Individual, Workplace, and Combined Effects Modeling of Employee Productivity Loss

Abstract

Objective: 

Systematically and simultaneously investigate a wide range of influences on workplace productivity loss.

Methods: 

Data were collected from 31,950 employees in the UK. Influences of employees’ socioeconomic characteristics, lifestyle, commuting, physical and mental health, well-being, and job and workplace environment were assessed using structural equation models, allowing systematic decomposition of the complex network of influences and creating new, deeper insights.

Results: 

Mental health, physical health, job characteristics, and support from organization are the most important (direct or indirect) determinants of employees’ productivity. 93% of the indirect influences are mediated through mental and/or physical health. Some influences that appear as strong predictors in simple models lose most of their explanatory power in more complex models with additional explanatory variables.

Conclusions: 

There is a need for a more tailored strategy to improve employees’ wellbeing as well as the overall organizational, work, and management culture.

Comprehension of employees’ productivity and performance has been a goal of academics and organizations alike for decades. Initially, the major research focus was on absenteeism, generally defined as not showing up for work, due to its evident implications for performance of organizations and relative ease of measurement.¹ However, more recently presenteeism has been gaining attention as it is suggested to cause higher aggregate productivity loss than absenteeism.^2,3 Presenteeism is defined in different ways in the literature, most often as being at work with decreased performance, productivity, and work quality due to a health problem or injury.^4,5

A multitude of absenteeism- and presenteeism-related factors has been suggested and empirically tested as determinants of productivity loss, including physical and mental health,^6,7 lifestyle,⁸ personal and family factors,⁹ relationships,¹⁰ work strain¹¹ or job, and workplace characteristics.¹² However, the studies so far have considered a narrow pathway of effects and have largely overlooked interrelations between influences. Moreover, limiting analysis to a small set of variables may conflate various separate, though often correlated, effects and hence fail to capture the main underlying explanations.

The present study aims to shed a light on this research gap through simultaneous analysis of an extensive set of workplace productivity factors. We utilize a unique, extensive cross-sectional dataset covering 29,928 employees across the United Kingdom collected by the Britain's Healthiest Workplace, an annual online survey of organizations and their employees in the UK, in 2017 (see Section Data for more information). Self-reported productivity, measured by the Work Productivity and Activity Impairment Questionnaire¹³ is used as the outcome variable; the determinants are evaluated using an advanced structural equation model (SEM), which combines the direct and indirect influences (eg, path analysis) with factor analysis—a technique to reduce the dimensionality of highly correlated variables into lower dimensional latent space.

More specifically, in line with previous studies, we first develop separate models with a limited number of variables to evaluate productivity, each emphasizing a different angle of either personal, job-related or workplace influences. The results, which are shown to be in line with the findings from the prior studies, allow us to assess the baseline influences. Second, we conflate all partial models to form a combined SEM framework in order to evaluate the most important influences and their path structure. Additionally, comparing the partial and combined models helps to make inference on the extent to which narrow-focused partial models are affected by the omitted-variable bias.

Given our limited control over workers’ participation in the underlying survey, our results may be affected by a self-selection bias despite the large size of our dataset. In order to provide evidence that this is not a particular issue, we compared characteristics of the survey respondents to a representative sample of the working population in the UK using data from the Health Survey for England¹⁴ across a multitude of variables, including age, sex, ethnicity, and body mass index (BMI). The results, presented in Tables S1 (Supplemental Digital Content 1, https://links.lww.com/JOM/A548) and S2 (Supplemental Digital Content 2, https://links.lww.com/JOM/A549) in the Online Appendix, show that both samples are similar in their characteristics and estimated influences are directionally equivalent. Further information on this and other potential limitations is provided in Section Limitations and their mitigation. A draft version of this study was previously published as.¹⁵

Existing Research

Employee productivity at the workplace is a measure encompassing both efficiency (time or other resources required for completing a task) and effectiveness (the degree to which objectives are achieved and/or targeted problems are solved). Research focus is often on relative productivity loss, for example, a comparison of an individual's performance to an optimal or past performance levels or to that of other employees.^16,17 Our study follows this direction and analyzes self-reported productivity loss compared with an optimal state. One common way to evaluate productivity loss is through absenteeism and presenteeism (eg, the share of time that an employee did not work or worked with a limited efficiency when compared with their contract or expectations). Below we discuss the most common individual and organizational factors associated with absenteeism and presenteeism.

