://raw.githubusercontent.com/Mixtape-Sessions/Advanced-DID/main/Exercises/Data/ehec_data.dta")
head(df,5)
## # A tibble: 5 × 5
##   stfips      year         dins yexp2      W
##   <dbl+lbl>   <dbl+lbl>   <dbl> <dbl>  <dbl>
## 1 1 [alabama] 2008 [2008] 0.681    NA 613156
## 2 1 [alabama] 2009 [2009] 0.658    NA 613156
## 3 1 [alabama] 2010 [2010] 0.631    NA 613156
## 4 1 [alabama] 2011 [2011] 0.656    NA 613156
## 5 1 [alabama] 2012 [2012] 0.671    NA 613156

The data is a state-level panel with information on health insurance coverage and Medicaid expansion. The variable dins shows the share of low-income childless adults with health insurance in the state. The variable yexp2 gives the year that a state expanded Medicaid coverage under the Affordable Care Act, and is missing if the state never expanded.

Estimate the baseline DiD

For simplicity, we will first focus on assessing sensitivity to violations of parallel trends in a non-staggered DiD (see below regarding methods for staggered timing). We therefore restrict the sample to the years 2015 and earlier, and drop the small number of states who are first treated in 2015. We are now left with a panel dataset where some units are first treated in 2014 and the remaining units are not treated during the sample period. We can then estimate the effects of Medicaid expansion using a canonical two-way fixed effects event-study specification,

where D is 1 if a unit is first treated in 2014 and 0 otherwise.

df <- read_dta("https://raw.githubusercontent.com/Mixtape-Sessions/Advanced-DID/main/Exercises/Data/ehec_data.dta")

#Keep years before 2016. Drop the 2016 cohort
df_nonstaggered <- df %>% filter(year < 2016 &
                                 (is.na(yexp2)| yexp2 != 2015) )

#Create a treatment dummy
df_nonstaggered <- df_nonstaggered %>% mutate(D = case_when( yexp2 == 2014 ~ 1,
                                                             T ~ 0))

#Run the TWFE spec
twfe_results <- fixest::feols(dins ~ i(year, D, ref = 2013) | stfips + year,
                        cluster = "stfips",
                        data = df_nonstaggered)


betahat <- summary(twfe_results)$coefficients #save the coefficients
sigma <- summary(twfe_results)$cov.scaled #save the covariance matrix


fixest::iplot(twfe_results)

Sensitivity analysis using relative magnitudes restrictions

We are now ready to apply the HonestDiD package to do sensitivity analysis. Suppose we’re interested in assessing the sensitivity of the estimate for 2014, the first year after treatment.

delta_rm_results <-
HonestDiD::createSensitivityResults_relativeMagnitudes(
                                    betahat = betahat, #coefficients
                                    sigma = sigma, #covariance matrix
                                    numPrePeriods = 5, #num. of pre-treatment coefs
                                    numPostPeriods = 2, #num. of post-treatment coefs
                                    Mbarvec = seq(0.5,2,by=0.5) #values of Mbar
                                    )

delta_rm_results
## # A tibble: 4 × 5
##         lb     ub method Delta    Mbar
##      <dbl>  <dbl> <chr>  <chr>   <dbl>
## 1  0.0241  0.0673 C-LF   DeltaRM   0.5
## 2  0.0171  0.0720 C-LF   DeltaRM   1  
## 3  0.00859 0.0796 C-LF   DeltaRM   1.5
## 4 -0.00107 0.0883 C-LF   DeltaRM   2

The output of the previous command shows a robust confidence interval for different values of . We see that the “breakdown value” for a significant effect is = 2, meaning that the significant result is robust to allowing for violations of parallel trends up to twice as big as the max violation in the pre-treatment period.

We can also visualize the sensitivity analysis using the createSensitivityPlot_relativeMagnitudes. To do this, we first have to calculate the CI for the original OLS estimates using the constructOriginalCS command. We then pass our sensitivity analysis and the original results to the createSensitivityPlot_relativeMagnitudes command.

originalResults <- HonestDiD::constructOriginalCS(betahat = betahat,
                                                  sigma = sigma,
                                                  numPrePeriods = 5,
                                                  numPostPeriods = 2)

HonestDiD::createSensitivityPlot_relativeMagnitudes(delta_rm_results, originalResults)

Sensitivity Analysis Using Smoothness Restrictions

We can also do a sensitivity analysis based on smoothness restrictions – i.e. imposing that the slope of the difference in trends changes by no more than M between periods.

