A quantification of the risk implied by a risk-sharing
agreement for Rheumatoid Arthritis drugs
University of Groningen
Master’s Thesis Econometrics, Operations Research and Actuarial Studies Supervisor: Dr. C. Praagman
A quantification of the risk implied by a risk-sharing
agreement for Rheumatoid Arthritis drugs
S.D. van Oeveren
May 5, 2020
Abstract
Contents
1 Introduction 5 2 Literature review 6 3 Methodology 7 3.1 Refundable patients . . . 8 3.2 Risk magnitude . . . 10 3.3 Risk sources . . . 10 4 Data 13 5 Results 15 5.1 Switchers . . . 15 5.2 Risk magnitude . . . 19 5.3 Risk sources . . . 19 6 Conclusion 25 7 Discussion 25 7.1 Limitations . . . 26 7.1.1 Data . . . 267.1.2 Method and results . . . 26
1
Introduction
Due to aging of the population as well as our growing wealth and the availability of new technolo-gies, overall health care costs are growing rapidly. Dutch health care costs are expected to double between 2015 and 2040 (Rijksinstituut voor de Volksgezondheid [19]). One of the fastest rising components of the health care costs are prescribed drug expenses. This rise is caused by a greater expense on existing drugs but also by the availability of new and expensive drugs. Currently almost 9 percent of the total medical specialists’ budget goes to expensive drugs and this percentage is still growing (Kennis -en Exploitatie centrum Offici¨ele Overheidspublicaties [14]). In the search for methods to control overall expenses and to improve cost-effectiveness of these new expensive drugs there has been a push to the implementation of value-based pricing and the interest in risk-sharing agreements has grown (Piatkiewicz et al.[16]).
Risk-sharing agreements are ”agreements concluded by payers and pharmaceutical companies to diminish the impact on the payer’s budget of new and existing medicines brought about by either the uncertainty of the value of the medicine and/or the need to work within finite budgets” as defined by Adamski et al. [1]. An example of a risk-sharing agreement is a discount on drugs provided by the pharmaceutical company, for an agreed upon time period to establish the value of the drug. Another example is a warranty form of risk-sharing where the response to a drug is evaluated per patient and the pharmaceutical company refunds the drug expenses for each patient that has a response below an agreed upon level. Pharmaceutical companies are interested in forms of risk-sharing agreements in order to be able to put their drugs earlier on the market, or to obtain a better competitive position. Payers are interested for the reason that they can provide patients with new technologies while controlling for the growing expenses.
The risk implied by risk-sharing agreements is caused by uncertainty regarding the effectiveness of the drug in the real-time population and under real-time circumstances. When risk-sharing agree-ments are incorporated at market entry the effectiveness of the particular drug is only known from the clinical trial. The way in which the practical effectiveness affects reimbursement also depends greatly on the exact specifications of the agreement. Most existing literature on risk-sharing agree-ments stresses the importance of the conceptual framework of risk-sharing agreeagree-ments, or focuses on the potential impact on social welfare. However, due to aforementioned reasons the existing lit-erature usually leaves the risk unspecified. This implies that despite the body of evidence regarding the importance of conceptual frameworks of risk-sharing agreements and the evidence on how the magnitude of risk affects the possible impact of risk-sharing agreements on social welfare (see next section), the evaluation of the actual magnitude and sources of risk enclosed in these risk-sharing agreements is falling behind. This contributes to the lack of knowledge about risk-sharing agree-ments which may prevent it from reaching its full potential.
One of the interesting drug groups for risk-sharing agreements are the Rheumatoid Arthritis (RA) drugs, and in particular the biological Disease-Modifying Antirheumatic Drugs (bDMARDs). Only little is known about the interchangeability of these drugs and only little research has been per-formed on the implementation of risk-sharing agreements for chronic disease drugs. This, despite the fact that bDMARDs are in the top 3 of most costly drugs per year (Kennis -en Exploitatie centrum Offici¨ele Overheidspublicaties[14]).
hospital a pharmaceutical company is to refund as a result of insufficient response to the drug. I determine the response to a drug based on existing claim registries resulting in an investigation on large scale. Methods to estimate this sort of risk on this kind of data are very scarce. In result, this study is unique and provides a first step in the examination of methods to perform research about the quantification of risk implied by risk-sharing agreements.
