1 INTRODUCTION

Periodic health checkups (PHCs), or periodic/annual medical examinations, have been a fundamental part of public health policy for decades. PHCs aim to detect the presence of diseases and the risk factors for diseases with the purpose of reducing morbidity and mortality.1 Notably, workplace PHCs have been considered a key element for occupational health and workers' well-being. However, the preventive effect of PHCs on health deterioration remains controversial. With regard to its impact on mortality, longitudinal studies have shown that participation in PHCs is associated with a lower mortality rate,2-4 whereas other studies have been skeptical of such benefit.5, 6 Recent systematic surveys, which focused on randomized controlled trials, concluded that PHCs tended to have little or no effect on the risk of mortality,1 which contrasts with the findings of previous review articles that demonstrated a beneficial effect of PHCs on the dispensation of subsequent medical care and more general health outcomes.7, 8

A challenging issue in examining the impact of PHCs on health outcomes is the endogeneity or self-selection of participation in PHCs. Previous studies have shown that the likelihood of participation in PHCs is associated with many factors, including an individual's socioeconomic status (SES),9, 10 health literacy,11 and lifestyle12 in addition to sex, age, and other sociodemographic factors. Without controlling for these factors, the estimation results cannot be used to assess the effects of PHCs.

This endogenous issue should be addressed explicitly, especially in Japan. The Japanese Industrial Safety and Health Act obliges each firm, regardless of its size, to provide PHCs to its employees13; non-regular employees are often exempted from this practice. Meanwhile, self-employed workers, family workers, and dependent spouses of employees are encouraged to participate in PHCs provided by the municipalities, but their participation is voluntary. Accordingly, regular full-time employees are more likely to participate in PHCs than are precarious employees and other individuals.14, 15 Occupational status is closely related to other SES factors, such as educational attainment and income, and is also associated with lifestyle and health behavior such as smoking, alcohol consumption, and physical activity.16-18 Hence, participation in PHCs may serve as a proxy for an individual's SES and its related factors. If this is the case, any observed correlation between PHCs and health may be substantially confounded by them.

In this study, we aimed to examine the impact of PHCs on health outcomes among middle-aged individuals using a population-based, fourteen-wave survey conducted in Japan. This study is expected to provide new insights into the relevance of PHCs in two ways.

First, it focused on whether, and the extent to which, PHCs would delay the onset of inpatient care for non-communicable diseases (NCDs), unlike most previous studies that focused on the mortality rate, which likely matters later in life. Middle-aged adults are likely to be exposed to increasing risks of NCDs,19 which may lead to disparities in health in older adults. One previous study examined the association between PHCs and the risk of NCDs, but it was a cross-sectional study targeting adults with disabilities.20 We further considered the impact of PHCs on self-rated health (SRH), which represents general health conditions21, 22 and activities of daily living (ADL), thus affecting the health-related quality of life.23 We examined whether, and to what extent, participation in PHCs would postpone the onset of poor SRH and ADL problems.

Second, this study explicitly assessed how the endogeneity of PHCs influences the evaluation of their effectiveness. Because it is difficult to conduct randomized controlled trials, two approaches were applied to address this issue. First, we controlled for potential confounders when estimating the Cox-proportional hazards model. Second, we applied the propensity score matching (PSM) method to these hazard models, regressing on the sample matched sets of PHC participants and non-participants who had similar propensity scores.24 We compared the estimation results with those of two previous studies3, 4 that used PSM methods to examine the impact of PHCs in different data settings.

In these statistical analyses, we aimed to evaluate the experience of successive participation in PHCs over the first three waves to obtain a more reliable estimate of the impact of participation in PHCs on health outcomes.

2 METHODS 2.1 Study sample

In this study, we used data obtained from a nationwide fourteen-wave panel survey, “The Longitudinal Survey of Middle-Aged and Elderly Persons,25” conducted by the Japanese Ministry of Health, Labour and Welfare (MHLW) every year from 2005 to 2016. Japan's Statistics Law required the survey to be reviewed from statistical, legal, ethical, and other viewpoints. We obtained survey data from the MHLW with its official permission; therefore, the current study did not require ethical approval.

