Introduction

After inpatient and outpatient care, pharmaceuticals account for more than about a sixth of health care expenditures in industrialized countries [1]. In the 1990s and early 2000s, annual increases in pharmaceutical expenditure growth rates (~5%) outpaced overall health spending growth rates (~4%) [2]. To tackle this problem, a number of countries introduced measures to contain pharmaceutical expenditure. Besides other regulations, drug budgets have been introduced as one instrument to control overall spending on the demand side of the pharmaceutical market [3]. Vogler et al. reported ten European countries that aimed to contain pharmaceutical spending through drug budgets in 2010 [4, 5].

The general idea of drug budgets is to define a maximum amount of money to be spent on pharmaceuticals ex ante and for a specific time period [4]. Drug budgets can be interpreted as a contract between third-party payers and physicians that limits or avoids ex post moral hazard behavior of the physician towards the payer and thereby reduces the risk of service costs that are ex ante not measurable (hold-up problem) [6, 7].

However, drug budgets may differ in the level at which they have been implemented as well as in their specific design: (a) at the national level, e.g., negotiated between the public system as a whole and the pharmaceutical industry (Ireland, Italy, France) and (b) at the regional level, i.e., negotiated or set for all physicians in a region (Sweden), as well as (c), as applied by the majority of countries, at the level of the prescribing physician, i.e., physician-specific prescribing targets (Czech Republic, Germany, Latvia, Slovakia, United Kingdom) [4, 8]. Also, enforcement of overspending differs across health care systems. Examples are audits, consultations, or claw-back mechanisms. This may also comprise financial rewards if targets are met as some type of pay for performance intervention [9].

Given the widespread application of drug budgets and similar instruments such as prescribing targets that link prescribing to financial bonuses or penalties, there is surprisingly little research on how physicians react. Existing evidence suggests that budget caps and targets reduce total cost whereas drug use remains constant [9]. Only for the UK fundholding system has it been shown that physicians achieve cost reductions mainly through increased substitution of brand-name products through generics. However, the quality of pharmaceutical care comprises multiple dimensions beyond generic substitution. Changes in other dimensions of prescribing such as the variety of products used or the share of pharmaceuticals prescribed that are deemed inappropriate in the elderly may be suboptimal compared with situations without a drug budget mechanism. Another notion is that no single study concentrates on differences in enforcement of drug budgets. Our study has two aims: (a) to analyze the impact of physician-level drug budgets on prescribing behavior to reveal how the reduction in pharmaceutical spending from drug budgets is achieved; and (b) to compare different modes of enforcing drug budgets.

To do so, we analyze panel data from 440 German outpatient physicians over a period of 7 years. We differentiate between the effect of drug budgets on the costs and quality of prescribing. We choose Germany as an example because the German design of drug budgets implies different enforcements depending on how much a physician overspends his or her budget. Annually, a random sample of 10% of physicians is chosen for an audit of prescription behavior. Although spending below 115% of the drug budget is not sanctioned at all, overspending above 115% may lead to timely audits or consultations. Overspending of more than 125% can trigger a financial claw-back mechanism.

Evidence from previous studies on drug budgets

A systematic review of the literature by Sturm et al. and an update of their literature search identified 18 studies that deal with the effect of pharmaceutical spending constraints at physician level in eight different settings and in six high-income countries [9, 10]. Of these, 14 addressed various drug budget policies. The majority of studies discussed the impact of general practitioner (GP) fundholding in the UK National Health Service between 1995 and 2002. The fundholder (i.e., the GP) received capitation, i.e., a lump-sum payment per patient. The payment was used to pay for all health care services delivered to patients including pharmaceutical care [11].

Surveys of patients and providers suggest that physicians with fundholding status prescribe less expensive pharmaceuticals than non-fundholding physicians [12, 13]. Dowell et al. [13] also spoke to patients whose pharmaceutical treatment had changed after their GP had opted for fundholding. Patients felt confident with cheaper treatment if the financial necessities of a switch had been explained by their physicians. The findings suggest that quality of prescribing does not seem to have been affected.

Studies that based their analyses on prescription data found ambiguous results for fundholding. Three studies found that GPs who are subject to fundholding reduce pharmaceutical spending significantly compared with those not subject to fundholding [14–16]. However, Harris and Scrivener reported that the differences were short term only, i.e., for the first 3 years [15]. Another set of (follow-up) studies examined fundholding over a longer time frame, but did not come to a consistent conclusion [17–21]. This is partly because the design of the UK fundholding mechanisms enabled self-selection. That means that physicians who prescribe efficiently per se may have signed up for fundholding disproportionately.

