xtpcse和reghdfe的结果应该用哪个

702

收藏 2023-04-10

我用的是n=13,t=36的面板数据。分别用stata的xtpcse和reghdfe两种方法做了双向固定效应模型，两种方法的结果大相径庭。求问应该选择哪种方法呢？我是个stata和计量小白，如果下边的做法和理论应用上有不足之处还请前辈们多多指教

谢谢大家！

最开始找代码做了reghdfe，结果还可以

. reghdfe l_r tmax_r tmin_r prcp_r tmax_r_square tmin_r_square prcp_r_square prcp
> _r*tmax_r prcp_r*tmin_r tmax_r*tmin_r,absorb(i.c i.y) vce(cluster i.c)  //tavg_
> r_rquare
(MWFE estimator converged in 2 iterations)

HDFE Linear regression                         Number of obs =       468
Absorbing 2 HDFE groups                         F( 9,    12) =    115.22
Statistics robust to heteroskedasticity          Prob > F       =    0.0000
                                                R-squared    =    0.7493
                                                Adj R-squared =    0.7144
                                                Within R-sq. =    0.0475
Number of clusters (c)    =       13       Root MSE       =    0.2271

                                    (Std. err. adjusted for 13 clusters in c)
-------------------------------------------------------------------------------
            |             Robust
               l_r | Coefficient std. err.    t    P>|t|    [95% conf. interval]
--------------+----------------------------------------------------------------
         tmax_r | 3.658484 .7934298    4.61 0.001    1.92975 5.387219
         tmin_r  |  -1.634608 .3739787 -4.37 0.001 -2.449438 -.8197785
         prcp_r | .0474761 .0266285    1.78 0.100 -.0105425 .1054947
tmax_r_square |  -.1005117 .0203735 -4.93 0.000 -.1449018 -.0561216
tmin_r_square |  -.0196665 .0065848 -2.99 0.011 -.0340136 -.0053195
prcp_r_square | .0000145 .0000317    0.46 0.655 -.0000545 .0000835
  prcp_rtmax_r | -.002593 .0012594 -2.06 0.062 -.0053369    .000151
prcp_rtmin_r | .0005757 .0007549    0.76 0.460 -.0010691 .0022205
  tmax_rtmin_r | .0881084 .0160395    5.49 0.000    .0531613 .1230555
         _cons | -24.0849 7.882248 -3.06 0.010 -41.25885 -6.910961
-------------------------------------------------------------------------------

然后昨天看了一篇论文，作者是做了一系列检验之后，发现误差项存在异方差、自相关和截面相关，后来选择用PCSE修正结果。我也复刻了一下

//--------------------------------自相关
. xtserial l_r tmax_r tmin_r prcp_r tmax_r_square tmin_r_square prcp_r_square prc
> p_r*tmax_r prcp_r*tmin_r tmax_r*tmin_r

Wooldridge test for autocorrelation in panel data
H0: no first order autocorrelation
F(  1,    12) =    49.909
         Prob > F =    0.0000

//--------------------------------异方差
. quietly xtreg l_r tmax_r tmin_r prcp_r tmax_r_square tmin_r_square prcp_r_squar
> e prcp_r*tmax_r prcp_r*tmin_r tmax_r*tmin_r,fe
. xttest2

Breusch-Pagan LM test of independence: chi2(78) =  1478.384, Pr = 0.0000
Based on 36 complete observations over panel units

//--------------------------------横截面自相关

. quietly xtreg l_r tmax_r tmin_r prcp_r tmax_r_square tmin_r_square prcp_r_squar
> e prcp_r*tmax_r prcp_r*tmin_r tmax_r*tmin_r,fe
. xtcsd,pesaran

Pesaran's test of cross sectional independence = 37.735, Pr = 0.0000

以我的理解是这三个检验结果都是拒绝原假设，误差项存在自相关、异方差和横截面自相关。
然后我按照在文献上的方法用PCSE方法做了固定效应模型，结果的系数非常不好

. xtpcse l_r tmax_r tmin_r prcp_r tmax_r_square tmin_r_square prcp_r_square prcp_
> r*tmax_r prcp_r*tmin_r tmax_r*tmin_r,corr(psar1)

Prais–Winsten regression, correlated panels corrected standard errors (PCSEs)

