Introduction

At the end of the 1990s (starting 1997) and the beginning of the 2000's sickness absence in Sweden increased dramatically and the total expenditures for the state rose by about 50% from 1999 to 2002 [1]. A significant amount of research has been conducted trying to understand the large increases in sickness absence; some researchers and policy makers have highlighted factors such as the ageing population, increased pressure in working life, sick leave as a disguise for unemployment, shortcomings in the sickness insurance system and rehabilitation services, and a more restrictive policy regarding early retirement [1]. Musculoskeletal disorders (MSDs) are the most common cause of sickness absence and being the disorder group causing the largest costs for society in the social insurance system, whereas the largest relative increase in sickness absence during this period was mostly due to an increase in sickness absence due to psychiatric disorders [2].

However, there is conflicting evidence regarding several of the suggested causes of the large increase; e.g. estimations indicate that the ageing population (the elderly have higher sickness absence) can only explain a very small share of the total increase in sickness absence [3]. Further, it may also be argued that it is quite unlikely that changes in rehabilitation services, increased pressure in working life etc. would change so dramatically in a few years as to explain the large increase in sickness absence seen for a few years starting 1997. Hence, it may be argued that the large increase in sickness absence during the above mentioned years is still not well understood.

Since 2002 the expenditures on sickness cash benefits have decreased by about one-third up until 2007 [4]. One reason for this is likely the stricter interpretation of the eligibility rules for sickness benefits. Still, in 2007, government expenditures on sickness cash benefits were approx. 27.1 billion Swedish kroner (approx. €2.5 billion)Footnote 1 [4]. As mentioned, MSDs are the most common causes of sickness absence in Sweden, followed by psychiatric disorders [2]. Further, in a European context, MSD is the major cause of work-related health problems. According to a European survey conducted in 2005, about 25% of the workers in the European Union countries reported back pain problems and 23% reported muscular pain problems. The problems were generally more common among blue-collar workers compared to white-collar workers [5]. Hence, determinants, treatments, and consequences of sickness absence due to MSD are important issues in terms of individual health, population health and costs for the social insurance and health care systems.

There is increasing evidence that staying active is an important part of a recovery process for individuals with MSD and related disability, and that total absence from work delays recovery [6–10]. Hence, a partial return to work from a sick-leave due to MSD may be beneficial for the individual’s health and lead to a quicker recovery of the lost work capacity [11]. Further, most of the employees are also satisfied with being on part-time sick leave [12]. The actual decision of sickness absence and entitlement to benefits is made by social insurance officers based on evaluations by physicians. Generally, the evaluation and judgment of the physician is followed by the social insurance officer. Nevertheless, part-time sick leave may be seen as a complex “treatment”, which requires an initial joint decision made by the individual, the employer, the physician, and the social insurance administrator [13], and actions and decisions (of the employee, colleagues and employer) to adjust both work time and work demands (during the treatment period). The rules for benefits the time data was drawn for this study stipulated that if entitled to sickness benefit the social insurance system replaced income from day 15 and forwards at 80% of the income up to an annual salary of 297,750 SEK ~ €27,620 (for the first day of sickness absence there is no income replacement, and for day 1–14 the employer is responsible for income replacement). The income replacement is for a large group of individuals on the labor market actually 90% due to agreements between unions and employer organizations (the income replacement for longer cases of sickness absence is nowadays lower).

More and more governments are also promoting part-time sick leave, expecting possibly to reduce the costs of the social insurance system by considering part-time sick leave as a “treatment” for individuals with certain conditions, such as MSD. For example, all the Scandinavian countries promote the use of part-time sick leave in various forms. In Sweden it has been possible to be on part-time sick leave of 50% since the beginning of the 1960s (extended to also include 25 and 75% in July 1990), although this policy did not receive much attention until the end of the 1990s when expenditure on sickness cash benefits increased dramatically as described above. In Sweden as well as in Norway and Denmark it is possible to start a part-time sick leave without any preceding history of full-time sick leave [14]. In Finland, employees can combine sick leave and working 40–60% of the daily working time only after the full-time sick leave has lasted for almost 3 months.