In the Swedish version of drug budgets, pharmaceutical spending is managed at the level of counties. Jansson and Anell found that physicians show higher cost awareness when subject to drug budgets [22]. Andersson et al. compared physicians working in public and privately owned health care facilities but limited their research to five therapeutic groups [23]. Adherence to guidelines increased significantly over time in three out of five high-volume therapeutic classes with stronger effects for publicly employed physicians compared with private practitioners when physicians were subject to drug budgets. Granlund et al. compared prescription patterns in two counties before and after the introduction of drug budgets [24]. They concluded that prescription behavior had not changed overall. Although the total number of prescriptions declined, the cost and quantity per prescription remained unchanged. An explanation for this behavior was that providers did not believe in the enforcement of drug budgets.

The German drug budget system was introduced in 1992 and started as spending caps that were negotiated at regional level but were not made accountable at physician level in the first place [25]. Since 2002, the system combines ex ante definition of the budget at physician level and a penalty mechanism if physicians are found to be overspending their budgets. Existing evaluations of this system refer to early periods between 1992 and 2002. Schreyögg and Busse found that prescription costs have significantly reduced overall [25]. In particular, the number of prescriptions filled for substances with disputed effectiveness—a measure of prescription quality—has declined. Assuming that referrals to specialists for drug therapy replaces primary care, Schöffski concluded that GPs increased referrals to specialists to avoid costs being accounted to their drug budgets. Thus, lower pharmaceutical expenditure by GPs was compensated by higher spending by specialists [26]. For the treatment of diabetes, Jünger et al. reported that drug budgets do not have an impact on actual treatment decisions, i.e., on prescription patterns [27].

In summary, previous studies have analyzed selected indicators (mainly costs) at a macro level, i.e., the overall change in pharmaceutical spending after the implementation of drug budgets. How the individual physician achieves this reduction in spending has only been analyzed using surveys, besides analyzing changes in generic share in the UK fundholding system. The fact that physicians typically face different patient structures in their practices has been widely neglected.

Methods

Data

Our study is based on a dataset from the IMS Disease Analyser. We used information on 440 outpatient practices (GPs and specialists) located in three German federal states, i.e., Berlin, Baden-Württemberg, and North-Rhine Westphalia, between 2005 and 2011. The dataset provides complete information about prescriptions (date of prescription, 5th-level ATC code, substance, brand name, gross pharmacy retail price), information on the prescribing physician (specialization, number of physicians in practice and number of employees in practice) as well as information on the patients in each practice (age, gender, diagnoses according to ICD-10-GM, and insurance status).

Calculation of drug budget and its utilization

We retrospectively calculated the drug budget for each physician. The drug budget of a physician in Germany depends on patient structure, patient volume, the physician’s specialization, and the federal state in which the physician works. In a first step, the Regional Association of SHI Physicians and the regional associations of the sickness funds negotiate a regional budget. In a second step, the regional drug budget is broken down to the physician level. To do so, the Regional Association of SHI Physicians calculates reference values for predefined patient subgroups by physician subspecialty per quarter based on previous pharmaceutical expenditure. These reference values are then multiplied by the number of patients in a subgroup the physician treated in a quarter, resulting in a quarterly drug budget at the physician level (see “Appendix Note 1” for a formal specification).

For example, a GP in the federal state of Berlin received EUR 41.54 in 2011 for a not-yet-retired patient per quarter, whereas a neurologist would have received EUR 170.20 for the same patient per quarter. As these figures differ by federal state, a GP in Baden-Württemberg would have received EUR 50.15 per quarter for the same type of patient, and a neurologist EUR 167.91.

Based on the information about the physician’s specialty, the characteristics of their patients, the number of patients that have visited the physician in a quarter, and the reference values provided by the regional associations of SHI physicians, we retrospectively calculated the drug budget of each physician in our sample on an annual basis since this is the time frame physicians may become subject to reviews.

To calculate the utilization of drug budgets, we identified annual pharmaceutical expenditure from prescriptions for each practice. As certain prescriptions are excluded when monitoring the physician’s compliance with drug budgets by indication or by substance name, we corrected for these exceptions. This is done to exclude rarely prescribed but very expensive prescriptions that would create distortions when controlling prescribing behavior.

