3.2. Data Sources and Sample
The data used in this study were obtained from a survey conducted by Tsinghua University. Survey teams commissioned by the University were dispatched in 2009 to conduct interviews in 17 villages in Deyang and Mianyang, that were classified by the government as the most heavily damaged (Jizhongzaiqu). The 17 villages were selected based on convenient sampling. As transportation was heavily affected after the earthquake, the survey teams could only visit the villages that were accessible by car. The survey was conducted between January and July 2009, with most of the households interviewed in July.
Forty households in each village were randomly selected based on a full list of households from the village committees. Hence, 680 households were sampled in this survey. A questionnaire was designed to collect the data. Information on the degree of earthquake damage, amount of aid received, a variety of post-earthquake assistance, and household socio-economic status were recorded in the survey.
Face-to-face interviews were conducted with every respondent by the survey team members who had been trained before the interviews. The respondent was randomly selected based on the Kish Table. Finally, 558 of the 680 households were interviewed successfully, meaning that the response rate of valid samples was 82.06%.
Table 1 presents the description of the sample. In
Table 1, most of household heads were male, accounting for 95.10% of the sample. The average age of the household head was 50.81 in 2009, and they had 6 years of schooling on average, having finished only elementary education. In terms of marriage, 87.07% of the household heads were married. The mean of household size was 3.20. A total of 5% of the sample households were classified by government as impoverished, and 6% were under the protection of wubao/dibao, a form of income protection.
3.3. Variables and Measurements
3.3.1. Dependent Variable
The dependent variable was house reconstruction. Among all the elements of disaster recovery, a house is a basic need for survivors and most important to a survivor’s life; thus, house reconstruction was selected as the dependent variable. In the questionnaire, the respondent was asked whether their house had been reconstructed successfully when the survey was conducted. This variable was treated as a dummy variable with yes = 1 and no = 0.
3.3.2. Independent Variable
We focused on social capital as the independent variable in this paper. Our study used three different measures of social capital:
- (1)
Social network: Size of Spring Festival network, defined as the number of people with whom a household interacted during Spring Festival in 2008. Spring Festival marks the beginning of the Chinese calendar and in 2008, the Spring Festival started on 7 February 2008, which is before the occurrence of Sichuan earthquake. Note here that those who lost their lives in the Wenchuan earthquake were not included in the network to reflect the social capital after earthquake. This variable is continuous.
- (2)
Government officials: Number of town government officials and village cadres with whom the households have close connections, which is also a continuous variable
- (3)
Party Membership: Communist party membership of the household head, which is a dummy variable with Party membership = 1 and otherwise = 0.
3.3.3. Variables as Channels
Four variables were treated as channels: (1) the amount of government aid received for housing reconstruction, which is continuous; (2) knowledge of the government aid program, which is an ordinal variable with no knowledge = 1, almost have no knowledge = 2, some knowledge = 3, and considerable knowledge = 4; (3) a dummy with yes = 1 and no = 0 indicating whether the households received support to build temporary housing; and (4) a continuous variable showing number of people providing monetary and material support.
Among all four channels, government aid obtained was the most important contributing factor to housing recovery. To further investigate how social capital channels increased government aid to the household, we examined how the monetary and material support received, and the knowledge about the government aid program, were correlated with the government aid obtained.
3.3.4. Control Variables
The control variables can be broadly classified into four categories: (1) Human capital was measured by years of schooling of the household head and number of household members possessing technical licenses, such as chef, plumber, or electrician licenses, which are both continuous. (2) Household wealth was estimated by the size of farmland which is a continuous variable and orchard ownership which is a dummy with yes = 1 and no = 0. Ideally, we could have been able to control for pre-earthquake annual household income and total household asset, which can be highly correlated with both social capital and status of housing reconstruction. Unfortunately, this information was not collected in the survey. (3) Socio-economic status, a dummy, was valued 1 if household was impoverished as classified by the government. Another measure of socio-economic status is a dummy of safety net protection (wubao/dibao), which was assigned a value of 1 if the household is under the protection of the safety net. (4) Some other control variables are considered, such as the size of household, gender, age, age-squared, marital status of household head, and status of residential registration (hukou). The survey also asked the households about the degree of housing damage and any incidence of mortality, which allowed us to estimate the effect of earthquake damage on housing reconstruction.
3.4. Model Specification
To begin, we only retained the households that reported the need to rebuild houses after the earthquake (95% of the overall sample). After refining our sample, we assigned a value of 1 if a household has a house reconstructed by the time of the interview, or 0 otherwise.
We used the simple ordinary least squares (OLS) model in all of our specifications and the primary dependent variable was a housing reconstruction dummy. Normally, as the dependent variable was a 0 or 1 dummy, the logit/probit estimation was warranted. However, we used the OLS model instead, mainly based on the considerations as follows.
At the beginning, we found that, in three of the sample villages, none of the households interviewed managed to reconstruct a new house, which means the responses to the dependent variable were all 0 in the three villages. Under this circumstance, logit/probit estimation would eliminate all the sample households in those three villages from the estimation after controlling for the village fixed effect, which would result in sample loss. As the number of respondents in those three sample villages was 90, reaching 16.13% of all the respondents, directly omitting such a large proportion of respondents would be inappropriate. Secondly, the effect of village on the dependent variable was quite significant. As a result, the village fixed effect should be controlled. Thus, comparing the disadvantages with the advantages, we finally decided to use the OLS model.
In our study, the measures of the Spring Festival network in 2008 and status of communist party membership of the household head were established before the earthquake. In addition, not all but most of the connections with the government officials were formed before the earthquake. Given that the interview occurred after the earthquake, our data could be subject to recall error. However, assuming the recall error should be random is reasonable, which will mitigate the magnitude of the estimated effects.
The regression estimates the impact of pre-earthquake social capital on post-earthquake recovery. Some confounding factors that can affect both household social capital and housing reconstruction may exist, e.g., household wealth and socio-economic status. To address this issue, we controlled for a list of control variables as shown above. Finally, we included the village fixed effect to account for any time-invariant village unobserved characteristics.
Since the application and allocation of housing subsidies and all other disaster recovery programs were administered at the village level, it is highly plausible that the error term in our specification was subject to arbitrary correlation within a village. Clustering errors at the village level did not work well as the samples only included 17 villages. Hence, we used block bootstrapping as suggested by Bertrand et al. [
38] to address the small cluster number issue. Standard robust clustering provides similar results, though the significance of the effects of social network on housing reconstruction and government aid dropped from 5% to 10%. Without clustering, the robust standard errors were even larger, which made the effect of social network on housing reconstruction insignificant, but the impact on housing subsidies received remained significant. The results suggest that, once controlling for the village fixed effect and other covariates, a negative intra-correlation existed among households in the same village for housing recovery. This may suggest that competition exists among village households for government resources for housing reconstruction. The results based on standard robust, both with and without clustering, are available upon request. Block bootstrapping may still lead to inconsistent estimates of standard error given the small cluster number. We also used wild bootstrapping method suggested by Cameron et al. [
39]. However, we needed to drop the village fixed effect from the regression equation as the method does not allow for fixed effect estimation. The overall results obtained were similar, even though the correlation between housing aid obtained and monetary support received from the social network became insignificant. The results are available upon request.