STRENGTHS AND LIMITATIONS OF THIS STUDY

The study examines correlations between subjective evaluations of well-being and physical health using a broad instrument, large datasets and multiple clinical health indicators.

The hierarchical structuring and clustering offer new insights into how well-being assessment could be used to identify different target groups for work well-being programmes.

The study sample consists solely of Finnish employees and the instrument used has not been validated.

We could only investigate correlations and not present any causal relationships.

Introduction

Occupational health challenges are changing as the workplace, demographics, technology and environment are transforming. Previously, occupational health and safety (OHS) measures focused on physical hazards.1 However, it has become evident that psychosocial hazards play a major role in occupational stress.2 Excessive working hours, non-standard forms of employment, remote work, telework and the ageing workforce all contribute to this change.1 In Finland, mental disorders recently overtook musculoskeletal disorders as the main reason for disability retirement,3 and mental disorder-related absences increased in Finland by 49% from 2016 to 2019.4 Similar trends can be seen in the increase of mental health-related disability claims in Norway, Sweden and Denmark.5 With a clear increase in psychosocial hazards, occupational health must find and incorporate new preventive measures that encompass a broader range of well-being-associated factors.6

Higher levels of overall well-being have been shown to increase labour productivity7 8; reduce unscheduled absences, presenteeism and employee turnover9; generate medical cost savings; and reduce absenteeism costs.10 Unfortunately, these benefits are rarely achieved, as currently many employee well-being interventions have only small, moderate, non-significant or inconsistent results.11–17 The reasons for this vary from study design and methodology to participation bias, missing data and short follow-up periods.11 14–16 Nonetheless, three important factors in employee well-being programmes have been identified. These are: (1) successful targeting of high-risk individuals, (2) aligning intervention with employee needs and (3) applying appropriate metrics for evaluating effectiveness.18 19

Currently, employees are assigned to workplace well-being intervention programmes based on health risk assessments, health-related behaviours, the presence of a certain condition or disease or a simple willingness to participate.17 18 20 The health risk assessments often focus on traditional clinical health indicators, which include blood pressure, pulse, height, weight, body mass index (BMI) and blood biomarkers.21 However, these clinical measurements are often costly and invasive, have a narrow focus on physical health, have poor scalability and tell little about employee needs. Subjective survey-based assessments, on the other hand, are cheap, easily implemented and repeated even in large cohorts, cause very little inconvenience to the employee and can incorporate a very broad perspective on physical, psychological and social well-being, bringing the focus to the individual’s experience of well-being.

This study aims to (1) investigate how a multifaceted subjective online well-being assessment could be used to identify distinct target groups for well-being interventions and (2) examine whether the assessment could, in some cases, be used instead of the traditional clinical health indicators.

MethodsStudy design, setting and data collection

This cross-sectional study used data from the Finnish healthcare provider Aava Medical Centre’s (Aava) well-being survey (Aisti). First, we identified different elements of well-being encompassed by the assessment by analysing the hierarchical structure of the assessment items. Second, we used the assessment to identify different well-being groups. Third, we used the hierarchical structure and identified groups to examine the associations between clinical health indicators and the Aisti assessment.

Aisti well-being assessment

Aisti is a survey-based digital well-being solution developed by Aava’s expert group, which consists of clinicians, psychologists and experts from other lifestyle areas. The assessment combines a comprehensive well-being screening and personalised feedback to facilitate a personal preventive health and well-being approach. The survey is intended for working-age people, contains 131 questions and takes approximately 15 min to fill out. Examples of the questions can be found in online supplemental file 1. The majority of the questions have been selected from the available research literature based on the research and clinical experience of the expert teams. The Aisti questionnaire has not been validated or pretested, but it includes some scientifically validated full instruments, such as the General Anxiety Disorder Scale,22 two-item Whooley questions for screening depression risk,23 the Alcohol Use Disorders Identification Test,24 the Presenteeism Scale25 and the single-item measure of physical activity.26 More details on the validated instruments can be found in online supplemental file 2.

