A population-based longitudinal cohort study was performed.
We included all the 3536 women in Sweden, aged 19-64 who were diagnosed with a first malignant neoplasm of breast (International Classification of Diseases 10th version (ICD-10) [40] code: C50) in 2010. Data were obtained from five nationwide registers as follows:
- The Board of Health and Welfare’s Cancer Register (all BC cases 1958-2010, diagnosis date, type, T, N, and M classifications[41]), Patient Register (main diagnosis, dates of in- and specialized outpatient care 2008-2010), and Cause of Death Register (dates 2010-2013);
- Statistics Sweden’s Longitudinal Integration Database for Health Insurance and Labor Market Studies (LISA) (age, educational level, marital status, family composition, birth country, occupational sector, geographical and type of living area in December 2009, emigration 2010-2012, not living in Sweden 2008 or 2009);
- National Social Insurance Agency’s Micro-data for Analyses of Social Insurance (MiDAS) (SA and DP benefits 2008-2013: dates, full- or part-time, main diagnosis).
Data were linked at individual level using the ten-digit personal identity numbers assigned to all residents in Sweden.
SA and DP benefits in Sweden
All people in Sweden ≥16 years, with an income from work or unemployment benefit, with reduced work capacity due to disease or injury can be granted SA benefit from the Social Insurance Agency.[42] The employers usually provide reimbursement for the first 14 days of a SA spell, which is why we do not have information on most SA spells ≤14 days. From day 8, a medical certificate issued by the treating physician is required. All residents aged 19-64 can be granted with DP if having long-term or permanent work incapacity due to disease or injury. SA and DP can be granted for full-time (100%) or part-time (25, 50, or 75%) of ordinary working hours. SA benefits cover 80% and DP 64% of lost income, up to a certain level.
Measures
We investigated two types of outcomes; DP and SA (for spells >14 days). SA and DP days were transformed into net days; e.g., 2 days on half-time SA or DP was counted as one net day. SA and DP diagnoses were coded by the certifying physician who assessed the patient’s condition and work capacity. Diagnoses were for some analyses classified into four categories: 1): BC (ICD10: C50), BC-related diagnoses (Z80, Z85, N61-N63), and other cancer (C00-D48), 2): mental diagnoses (F00-F99, Z73), 3): other diagnoses (all remaining ICD codes), and 4): missing information. The outcome in the predictive model was defined as starting a new SA spell >14 days due to one of the following SA diagnoses (C00-D48, Z80, Z85, N61-N63, F00-F99, or Z73) during the time-window of 14 days before to 29 days after the BC diagnosis. This time window was based on the frequencies of start of new SA spells in the full cohort, in relation to diagnosis date (T0). For some women there was a delay before the diagnosis was included in the Cancer Register (even if the women were informed) and for others, treatment did not start until weeks later. The reason for including “diagnoses related to BC” and “other cancer diagnoses” in the predictive model was that sometimes a broader category of cancer diagnoses is given in the medical certificate.[43] Mental diagnoses were also included in the predictive model as that a cancer diagnosis might lead to anxiety or depression.[44, 45]
The included sociodemographic, disease-related, and comorbidity covariates (listed in Table 1) were selected for the predictive model based on previous findings regarding factors influencing SA and RTW.[13, 14, 17, 20, 23, 25, 26, 29] Missing information on educational level was coded as elementary school. Cancer-stage groups were assigned using the TNM Classification of Malignant Tumours[41] and categorized as: T0N0M0+stage 0+I, stage II, stage III+IV, and missing all TNM (with no T, N, or M information), respectively. When T, N, or M information was missing in one or two of the categories or classified as ‘X’ (assessment not possible), the value was set to 0. If more than one tumour was registered, with different diagnosis dates in 2010, the most advanced tumour was selected. The main ICD-10 diagnoses for healthcare were coded by the treating physicians. Healthcare due to uncomplicated delivery (O80) or not related to morbidity (e.g., screening) was excluded.
Statistical analyses
The mean number of SA and of DP net days/year, respectively, were calculated for all women, using the BC diagnosis date (T0) as reference, for the two years before T0 and three years after T0 (Y-2 to Y+3). This was done for all SA and DP as well as for the four SA/DP diagnostic categories mentioned above. The annual numbers and proportions of women with SA/DP due to the different diagnoses were also calculated. The denominator used in these calculations varied somewhat over the years due to the exclusion of women (turning 65 years, emigration, or death).
In the predictive model regarding risk of new SA related to time of diagnosis, 2954 women were included. For those analyses we excluded the 521 women (14.7%) already on SA or on DP for full-time or nearly full-time (75-100%) at T0. Additionally, 61 women were excluded due to lack of covariate information, or because of extreme values on some of the continuous variables, e.g., number of healthcare visits or inpatient days.
The risk of a new SA spell due to BC or related diagnoses, other cancer diagnoses, or mental diagnoses was modelled using multivariable logistic regression[46, 47] with a logistic model formulated as follows: log[p(yi=1)/p(yi=0)]=xi'β where yi denotes the SA status of individual i, and xi is a vector of observed covariates. Natural cubic splines[48-50] were used to model potentially nonlinear effects of continuous covariates. The five variables that were modelled using splines were: age and number of previous: SA days, DP days, outpatient healthcare visits, and inpatient days, respectively, in the two pre-diagnostic years. An optimal threshold c was selected, such that predicting SA=1 whenever the fitted probability was above c, minimized the sum of false positive (FP) and false negative (FN) and maximized the proportion of correctly classified observations. Also, the receiver operating characteristic (ROC) was calculated.