Our data

We used Japanese longitudinal data from the Japanese Study of Health, Occupation and Psychosocial Factors Related Equity (J-HOPE) (Eguchi et al. 2015; Inoue et al. 2016, 2018; Miyaki 2013; Tsutsumi 2011). Appendix Table presented the J-HOPE raw data sample status by company (anonymized by industry), male sample status in the raw data, and sample status in our analysis data. J-HOPE surveyed 6400 employees at 10 arbitrary companies in 2010 and at 13 arbitrary companies between 2011 and 2013 (10,228 employees in 2013). In this dataset, we restricted our study sample to nine companies with individuals who were promoted to managerial positions, to men, and to individuals who could be tracked for four years; further, we excluded from the analysis individuals who were in managerial positions at the first period in 2010, individuals who transferred from managerial to other positions, individuals whose working hours are less than 30, and individuals who indicated that their education had decreased. Individuals with missing data for each variable used in the analysis were also excluded. Finally, four waves of complete panel data for 1454 individuals (5816 in total for the four waves) were used in the analysis.

Treatment variable: promotion to managerial positions

J-HOPE does not contain a direct question regarding promotions. Among the eight occupational categories, managerial positions are included in one of these occupational categories. The analysis was conducted by considering promotion to managerial occupations in 2011 and later as the promotion to managerial positions,” which is the focus of our study. The data exclude individuals who have been in managerial positions since 2010, the first period, and those who have changed from managerial positions to other types of jobs. It is difficult to assume that the effects of promotion from other occupations to management, and from management to other occupations, are identical. Therefore, we focused on the former to avoid contaminating the effects we desired to estimate.

Step 1’s outcome variables: mental health and satisfaction

Kessler et al.’s (2002) Kessler-s (K6) score, which has been validated for Japan (Furukawa et al. 2008), was used in our analysis as both a continuous and binary variable that takes the value of 1 when the score is 5 or higher. Scores of 5 or higher indicate that individuals have depressive and anxiety tendencies. Higher scores indicate worse mental health.

We also used the results of a study employing a four-point scale on satisfaction with work and family life (1 = unsatisfied, 2 = somewhat unsatisfied, 3 = fairly satisfied, and 4 = satisfied; Shimomitsu 2000). We used the index as a continuous variable in our analysis.

Step 2–1’s outcome variables: potential negative mechanisms

The increase in working hours negatively affects mental health (Virtanen et al. 2011; Amagasa and Nakayama 2013; Kim et al. 2016; Afonso et al. 2017; Ogawa et al. 2018). J-HOPE asked about weekly working hours using the following option: 30 h or less, 31–40 h, 41–50 h, 51–60 h, and 61 h or more. We used these responses in our analysis as continuous values from 1 to 5, beginning with less than 30 h as 1.

J-HOPE used the Japanese version index of job demand in the job-demand control model (Karasek 1979; Kawakami et al. 1995), effort-reward-imbalance model (ERI; Siegrist 1996; Tsutumi et al. 2011), and job-insecurity model (Karasek et al. 1998; Kawakami et al. 2004). Higher scores indicate worse mental health. These scores were used in our analysis as continuous variables.

Step 2–2’s outcome variables: potential positive mechanisms

Apouey and Clark (2015) and Kim and Koh (2021) used lottery winnings as the context for exogenous income and indicated that increases in exogenous income positively impact individuals’ mental health in the short run, besides having long-term effects. J-HOPE does not ask about individual or annual household income. Although the movements in individual and household incomes do not necessarily coincide depending on the trends of people living together, the household income response was used as the outcome variable because it is expected to be generally consistent. J-HOPE asked about annual household income using the following choices: less than ¥2.99 million, ¥3 million to ¥4.99 million, ¥5 million to ¥7.99 million, ¥8 million to ¥9.99 million, ¥10 million to ¥14.99 million, and ¥15 million or more. As there was a marked difference in the range of each item, the median value for each option was log-transformed (¥15 million was used for ¥15 million and above) and used in the analysis as a continuous value.

J-HOPE utilized the Japanese version of the job control index in the job-demand control model (Karasek 1979; Kawakami et al. 1995), work engagement (Schaufeli et al. 2006; Shimazu et al. 2008), and job role clearness (Coverman 1989). These scores were used in our analysis as continuous variables. Increasing these scores is thought to have a positive effect on mental health. These scores were used in our analysis as continuous variables.

Controlling time variant confounders

The birth of a child may have a positive or negative effect on mental health. Since we do not have access to data on the number of children, we use the total number of family members, which is a proxy variable in the survey’s acquisition information, as a control variable. Since the individual and time confounder are considered, we do not use any other control variables.

Promotions to managerial positions in Japan

The Global Employment Status and Growth Attitudes Survey (Persol Research and Consulting Co. Ltd. 2022) was conducted in 18 countries (13 Asia-Oceania countries plus the United States, the United Kingdom, Germany, France, and Sweden). Respondents were asked, “Do you feel you would like to be a manager in your current company?” and their responses were rated on a five-point scale. The results showed that Japan had the lowest percentage of positive responses (i.e., “agree” and “somewhat agree”), at 19.8%. As Australia’s next lowest rate was 38.0% and all countries’ average rate was 58.6%, Japan could be a negative country in terms of promotion to managerial positions.