The first wave of the survey was conducted among individuals aged 50–59 years (born between 1946 and 1955). A total of 34 240 individuals responded (response rate: 83.8%). The second to 14th waves of the survey were conducted each year from 2006 to 2018, and 20,677 individuals were retained in the fourteenth wave of the survey.

2.2 Inclusion and exclusion criteria

We focused on the respondents who remained at least until the third wave, in order to know their experience in PHC participation over the first three waves. Then, we removed respondents who had already started receiving inpatient care in the first three waves for each NCD, because we could not identify the timing of the onset of inpatient care. For poor SRH and ADL problems, we similarly removed respondents who had already received inpatient care at least once in the first three waves.

In total, the longitudinal data of 29 770 respondents (15 399 men and 14 371 women) were used in this study. The number of respondents used in statistical analysis for each health outcome ranged from 15 467 (dyslipidemia) to 29 369 (ADL problems) (see Table 2 for more detail). We divided respondents into two groups: PHC participants (who participated in PHCs successively over the first three waves) and non-participants (others).

2.3 Periodic health checkups

During the survey, the respondents were asked whether they had participated in PHCs—including the “Ningen Dock” (the comprehensive health checkup system)—in the previous year. We constructed a binary variable for PHCs by allocating 1 to respondents who participated in PHCs successively over the first three waves (2005–2007) and 0 to others.

2.4 Health outcomes

We considered the onset of inpatient care for five types of NCDs (diabetes, heart disease, stroke, hypertension, and dyslipidemia) as well as SRH and ADL. We focused on the onset of inpatient care rather than the initial diagnosis or onset of outpatient care, despite the availability of both data in the survey. This is because the initial diagnosis and the onset of outpatient care may represent detection of illness or the start of medical intervention induced by the reported PHC results rather than their health outcomes.26 We defined the timing of the onset of inpatient care as the wave in which the respondents answered at the first time in the survey that they had an experience of hospitalization over the past one year. Thus, the onset of inpatient care in this study was based on the respondents' self-report. Regarding SRH, the respondents were asked to rate their current health condition as follows: 1 (very good), 2 (good), 3 (somewhat good), 4 (somewhat poor), 5 (poor), or 6 (very poor). We constructed a binary variable for poor SRH by allocating 1 to those who chose 4–6 and 0 to the rest of the respondents. We also constructed a binary variable for ADL problems by allocating 1 to those who answered that they needed assistance in at least one of the 10 ADLs (such as walking, getting in and out of bed, and getting into and out of a chair).

2.5 Potential confounders

As potential confounders, we considered (i) educational attainment, household spending, and occupational status as SES factors; (ii) smoking, alcohol consumption, and no physical activity as health behavior; and (iii) age, all of which were evaluated in the third wave. For educational attainment, we constructed binary variables for graduating from junior high school, high school, junior college, college, or above. We also merged the respondents who graduated from other schools and those who did not respond to the questions in one group and constructed a binary variable for them. Household spending was adjusted for household size by dividing it by the square root of the number of household members. We categorized them into quartiles and constructed binary variables for each quartile. For respondents who did not answer questions about household spending, we allocated a binary variable to unanswered questions. For occupational status, we constructed binary variables for managers, regular employees, non-regular (or precarious) employees (such as part-time, temporary, and contract workers), others, and not working. Moreover, we constructed binary variables for smoking, alcohol consumption, and no physical activity by allocating 1 to respondents who answered that they were currently smoking, consuming alcohol almost every day, and not performing physical activity, respectively, and 0 to the rest of the respondents. Finally, we constructed binary variables for each age group in the third wave.

2.6 Descriptive analysis

For descriptive analysis, we compared SES and health behavior between PHC participants and non-participants, and then compared the prevalence of the onset of inpatient care for each NCD, poor SRH, and ADL problems in the third and fourteenth waves between the two groups.

2.7 Propensity score matching

For PSM, we initially computed the propensity scores by estimating the logistic regression model to explain the PHCs based on a respondent's SES and health behavior. Then, we conducted simple nearest-neighbor matching with one neighbor without caliper.24 We matched each PHC participant with a non-participant whose propensity score was closest to that of the participant. Some non-participants may have had two or more matching participants, while others may have had no matches and were excluded from the analysis. We counted the number of matches for each non-participant and used it as the frequency weight in estimating Cox-proportional hazard model (Model 3), which will be explained below. We allocated one as a frequency weight to each PHC participant.