Dataset 1: Aisti database

The Aisti online well-being assessment has been used in multiple companies in Finland to assess employee well-being since 2018. To analyse the structure of the assessment and to identify well-being groups, we used a complete responses dataset (n=2990), in which each respondent had answered all the available 131 questions. The dataset included answers from 1505 women, 1464 men and 21 employees who stated their gender to be ‘other’. The mean age was 45 (SD 10 years). There were both office and manual workers from various industries (eg, technology, manufacturing and healthcare). Approximately 65% of respondents can be characterised as office workers and 35% as manual workers. Usually, the companies advertised to their employees beforehand that a well-being survey was to be performed. The participating companies sent Aava the email addresses of the employees via a secure email portal. Personalised links to the survey were sent via email to all the employees of the companies, which allowed them to sign into the survey portal and fill in their information and the survey. An informed consent for data processing was collected at the beginning of the online survey. The employees received three reminders of the survey. The response rate varied between 60% and 90%.

Dataset 2: an intervention study

To assess the associations between the questionnaire and clinical health indicators, we used data from the Virta1 randomised controlled trial (Clinical trials NCT04633876), which was approved by the Helsinki University Hospital District Ethical Committee on 4 September 2019 (project no. HUS/2093/2019). The trial was developed by Aava for working-age adults who were at an increased cardiometabolic risk and was conducted between March 2020 and June 2021. The trial had participants from three different companies that were occupational health clients of Aava in Finland. Two of the companies were in the technology industry and had only office workers (n=647). The third company had both manual (n=31) and office (n=35) workers from the food industry.

All employees were invited to a screening phase but to be eligible to participate, an employee had to be 18–65 years old, speak and understand Finnish or English and have given written informed consent. The participants were told that they had the right to withdraw from the study at any point without consequences. The exclusion criteria included (1) history of a major cardiovascular event during the past 6 months, (2) having type 1 or type 2 diabetes, (3) history of a malignant disease during the past 5 years, (4) use of lipid-lowering or obesity medications, (5) use of a cardiac pacemaker or history of atrial fibrillation, (6) pregnancy and (7) plan to travel more than 1 day a week during the intervention period.

A total of 713 participants took part in the screening phase, of which 448 were male, 261 were female and four participants chose their gender to be ‘other’. The mean age was 46 years (SD 9 years). All participants were invited to fill out the online Aisti survey and participate in an on-site measurement of the clinical health indicators (online supplemental file 3). The proportion of missing values varied between 0.6% and 0.8% for the clinical indicators and the response rate for the Aisti well-being assessment was 98.2% (700 out of 713 participants). Only complete data for each clinical health indicator was used in the analyses.

Step 1: Identifying the elements encompassed by the assessment

To identify which well-being-associated factors the assessment captured, we analysed the hierarchical structure of the questionnaire items. We first identified the pairwise Spearman correlations between the individual questions. Two questions with a maximum pairwise correlation of <0.2 were excluded from further analysis. For the remaining 129 questions, distances were calculated as 1—correlation and the questions were hierarchically clustered using Ward’s clustering criterion (function hclust in R). The questions were clustered into groups of highly correlating questions at four levels and these groups were interpreted as measuring or characterising the same underlying phenomenon. The question clusters obtained were named based on the questions that formed them. The highest level of clustering was termed the sector, the second-level subsector, the third-level theme and the finest-level subtheme. We calculated the average score within each question cluster for each respondent. We also calculated the average correlation between each question and the cluster score to compare the internal consistency between each question within each question cluster.

Step 2: Identifying different target groups for well-being interventions

As the number of questions per identified theme varied, we calculated a theme score for each respondent as an average of all the questions within the theme. No weighting of the sample was needed. The theme averages were used to calculate the Euclidean distances between the respondents (function dist in R). The respondents were clustered into five groups using partitioning around medoids based on the theme score distances (function pam (PAM), R package cluster). The group profiles in the average theme scores were used to interpret the groups. The clustering was visualised using principal component analysis of the theme scores (capscale function, R package vegan).

Step 3: Examining associations between Aisti assessment and clinical health indicators

We evaluated how the questions, either themselves or grouped into sectors or themes, were associated with clinical health indicators. First, we analysed how the overall interindividual distances in the questionnaire space were associated with clinical health indicators, using principal component analysis and correlating the health indicators with the first principal component scores. The analysis was stratified by gender. Second, we tested the ability of the questionnaire to explain the variation in clinical health indicators and identified sets of questions or themes that together best explained the health indicators. Third, for each clinical health indicator, we selected the most significantly associated questions for a full linear regression model, which we reduced stepwise using Akaike Information Criterion to arrive at the best model. The variables were normally distributed, and no patterns were evident in the residuals, indicating that the data followed model expectations. Finally, we compared the differences in the health indicators of the well-being groups using an analysis of variance.