Despite the interest in and use of part-time sick leave as a treatment for individuals on sick leave, there are hardly any studies evaluating the impact on recovery. A Norwegian cluster-randomized study on “active sick leave”, which implies returning to an adjusted work environment with the assistance of social security, showed no beneficial effects [15]. A recent Swedish study used observational data to analyze part-time sick leave as a treatment method in Sweden for individuals on sick-leave [13]. They used a discrete choice one-factor model and instrumental variables to control for the non-random assignment of individuals into part-time and full-time sick leave. Their results indicate that part-time sick leave is associated with an increase in the likelihood of recovery to full work capacity for sickness spells longer than 150 days. We use a somewhat similar approach, but we focus on individuals on sick leave with MSD by analyzing part-time sick leave (rather than full-time) as an intervention to affect the outcome (full recovery of the lost work capacity) for employees who were on sick leave due to MSD in Sweden at the beginning of the 2000's.

Materials and Methods

Data

We analyze a subsample of 1,170 employed people who were on (part-time or full-time) sick leave due to MSDs. They were selected from the 2002 sample of the RFV-LS database of the National Agency of Social Insurance in Sweden, which contains data on 5,000 individuals and is representative for all the residents registered with the social insurance office in Sweden.Footnote 2 Hence, that we have e.g. 40% males in the sample reflects the lower share of males on sick-leave in the population. All individuals in the analyzed sample, were 20–64 years old, were employed and started a sickness spell due to MSDs between 1 and 16 February 2001. We excluded all employees who ended their sick leave because of incarceration, emigration, or participation in a rehabilitation program. The joint data has been approved by an appropriate ethics committee (in 2003) and have therefore been performed in accordance with the ethical standards laid down in the 1964 Declaration of Helsinki. The data has been unidentified and does not contain personal ID numbers.

Table 1 contains means and standard deviations for the variables used in the estimations.

Table 1 Descriptive statistics
Full size table

The outcome variable (Full recovery) is a dummy variable taking the value 1 if the individual is back at work on a certain day in the future with full recovery of the lost capacity. We analyze whether or not full recovery is reached within different time periods after the spell started (30, 90, 150, 210, 270, and 330 days), which are calibrated with general guidelines used for sick listing. Table 2 in shows summary statistics of this variable for full-time and part-time sick leave. These two categories are defined by the degree of sick leave at day 15, the first day when the sickness insurance covers the employees’ sick leave (after 14 days covered by the employer). The part time dummy variable takes the value 1 for all employees who started their period covered by the sickness insurance with 25, 50, or 75% sick leave, and it takes the value 0 for all employees that started with a 100% sick leave. Only 12% of the employees who were on sick leave due to MSDs, started their period covered by sickness insurance on part-time (Table 1), and their recovery is much slower than their “peers” who started with full time as seen in Table 2 below.

Table 2 Cumulative share of employees who finished with full recovery (in percent), by type of sick leave (new full recoveries for each segment in brackets)
Full size table

The Empirical Strategy

The process of sick listing is complex, and its output in the form of the duration and the degree of sick leave, cannot be considered exogenous factors for the recovery process. This means that there are unobserved or omitted variables in the equation of interest that are correlated with the part-time dummy (treatment) variable that affects the outcome (Full recovery). If there is endogeneity, this implies inconsistency and biased estimates in finite sample sizes. Therefore, we consider that the response binary variable (Full recovery) is simultaneously determined with a dichotomous regressor (the degree, or the type of sick leave: part-time or full-time) [16, 17]. In this setting, it is important to analyze the relative performance of alternative exogeneity tests since their finite sample properties are unknown. In this way, we can analyze whether or not an association between part-time sick leave and full recovery is due to a causal effect or to selection. For example, if it is found that full-time sick-leave is associated with quicker recovery (as the raw summary statistics in Table 2 suggest), this may be due to the beneficial effect of being on full-time sick leave (causal) or to a selection effect such that individuals with a higher likelihood of recovery are assigned to full-time sick-leave. For example, given the general guidelines, an individual with certain types of musculoskeletal-related acute neck pain may be assigned to full-time sick leave regardless of his/her job task, but has a higher probability of full recovery regardless of the degree of sick leave for some job tasks. Whereas an individual with mild chronic low back pain may be assigned to part-time sick leave but have bleaker chances of recovering to full work capacity in a reasonable time horizon. If it is a selection effect that drives the association, a policy prescription of assigning more individuals to part-time or full-time sick-leave will not have any beneficial effect on recovery. On the other hand, if it is a causal relationship, a policy prescription of assigning more individuals to part-time or full-time sick-leave is likely to have beneficial effects on recovery times.Footnote 3

