Interactive Fixed Effects (ife) Package

The ife package contains code to estimate treatment effects in a setup where a researcher has access to panel data (or, hopefully in the near future, repeated cross sections data) and where untreated potential outcomes are generated by an interactive fixed effects model.

The package is not currently available on CRAN, but the development version of the package can be installed from github by

# install.packages("devtools")
devtools::install_github("bcallaway11/ife")

Example

Next, we provide a brief example using the application from Callaway and Karami (2021).

res <- ife(yname="earn",
           gname="first.displaced",
           tname="year",
           idname="id",
           data=job_displacement_data,
           nife=1,
           xformla=~EDUC + race + gender,
           zformla=~EDUC + race + gender + afqt,
           ret_ife_regs=TRUE,
           anticipation=1,
           cband=FALSE,
           alp=0.10,
           boot_type="multiplier",
           biters=1000,
           cl=10)

summary(res)
#> 
#> Overall ATT:  
#>        ATT    Std. Error     [ 90%  Conf. Int.]  
#>  -3976.881      1141.811  -5854.993   -2098.769 *
#> 
#> 
#> Dynamic Effects:
#>  Event Time   Estimate Std. Error     [90%  Conf. Band]  
#>          -6   -15.0413   791.5992 -1317.106   1287.0234  
#>          -4  -131.3128   701.6462 -1285.418   1022.7925  
#>          -2  -664.5203   580.2852 -1619.005    289.9638  
#>           0 -3946.5861   986.6150 -5569.423  -2323.7488 *
#>           2 -4281.1615  1537.6912 -6810.438  -1751.8846 *
#>           4 -1454.0791  2591.7845 -5717.185   2809.0271  
#> ---
#> Signif. codes: `*' confidence band does not cover 0
ggpte(res) + ylim(c(-7000,7000))

We also have some code for running individual-specific linear trends models. These are a special case of the interactive fixed effects models that we consider in the paper, but where the factors are restricted to be equal to . We mostly argue against these sorts of models in the paper, but one advantage is that they do not require any restrictions/assumptions about finding a covariate whose effects do not change over time.

This code also implements a version of linear trends that is specific to untreated potential outcomes. Presumably, many of the same criticisms (and perhaps more actually) in recent papers about implementing DID with a two-way fixed effects regression likely apply when one includes individual-specific linear trends in the same sort of specification. The code we provide here circumvents those issues.

lt_res <- linear_trends(yname="earn",
                        gname="first.displaced",
                        tname="year",
                        idname="id",
                        data=job_displacement_data,
                        xformla=~EDUC + race + gender,
                        anticipation=1,
                        cband=FALSE,
                        alp=0.10,
                        boot_type="multiplier",
                        biters=1000,
                        cl=10)

summary(lt_res)
#> 
#> Overall ATT:  
#>        ATT    Std. Error     [ 90%  Conf. Int.]  
#>  -3463.467      1350.412  -5684.697   -1242.236 *
#> 
#> 
#> Dynamic Effects:
#>  Event Time   Estimate Std. Error     [90%  Conf. Band]  
#>          -6   -17.6954   823.6747 -1372.520   1337.1290  
#>          -4   -69.4965   702.3175 -1224.706   1085.7130  
#>          -2  -586.0360   654.9993 -1663.414    491.3420  
#>           0 -3599.9237  1205.0032 -5581.978  -1617.8697 *
#>           2 -3628.0465  1808.2979 -6602.432   -653.6612 *
#>           4   326.6318  3178.3581 -4901.302   5554.5656  
#> ---
#> Signif. codes: `*' confidence band does not cover 0
ggpte(lt_res) + ylim(c(-7000,7000))