Estimation methodPrevious studies’ method: conventional TWFE and event study

Previous studies typically performed conventional TWFE estimation using panel data as follows:

$$_=_+_+bPostTrea_+erro_$$

(1)

PostTreat denotes a binary indicator that continues to take the value of 1 after an individual has been treated, and \(_\) (j = i,t) denotes the fixed-effect term of the j factor. Instead of ordinary least squares estimation, TWFE has been widely used in empirical studies to address endogeneity issues successfully because it controls for individual and time-derived confounding factors that are not observable in the data. In the context of promotion effects, studies have used mental and physical health for Y in Eq. (1) (Anderson and Marmot 2012; Boyce and Oswald 2012; Johnston and Lee 2013; Sato 2015; Nyberg et al. 2017; Ikesu et al. 2021). Since TWFE was considered an extension of difference-in-difference (DID) estimation, the coefficient value b in (1) can be considered as average treatment effect on treated (ATT) as well as DID estimation.

Along with TWFE, a dynamic decomposition of the onset of the TWFE effect, called an event study, was estimated as follows:

$$_=_+_+\sum\limits_^_La_^+\sum\limits_^_Lea_^+erro_$$

(2)

\(La_^\) denotes a binary indicator that takes the value of 1 if individual i is in period t before the gth period of treatment, and \(Lea_^\) denotes a binary indicator that takes the value of 1 if individual i is in period t after the pth period of treatment. The baseline is generally measured one period before treatment. This dynamic estimation of (2) allows us to confirm the prior parallel trend and determine how long the effect lasts and whether it occurs with a lag. The presence or absence of a statistical significance in \(_\) suggests whether a conditional parallel trend could hold in g prior period from treatment (if it could be satisfied, the results could be interpreted as a causal relationship). \(_\) was interpreted as dynamic ATT after treatment period p. The promotion effect literature has been used as an indicator of the number of terms before or after a promotion. For example, Johnston and Lee (2013) used the Household, Income and Labour Dynamics (HILDA) survey data for Australia to show that mental health deteriorates two years after a promotion.

Recent modified estimation method: Callaway and Sant’Anna’s (2021) estimation

However, the recent advances in the methodological and theoretical understanding of TWFE have revealed the difficulties in considering the TWFE coefficient (1)’s b to be ATT (De Chaisemartin and d’Haultfoeuille 2022; Borusyak et al. 2021; Goodman-Bacon 2021; Baker et al. 2022). Goodman-Bacon (2021) showed that if the sample is divided into three groups—never treated, early treated, and later treated—the TWFE can be broken down into a 2 × 2 DID comparison for each group. It pointed out that TWFE contains a DID estimation in which the early treated group’s post-treatment status is considered the control group, as a so-called “forbidden comparison.” Therefore, if there were some heterogenous treatment timing, TWFE could be biased. Borusyak et al. (2021) point out that the occurrence of such bias remains even with event study specification.

Callaway and Sant’Anna’s (2021) method was developed for treatments with staggered timing. They modified TWFE not to conduct forbidden comparisons and their method estimates ATT based only on appropriate comparisons. Our Appendix presents a detailed discussion of the analysis method (TWFE’s bias), Callaway and Sant’Anna’s (2021) method, and their definition of ATT.

Our estimation in this research

The timing of the promotion to managerial positions was heterogeneous for individuals within the data period. In this context, we confirmed dynamic and static treatment effects as our main result by considering the staggeredness in treatment timing and using Callaway and Sant’Anna’s (2021) specification. We also attached static TWFE results of Eq. (1) to compare the findings. We favored dynamic over static results. We did not judge the presence of a causal effect based only on the presence of statistical significance. We display the dynamic effects as figures. We argue that causal effects exist when the pre-treatment motion is stable and post-treatment coefficient deviates from pre-treatment coefficientsFootnote 3.

The analysis was conducted in two steps. In Step 1, we estimated the dynamic and static treatment effects without controlling for covariates, using mental health status and job/life satisfaction as outcome variables. In Step 2, we estimated the dynamic treatment effect using potential factors underlying the mechanism for mental health status as the outcome variable y, for which previous studies have suggested a positive or negative relationship. Previous studies have shown that income, job control, work engagement, and high job clarity have positive effects on mental health (Apouey and Clark 2015; Kim and Koh 2021; Karasek 1979; Kawakami et al. 1995; Schaufeli et al. 2006; Shimazu et al. 2008; Coverman 1989). Since personal income was not included in our data, we substituted it for household income, which was included in our data. Previous studies have shown that working hours, job demands, effort-reward imbalance (ERI), and high job insecurity negatively affect mental health (Virtanen et al. 2011; Amagasa and Nakayama 2013; Kim et al. 2016; Afonso et al. 2017; Ogawa et al. 2018; Karasek 1979; Kawakami et al. 1995, 2004; Siegrist 1996; Tsutsumi et al. 2001; Karasek Jr et al. 1998).

In Step 1, we estimated the aggregated overall effect of promotion on mental health and satisfaction, including the effects of various pathways. Step 2 explored the mechanisms responsible for the total effect obtained in Step 1. Because of data limitations, we could check the results only for the four positive and four negative elements.

We made the following three estimations to check the robustness of the results: First, we estimated the probability of being a worker promoted during 2011–2013. We used the 2010 information of the K6 total score, working hour, ERI score, log household income, work engagement score, age, years of experience, number of household members, education level, and firm. We constructed a 1:1 matched sample settled caliper to 0.001 and used Callaway and Sant’Anna’s (2021) estimation.Footnote 4 Second, we used Sun and Abraham’s (2021) approach as a robust method to address the staggeredness of treatment timing other than Callaway and Sant’Anna’s (2021) method. Third, we performed conventional TWFE event studies.