Dual trajectories of short-term and long-term sickness absence and their social- and health-related determinants among women in the public sector

Abstract Background Short- and long-term sickness absence (SA) vary in their determinants. We examined short- and long-term SA contemporaneously as two interconnected phenomena to characterize their temporal development, and to identify employees with increasing SA at an early stage. Methods We extracted 46- to 55-year-old employed women from the Helsinki Health Study occupational cohort during 2000–17 (N = 3206) and examined the development of short- (1–14 days) and long-term (>14 days) SA using group-based dual trajectory modelling. In addition, we investigated the associations of social-, work- and health-related factors with trajectory group membership. Results For short-term SA, we selected a three-group solution: ‘no short-term SA’ (50%), ‘low frequency short-term SA’ (40%), and ‘high frequency short-term SA’ (10%) (7 spells/year). For long-term SA, we also selected three trajectory groups: ‘no long-term SA’ (65%), ‘low long-term SA’ (27%), and ‘high long-term SA’ (8%). No SA in the short-term SA model, indicated a high probability of no SA in the long-term model and vice versa. The developmental pattern was far less certain if participant was assigned to a trajectory of high SA in either one of the models (short- or long-term SA model). Low occupational class and poor health behaviours were associated with the trajectory groups with more SA. Conclusion SA does not increase with age among most employees. If either SA rate was high, the developmental patterns were heterogenous. Employers’ attention to health behaviours might aid in reducing both short- and long-term SA.


S
ickness absence (SA) allows employees to recuperate, but simultaneously places a financial burden on society, with relatively high SA rates in Europe. 1 The determinants behind SA are multifaceted; in addition to socioeconomic, health-, and work-related factors, such as occupational demands, [2][3][4][5][6] there is variation associated with political changes in SA benefits, the economy and perceived job security. 7,8hort-term SA (1-14 days) is mainly considered to be absence caused by minor infectious diseases and indolent conditions. 9ontrary to this, studies on municipal employees have identified that earlier short-term SA precedes long-term SA 10,11 and permanent work disability. 12It has also been proposed as a coping behaviour to deal with work-related strain. 64][15][16] Multiple health behaviours, 3,17 socioeconomic factors 2,18 and work environment factors 4 are widely studied risk factors of long-term SA.Older workers and women have more long-term SA, 5,19 but this simplification does not fully represent the variation within these groups.Our previous study on the developmental patterns of long-term SA identified three distinct trajectory groups among 50 to 60-year-old municipal employees. 20espite their intertwined relationship, short-and long-term SA are often studied with short-term SA as the predictor and long-term SA 21 as the outcome and only few studies have analysed their coexistence and development.We consider short-and long-term SA as coexisting but linked phenomena that reflect underlying characteristics of work ability with potentially differing predictors.To model this, we utilize group-based dual trajectory modelling 22 which allows the joint analysis of two separate outcomes without combining them into a single measurement.With dual trajectory analysis an elaborate picture of the connections between trajectory groups can be given.For example, a customary interpretation of a correlation coefficient is that it applies equally to all participants, while alternatively, for some subpopulations the association might be little and for others larger, which can be examined with dual trajectory analysis.Further, examination of the trajectory groups determinants enables studying differences between occupational classes and health outcomes.An analysis of the contemporaneous occurrence and development of these two SA outcomes could facilitate better understanding of SA as a phenomenon.To the best of our knowledge, this is the first study employing dual trajectory modelling in studying short-and long-term SA.
In this study, we examined the dual developmental trajectories of short-and long-term SA among women working in the public sector and considered the trajectories' connections with a range of social-, work-, and health-related factors.

Methods
Our study is part of the ongoing Helsinki Health Study (HHS) that examines health and well-being of the employees of the City of Helsinki, Finland.The HHS cohort consists of employees of the City of Helsinki who turned 40, 45, 50, 55 and 60 years old at study baseline during years 2000-02 (N ¼ 8960, response rate 67%, 80% women). 23Follow-up questionnaires were mailed in 2007 (response rate 83%), 2012 (response rate 79%) and 2017 (response rate 82%).

Study population
We extracted a subsample of the HHS cohort with informed consent to link their survey data to the register comprising information on SA and retirement.We included only women and only years during which their employment contract lasted through the year for the reliability of SA information (N ¼ 5004).We focused on women firstly because most public sector employees in Finland are women and our cohort corresponds to that.Secondly, the largest employment sectors-healthcare and social-are dominated by women.
Based on our previous findings, 20 we included employees who were 46-55 years old during years 2000-17 (N ¼ 4249, excluding N ¼ 755).Participants' SA data were not considered after retirement (excluding N ¼ 55).We further required three years follow-up on SA information either separately or continuously (excluding N ¼ 988).In effect, the minimum follow-up was 3 years, and the maximum 10, with a mean follow-up time of 8.5 years.The final analytical sample size was 3206.

Sickness absence data
The SA information was derived from the City of Helsinki personnel register.We used two outcomes: short-term SA (1-14 days) and long-term SA (>14 days).
The outcomes were the annual number of short-term SA spells and the annual number of long-term SA months.We modelled short-term SA using spells instead of cumulative short-term SA days, since previous research suggests that work disability is initially marked by recurring short-term SA spells. 10,11,24Long-term SA days per year were transformed to account for months due to analytical purposes as follows: 0-13 days ¼ 0 months, 14-29 days ¼ 1 month, 30-59 days ¼ 2 months, etc., and 12 months ¼ 330 or more SA days.

Sociodemographic and socioeconomic factors
Age and marital status were gathered from the baseline questionnaire.We dichotomized marital status as cohabiting (married/ cohabiting) or non-cohabiting (divorced/widowed/single). Occupational class was derived from the employer's personnel register at baseline and classified as follows: professionals and managers (such as teachers, doctors, managers with subordinates), semiprofessionals (such as nurses, foremen), and routine non-manual and manual employees, which combines occupational classes routine non-manual employees (such as child minders, assistant nurses) and manual workers (such as transport workers, cleaners). 25We combined routine non-manual and manual employees because the most common occupations in both groups share a similar characteristic (physically demanding job).

Work-related factors
We extracted information on the employment sector from the employer's personnel register.Four occupational sectors were formed: (i) healthcare (primary healthcare, hospital services), (ii) social (social welfare services, elderly care), (iii) education (primary and secondary school, vocational training) and (iv) other sectors (culture, leisure, rescue, financial administration, transport).
Other work-related factors were gathered from the questionnaires.Work was categorized as regular daytime work (including daytime work with on-call nightshifts) or shift work (including shift work with regular night shifts, regular night work, and work type other).Work-home satisfaction was asked by the question 'How satisfied are you with combining paid work and family?'.Replying satisfied or somewhat satisfied was classified as (work-home satisfaction) 'satisfied' and replying somewhat not satisfied or not satisfied as 'dissatisfied'.

Health-related factors
Smoking was dichotomized into non-smoking (including exsmoking) and smoking (current smoking, daily or occasionally).Body mass index (BMI) was calculated from self-reported height and weight and categorized into healthy weight (BMI <25), overweight (25BMI< 30), and obesity (BMI 30).Leisure-time physical activity was estimated using metabolic equivalent (MET) hours 26 and divided into: low (<14 MET-h/week), intermediate (14 MET-h/week with only moderate activity) and high (14 MET-h/week including vigorous activity).The Jenkins four-item questionnaire was used to indicate sleep problems. 27,28If at least 1 of the 4 symptoms occurred at least 15 times within the past 4 weeks, the participant was categorized as having sleep problems.Covariates were derived from all the questionnaires and the mode was used to determine the most common answer for each participant.

