Setting, design and study population
The National Social Insurance covers all Swedish permanent residents between 16–65 years of age, whether citizens or not [17]. The insurance covers access to primary or hospital care at heavily subsidised rates, the right to see any physician of one’s own choosing, to have sickness benefits for income loss in case of reduced work capacity due to injury or disease and many other benefits. At the time of the study, there was no limit of the time a patient could be sick-listed.
The study protocol used in this study has been described previously [1]. Briefly, the study was designed as a three-year prospective, cohort study and was performed at one of the primary health care centres in Eskilstuna, Sweden, with ten general practitioners serving a population of approximately 25,000 residents. A total of 943 patients (482 women and 461 men), who were 18 to 63 years of age, sickness-certified by a general practitioner at the centre at any time from 1 January until 31 August 2004 and who gave their informed consent of participation were included in the study. Patients already included in a medical or vocational rehabilitation programme were excluded.
Baseline data
Baseline data obtained from the sickness certificate included the age, sex, occupational status (in gainful work or not), sick leave diagnosis according to the WHO International Classification of Diseases (ICD-10) [18] and degree of sick leave (25%, 50%, 75% or 100%). Information regarding the sickness absence track record during the 365 days preceding the baseline examination was obtained from the National Social Insurance Agency database. The data included sick leave diagnoses, first and last day of each sick spell, information on marital status, salary, whether born in Sweden, and for immigrants, Swedish citizenship status.
At the time of the study, patients could self-certify the first seven days of a sick leave. If the sick leave protracted beyond this point, a physician’s sickness certificate was needed. For this reason, the sick leave information was verified with the primary healthcare medical records and completed with self-certified days.
A manual classification of the chances of concluding the ongoing sick leave period on expected time was made based on any of the following variables: a sickness certification track record during the last year of more than 28 days, being sickness-certified at baseline because of musculoskeletal disease (ICD code M) or a psychiatric disease (F), being unemployed, being older than 45 years and being a woman. Of the 943 patients, 496 were classified as low risk of not concluding their sick leave on time, 277 as having a moderately high risk and 170 as having a high risk.
Intervention
The 170 patients in the high-risk group then passed an examination and assessment of medical needs and workability by a multidisciplinary medical team, including a physician, a physiotherapist, an occupational therapist and a social worker (Figure 1).
The individual rehabilitation programmes, as agreed with the patients, were then implemented by the multidisciplinary medical team and the stakeholders from the local Social Insurance Agency, the local Employment Agency and a social worker from the city of Eskilstuna. A sick leave conclusion programme typically included physiotherapy, physical exercise, stress and pain management, meetings with the workplace management, work modifications, gradual return to work or workability training. The 773 subjects with moderately high or low risk received standard treatment as recommended by their general practitioners.
Follow-up data
Information on sickness absence during the 3-year follow-up from baseline was obtained from the National Social Insurance Agency database, including sick leave diagnoses and first and last day of each sick spell, and whether a disability pension was granted during follow-up. Information on vital status and date of death for those who died (n = 6) was obtained from the National Cause of Death register. The Regional Ethics Review Board in Stockholm, Sweden approved the study, Act number 2008/980-31/3.
Statistical considerations
Data was analysed with the Statistical Analysis System (SAS) software, version 9.3. No data were missing. The outcome of the study was sick leave conclusion and the day after baseline when it occurred. In order to decide on that day the method proposed by Bogefeldt et al. was used [19]. The three-year sick-leave follow-up period was converted into a day-by-day matrix starting with variable ‘day 1’ (baseline day) and ending with variable day ‘1096’ (end of follow up). Each variable measured whether the subject was on sick leave (= 1) or not (= 0) on that day.
Based on this matrix a sick leave conclusion variable for the sick leave period in effect at baseline was computed. Two criteria were applied for each sick spell: (a) the sick spell was followed by a sick-leave-free interval of more than 28 days, regardless of the length of any following sick spell; (b) the sick spell was followed by a sick-leave-free interval of more than 7 days, and that interval had to be longer than the next sick spell. When at least one of the criteria was fulfilled, sick leave conclusion was presumed to have occurred on the first non-sick-leave day. If none of the criteria were satisfied at the end of follow up, sick leave conclusion presumably had not occurred.
The study population was not randomised into an intervention and a control group, initially, since the study was not intended to be a scientific one. A post hoc control group, as similar to the intervention group as possible, was created by means of a propensity score. Propensity score, first proposed by Rosenbaum and Rubin in 1983 [20], implies that matching may be performed based on an unlimited number of matching variables that are weighed together into a propensity score.
A prerequisite for matching in this case was that the manual classification of individuals into a high-risk group versus a moderate-to-low-risk group was imperfect, as it usually is when no explicit variable weights are used. When logistic regression was used to compute predicted risk based on the same variables that were used manually, but now graded (given weights), a substantial overlap of risk score was found, primarily between the high-risk and the moderate-risk groups.
This circumstance then allowed computation of a propensity score with nominal logistic regression using the rehabilitation (high-risk) group (code 1) and all others (code 0) as dependent variable, and age, sex, number of sick leave days last year, sick leave diagnosis, degree of sick leave, whether born in Sweden, whether a Swedish citizen and marital status as independent variables. Based on an analysis of the impact of the various sick leave diagnoses on sick leave conclusion previously published [1], the latter were ranked from –2 (largest impact) to +3 (least impact). In this way, all variables entered into the logistic regression assumed to carry a risk for not concluding sick leave at the expected time, were collected into one measure, the propensity score.
Subjects in the non-rehabilitation group were then matched to subjects in the rehabilitation group by propensity score to form potential control groups. Mean (standard deviation, SD) propensity score in the rehabilitation group was 0.301 (0.173), in the first matched control group 0.293 (0.162), in the second control group 0.223 (0.100), in the third control group 0.119 (0.034) and in the fourth control group 0.051 (0.034). The scores of the first and second control groups were thus fairly similar to the rehabilitation group and were combined into a common control group. The rehabilitation group (n = 170) and the control group (n = 340) constituted the study population of this report.
Simple differences between the rehabilitation and the control group were tested with Student’s t-test for continuous variables and the chi-squared test for discrete variables. According to the SAS ‘life test’ procedure, there were no close proportional hazards regarding sick leave conclusion across the total follow-up time. The latter was therefore divided into days 1–14 (when rehabilitation activities had not started), days 15–112 (when most rehabilitation activities were performed), days 113–365 days (when most rehabilitation activities were finished), and days 366–1096 (long-term follow up). For each of these partial follow-up periods, the hazard rates were approximately proportional.
The effect of the vocational rehabilitation programme was evaluated with conditional proportional hazards regression, one analysis for each partial follow-up period, where conclusion of sick leave and the time when it occurred were entered as dependent variables, and group allocation was entered as the independent variable, as well as individual propensity scores to further adjust for the potential remaining risk differences between the groups. To check the results for dependence on remaining propensity score differences, the analyses were repeated using only the first matched group as the control group. The results were the same as shown below, except that measures of dispersion were somewhat wider.
The analysis provided hazards ratios (rehabilitation group versus control group) and 95% confidence limits, Wald’s chi-squared (a measure of exposure impact on outcome) and p-values. All tests were two-tailed, and the significance level was set at p < 0.05.