The effect of warning signals from health check-ups on modifiable lifestyle risk factors: evidence from mandatory health check-ups for employees in Japan

ABSTRACT

Health check-ups provide information on disease risk for individuals. It is assumed that such negative health information will lead to the adoption of healthier lifestyles. However, the relationship between the information provided by health check-ups and subsequent lifestyle modifications remains unclear. This study investigates whether warning signals that people receive after health check-ups lead to modified smoking and drinking behaviours over 10 years after the check-ups, using a longitudinal nationwide survey of middle-aged people conducted from 2005 to 2018 in Japan. The panel nature of the data enabled me to control for unobserved individual heterogeneity as individual fixed-effects. The results show that negative health information provided by check-ups reduces smoking and drinking. The effects were found to persist more than 10 years after the check-ups. It was also found that older, more educated, and higher-income individuals make significant reductions in these behaviours after receiving warnings. These findings suggest that negative health information from health check-ups may steer lifestyles in a healthier direction.

KEYWORDS:

• Health check-ups
• health information
• smoking
• drinking
• Japan

JEL CLASSIFICATION:

• I10
• I18
• I19

I. Introduction

A report of the World Health Organization (2020) identified modifiable lifestyle factors such as smoking and drinking to be a leading cause of global deaths. However, people are often unwilling to engage in healthy lifestyles. One explanation is that individuals are imperfectly informed about their own health (Kenkel 1991). Limited knowledge about their own health is of particular concern when individuals with health risks are unaware of problems, especially in the absence of warning signals or symptoms. One strategy to improve individuals’ knowledge about their own health is to implement a public health check-up programme. The policy assumes that personal health information provided by health check-ups will lead to healthier lifestyles and eventually help prevent diseases.

However, there is little consensus on the effects of health information received from health check-ups on lifestyle modifications. Some find little evidence (e.g. Fukuma et al. 2020; Iizuka et al. 2021), while others find significant improvements (e.g. Kang, Kawamura, and Noguchi 2021; Oikawa et al. 2023). Additionally, assuming that such effects exist, evidence on whether they are transitory or persistent is very limited. Furthermore, little is known about who is more likely to modify their lifestyles in response to health information from check-ups.

This study investigates whether warning signals from check-ups modify smoking and drinking behaviours over a period of 12 years after the check-ups using a longitudinal survey of Japanese middle-aged people from 2005 to 2018. Additionally, I argue that the lifestyle changes induced by the warning signals are heterogeneous. I focus on individuals who have not previously received health warnings from check-ups, following Iizuka et al. (2021).

This study’s setting is mandated health check-ups for employees in Japan, which brings some advantages. First, mandatory check-ups circumvent concerns about self-selection for health check-ups. I conducted balanced tests to ensure that among the baseline characteristics, there is no difference between respondents who received warning signals from check-ups and who did not. Typically, in other settings, participation in check-ups is voluntary. Accordingly, their receipt of check-ups is likely to be correlated with unobserved factors such as health preferences; thus, endogeneity may exist. Second, I can use individual-level panel data, which permit the control of unobserved heterogeneity as individual fixed-effects. Additionally, the panel nature of the data enabled me to examine a falsification test and the concern about non-random attrition to strengthen the findings.

II. Materials and methods

Data sources

This study used panel data from a nationwide, population-based survey, the Longitudinal Survey of Middle-aged and Elderly Persons (LSMEP), which began in 2005 with a sample of individuals aged 50–59 years, and was conducted by the Ministry of Health, Labour and Welfare (MHLW) in Japan. I used the surveys conducted between 2005 and 2018. The respondents are surveyed annually in November. The survey asks questions about each respondent’s current lifestyle behaviours. It also asks if a respondent received health check-ups within one year and whether any health warnings were received after the check-ups.