This paper uses a two-stage recursive bivariate probit model to try to address the problem of an endogenous regressor (e.g. [18]). The recursive structure builds on a first reduced form equation for the potentially endogenous dummy and a second structural form equation determining the outcome of interest:

$$ \begin{gathered} y_{1i}^{*} = \beta^{\prime }_{1} x_{1i} + u_{1i} , \hfill \\ y_{2i}^{*} = \beta^{\prime }_{2} x_{2i} + u_{2i} = \delta_{1} y_{1i} + \delta^{\prime }_{2} z_{2i} + u_{2i} , \hfill \\ \end{gathered} $$
(1)

where \( y^{*}_{ 1i} \) and \( y^{*}_{ 2i} \) are latent variables, y 1i (the part-time indicator) and y 2i (the full recovery indicator) are dichotomous variables observed according to the rule:

$$ \left\{ \begin{gathered} y_{ji} = 1\quad {\text{if}}\,y_{ji}^{*} > 0, \hfill \\ y_{ji} = 0\quad {\text{if}}\,y_{ji}^{*} \le 0,\;j = 1,2. \hfill \\ \end{gathered} \right. $$
(2)

x 1i and z 2i are vectors of exogenous variables,Footnote 4 \( \beta^{\prime }_{1} ,\beta^{\prime }_{ 2} = \left( {\delta_{ 1} \delta^{\prime }_{ 2} } \right) \) and δ 2 are parameter vectors, δ 1 is a scalar parameter, and the error terms (u 1i , u 2i ) are identically distributed as bivariate normal with zero mean, unit variance and correlation coefficient ρ, independently across observations. Inference on the (K × 1) parameter vector \( \beta = \left( {\beta^{\prime }_{ 1} ,\beta^{\prime }_{ 2} ,\rho } \right)^{\prime } \) can be made by the maximum-likelihood method.

The main difficulty in the statistical approach is finding appropriate instruments in Eq. 1. A good instrument needs to have a causal effect on the behavioral variable, i.e., selection into part-time or full-time sick leave, but not a direct causal effect on the outcome variable, i.e. full recovery [17]. We argue that the type of employer and occupation is important for the possibility of being assigned to part-time sick leave or not. Certain jobs and occupational types have conditions that make it very difficult for employees to be on part-time sick leave. These may be in small establishments with only one or very few employees, but may also be larger offices or labs that have only one employee who can perform certain tasks, i.e., the job requires full-time attendance. In these cases, the employer can pose two alternatives to the employee, either continue to work full-time (more than the health capacity allows) or being on full-time sick leave. This implies that the type of employer and occupation can have an effect on the likelihood of being assigned to part-time sick leave when working capacity is less than 100%. Further, we (expect and) assume individuals do not self-select into different types of employers due to the possibility of (in the future) being on part-time sick leave. It seems unlikely that individuals would base their job careers on such considerations. The conclusion from these arguments is that the type of employer may have a direct causal effect on the likelihood of being assigned to part-time sick leave, but should not have a direct systematic effect on the likelihood of full-recovery from the sick-leave. The variables of occupational type (see Table 1) are the closest to the employer type in our data, and therefore we use them as instruments in our model. It would have been beneficial to have more direct data on employer type (such as number of employees in the firm), but this is unfortunately not linked to the register data accessed for this study.

The potential problem with our instruments is that the type of employer may have a causal effect on the type of musculoskeletal injury the individual is diagnosed with, which may be related to the likelihood of recovery. However, the analyzed sample shows low level of correlation between the instruments and the analyzed outcome (i.e., full recovery). This might be the case for the entire population, given that in Sweden the workers and their working conditions and environment are protected by The Work Environment Act (i.e., Arbetsmiljölag 1977:1160). It is always the employer who is responsible for the operation being conducted in such a way that ill-health and accidents are prevented, and a satisfactory work environment achieved. The task of the Work Environment Authority is to verify that the employer lives up to the stipulations made in the Work Environment Act and in the Provisions issued by the Authority itself. This verification is usually based on inspection.

