On the Direct and Indirect Effects of the Great East Japan Earthquake on Self Rated Health through Social Connections: Mediation Analysis

Abstract

The Great East Japan Earthquake created health hazards for many people. Using Panel Data gathered in Iwanuma city, Japan, at two points in time (in 2010 before the quake, and in 2013 after the quake), we found that the high degree of housing damage negatively affected victims’ self rated health (SRH) (direct effect), and decreased the levels of their social connections, which in turn also had a harmful effect on their SRH (indirect effect). We also found that although the direct impacts of earthquakes disappear relatively quickly, the harmful indirect effects associated with a decrease in social connections are slower to dissipate. We conducted a first-difference two-step GMM estimation to consider the possible problem of endogeneity. The results support the above conclusion, and show that in the short-term, the indirect impacts of the earthquake accounted for 55% of all the impacts experienced.

1. Introduction

The Great East Japan Earthquake that occurred in March 2011 caused enormous damage. A Japanese National Police Agency report (2017) confirmed 15,894 deaths, 6,152 injured, and 2,562 people missing across twenty prefectures. People who were forced to change their residence because their homes had been burned or otherwise destroyed lost their connections with other people (Hikichi et al., 2017). Therefore, it can be concluded that the earthquake had both direct adverse health effects on its victims, and indirect negative effects due to reduced social connections (Hikichi, Aida, Tsuboya, Kondo, & Kawachi, 2016). In this paper, by analyzing self-rated health (SRH) outcomes, we empirically investigate the Great East Japan Earthquake’s direct and indirect impacts on houses.

We found that a high degree of damage to houses negatively affected victims’ SRH (direct effect), and decreased the level of their social connections, which precipitated harmful effects on their SRH (indirect effect). In this paper, we use the term social connections (hereafter SC) as a concept related to social capital. It has also been shown that although the direct impacts of earthquake damage disappear relatively quickly, the indirect effects of a decrease in social connections remain for some time. In other words, the direct health effects caused by earthquakes resolve in the short term, but the earthquake’s negative aftereffects, which include a reduction in social connections, may continue to be felt for a relatively long period of time. We conducted a sensitivity analysis of the robustness of the results noted above. The robustness check shows that the results of the mediation analysis are established only within a fairly limited range. This means that the correlation between the SC and SRH error terms must be less than 0.068, in order to allege these results. This suggests the need to conduct further verification analysis using methods that will prove a causal relationship. We then performed an Instrumental Variable (IV) estimation, to verify the robustness of the assumptions that are necessary to assure the validity of the estimated results of the mediation analysis.

The extent to which the houses were destroyed is a natural type of experiment, and we may be able to regard the damage as a random event that occurred irrespective of the individuals’ attributes. However, as for SC, it may be difficult to accept these suppositions, because the following counter-arguments must also be considered: Firstly, there may be the opposite causality that you can interact with people because you are fine (a reverse causality problem). Second, there may be a correlationship between the ability to actively interact with people and the ability to be healthy, so these variables are correlated (an omitted variables problem). If such unobserved factors exit, false observation results are obtained.

In this paper, we use geographical information describing the area in Iwanuma City where the respondents lived before the earthquake. That is, the distance from the coast ≒ the magnitude of the effects of the tsunami is used as the principal instrumental variable. In addition, we also use the type of ownership (such as owned or rented) and the type of building (such as single-family or apartment) as instrumental variables. The results of the IV estimation show that there are larger mediation effects, and that in the short-term, the indirect impacts of the earthquake made up about 55% of all the impacts. We also used these instrument variables to test whether the extent of housing damage and the SC variables are endogenous variables. The results show that the extent of housing damage is exogenous, and SC is endogenous to SRH.

In recent years, a large number of papers have examined the influences of social connections on health (Choi et al., 2014; De Silva, McKenzie, Harpham, & Huttly, 2005; Gargiulo & Benassi, 1999; Lund et al., 2010). In particular, after Kawachi Kennedy, Lochner, and Prothrow-Stith (1997) showed that social connections are related to overall mortality, the relationship between high levels of social connections and good health outcomes has become the subject of intensive research. Social connections can be captured at the individual or regional level. In this paper, we focused on the individual’s level of social connections, mainly because we assessed the earthquake’s impact in a small area, Iwanuma city, which has an area of 60.45 km².