The baseline survey for this study is 2006.¹ A key explanatory variable is the dummy variable for health warnings from health check-ups that a respondent received between November 2006 and October 2007, as reported in the 2007 survey. The main outcomes are lifestyle changes between the 2006 and 2007 surveys that measure behavioural changes within one year after the health check-ups. I also examine longer-term effects (within four, eight, and twelve years after the health check-ups) on lifestyle changes. The variables and sample selection criteria are explained in Appendix A.

Columns (1) and (2) of Table 1 summarize the baseline characteristics of respondents who received health warnings and those who did not, respectively. Columns (1) and (2) show mean values with standard deviations in brackets. Column (3) shows the p-values for mean differences. The table shows that among the baseline characteristics, there is no difference between those who received warning signals from check-ups and those who did not.

Table 1. Baseline characteristics in the 2006 survey.

	(1)		(2)
	Not received health warnings reported in the 2007 survey (n = 3199)		Received health warnings reported in the 2007 survey (n = 1008)		(3)
Variables	Mean	(S.D.)	Mean	(S.D.)	$\mathit{p}$-value for mean differences
Panel A: Baseline individual characteristics
Age	55.66	(2.67)	55.81	(2.72)	0.12
Female	0.29	(0.45)	0.29	(0.45)	0.86
Undergraduate/postgraduate education	0.35	(0.48)	0.37	(0.48)	0.31
Married	0.86	(0.35)	0.88	(0.33)	0.26
Household size	3.27	(1.45)	3.21	(1.44)	0.29
Household monthly income (in JPY 10,000)	33.30	(53.77)	31.99	(40.52)	0.48
Executive officer	0.09	(0.29)	0.09	(0.29)	0.96
Weekly working days	5.25	(0.66)	5.26	(0.73)	0.96
Weekly working hours	44.19	(10.64)	44.60	(11.42)	0.31
Wage rate (in JPY)	2712.7	(7725.0)	2498.8	(4235.8)	0.42
Firm size fewer than 100 employees	0.46	(0.50)	0.47	(0.50)	0.53
Panel B: Baseline lifestyle variables
Current smoker	0.37	(0.48)	0.36	(0.48)	0.56
Number of cigarettes smoked	6.77	(9.75)	7.15	(10.46)	0.29
Frequency of drinking	0.31	(0.46)	0.33	(0.47)	0.23
Quantity of drinks	1.12	(1.09)	1.16	(1.06)	0.28

Household monthly income is divided by the square root of household size.

Model specification

The model to test whether health warnings from check-ups affect lifestyle variables is written as follows:

(1) $Y_{\mathit{it}}=\mathit{\beta Warnin}{g}_{\mathit{it}}+\mathit{\pi }{X}_{\mathit{it}}+{\delta }_i+{\alpha }_t+u_{\mathit{it}}$(1)

$Y_{\mathit{it}}$ is a lifestyle variable for individual i at time t; $\mathit{Warnin}{g}_{\mathit{it}}$ is a dummy variable indicating whether individual i reported receiving health warnings from health check-ups; $X_{\mathit{it}}$ is time-varying individual characteristics; and $u_{\mathit{it}}$ is an error term. I account for employee-level unobserved factors and aggregate-level time effect using employee and year fixed effects, ${\delta }_i$ and ${\alpha }_t$. The use of employee fixed effects can eliminate the bias arising from time-invariant employee-specific confounding factors. Parameter $\mathit{\beta }$ captures the effect of health information from check-ups on lifestyle changes.

III. Results and discussion

Table 2 summarizes the estimates for running Equation (1)(1) $Y_{\mathit{it}}=\mathit{\beta Warnin}{g}_{\mathit{it}}+\mathit{\pi }{X}_{\mathit{it}}+{\delta }_i+{\alpha }_t+u_{\mathit{it}}$(1) .² The estimate β from Equation (1)(1) $Y_{\mathit{it}}=\mathit{\beta Warnin}{g}_{\mathit{it}}+\mathit{\pi }{X}_{\mathit{it}}+{\delta }_i+{\alpha }_t+u_{\mathit{it}}$(1) is reported with standard errors clustered at an employee level in parentheses for all the regression tables. The main analysis results in Column (1) (one year after the check-ups) show that health warnings from check-ups significantly decreased both the probability of being a current smoker and the number of cigarettes smoked at the 1% level. I also found that health warnings were associated with a significant reduction in drinking behaviours.