Further, the evaluation of interest is to use the estimates from Eqs. 1 and 2 above to say something about the average treatment effect (ATE) and the treatment on the treated effect (TT).We have that:

$$ \begin{gathered} {\text{ATE}} = \Pr (y_{1} = 1|x) - \Pr (y_{0} = 1|x), \hfill \\ {\text{TT}} = \Pr (y_{1} = 1|d = 1,x) - \Pr (y_{0} = 1|d = 1,x), \hfill \\ \end{gathered} $$
(3)

where y is the outcome variable (1 for full recovery and 0 for not full recovery), x are covariates and d is the treatment (here part-time sick leave). The ATE is computed by utilizing the partial effects of all individual observations and taking the sample means. The ATE tells us the average difference between the probability that the individual will fully recover after part-time sick leave and the probability that the individual will fully recover after full-time sick leave. The TT is just the average effect of treatment only on those who have been treated [18, p. 48].

Results

Selection into Treatment

Table 3 shows the estimated coefficients of the probit model of the selection into treatment, which offers a straightforward way to examine the presence of non-random selection into treatment.

Table 3 Coefficient estimates for selection equation (dependent variable: part-time sick leave)
Full size table

Several of the estimated coefficients are statistically different from zero, which indicates that individuals under treatment differ significantly from non-participants with respect to observable characteristics.

The oldest age-group (56–65) is more likely to be on part-time sick leave, which is also true for the youngest age group for longer spells. Otherwise there are no statistically significant differences between the age groups. Females are more likely and married people less likely to be on part-time sick leave. People born in Sweden are also more likely to be on part-time sick leave. Individuals who are sick-listed by a company physician (compared to a public one) are more likely to be on part-time sick leave, while individuals who are sick listed by private and specialist physicians (compared to public ones) are less likely to be on part-time sick leave.

The estimated parameters for the occupational type (our instruments) are statistically different from zero for four out of six occupational groups: Legislators/senior officials (positive), Craft and related trades (positive), Plant/machine operators (negative), Elementary occupations (positive). Thus, we passed the first test of having a valid instrument only for these occupations: The instrument should be correlated with the treatment decision, and not affect the outcome directly, but only indirectly through the treatment variable. In order to ensure the validity of the instruments, we also tested the collective significance of all the instrumental variables in the first-stage regression, with the likelihood of being on part-time sick leave as the outcome variable. We rejected (P = 0.017) the fact that the instruments jointly do not have any explanatory power regarding the likelihood of being assigned to part-time sick leave [20]. Further, we performed an informal test for the exclusion of the instruments in a probit model for full recovery (at 30–330 days) based on including the instruments together with the part-time dummy [21]. In this model the instruments should be jointly insignificant, which we cannot reject, with P = 0.27–0.54. Taken together, this gives us some confidence about the validity of our instruments.

Outcome Equation

The outcome equation is the second step in the bivariate recursive probit regression and shows the impact on full recovery within a specific time period. As explained in section “The Empirical Strategy”, the part-time variable is instrumented by the occupational type variables. Table 4 below shows the results.

Table 4 Coefficient estimates for outcome equation (dependent variable: full recovery within the time-period)
Full size table

The impact of the part-time variable is positive, large in magnitude and highly statistically significant. This implies that being assigned to part-time sick leave seems to increase the likelihood of full recovery. The coefficient is relatively similar across the different lengths of time analyzed, from 1.50 for spells lasting equal to or less than 30 days to 1.20 for spells lasting equal to or less than 330 days. The results, using our specification and instrumental variables, go against the raw data (Table 2) that showed that individuals on part-time sick leave have a lower likelihood of recovery within each time period.

Other results in Table 4 show that males are more likely to recover from sick leave and the age-pattern shows an expected one, i.e., older individuals are less likely to recover (compared to the youngest age-group). The lowest likelihood of full recovery is found among the oldest individuals (age-group 56–65), while being married is positively associated with full recovery. Having been on sick leave the previous year (Previous sick) is negatively associated with full recovery. Also, being sick-listed by a company physician (compared to public) is associated with a lower likelihood of full recovery.

Average Treatment Effects and Treatment on the Treated

As stated in section “The Empirical Strategy”, the evaluation measures that we calculate are ATE as well as TT. The calculations of the treatment effects are shown in Table 5, and they are calculated separately for each time-interval (cumulative).