Lifestyle, Physical, and Mental Well-Being

Health assessment includes factors such as diseases, biometric indicators, life satisfaction, anxiety, depression or fatigue, and is substantially affected by one's lifestyle (eg, physical activity, nutrition, alcohol use, smoking, sleeping patterns, or other wellness behaviors).¹⁸ On the other hand, one's lifestyle depends on his or her health; this inherently creates two layers of effects in which both health and lifestyle affect productivity but also each other.

Some of the studies showing the negative effects of poor physical health include^19,20 who analyze obesity^21,22; who look specifically at individuals with diabetes; or²³ and²⁴ who target musculoskeletal conditions. Similarly, a number of mental health indicators have been used as productivity determinants. For instance,²⁵ show the importance of personal problems, financial concerns, and depression, while,^6,26 or²⁷ found positive correlation between indicators of happiness and productivity. Finally, lifestyle indicators have then been evaluated (eg, see,²⁸ who look at physical activity and psycho-social health^29,30; who look at the burden of alcohol consumption; or,^31,32 who estimate the impact of insufficient sleep). All of the risk factors have been shown to have statistically significant effect on productivity.

Job Attitude and Characteristics

Job attitude is a set of feelings toward, beliefs about, and attachment to one's job. Research shows that it is mainly determined by job characteristics,³³ yet it may also be affected by a worker's characteristics or emotional moods.³⁴ The determinants of job attitudes have been studied, for example.^35–37 Moreover, job characteristics may also have an impact on lifestyle, as well as physical and mental health, in turn affecting productivity indirectly and highlighting the need to be considered simultaneously.

Job attitude can be measured by job satisfaction or employee engagement. Work engagement has received particular attention in the empirical literature (eg,^16,38–41) and has been shown to have strong causal influence on workplace productivity. However, as we discuss further in our study, this may be partially due to omission of other relevant factors in the analysis. The strong influence of job satisfaction and other job characteristics, such as work-related stress, on productivity has been shown.^34,42–44

Workplace Characteristics

Workplace and organizational characteristics represent the broader attitude of an employer and its management toward its employees. Examples include company values, support from management, well-being offerings, fairness in treatment, and appraisal or open and honest communication. Again, such factors are generally beyond an individual's control although he or she may be affected by attitude toward job and workplace.

The influence of workplace and organizational factors has been investigated, for example,^12,45–48 who provide a systematic literature review investigating the effect of on-site workplace health and well-being interventions on employee productivity, showing that organizations with better well-being offer see lower productivity losses.

Analytical Approach

Method of Analysis

Many empirical studies presented in the literature review focus on a limited number of factors when analyzing influences on workplace productivity loss; this risks misinterpretation of results due to the omitted variable bias. In this study, we use integrated SEMs to simultaneously estimate the complex network of influences of physical and mental health, lifestyle, personal and family factors, and job and workplace characteristics on relative workplace productivity loss. SEM is a theory-driven data analytical approach for evaluation of a priory specified hypotheses about causal relations among measured and/or latent variables.^49,50 SEMs allow systematic decomposition of the complex network of influences, where the effects of each variable can be examined in turn; the framework is thus particularly useful in evaluating productivity determinants.

The analysis is done in two steps. First, highly correlated influences, which are difficult to model as separate explanatory variables are grouped to construct distinct latent variables such as lifestyle, mental health, and physical health; exploratory factor analysis (EFA) is used to investigate how the latent variables are formed and manifested into the observed data space. Second, causal links between the constructed factors and directly observed variables are specified in a structural model. Here, both direct and indirect (mediated through other variables) impacts can be modeled to create a network of effects on productivity. Capturing commonality across explanatory variables by creating latent variable indicators also helps to avoid potential measurement errors and endogeneity issues.⁵¹

In what follows, we first develop three separate SEMs, namely personal, job, and workplace models, based on the conceptual structures suggested in the literature (refer to section Existing research). This involves employing factor analysis to construct a set of latent variables (factors) to measure the main conceptual elements (eg, mental health or job attitudes), followed by developing separate SEM frameworks, one for each individual model, to test influences of the constructed latent variables alongside standalone-observed variables on productivity. The individual models can be used to directly compare our findings using a limited set of variables against those reported in the prior literature and build a base for comparison against the full model.

Second, we combine all individual models into a single combined SEM framework, reducing the omitted variable bias and allowing to account for additional interrelations across variables. Consequently, we can postulate the relative importance of influences and highlight those with the most significant effect on productivity loss, as well as any differences compared with the individual models. The theoretical foundations of our model are based on,^16,52,53 who construct various individual- or workplace-related models. Their analysis suggests that personal resources, job demands, and job resources, which are estimated independently, affect employee's health and job attitudes.