It is difficult to enumerate all the studies that have been conducted on the relationship between social connections and health. This paper will introduce only Kawachi, Subramanian, and Kim (2008), who surveyed these studies. Several studies have examined the relationship between social connections and health in disaster situations. These include Ali, Farooq, Bhatti, and Kuroiwa (2012), Beaudoin, (2007, 2011), Beiser, Wiwa, and Adebajo (2010), Wind, Fordham, and Komproe (2011), Wind and Komproe (2012), Fergusson, Horwood, Boden, and Mulder (2014), Frankenberg et al. (2008), Kumar et al. (2007), and Van Griensven et al. (2006), but, with the exception of Fergusson et al. few studies have dealt with panel data gathered before and after the occurrence of natural disasters. In addition, Hikichi, Aida, Tsuboya, Kondo, and Kawachi (2016a), Hikichi et al. (2016b), and Tsuboya et al. (2016) used the data analyzed in this study in a previous study. This article adopts an original approach by considering the broad concept of SRH as outcomes, an analysis that has not been conducted elsewhere.

In this study, we use subjective variables based on recollections to ask questions about changes in health conditions before and after the disaster. This step is often considered a weak point in the research process. However, in a study examining the quality of life (QOL) this point can be advantageous, because subjective variables evaluate their own health based on some criteria present in their own minds. This also applies to subjective variables such as life satisfaction. After experiencing a huge influence such as the Great East Japan Earthquake, the health conditions of the people in the respondents’ circle have often deteriorated. Based on such circumstances, the respondents’ subjective health levels will have great value, even if their health levels have not changed. Also, if they become more aware of the value of health, they will be biased to report that their health levels have increased. In other words, by changing the evaluation criteria, it becomes difficult to evaluate changes in subjective variables even within the same individual. When asking respondents to recollect changes before and after the disaster, we may be able to obtain more accurate information, since the evaluation criteria have not been changed.

Several studies on life satisfaction and happiness have reported that changes in these evaluation criteria actually occurred as a result of the impacts of the Great East Japan Earthquake. Some papers report that this earthquake increased the level of happiness in Japan (Ishino, Ogaki, Kamesaka, & Murai, 2011; Yamamura, Tsutsui, Yamane, Yamane, & Powdthavee, 2015), or at least no statistically significant nation-wide drop in happiness has been observed after the disaster (Tiefenbach & Kohlbacher, 2015). We think the reason these results were observed is that the logic described above is working, because there is little objective evidence that social and economic conditions improved after the earthquake. Unless the evaluation criteria are unified when subjective evaluations are made, an accurate analysis cannot be conducted. Research on anchoring vignettes is underway, as a means of dealing with these problems (Grol-Prokopczyk, Freese, & Hauser, 2011; Salomon, Tandon, & Murray, 2004).

2. Data

The Japan Gerontological Evaluation Study (JAGES) is a nationwide cohort study initiated in 2010, by a research group established for the purpose of prospectively investigating personal and regional predictors of healthy aging. We used two of the JAGES cohort’s investigations conducted in 2010 (baseline) and 2013 (following the aftermath of the Great East Japan Earthquake) for this study. This study’s research profile has been described in detail earlier by Hikichi et al. (2016a; 2016b), and Tsuboya et al. (2016). Iwanuma City, with a total population of 4,4187 in 2010 and the site where this survey was conducted, is a coastal municipality in Miyagi prefecture, and the Great East Japan Earthquake and Tsunami occurred on March 11, 2011.

We conducted a survey of all residents aged 65 and above who were non-recipient of long-term care insurance in Iwanuma (August 2010) (n = 8576), and used the official residence register provided by the city office. In our survey, we investigated residents’ awareness of social unity in the community, and their personal characteristics. The baseline survey was completed in August 2010, before the earthquake. The response rate to the baseline survey was 59.0% (n = 5,058). This rate is slightly higher than the average for this type of community survey (Brick & Williams, 2013; Sinclair, O’Toole, Malawaraarachchi, & Leder, 2012).

