3.1. Basic Principles of the DID Model
The difference-in-difference (DID) is a policy evaluation technique widely used in sociology. Its basic principle is to start from a counterfactual framework to evaluate the changes of a certain observed factor Y under the two situations of policy occurrence and policy failure [
27]. The DID model supposes that an exogenous policy shock can divide the sample into two groups, namely the Treat Group that is affected by the policy and the Control Group that is not affected by the policy. If there is no significant difference between the observed factor Y in the Treat Group and the Control Group before the policy shock, the change in Y of the Control Group before and after the policy can be regarded as the situation of the Treat Group when it did not receive the policy shock (called the counterfactual result). In this way, by comparing the change in Treat Group Y (D1) and the change in Control Group Y (D2), the actual impact of policy shocks on the observed factor Y (DID = D1 − D2) can be calculated.
When using the DID model to evaluate the performance of policy implementation, it should be firstly assumed that the model meets the linear regression assumption. In addition, three premises should be met [
28], that is, the policy shock has no significant impact on the Control Group, factors other than policy shocks have the same impact on the Treat Group and the Control Group, and the observed variables in the Treat Group and the Control Group have stable distribution characteristics and do not change with time.
In 1985, Princeton University scholars Ashenfelter and Card introduced the DID model for the first time in a project evaluation article [
29]. Subsequently, the model was widely used in the fields of econometrics and sociology [
30,
31]. According to the basic principles of the DID model, the following two key conditions must be met when using the DID model: First, there must be a pilot policy impact, otherwise it is impossible to find a Treat Group and a Control Group for the researched policy. That is to say, those policies that are implemented for all objects at one time are not suitable for DID analysis. Second, considering that policy effects often take a while to appear, a panel data set of at least two years (one year before and after policy implementation) must be available to realize parameter estimation. Suppose there are the following two periods of panel data:
where
,
, and
are regression parameters to be estimated in this equation.
is a dummy variable reflecting the policy implementation period (
, if
t = 2, indicating that the policy has been implemented;
, if
t = 1, indicating that the policy has not been implemented).
is a dummy variable reflecting the policy implementation object, and its value is shown in formula (2)
is the unobservable individual characteristic, and
is the residual item.
According to formula (2), when t = 1, the Treat Group and the Control Group are treated the same, that is, is equal to 0; when t = 2, the of the Treat Group is set to 1, while the of the Control Group is still equal to 0.
If the treatment of policy shocks cannot be completely randomized,
may be related to
, resulting in inconsistencies in the results of least squares regression (OLS). Since the panel data is used, the original formula (1) can be first-differenced to eliminate
:
Using the OLS to estimate the above formula (3), the consistency estimator
can be obtained. According to the same reasoning as the difference estimator:
where
is the average value of the observations in each group. This method is called Difference-in-Difference estimator (abbreviated as DID), denoted as
. The DID estimator has eliminated the effects of pretreatment differences between the Treat Group and the Control Group, as shown in
Figure 6.
For DID estimators, other explanatory variables
can also be introduced, namely:
where
are the regression parameters to be estimated in this equation. The DID model with
as the explained variable is not applicable to multi-period data.
Returning to the panel model with
as the explained variable, without considering the explanatory variables, and still assuming two-period panel data, it can be considered that the original Equation (1) is equivalent to the following two-period panel model:
where
,
,
and
are the regression parameters to be estimated in this equation, and
. is the residual item.
. is the dummy variable of the Treat Group (
. when the individual
i belongs to the Treat Group;
. when the individual
i belongs to the Control Group).
. is a dummy variable reflecting the period of policy intervention (
. if
t = 2 and
if
t = 1). The interactive item
(valued 1 if the individual
i belongs to the Treat Group and
t = 2, otherwise it values 0). The grouping dummy variable
represents the difference between the Treat Group and the Control Group itself. The time dummy variable
represents the difference between the two periods before and after the policy intervention. The interaction term
. is the net effect of the policy. If there are other explanatory variables, they can be directly put into the above equation.