The next section discusses the existing literature on risk-sharing agreements. Section 3 opens with a framework explicating some background on RA drugs and the drug market. The rest of the section explains the methods being used in this research. Section 4 subsequently describes the data. In section 5 the results will be presented which lead to the conclusions drawn in section 6. Finally, in section 7 the limitations of the study and recommendations for further research will be discussed.
2
Literature review
The number of implemented risk-sharing agreements over the world has been rising (Towse et al. [25]). Current existing risk-sharing agreements can be categorized as either financial based (price volume agreements, patient access schemes) or performance/ outcome based (Saggia M. [20]). Price volume agreements are focused on the control of expenditures by forcing pharmaceutical companies to reimburse over-budget situations. An example was introduced in Australia with price reductions if the sales exceeded a pre-agreed volume (Adamski et al. [1]). With the use of patient access schemes it is common that a pharmaceutical company provides free drugs or drugs with a discount for an agreed upon time period in order to enhance the value of the drug. Also, the use of price caps is a commonality in patient access schemes. An example of such a price cap was implemented in the UK where the first 14 injections of Ranibizumab were paid for by the national health service in case a patient’s response was ”adequate”. The costs of the exceeding number of injections were reimbursed by the pharmaceutical company (Neumann et al. [15]).
Performance/ outcome based schemes imply new price negotiations after a trial period win which the performance/ outcome is established, or a refund of drug expenses based on average population or patient specific outcomes. In Australia the price of Bosentan was adjusted with the ratio of costs and the Quality-Adjusted life year (QALY), so that the price increased once the treatment with Bosentan showed to result in a greater QALY (Wlodarcyzk et al. [27]). Another example is the agreement of a refund made by ”Merck” in the US. The costs of their drug against benign prostatic hyperplasia were refunded for patients that did not show disease reduction within six months or needed surgery within two years (Adamski et al. [1]).
As can be concluded from this overview, multiple risk-sharing agreements have been implemented, but each with their own specifications. Esp´ın et al. [8] conducted a survey among 6 European coun-tries and concluded that only some councoun-tries have started the evaluation of risk-sharing agreements, but no results have emerged yet. This supports the statement of Towse et al. [24], [25] that there is a lack of open access evidence on the success of already implemented risk-sharing agreements. A range of theoretical studies have been conducted on the impact of risk-sharing agreements. Economic approaches resulting in recommendations on best practices have been performed by for example Gon¸calves et al. [12] and Towse et al. [24]. They both conclude that there are mainly a lot of challenges in the implementation and that the benefit on social welfare is only obtained in certain situations.
so-cial welfare is subsequently evaluated. Antonanzas et al. [3] built upon this model by including bargaining power and approaching the situation in a game theoretical manner. A Markov model is used by Carletto et al. [5] taking the probability of drug response as the transition probability and Zaric and Xie [29] use a two-period model estimating a pharmaceutical company’s profit and social benefit based on some expected distribution of the drug’s effectiveness.
The great variety in approaches and specifications of these studies results in sometimes conflicting and not comparable conclusions. However, they still provide some insight in the factors influencing the impact of risk-sharing agreements.
Overall there is a lack of empirical evidence on the performance of the existing risk-sharing agree-ments and the theoretical studies on the impact of risk-sharing agreeagree-ments are all based on some assumption of the distribution of a drug’s effectiveness, leaving the actual risk undefined.
3
Methodology
This section starts with a framework providing background information about RA drugs and the drug market. The frame gives context to the subsequent methodology description, but could be skipped by readers with an extensive medical background.
