Based on correlation between the volume components and medical outcome minimum volume thresholds for specific medical treatments have been implemented by the Federal Joint Committee (Gemeinsamer Bundesausschuss, G-BA) in Germany. Decision making in-cludes both, the catalogue of procedures and the mode of participation in the delivery of enclosed procedures, respectively. The prospective permission to perform enlisted proce-dures relies on a prognosis about surpassing the defined threshold in the following year to be based on the case number of the previous year. According to this prospective ap-proach new thresholds unfold there effects prior to that. The contribution explains opera-tion mode of the rules enacted by the G-BA and provides first insides on the effects.

For a range of medical services provided, an association between the number of procedures carried out at a medical centre, i. e. the volume, and the outcomes of patients treated at that centre has been shown. Since 2003, the Federal Joint Committee (G-BA) sets binding minimum volume standards for certain planned inpatient services for a hospital to be permitted to deliver this service. When considering introducing a minimum volume standard, the G-BA can commission IQWiG with determining whether there is sufficient scientific evidence of any volume-outcome association with respect to this service. This presentation will give an overview of the methods used by IQWiG to assess the validity of studies investigating volume-outcome associations in healthcare.

As a first step, any relevant clinical evidence is identified in a literature search. Evidence from observational studies as well as controlled intervention studies (the intervention being the setting of a minimum volume) is considered. The studies have to meet a set of inclusion and exclusion criteria to ensure a minimum level of quality. In a next step, the selected studies are assessed for a range of quality aspects in order to rate their explanatory power. These include the quality of the data, details of statistical modelling, e.g. accounting for clustering, and the quality of reporting. Each reported result for any outcome relevant to the patient is subsequently assessed separately for its usability based on the analysis carried out. Then, the results for all relevant outcomes are extracted and compared. Due to the diversity of the studies included, results can generally only be synthesised qualitatively.

Despite the ongoing strong interest in associations between quality of care and the number of cases (volume) of healthcare providers, a unified statistical framework for analyzing them is missing. Also, many studies suffer from poor statistical modelling choices such as the discretization of volume into groups.

In this talk, we present a flexible, additive mixed model on the level of the individual patients for studying volume-outcome associations in health care. We treat volume as a continuous variable and model its effect on the considered outcome through penalised splines. We adjust for different case-mixes by including patient-specific risk factors. Furthermore, we take into account clustering on the provider level through random intercepts. Using that approach, we obtain a smooth volume effect as well as volume-independent provider effects. A comparison of these two quantities gives insight into the sources of variability of quality of care. All effects are estimated in a unified framework allowing for adequate uncertainty quantification.

Depending on the estimated association from data, our approach also enables the estimation of potential threshold values for the volume based on a break point model. For instance, that is of interest when investigating administrative requirements on the minimum provider volume. Furthermore, given a potential minimum provider volume, it is also possible to evaluate the statistical effect on the number of adverse outcomes.

We illustrate our approach through an example based on German health care data.

The association between surgical, chemotherapy and radiotherapy hospital volume and intermediate‑term survival was evaluated in a registry cohort of 18,938 patients (age 19-85) with epithelial breast cancer, diagnosed between Jan 2014 and Dec 2016, resident and treated in five out of eight Lombardy Agencies for Health Protection (AHP) including 78% of Lombardy residents. Cancer characteristics at diagnosis were retrieved from the AHP registries and information on treatment from hospitalization, outpatient and drug databases. Follow-up was investigated through census at Dec 2019. N=789 patients were excluded because of no registered treatment , and 651 because treatments had been performed outside the study area. Of the 17,498 included patients 95% were surgically treated, 38% received CT and 56% RT. We tried to determine which volume was the best proxy of hospital quality. For this purpose, both previous-year and cumulated 3-year volumes were used, considering both all breast surgeries and specific breast cancer surgeries while, as regards to chemotherapy and radiotherapy, both the total and the specific volume for breast cancer delivered from the hospital. Each patient was assigned, for each treatment, the volume of the facility where it was delivered. Only hospitals with a previous-year specific volume of at least five were analysed: 75 for surgery, 24 for radiotherapy, and 61 for chemotherapy. Previous year specific volumes I and III quartiles were (patient/year): 111-596 for surgery, 99-424 for chemotherapy and 213-641 for radiotherapy. The association between volume and death was then estimated using a Cox model with hospital as a random effect. The Hazard Ratio (HR) of death and its relative p values, the AICs of the linear models, and the AICs and graphical trends of the models with the volume as a spline (natural cubic with 3 to 5 knots) were then compared for the different treatment volumes. The AICs of the models with splines were all smaller than those with the linear variable. Both at unadjusted and adjusted analysis, the association between chemotherapy volume and outcome was not significant. For surgery and radiotherapy, the specific average volume of the previous year was chosen for the adjusted analyses, including the following potential confounders: age, stage, morphology, comorbidity index, educational qualification, grading, emergency diagnosis. Predictive mean matching imputation was performed for stage, grading and educational qualification because of missing data (<25% for all). The models with and without random intercept for the hospital were compared, using the likelihood ratio test, which was highly significant both for the linear volume model and for those with splines (p <0.00001). Therefore, the random effects model with the linear volume was compared with those with the volume inserted as a spline with 3 and 4 nodes using the AIC, calculated on the likelihood after integration of the random effect. The HR of death for every 100 unit of volume increase was 0.977 (p=0.0004) for surgery and 0.960 (p=0.04) for radiotherapy, the latter excluding stage IV patients. However, the association was not linear and, based on the AIC, the model with 3 knots was chosen for both.