Patients and public involvement

None.

ResultsHierarchical structure of the Aisti assessment

Hierarchical clustering revealed that the questions could be divided into groups at four levels: 2 sectors, 6 subsectors, 10 themes and 25 subthemes (figure 1). The two sectors were termed Mental well-being and Physical well-being: the former was further divided into the Mental balance and Resources for change subsectors, and the latter into the Lifestyle, Eating habits, Performance orientation and Health subsectors. The internal consistency of the Mental well-being categories was high, whereas Physical well-being comprised themes with low intercorrelations (figure 1).

Figure 1

Hierarchical clustering of questions categorised into four levels. The colour and number of each category indicate the average correlation coefficient between all the questions in the category and the average category score.

Identifying different target groups for well-being interventions

Principal component analysis of the theme scores showed that the first principal component correlated strongly with the total questionnaire score, whereas the second correlated with exercise habits (positively) and work satisfaction (negatively). Based on the average theme scores, the respondents were clustered into five groups using PAM (figure 2).

Figure 2

Principal component analysis of theme averages and five well-being groups: Multiple challenges (red), Recovery challenges (orange), Hard on oneself (purple), Lifestyle challenges (blue) and Good well-being (green) defined by function pam clustering.

The first group, the Good well-being group (n=755), consisted of respondents who had high values in all themes and subthemes as well as low presenteeism (86.91±10.02), indicating good well-being (figure 3). The second group, the Lifestyle challenges group (n=758), had good Mental well-being and low presenteeism (80.60±11.13) but low scores in the themes of Physical activity and Lifestyle. The group had low values, especially in the subthemes Eating behaviour, Screentime, Time for oneself and Daily activity. The third group, the Hard on oneself group (n=673), had high Physical activity scores but average Mental well-being scores, low presenteeism (79.61±12.04) and low scores in the Performance orientation theme, implying demanding characteristics and low psychological flexibility. This can also be seen in low Work satisfaction and Attitude towards exercise subtheme scores. The fourth group, the Recovery challenges group (n=454), had medium-to-low scores in all themes and more presenteeism (72.82 ± 15.1). The group had low values, especially in Happiness, Satisfaction with salary, Relationships, Energy for daily life, Work satisfaction, Resilience, Financial prudence, Mental well-being challenges, Time for oneself, Attitude towards exercise and Workload. Lastly, the fifth group, the Multiple challenges group (n=350), had very low overall scores and higher presenteeism (68.99±17.42).

Figure 3

Different well-being groups and subjective well-being results for men (blue) and women (purple). Reco, Recovery challenges; Good, Good well-being; Life, Lifestyle challenges; Hard, Hard on oneself; Mult, Multiple challenges.

Associations between questionnaires and clinical health indicators

Unsupervised principal component analysis of the questionnaire data revealed the same two principal topics in Dataset 2 that were identified using clustering in Dataset 1: Mental well-being and Physical well-being. The first principal component was strongly correlated with Mental well-being (correlation=0.97 and p<0.0001). The second principal component correlated most strongly with the Physical activity theme (correlation=0.63 and p<0.0001). The third component correlated with the health theme (correlation=0.42 and p<0.0001). The fourth component correlated most strongly with the Financial prudence subtheme (correlation=0.71 and p<0.0001). The fifth component correlated with illness (correlation=0.39 and p<0.0001) and the sixth with Performance orientation (correlation=0.61 and p<0.0001). Among both the women and the men, the second principal component was strongly correlated with BMI (women: correlation=0.50 and p<0.0001; men: correlation=0.27 and p<0.0001), waist circumference (women: correlation=0.50 and p<0.0001; men: correlation=0.29 and p<0.0001) and diastolic blood pressure (women: correlation=0.38 and p<0.0001; men: correlation=0.28 and p<0.0001). Among women, there was also a significant correlation with blood glucose (women: correlation=0.36 and p<0.0001; men: correlation=0.08 and p=0.08).