Utilization of drug budget of physician i in year t is defined as corrected annual pharmaceutical spending divided by the drug budget. To reduce bias from extreme prescribers in our sample, we deleted observations outside the 99th percentile, i.e., with a utilization lower than 19.2% or higher than 342.2%.

Outcome measurement

To analyze the impact of drug budgets on prescribing behavior, we differentiated between (I) cost of prescribing and (II) quality of prescribing. We measured cost of prescribing by (I.a) the generic share of prescriptions among those prescriptions eligible for generic prescribing, (I.b) the number of prescriptions per visit, (I.c) the number of branded prescriptions that were not eligible for generic substitution, and (I.d) the concentration of generic brands among generic prescriptions measured according to the Herfindahl index.

For generic share (I.a), we assumed that the most effective way for the physician to lower the cost of prescribing is to support generic substitution. In particular, this should be valid for GPs as a wide range of generics exists among most of the prescribed substances. Furthermore, the generic share is an indicator of the price awareness of the physician that can be influenced directly. We measured generic prescribing if the physician had written INN or a generic product on the prescription. Given the constraints posed upon the physician by drug budgets, we expect the number of prescriptions per visit (I.b) and the number of branded prescriptions for which no generics exist (I.c) to decrease. We would also expect the same for the concentration among generic brands (I.d) as the physician may compromise personal preferences for specific manufacturers in order to prescribe less expensive products.

Quality of prescribing was measured by (II.a) the concentration of substances among all substances (Herfindahl index), (II.b) the concentration of brands among prescriptions not eligible for generic substitution (Herfindahl index) and (II.c) the share of potentially inappropriate medication for patients aged 65 years and above. This share is derived from a list of substances that show side effects in the elderly or interaction with other substances usually taken by the elderly. It was originally proposed by Beers et al. in 1991 for long-term care in nursing home residents in the United States [28]. Since then, it has been updated and adapted in countries such as France, Austria, Norway, Ireland, or Germany [29]. It includes substitutes to be used instead.

If quality was affected by drug budgets, we would expect the concentration of substances to vary depending on previous utilization of drug budgets. Here, we assume that physicians believe that their patients have been treated well in the last period. Therefore, they will not change their treatment strategy. Thus, (II.a) and (II.b) should not be affected by the utilization of drug budgets. However, the share of potentially inappropriate medication (II.c) should decline as these are potentially harmful substances. A physician may cease to prescribe these at any time to increase the quality of prescribing.

Empirical strategy

We hypothesized that our j indicators for cost or quality of prescribing of physician i in period t (outcome j,i,t ) depend on the utilization of the drug budget of physician i in period t − 1 (utilization i,t − 1 ). We control for our indicator j in previous period t − 1 (outcome j,i,t − 1 ), the patient structure of physician i in period t (pat_controls i,t ) and physician characteristics (phys_controls j,t ). We therefore estimated the following lagged dependent-variable regression models for each of our indicators within the general linear model framework:

$$ {\text{outcome}}_{j,i,t} = f({\text{utilization}}_{i, t - 1} ;{\text{outcome}}_{j, i, t - 1} ;{\text{pat\_controls}}_{j,t} , {\text{phys\_controls}}_{j,t} ). $$

Our main variable of interest was the utilization of the drug budget of our physician i in the previous period (t − 1) while controlling for patient structure and physician characteristics in period t.

We controlled for (a) physician characteristics (i.e., physicians’ specialization, single or group practice), (b) patient structure of the physician (i.e., mean age and gender of patients, the average number of visits per patient, the comorbidity according to pharmaceutical based metric groups) and (c) federal state fixed effects as well as time fixed effects (years). Pharmacy-based metric groups are a measure of patient morbidity based on prescription data [30]. We chose to prefer pharmacy-based metric groups as they have been validated for the outpatient setting, whereas the Elixhauser index [31, 32] or the Charlson comorbidity index [33] are validated for inpatient settings only.

For each regression model, we estimated two specifications: (a) outcomes as a (non-linear) function of utilization in period t − 1 to analyze utilization on our outcome measures and (b) outcomes as a (non-linear) piecewise function for utilization in t − 1 to be between 0 and 115, 115, and 125% and above 125%, respectively, to test whether the results differ by type of enforcement (no formal enforcement, audit/consultation, becoming financial liable).

Model specification and sensitivity analysis

After including the variable of interest and all control variables, we tested different specifications of our variable of interest utilization in period t − 1, i.e., quadratic and cubic specification. We then applied forward selection to include interaction effects as well as quadratic functional forms of the control variables. We inserted control variables in the order of them increasing model fit if below the threshold of p < 0.2.

To test the distribution of our general linear models, we computed the Modified Park test for each regression model. In addition, we performed Pregibon’s link test to test the optimal link function. To account for serial correlation, i.e., clustering at the physician level, we considered different variance–covariance structures. For each model, we chose among different structures, i.e., variance components, compound symmetry, unstructured, Toeplitz and Toeplitz with two bands, heterogeneous Toeplitz and first-order autoregressive moving average according to the pseudo maximum likelihood. “Appendix Note 2” provides an overview of model specifications for our eight regression models.

To establish whether our results were robust, we first performed the same regression models with different exclusion criteria for outliers, i.e., 10%/90% percentile. By this method, we were able to observe that the results are not only driven by some special physicians but also correspond to drug budgets. Second, we also excluded observations based on Cook’s distance (i.e., excluding any observation with an influence of more than 4/n).

Results

Table 1 provides descriptive statistics of our study sample and our outcome measures for cost and quality of prescribing. Across all physicians and years, the mean utilization of drug budgets was 92.3% (STD 100.5). Some 52.4% of physicians did not overspend their budget at all, whereas 26.4% of physicians spent 115–125% of their budget, so potentially facing an audit and/or a consultation. A total of 21.1% exceeded the budget by more than 125%, i.e., they could become liable for overspending.

Table 1 Sample characteristics and summary statistics
Full size table

The results of the regression models (Table 2) suggest that the utilization of the drug budget influences the cost and quality of prescribing, but the level of enforcement does not. Predictions of our outcomes depending on the utilization of the drug budget in the previous year with all other variables set to their mean are depicted in Fig. 1.

Fig. 1
figure 1

Effects of drug budget utilization and different levels of enforcement on the cost and quality of prescribing. Solid lines represent the level of the outcome measure analysed for the level of drug budget utilization. Dashed lines represent the level of the outcome measure by the level of enforcement. Dotted lines show the distribution of physicians

Full size image
Table 2 Regression results
Full size table

Our indicators for cost of prescribing in column I of Table 1 show that the utilization of drug budgets significantly influences generic share (p = 0.0246) and the concentration among generic brands (p = 0.0056). When the utilization of drug budgets in year t − 1 with all other variables set to their mean increases by 10%, the generic share increases by 0.04% in year t. On a national basis, this corresponds to about 207,000 prescriptions eligible for generic substitution. However, the number of prescriptions per visit and the number of branded prescriptions per visit remain unaffected by the utilization of the drug budget in the previous year.

For quality of prescribing, utilization of the drug budget affects two of the three measures. The concentration among therapeutic substances decreased if utilization increased (p < 0.001). This means that physicians prescribed a larger variety of substances. We also observed that the share of prescriptions deemed inappropriate for the elderly increased (p < 0.001) if the utilization of the drug budget increased.

Differentiating by means of enforcement (i.e., audit/consultation vs. no enforcement and becoming liable for overspending vs. no enforcement), as shown in columns II and III of Table 2, we found no significant effect for any of our indicators that describe the cost and quality of prescribing.

Results were robust against the sensitivity analyses that we performed. Excluding outliers based on the 10 and 90% percentiles for the variable utilization of drug budgets did not change the results.

Discussion

Our results suggest that physicians are sensitive to the utilization of their drug budget in the previous year. Indicators for cost of prescribing show that physicians tend to avoid costs by increasing the generic share and selecting generic brands more carefully. At the same time, however, the total number of prescriptions does not change. This suggests that physicians seek to continue prescribing if considered necessary. The fact that we found no significant effects for the number of branded prescriptions per visit supports this idea.

