Database
We investigated a 1/97th random sample of the National Inter-Scheme Information System on Health Insurance (Système national d'information inter-régimes de l'Assurance maladie, SNIIRAM) covering 98% of the French population, called the Echantillon Généraliste de Bénéficiaires (EGB) [20]. The EGB is a medico-administrative database which constitutes a representative sample of the French population in terms of age and gender. It comprises more than 660,000 individuals, whether they receive healthcare or not [20]. The EGB is an open cohort that is continuously updated with new beneficiaries and newborn infants. The EGB contains exhaustive information on all outpatient care performed (volumes) and reimbursed (values) by the national health insurance. It also contains information on patient-specific administrative data, such as date of birth, gender, place of residence, and conditions for reimbursement of care (total or partial coverage of insured persons). The EGB is linked to the private and public hospital discharge database (Programme de Médicalisation des Systèmes d’Information, PMSI). The PMSI is a medico-administrative hospital discharge database set up to evaluate the costs of hospital stays according to different Diagnosis Related Groups (DRGs). The PMSI provides exhaustive information on public and private hospital care in France, such as diagnoses (coded by physicians, using the International Classification of Diseases, 10th version (ICD-10)), underlying comorbidities, dates and lengths of stay. The PMSI also contains information on outpatient visits and technical medical acts performed in hospitals, available since 2013.
Study population
All women who purchased a pharmaceutical treatment used in infertility (see Additional file 1) in 2014 and did not receive any of these treatments in the previous three years (rolling year between 2011 and 2014) were included in the study, at the date of their first purchase. The date of inclusion in the study was the date of purchase of the first infertility treatment in 2014. The study population was limited to women aged 18 to 50 in 2014, who were living in mainland France and were affiliated to the general scheme, and who did not have a long-term disease (defined as a disease in which the severity and/or the chronicity require a long-term and particularly costly treatment), including cancer, during the 3-year follow-up period after the beginning of inclusion (rolling year between 2014 and 2017). When a woman gave birth, her follow-up was censored at the time of early pregnancy measured by the estimated date of the first day of pregnancy.
Study design
A self-controlled before-after analytic design was used to evaluate the economic burden of infertility. The difference of expenditure in overall healthcare resource utilisation was calculated for each patient, by semester. The healthcare resources considered were hospitalisation (private and public), pharmacy, consultations, technical acts, biology, others (including nursing care, midwifery, physiotherapy, dental care, transportation, medical devices and services, and cash benefits). They were considered during the 2 semesters preceding the date of inclusion (pre-treatment) and the following 6 semesters (post-treatment) (rolling year between 2013 and 2017). To control for any changes over time in healthcare resource utilisation independent of infertility, we selected a group of matched controls and conducted the same expenditure assessment 2 semesters before and 6 semesters after a matched index date. Infertility-associated expenditure was derived as a difference-in-difference (DiD), in which the difference between expenditures for patients treated for infertility and non-treated controls were regarded to be associated with the infertility event.
The matching method
We used the exact matching method to select the control group. The exact matching method is a method that associates one or more controls with identical matching characteristics with the cases (the treatment group) [21]. The control group was constituted in several phases. Firstly, we selected all women aged between 18 and 50 in 2014, who did not receive any fertility treatment between 2011 and 2017, and who were living in mainland France and affiliated to the general scheme. Just like the women treated for infertility, we limited the control population to women who did not have a long-term disease or were not treated for cancer during the follow-up period. Secondly, for each case in the treatment group, controls were selected at random on each of these four variables collected in 2013: age (in six categories), Universal Health Coverage (Couverture Maladie Universelle, CMU) as a dummy variable, the quintile of social deprivation index (Indice de désavantage social, FDep13), the quintile of the Local Potential Accessibility to gynaecologists (Accessibilité potentielle localisée, APL). An individual could benefit from CMU in 2013 if he/she was legally resident in France for more than 3 months and if he/she had resources below a ceiling based on the composition of the household. CMU exempts patients from advance payment of expenses. It is used here as a proxy for individual socio-economic status. FDep13 is an ecological measure that characterises the socio-economic commune in which individuals live [22]. APL measures the spatial adequacy between the supply and the demand for healthcare at the city level [23]. This indicator takes into account access to practitioners based on distance, practitioners’ volume of activity, and service use rates differentiated by population age structure. APL is expressed in terms of full-time equivalent per 100,000 inhabitants.
Controls matched to a case—a woman who gave birth during follow-up—were censored at the time of the case’s early pregnancy. Controls matched who themselves gave birth during the follow-up were not censored.
Economic analysis
We carried out a DiD regression estimation [24,25,26] under the assumption that the differences between the groups (cases and controls) would have remained constant without treatment. DiD was implemented as an interaction term between time and treatment group dummy variables in a linear regression model, as follows (see Additional file 2):
$$y_{ij}=\alpha+\beta{\text{I}}_i+{\textstyle\sum_{j\neq-1}}\lambda_j{\text{S}}_{ji}+{\textstyle\sum_{\begin{array}{c}j\neq-1\end{array}}}\delta_j{\text{S}}_{ji}\times{\text{I}}_i+\varepsilon_{ij}$$
(1)
where yij is the total expenditure of ith woman at the jth semester;
Ii is an indicator of treatment group, case group (Ii = 1), or control group (Ii = 0);
Sji represents the jth semester of the ith woman for j = 1, . . . , 8:
where \(t=1,\dots ,48\) is the number of months since January 2013;
We chose the -1 semester (\({\text{S}}_{-1i}\), 1 to 6) as the baseline.
\(\alpha\) is the average expenditure in the control group at semester -1;
β is the average expenditure (all expenditures combined) differential between cases and controls at semester -1;
\({\lambda }_{j}\) is the average expenditure (all expenditures combined) differential between semester j and semester -1, in the control group;
\({\delta }_{j}\) is the DiD between cases and controls between semester \(j\) and semester -1, i.e. the infertility-associated expenditures.
This model made it possible to calculate the infertility-associated costs in the semester preceding the beginning of treatment (Semester 0) and those in the following 6 semesters, and also the infertility-associated expenditures per woman and per 10,000 women over the 3.5-year follow-up period. Using the DiD method, we also examined the different expenditure items separately: hospitalisation (private and public), pharmacy, consultations, technical acts, biology, others (including nursing care, midwifery, physiotherapy, dental care, transportation, medical devices and services, and cash benefits), and their distribution over time.
By including semester 0 as part of the treatment period, we assumed that this is a period during which women were more likely than usual to seek care because of their fertility disorders (discomfort, anxiety, etiological assessment, etc.).
We used the MIXED procedure in SAS 9.4. In the analysis of response profiles, no specific time trend is assumed. Instead, the times of measurement are regarded as levels of the discrete factor. In order to take into account repeated data related to the same women, we computed the empirical (‘sandwich’) estimator of the covariance to correct for any misspecification of the covariance [27, 28]. The data were assumed to be Gaussian, and their likelihood was maximised to estimate the model parameters.
A societal perspective was adopted. The expenditures (95% confidence interval) were converted into 2020 euros, no discount rate was applied.
Ethics
Access to the EGB (pseudonymous data) is subject to prior training and authorisation. The EGB was approved by the French National Commission for Data Protection and Liberties (Commission nationale de l'informatique et des libertés, CNIL).