The clinical health indicators had significant correlations with several subthemes (figure 4). In both genders, the Energy for daily life, Eating behaviour, Physical activity, Snoring and Blood pressure subthemes were strongly associated with weight (BMI, weight, neck and waist circumference) and certain metabolic indicators, such as triglycerides. The questions measuring flexible attitudes towards eating were highly predictive of weight, but the questions about diet composition were not (figure 4). Among the men, the Alcohol and tobacco use, Amount of sleep and Work schedule subthemes were more strongly correlated with the clinical health indicators, whereas among the women, these associations were weak. However, in women, Time for oneself, Satisfaction with salary and Work satisfaction had a stronger link with multiple weight indicators compared with men. All in all, the Mental well-being sector was more strongly associated with clinical health indicators among women than among men. The Aisti well-being assessment was particularly strongly associated with the weight indicators, with the responses explaining 63% of the variation in BMI and 52% of waist circumference (online supplemental file 4).

Figure 4

Investigated health indicators’ correlations with different subthemes for (A) women and (B) men. The subthemes are coloured by the subsector. Significance key: ***0.001, **0.01 and *0.05.

We characterised the well-being groups in Dataset 2 by comparing the groups’ clinical health indicators (figure 5). All the groups were compared against the Good well-being group, men and women, separately. In general, the Good well-being group usually did not have problems with their clinical health indicators, while respondents in the Multiple challenges group had various unfavourable health indicators. These include low lipoprotein A1 (LIPOA1) and high-density lipoprotein cholesterol, elevated C reactive protein, waist and neck circumference and BMI. The Recovery challenges and Lifestyle challenges groups also had elevated weight indicators. The Hard on oneself and Good well-being groups did not differ from each other in terms of clinical health indicators, even though, based on the survey, they partly experienced different kinds of well-being challenges (figure 3).

Figure 5

Different well-being groups and clinical health indicators for men (blue) and women (purple). Reco, Recovery challenges; Good, Good well-being; Life, Lifestyle challenges; Hard, Hard on oneself; Mult, Multiple challenges

Discussion

We assessed how well a comprehensive, subjective well-being assessment could facilitate surveying the physical and mental health and well-being of large groups of employees. At workplaces, such assessments help indicate potential overall problems and identify at-risk individuals who may benefit from specific interventions or more throughout health examinations. Our results show that questionnaire responses indicate several groups with distinct characteristics and risks, which will be helpful when developing targeted interventions. Furthermore, the subjective assessment responses were strongly correlated with physical health, measured by blood values and anthropometry, which suggests that a questionnaire-based well-being evaluation may be more useful for screening large populations than more laborious in-person health checkups. We also identified novel well-being groups whose clinical health indicators did not differ while they clearly had differences in their subjective evaluations, emphasising the need to assess a broader range of well-being factors in OHS.

In the present analysis, we identified five distinct well-being groups: Good well-being, Hard on oneself, Lifestyle challenges, Recovery challenges and Multiple challenges. While the people in the Good well-being group generally had good values both in their subjective evaluations and clinical health indicators, the situation for the Multiple challenges group was the opposite. We also identified a group (Hard on oneself) that did not differ from the Good well-being group in terms of clinical health indicators but did differ in their subjective evaluations of well-being and experienced higher presenteeism, demonstrating the utility of a subjective well-being assessment in identifying target groups that clinical measurements fail to capture. Hence, while people in the Good well-being and Hard on oneself groups would benefit from traditional clinical measurements little, people in the Multiple challenges group would benefit a lot. This shows that an online well-being assessment could be potentially used to screen larger populations and target more costly and invasive health examinations to groups that benefit from them the most.

Despite Hard on oneself group not potentially needing a physical examination, their higher presenteeism can be considered the first signal of declining workability. Identifying this group may enable a shift in the focus of work disability prevention to a significantly earlier stage than was previously possible. As the results show, people in Lifestyle challenges and Recovery challenges groups have already started to portray physical health challenges. While people in the Lifestyle challenges group would significantly benefit from interventions aiming to increase physical activity and enhance eating habits, people in Recovery challenges would require balancing the workload and enabling time for themselves to be able to make changes to their lifestyles. The people in the Multiple challenges group would require more intensive, personalised guidance to enhance their physical and mental health.