2.8 Cox-proportional hazard model analysis

In the regression analysis, we estimated three Cox-proportional hazards models (Models 1–3) to compute the hazard ratio (HR), with its corresponding 95% confidence interval (CI), for the onset of inpatient care for each NCD, poor SRH, and ADL problems in men and women. We defined the duration of follow-up as the length between the third and follow-up waves; for example, the fifth wave corresponded to the duration of two years.

Model 1 estimated the HR for each health outcome, adjusted only for baseline age. Model 2 additionally controlled for SES and health behavior. Model 3 replaced the original sample with the sample matched by PSM using the number of matches as the frequency weight, in addition to controlling for baseline age, SES factors, and health behavior. In this model, we also estimated the robust variances that accounted for clustering within the matched sets.27 The Stata software package (Release 17) was used for all statistical analyses.

We conducted two supplementary analyses. First, we examined how the results depended on the definition of PHC participation. To this end, we redefined PHC participants as those who underwent health checkups at least once over the first three waves and checked the robustness of the estimation results. Second, in addition to separate analyses for men and women, we directly examined sex differences using the entire sample. Specifically, we included a binary variable for women and its interaction term with PHC in regression models and examined their statistical significance.

3 RESULTS 3.1 Descriptive analysis

Table 1 summarizes the key features of PHC participants and non-participants evaluated in the third wave, showing that 55.9% of the study sample were PHC participants. For each item in the category, we examined the presence of bias toward PHCs. Initially, we computed the proportions of PHC participants and non-participants for each category. As seen in this table, men, regular employees, higher educational attainment, and higher household spending tended to be associated with PHCs, while women, self-employed, not working, lower educational attainment, and lower household spending tended to be associated with no PHCs. None of the three types of unhealthy behavior had any bias.

TABLE 1. Key baseline features of periodic health checkup participants and non-participants PHC participants Non-participants All Sex Men 8907 (62.0%) 5464 (38.0%) 14 371 (100%) Women 7720 (50.1%) 7679 (49.9%) 15 399 (100%) Occupational status Manager 939 (63.8%) 532 (36.2%) 1471 (100%) Regular employee 7302 (78.6%) 1993 (21.4%) 9295 (100%) Non-regular employee 3640 (56.1%) 2843 (43.9%) 6483 (100%) Self-employed 1326 (36.1%) 2349 (63.9%) 3675 (100%) Other 952 (40.5%) 1399 (59.5%) 2351 (100%) Not working 2468 (38.0%) 4027 (62.0%) 6495 (100%) Educational attainment Junior high school 2618 (47.1%) 2935 (52.9%) 5553 (100%) High school 9549 (56.0%) 7498 (44.0%) 17 047 (100%) Junior college 1196 (55.4%) 963 (44.6%) 2159 (100%) College or above 3090 (66.5%) 1556 (33.5%) 4646 (100%) Other or unanswered 174 (47.7%) 191 (52.3%) 365 (100%) Household income 1st quartile 3111 (50.6%) 3200 (52.0%) 6148 (100%) 2nd quartile 3951 (57.8%) 2991 (43.8%) 6836 (100%) 3rd quartile 3714 (61.2%) 2432 (40.1%) 6064 (100%) 4th quartile 4492 (61.9%) 2857 (39.4%) 7259 (100%) Unanswered 1359 (49.5%) 1663 (60.6%) 2743 (100%) Health behavior Smoking 4357 (53.4%) 3800 (46.6%) 8157 (100%) Alcohol consumption 5371 (60.0%) 3588 (40.0%) 8959 (100%) No physical activity 15 841 (56.1%) 12 374 (43.9%) 28 215 (100%) Age at baseline (years) M 54.7 (SD 2.7) M 54.7 (SD 2.7) M 54.7 (SD 2.7) N 16 627 (55.9%) 13 143 (44.1%) 29 770 (100%)

Table 2 compares the onset of inpatient care for each NCD as well as poor SRH and ADL problems between PHC participants and non-participants in the fourth to fourteenth waves among men and women. We did not adjust for other factors and ignored the differences in the timing of events. In men, PHCs were negatively associated with the risks of diabetes, stroke, hypertension, poor SRH, and ADL problems and were positively (albeit p > .05) associated with dyslipidemia. In women, no association was observed except for the risk of problems in performing ADLs, which was negatively associated with PHCs.