Data

We use the data collected through the 2017 Britain's Healthiest Workplace survey of organizations and their employees in the UK. The survey is open to all organizations in the UK with more than 20 employees from any industrial sector; participating organizations self-select to the survey and distribute the survey links to their employees. There is no fee for participation nor a selection process for participants. All employees aged 18+ are allowed to complete the survey, yet their participation is voluntary and results are anonymized. The survey covers personal, social, lifestyle, job and workplace areas. Since its inception in 2013, more than 370 organizations and 124,000 individual employees participated in the study.

Our data consists of the entire 2017 cross-sectional dataset of 31,950 employee responses across all 173 participating organizations. The sectoral breakdown of the organizations and their respondents is provided in Table S3 (Supplemental Digital Content 3, https://links.lww.com/JOM/A550) in the Online Appendix. After excluding responses with missing data, the final dataset consists of 29,928 individuals. With more than 90 indicators per individual in the original dataset, each variable was carefully considered for inclusion in the study based on the intuition, statistical tests of wide range of models, and the prior literature, eventually limiting the number of included variables to 36 plus six controls. The variables were allocated to the three theoretical frameworks discussed above: the personal, job, and workplace. Each individual model thus consists of variables intuitively related to each other (eg, diet, alcohol consumption, smoking and sleep patterns; or job satisfaction, job-related stress, isolation, and safety at the workplace) that have also been proven interlinked in the previous studies.

The full list of variables is presented in Table 1 with their descriptive statistics reported in Table S4 (Supplemental Digital Content 4, https://links.lww.com/JOM/A551) and Table S5 (Supplemental Digital Content 5, https://links.lww.com/JOM/A552) in the Online Appendix. Most of the variables represent a single question in the survey, while several variables represent composite indices—a combination of multiple questions—created and validated in the prior literature. For instance, mental well-being is measured using the Kessler Psychological Distress Scale,⁵⁴ whereas work engagement is measured using the Utrecht Work Engagement Scale.⁵⁵ Our main outcome variable, relative employee-level productivity loss, is a self-reported variable captured using the Work Productivity and Activity Impairment questionnaire.¹³ The list of composite variables and their structure is shown in Table S6 (Supplemental Digital Content 6, https://links.lww.com/JOM/A553).

TABLE 1: List and Categorization of Variables Used in the Analysis

In order to simplify interpretation of results, we transformed the variables where appropriate so that, for all variables, higher values represent the preferable outcomes. For instance, the question on feeling isolated at workplace, originally on a 0 (never) to 4 (always) scale was transformed to 0 (always) to 4 (never) scale. Productivity is measured relative to the maximum potential individual performance at the workplace and is reported on a (0; 100%) scale, where 100% equals maximum productivity and lower values represent presence of self-reported absenteeism and/or presenteeism. The analysis was done in Stata 15.⁵⁶

Analysis

Factor Analysis

Given the high level of interrelation among variables in our dataset, they cannot be treated as independent; this prompts us to construct latent variables (factors) to represent the conceptual content of the interrelated indicators. We use EFA to establish associations among variables and create variable clusters defining composite, latent variables that are better capable of representing the pattern of influences than any of the constituents.⁴⁹

We performed EFA for each variable category presented in Table 1 to classify the associated variables into an optimal number of factors. We examined goodness of fit for the constructed factors iteratively, starting with simple models with no covariance structure between the factor indicators’ measurement errors and moving toward more complex models where correlations between some or all of the measurement errors are controlled for. In each step, the Comparative Fit Index (CFI) and difference in the chi-squared statistic in relation to the associated difference in the degrees of freedom are used to assess suitability of the model.⁵⁷ The final constructed factors with the best goodness of fit indices are then used in the SEM frameworks to analyze the influences on productivity loss. Additional information on the procedure and results are summarized in Tables S7–S9 (Supplemental Digital Content 7, https://links.lww.com/JOM/A554, Supplemental Digital Content 8, https://links.lww.com/JOM/A555, Supplemental Digital Content 9, https://links.lww.com/JOM/A556) in the Online Appendix.

