Influence of physician networks on prescribing a new ingredient combination in heart failure: a longitudinal claim data-based study

In 2017, the prevalence of heart failure (HF) in Germany was 3.4% [1]. Due to increases in life expectancy and improved treatment, the prevalence of HF is expected to rise. Cardiovascular diseases, including HF, are among the most frequent causes of death worldwide [2]. Each year in Europe, cardiovascular diseases cause more than 4 million deaths—45% of all deaths [3]. Drug therapy is the most critical pillar in the treatment of HF. Since the treatment guidelines were updated in 2016, the European Society of Cardiology and the national guidelines in Germany have recommended the prescription of a combination of sacubitril and valsartan, if symptoms persist despite administration of basic therapy [4, 5]. Sacubitril/valsartan, which combines an angiotensin receptor and neprilysin inhibitor (ARNI), has been available in Germany since 2016 [6]. According to the data from statutory health insurance in Germany, 3.7 million defined daily dose (DDD) of ARNI were prescribed in the first year of approval, while in 2019, 34.1 million DDD were already documented [7, 8]. The PARADIGM-HF study demonstrated that ARNI reduces mortality due to HF more than the angiotensin-converting enzyme inhibitors enalapril [9]. Adoption of new medications, including ARNI, is often slow [10‐12]. Wachter et al. [13] showed in their longitudinal cohort study that most patients did not receive the recommended dose for ARNI. To achieve the best possible treatment for patients, it is important to optimize the adoption of effective medicines.

Many factors influence the uptake of clinical practices among physicians, such as the prescribing of medication. In addition to patient factors and the benefit–harm profile of practices, individual physicians’ beliefs and routines play a role [14]. Their beliefs and routines may be particularly relevant if the practice is new, controversial (e.g. due to high cost), or associated with high uncertainty regarding clinical effects. Individual characteristics of physicians, such as sex, professional experience, and medical specialty designation, may influence the adoption of a new drug. For instance, cardiologists respectively specialists tend to adopt new drugs earlier than primary care physicians [14, 15]. Physicians’ beliefs and routines are also shaped by contextual influences, such as collaboration and exchange with peers [14, 15]. A systematic review by Lublóy [14] summarizes how social interactions among physicians were associated with the adoption of new drugs. The studies included in Lublóy’s review were predominantly from the USA and did not consider drugs used in cardiology. Donohue et al. [16] describes the influence of a peer network on the adoption of new drugs: a 10% increase in peer adoption in the patient-sharing network led to a 5.9% (standard error [SE] = 1.5%, p < 0.001) increase in physician adoption of dabigatran, a 8.3% (SE = 1.51%, p < 0.001) increase in adoption of sitagliptin, and a 7.84% (SE = 2.93%, p < 0.001) increase in adoption of aliskiren.

By treating common patients, it is likely that new recommendations will be adopted by physicians in line with social contagion. Social contagion is understood as the process in which one person adopts a new attitude or behaviour from another person either because the actor is directly influenced by them in the context of an interaction relationship (cohesion) or because the actor is in a structurally similar position within social networks (structural equivalence) [17].

In this study, we explored the role of collaboration and exchange between ambulatory care physicians in their prescription of sacubitril/valsartan, a new ingredient combination in Germany. In particular, we aimed to provide a quantitative estimate of the impact of physicians’ patient-sharing networks on the prescription of medication.

The prescription of ARNI has continuously increased since its approval of access to the market in 2016. In the first year of approval (2016), 1127 (13.5%) physicians prescribed ARNI. Prescription of the new drug increased in the following years: in 2017, it was prescribed by 2110 (25.2%) physicians, and in 2018, it was prescribed by 3075 (36.5%) (see Table 1).

For the shared patients’ network in 2017, we initially identified 3,123,514 insured users of health services and 8370 ambulatory care physicians. These physicians had 144,636 connections between one another (that is, shared patients), 443 of whom were not active in Baden-Wurttemberg in 2018, leaving 7927 physicians with the main outcome prescription ARNI in 2018 and who shared at least five patients with other physicians. The largest group of physicians were family physicians with general practitioners, which together were 7071 (89.2%) physicians. Table 2 shows comparisons between prescribers and non-prescribers. While only 4.5% of cardiologists in 2017 were represented in the overall group of physicians, the number rose to 7.1% within the prescriber group.