Table 2. The effects of warning signals from health check-ups on behavioural changes.

	Main	Long-term
	1 year after check-ups (2007—2006)	4 years after check-ups (2010—2006)	8 years after check-ups (2014—2006)	12 years after check-ups (2018—2006)
Outcome	(1)	(2)	(3)	(4)
Current smoker
$\hat{\beta }$	−0.023***	−0.005	−0.020	−0.012
	(0.008)	(0.012)	(0.015)	(0.017)
n	4179	3596	3167	2930
Number of cigarettes smoked
$\hat{\beta }$	−0.507***	−0.584**	−0.621*	−0.702**
	(0.149)	(0.257)	(0.317)	(0.357)
n	4169	3584	3156	2919
Frequency of drinking
$\hat{\beta }$	−0.024**	−0.032**	−0.039**	−0.044**
	(0.012)	(0.014)	(0.016)	(0.018)
n	4178	3595	3166	2927
Quantity of drinks
$\hat{\beta }$	−0.047*	−0.001	−0.004	−0.012
	(0.025)	(0.028)	(0.031)	(0.035)
n	4138	3562	3133	2907

Significance levels: ***p < 0.01, **p < 0.05, *p < 0.10.

For long-term analysis, Columns (2), (3), and (4) show the effects within four, eight, and twelve years after the check-ups, respectively. The long-term effects show that health warnings significantly decreased the number of cigarettes smoked and the frequency of drinking. These effects were comparable to the corresponding main effects in Column (1), indicating that the effects were persistent more than 10 years after receiving a health warning. In contrast, the results for current smokers show the effects fading with time, suggesting the difficulty in sustaining smoking cessation.

As a robustness check, I conducted a falsification test that examines the effect of a health warning reported in the 2007 survey on lifestyle changes between the 2005–2006 surveys. As expected, no evidence of lifestyle changes prior to the baseline was found (Appendix B). As another robustness check, I examined the concern that the non-random attrition of particular types of individuals can be a potential confounder that may bias the estimates. As expected, no evidence of non-random attrition was found (Appendix B).

Table 3 summarizes the main results split by different subsamples. I divide the sample by median age (Columns (1) and (2)), educational attainment (Columns (3) and (4)), and median income (Columns (5) and (6)). The coefficients among older, more educated, and higher-income individuals were larger in magnitude and indicated a more statistically significant association compared to their counterparts. The heterogeneous effects of age may possibly reflect differences in the severity of health risks among age groups. Additionally, the effects concentrated among more educated individuals are consistent with the previous findings (Oikawa 2023). Furthermore, the results for higher-income individuals are consistent with previous findings that demonstrate similar heterogeneity across income groups in response to hypertension diagnoses (Zhao, Konishi, and Glewwe 2013).

Table 3. Heterogeneous effects (within one-year change).

	Age		Educational Attainment		Household income
	Age $\ge$ Median	Age $<$ Median	Undergraduate/ postgraduate education	High School graduation or less	Income $\ge$ Median	Income $<$ Median
Outcome	(1)	(2)	(3)	(4)	(5)	(6)
Current Smoker
$\hat{\beta }$	−0.035***	−0.012	−0.028**	−0.019**	−0.029***	−0.017
	(0.012)	(0.011)	(0.014)	(0.010)	(0.009)	(0.013)
n	2113	2066	1486	2693	2136	2043
Number of cigarettes smoked
$\hat{\beta }$	−0.716***	−0.301	−0.678***	−0.386**	−0.557***	−0.464*
	(0.225)	(0.188)	(0.247)	(0.183)	(0.182)	(0.239)
n	2107	2062	1484	2685	2131	2038
Frequency of drinking
$\hat{\beta }$	−0.046***	0.001	−0.042**	−0.013	−0.045***	−0.003
	(0.016)	(0.017)	(0.019)	(0.015)	(0.016)	(0.016)
n	2112	2066	1488	2690	2136	2042
Quantity of drinks
$\hat{\beta }$	−0.078**	−0.010	−0.030	−0.056*	−0.056	−0.036
	(0.034)	(0.037)	(0.039)	(0.032)	(0.035)	(0.036)
n	2086	2052	1478	2660	2123	2015