Statistical methods
Group-based dual trajectory modelling utilizes a group-based trajectory model approach which assumes that the population is formed by separate subpopulations (trajectory groups), 22 but instead of studying one outcome, it models two outcomes that are distinct but related in their development.
The dual trajectory model estimates three main parameters to describe individuals' development through time across both outcomes (short-and long-term SA): (i) determining the ideal number of trajectory groups for both sets of measurements, (ii) the probability of membership in each trajectory group and (iii) the probabilities that link membership in trajectory groups across the two measured phenomena. 22The model specifies both trajectories at the same time and the maximum posterior probability assigns participants to the trajectory group with the highest posterior probability in each trajectory model. 29he annual number of short-term SA spells and the annual number of long-term SA months were used as outcomes with a zeroinflated Poisson distribution given that during multiple years the participants had zero SA spells or no long-term SA.First, both SA outcomes were modelled independently.Models were tried with 1-7 trajectory groups and up to third-order polynomials.The number of trajectory groups was based on Bayesian Information Criterion (BIC), Akaike Information Criteria (AIC), entropy, and preference for clinically plausible models that produced trajectory groups with no fewer than 5% of the total sample.A model with three trajectory groups was selected for both outcomes (Supplementary material).
The dual trajectory model was fitted using starting values from the two models.Those with more SA were more likely to drop-out during the follow-up.To account for this non-random attrition, we modelled dropout-probability with dependence on two previous responses. 30We selected a dual model with three trajectories for both SA outcomes (Supplementary material).Mean posterior probabilities were >0.90, indicating a good model fit and a low risk of false classification.
Information on the trajectory group memberships was combined with survey and register data.The association of social-, work-and health-related factors with trajectory group membership was examined using Chi-square tests.We additionally fitted a multinomial logistic regression model of the trajectory group memberships and work-related factors.The trajectory analyses were computed with Stata 17 software's traj-command.R was used for other analyses.
The ethics committees of the Department of Public Health, the University of Helsinki (decision 30 November 1998) and the health authorities of the City of Helsinki (decision 5 October 1999) have approved the Helsinki Health Study.

Results
Routine non-manual and manual employees were the largest occupational class, accounting for half of the employees (table 1).

Short-and long-term sickness absence 323
Healthcare and social sectors were the two largest occupational sectors, accounting for two-thirds of the employees.One-fifth of the employees reported shift work, smoking and sleep problems, and two-thirds were cohabiting.One-fifth reported low leisure-time physical activity.

Short-term SA trajectories
For short-term SA, we selected a model with three trajectory groups (figure 1).Two trajectory groups with the fewest spells-'no shortterm SA' (50%, around 1 spell/year) and 'low frequency short-term SA' (40%, around 3 spells/year)-comprised 90% of the population.The third group, 'high frequency short-term SA', comprised 10% of the population and was well-separated from the two other groups by having a higher frequency of SA spells (around 7 spells/year).Trajectory groups describe the number of short-term SA spells in relation to time (here, age).No considerable change with age was noticed in the frequency of short-term SA spells in any of the trajectory groups (Supplementary material for individual trajectories).

Long-term SA trajectories
For long-term SA, we selected a model with three trajectory groups (figure 1).Employees in the largest group, 'no long-term SA' (65%), had no long-term SA.Individuals in the second largest group, 'low long-term SA' (27%), had on average 2-3 weeks of SA with a gently increasing trend.Individuals in the smallest group, 'high long-term SA' (8%), had an increasing trend with SA rising from one month to four months per year during follow-up.

Relationship between the SA trajectories
The dual trajectory model output produces a summary on the interrelationship of the outcomes (table 2).Regarding long-term SA conditional on short-term SA, those assigned to the trajectory group 'no short-term SA' had an 85% probability of being assigned to the 'no long-term SA' trajectory group and only 3% to 'high long-term SA', meaning that there exists a clear interrelationship with the trajectory groups as employees taking no short-term SA were very likely to follow a path without SA in the long-term SA model.In contrast, when assigned to the trajectory group 'high frequency short-term SA' the same probabilities were 24% and 23%, respectively.This embodies uncertainty in the long-term SA trajectories of employees who take multiple short-term SA spells.This is denoted by the equal probability of following either the trajectory with no long-term SA (24%) or the trajectory with high long-term SA (23%), while there is still a 53% probability of following a path with low long-term SA.
Second, looking at short-term SA conditional on the long-term SA trajectory group, those assigned to the 'no long-term SA' trajectory group had a 66% probability of being assigned to 'no short-term SA' and only 4% probability of being assigned to the 'high frequency short-term SA' trajectory group.When assigned to the 'high longterm SA' trajectory group, the probabilities were 20% and 28%, respectively.That is, the probability of belonging to any specific short- term SA group when assigned to the 'high long-term SA' trajectory group was far less certain, implying population heterogeneity in the developmental course of SA.Lastly, the joint probability memberships show a fair amount of overlap but also unexpected developmental patterns.Having more short-term SA spells is linked to a higher possibility of more longterm SA, but the trajectories do not always develop in the same direction; those having frequent short-term SA spells do not always end up with high long-term SA and not all participants leading to the 'high long-term SA' trajectory group could have been foretold by their short-term SA patterns.