The Great East Japan Earthquake and tsunami occurred seven months after the cohort baseline was established, and Iwanuma City is located about 80 km west of the epicenter of the earthquake. The tsunami killed 180 people (out of a total of 44,814 people), and 48% of the land area was flooded. A follow-up survey of victims was conducted about 2.5 years after the disaster (beginning in October 2013). Informed consent was obtained at the time of data collection. The response rate of the follow-up survey in 2013 was 82.1% (3594/4380).

The survey protocol was reviewed and approved by the human subjects committee of the Harvard T. H. Chan School of Public Health, as well as the human subjects committees of Tohoku University, Nihon Fukushi University, and Chiba University. The definitions of variables and descriptive statistics of the data used in this paper are provided in Table 1.

Table 1:

The Definitions of Variables and Summary Statistics

					2010						2013
Variable	Obs	Mean	Std. Dev.	Min	Max		Obs	Mean	Std. Dev.	Min	Max
[Outcome Variables]
Self Rated Health SRH1							3,500	2.77	0.57	1	5
Self Rated Health SRH2							3,504	2.87	0.58	1	5
Self Rated Health SRH3							3,499	2.86	0.63	1	5
[Mediation Variables]
Social Connections (Total : SC=SC1+・・・+SC6)	3,211	21.70	3.71	6	30		3,400	21.27	3.85	6	30
SC1 : How often do you see your friends?
1:Rarely, 6 : Almost everyday	3,452	3.73	1.49	1	6		3,518	3.63	1.57	1	6
SC2 : How many friends / acquaintances have you seen over the past month?
Number of friends who met, 1 : none, 5 : 10 or more persons	3,422	3.58	1.28	1	5		3,508	3.42	1.35	1	5
SC3 : Do you think that people living in your community can be trusted in general?
1 : Not at all, 5 : Very much	3,525	3.75	0.77	1	5		3,512	3.74	0.72	1	5
SC4 : Do you think people in your community try to help others in most situations?
1: Not at all, 5 : Very much	3,492	3.54	0.83	1	5		3,484	3.53	0.80	1	5
SC5 :How attached are you to the community in which you live?
1: Not at all, 5 : Very much	3,521	4.00	0.83	1	5		3,518	3.95	0.82	1	5
SC6 : What kind of relationship are you with local neighbors in the area?
1 : Nothing 2 : Minimal relationship with a greeting degree 3 : Socializing to the degree to talk standing on a daily basis 4 : Some cooperate with each other in terms of living, such as cooperating and lending and borrowing daily necessities	3,486	3.02	0.70	1	4		3,529	2.94	0.72	1	4
[Independent Variables]
<Degree of Overall Housing Damage>
Completely destroyed	3,466	0.05	0.21	0	1
Major : requires major repairs	3,466	0.04	0.19	0	1
Minor : habitable with repairs	3,466	0.07	0.26	0	1
Affected	3,466	0.43	0.50	0	1
No damage	3,466	0.41	0.49	0	1
DAMAGE : completely destroyed	3,466	0.05	0.21	0	1
<House ownership>
Owned house/condominium	3,599	0.89	0.32	0	1
Private rental apartment/housing	3,599	0.03	0.18	0	1
Public rental apertment/housing	3,599	0.04	0.20	0	1
Company providing apertment/housing	3,599	0.01	0.08	0	1
Others	3,599	0.03	0.17	0	1
<Type of housing structure>
Detached housing	3,476	0.93	0.25	0	1
Tenement, Others	3,476	0.01	0.11	0	1
Apartment building	3,476	0.05	0.23	0	1
<Family Configuration>
Living alone	3,508	0.10	0.30	0	1		3,514	0.14	0.35	0	1
Living with spouse	3,588	0.31	0.46	0	1		3,606	0.32	0.47	0	1
Living with others	3,588	0.59	0.49	0	1		3,606	0.54	0.50	0	1
<Residential Area>
Area1 : Mid Iwanuma area	3,606	0.57	0.49	0	1		3,567	0.60	0.49	0	1
Area2 : East Iwanuma area	3,606	0.16	0.36	0	1		3,567	0.12	0.33	0	1
Area3 : West Iwanuma area	3,606	0.27	0.44	0	1		3,567	0.27	0.45	0	1
<Other Variables>
Loss of family member	3,567	0.26	0.44	0	1		3,567	0.26	0.44	0	1
Equivalent household income(10000Yen)	2,940	228.32	140.62	8.84	1300		3,025	226.46	144.73	11.18	1300
State of work	3,169	0.18	0.38	0	1		3,451	0.13	0.34	0	1

・Equivalent household income is calculated as : Equivalent household income=household income/√(number of family members)

3. Causal/Sensitivity Mediation Analysis

3–1. Causal/Sensitivity Mediation Analysis

The Mediation Analysis Framework is shown in Figure 1 below.

Figure 1:

The Relationship Between Treatment, Mediation, and Output

In our analyses, these acronyms correspond to the following variables:

Y: State of Health (Self Rated Health)

M: Social Connections

T: Degree of Overall Housing Damage

SRH is composed of SRH1, SRH2, and SRH3. These are the responses to the questions asked about changes in respondents’ health conditions before and after the disaster. More specifically, SRH 1 is the response given to the following question: How had your health condition changed about 1 month after the earthquake (around April 2011) compared to how it was before the earthquake?

SRH2: How had your health condition changed about 1 year after the earthquake (around March 2012) compared to how it was before the earthquake?

SRH3: How has your current health condition changed compared to how it was before the earthquake?

Possible responses to the above questions were the following: 1. getting very bad, 2. getting somewhat worse, 3. did not change, 4. getting somewhat better, and 5. getting better.

As for the detailed descriptions of the social connections variables, see Table 1. The degree of overall housing damage shows the extent to which a dwelling has been damaged, the result of receiving approval by the administration. In this paper, we analyze responses using a dummy variable that takes 1 if it is judged that a house has been completely destroyed, and 0 otherwise.

Specifically, we estimate the following equations:

$$M_{it}={\alpha }_1\cdot t+{\alpha }_2\cdot T\cdot t+X_{it}{\alpha }_3+U_i{\alpha }_4+{\epsilon }_1$$ (1)

$$Y_{it}={\beta }_1\cdot t+{\beta }_2\cdot M_{it}+{\beta }_3\cdot T\cdot t+X_{it}{\beta }_5+U_i{\beta }_4+{\epsilon }_2$$ (2)

Subscript t takes 1 if year = 2013, and it takes 0 if year = 2010. U represents unobserved, time invariant factors. X is the set of explanatory variables: family configuration, experiencing the loss of a family member, equivalent household income, and work status. To eliminate the effects of unobserved factors, by differencing (1) and (2) respectively, we get the following equations:

$$\mathrm{\Delta }{M}_i=M_{it}-M_{it-1}={\alpha }_1+{\alpha }_2\cdot T+\mathrm{\Delta }{X}_i{\alpha }_3+\mathrm{\Delta }{\epsilon }_1$$ (3)

$$\mathrm{\Delta }{Y}_i=Y_{it}-Y_{it-1}={\beta }_1+{\beta }_2\cdot \mathrm{\Delta }{M}_i+{\beta }_3\cdot T+\mathrm{\Delta }{X}_i{\beta }_5+\mathrm{\Delta }{\epsilon }_2$$ (4)

We estimated equations (3) and (4) using the causal/sensitivity mediation analysis¹ developed by Imai, Keele, and Yamamoto (2010). To conduct mediation analysis and identify the effects of the related variables, it is necessary to satisfy the following assumptions:

Assumption 1 (Sequential Ignorability: Imai, Keele, and Yamamoto, 2010)

$$1-1.\{\text{Y}(\text{t},\text{m}),\text{M}(\text{t})\}\perp \text{T}\mid \text{X}=\text{x}$$

$$1-2.\text{Y}(\text{t},\text{m})\perp \text{M}(\text{t})\mid \text{T}=\text{t},\text{X}=\text{x}$$

where X, as we already noted, is a vector of the observed pretreatment confounders, 0 < Pr (T = t | X = x) and 0 < p(M = m | T = t, X = x) for t = 0 and 1, and all x and m are in support of X and M, respectively. Assumption 1–1 seems to be satisfied because quakes are natural incidents (natural experiments). We will investigate the validity of this assumption later.

Assumption 1–2 is problematic. It requires that the correlation (ρ = corr(Δε₁, Δε₂)) between Δε₁ and Δε₂ is 0. However, we cannot test this directly with observational data. This leads to the need to test the validity of the outcome using sensitivity analysis, carried out using a quasi-Bayesian Monte Carlo algorithm (King, Tomz, & Wittenberg, 2000).

The results are shown in Table 2. DAMAGE equals 1 when the house was “completely destroyed,” and it is otherwise 0. The coefficients of DAMAGE in columns (1), (2), and (3) represent the arrow B (direct effect) in Figure 1. The coefficient of social connections in columns (1), (2), and (3) represents the arrow C. The coefficient of DAMAGE in column (4) represents the arrow A. The sign of arrow A is significantly negative. The sign of arrow B is significantly negative, and the sign of arrow C is significantly positive. Arrow A times arrow C equals the mediation effect. Arrow B indicates the direct effect. The direct and indirect effects combined are the total effect. The mediation effect accounted for 4.3% to 13% of the total effect.

Table 2:

Mediation Analysis

	(1)	(2)	(3)	(4)
	SRH1	SRH2	SRH3	SC
DAMAGE	−0.432***	−0.195*	−0.00947	−1.740***
	(0.0914)	(0.110)	(0.114)	(0.496)
SC	0.0112***	0.00721*	0.0120***
	(0.00374)	(0.00409)	(0.00449)
Loss of family member	−0.0814***	−0.0967***	−0.0143	0.0855
	(0.0266)	(0.0289)	(0.0317)	(0.160)
Family configuration (base : living alone)
Living with spouse	0.0460	0.107	0.0414	−0.533
	(0.0640)	(0.0735)	(0.0706)	(0.381)
Living with other persons	0.0678	0.127*	0.0243	−0.242
	(0.0614)	(0.0708)	(0.0673)	(0.385)
Equivalent household income(Yen)	0.852	−0.330	0.308	4.188
	(0.944)	(1.084)	(1.127)	(6.280)
State of work	0.0301	0.0283	0.0189	−0.0595
	(0.0347)	(0.0396)	(0.0405)	(0.229)
Constant	2.830***	2.911***	2.874***	−0.281***
	(0.0125)	(0.0129)	(0.0143)	(0.0792)
Observations	2,136	2,139	2,134	2,134
R-squared	0.039	0.017	0.004	0.012
Average Causal Mediation Effect	−0.02	−0.01	−0.02
Direct Effect	−0.44	−0.20	−0.01
Total Effect	−0.46	−0.21	−0.03
% of Total Effect mediated	4.3%	5.9%	13%

Robust standard errors in parentheses :

^***p<0.01

^**p<0.05

^*p<0.1

Table 2 shows that the influence of DAMAGE decreases with the lapse of time. On the other hand, the influence of SC is also constant, even if time passes. As a result, the indirect effect through the SC increased by 4.3%, 5.9%, and 13% as a whole.

Figure 2 shows the robustness check result of Assumption 1–2 of column (1). The mediation effect is shown to be 0 when ρ reaches 0.0683. When SRH2 and SRH3 are the dependent variables, the values are 0.0415 and 0.0628 respectively. This cannot be said to be robust, and shows that it is necessary to further improve the accuracy of the verification of causality, and the influence on SRH caused by the rise in the SC level.

Figure 2:

Sensitivity Analysis of Column (1) in Table 2

3–2. Instrumental Variable and Two-step GMM Estimation

We now return to the issue of the validity of Assumptions 1–1 and 1–2. As discussed in the Introduction, we cannot deny the possibility that a confounding bias is occurring. Therefore, we employ the area dummy expressing the distance from the coastline as our instrument. Hikichi et al. (2016b) conducts instrumental variable analysis with almost the same idea. The area closest to the coast is the East Iwanuma area. Because the damage inflicted by the tsunami was terrible, there is a high probability that houses in this area were completely destroyed. There is no rational basis for reasoning that such a location correlates with the changing trend in SRH. Therefore, we adopt these areas as the principal instrument. In this paper, we also use the type of house ownership (such as owned or rented) and the type of building (such as single-family or apartment) as instrumental variables.

Housing ownership styles such as owned house/condominium or rented house/condominium have little influence on subsequent changes in health conditions. Also it is assumed that the structure of a detached house or an apartment will have a very limited effect on health conditions. On the other hand, these variables are thought to affect the level of DAMAGE and SC. For example, as for SC, if the type of house ownership is rental, the inhabitants are regarded as having a weak intention to continue living there, and they are relatively free to move. These circumstances weaken the incentive to build relationships with the neighboring people, and so will lead to a low SC level. As for DAMAGE, because condominiums are often more robust buildings than detached houses, it is assumed that those who live in condominiums suffer less damage than those living in detached houses.

The equations to estimate are equations (3) and (4). If the assumption of homoskedasticity in the error term holds, and there is no serial correlation between samples, the GMM estimator coincides with the 2SLS estimator. However, it is impossible to assume that there is no serial correlation among people. Also there is no rational basis to assume homoskedasticity. Therefore, using two-step GMM estimation, we can obtain a more efficient and consistent estimator, compared to 2SLS.

We begin by estimating that in the first stage both DAMAGE and SC are endogenous variables. The results of the first stage are shown in columns (1) and (2) of Table 3. As a result of the endogenous test in Table 4, we will show later that DAMAGE was judged to be an exogenous variable, and SC was judged to be an endogenous variable. Then, in column (3), the results of the first stage show SC as an endogenous variable and DAMAGE as an exogenous variable. We have to check whether or not the problem of weak instruments (Bound, Jaeger, & Baker, 1995; Staiger & Stock, 1994) occurs in the first stage. The results are shown in Table 3².

Table 3:

First Stage Estimation

	(1)	(2)	(3)
	DAMAGE	SC	SC
DAMAGE			−1.245**
			(0.605)
<IV>
Area (base : mid and west Iwanuma area)
East Iwanuma area	0.199***	−0.809***	−0.562**
	(0.0223)	(0.248)	(0.272)
House ownership (base : owned house/condominium)
Private rental apartment/housing	0.187***	−1.226**	−0.994**
	(0.0401)	(0.491)	(0.503)
Public rental apertment/housing	0.214***	0.234	0.500
	(0.0470)	(0.568)	(0.589)
Company providing apartment/housing	0.301***	0.421	0.796
	(0.108)	(1.013)	(1.050)
Others	0.388***	−0.357	0.126
	(0.0614)	(0.647)	(0.710)
Type of housing structure (base : detached housing)
Tenement, others	−0.224***	−1.411**	−1.689**
	(0.0464)	(0.668)	(0.687)
Apartment building	−0.152***	0.916*	0.727
	(0.0320)	(0.496)	(0.504)
<Other explanatory variables>
Loss of family member	0.0305***	0.143	0.181
	(0.00855)	(0.164)	(0.163)
Family configuration (base : living alone)
Living with spouse	0.00405	−0.705*	−0.700*
	(0.0196)	(0.364)	(0.362)
Living with other persons	−0.0242	−0.400	−0.430
	(0.0201)	(0.360)	(0.357)
Equivalent household income(Yen)	0.106	6.12	6.25
	(0.295)	(5.59)	(5.64)
State of work	0.00443	−0.111	−0.106
	(0.0128)	(0.229)	(0.228)
Constant	−0.0119***	−0.248***	−0.263***
	(0.00260)	(0.0825)	(0.0826)
Observations	2,126	2,126	2,126
F test of excluded instruments	22.42	5.38	4.27
P-value	0.000	0.000	0.000
R-squared	0.023	0.349	0.027

Robust standard errors in parentheses :

^***p<0.01

^**p<0.05

^*p<0.1

Table 4:

Two-step GMM Estimation

	(1)	(2)	(3)	(4)	(5)	(6)
Dependent variable	SRH1	SRH2	SRH3	SRH1	SRH2	SRH3
Endogenious variable	DAMAGE, SC	DAMAGE, SC	DAMAGE, SC	SC	SC	None
Estimation method	GMM	GMM	GMM	GMM	GMM	OLS
Endogeneity test of DAMAGE	0.2372	0.5616	0.5628
Endogeneity test of SC	0.0478	0.0223	0.2297	0.0176	0.0113
Hansen J statistic	0.1482	0.0763	0.5596	0.1789	0.1261
DAMAGE	−0.494**	−0.114	0.157	−0.279**	0.00115	−0.00426
	(0.205)	(0.237)	(0.207)	(0.136)	(0.155)	(0.114)
SC	0.0849**	0.0906**	0.0529	0.103**	0.101**	0.0132***
	(0.0409)	(0.0429)	(0.0332)	(0.0415)	(0.0410)	(0.00445)
Loss of family member	−0.0782**	−0.0956***	−0.0221	−0.0874***	−0.102***	−0.0147
	(0.0318)	(0.0345)	(0.0337)	(0.0325)	(0.0334)	(0.0318)
Family configuration (base : living alone)
Living with spouse	0.0681	0.138	0.0761	0.0812	0.144*	0.0499
	(0.0767)	(0.0867)	(0.0739)	(0.0803)	(0.0876)	(0.0704)
Living with other persons	0.0641	0.134*	0.0594	0.0781	0.141*	0.0387
	(0.0718)	(0.0813)	(0.0684)	(0.0748)	(0.0817)	(0.0669)
Equivalent household income(Yen)	0.745	−0.776	−0.379	0.702	−0.804	0.0222
	(1.1)	(1.2)	(1.16)	(1.16)	(1.21)	(1.12)
State of work	0.0306	0.0380	0.0216	0.0291	0.0350	0.0195
	(0.0398)	(0.0453)	(0.0408)	(0.0417)	(0.0460)	(0.0405)
Constant	2.859***	2.939***	2.886***	2.861***	2.941***	2.874***
	(0.0174)	(0.0183)	(0.0169)	(0.0184)	(0.0184)	(0.0144)
Observations	2,126	2,129	2,124	2,126	2,129	2,124
Mediation Effect				−0.347	−0.126	−0.016
Direct Effect	−0.494	−0.114	0.1 57	−0.279	0.001	−0.004
Total Effect				−0.626	−0.125	−0.021
				55.4%	99.1%	79.0%

Robust standard errors in parentheses :

^***p<0.01

^**p<0.05

^*p<0.1

The East Iwanuma area significantly affects both DAMAGE and SC. In the case of DAMAGE, when the living units were rentals, for example, they suffered extensive damage. This seems to have been due to the fact that the building owners had a stronger incentive to construct robust buildings and perform earthquake-resistant reinforcement work when they lived in the buildings themselves. In the case of rental housing, it is difficult to undertake earthquake-resistant reinforcement work while residents are living in the space, and rental income is lost while the work is being performed, so the economic incentive to do so is very weak. Generally speaking, in Japan, buildings built as rental units are often more vulnerable to earthquake damage than buildings built for sale. With respect to a building’s structure, it has been shown that collective housing is more resistant to earthquake damage than single-family houses.

Meanwhile, the level of SC declined if the respondents lived in private residential housing. This seems to be due to the tendency of people with a high level of commitment to the area to buy houses. Regarding a building’s structure, if it is neither a detached house nor an apartment, that is, a “tenement or other,” the level of SC is lowered significantly. It is thought that this is because these buildings are often suitable for temporary residences, and so are a type similar to a very simple inn, and it is thought that they will engender a low commitment to the community. We can posit that the problem of weak instruments does not arise from column (1), however, you should be aware that although columns (2) and (3) have a p-value of 0, the F value has a somewhat low value of 5.38 or 4.27.

Based on the results of the endogenous tests in columns (1), (2), and (3), we estimate columns (4), (5), and (6), respectively. Based on the results of columns (1) and (2), we estimate columns (4) and (5), with SC being treated as an endogenous variable, and DAMAGE as an exogenous variable. Because both SC and DAMAGE were judged to be exogenous variables in column (3), in column (6) we use both variables as exogenous, and the results of the OLS estimation are presented. From these estimation results, as Baron and Kenny (1986) show, the direct effect, mediation effect, and total effect are calculated. That is, the mediation effect is represented by the product of α₂ and β₂, that is, α₂β₂ in equations (3) and (4). As a result, the mediation effect accounted for about 55.4% of the total effect as for SRH1. The proportion of the mediation effect is largely higher than the results shown in Table 2. If the resident’s area at the time of the earthquake is taken as a reliable instrumental variable, the results of the tests of over-identifying restrictions are reliability tests for other instrumental variables.

In other words, the type of ownership and the type of building structure can be regarded as appropriate instrumental variables. The results of the tests of over-identifying restrictions are satisfactory, except for column (2). In Table 4, however, column (5) does not have this problem, so we have judged that it is not an important problem. Looking at Table 4, the strength of the influence’s indirect effect is clear. When the explained variable is SRH 3, the indirect effect’s proportion of the whole is about 80%.

4. Conclusion

In this research, we found that a high degree of housing damage not only affected victims’ SRH (direct effect) negatively, but it also decreased the level of social connections, which subsequently had a harmful effect on SRH (indirect effect). We also found that although the direct impact of earthquake damage disappears relatively quickly, the indirect effects of a decrease in social connections remain for a while. To consider the possible problem of endogeneity, we conducted a first-difference two-step GMM estimation. The results of this estimation support the above conclusion, and show that in the short-term, an earthquake’s indirect impacts make up about 55% of the total impacts.

These results indicate that from a long-term perspective, it is important to implement measures that will prevent a decrease in SC levels to minimize the total damage inflicted on people’s health by an earthquake disaster. In terms of long-term effects, it is likely that people who interact with others are more likely to have access to information that is beneficial to maintaining and improving their health, and that having someone close by to talk to may help protect their health.

Not only did we find that community ties that existed before the earthquake mitigated health problems, it was also shown that living in a community with rich ties protects the health of people who tend to be isolated and say, ‘Leave me alone’.

Thus, the importance of raising the level of social capital in the community even before the earthquake occurred was demonstrated. Activities such as encouraging people to participate in salons were found to be important for this purpose(Ichida et al.(2013)), and spreading these activities is considered to be necessary to reduce health hazards from disasters.

The following questions remain, however. How long will the indirect impact on SC be sustained? How will the results differ when analyzing a broader range of people? How robust are the results obtained in this paper, which used the recollection method? These questions provide scope for further research.

We appreciate the support and cooperation of the Iwanuma Mayor’s office, and the staff of the Department of Health and Welfare of Iwanuma city government. This work was supported by a grant from the National Institutes of Health (R01 AG042463); Grants-in-Aid for Scientific Research from the Japan Society for the Promotion of Science (KAKENHI 15H01972, KAKENHI 23243070, KAKENHI 22390400, KAKENHI 22592327 and KAKENHI 24390469); a Health Labour Sciences Research Grant from the Japanese Ministry of Health, Labour and Welfare (H22-Choju-Shitei-008, H24-Choju-Wakate-009 and H28-Chouju-Ippan-002); and a grant from the Strategic Research Foundation Grant-Aided Project for Private Universities from the Japanese Ministry of Education, Culture, Sports, Science and Technology (S0991035).