delta_sd_results <-
  HonestDiD::createSensitivityResults(betahat = betahat,
                                      sigma = sigma,
                                      numPrePeriods = 5,
                                      numPostPeriods = 2,
                                      Mvec = seq(from = 0, to = 0.05, by =0.01))

delta_sd_results
## # A tibble: 6 × 5
##         lb     ub method Delta       M
##      <dbl>  <dbl> <chr>  <chr>   <dbl>
## 1  0.0259  0.0607 FLCI   DeltaSD  0   
## 2  0.0132  0.0787 FLCI   DeltaSD  0.01
## 3  0.00286 0.0907 FLCI   DeltaSD  0.02
## 4 -0.00714 0.101  FLCI   DeltaSD  0.03
## 5 -0.0171  0.111  FLCI   DeltaSD  0.04
## 6 -0.0271  0.121  FLCI   DeltaSD  0.05
createSensitivityPlot(delta_sd_results, originalResults)

We see that the breakdown value for a significant effect is M ≈ 0.03, meaning that we can reject a null effect unless we are willing to allow for the linear extrapolation across consecutive periods to be off by more than 0.03 percentage points.

Sensitivity Analysis for Average Effects or Other Periods

So far we have focused on the effect for the first post-treatment period, which is the default in HonestDiD. If we are instead interested in the average over the two post-treatment periods, we can use the option l_vec = c(0.5,0.5). More generally, the package accommodates inference on any scalar parameter of the form θ = lvecτpos**t, where τpos**t = (τ1,…,τ)′ is the vector of dynamic treatment effects. Thus, for example, setting l_vec = basisVector(2,numPostPeriods) allows us to do inference on the effect for the second period after treatment.

delta_rm_results_avg <-
HonestDiD::createSensitivityResults_relativeMagnitudes(betahat = betahat,
                                    sigma = sigma,
                                    numPrePeriods = 5,
                                    numPostPeriods = 2, Mbarvec = seq(0,2,by=0.5),
                                    l_vec = c(0.5,0.5))

originalResults_avg <- HonestDiD::constructOriginalCS(betahat = betahat,
                                                  sigma = sigma,
                                                  numPrePeriods = 5,
                                                  numPostPeriods = 2,
                                                  l_vec = c(0.5,0.5))

HonestDiD::createSensitivityPlot_relativeMagnitudes(delta_rm_results_avg, originalResults_avg)

Regressions with controls

The parameters betahat and sigma should be the coefficients and variance-covariance matrix for the event-study plot only. Sometimes we may have regression results that contain both the event-study coefficients and coefficients on other auxilliary variables (e.g. controls). In this case, we should subset betahat and sigma to the relevant coefficients corresponding to the event-study. An example is given below:

# CREATE A FAKE CONTROL FOR ILLUSTRATION
set.seed(0)
df_nonstaggered$control <- rnorm(NROW(df_nonstaggered), 0, 1)

#Run the TWFE spec with the control added
# (Note that TWFEs with controls may not yield ATT under het effects; see Abadie 2005)
twfe_results <- fixest::feols(dins ~ i(year, D, ref = 2013) + control | stfips + year,
                              cluster = "stfips",
                              data = df_nonstaggered)


betahat <- summary(twfe_results)$coefficients #save the coefficients
sigma <- summary(twfe_results)$cov.scaled #save the covariance matrix

#Subset the coefficients to exclude the control, which here is in position 8
betahat <- betahat[-8]
sigma <-sigma[-8,-8]

HonestDiD::createSensitivityResults(betahat = betahat,
                                    sigma = sigma,
                                    numPrePeriods = 5,
                                    numPostPeriods = 2,
                                    Mvec = seq(from = 0, to = 0.05, by =0.01))
## # A tibble: 6 × 5
##         lb     ub method Delta       M
##      <dbl>  <dbl> <chr>  <chr>   <dbl>
## 1  0.0264  0.0607 FLCI   DeltaSD  0   
## 2  0.0128  0.0788 FLCI   DeltaSD  0.01
## 3  0.00186 0.0908 FLCI   DeltaSD  0.02
## 4 -0.00814 0.101  FLCI   DeltaSD  0.03
## 5 -0.0181  0.111  FLCI   DeltaSD  0.04
## 6 -0.0281  0.121  FLCI   DeltaSD  0.05

Staggered timing

So far we have focused on a simple case without staggered timing. Fortunately, the HonestDiD approach works well with recently-introduced methods for DiD under staggered treatment timing. Below, we show how the package can be used with the fixest package implementing Sun and Abraham or with the did package implementing Callaway and Sant’Anna. (See, also, the example on the did package website.) We are hoping to more formally integrate the did and HonestDiD packages in the future – stay tuned!