Rheumatoid Arthritis drugs
Rheumatoid Arthritis (RA) is a long-term autoimmune disorder that primarily causes painful, swollen, and stiff joints. Controlling RA implies constant monitoring and treatment. For the treatment of RA there are pain reducing drugs such as on-steroidal anti-inflammatory drugs (NSAIDs) and drugs that reduce disease progression such as the traditional Disease-modifying antirheumatic drugs (DMARDs), or new biological Disease-modifying antirheumatic drugs (bD-MARDs) and Janus Kinase (JAK) inhibitors. The choice of the prescribed drug depends on the stadium and severity of a patient’s RA. That being said, normally patients start with a com-bination of an NSAID and DMARD. When progression does not reduce doctors can prescribe a switch to a different DMARD in combination with the NSAID or switch to a bDMARD or JAK inhibitor. The latter two are often prescribed in combination with a conventional DMARD such as Methotrexate. A switch within the category bDMARDs or JAK inhibitors can also occur. The progression reduction could stay away because of initial non-response to the drug or due to the forming of anti-bodies over time. In The Netherlands the webpage of the Farmacotherapeutisch Kompas (FK) provides, among other drug information, guidelines about the period after which the response of the drug should be evaluated and the continuation of the treatment process should be determined. The FK is affiliated with the Dutch National Health Care Institute and the proposed guidelines are scientifically based as they are established by a panel of pharmacists and doctors. The aim of controlling RA often involves trying multiple drugs and combinations. Consecutive drugs or drug combinations are indicated as different treatment lines. (Farmacotherapeutisch Kompas [9])
on the European market the biosimilars are tested for their similarity to the bDMARD. The first RA biosimilar was approved in 2013 and currently there are biosimilars available for all bDMARDs. The approval of biosimilars to the European market has improved the competition on the RA drug market and resulted in price decreases. (Smolen et al. [21])
Both the parties involved in the determination of a drug price, as well as the payment structure differ significantly per country. In The Netherlands the pharmaceutical company can initiate a price at entry whereafter the Dutch government states a maximum price for the drug. Sub-sequently the hospitals (or hospital groups) bargain with the pharmaceutical company over the price, after which they also have to bargain with the insurance companies regarding the reimbursement. (K. Wessels and I. Doude van Troostwijk [13])
Patients switch drugs whenever the prescribed drug causes an excessive burden from side effects, or when the progression reduction stays away due to either initial non-response or the forming of anti-bodies. Within a warranty risk-sharing agreement the pharmaceutical company should refund the drug expenses in the first two situations. In order to extract those two situations from an observable switch, an evaluation period will be determined in the form of a time threshold. Initial drug respondents that formed antibodies over time are in this way separated from other patients with observable drug switches.
In this research an insufficient response to a drug will thus be estimated by the switch from and to RA specific drugs that are registered as expensive drugs within a given time period after the first issue. Different treatment lines but with the same drug will be distinguished. Patients issued a particular initial drug that switch to a subsequent drug and then back to the initial one will thus be stated to have been on 3 different treatment lines and made 2 switches. The date of switch will be established as the first issue date of the new drug. The determination of the threshold will be based on the exploration of whether the evaluation of drug response is in compliance with the guidelines as stated by the Farmacotherapeutisch Kompas (FK).
The method of the exploration of this evaluation period will be discussed in the first subsection. The second subsection explains how the magnitude of the risk implied by a risk sharing agreement will be determined and the final subsection explicates the way of establishing this risk’s source.
3.1
Refundable patients
In order to investigate the practice of the guidelines as set by the FK, the distribution of patients’ treatment periods with a particular drug until switching will be investigated. When the evaluation period is applied, this distribution is expected to show a drop of the number of patients switching drugs just after the evaluation period. This drop could occur with some days of delay due to situ-ations such as where the evaluation appointment was postponed due to scheduling issues or other circumstances. The form of the distribution before and after the evaluation period have no strict expected forms besides a small number of switchers around zero days.
A patient’s treatment period will be determined by summing the time between drug issues. For the last drug issue there is no subsequent one and the treatment period resulting from this issue is thus not included.
in-Figure 1: Illustration of selected observations for the analysis, marked green. The bottom line represents a hospital’s time span over which data is available for this research. The actual time span can differ in practice. All lines above represent possible observed patient time spans and their treatment lines in the particular hospital. The number of treatment lines could be expanded in practice. Unknown drug history and future drug issues are indicated by a fading grey line. Situation A represents patients with unobservable drug history and situation B patients with an observable first drug issue.
formation would result in a bias of the treatment period until switch. Hence, the evaluation of the compliance with the evaluation period should be solely based on patients of which the exact treatment period with a drug is known.