After the exclusion of physicians with a constraint of 1, the network properties for 7831 physicians were calculated and included in the subsequent regression analysis. As previously described, we suspect that these physicians are located on the outskirts of Baden-Wurttemberg and no reliable network can be created, as no data outside of Baden-Wurttemberg was available. Table 3 shows the network characteristics subdivided according to prescription of ARNI in 2018. Prescribers in 2018 had more contacts with physicians who had already prescribed the new drug in 2017 as well as non-prescriber in the network. Non-prescribers in 2018 had a higher network constraint: median 0.08 (IQR 0.06) compared to prescriber median 0.06 (IQR 0.03). The prescribers exhibited a higher betweenness centrality median 0.03 (IQR 0.14) compared to the non-prescriber median 0.01 (IQR 0.08) in the network.

Univariable and multivariable regression analyses showed no significant associations between prescribing and physician age. For this reason, we excluded age from the final model because it acted only as a covariate and no change in the explained variation was registered. The test for multicollinearity showed a very high correlation between the variables ‘link to ARNI prescriber’ and ‘degree centrality’ eta² = 0.69. For this reason, we excluded degree centrality from the full model. Furthermore, constraint correlates negatively (eta² = 0.56) with ‘link to ARNI prescriber’. The full final model shows an average marginal effect [AME] of 0.04 (confidence interval [CI] 95% 0.01–0.06), p < 0.01 for a medium physician’s connection and an AME of 0.07 (CI 95% 0.05–0.10), p < 0.001 for many links to an ARNI prescriber in comparison to few connections. In addition, constraint had an AME of − 0.05 (CI 95% − 0.07 to − 0.03), p < 0.001, which affects ARNI prescribing in 2018. The effects of betweenness centrality were not significant in the final model (see Table 4).

As a sensitivity analysis, we calculated a model without the network property constraint, since a correlation of eta² = 0.56 of the variable’s constraint and ‘link to ARNI prescriber’ was shown. In this model, similar to the previously described model, medium links to ARNI prescriber show an AME 0.07 (CI 95% 0.05–0.09), p < 0.001, and many links to ARNI prescriber an AME 0.12 (CI 95% 0.09–0.14), p < 0.001, of being adjusted for the variables already described. In addition, we performed a subgroup analysis with only the 2017 non-prescribers, which confirmed our previous regression model results (see Additional file 2). A further sensitivity analysis with a zero-inflated model shows that the number of heart failure patients had a greater impact on the presence of excessive zeros than on prescribing behaviour.

Controlled for other factors, such as physicians’ specialty, this study found associations between professional networks of physicians and the adoption of a new drug. Physicians who have many connections to other physicians who already prescribed ARNI, who are well integrated into the patient-sharing network (high-degree centrality), and who have many connections to others who are not connected with each other (low constraint) were more likely to prescribe ARNI. Overall, the study suggests that exchange on the basis of shared patients influences physicians’ beliefs and routines regarding the prescription of medication.

Our study showed that prescriptions of the new drug combination (ARNI) are gradually increasing. This increase was also found in other studies [10‐12, 22]. Previous studies [14] showed that specialists adopted new practices earlier than general practitioners. This is confirmed by our results for the new drug ARNI. In their qualitative study of prescribing behaviour after myocardial infarction, Freier et al. [23] describe how general practitioners feel uncertain about prescribing medications due to side effects, intolerances, comorbidities, and the number of guideline recommendations. This study also reported that general practitioners often ask cardiologists for advice and adopt the prescribing behaviour of cardiologists whom they consult. In addition, general practitioners may purposefully reduce the number of different drugs that they prescribe in order to learn more effectively about the benefits and risks of these drugs. Cardiologists tend to share patients with many other physicians and they tend to prescribe new cardiological medication earlier than primary care physicians, so the identified effects may not reflect network mechanisms only. However, the effects of high degree centrality were also found for general practitioners and family doctors, as they existed independently of physicians’ specialty. This supports the hypothesized network effects on medication prescribing decisions. Physicians’ mutual connections through shared patients may influence their opinions on prescribing ARNI, which may be explained by mechanisms such as social influence (e.g. imitation), opinion leadership, and dependency with respect to patient referrals. The strength of the effect of the network on prescribing in the observed 2 years may be described as small to moderate, but the accumulation over several years implies that it ultimately has a substantial impact.