TABLE 2. Medical care experiences over the fourth and fourteenth waves by participants and non-participants in periodic health checkups over the first three waves All PHC participants Non-participants Difference in proportion N N Onsets Proportion (A) N Onsets Proportion (B) (B) − (A) SE p Men Diabetes 8796 5963 175 (2.9%) 2833 108 (3.8%) 0.9% (0.4%) .029 Heart disease 8433 5741 276 (4.8%) 2692 137 (5.1%) 0.3% (0.5%) .577 Stroke 8241 5624 143 (2.5%) 2617 102 (3.9%) 1.4% (0.4%) <.001 Hypertension 8246 5622 179 (3.2%) 2624 109 (4.2%) 1.0% (0.5%) .023 Dyslipidemia 8091 5505 53 (1.0%) 2586 15 (0.6%) −0.4% (0.2%) .079 Cancer 7754 5245 437 (8.3%) 2509 168 (6.7%) −1.6% (0.6%) .012 Poor SRHa 12 161 7712 979 (12.7%) 4449 677 (15.2%) 2.5% (0.7%) <.001 ADLb problems 14 204 8846 265 (3.0%) 5358 201 (3.8%) 0.8% (0.3%) .014 Women Diabetes 8183 4563 66 (1.4%) 3620 66 (1.8%) 0.4% (0.3%) .179 Heart disease 7849 4394 65 (1.5%) 3455 52 (1.5%) 0.0% (0.3%) .926 Stroke 7664 4297 60 (1.4%) 3367 48 (1.4%) 0.0% (0.3%) .914 Hypertension 7610 4266 96 (2.3%) 3344 81 (2.4%) 0.2% (0.3%) .622 Dyslipidemia 7376 4160 43 (1.0%) 3216 27 (0.8%) −0.2% (0.2%) .394 Cancer 7028 3941 231 (5.9%) 3087 175 (5.7%) −0.2% (0.6%) .731 Poor SRH 13 226 6770 840 (12.4%) 6456 833 (12.9%) 0.5% (0.6%) .392 ADL problems 15 165 7635 316 (4.1%) 7530 368 (4.9%) 0.7% (0.3%) .026 a Self-rated health. b Activities of daily living. 3.2 Propensity score matching

To explain the PHCs to compute the propensity scores, we estimated the logistic regression models. Table 3 summarizes the estimation results of the logistic regression models, showing that PHCs were negatively associated with occupational status other than regular employees, lower educational attainment, and lower household spending for both men and women. PHCs were also negatively associated with smoking in both men and women. The results in Table 3 suggest that a higher propensity score corresponded to higher SES and non-smoking status.

TABLE 3. Estimation results of logistic regression models to explain the successive participation in periodic health checkups over the first three waves Men Women HR (95% CI) HR (95% CI) Occupational status Regular employee 1 1 Manager 0.48* (0.42, 0.55) 0.30* (0.23, 0.38) Non-regular employee 0.37* (0.33, 0.42) 0.37* (0.33, 0.41) Self-employed 0.15* (0.14, 0.17) 0.17* (0.14, 0.20) Other job 0.22* (0.18, 0.27) 0.19* (0.16, 0.22) Not working 0.17* (0.15, 0.20) 0.17* (0.15, 0.19) Educational attainment College or above 1 1 Junior high school 0.62* (0.55, 0.70) 0.73* (0.62, 0.85) High school 0.91*** (0.83, 1.00) 0.84*** (0.73, 0.97) Junior college 0.93 (0.72, 1.20) 0.92 (0.78, 1.08) Other 0.75 (0.55, 1.03) 0.60** (0.42, 0.85) Household spending 4th quartile [highest] 1 1 1st quartile 0.65* (0.58, 0.73) 0.82* (0.74, 0.90) 2nd quartile 0.80* (0.72, 0.90) 0.96 (0.87, 1.06) 3rd quartile 0.93 (0.83, 1.04) 0.99 (0.90, 1.10) Unanswered 0.59* (0.51, 0.68) 0.68* (0.60, 0.77) Health behavior Smoking 0.73* (0.68, 0.79)