6
$\begingroup$

I'm trying to re-create an analysis done using Stata function xtreg (though I don't have the code) with R package plm, and I'm having trouble translating between the two. A minimal example of the model I'm trying to estimate would be as follows:

$$ \text{Dependent}_{i,t} = \beta_0 + \gamma_i + \delta_t + \beta_1*\text{Dependent}_{i,t-1} + \beta_2*\text{Independent}_{i,t} + \epsilon_{i,t} $$

where $\gamma_i$ is a control for the cross-sectional factor (States), $\delta_t$ is a control for each time period (Year), $\text{Dependent}_{i,t}$ is a dependent variable in the form of a proportion, and $\text{Independent}_{i,t}$ is an attribute of a record that changes with both the cross-sectional and time component of the data.

I've attempted to estimate the coefficients for the model using R with the following (generalized) code:

model <- plm(Dependent_it ~ Independent_it + lag(Dependent_it,1), 
             data   = data, 
             method = "within", #fixed effects model
             effect = "twoway", #does the gamma_i and delta_t parts (I think)
             index  = c("State", "Year")
             )

I've tried to re-create it in Stata using the following (after running xtset State Year, yearly of course)

xtreg Dependent Independent l.Dependent i.Year, fe

I get different results, so I must be doing something wrong with either xtreg or plm, or both. Can someone enlighten me as to if I've coded the above model appropriately in either language?

$\endgroup$
1
  • 3
    $\begingroup$ You should keep in mind that the within estimators requires a strict exogeneity assumption which is necessarily violated if you include lagged outcomes as regressors, e.g. www.nber.org/WNE/lect_2_linpanel.pdf‎ $\endgroup$ Aug 29 '13 at 12:51
10
$\begingroup$

Welcome to the site, @gwatson! You are right that effect = "twoways" sets up both "individual" and "year" effects.

I tested with Produc data from R package plm and found the main results are the same (see the codes and outputs below). The only apparent difference I found is the year effect, which is caused by contrast (xtreg sets the first year as reference, while plm directly estimates the effect for each year).

## R code
data("Produc", package = "plm")
zz <- plm(gsp ~ unemp + lag(gsp), data = Produc, index = c("state","year"), method = "within", effect = "twoways")
summary(zz)

## plm output
Coefficients :
            Estimate  Std. Error  t-value  Pr(>|t|)    
unemp    -5.4525e+02  6.8611e+01  -7.9469 7.614e-15 ***
lag(gsp)  1.0125e+00  9.1789e-03 110.3029 < 2.2e-16 ***


## Stata code
use Produc, clear
xtset state year, yearly
xtreg gsp unemp l.gsp i.year, fe

## xtreg output
------------------------------------------------------------------------------
         gsp |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
       unemp |   -545.246   68.61136    -7.95   0.000    -679.9537   -410.5383
      gsp L1.|   1.012464   .0091789   110.30   0.000     .9944422    1.030485
-------------+----------------------------------------------------------------
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.