The selected treatment lines for the analyses are illustrated in figure 1. This research is based on claim registries which are provided by hospitals. The bottom line in figure 1 gives an example of a time span over which a hospital has provided data and could differ in practice. For patients with observed drug issues over the complete time span of their hospital (situation A) the start date of the first drug and the end date of the last drug are unknown. Therefore, only treatment lines enclosed by others should in this case be incorporated in the analysis. Hence, only the second drug of patient 3 is selected. For patients with an observed first drug issue later in their hospital’s time span (situation B) the starting date of the first treatment line is known. The end date of the last treatment line, however, remains unknown. Therefore, in this case all treatment lines, except the last observed could be included. The first drug of patient 5 and the first and second drug of patient 6 are thus selected.
3.2
Risk magnitude
For a pharmaceutical company the performance of the drug for a particular population and under certain application is known from the clinical trial. There is however uncertainty about the real-time population and the unknown practical application of the drug. I expect that the differences of both elements can be captured by distinguishing between hospitals and I assume that pharmaceu-tical companies enter agreements with individual hospitals. Therefore, the risk resulting from the uncertainty will be investigated by inspecting the spread of the percentage of refundable patients per hospital. A small spread resembles little differences between hospitals and thus a small risk for the pharmaceutical company. A great spread, however, implies that there are great differences between hospitals which are probably caused by either differences in the treatment population and their responses to the drug or the application of the drug.
The establishment of refundable patients within a hospital constitutes the comparison of a patient’s treatment period with drug to the threshold. In figure 1 it is shown that the analysis of the treat-ment period before a switch ( subsection 3.1) could only be based on patients with an explicit treatment start and end date. However, for the determination of the refundable patients the exact start and end dates are not strictly necessary. It suffices only knowing whether the treatment period was longer or shorter than the time threshold. Therefore, the left tails in situation A and the right tails in situation A and B of figure 1 could be included as long as they are observable for at least the time span of the threshold. The data selected thus depends on the drug specific threshold. All refundable patients within a hospital will be compared with the total number of patients per hospital on the particular drug in order to create percentages of the refundable patients per hospital. Great differences in the total number of patients treated with a drug per hospital imply that some percentages are based on a great number of observations and can thus be stated to be robust while others are solely based on a few and could not be stated as robust. The spread of the percentage of refundable patients per hospital will be investigated by a histogram. To indicate the robustness of the drawn percentages, the number of patients underlying the percentage will be marked by colour. In order to observe the impact of the chosen threshold, the distribution of the percentage of re-fundable patients per hospital will be investigated for different choices of the threshold. This will be done by plotting the median, first quantile and third quantile of the distribution for different choices of the thresholds.
3.3
Risk sources
The analysis above provides an insight in the magnitude of the risk implied by a warranty risk-sharing agreement. However, the question remains if the source of the risk is mainly a result of the uncertainty about the real-time population or the practical application of the drug. In order to obtain an insight into the division of the risk over these two components, multiple regressions will be performed estimating a switch before threshold. The regressions will be built up, starting with a correction for person characteristics influencing the probability of a switch before the time thresh-old, a case-mix correction. Hereafter the model will be expanded with a dummy variables for each hospital to investigate whether after case-mix correction there is an impact of being under treatment in a particular hospital. Subsequently, a first attempt will be made with the establishment of the hospital characteristics and treatment processes that cause the potential differences in the impact of hospitals. All regressions predict a switch before threshold by a logistic model. Consider the set of all patients P in the data and the subsets Pd ⊂ P of all patients treated with drug d = 1, .., D.
patient with multiple treatment lines of the same drug d will be considered as if there are multiple patients with the same person characteristics.
The characteristics to be included will be based on the available data and the existing literature on important factors influencing the outcomes of drug.