Sun and Abraham

First, we import the function we created for extracting the full-variance covariance matrix from fixest with sunab:

#' @description
#' This function takes a regression estimated using fixest with the sunab option
#' and extracts the aggregated event-study coefficients and their variance-covariance matrix
#' @param sunab_fixest The result of a fixest call using the sunab option
#' @returns A list containing beta (the event-study coefficients),
#'          sigma (the variance-covariance matrix), and
#'          cohorts (the relative times corresponding to beta, sigma)

sunab_beta_vcv <-
function(sunab_fixest){

  ## The following code block extracts the weights on individual coefs used in
  # the fixest aggregation ##
  sunab_agg   <- sunab_fixest$model_matrix_info$sunab$agg_period
  sunab_names <- base::names(sunab_fixest$coefficients)
  sunab_sel   <- base::grepl(sunab_agg, sunab_names, perl=TRUE)
  sunab_names <- sunab_names[sunab_sel]
  if(!base::is.null(sunab_fixest$weights)){
    sunab_wgt <- base::colSums(sunab_fixest$weights * base::sign(stats::model.matrix(sunab_fixest)[, sunab_names, drop=FALSE]))
  } else {
    sunab_wgt <- base::colSums(base::sign(stats::model.matrix(sunab_fixest)[, sunab_names, drop=FALSE]))
  }

  #Construct matrix sunab_trans such that sunab_trans %*% non-aggregated coefs = aggregated coefs,
  sunab_cohorts <- base::as.numeric(base::gsub(base::paste0(".*", sunab_agg, ".*"), "\\2", sunab_names, perl=TRUE))
  sunab_mat     <- stats::model.matrix(~ 0 + base::factor(sunab_cohorts))
  sunab_trans   <- base::solve(base::t(sunab_mat) %*% (sunab_wgt * sunab_mat)) %*% base::t(sunab_wgt * sunab_mat)

  #Get the coefs and vcv
  sunab_coefs   <- sunab_trans %*% base::cbind(sunab_fixest$coefficients[sunab_sel])
  sunab_vcov    <- sunab_trans %*% sunab_fixest$cov.scaled[sunab_sel, sunab_sel] %*% base::t(sunab_trans)

  base::return(base::list(beta    = sunab_coefs,
                          sigma   = sunab_vcov,
                          cohorts = base::sort(base::unique(sunab_cohorts))))
}
# Run fixest with sunab
df$year_treated <- ifelse(is.na(df$yexp2), Inf, df$yexp2)
formula_sunab <- dins ~ sunab(year_treated, year) | stfips + year
res_sunab <- fixest::feols(formula_sunab, cluster="stfips", data=df)
fixest::iplot(res_sunab)

# Extract the beta and vcv
beta_vcv <- sunab_beta_vcv(res_sunab)

# Run sensitivity analysis for relative magnitudes
kwargs <- list(betahat        = beta_vcv$beta,
               sigma          = beta_vcv$sigma,
               numPrePeriods  = sum(beta_vcv$cohorts < 0),
               numPostPeriods = sum(beta_vcv$cohorts > -1))
extra <- list(Mbarvec=seq(from = 0.5, to = 2, by = 0.5), gridPoints=100)

original_results <-
  do.call(HonestDiD::constructOriginalCS, kwargs)

sensitivity_results <-
  do.call(HonestDiD::createSensitivityResults_relativeMagnitudes,
          c(kwargs, extra))

HonestDiD::createSensitivityPlot_relativeMagnitudes(sensitivity_results,
                                                    original_results)

Callaway and Sant’Anna

First, we import the function Pedro Sant’Anna created for formatting did output for HonestDiD:

#' @title honest_did
#'
#' @description a function to compute a sensitivity analysis
#'  using the approach of Rambachan and Roth (2021)
#'
#' @param ... Parameters to pass to the relevant method.
honest_did <- function(...) UseMethod("honest_did")

#' @title honest_did.AGGTEobj
#'
#' @description a function to compute a sensitivity analysis
#'  using the approach of Rambachan and Roth (2021) when
#'  the event study is estimating using the `did` package
#'
#' @param es Result from aggte (object of class AGGTEobj).
#' @param e event time to compute the sensitivity analysis for.
#'  The default value is `e=0` corresponding to the "on impact"
#'  effect of participating in the treatment.
#' @param type Options are "smoothness" (which conducts a
#'  sensitivity analysis allowing for violations of linear trends
#'  in pre-treatment periods) or "relative_magnitude" (which
#'  conducts a sensitivity analysis based on the relative magnitudes
#'  of deviations from parallel trends in pre-treatment periods).
#' @param gridPoints Number of grid points used for the underlying test
#'  inversion. Default equals 100. User may wish to change the number of grid
#'  points for computational reasons.
#' @param ... Parameters to pass to `createSensitivityResults` or
#'  `createSensitivityResults_relativeMagnitudes`.
honest_did.AGGTEobj <- function(es,
                                e          = 0,
                                type       = c("smoothness", "relative_magnitude"),
                                gridPoints = 100,
                                ...) {

  type <- match.arg(type)

