Healthcare and productivity cost of osteoporosis: a Danish register-based quasi-experimental study

Administrative registers encompassing all individuals residing in Denmark were accessed through the Statistics Denmark Research Service. The study population was identified from the National Patient Register (NPR) and Danish National Prescription Register [13, 14]. The NPR captures all contacts with the hospital system in Denmark.

The osteoporosis group consisted of individuals born in 1930–1950 with an osteoporosis diagnosis or an osteoporotic fracture occurring between 2000 and 2021. The birth cohort and follow-up period were chosen to ensure a large enough sample of patients with osteoporosis was captured and was observed for sufficiently long, such that we had information on patients ranging in age from 50 until 89. The sample size of patients aged 90 + was too small to render statistically significant results (Supplementary Fig. 2). Osteoporosis diagnosis was defined as International Classification of Diseases, 10th edition (ICD-10) codes M80–M82 including primary and secondary diagnoses after the age of 50. An osteoporotic fracture was defined as hip fractures (ICD-10: S72, S72.0, and S72.1), vertebral fractures (ICD-10: M48.4, M48.5, M49.5, S22.0, S22.1, S32.0, S32.7, and S32.8), upper arm fracture (ICD-10: S42.2 and S42.3) and wrist fractures (ICD-10: S52.5, S52.6, S52, S52.5B, and S52.9) [10, 15, 16] including primary and secondary diagnoses. Further, individuals prescribed pharmaceuticals bisphosphonate (ATC-codes: M05BA01, M05BA04, M05BA06, M05BA07, M05BB01, M05BB03, and M05BA08 with more than nine months between prescriptions), strontium ranelate (ATC-code: M05BX03), denosumab (ATC-code: M05BX04) with more than five months between prescriptions, romosozumab (ATC-code: M05BX06), selective estrogen receptor modulator (SERM) (ATC-code: G03XC01), and parathyroid hormone analogue (ATC-code: H05AA02) [16, 17] were also included in the osteoporosis population. Hospital-based medical treatment with the following regiments also qualified individuals for inclusion in the osteoporosis group: with zoledronic acid (intravenous [IV] infusion) > 9 months between injections, or denosumab (subcutaneous injection) (> 5 months between injections), or ibandronate (IV infusion) (> 3 months between injections) [18]. No explicit exclusion criteria were applied in this study. This approach was chosen because trauma mechanisms are not coded in Danish registers, and cancer patients’ fractures are predominantly due to co-existing osteoporosis rather than metastases.

The control group was selected from all individuals born in 1930–1950 not meeting any of the aforementioned criteria and alive at their index date.

Medical care costs were based on the data of the study population’s utilisation of healthcare services for the period of 01.01.2000–31.12.2021 and were identified from several Danish national registers. Information on primary healthcare costs were obtained from the National Health Insurance Service Register (NHSR) [19]. The National Patient Register was used to identify all inpatient admissions, outpatient visits and diagnosis codes in accordance with the ICD-10 codes, and costs associated with delivery of these services paid for in Diagnostic Related Groups (DRGs). Data on the cost of community-filled prescription medicines was retrieved from the Danish National Prescription Register.

Moreover, the National Population Register, which holds detailed demographic information on all individuals residing in Denmark, was used to identify the demographic characteristics for study population, including marital status, municipality of residence, and region. Furthermore, data on emigration, death, income, educational level were also linked to the study population.

Productivity gains were estimated for the period from the beginning of observation until the individuals reach the age of 67, a conservative estimate for retirement in Denmark [20]. The estimation was done by applying the average wage of €48 per hour [21] to those who were recorded as working in each quarter during the aforementioned period. The average rate was selected to avoid introducing differences in labour market participation rate, payments, and retirement age, in order to maintain generalisability of our findings. The data on employment status was obtained from the social transfer register [22].

Propensity score matching was used to create a matched control group. The propensity score (the probability of having osteoporosis) was estimated using multiple logistic regression. The model consisted of confounders associated with an increased risk of having osteoporosis which were available in the register data. These included: age, gender, income, educational level, marital status, and Charlson comorbidity index (CCI) [23].

One-to-one nearest neighbour matching without replacement was employed [24]. Reduction of bias in the matched sample was tested with chi-squared and t-tests, as appropriate, as well as by comparing reduction in standardised differences [25]. All unmatched cases, both controls and osteoporosis patients, were excluded from further analyses. The index date of osteoporosis patients was the date of first diagnosis or relevant hospital admission, whichever occurred first. The index date was transferred to their matched controls such that the entire analysis population had an index date.

A quasi-experimental difference-in-differences (DID) approach was applied to estimate healthcare and productivity costs attributable to osteoporosis and osteoporosis fractures [26]. The total costs were compared before and after their index date. Healthcare costs included all medical (hospital care, primary care, and community-dispensed medication) and health-related social costs (home help, home nursing, rehabilitation and nursing homes). Productivity costs were defined as loss of income and only occurred until age 67 and were therefore analysed separately.

A DID analysis is a suitable choice for the analysis of observational data, where the levels of the compared groups can differ also prior to the observed event. If the treatment and control groups display parallel trends before the event, the development of the control group serves as a counterfactual to the development of the treatment group after the event, with the event being the index day. If the trend of the treatment group deviates from the counterfactual, it is assumed to be caused by the event (here occurrence of osteoporosis).