Anderson et al. [2], Radovits et al. [18] and Weyend et al. [26] investigate the effectiveness of bDMARDs and prove that sex and age are significant predictors. However, the same literature corrects in their models for disease duration. This variable is not available in the dataset and will thus not be included. As a consequence, the impact of disease duration will partly be captured in the impact of age since older patients are likely to have a greater disease duration. This makes the linearity of age doubtful. Covering the possibility of this problem, age will be incorporated in the form of age groups with the groups based on the quantiles of the observed ages. Sex will be incorporated as a dummy being one for women and zero for men.
Another person characteristic that will be included for the case-mix correction is social economic class, of which the importance is shown by Anderson et al. [2]. A patient’s social economic class will be estimated by the average income of its zip code area.
Comorbidity is the presence of multiple health complications. Radner H.[17] stresses the importance of comorbidity in rheumatism on outcomes such as physical function, quality of life and mortal-ity. These factors are likely to influence a switch between drugs. Therefore I include a patient’s comorbidity into the case-mix correction. Radner H. [17] uses the Charlson score (Charlson et al. [7]) as a proxy for the severity of comorbidity. This score is composed by assigning weights to different health issues as well as different ages to predict the chance of one year survival. However, another broadly used measurement of comorbidity which for some cases is even proven to be a better proxy (Chang et al. [6]) is the Elixhauser score (Thompson et al. [22]). Similarly to the Charlson score, it is composed by assigning scores to different health issues but excludes the effect of age. By simultaneously including age and the Charlson score there would be great correlation between the variables in the case of absence of comorbidities. For these reasons, the Elixhauser score will be included in the model as the measurement for comorbidities. The correlation between all incorporated patient specific variables will be investigated.
Let there be, g = 1, ..., G age groups and social classes k = 1, ..., K. All models predict a patient’s switch before threshold denoted by the (N × 1)-vector sw with elements swi being one when a
patient i switched before threshold and zero otherwise. Let a patient’s age group be denoted by element Aigof the (N × G)-matrix A being one if patient i is in age group j and zero otherwise. Let
in the same manner, element Sikof (N ×K)-matrix S denote a patient’s social class and element wi
of the (N × 1)-vector w denote whether patient i is a woman. The (N × 1)-vector e with elements ei denotes a patient’s Elixhauser score. Then the first logit model is defined as follows:
swi= β0+ β1wi+ G X g=1 β2gAig+ β3ei+ K X k=1 β4kSik+ i, (1)
with the (N × 1)-vector of error terms.
z. Incorporating this variable in the regression model of equation 1 gives the logit model swi= β0+ β1wi+ G X g=1 β2gAig+ β3ei+ K X k=1 β4kSik+ Z X z=1 β5zHiz+ i. (2)
The differences in hospital coefficients are investigated by comparing all individual combinations of hospital dummy coefficients. The hypothesis of equivalence of the coefficients will be tested by use of an F-test and thus creates a Z × Z triangular matrix of all test values between hospital combinations.
A first attempt in quantifying the factors underlying the differences in impact of the hospital dummies will be made by replacing the hospital dummies with two variables capturing some hospital variation. Firstly, the impact of a hospital’s total number of treated patients on surgical outcomes has been proven by Flood et al. [10]. Its influence on medical outcomes is less clear (Flood et al. [10]), nevertheless the average number of treated patients per year with drug d in hospital z will be incorporated as an indicator of a hospital’s expertise and experience with a specific bDMARD. Consider herefore the (Z × D)-matrix N P of the total number of treated patients with drug d per hospital z. Secondly, a dummy will be included stating whether the issued drug was a patient’s first drug in order to determine whether the stage of a care process influences the probability of a switch before threshold. Whether the issued drug d was a patient i’s first drug issue will be denoted by the elements fi of the (N × 1)-vector f . Incorporating these variables by replacing the hospital
dummy in equation (2) gives the final logit model:
swi= β0+ β1wi+ J X g=1 β2gAig+ β3ei+ K X k=1 β4kSik+ Z X z=1 β5zHizN Pzd+ β6fi+ i. (3)
