home workflow spillover spmatrix spregress spxtregress

Panel-data Spatial Autoregression : spxtregress


Spatial static panel-data models

Two models

spxtregress can fits two different static spatial-panel-data autoregression models with fixed effects estimator (fe) or random effects estimator (re).

Model I (the default)

$$ \begin{align} y_{nt} &= \lambda W y_{nt} + X_{nt} \beta + \sum_{k=1}^K W_k X_{nt}\gamma_k + {\color{red}{c_n}} + u_{nt} \nonumber \\ u_{nt} &= \rho M u_{nt} + \epsilon_{nt} \label{eq:model_i} \end{align} $$

Model II (with option -sarpanel-)

$$ \begin{align} y_{nt} &= \lambda W y_{nt} + X_{nt} \beta + \sum_{k=1}^K W_k X_{nt}\gamma_k + u_{nt} \nonumber \\ u_{nt} &= \rho M u_{nt} + {\color{red}{c_n}} + \epsilon_{nt} \label{eq:model_ii} \end{align} $$

Comments


Fixed-effects estimator

Two models are the same for FE estimator

We write Model I and II by explicitly writing out the spatial error terms.

$$ \begin{align*} y_{nt} &= \lambda W y_{nt} + X_{nt} \beta + \sum_{k=1}^K W_k X_{nt}\gamma_k + {\color{red}{c_n}} + (I - \rho M)^{-1}\epsilon_{nt} \\ y_{nt} &= \lambda W y_{nt} + X_{nt} \beta + \sum_{k=1}^K W_k X_{nt}\gamma_k + {\color{red}(I-\rho M)^{-1}c_n} +(I - \rho M)^{-1}\epsilon_{nt} \end{align*} $$

We can just regard ${\color{red}(I-\rho M)^{-1}c_n}$ as another form of fixed effect.

FE cannot estimate a time invariate term

The spxtregress, fe use a within-type transformations (Helmert transformation) to remove the fixed effects $c_n$. However, this transformation also removes all the time invariate terms. So FE cannot estimate a time invariate term.


. spxtregress hrate  gini i.year, fe  dvarlag(W) errorlag(W)
  (1016 observations)
  (1016 observations used)
  (data contain 254 panels (places) )
  (weighting matrix defines 254 places)

Performing grid search ... finished 

Optimizing concentrated log likelihood:

Iteration 0:   log likelihood = -2569.5658  
Iteration 1:   log likelihood = -2569.3675  
Iteration 2:   log likelihood = -2569.3668  
Iteration 3:   log likelihood = -2569.3668  

Optimizing unconcentrated log likelihood:

Iteration 0:   log likelihood = -2576.1229  
Iteration 1:   log likelihood = -2574.3725  
Iteration 2:   log likelihood = -2574.3683  
Iteration 3:   log likelihood = -2574.3683  

Fixed-effects spatial regression                Number of obs     =      1,016
Group variable: _ID                             Number of groups  =        254
                                                Obs per group     =          4

                                                Wald chi2(2)      =      78.53
                                                Prob > chi2       =     0.0000
Log likelihood = -2574.3683                     Pseudo R2         =     0.0017

------------------------------------------------------------------------------
       hrate |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
hrate        |
        gini |   34.71117   9.312023     3.73   0.000     16.45994     52.9624
             |
        year |
       1970  |          0  (omitted)
       1980  |          0  (omitted)
       1990  |          0  (omitted)
-------------+----------------------------------------------------------------
W            |
       hrate |  -.8918254   .1171214    -7.61   0.000    -1.121379   -.6622718
     e.hrate |   .7645118   .0574604    13.31   0.000     .6518915    .8771321
-------------+----------------------------------------------------------------
    /sigma_e |   6.414088   .2154119                      6.005484    6.850493
------------------------------------------------------------------------------
Wald test of spatial terms:          chi2(2) = 197.10     Prob > chi2 = 0.0000

. estat impact

progress   :100% 

Average impacts                                 Number of obs     =      1,016

------------------------------------------------------------------------------
             |            Delta-Method
             |      dy/dx   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
direct       |
        gini |   38.06052   10.15738     3.75   0.000     18.15242    57.96863
-------------+----------------------------------------------------------------
indirect     |
        gini |  -18.15548   5.067272    -3.58   0.000    -28.08715   -8.223809
-------------+----------------------------------------------------------------
total        |
        gini |   19.90504   5.550587     3.59   0.000     9.026093      30.784
------------------------------------------------------------------------------

Note that the variable year is omitted in the estat impact.


Random-effects estimator

Key difference between two models for RE ?

$$ \begin{align*} y_{nt} &= \lambda W y_{nt} + X_{nt} \beta + \sum_{k=1}^K W_k X_{nt}\gamma_k + {\color{red}(I-\rho M)^{-1}c_n} +{\color{blue}(I - \rho M)^{-1}\epsilon_{nt}} \end{align*} $$

. qui spxtregress hrate gini i.year, re dvarlag(W) 

. estimate store model_1

. qui spxtregress hrate gini i.year, re dvarlag(W) sarpanel

. estimate store model_2

. estimate table model_1 model_2

----------------------------------------
    Variable |  model_1      model_2    
-------------+--------------------------
hrate        |
        gini |  4.3603851    4.3603851  
             |
        year |
       1970  |   2.230819     2.230819  
       1980  |   2.729887     2.729887  
       1990  |  1.9152529    1.9152529  
             |
       _cons |   3.186456     3.186456  
-------------+--------------------------
W            |
       hrate |  .20003314    .20003314  
-------------+--------------------------
sigma_u      |
       _cons |   2.625071     2.625071  
-------------+--------------------------
sigma_e      |
       _cons |  7.1365439    7.1365439  
----------------------------------------


estat impact

impacts averaged over all the years

. qui spxtregress hrate gini, re dvarlag(W) 

. estat impact

progress   :100% 

Average impacts                                 Number of obs     =      1,016

------------------------------------------------------------------------------
             |            Delta-Method
             |      dy/dx   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
direct       |
        gini |   17.53096   5.208656     3.37   0.001     7.322179    27.73974
-------------+----------------------------------------------------------------
indirect     |
        gini |   4.508373   1.654764     2.72   0.006     1.265095    7.751652
-------------+----------------------------------------------------------------
total        |
        gini |   22.03933   6.513159     3.38   0.001     9.273774    34.80489
------------------------------------------------------------------------------

impacts averaged over selected year

. estat impact if year == 1960

progress   :100% 

Average impacts                                 Number of obs     =        254

------------------------------------------------------------------------------
             |            Delta-Method
             |      dy/dx   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
direct       |
        gini |   17.53096   5.208656     3.37   0.001     7.322179    27.73974
-------------+----------------------------------------------------------------
indirect     |
        gini |   4.508373   1.654764     2.72   0.006     1.265095    7.751652
-------------+----------------------------------------------------------------
total        |
        gini |   22.03933   6.513159     3.38   0.001     9.273774    34.80489
------------------------------------------------------------------------------