smskim / ivreghdfe

Run IV/2SLS with many levels of fixed effects (i.e. ivreg2+reghdfe)

IVREGHDFE: reghdfe + ivreg2 (adds instrumental variable and additional robust SE estimators to reghdfe)

• Jump to: usage benchmarks install

This package integrates reghdfe into ivreg2, through an absorb() option. This allows IV/2SLS regressions with multiple levels of fixed effects.

Recent updates

• version 1.1.2 29Sep2022:
  • Fix bug #44; where cluster(...) worked but vce(cluster ...) was silently ignored
  • Fix bug #46; small correction to e(cmdline)
• version 1.1.1 14dec2021:
  • Add experimental support for margins postestimation command.
• version 1.1 26feb2021:
  • Update ivreg2 dependency from 4.1.10 9Feb2016 to 4.1.11 22Nov2019.
  • Update reghdfe dependency from 5.9.0 03jun2020 to 6.0.2 25feb2021
  • Before, reghdfe options had to be passed as suboptions of absorb(). Now they are passed directly as normal options
  • Note that some options are slightly different in reghdfe v6 (e.g. the exact technique used is set through the technique() option, following Stata convention).
  • Note that there might be a tiny difference in the SE estimates of ivreghdfe wrt those in reghdfe when both are used to run OLS instead of IV.
    • This happens if we have clustered standard errors, and the fixed effects are nested within the clusters.
    • Then, when computing the small sample adjustment q, reghdfe divides by (N-K-1) while ivreg2 (and thus ivreghdfe) divides by (N-K)
    • reghdfe does so to keep consistency with the small sample adjustment done by xtreg
    • For more details see comment in code ("minor adj. so we match xtreg when the absvar is nested within cluster")

Comparison with other commands

As seen in the table below, ivreghdfe is recommended if you want to run IV/LIML/GMM2S regressions with fixed effects, or run OLS regressions with advanced standard errors (HAC, Kiefer, etc.)

Command	regress	areg	reghdfe	ivreg2	ivreghdfe
Models:	OLS	OLS	OLS	OLS, IV, LIML, GMM2S, CUE	OLS, IV, LIML, GMM2S (not CUE!)
Fixed effects?	-	One-way	Multi-way	-	Multi-way
Cluster SE?	One-way	One-way	Multi-way	Two-way	Two-way
Additional SEs:	-	-	-	AC, HAC, Kiefer, Driscol-Kraay, etc.	AC, HAC, Kiefer, Driscol-Kraay, etc.
(Speed) Time without FEs:	1x	-	2x	3.7x	4.3x
(Speed) Time with one FE:	-	6.3x	2.1x	-	4.6x

(Benchmark run on Stata 14-MP (4 cores), with a dataset of 4 regressors, 10mm obs., 100 clusters and 10,000 FEs)

Installation

ivreghdfe requires three packages: ivreg2, reghdfe (version 5.x) and ftools. Run the lines below to install everything you might possibly need:

* Install ftools (remove program if it existed previously)
cap ado uninstall ftools
net install ftools, from("https://raw.githubusercontent.com/sergiocorreia/ftools/master/src/")

* Install reghdfe
cap ado uninstall reghdfe
net install reghdfe, from("https://raw.githubusercontent.com/sergiocorreia/reghdfe/master/src/")

* Install ivreg2, the core package
cap ado uninstall ivreg2
ssc install ivreg2

* Finally, install this package
cap ado uninstall ivreghdfe
net install ivreghdfe, from(https://raw.githubusercontent.com/sergiocorreia/ivreghdfe/master/src/)

Advice

This code just modifies ivreg2 adding an absorb() option that uses reghdfes Mata functions (see this link for the line-by-line differences). When used, absorb() will also activate the small, noconstant and nopartialsmall options of ivreg2 (basically to force small sample adjustments, which are required as we might have a substantial number of fixed effects).

You can also use all other reghdfe options as normal options of ivreghdfe (e.g. tolerance, choice of transform, etc.):

sysuse auto, clear
ivreghdfe price weight (length=gear), absorb(turn trunk) tol(1e-6) accel(sd)

This is gives the same result as using the old version of reghdfe (but slower):

reghdfe price weight (length=gear), absorb(turn trunk) tol(1e-6) accel(sd) version(3)

Residuals

To save residuals, do this:

sysuse auto
ivreghdfe price weight, absorb(trunk) resid(myresidname)

You can also use the other predict options of reghdfe, such as d:

predict d, d