Our results are in line with those of Donohue et al. [16] and Sundmacher et al. [24], who showed that physicians whose peers prescribed the new drug are more likely to adopt the new medication. In the classic study ‘Medical Innovation: A Diffusion Study’, Coleman et al. [25] describe that social interactions accelerate the adoption of tetracycline, a broad-spectrum antibiotic, in the first few months after market approval. Roger’s ‘diffusion of innovation’ theory [26] also points to the effect of social influence. Central actors play a major role in the adoption process. Through their many connections, they receive information more quickly and thus adopt new innovations earlier than less central actors. As measured by degree centrality, we showed that prescribers had more linkages in the network than non-prescribers. This is consistent with the findings of Iyengar et al. [27]. They showed the adoption of a new drug for treatment of viral infection. Prescribers had more incoming connections (high in-degree) with other physicians versus non-prescribers. The analysis took place in a network in which referrals form the links.

Some network mechanisms explain the increasing homogeneity of beliefs and behaviours in a network, but homogeneity itself does not necessarily imply a specific direction of change (e.g. increasing prescribing of a new drug). The latter requires a driving force, such as trusted clinical guidance or a trusted opinion leader. Having many connections is one feature of clinical opinion leadership [28, 29]. Valente and Yon [30] set in their simulation study the starting point for the spread of an innovation among physicians with high in-degree, who are referred to as opinion leaders. This led to a faster and farther spreading adoption of the innovation compared to a randomly selected starting point.

In our study, physicians who have many connections to other physicians who are not mutually connected, as measured by constraint, were more likely to be connected to physicians who prescribed the new drug than physicians who were strongly involved in the network. These low-constrained networks are more open to new information. According to Granovetter’s theory of weak ties [31], individuals on bridges of networks can bring new information into a network and thus promote for example the adoption of a new drug. In a closed network with a high constraint, there are no such persons and new information might not reach the network members. Such information can reach physicians at so-called quality circles [32], for example. Quality circles are meetings between physicians where freely selected topics are discussed in small groups and experiences are exchanged [33]. In a qualitative study of antibiotic prescribing behaviour, Poss-Doering et al. [34] showed that primary care networks can lead to more guideline-compliant prescribing behaviour through social influence and social support.

In contrast to degree centrality or constraint, betweenness centrality did not explain the uptake of ARNI. This could be explained by individuals with high betweenness centrality have power over others because they can control the flow of information between two people. Since this is not accompanied by social contagion, our theoretical considerations are confirmed.

Structured care models such as the disease management programme for coronary heart disease were included as a covariate in the analysis, because participation in this programme is more likely to lead to the prescription of the new drug than non-participation. Roehl et al. [35] show that physicians in such programmes are more likely to implement the recommendations of the guidelines than those who are not in the programme.

A major strength of the study is the longitudinal observational design, in which predictors preceded outcomes in time. This design facilitates a causal interpretation of the correlations, although we cannot fully rule out confounding. We were able to quantify network effects in a large real-world study. A further strength of the study is the inclusion of the whole network of physicians in a large region through claims data from a health insurer. Moreover, and typical for claims data, the data is nearly complete with only a few missing values for location.

Nevertheless, a limitation of claims data is that it is not possible to verify the indication for prescribing the new medication, because there are only International Statistical Classification of Diseases and Related Health Problems (ICD-10) presented and no clinical data. For example, the ejection rate of the heart is not shown in the health insurance data. Also, our study did not include patient outcomes, such as the number of patients who benefit from a prescription of ARNI. There is a possibility that some physicians do not have the opportunity to prescribe the new drug or that they do not have patients who need the drug. Another possible confounder may be links to pharmaceutical companies. The data do not allow any statement on a possible influence by ‘advertising campaigns’. Finally, we used data from one health insurer only, which is known to underrepresent specific population groups such as citizens with high income and specific employees in public services who might be the first to be treated with this new drug. However, we expect that this only mildly affected our results, as nearly all physicians were included in our analysis, and on the physician level, no structural bias is to be expected. An analysis with the data from 2019 would also have been desirable, but the data were not available at the time of analysis.

Our study provides evidence that physicians’ prescribing behaviour and uptake of a new drug are influenced by connections to other physicians via shared patients. Physicians with more connections to physicians who already prescribed ARNI were more likely to prescribe the combination of sacubitril/valsartan, independent of physicians’ specialty.