Group variable: c                            Number of obs    =       468
Time variable: y                            Number of groups  =       13
Panels:          correlated (balanced)       Obs per group:
Autocorrelation:  panel-specific AR(1)                      min =       36
                                                            avg =       36
                                                            max =       36
Estimated covariances    =       91       R-squared       =    0.9498
Estimated autocorrelations =       13       Wald chi2(9)    =    20.80
Estimated coefficients    =       10       Prob > chi2    =    0.0136

-------------------------------------------------------------------------------
            |          Panel-corrected
      l_r       | Coefficient std. err.       z    P>|z|    [95% conf. interval]
--------------+----------------------------------------------------------------
         tmax_r |  -.1140283 1.972278 -0.06 0.954 -3.979622 3.751566
         tmin_r | .9949556 .9287658    1.07 0.284    -.825392 2.815303
         prcp_r |  -.0260159 .0630197 -0.41 0.680 -.1495322 .0975004
tmax_r_square | .0111037 .0527673    0.21 0.833 -.0923183 .1145256
tmin_r_square | .0055959 .0182056    0.31 0.759 -.0300865 .0412782
prcp_r_square | .0000282 .0000626    0.45 0.652 -.0000945    .000151
prcp_rtmax_r | .0016126 .0035213    0.46 0.647 -.0052891 .0085143
prcp_rtmin_r |  -.0013843 .0019009 -0.73 0.466    -.00511 .0023415
  tmax_rtmin_r |  -.0446226 .0479093 -0.93 0.352 -.1385231 .0492778
         _cons | 5.190663 18.32978    0.28 0.777 -30.73504 41.11637
-------------------------------------------------------------------------------
      rhos =  .3946532  .0071098  .5005875  .5908927  .4767165 ...  .7193214
-------------------------------------------------------------------------------

扫码加我拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

全部回复

917968079

2023-4-10 19:10:06

你这个时候cluster的标准误肯定是不一致的，cluster的数量太少了

扫码加我拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

马红梅加油

2023-4-10 21:48:57

917968079 发表于 2023-4-10 19:10
你这个时候cluster的标准误肯定是不一致的，cluster的数量太少了

谢谢您的回复~您的意思是我这个数据不适合用reghdfe的cluster方法是嘛？
后来我又做了robust的，结果如下，
还请您帮我看一下下边的结果和xtpcse的结果哪个更合适呢？

. reghdfe l_r tmax_r tmin_r prcp_r tmax_r_square tmin_r_square prcp_r_square prcp_r*tmax_r pr
> cp_r*tmin_r tmax_r*tmin_r,absorb(i.c i.y) vce(robust)  //tavg_r_rquare
(MWFE estimator converged in 2 iterations)

HDFE Linear regression                         Number of obs =       468
Absorbing 2 HDFE groups                         F( 9, 411) =    3.18
                                                Prob > F       =    0.0010
                                                R-squared    =    0.7493
                                                Adj R-squared =    0.7151
                                                Within R-sq. =    0.0475
                                                Root MSE       =    0.2269

-------------------------------------------------------------------------------
            |             Robust
                  l_r | Coefficient    std. err.    t       P>|t|    [95% conf. interval]
--    ------------+----------------------------------------------------------------
         tmax_r | 3.658484 .9928073    3.68 0.000    1.706871 5.610098
         tmin_r |  -1.634608 .5336433 -3.06 0.002 -2.683619 -.5855974
         prcp_r | .0474761 .0343795    1.38 0.168 -.0201056 .1150578
tmax_r_square |  -.1005117 .0274376 -3.66 0.000 -.1544473 -.0465761
  tmin_r_square |  -.0196665 .0108181 -1.82 0.070 -.0409322 .0015992
  prcp_r_square | .0000145 .0000365    0.40 0.691 -.0000572 .0000863
prcp_rtmax_r | -.002593 .0019015 -1.36 0.173 -.0063309    .001145
prcp_rtmin_r | .0005757 .0011683    0.49 0.622    -.001721 .0028724
tmax_rtmin_r | .0881084 .0259382    3.40 0.001    .0371204 .1390964
      _cons | -24.0849 8.915233 -2.70 0.007 -41.61005 -6.559761
-------------------------------------------------------------------------------

扫码加我拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

马红梅加油

2023-4-10 21:51:42

我还有个疑惑，就是长面板数据做固定效应模型的时候一定要进行误差自相关和异方差检验吗？感觉好难啊

扫码加我拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

扫码加我 拉你入群

扫码加我 拉你入群

扫码加我 拉你入群

扫码加我 拉你入群

分享

扫码加好友，拉您进群

扫码加我拉你入群

扫码加我拉你入群

扫码加我拉你入群

扫码加我拉你入群