Hello all,
For my master's thesis, I am analyzing the effect of mitigation finance on emissions in developing countries. I have a mostly balanced panel with T=17 and N=127 (2149 observations in total). I regress CO2 emissions per capita on the lag of mitigation finance and various controls. I also use a shift-share type instrument to address potential endogeneity of mitigation finance.
Running my regressions using different fixed effects specifications leads to confusing results that I am struggling to interpret. I ran the following three commands in Stata for a pooled OLS, country fixed effects, year fixed effects and both:
The results show a positive and significant effect in pooled OLS, country fixed effects and year-fixed effects models, but not in the two-way fixed effects model. In the two-way fixed effects model, the effect not only changes in significance but also in magnitude and direction. I am struggling to understand why all pooled OLS, country fixed effects and year-fixed effects are consistent, but the two-way model deviates. As I understand it, the two-way model should be a weighted average of country-fixed effect and year-fixed effects estimator (see Kropko J, Kubinec R (2020) Interpretation and identification of within-unit and cross-sectional variation in panel data models. PLoS ONE 15(4): e0231349. https://doi.org/10.1371/journal. pone.0231349). But in fact, the estimator from the two-way model lies outside of the two estimators from one-way models.
Regression results:
The phenomenon persists when using robust errors instead of cluster(country) and also if I regress without instrumentation. The same phenomenon also occurs when using a time-trend variable instead of time fixed effects.
Are there any possible explanations for this phenomenon that I am overlooking? Could multicollinearity play a role (see VIF below)?
Did I correctly specify these models in Stata or is there maybe a mistake that might cause the phenomenon?
How worried should I be about the validity of my results?
Any help or thoughts would be greatly appreciated!
For my master's thesis, I am analyzing the effect of mitigation finance on emissions in developing countries. I have a mostly balanced panel with T=17 and N=127 (2149 observations in total). I regress CO2 emissions per capita on the lag of mitigation finance and various controls. I also use a shift-share type instrument to address potential endogeneity of mitigation finance.
Running my regressions using different fixed effects specifications leads to confusing results that I am struggling to interpret. I ran the following three commands in Stata for a pooled OLS, country fixed effects, year fixed effects and both:
Code:
ivreg2 logCO2_perc (l4mitfin_perc = l4instrument) l4ODAexmitfin_perc urbanpop urbanpop2 industry trade oilprice FDInet im_* if exclude==0, cluster(country) ivreg2 logCO2_perc (l4mitfin_perc = l4instrument) l4ODAexmitfin_perc urbanpop urbanpop2 industry trade oilprice FDInet im_* i.country if exclude==0, cluster(country) ivreg2 logCO2_perc (l4mitfin_perc = l4instrument) l4ODAexmitfin_perc urbanpop urbanpop2 industry trade FDInet im_* i.Year if exclude==0, cluster(country) ivreg2 logCO2_perc (l4mitfin_perc = l4instrument) l4ODAexmitfin_perc urbanpop urbanpop2 industry trade FDInet im_* i.country i.Year if exclude==0, cluster(country)
Regression results:
Code:
CO2 per capita on mitigation finance and controls using POLS, country FE, year FE and two-way
--------------------------------------------------------------
POLS Country~E Year FE Two-way~E
--------------------------------------------------------------
l4m~n_perc 0.0104*** 0.00197*** 0.0108*** -0.000232
(3.80) (2.81) (3.22) (-0.38)
l4ODAexm~c 0.00000759 0.0000281* 0.000000959 0.0000228
(0.05) (1.67) (0.01) (1.37)
urbanpop 0.0673*** 0.138*** 0.0673*** 0.107***
(3.88) (6.12) (3.88) (4.80)
urbanpop2 -0.000297* -0.000971*** -0.000297* -0.000890***
(-1.84) (-5.35) (-1.84) (-4.98)
industry 0.0246*** 0.00519** 0.0248*** 0.00666***
(3.54) (2.14) (3.56) (2.97)
trade 0.00861*** -0.0000413 0.00860*** 0.000391
(3.48) (-0.06) (3.45) (0.58)
oilprice -0.000729 0.000165
(-1.26) (0.67)
FDInet -0.00264 0.00134*** -0.00260 0.00180***
(-1.11) (4.25) (-1.11) (5.25)
_cons -3.728*** -4.286*** -3.789*** -3.039***
(-8.95) (-6.05) (-9.10) (-4.30)
--------------------------------------------------------------
N 2149 2149 2149 2149
adj. R-sq 0.534 0.984 0.530 0.985
--------------------------------------------------------------
t statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.01
Are there any possible explanations for this phenomenon that I am overlooking? Could multicollinearity play a role (see VIF below)?
Did I correctly specify these models in Stata or is there maybe a mistake that might cause the phenomenon?
How worried should I be about the validity of my results?