When
t = 1, the equation can be written as:
When
t = 2, the equation can be written as:
Subtract the two equations to get:
This equation is exactly the same as the difference equation shown by formula (3). Performing the OLS estimation on this equation, the obtained is the coefficient of the interaction term , which is called the DID estimator. The advantage of the DID method is that it simultaneously controls time-specific effects and group-specific effects .
3.2. Model Hypothesis and Sample Data
The main purpose of this section is to examine the implementation effects of China’s SSL industry policies from the perspective of energy conservation and emission reduction. China’s electricity consumption for lighting accounts for about 14% of the entire society’s electricity consumption, which is an important area of energy conservation and emission reduction. The 2016–2017 China SSL Products Market Research Report shows that China’s LED lamp market share in 2017 was 69.1%, of which the market share of LED lamps in indoor lighting was 68.8%. In 2016, China’s LED lamp market share was 48.7%, of which the market share of LED lamps in indoor lighting was 47.7% [
32]. In 2016 and 2017, the market share of LED lamps in outdoor lighting street lamps was about 20%, and the market share of night scene lighting LED was 100%.
The total electricity consumption of lighting in 2017 was 885.8 billion kWh, accounting for 14.04% of the total electricity consumption of the whole society, and the total electricity consumption of lighting in 2016 was 839.9 billion kWh, accounting for 14.19% of the total electricity consumption of the whole society. Due to the improvement of urban planning and construction in 2016–2017, the proportion of electricity consumption for lighting in the total electricity consumption of society has dropped significantly. The above data show that the use of SSL products has played an important role in national energy conservation and emission reduction [
33].
Based on the above analysis, the hypothesis (H) is made: the SSL industry policy promotes energy conservation and emission reduction.
In order to verify the above hypothesis, this section applies the DID method to build am SSL industry policy effect evaluation model. From the point of view of policy implementation, 2009 and 2010 are turning years for the development of China’s SSL industry. As the National Development and Reform Commission issued the Guidance for the Development of SSL Energy-saving Industry in 2009 and initiated the preparation of the Roadmap for the Phase-out of Incandescent Lamps in China in 2011, China’s SSL industry has begun to enter a period of rapid growth, and various regions have introduced targeted SSL industry policies (industry planning, standard systems, fiscal and tax policies, and technology roadmaps) [
34]. Although the timing of the introduction of different policies in different regions is different, the years when the relevant policies were issued specifically for the SSL industry were concentrated in 2010. Therefore, without loss of generality, this section sets 2010 as the starting year for the implementation of SSL industry policies.
From the perspective of policy implementation cities, currently, China has formed seven national SSL engineering industrialization bases in cities of Shanghai, Dalian, Nanchang, Xiamen, Shenzhen, Yangzhou and Shijiazhuang. In addition, the Yangtze River Delta, Pearl River Delta, Fujian Delta and northern regions have become gathering places for the development of China’s LED industry [
35]. In the above-mentioned areas, the policy system (including industrial development planning, standard system, fiscal and taxation policies) specifically for the SSL industry is relatively complete, and policy support is relatively strong. In other regions, there are fewer relevant policies for the SSL industry, especially the lack of targeted SSL industry policies, and the policy effects are not significant.
Based on national statistical yearbooks, local statistical yearbooks, and annual statistical bulletins of local governments, we extract data on the economic and social development of cities with a population of more than 1 million in China from 2008 to 2019, and set the exclusion criteria:
Cities lacking a large-scale SSL industry; cities with incomplete annual data of variables; cities with outliers; considering that the overall development of the SSL industry in the western region is relatively weak, in order to avoid overestimating the impact of the SSL industry, the western region is not included.
Finally, 85 sample cities were selected. This paper identifies cities with perfect or complete SSL industry policies based on the standards in the Research Report on the Development Policy Framework of Green Lighting Industry issued by China SSL Promotion Project [
26]. According to the report, cities with the perfect or complete SSL industry policies refer to the cities that established targeted regional SSL industry development plans, strictly implemented the SSL industry standard system, and have rich economic incentive policies on SSL industry. According to this standard, 25 of the 85 cities can be called cities with perfect or complete SSL industry policies (called policy implementation cities), while the SSL industry policies in other 60 cities cannot be called perfect or complete, and these cities are called the non-policy cities. The
Appendix Table A1 of this paper reports the list of all analyzed cities.