Most categories in Table 1 can be represented by a single latent variable. For instance, mental health consists of four observed indicators: mental health self-assessment, Kessler Psychological Distress Scale, life satisfaction, and measure of financial concerns. On the contrary, lifestyle variables, which show little intercorrelation, were found unfit for constructing common latent variables and are therefore considered as separate observed independent variables in the analysis. Through factor analysis, we eventually reduced dimensionality of our data space to 19 variables (consisting of seven factors, six observed variables, and six control variables—age, sex, ethnicity, education, income, and job position). In the SEM diagrams presented further (see Figures 1–4), the variables are represented as follows: individual independent variables (boxes) and latent variables (ellipses) used in the final models are shown with white background; variables forming each latent variable are shown with dark background. We discuss each of these models in turn in the following sections.

FIGURE 1: Personal model. The diagram represents the model with the best fit to the observed data. All coefficients are standardized and represent change (in standard deviations) in a dependent variable per standard deviation increase in the predictor variable. Only coefficients with significant direct or indirect influence at less than 0.001 level are shown. Variables with dark background represent indicators associated with latent variables.

FIGURE 2: Job model. The diagram represents the model with the best fit to the observed data. All coefficients are standardized and represent change (in standard deviations) in a dependent variable per standard deviation increase in the predictor variable. Only coefficients with significant direct or indirect influence at less than 0.001 level are shown. Variables with dark background represent indicators associated with latent variables.

FIGURE 3: Workplace model. The diagram represents the model with the best fit to the observed data. All coefficients are standardized and represent change (in standard deviations) in a dependent variable per standard deviation increase in the predictor variable. Only coefficients with significant direct or indirect influence at less than 0.001 level are shown. Variables with dark background represent indicators associated with latent variables.

FIGURE 4: Combined model. The diagram represents model with the best fit to the observed data. Uncorrelated error variables are omitted for clarity of presentation. Solid black lines highlight pathway coefficients greater than 0.1. All the paths shown are statistically significant at less than 0.001 level.

Individual Models of Workplace Productivity

In this section we analyze the influences on workplace productivity loss. First, we develop three individual SEMs broadly based on the frameworks established in the literature. Subsequently, we combine the three models into a single framework in order to account for potential interrelations.

For each individual model, we tested various SEM structures with different directional effects in order to examine the one with the best fit to the data. The Akaike information criterion and Bayesian information criterion measures were used to determine model fits. Throughout the paper, only the statistically significant influences (P < 0.01) in the optimal models are reported in the path diagrams.

Personal Model—Physical and Mental Well-Being, Lifestyle

The path analysis for the personal model is depicted in Fig. 1. All coefficients are standardized so that they can be directly comparable. In addition to the direct effects, such as that of mental health on productivity (with the coefficient of 0.296), one major advantage of SEM is the estimation of indirect influences, which are quantified by multiplying the coefficients along the SEM paths. In our reported results, all indirect effects are standardized after being estimated from multiplication of direct influences. The total or combined effects are then calculated by adding up the direct as well as all indirect effects. The full list of direct, indirect, and combined effects and comparisons with the combined model are provided in Table 2. For instance, as shown in the personal model column, the indirect effect of mental health on productivity via physical health and work engagement is 0.124.

TABLE 2: Direct and Indirect Influences on Productivity

Mental and physical health have the strongest direct influences on workplace productivity. Additionally, mental health has a strong indirect effect mainly mediated through physical health (86% of total indirect effect). This can be calculated using the combined effect of 0.441, which is the sum of the direct effect (0.296) and all indirect effects (0.145), compared with the indirect effects mediated through mental health (0.124, see Table 2 for more information). Other influences are negligible, with only work engagement having statistically significant direct effect on productivity. All pathway coefficients, except for the one between smoking and work engagement, are positive, meaning that better health and lifestyle generally increase work engagement and productivity. Most of the control variables do not have statistically significant effect on productivity (refer to Table S10 [Supplemental Digital Content 10, https://links.lww.com/JOM/A557] in the Online Appendix for full results).

To compare our results with the prior literature, we developed a model with only work engagement as an explanatory variable in addition to the full set of controls (eg, the effects of lifestyle, mental health, and physical health are disregarded). The estimated standardized coefficient of 0.267 (see Table S10 [Supplemental Digital Content 10, https://links.lww.com/JOM/A557] in the Online Appendix) is in line with the overall findings from previous studies (refer, eg, to Refs.^38–40). However, after including lifestyle and health variables in the model, as shown in Fig. 1, the influence of work engagement becomes negligible, suggesting that mental and physical health are in fact the major influences on productivity loss and it is their effects which are captured by work engagement when they are excluded from the model.