We also identified highly significant correlations between metabolic health indicators and several subjective assessment subthemes. Correlations between clinical health indicators and subjective measures have been found before, for instance, by the subjective workability index, which predicts disability pensions, and the work ability score, which predicts long-term sickness absences.27 28 However, our study also strengthened some unexpected correlations. For instance, although the questions measuring flexible attitudes towards eating were highly predictive of weight, the questions relating to diet composition were not. This observation is in line with previous observations of predictors of successful long-term weight management, which tend to be more behaviour-related than food-related.29 30 However, it would be important to review the questions related to diet composition and assess whether the questions alone provide sufficiently credible data. Our results regarding BMI and subjective well-being are also in line with those of previous studies.31 The stronger association among the women is likely explained by the previously observed increased prevalence of problematic eating behaviours among them, such as emotional eating and loss of control over eating.32

Overall, our study highlights the importance of OHS paying attention to a broad range of well-being factors, as they are linked to physical and mental health. For instance, in the ‘Hard on oneself’ group, we can start seeing signs of declining workability, such as lowering energy for daily life, work satisfaction, resilience, time for oneself and increasing workload and schedule challenges. Furthermore, respondents who experienced well-being challenges seemed to also experience more clinical health challenges, which should also be examined in more detail. Hence, our results can help enhance the targeting of employee well-being programmes and help align them with employees’ needs by presenting a new, innovative clustering method that considers a broader range of well-being factors. This kind of subjective assessment offers a promising way to track long-term and seasonal changes in large populations, providing OHS researchers and practitioners with an up-to-date, detailed understanding of the well-being challenges in the workforce.

Strengths and limitations

We acknowledge the lack of validation of the Aisti survey as one possible limitation, although we consider it to be justified for the study. We wanted to examine employee well-being and correlations with different health indicators from the widest possible angle and the 131-question Aisti survey had been found useful in occupational health practice. Due to the nature of our analyses, we could not present any causal relationships but instead show only correlations. Furthermore, we did not consider the type of work the employees were performing (eg, office and manual) when conducting the analyses, which may affect the results. There is also a possibility of some sampling bias, as people who consider their well-being to be excellent or reciprocally very bad might not answer the survey. However, in this case, we consider the effect to be most likely small, as our sample included people with both very low and high physical health. We also focused only on examining gender in the analyses, although in the future it could be beneficial to look at other confounders as well. As only a very small amount of data was missing, we consider that the missing data has not affected the results. As our focus was on companies with mainly Finnish employees, the results may not be directly generalisable outside Finland or to, for example, other socioeconomic groups (eg, the unemployed). Furthermore, it would be beneficial to test the clustering on a larger dataset and with a different sample of employees. Nevertheless, we consider the use of a broad instrument, our large dataset and research on multiple physical health indicators as strengths of this study.

Conclusions

We examined whether a subjective well-being assessment can identify different target groups on the basis of their distinct well-being challenges and investigated the links between subjective assessment and clinical health indicators. It is becoming increasingly important to focus on a broader scale of OHS factors that affect employee well-being as the challenges of the field become more and more complex. It is important to find new tools to examine employee well-being and to find ways to better identify underlying problems and the right target groups for different interventions. Currently, the field is filled with different options for targeting, planning and measuring the effectiveness of interventions. However, we still lack cost-effective ways to monitor the success of such programmes. We hope that this study offers OHS researchers and practitioners new insights into how broad-scale subjective assessments could be used to identify different target groups and their needs, as well as to assess the effectiveness of follow-up interventions.

Data availability statement

Data may be obtained from a third party and are not publicly available. Data subject to third party restrictions.

Ethics statementsPatient consent for publication

Not applicable.

Ethics approval

This study involves human participants and used data from the Virta1 RCT trial (Clinical trials NCT04633876), which was approved by the Helsinki University Hospital District Ethical Committee on 4 September 2019 (project no. HUS/2093/2019). Participants gave informed consent to participate in the study before taking part.

Acknowledgments

Special thanks to Tuija Oksanen and Mauri Honkanen for managing and executing the Virta1 trial and Alice Lehtinen for language editing.