arlionn / sumhdfe

Summary and diagnostic information for evaluating within-fixed-effect variation.

Sumhdfe is a Stata package that produces summary and diagnostic information of linear fixed effect models. It shows:

• The frequency of fixed effects
• How many groups (e.g., firms) have no variation within fixed effects
• The residual within-fixed-effect variation of the regression variables

It is currently in beta version, so all comments and suggestions are welcome.

For a discussion of within-fixed-effect variation, and the underlying issues that sumhdfe addresses, see deHaan (2021). Similarly, if you find these diagnostics to be useful, please cite:

**deHaan, Ed. (2021). Using and Interpreting Fixed Effects Models. ** Available at SSRN: https://ssrn.com/abstract=3699777.

Authors

• Sergio Correia
• Ed deHaan
• Ties de Kok

Table of contents

• Installation
• Usage & Features
• Pending Items
• Changelog
• Questions?

Installing sumhdfe

Sumhdfe requires the latest development versions of reghdfe and ftools to be installed prior to installation.

To install these packages and sumhdfe, follow the steps below:

cap ado uninstall ftools
cap ado uninstall reghdfe
cap ado uninstall sumhdfe

net install ftools, from("https://raw.githubusercontent.com/sergiocorreia/ftools/groupreg/src/")
net install reghdfe, from("https://raw.githubusercontent.com/sergiocorreia/reghdfe/reghdfe6/src/")
net install sumhdfe, from("https://raw.githubusercontent.com/ed-dehaan/sumhdfe/master/src/")

Usage & Features

Example usage

Sumhdfe can be used in one of two ways:

1. As a postestimation command following reghdfe
2. As a standalone command

Post-estimation version

First run reghdfe and then run sumhdfe. A simple example is show below, see the Stata help file for additional examples.

use "https://raw.githubusercontent.com/ed-dehaan/sumhdfe/master/sumhdfe_demo_data.dta", clear
reghdfe y x1 x2  , a(firm year) 
sumhdfe

Standalone version

Run sumhdfe directly.

use "https://raw.githubusercontent.com/ed-dehaan/sumhdfe/master/sumhdfe_demo_data.dta", clear
sumhdfe y x1 x2  , a(firm year)

Default output

The sumhdfe command will provide four panels by default.

Panel A - summary statistics - reghdfe

Panel A provides summary statistics for the sample used in reghdfe.

Example:

Notes:

• It can be customized similar to estat summarize
• N includes singletons, so it differs from N shown in the reghdfe output

Panel B - summary statistics - fixed effects

Panel B provides summary statistics for the fixed effects themselves.

Example:

Notes:

• Interpretation of the above example:
  • There are 189 unique firms within the firm fixed effects, 28 of which are singletons (i.e., appear just once). An individual firm has between 1 and 8 observations.
  • There are 39 unique years within the year fixed effects, 8 of which are singletons.
  • Iterating across both firm and year eliminates 2 more "joint singletons," for a total of 38 singletons eliminated from the reghdfe output.

Panel C - Groups without any within-fixed-effect variation

Panel C quantifies how often each variable is constant within a given fixed effect group (such as within a given firm). These observations can have unexpected effects on regression coefficients and, if numerous, should be carefully evaluated.

Example:

Notes:

• Interpretation of the above example:
  • Variable x1 has (623-38=) 585 observations excluding singletons.
  • Within the non-singleton data, 58 firms have no variation in x1; i.e., each firm has the same x1 in all years. Those 58 firms relate to 217 observations.
  • X1 is constant within 4 years, relating to 28 observations.

Panel D - Variation lost (absorbed) due to fixed effects

Panel D shows how much variation in each variable is lost (or absorbed) due to the fixed effects, in terms of both standard deviations and r-squared.

Example:

Notes:

• Interpretation of the above example:
  • The standard deviation of x1 is 79.7 in the pooled sample (as also showed in Panel A), but the within-fixed-effect standard deviation of x1 is 22.7. Thus, the within-fixed effect variation of x1 is roughly 28.4% of the pooled sample.
  • In terms of r-squared, the firm fixed effects explain roughly 87% of the variation in x1 while the year fixed effects explain roughly 13%. Combined, the fixed effects explain 92.4% of the variation in x1.
    • Technical note: the r-squared is relative to the sample including singletons, for which the r-squared is mechanically equal to 100%.

Optional outputs

Histogram

The histogram(#) option tabulates the frequencies of observations within a fixed effect grouping.

Example:

For example, sumhdfe, histogram(1) shows the frequencies of observations for the first fixed effect grouping listed within a(firm year), which in this case if firm. You can also specify the fixed effect name; for example sumhdfe, histogram(year).

Additional options

For additional examples and additional options, see the stata help file with help sumhdfe, or its online version.

Pending Items

1. Allow for easy export of each table to csv/excel/tex
2. Tutorial/documentation with real-world example
3. Add an option to visually compare the pooled- and within-fixed-effect variation in a variable. In the meantime, it can be manually done as follows:

use "https://raw.githubusercontent.com/ed-dehaan/sumhdfe/master/sumhdfe_demo_data.dta", clear
qui: reghdfe y x1 x2, a(firm year)
qui: reghdfe x1 if e(sample), a(firm year) resid
twoway (histogram x1, fcolor(green%75) lcolor(none)) (histogram _reghdfe_resid, ///
fcolor(navy%70) lcolor(none)), legend(on order(1 "x1" 2 "within-FE x1"))

Changelog

(will be added as new versions are posted)

Questions and bug reports

If you have questions or experience problems please use the issues tab of this repository.

Known bugs: