This study used data obtained from a population-based, nationwide internet survey conducted as a research project of the Cabinet Office (CAO) of the Japanese government in October 2019 and again in February 2020. The survey was conducted in accordance with Japan’s Statistics Law, which governs the statistical, legal, ethical, and other rules for surveys conducted by the government. Informed consent was obtained from all participants. We obtained the data of the survey with the permission of the CAO; therefore, ethics approval was not required for the current study.
We distributed the questionnaires to the registrants of an internet survey company. We planned to collect data from approximately 15,000 participants: around 10,000 from the survey in 2019 and the remaining 5,000 from the survey in 2020. We divided the targeted sample into two groups. First, we distributed 11,280 registrants equally between each of the 47 prefectures, between men and women, and among five age groups (aged 15–24, 25–34, 35–44, 45–59, and 60+). Thus, each prefecture-gender-age group consisted of 24 individuals. Next, we allocated the remaining 4,245 registrants to each gender-age group in each prefecture in proportion to each prefecture’s actual population size. When we closed the survey, we had obtained data from 15,574 participants. It should be noted that this construction of the dataset made the residents living in the metropolitan areas underweighted compared to the actual population.
We used municipality, which is the basic unit of local administration in Japan, for a unit of area. From a total of 1,741, the current study collected data from 1,273 municipalities with the number of participants ranging from 1 to 257 (mean [M] 12.2 and standard deviation [SD] 25.4). Excluding data from municipalities with less than ten participants and also participant data missing key variables, we used data from 12,461 participants (6,157 men and 6,304 women) living in 366 municipalities, where the number of participants from each municipality ranged between 1 and 257 (M 34.0 and SD 38.5).
Regarding SRH, the survey asked participants to answer questions about their health condition by selecting one of the following possible answers: good, somewhat good, average, somewhat poor, and poor. We constructed a continuous variable of SRH by allocating 1 to poor and 5 to good (the higher, the better). The survey also asked the participants to score their LS on an 11-point scale (0 = not satisfied at all to 10 = highly satisfied). Correspondingly, we constructed a continuous variable of LS (the higher, the more satisfied).
We also considered health behavior and interactions with others. Regarding health behavior, the survey asked the participants whether they were usually doing the following for their health: 1) eating a balanced diet, 2) exercising, 3) getting enough sleep, 4) refraining from smoking, 5) refraining from excess alcohol consumption, 6) avoiding the accumulation of stress, 7) going for regular checkups, 8) doing nothing in particular, and 9) other. We constructed a continuous variable of health behavior by adding up the number of chosen items from 1–7. As for interactions with others, the survey asked participants how often they interacted with friends or others. We constructed a continuous variable of interactions with others by allocating 7, 7/2, 1, 2.5/4, 1/4, 1/16, 1/48, and 0 to almost every day, three or four times a week, once a week, twice or thrice a month, sometimes a year, once a year, and no one to interact with, respectively.
We considered individual-level deprivation in terms of income, education (graduated from junior high school only), and job status (unemployed). Regarding income, we adjusted the reported amount by household size by dividing household income by the square root of the number of household members. We subsequently defined income poverty as an adjusted household that fell below the official poverty line, which is 1.22 million JPY at 2015 price, as defined by the Japanese Ministry of Health, Labor and Welfare (MHLW) . As individual-level covariates, we considered gender, age group (29 or below, 30s, 40s, 50s, and 60 or above), marital status (having a spouse or not), and survey years (2019 or 2020).
At the municipality level, we selected seven indicators: 1) the percentage of unemployed persons, 2) the percentage of persons who had an educational attainment of college or above, 3) taxable income per capita, 4) the percentage of owned houses, 5) the percentage of households with floor space per capita that was below the minimum level (defined by the Ministry of Land, Infrastructure, Transport and Tourism ); 6) the percentage of single-parent households, and 7) the percentage of aged single households. The choice of indicators largely followed those of preceding studies [4, 21–23]. Specifically, the indicators 1) to 4) correspond to general socioeconomic conditions, 5) represents the extent of household overcrowding, and 6) and 7) represent prevalence of vulnerable groups. We downloaded this data from the website database for municipality-level socioeconomic indicators, which are based on government surveys conducted around the year 2015. This official database is provided by the Ministry of Internal Affairs and Communications .
We computed the municipality-level deprivation indices in two ways. First, we conducted the z-scoring method by summing the z-scores of each indicator. In this calculation, we revered the signs of taxable income per capita and the percentage of owned houses, both of which were expected to make a negative contribution to municipality-level deprivation. Second, we conducted a principal component analysis (PCA) and selected components for which the eigenvalue of the correlation matrix was more than one as significant dimensions. Two or three principal components were expected to be obtained as significant dimensions. We calculated scores based on each of these components as well as their sum as deprivation indices. We categorized each deprivation index into tertiles: low, moderate, and high.
For the statistical analysis, we estimated multilevel linear regression models to explain SRH/LS scores by binary variables of moderate and high levels of each deprivation index (using low deprivation as a reference) along with individual-level deprivation and covariates. We compared the results across different methods of constructing the deprivation index.
We then conducted a multilevel mediation analysis to examine whether and to what extent health behavior and interactions with others mediated the association between municipality-level deprivation and SRH/LS. In the case of health behavior for SRH, we estimated a structural equation model consisting of 1) a model to explain health behavior by deprivation, 2) a model to explain interactions with others by deprivation, and 3) a model to explain SRH by deprivation, health behavior, and interactions with others, and then calculated the mediation effects of health behavior and interactions with others based on the estimated parameters. Denoting the scores of health behavior, interactions with others, and SRH as HBEHAV, INTERACT, and SRH, respectively, and the binary variables of moderate and high deprivation as Moderate and High, respectively, we estimated the structural equation model for individual j living in municipality i:
HBEHAVij = β11Moderarei + β12Highi + (individual-level variables) + u1i + ε1ij,
INTERACTij = β21Moderatei + β22Highi + (individual-level variables) + u2i + ε2ij,
SRHij = β31Moderatei + β32Highi + γ1HBEHAVij + γ2 INTERACTij
+ (individual-level variables) + u3i + ε3ij,
where u denotes municipality-level fixed effects, and ε an error.
Based on the estimated regression coefficients, we calculated the mediating effects. For the impact of living in municipalities with high deprivation on SRH, we derived the mediating effects of health behavior and interactions with others as β12γ1 and β22γ2, respectively, using participants living in the municipalities. If the mediating effect was found significant, we further computed its proportion out of the entire impact of living in highly deprived municipalities on SRH as β12γ1/(β32 + β12γ1 + β22γ2) × 100 % in the case of health behavior. Similarly, we computed the mediating effects of living in moderately deprived municipalities. We also repeated the same estimation procedure for LS. We used the software package Stata (Release 16) for the statistical analysis. Finally, we repeated a similar analysis by replacing the variable of overall health behavior with a binary variable of each of seven types of health behavior and examined their relative importance for self-rated health and life satisfaction.