Significance levels: ***p < 0.01, **p < 0.05, *p < 0.10.

IV. Conclusions

In summary, I find that negative health information from health check-ups is associated with a reduction in smoking and drinking in middle-aged employees in Japan. Furthermore, I find that the information induces long-term behavioural modifications. Finally, a distinct pattern of heterogeneity is found: those who are older, more educated, and with higher incomes are more responsive to negative health information in terms of lifestyle changes. These results should be considered alongside the following limitation. If unobserved time-varying factors may exist, the methods employed in this study cannot account for them. Valid instrumental variables are required to handle this concern; however, to the best of my knowledge, they are not available in this study’s setting.

Ethical approval

This study has received official approval for the use of secondary data from the Statistics and Information Department of the MHLW under Tohatsu-1005–2 as of 5 October 2020, and this research was conducted with permission from the Ethics Review Committees of Waseda University (approval no. 729–420).

Acknowledgements

The author thanks Toru Asahi, Toshihide Awatani, Mindy Fang, Rong Fu, Cheolmin Kang, Akira Kawamura, Haruko Noguchi, Masato Oikawa, and Kenichiro Tamaki for extensive discussions and suggestions.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Data used in this study cannot be shared publicly because of Article 33 of Japan’s Statistics Act (Act No. 53 of 2007). Data are available based on the criteria set by the Ministry of Health, Labour, and Welfare (contact via https://www.mhlw.go.jp/toukei/sonota/chousahyo.html) for researchers.

Additional information

Funding

This study was financially supported by a Grant-in-Aid for Scientific Research Project funded by the Ministry of Health, Labour, and Welfare (MHLW): “An empirical study on the socioeconomic impact of lifestyle-related disease prevention by industry and region (19FA1013)” (PI: Haruko Noguchi). This study was also financially supported by JSPS KAKENHI (19H05487, 20K20418, 22K01539).

Notes

¹ I used the 2005 survey for a falsification test (Appendix B).

² Physical inactivity is another possible outcome. Following Kumagai and Ogura (2014), I used the dummy variable for regular physical activity and found no significant effects (Appendix A).

Appendix A.

Explanation of explanatory variables, outcome variables, and sample selection

1. Explanatory variables

A key explanatory variable is the dummy variable for health warnings from health check-ups that a respondent received between November 2006 and October 2007, as reported in the 2007 survey. Health check-up results that are outside the normal range typically notify the need for follow-up medical care, confirmatory tests, or counselling to encourage the adoption of healthy lifestyles. In this study, these notifications are defined as warning signals from health check-ups. It should be clarified that I did not observe the exact values of the test results, such as the level of fasting blood sugar and blood pressure, in the survey data. Thus, I have not identified which type of health risk (risk of diabetes, hypertension, hyperlipidaemia, or other diseases) was determined for the individual.

The other explanatory variables include time-varying individual characteristics, namely, marital status, household size, household income, a dummy variable for executive officer, weekly working days, weekly working hours, wage rate, and firm size. As geographical information is not available for the period between the 2006 and 2018 versions of the LSMEP, this prevented me from controlling regional factors in the empirical model. The estimation results of other explanatory variables are available upon request.

For the estimation, as a robustness check, I exclude labour market characteristics (household income, a dummy variable for executive officer, weekly working days, weekly working hours, wage rate, and firm size) since these variables may be endogenous in the model of lifestyle outcomes. I find that my baseline results are barely affected by excluding labour market characteristics (the results are not shown).