  # Make sure that user is passing in an event study
  if (es$type != "dynamic") {
    stop("need to pass in an event study")
  }

  # Check if used universal base period and warn otherwise
  if (es$DIDparams$base_period != "universal") {
    stop("Use a universal base period for honest_did")
  }

  # Recover influence function for event study estimates
  es_inf_func <- es$inf.function$dynamic.inf.func.e

  # Recover variance-covariance matrix
  n <- nrow(es_inf_func)
  V <- t(es_inf_func) %*% es_inf_func / n / n

  # Check time vector is consecutive with referencePeriod = -1
  referencePeriod <- -1
  consecutivePre  <- !all(diff(es$egt[es$egt <= referencePeriod]) == 1)
  consecutivePost <- !all(diff(es$egt[es$egt >= referencePeriod]) == 1)
  if ( consecutivePre | consecutivePost ) {
    msg <- "honest_did expects a time vector with consecutive time periods;"
    msg <- paste(msg, "please re-code your event study and interpret the results accordingly.", sep="\n")
    stop(msg)
  }

  # Remove the coefficient normalized to zero
  hasReference <- any(es$egt == referencePeriod)
  if ( hasReference ) {
    referencePeriodIndex <- which(es$egt == referencePeriod)
    V    <- V[-referencePeriodIndex,-referencePeriodIndex]
    beta <- es$att.egt[-referencePeriodIndex]
  } else {
    beta <- es$att.egt
  }

  nperiods <- nrow(V)
  npre     <- sum(1*(es$egt < referencePeriod))
  npost    <- nperiods - npre
  if ( !hasReference & (min(c(npost, npre)) <= 0) ) {
    if ( npost <= 0 ) {
      msg <- "not enough post-periods"
    } else {
      msg <- "not enough pre-periods"
    }
    msg <- paste0(msg, " (check your time vector; note honest_did takes -1 as the reference period)")
    stop(msg)
  }

  baseVec1 <- basisVector(index=(e+1),size=npost)
  orig_ci  <- constructOriginalCS(betahat        = beta,
                                  sigma          = V,
                                  numPrePeriods  = npre,
                                  numPostPeriods = npost,
                                  l_vec          = baseVec1)

  if (type=="relative_magnitude") {
    robust_ci <- createSensitivityResults_relativeMagnitudes(betahat        = beta,
                                                             sigma          = V,
                                                             numPrePeriods  = npre,
                                                             numPostPeriods = npost,
                                                             l_vec          = baseVec1,
                                                             gridPoints     = gridPoints,
                                                             ...)

  } else if (type == "smoothness") {
    robust_ci <- createSensitivityResults(betahat        = beta,
                                          sigma          = V,
                                          numPrePeriods  = npre,
                                          numPostPeriods = npost,
                                          l_vec          = baseVec1,
                                          ...)
  }

  return(list(robust_ci=robust_ci, orig_ci=orig_ci, type=type))
}
###
# Run the CS event-study with 'universal' base-period option
## Note that universal base period normalizes the event-time minus 1 coef to 0
cs_results <- did::att_gt(yname = "dins",
                          tname = "year",
                          idname = "stfips",
                          gname = "yexp2",
                          data = df %>% mutate(yexp2 = ifelse(is.na(yexp2), 3000, yexp2)),
                          control_group = "notyettreated",
                          base_period = "universal")

es <- did::aggte(cs_results, type = "dynamic",
                 min_e = -5, max_e = 5)

#Run sensitivity analysis for relative magnitudes
sensitivity_results <-
  honest_did(es,
             e=0,
             type="relative_magnitude",
             Mbarvec=seq(from = 0.5, to = 2, by = 0.5))

HonestDiD::createSensitivityPlot_relativeMagnitudes(sensitivity_results$robust_ci,
                                                    sensitivity_results$orig_ci)

Additional options and resources

See the previous package vignette for additional examples and package options, including incorporating sign and monotonicity restrictions, and combining relative magnitudes and smoothness restrictions.

You can also view a video presentation about this paper here.

Authors

Acknowledgements

This software package is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant DGE1745303 (Rambachan) and Grant DGE1144152 (Roth). We thank Mauricio Cáceres Bravo for his help in developing the package.

About

Robust inference in difference-in-differences and event study designs

Topics

Resources

License

Unknown, MIT licenses found

Licenses found

Unknown
LICENSE
MIT
LICENSE.md

Stars

Watchers

Forks

Packages

No packages published

Contributors 4

  •  
  •  
  •  
  •  

Languages