Job Model—Attitudes Towards Workplace

The job model is depicted in Fig. 2. Commuting time is statistically significant though it has a small impact on productivity and all other variables (note that higher values represent shorter commuting time). Work patterns also have relatively small impact on both work engagement and productivity. On the other hand, the effect of job characteristics—both direct and indirect—is substantially bigger and positive, meaning that employees facing less stress at work and those more satisfied with their job are, on average, more productive. While this confirms previous findings of, for example,^34,42,43 we will see later that the direct effect decreases by more than 50% once other variables are introduced. Regarding work engagement, its effect on productivity is higher than that in the personal model, yet it is relatively small compared with that of job characteristics.

Workplace Model—Workplace and Organizational Characteristics

In line with what was suggested in the previous literature (eg,^12,45), the support from an organization and its managers are both statistically significant predictors of work engagement and productivity (refer to Fig. 3). Both factors also affect productivity indirectly, through work engagement. Note that the pathway coefficient between work engagement and productivity is twice as high than in the previous models and generally more in line with the simple model presented in Table S11 (Supplemental Digital Content 11, https://links.lww.com/JOM/A558). The results thus again suggest that there are other unobserved effects affecting the results and that such simple models fail to properly capture the actual pathway of effects. Influences of control variables are in line with the previous models with sex, age, and income being statistically significant predictors of all four dependent variables.

The Combined Model of Workplace Productivity

By integrating our findings from the individual models we developed a joint, comprehensive model of workplace productivity to determine the relative importance of influences. This is again done iteratively; the individual models are linked in one combined SEM framework and an optimal model is found after testing alternative model structures and comparing the Akaike information criterion and Bayesian information criterion values. That is, we took all statistically significant direct and indirect influences, keeping previously discovered pathway of effects. We then added all possible new links and tested both the significance of influences and the goodness of fit of the combined models in order to choose the one best describing our data. For instance, smoking was identified as a codeterminant of physical health, which in turn affects work engagement and, both directly and indirectly, productivity. Equally, work patterns affect work engagement and productivity in the second individual model. In the combined model, we further tested the relationship between smoking and work patterns, physical health and work patterns, and whether work engagement remains a statistically significant determinant of productivity when additional variables are included. The analysis showed that none of the new pathways (including the work engagement–productivity link) were statistically significant and were therefore not included in the final model. The resulting model is illustrated in Fig. 4 and the direct and indirect influences on productivity are presented in Table 2.

The results show that, once all the factors established in the three individual models are combined, the most important influences on workplace productivity are: mental health, physical health, job characteristics, and support from organizations, all of which are indeed also major influences highlighted in the separate individual models. Of these, mental health stands out as the most important determinant. On the other hand, working patterns show relatively small influence on workplace productivity, suggesting that the way employees work is no longer a key factor.

As in the individual models, most of the indirect effects (93%) in the model are mediated through mental and/or physical health. For instance, more than 50% of the influence of job characteristics, the second most important productivity determinant, is mediated through mental health. Second to mental health, physical health has the strongest direct effect on productivity; it also acts as a mediating factor specifically for the indirect influences of mental health and job characteristics.

Note that majority of pathway coefficients are smaller in size in the combined model compared with the individual models. This is in line with our expectations as the simpler models tend to overestimate the observable influences by capturing some of the effects of omitted variables. For instance, the direct effect of mental health is 9% weaker when compared with the individual model; for physical health, the difference is 7% and then substantially greater for all other factors (eg, 52% for job characteristics or 73% for support from organization). These findings highlight the potential misinterpretation of influences when only parts of the causal paths are modeled.

The benefit of a more inclusive combined model can be better perceived by examining the effect of work engagement, which appears as a significant productivity determinant in the individual models but completely lost its significance in the combined model. The effect of work engagement was strongest in the job and workplace models, where it additionally served as a mediating variable for job characteristics and for support from organization and managers. For the combined model, however, the role of work engagement has been diminished and is replaced with a more complex path structure. In the combined model, support from organization and managers affects job characteristics, which in turn has a direct influence on productivity, but it also exhibits a significant indirect influence through mental and physical health. This suggests that, arguably, the simpler individual models have overestimated the impact of work engagement by capturing the effects of other potentially important influences which were not included in the models.