2. Outcome variables

Outcome variables are the changes in lifestyle behaviours. Regarding alcohol consumption, two indices were used: (i) frequency of drinking and (ii) quantity of drinks per occasion (1 unit = 16.9 fl oz [500 ml] of beer). The respondents were asked about the frequency of drinking with the following options: 1, everyday; 2, five or six days a week; 3, three or four days a week; 4, one or two days a week; 5, a few days a month; 6, hardly drink; and 7, do not drink at all. The frequency of drinking is a dummy variable that takes a value of one if the person drinks every day. The respondents were asked about the quantity of drinks with the following options: 1, less than one can of beer or its equivalent a day; 2, one or more and less than three cans of beer or its equivalent a day; 3, three or more and less than five cans of beer or its equivalent a day; and 4, five or more cans of beer or its equivalent a day. The question about quantity drunk was only asked to individuals who drink at least a few days a month; the value of this variable was replaced by zero if the respondent answered that they drink less than a few days a month. Regarding smoking habits, two indices were used: (i) whether the respondent is currently a smoker or not and (ii) number of cigarettes smoked per day. Current smokers were asked about the number of cigarettes with the following options: 1, 10 or less cigarettes per day; 2, between 11 and 20 cigarettes per day; 3, between 21 and 30 cigarettes per day; and 4, 31 and more cigarettes per day. The number of cigarettes smoked was replaced by zero if the respondent was a non-smoker. Note that in order to treat the discrete variable for the quantity of drinks and the number of cigarettes as continuous, each category was assigned the midpoint of the upper and lower boundaries for that category.

The survey also covers physical inactivity, which is another possible lifestyle risk factor. I examined the effects of health warnings from health check-ups on participation in regular physical activity (RPA) by RPA classification, following Kumagai and Ogura (2014). No evidence indicated that health warnings from check-ups were associated with changes in physical activity (see Table A1).

Table A1. The effects of warning signals from health check-ups on changes in physical activity.

	Main		Long term
	1 year after check-ups		4 years after check-ups	8 years after check-ups	12 years after check-ups
Outcome	(1)		(2)	(3)	(4)
Panel A: Regular physical activity
Regular physical activity
$\hat{\beta }$	−0.016		−0.020	−0.024	−0.029
	(0.017)		(0.021)	(0.020)	(0.025)
n	4051		3525	3105	2875
Panel B: By intensity of regular physical activity
Light regular physical activity
$\hat{\beta }$	−0.012		−0.024	−0.022	−0.038
	(0.018)		(0.020)	(0.023)	(0.026)
n	4051		3525	3105	2875
Moderate or vigorous regular physical activity
$\hat{\beta }$	0.005		0.027	−0.023	0.001
	(0.014)		(0.018)	(0.023)	(0.023)
n	4051		3525	3105	2875

The dummy variable for regular physical activity takes on the value of one if individuals took part in sports more than twice a week during their leisure time, and zero otherwise. The intensity of exercise was classified as follows: light (stretch, light gymnastics), moderate (walking, jogging), or vigorous (aerobics, swimming). Following Kumagai and Ogura (2014), moderate and vigorous are combined. Significance levels: ***p < 0.01, **p < 0.05, *p < 0.10.

3. Sample selection

Considering that this study covers mandated health check-ups for employees, the analysis sample was restricted to employees. The analysis sample was also restricted to respondents with no missing values for the results of check-ups reported in the 2007 survey. This information was used to create a key explanatory variable in the empirical analysis. Furthermore, because this study examines the effects of warning signals from health check-ups on lifestyle changes after the baseline 2006 survey, I excluded respondents who reported that they had received any health warnings from check-ups reported in the 2005 and 2006 surveys. The final sample comprised 4207 individuals.

Appendix B.