Generalised estimating equations [27, 28] with individual-level fixed effects were applied to allow for more flexibility in the common trend assumption [26]. The choice of the DID approach enables us to both control for unobserved covariates, as each individual serves as their own control, and to circumvent the issue of right-skewed cost data which otherwise requires parametric analytical approaches or log-transformation of data and subsequent interpretation complications [29]. The assumption of parallel trends in outcomes during the pre-treatment period was tested using graphical examination and placebo tests [30]. Three DID models were fitted: unadjusted (Model 1); adjusted for sex, comorbidity, and age (Model 2); and adjusted for sex, comorbidity, age, marital status, educational level, and income (Model 3) to control for variation between the two groups unadjusted by the matching.

All costs were readjusted into 2022 Euros (€) using the net price index [21] and an exchange rate of 7.45 DKK = 1 Euro (€). All statistical analyses were performed on Stata Version 18.0 [31].

The matching process successfully matched 323,752 of 667,290 individuals from the osteoporosis group with 323,752 controls (Table 1). The matching process balanced the differences in age, sex, and income between the two groups, while statistically significant but numerically small differences in marital status, education and comorbidity remain (Table 1). After matching, both groups were made up of 73.2% women, and had a mean annual income of about €36,000. Fifty-eight percent of the osteoporosis group were married or in a partnership, compared to 56.7% controls. The osteoporosis group had a mean CCI score of 1.299, compared to 1.276 for the control group.

The majority of the osteoporosis group was identified through medication prescription pattern (n = 101,189) and a wrist fracture (n = 73,463). The mean age at diagnosis was 69.4 although it varied by how the diagnosis was identified, with patients identified through denosumab and zoledronic acid prescriptions being older: 75 and 74 years, respectively.

Annual productivity gains were lower for the osteoporosis group compared to the matched controls from the beginning of the observation period until reaching the age of 59, after which the costs for the two groups converged and dropped significantly (Fig. 3). Then, 175,566 individuals from the control and 171,354 from the osteoporosis group were recorded as working during the productivity observation period. After age 65, most people are retired and there is no detectable difference between the groups.

Bild vergrößern

The mean annual productivity gains were 18% lower for the osteoporosis group than for the reference group after controlling for age, gender, income, educational level, and marital status (OR 0.82, 95%CI 0.81–0.82) (Table 4). This corresponds to an average of €3,883 less earnings, annually, until the age of 66 for the osteoporosis group (Table 4). The gap in productivity gains between the two group narrows when approaching retirement, decreasing from around €22,500 at age 50 to around €,2000 at age 65 (Fig. 4).

Bild vergrößern

The findings of this study are subject to some limitations. The identification algorithm has been composed based on previous studies and with inputs from clinicians. As an algorithm, it may have suboptimal sensitivity and specificity. We have included fractures that are predominantly osteoporotic when occurring in this age group, although some may indeed be caused by severe accidents. We have included pharmaceuticals usually dispensed to osteoporotic patients but excluded drugs dispensed to both cancer patients and osteoporosis patients when the administration frequency suggested that the drug was used for the treatment of cancer, for fear of misclassifying oncological patients as osteoporosis patients.

The propensity score was based on administrative observational data, and lacked some clinical predictors of osteoporosis risk, such as family history, calcium intake, and medication affecting bone metabolism, though our difference-in-differences design helps address time-invariant unobserved factors. Propensity score matching did not completely balance the differences in observed covariates between groups, in terms of education, comorbidity, and marital status. However, the absolute differences were small, and we believe that individual fixed-effects models used in the analysis sufficiently control for these differences. As this study was register-based, it was not possible to include out-of-pocket healthcare expenses associated with osteoporosis. However, as Denmark has a universal healthcare system free at the point of contact for users, with the exception of small co-payments for prescription medication which has a maximum ceiling of €610 per year [36], we believe that the costs captured in the presented analysis should capture the vast majority of healthcare costs incurred by all individuals. For the same reason, it was not possible to capture informal care costs; future studies should employ prospective data collection to capture the economic burden on informal care in osteoporosis.

It is important to acknowledge that not all included fractures were necessarily incident, as it was possible for older persons in our cohort to have experienced a fracture prior to inclusion period of 2000 to 2021. This may mean that costs of osteoporosis fractures are underestimated, making our results conservative. Moreover, the presence of the younger age groups in our cohort does represent an incident population, and propensity score matching should ensure better comparability between groups despite this limitation.

Our study only included vertebral fractures which were recorded through hospital contact. Many vertebral fractures are either asymptomatic or managed in primary care settings without hospital referral. These unreported fractures can still result in healthcare utilisation and productivity losses through chronic pain management, physiotherapy, medication use, and reduced work capacity. Therefore, our cost estimates for vertebral fractures may represent a conservative reflection of the true societal burden.

The study is also characterised by a number of strengths. The robustness and reliability of administrative register data has been documented and reported as highly suitable for health services research [11, 13, 37]. The study also utilises a unique dataset of an entire population of people with osteoporosis, observed over a 21-year period, providing a unique opportunity to draw conclusions about healthcare costs associated with the disease. Finally, we believe that the double robust methodology of matching and difference-in-difference analysis permits us to make claims of causal attribution of costs to osteoporosis.