影響交通事故率高低的因素與因果推論？

原始資料來源

交通部統計查詢網
http://stat.motc.gov.tw/mocdb/stmain.jsp?sys=100

Idea 構想：

看看交通事故率的高低，是否和以下因素有關 與 猜想背後的因果關係

1. 新手駕駛=>看駕照數的增減
2. 景氣好壞 =>失業率/進出口金額
3. 空氣品質 =>看環保署空氣汙染指標
4. 天候氣溫 =>氣溫日照天數
5. 官方稽查 =>稽查開罰件數


Disclaimer: 僅為統計Kuso應用

 

 

 

Part 1 Cross-Sectional Separated Regression

假說1

新手駕駛(大小客車駕照) 

Dependent Variable: DYA1
Method: Least Squares
Sample (adjusted): 2001M02 2016M04
Included observations: 183 after adjustments
Variable	Coefficient	Std. Error	t-Statistic	Prob.
DLA	4.881779	6.423653	0.759969	0.4483
DLB	-1.922808	6.677659	-0.287946	0.7737
R-squared	-0.007175	Mean dependent var		1.204678
Adjusted R-squared	-0.012740	S.D. dependent var		9.032821
S.E. of regression	9.090177	Akaike info criterion		7.263135
Sum squared resid	14956.27	Schwarz criterion		7.298211
Log likelihood	-662.5768	Hannan-Quinn criter.		7.277353
Durbin-Watson stat	2.987217

呵，解釋力很糟

假說2

景氣好壞
解釋力35% 

Dependent Variable: DYA1
Method: Least Squares
Sample (adjusted): 2001M02 2016M04
Included observations: 183 after adjustments
Variable	Coefficient	Std. Error	t-Statistic	Prob.
E1	0.224664	0.120185	1.869316	0.0632
DE2	0.002045	0.000205	10.00076	0.0000
DE3	-0.858224	0.601087	-1.427787	0.1551
R-squared	0.358201	Mean dependent var		1.204678
Adjusted R-squared	0.351070	S.D. dependent var		9.032821
S.E. of regression	7.276499	Akaike info criterion		6.823434
Sum squared resid	9530.539	Schwarz criterion		6.876049
Log likelihood	-621.3442	Hannan-Quinn criter.		6.844762
Durbin-Watson stat	2.719075

對交通事故率與外銷出口的金額的關聯性非常顯著(景氣好，事故率高)

 

假說3

空氣品質(霧霾)

解釋力5%

Dependent Variable: DYA1
Method: Least Squares
Sample (adjusted): 2001M02 2016M04
Included observations: 181 after adjustments
Variable	Coefficient	Std. Error	t-Statistic	Prob.
DAQ1	-0.369524	0.354330	-1.042880	0.2984
DAQ2	-0.115151	0.083183	-1.384303	0.1680
DAQ3	0.353249	0.113126	3.122610	0.0021
R-squared	0.056868	Mean dependent var		1.378677
Adjusted R-squared	0.046271	S.D. dependent var		8.905764
S.E. of regression	8.697285	Akaike info criterion		7.180334
Sum squared resid	13464.41	Schwarz criterion		7.233348
Log likelihood	-646.8203	Hannan-Quinn criter.		7.201827
Durbin-Watson stat	2.813369

空氣品質不好、能見度低，所以事故率會高？

亦或
空氣品質好，大家開快車，所以事故率高？！

 

 

假說4

天候氣溫
解釋力10% 

Dependent Variable: DYA1
Method: Least Squares
Sample (adjusted): 2001M02 2016M04
Included observations: 183 after adjustments
Variable	Coefficient	Std. Error	t-Statistic	Prob.
DW1	0.062103	0.019511	3.182986	0.0017
DW2	-0.035565	0.046516	-0.764575	0.4455
DW3	0.162819	0.147772	1.101827	0.2720
DW4	0.094797	0.188125	0.503903	0.6149
R-squared	0.101875	Mean dependent var		1.204678
Adjusted R-squared	0.086823	S.D. dependent var		9.032821
S.E. of regression	8.631791	Akaike info criterion		7.170397
Sum squared resid	13336.90	Schwarz criterion		7.240549
Log likelihood	-652.0913	Hannan-Quinn criter.		7.198833
Durbin-Watson stat	2.825518

日照天數與事故率顯著相關(日照天數多，大家出遊多，所以事故率高？)

 

假說5

官方執法稽查

解釋力30% 

Dependent Variable: DYA1
Method: Least Squares
Sample (adjusted): 2001M02 2016M04
Included observations: 183 after adjustments
Variable	Coefficient	Std. Error	t-Statistic	Prob.
DP1	0.037243	0.006352	5.863581	0.0000
DP2	0.002186	0.004392	0.497629	0.6193
R-squared	0.309954	Mean dependent var		1.204678
Adjusted R-squared	0.306141	S.D. dependent var		9.032821
S.E. of regression	7.524176	Akaike info criterion		6.884988
Sum squared resid	10246.99	Schwarz criterion		6.920065
Log likelihood	-627.9764	Hannan-Quinn criter.		6.899207
Durbin-Watson stat	2.377284

關係顯著，但兩者是正相關，而非負相關

 

 

Part 2 Cross-Sectional Pooled Regression

以上這些因素都放在一起回歸的時候

哪個因素會顯著、relevant?

 

Dependent Variable: DYA1
Method: Least Squares
Sample (adjusted): 2001M02 2016M04
Included observations: 183 after adjustments
Variable	Coefficient	Std. Error	t-Statistic	Prob.
C	1.338444	0.695792	1.923625	0.0560
DLA	-1.091594	2.428728	-0.449451	0.6537
DE2	0.001414	0.000218	6.492952	0.0000
DAQ3	0.031253	0.041543	0.752312	0.4529
DW1	0.038969	0.010532	3.700053	0.0003
DP1	0.020305	0.004571	4.441803	0.0000
R-squared	0.491222	Mean dependent var		1.204678
Adjusted R-squared	0.476850	S.D. dependent var		9.032821
S.E. of regression	6.533360	Akaike info criterion		6.623957
Sum squared resid	7555.208	Schwarz criterion		6.729186
Log likelihood	-600.0921	Hannan-Quinn criter.		6.666612
F-statistic	34.17849	Durbin-Watson stat		2.317044
Prob(F-statistic)	0.000000

新手駕駛、空氣品質不顯著

景氣、天候與官方稽查顯著正相關

 

Part3 Longitudinal Regression

確認不同因素的先後關係以判斷因果關係的解釋

e.g.

景氣好=>車輛數增加=>空物污染增加=>造成車禍事故率高？

車禍事故率高=>官方稽核增加 亦或 官方稽核增加=>車禍事故率低

 

 

Pairwise Granger Causality Tests
Date: 07/14/16   Time: 14:29
Sample: 2001M01 2016M04
Lags: 1
Null Hypothesis:	Obs	F-Statistic	Prob.
DYA1 does not Granger Cause DP1	182	14.0995	0.0002
DP1 does not Granger Cause DYA1		4.86715	0.0286

Pairwise Granger Causality Tests
Date: 07/14/16   Time: 14:30
Sample: 2001M01 2016M04
Lags: 7
Null Hypothesis:	Obs	F-Statistic	Prob.
DYA1 does not Granger Cause DP1	176	8.10248	2.E-08
DP1 does not Granger Cause DYA1		2.51959	0.0174

是因為上個月事故多，官方才會加強稽查
不是官方稽查有助於減少事故發生！