Robustness checks

I performed two robustness checks. One is a falsification test. The other is a robustness check regarding potential endogeneity due to non-random attrition.

1. Falsification test

I tested whether there was any lifestyle change prior to the baseline survey, which included changes in lifestyles between the 2005/2006 surveys. By design, lifestyles prior to the baseline should not be affected. As reported in Table B1, I found no evidence that health warnings from check-ups between November 2006 and October 2007 affected pre-treatment changes in smoking and drinking behaviours between 2005–2006.

Table B1. Falsification test: changes in lifestyles before baseline year 2006.

	Placebo	Main
	1 year before check-ups (2006—2005)	1 year after check-ups (2007—2006)
Outcome	(1)	(2)
Current smoker
$\hat{\beta }$	−0.010	−0.023***
	(0.007)	(0.008)
n	4175	4179
Number of cigarettes smoked
$\hat{\beta }$	−0.198	−0.507***
	(0.127)	(0.149)
n	4167	4169
Frequency of drinking
$\hat{\beta }$	0.003	−0.024**
	(0.011)	(0.012)
n	4168	4178
Quantity of drinks
$\hat{\beta }$	−0.014	−0.047*
	(0.027)	(0.025)
n	4098	4138

The estimate β from Equation (1)(1) $Y_{\mathit{it}}=\mathit{\beta Warnin}{g}_{\mathit{it}}+\mathit{\pi }{X}_{\mathit{it}}+{\delta }_i+{\alpha }_t+u_{\mathit{it}}$(1) is reported with standard errors clustered at the employee level in parentheses. Significance levels: ***p < 0.01, **p < 0.05, *p < 0.10.

2. A robustness check regarding potential endogeneity due to non-random attrition

The average attrition rate between the 2006 and 2018 surveys was fairly low, at 3.8% in each wave; however, the attrition of particular types of individuals can be a potential confounder that may bias the estimates. For instance, it may be possible that individuals who received health warnings from health check-ups but did not modify their lifestyles in response to the warnings may be less likely to respond to surveys after the baseline due to worsening health. Eventually, such individuals may be more likely to be excluded from the sample in a long-term analysis. If this is the case, non-random attrition can overestimate the effects in this study.

I examined this concern for a long-term analysis (i.e., four, eight, and twelve years after the health check-ups). Considering that the analysis sample was restricted to individuals with no missing values about the result of check-ups reported in the 2007 survey, this sample selection restricted me from examining the attrition between the baseline survey in 2006 and the 2007 survey.

To address this concern, I regressed a dummy variable for attrition that took a value of one if the individual was not observed in the subsequent survey after the baseline survey and zero otherwise, on the dummy variable for health warnings from check-ups and the same set of variables as in Equation (1)(1) $Y_{\mathit{it}}=\mathit{\beta Warnin}{g}_{\mathit{it}}+\mathit{\pi }{X}_{\mathit{it}}+{\delta }_i+{\alpha }_t+u_{\mathit{it}}$(1) , in order to check whether the attrition was correlated with health warnings. If non-random attrition occurs in long-term analysis, the coefficients in Table B2 are expected to show statistically significant effects. This approach is based on Voors et al. (2012). The results in Table B2 show that attrition was not systematically correlated with receiving health warnings.

Table B2. The effects of health warnings from health check-ups on attrition.

	Attrition dummy for 4 years after check-ups (2006—2010)	Attrition dummy for 8 years after check-ups (2006—2014)	Attrition dummy for 12 years after check-ups (2006—2018)
Outcome	(1)	(2)	(3)
Key explanatory variable:
Health warnings dummy	−0.0036	0.0019	0.0029
	(0.0023)	(0.0022)	(0.0031)
n	4207	4207	4207

Significance levels: *** p<0.01, ** p<0.05, * p<0.10. Standard errors clustered at the employee level are reported in parentheses. In this study’s sample, the fraction of attrition between 2006–2010 is 14%, between 2006–2014 is 24%, and between 2006–2018 is 31%.