谢谢连老师亲切的答辩, 真是胜读十年Manual啊。。
1) 我的估计模型的右边自带被解释变量的一阶滞后差分项。您的课程中有用OLS和FE模型做SystemGMM估计效率性的标准,我用固定效果模型(最下偏估计)比较了加和不加一阶滞后项的模型。结果如下, OLS的估计跟FE的结果差不多。
. xtreg fc3 dl.fc3 dcon3 dep rfc3 i ta frn1, fe r cluster(id)
Fixed-effects (within) regression Number of obs = 348
Group variable: id Number of groups = 29
R-sq: within = 0.8725 Obs per group: min = 12
between = 0.9353 avg = 12.0
overall = 0.9040 max = 12
F(7,28) = 103.30
corr(u_i, Xb) = 0.5613 Prob > F = 0.0000
(Std. Err. adjusted for 29 clusters in id)
------------------------------------------------------------------------------
| Robust
fc3 | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
fc3 |
LD. | .8284162 .177271 4.67 0.000 .465293 1.191539
|
dcon3 | -.111504 .0654094 -1.70 0.099 -.2454892 .0224811
dep | .9851547 .0847462 11.62 0.000 .81156 1.158749
rfc3 | -.0838315 .0444635 -1.89 0.070 -.1749107 .0072478
i | -3.876992 .4710413 -8.23 0.000 -4.841876 -2.912107
ta | 2.829591 1.227325 2.31 0.029 .3155299 5.343653
frn1 | 1.913851 .555001 3.45 0.002 .7769828 3.050719
_cons | -.327994 .3515896 -0.93 0.359 -1.048193 .3922046
-------------+----------------------------------------------------------------
sigma_u | .18043037
sigma_e | .10681976
rho | .74046832 (fraction of variance due to u_i)
------------------------------------------------------------------------------
. xtreg fc3 l.fc3 dl.fc3 dcon3 dep rfc3 i ta frn1, fe r cluster(id)
Fixed-effects (within) regression Number of obs = 348
Group variable: id Number of groups = 29
R-sq: within = 0.9682 Obs per group: min = 12
between = 0.9995 avg = 12.0
overall = 0.9908 max = 12
F(8,28) = 1380.57
corr(u_i, Xb) = 0.8196 Prob > F = 0.0000
(Std. Err. adjusted for 29 clusters in id)
------------------------------------------------------------------------------
| Robust
fc3 | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
fc3 |
L1. | .9294488 .0371141 25.04 0.000 .853424 1.005474
LD. | .4826679 .0735649 6.56 0.000 .331977 .6333587
|
dcon3 | .0393039 .0276015 1.42 0.166 -.0172351 .0958429
dep | -.0165078 .0540171 -0.31 0.762 -.1271569 .0941413
rfc3 | .0065462 .0178644 0.37 0.717 -.0300474 .0431398
i | .193571 .2192272 0.88 0.385 -.2554954 .6426375
ta | .8479162 .6881883 1.23 0.228 -.5617736 2.257606
frn1 | -.6416758 .3801849 -1.69 0.103 -1.420449 .1370977
_cons | .2978077 .2229683 1.34 0.192 -.1589222 .7545375
-------------+----------------------------------------------------------------
sigma_u | .03987484
sigma_e | .05339258
rho | .35804662 (fraction of variance due to u_i)
------------------------------------------------------------------------------
根据动态面板估计理论, 这两个估计不应该有这么大的差异。这应该怎么解释啊?是模型本身和数据的问题吗?
2) 还有一个问题, 用对数转换平稳序列的时候, Ln(X+a)和 Ln((X+a)/a)两种方法中那一种比较合理? 各个序列平均的移动对估计结果没有太大影响(除系数大小变化外)吧?
3) gmm()选项中 Collapse 选项对小样本的估计结果有无大的影响啊??
等候您的答复。再次感谢!!!
|