Table 5 Average treatment effects and treatment effects on the treated (SD in parenthesis)
Full size table

The results show strong positive average treatment effects that are also statistically significant. The ATE is highest for the shorter time period (0.52 for 30 days or less) but also substantial for the longest time period analyzed (0.25 for 330 days or less). The ATE is the average of the individual treatment effects in the relevant population and should be interpreted here to mean that, on average, individuals who are sick-listed for a musculoskeletal disorder have a 0.25 higher likelihood of full recovery if assigned to part-time sick leave rather than full-time sick leave (330 days or less). The TT results are not statistically significant, and hence we cannot reject the hypothesis that they are equal to zero.

That ATE is positive and significant, while TT is not, may be explained by the fact that the observable characteristics of people in the treatment (part-time) group are generally strongly associated with a lower likelihood of full recovery. Hence, this implies that their treatment effect is less than would be the case for a random pool of patients who are sick-listed [22].

Discussion

This paper estimates average treatment effects and treatment effects on the treated, with regard to being on part-time sick leave rather than full-time sick leave for patients with MSDs. The interest in this question stems from the research findings that activity and some connection to the labor market may be beneficial for the recovery of patients with MSDs, and from the fact that it is an advocated policy in the Swedish institutional setting to use part-time sick leave for this patient population when deemed possible. The raw data indicates that individuals assigned to part-time sick leave are less likely to fully recover compared to individuals assigned to full-time sick leave. However, this is not particularly surprising considering that individuals on part-time sick leave have observable characteristics that are associated with a generally lower likelihood of full recovery (such as being female and being older).

In our empirical approach we use an instrumental-variable approach to handle the endogeneity problem, and we instrument the sick-leave type by occupational type. The results indicate that the average treatment effect for full recovery within 330 days or less is 25 percentage points. The effect on the treated population is smaller, which we would not expect if patients could rationally self-select into the different alternatives. However, this decision is a complex procedure involving the individual, the employer, the physician, and the social insurance administrator. As stated above, we find that individuals on part-time sick leave have observable characteristics that are associated with a generally lower likelihood of full recovery, which may explain the result that TT < ATE.

Hence, the results from our model run counter to the raw data, i.e. our model results indicates that assigning individuals to part-time sick leave is associated with a higher likelihood of full recovery. One reason for this has already been given, considering that individuals that are assigned to part-time sick leave have observable characteristics that for other reasons are associated with lower likelihood of full recovery (females, elderly).

Finally, it could also be mentioned that we saw differences in sick-listing practices across physician categories; something also reported previously [23–26]. Occupational health care physicians are a category that might handle the situation better than average because of their knowledge of and contact with work-places, which might give them a better basis for their decision when evaluating the patient’s ability to work. They had longer certification periods than GPs but used partial sick-listing more frequently, which is consistent with a previous Swedish study indicating that individuals sick listed by a GPs had the (on average) shortest time to full recovery [24]. But as indicated here, the counterfactual (treating more with full-time sick leave) may have been worse for patients consulting the company physicians. It should be noted that the reason that sickness periods are longer for patients consulting the company physician is not particularly well understood, but it may be explained both by confounding variables such as type of patient, it has been shown that patients consulting company physicians may be older than patients consulting the GP for certification of sickness absence [24], as well as selection, e.g. different physicians may have different interests which may influence the type of patients that choose to consult them.

This study obviously has some limitations. Even though the instrumental variables used to handle the endogeneity problem were jointly significant in the first stage of the regression they were fairly weak, which creates a potential problem of bias that may be quite large, e.g. theoretically IV estimators may be more biased than standard OLS estimators [27, 28]. This is a particularly relevant potential limitation considering that we found quite strong effects such as the IV estimations goes in another direction of the “raw data” as discussed above.

Further, in future research it would be beneficial to have a larger sample size, considering that the number of individuals with part-time sick leave in our sample is rather limited (133 individuals). Also, to overcome the difficulties with observational data in general, using randomized controlled trials (RCT) to evaluate part-time sick leave as a treatment method would handle the endogeneity problem in a more convincing matter (but may of course create other problems). A study protocol of an RCT to evaluate part-time sick leave has been published, but to our knowledge no such study has yet published any results [14].