Previous studies have focused on the cost per item prescribed and changes in overall pharmaceutical spending [9]. The results provided for Germany confirm the notion that drug use expressed as the number of prescriptions per visit does not change, whereas the cost of prescribing changes when a drug budget mechanism is put in place. For example, Granlund et al. found no change in the cost and quantity of prescriptions after the implementation of drug budgets for the Swedish system of drug budgets [24]. Our findings on the number of prescriptions per patient visit are the same in the sense that physicians do not modify the overall quantity of prescribing as expressed by the number of prescriptions per visit. It has to be noted though that total pharmaceutical spending in Germany has increased continuously in our study period, i.e., between 2005 and 2011 [1]. Because drug budgets interfere with other types of regulation such as, for example, preferred supplier contracts [34], we cannot make a statement about the total effect of drug budgets on pharmaceutical spending.

In terms of economic efficiency, our results suggest that physicians incur changes in the cost of prescribing through generic substitution (i.e., the generic share and the concentration of generic brands), which has also been observed for the UK fundholding system [19, 21], but has not been analyzed in Sweden. Generic substitution seems to be one viable means of complying with drug budget regulation across different settings and reducing the cost of prescribing [9]. Although explicit policies to encourage or mandate generic substitution also exist in Germany [5], our results suggest that physicians increase their generic share even more when they face sanctioning because of overspending of the drug budget. However, effect sizes appear to be rather small. Still, when interpolated to the German pharmaceutical market with 625 million prescriptions annually in 2011 [35], an increase in generic prescriptions of 0.04%, i.e., about 207,000 prescriptions eligible for generic substitution in the off-patent market, will yield substantial cost savings.

Against our expectations, we identified changes in the concentration of therapeutic substances and the share of potentially inappropriate medication in the elderly. This suggests that physicians reconsider the different treatment options available to make adjustments to prescribing costs. Although relying on a larger variety of therapeutic substances does not imply worse quality of treatments per se, increasing the share of medication deemed inappropriate for the elderly suggests a trade-off in the cost and quality of prescribing.

Comparing the German drug budget system with others [5], the enforcement mechanism analyzed as a form of sanctioning is one of the stricter regulations for implementing drug budgets. In contrast to the German system, the UK fundholding mechanism and the Swedish drug budgets raise the prospect of using potential savings in drug budgets for future patient benefit or as additional personal income [16, 23]. Behavioral theory suggests that individuals are typically more risk averse towards facing losses instead of gains [36]. Accordingly, the incentive not to overspend the budget should be higher in Germany. Still, we did not find evidence that varying levels of sanctioning lead to substantial changes in the cost and quality of prescribing.

Potentially, the design of the enforcement mechanism may provide explanations for our results: (1) Only a random sample of 10% of physicians is subject to annual audits. This means that the likelihood of becoming financially liable is generally very low (statistically once every 10 years). (2) Drug budgets are calculated rather generously. The mean utilization of the drug budget in our sample was 92.7%, whereas potential sanctions start above 115% (consultation on prescription behavior) and above 125% (becoming financially liable). (3) Even when exceeding budget by 125%, physicians may excuse themselves, e.g., for having a special patient structure. As long as the physician has always prescribed efficient, i.e., a generic when available, no clawback can be enacted. In fact, very few physicians were ever held financially liable in practice [37]. (4) The time lag between overspending and sanctioning can be very long. Besides, the IT infrastructure typically does not provide instant information about the utilization of the drug budget in a given year. For this reason, several physician associations have put forward initiatives to monitor prescribing behavior at lower levels and in a more timely manner, e.g., by measuring generic share for selected therapeutic classes [38]. The situation is perhaps similar to the Swedish system of drug budgets where the enforcement mechanism was not considered credible by physicians [24].

This study has several limitations. First, we only observe prescriptions. As patients may not have filled every prescription, the true utilization of drug budgets is likely to be lower. Second, we cannot observe whether a physician in our sample was selected for an audit or has been liable for overspending in a particular year. Third, the list of medications deemed inappropriate for the elderly is not legally binding such that knowledge about potentially inappropriate medication may vary across prescribers. Finally, our physician panel covers only three federal states of Germany representing about 40% of the total population. Compared with national data from 2005, the physicians in our sample are older. Male physicians and GPs are somewhat over-represented.

Conclusions

This study provides evidence on how physicians react to drug budgets as a means to contain pharmaceutical spending when facing a sanctioning mechanism. Physicians seem to reconsider the treatment options more carefully to achieve lower costs of prescribing while the number of prescriptions is unaffected, especially by increasing the generic share and selecting generic brands more thoroughly.