Code:
vif, uncentered
Variable | VIF 1/VIF
-------------+----------------------
l4mit~n_perc | 1.49 0.669579
l4ODAexmit~c | 3.07 0.325313
urbanpop | 1233.25 0.000811
urbanpop2 | 1177.62 0.000849
industry | 60.76 0.016458
trade | 60.68 0.016479
FDInet | 1.82 0.550937
Year |
2005 | 2.01 0.496449
2006 | 2.04 0.489568
2007 | 2.07 0.483273
2008 | 2.10 0.475808
2009 | 2.11 0.474606
2010 | 2.13 0.468701
2011 | 2.17 0.460143
2012 | 2.21 0.452265
2013 | 2.26 0.442214
2014 | 2.33 0.429059
2015 | 2.36 0.424420
2016 | 2.43 0.411883
2017 | 2.51 0.398927
2018 | 2.55 0.391649
2019 | 2.60 0.384961
2020 | 2.70 0.370296
country |
3 | 3.15 0.317913
4 | 2.65 0.377846
5 | 2.28 0.439513
6 | 8.94 0.111903
7 | 2.51 0.398990
8 | 2.47 0.405602
9 | 9.01 0.110954
10 | 1.88 0.532858
11 | 1.92 0.519542
12 | 4.12 0.242594
13 | 1.99 0.503501
14 | 1.94 0.514769
15 | 2.04 0.489394
16 | 2.60 0.385322
17 | 1.92 0.521321
18 | 2.50 0.400220
19 | 6.23 0.160591
20 | 1.75 0.571206
21 | 1.36 0.736306
22 | 2.29 0.437271
23 | 1.99 0.502342
24 | 1.99 0.501871
25 | 2.02 0.494776
26 | 1.86 0.538810
27 | 7.31 0.136851
28 | 2.19 0.457566
29 | 4.53 0.220887
30 | 1.87 0.535986
31 | 1.98 0.504399
32 | 3.02 0.331261
33 | 3.36 0.297850
34 | 1.94 0.514729
35 | 2.00 0.500838
37 | 3.68 0.271986
38 | 2.78 0.359872
39 | 3.77 0.264990
40 | 2.38 0.420112
41 | 2.01 0.498233
42 | 2.59 0.386649
43 | 3.86 0.259181
45 | 1.93 0.518332
46 | 1.65 0.605811
47 | 2.07 0.483948
48 | 7.87 0.127013
49 | 2.09 0.479268
50 | 2.03 0.493237
51 | 1.93 0.517554
53 | 1.93 0.517781
54 | 1.90 0.526729
55 | 1.97 0.506905
56 | 3.23 0.309750
57 | 1.97 0.507193
58 | 2.03 0.492640
59 | 1.87 0.533884
60 | 2.17 0.460164
61 | 3.46 0.288802
62 | 3.57 0.280183
63 | 1.98 0.504422
64 | 7.22 0.138425
65 | 2.10 0.476286
66 | 1.71 0.583344
67 | 2.10 0.476679
70 | 2.02 0.496081
71 | 1.90 0.527396
72 | 7.18 0.139245
73 | 2.22 0.449883
75 | 5.56 0.179847
76 | 1.89 0.528715
78 | 3.79 0.263559
79 | 2.54 0.393036
80 | 1.88 0.532167
81 | 14.57 0.068631
82 | 4.47 0.223849
83 | 1.94 0.516269
84 | 2.00 0.500903
85 | 4.47 0.223489
86 | 3.03 0.330309
87 | 1.98 0.505779
88 | 2.70 0.370008
89 | 2.42 0.413289
90 | 2.09 0.479397
91 | 1.90 0.525179
92 | 1.93 0.519043
93 | 1.91 0.523487
95 | 1.54 0.650987
96 | 2.05 0.488564
97 | 1.52 0.658227
98 | 1.97 0.506488
99 | 2.10 0.475925
100 | 5.50 0.181975
101 | 1.95 0.513663
103 | 2.66 0.376278
104 | 2.47 0.405204
105 | 2.20 0.454392
106 | 4.14 0.241429
107 | 1.93 0.517076
108 | 1.52 0.656904
109 | 1.69 0.592417
111 | 6.50 0.153947
112 | 1.91 0.523119
113 | 2.05 0.487440
114 | 2.84 0.352640
115 | 2.02 0.496253
116 | 2.17 0.459967
117 | 1.75 0.572224
119 | 2.31 0.433559
121 | 1.71 0.583653
125 | 1.94 0.516059
126 | 2.88 0.346829
128 | 1.82 0.549997
129 | 1.82 0.549126
130 | 2.10 0.476939
131 | 1.94 0.516697
132 | 1.89 0.528156
133 | 1.81 0.553547
135 | 2.62 0.381430
136 | 3.17 0.315633
137 | 2.30 0.434830
140 | 1.65 0.606500
141 | 2.77 0.361197
142 | 10.78 0.092807
143 | 1.94 0.515434
144 | 1.98 0.505916
146 | 2.22 0.450265
149 | 1.92 0.520602
150 | 1.86 0.538124
im_industry | 1.43 0.700963
im_FDInet | 1.94 0.516729
im_trade | 2.46 0.405749
-------------+----------------------
Mean VIF | 19.35
Comment