Determinants of trajectory group membership
Employees in the lowest occupational class were more prevalent in the trajectories with high SA in both SA models (table 1).The healthcare and social sectors showed increased prevalence in trajectory groups with high SA.Similarly, shift work, being dissatisfied in combining paid work with family life, and non-cohabiting were more prevalent in the trajectory groups with low or high short-or long-term SA.Regarding health-related factors, smoking, sleep problems, higher BMI and low leisure-time physical activity were more prevalent in the low and high SA trajectory groups than no SA trajectory groups.The differences were clearer in the long-term SA model.
Differences between employment sectors remained when adjusting for occupational class and work type in the multinomial logistic regression model for trajectory group membership and work-related factors (Supplementary table).The social sector was associated with a higher likelihood of being assigned to 'low frequency short-term SA' and 'high frequency short-term SA' trajectory groups and healthcare to the 'high frequency short-term SA' trajectory group in the short-term SA model.In the long-term SA model, the healthcare and social sectors were associated with a higher likelihood of being assigned to the 'high long-term SA' trajectory group compared with employment sector 'other'.

Sensitivity analysis
Attrition during follow-up was highest in the 'high long-term SA' trajectory group (median follow-up 7 years).This reflects the general idea of long-term SA preceding disability pension and the legislation in Finland according to which an employee must have been on longterm SA prior being granted disability pension.We tested if the trajectories would change when analysing shorter time periods.The short-term model produced similar trajectories when analysed separately during ages 46-50 and 51-55.The long-term model produced similar trajectories in shape, but the high long-term SA trajectory was around 3% smaller.Hypothetically, fewer employees had time to move to disability pension during a shorter period.We tested the model with both genders and the trajectories were not significantly altered.

Discussion
We discovered that SA rates did not increase with age among most employees.The number of short-term SA spells did not increase with age and only a tenth of the employees were assigned to a high frequency short-term SA trajectory.A third of the employees had some increase in long-term SA with age, but only 8% of all the employees were assigned to a trajectory group with notably increasing long-term SA.The short-and long-term SA trajectories were interconnected, and population heterogeneity was noticed.The descriptive results were as expected; lower occupational class and Notes: Conditional probability describes how likely an individual is to be in each long-term SA trajectory group if a person is known to be in a given short-term SA trajectory group.Joint probability describes how likely overall it is for a person to fall into any combination of the short-and long-term SA trajectory groups.Joint probability membership is presented for every trajectory group combination, nine joint probabilities, that sum up to one.The names express the trajectory groups' name and the percentage (%).SA, sickness absence.
Short-and long-term sickness absence 325 poorer health behaviours were associated with the trajectory groups with more SA.Employees in healthcare and social sectors were more likely to be assigned to a high SA trajectory group even when adjusting for occupational class and work arrangements.Few previous studies have examined SA trajectories among women.A Spanish study identified three trajectories with most employed women being assigned to a low stable trajectory, 31 the percentage of which is in line with our results.A Finnish study examined the combined short-and long-term SA trajectories of municipal and private sector employees among both men and women, and the smallest group with the most SA had around 60 days absence per year, which increased only slightly during follow-up. 32This is roughly the mean SA length in our trajectory group 'high long-term SA'.They, however, used calendar years, whereas we examined SA with age using a longer follow-up.
Stability in the number of short-term SA spells with age reflects the general understanding of SA; short-term SA is either selfcertified or administered due to minor diseases or in the beginning of a more severe disease and short-term SA spells are more associated with motivational factors. 33However, not only individual differences, but also occupational class differences are likely to explain this pattern, as those in higher occupational classes might more easily be able to work while sick and have better possibilities to modify working tasks.In the smaller subgroups with high shortterm SA, SA could perhaps be reduced by targeting work arrangements and motivational factors.If so, the frequency of SA spells would be utilizable in screening while considering that the developmental patterns of short-term SA show heterogeneity in the longterm SA patterns.
Consistent occupational class differences in SA were found.Over 70% of the participants assigned to high SA trajectory groups were routine non-manual employees and manual workers while this occupational class accounts for half of the study population.Especially long-term SA is strongly associated with health and these findings reflect on the general health disparities between occupational classes.
A recent Swedish study found that employees in femaledominated workplaces are at higher risk of SA compared with gender-equal or male-dominated workplaces. 34Our findings suggest a similar pattern in our data, though limited to only one large public sector employer in Finland.Two occupational sectors where the majority of employees are women-healthcare and social-appeared more prone to prolonging SA or more frequent short-term SA spells.Reasons for this can be speculated; job descriptions in healthcare and social sectors are both mentally and physically demanding, and employees might do shift work or suffer from infectious diseases.In addition, staff shortages or ethical dilemmas might affect employees' motivation and increase the likelihood of SA.On the other hand, no SA differences were noticed in the employment sector teaching, which is also a sector with more women and contact with infectious diseases.Differences might exist in absence culture or in the use of SA to cope with work-related strain.
Multiple work-and health-related factors have been linked to a higher risk of short-or long-term SA and prolonging absence among women, and our results correspond to that.Among nurses, working arrangements such as night shift work have been associated with longterm SA in Denmark and Finland, 35 and long weekly working hours with short-term SA 36 in Finland.Our results with a novel method support the previous well-established association of health-related factors and SA. 3 The predictors of short-and long-term SA were largely similar, with those reporting smoking, low leisure-time physical activity, poorer sleep and higher BMI being more prevalent in groups with higher SA rates.More adverse health-related factors and SA accumulate among a minority of employees.
The double burden hypothesis has been proposed as the explanation for women's higher SA prevalence. 18,19A previous review suggests measuring experienced family stressors instead of the number of children as a better tool to examine this burden of multiple roles 37 and previous studies have reported an association between work-family conflict and health. 38Our results support these findings, those with more SA more often replied being dissatisfied with combining paid work and family.