Considering the four individual lifestyle variables—alcohol consumption, smoking, sleep length, and commuting time—only commuting time is estimated to have statistically significant direct effect on productivity, while even the indirect effect of alcohol consumption is not significant when all other factors are included in the analysis. Looking at Table S10 (Supplemental Digital Content 10, https://links.lww.com/JOM/A557) in the Online Appendix, the influence of control variables remains remarkably consistent with those found in individual models. Sex, age, income and selected ethnicity and job position indicators have statistically significant influence on productivity. In particular, men, older employees, and those with higher income tend to report higher productivity.

DISCUSSION

Discussion and Conclusions

In this study, we present a novel insight into understanding employees’ productivity by developing a new conceptual model augmenting simpler frameworks assessed in the prior literature. This is possible through exploiting granularity of the 2017 Britain's Healthiest Workplace survey, which provide detailed information on a large set of socioeconomic and workplace characteristics, as well as various personal and institutional variables for more than 30,000 employees.

Our principal findings are threefold. First, our results show that mental health, physical health, job characteristics, and support from organizations are the most important determinants of employees’ productivity. This highlights a strong case for promoting workplace interventions aimed at improving employees’ wellbeing and the overall organizational, work, and management culture.

Second, our study shows that the network of influences affecting employees’ productivity is more complex than what hitherto presented in the literature. Disentangling the pathway of influences, we show that a large proportion of effects that support from the organization and managers, as well as workplace conditions and attitudes more generally, have on workplace productivity are mediated through mental and physical health. This highlights the need for a more tailored strategy to improve employees’ well-being. Indeed, employers typically focus on addressing the symptoms of poor mental and physical health through investing in comprehensive medical benefit packages as well as employees’ assistance programs. Our study suggests that it is equally or even more important to address the source of such problems through supportive management, promoting more inclusive work atmosphere and improving job satisfaction in a healthy work environment.

Third, from the technical perspective, our study shows that the simple individual models carry the risk of overestimating influences as a result of neglecting important direct and indirect influences. For instance, even though work engagement appears as a strong predictor of workplace productivity in the individual models, inclusion of other variables such as health, job, and workplace characteristics reduce its explanatory power a negligible amount. Similar studies aimed at assessing productivity or associated employee or workplace factors should therefore carefully design the evaluation framework to minimize the potential omitted variable bias.

Limitations and Their Mitigation

Despite using a very comprehensive dataset, the present data and analytical approach have their limitations. In particular, the participation rates per organization are unknown as the full lists of eligible employees were not disclosed. We have carried out extensive tests to understand and tackle the associated potential limitations. First, given that our results may be affected by a self-selection bias, we compared characteristics of the survey respondents to a representative sample of the UK working population from the Health Survey for England¹⁴ across a multitude of variables, including age, sex, ethnicity, and BMI. The results, presented in Table S1 (Supplemental Digital Content 1, https://links.lww.com/JOM/A548) in the Online Appendix, show that both samples are similar in their characteristics. Additionally, we run two equivalent regressions with life satisfaction as the dependent variable using both samples to compare the influences. The results of this test, shown in Table S2 (Supplemental Digital Content 2, https://links.lww.com/JOM/A549), show that the estimated influences are similar, with the same direction of effects and level of significance. The only exception is ethnicity; our dataset contains 10 percentage points more white individuals and proportionally less blacks, Asians, and individuals of mixed ethnicity. Consequently, white ethnicity appears as a statistically significant variable in the regression using the study dataset, but it is reported as insignificant in the representative dataset (although with a positive coefficient in both cases). Given that ethnicity has rarely appeared as a statistically significant variable in the main SEM results, we argue that the outcomes are representative of the general working population in the UK.

Second, measurement errors in our study are controlled for using latent variables that capture commonality across the correlated indicators. However, a separate potential issue with self-reported variables, such as those in our dataset, is that some variables may be consistently under- or over-stated as a result of psychological biases. To control for this, we modeled correlations among error terms in the exploratory factor analysis and, where statistically significant, we included them in the final SEMs.

Third, to control for a possible endogeneity bias and structural ambiguity, we estimated numerous model structure with pathways in opposite directions to check which model structure would fit the data best and used that for our final estimates.

Finally, our main outcome variable—productivity loss assessed using the Work Productivity and Activity Impairment questionnaire¹³—is self-reported and thus prone to biases as documented by.⁵⁸ Notwithstanding that, self-reported relative productivity loss assessments are extensively used in the literature (eg, Refs.^59,60). It will be useful for future studies to examine alternative survey designs to come out with the best approach for recording productivity metrics which can minimize the potential biases in assessment of physical and mental health influences.