Strengths and limitations
Strengths of the study included a large cohort data, a novel method to study SA patterns and the possibility to include four welldocumented health-related risk factors for SA. 3,39gure 1 The developmental trajectories of sickness absence (SA) of employed women at the city of Helsinki between ages 46-55 (N ¼ 3206).Left panel: The developmental trajectories for the number of short-term SA spells per year during ages 46-55: 'no short-term SA' (50%), 'low frequency short-term SA' (40%) and 'high frequency short-term SA' (10%).Right panel: The developmental trajectories of the number of long-term SA months per year during ages 46-55: 'no long-term SA' (65%), 'low long-term SA' (27%), 'high long-term SA' (8%).SA, sickness absence; freq., frequency The group-based dual trajectory model produces an informative picture of SA development; however, trajectories are always an approximation of reality.The model divides the population into distinctive trajectories with as much similar developmental features as possible.Trajectory group membership is assigned according to which group the individual has the highest posterior probability of belonging to.Conclusions on the trajectories' determinants should therefore be made with care.The dual model's conditional and joint probabilities provide information on the distribution of probabilities and heterogenic continuity, but we cannot assess causality.
We analysed SA development among employees in midlife, thus selecting those healthy enough to stay in the work force.The healthy worker effect should therefore be considered, meaning that people with a diminished work ability might have been selected out of the population prior to our study. 40Exiting the study was also more likely among employees with more long-term SA.We did not have the specific diagnostic codes for the SA.We have, however, analysed the most common diagnosis-specific long-term SA trajectories in our previous publication. 20

Conclusions
We discovered that the development of short-and long-term SA is interconnected but that these outcomes do not always develop in the same direction.SA does not increase among most ageing employees, but when the SA rate is high, SA develops in varying ways.Adverse health-related factors appear as significant determinants for both short-and long-term SA, and there were differences in proneness to SA between employment sectors.

Table 1
Characteristics of the study population: sample of women employed at the city of Helsinki aged 46-55 years during years 2000-17 (N ¼ 3206) and the determinants of sickness absence (SA) trajectory groups; Panel A: short-term SA model.Panel B: long-term SA model Notes: Values express N (%).SA, sickness absence; BMI, body mass index.

Table 2
The linkage between short-term and long-term SA trajectory groups presented by conditional and joint probabilities from the group-based dual trajectory model Probability of long-term SA trajectory group (k) conditional on short-term SA trajectory group ( j) [p kjj ]