The predictive power of the models and for each drug will be investigated to be able to estimate their quality. The total number of observations per drug are split into a 70% training set (T r) and 30% test set (T e). Let swcj denote the probability of a switch before threshold of patients
j = 1, .., M ∈ T e on drug d based on the set of model coefficients estimated by the training set T r. The observed percentage of refundable patients (ORP) per hospital z will be compared to the predicted percentage of refundable patients (PRP). Consider q = 1, ..., Q training and test sets. Then ORP and PRP for a hospital z are calculated by
ORPz= PQ q=1 PM j=1(Hjzswj) PQ q=1 PM j=1Hjz , (4) PRPz= PQ q=1 PM j=1(Hjzswcj) PQ q=1 PM j=1Hjz . (5)
correctly predicted non-switchers before threshold (sensitivity) will be split. Consider the subsets y = 1, ..., M1∈ T e and l = 1, ..., M0∈ T e of all patients with and without switch before the
thresh-old respectively. Since switching before the threshthresh-old is a binary variable the comparison with the estimated swcj cannot be made directly. Accordingly a random number from a Bernoulli
distribu-tion with a probability of successswcj will be drawn for each patient where after the percentage of
correctly estimated outcomes will be determined. This will be done B times and over Q different training and test sets resulting in the following formulas for the specificity and sensitivity:
Specificity = 1 Q Q X q=1 1 B B X b=1 1 M1 M1 X y=1 1 − Bern.(swcy) − swy , (6) Sensitivity = 1 Q Q X q=1 1 B B X b=1 1 M0 M0 X l=1 1 − Bern.(swcl) − swl . (7)
4
Data
The analysis is based on data from 58 Dutch hospitals affiliated with LOGEX B.V., a company that performs data analysis to improve health care outcomes and performances. This study is performed to improve LOGEX analysis and tools.
The registration of health care activities is required by the Dutch healthcare system to be bundled into so called DBC-DOT Zorgproducten for reimbursement. Within this reimbursement structure, the issues of expensive drugs have to be registered separately as an add-on. This implies that they are, in contrast to other drug issues, observable in claim registries. The data being used comprises these claim registries about patients’ expensive drug issues. Confidentiality of the data causes that in this study there will be referred to drug A to E rather than specific drugs.
This study focuses on rheumatism and therefore only expensive drug issues within the diagnosis group ”Rheumatoid Arthritis and related diseases” were selected for the analysis. Besides RA, this diagnosis group as well comprises other arthritis-like diagnoses such as Juvenal Arthritis and Spondyloarthritis. The selection based on this diagnosis group also takes into account all drug issues for RA related diseases. Keeping the focus solely on RA specific drugs, all drugs were investigated manually and all drug issues that could not be categorized as JAK inhibitors, DMARDs, bDMARDs or NSAIDs were excluded.
For some of the drug issues, the total dose was observed to be negative. This kind of drug issues are probably used to make corrections on previous registered drug issues. To account for this, the total dose of exactly identical drug issues were summed per patient and date. Drug issues that after this modification still showed a zero or negative total dose were considered as data errors and excluded from the set. Finally the data consists of 58 hospitals, 32,995 unique patients and in total 455,364 drug issues with a median of 11 drug issues per patient.
The observable time periods of patients differ as well. Only 8.6% of all patients show to have expensive drug issues over the entire time span of their hospital, measured as a drug issue in the first quarter of their hospital’s first observation and a drug issue in the last quarter of their hospital’s last observation. The average observable time span of patients’ expensive drug issues in the data is 900 days with a maximum of 2845 days and a minimum of one single drug issue and thus 1 day.
Table 1: Summary statistics of the baseline data
Hospitals mean st. dev. min max Number of hospitals
Number of treated patients 569 516 2 3260 58
Observable time span 1495 days 531 days 342 days 2894 days 58
Patients mean st. dev. min max Number of patients
Age 56.6 15.6 0 99 32988
Social class 2.0 0.80 1 3 32725
Elixhauser score 1.22 0.68 0 10 32988
Number of drug issues 13.3 11.7 1 184 32995
Observable time span 900 days 614 days 1 day 2845 days 32995 Percentage of total patients
Men 35.2 % 11622
Women 64.7 % 21363
Table 2: Number of treatment lines per patient
Number of treatment lines Number of Patients Percentage patients of total
1 25726 77.31% 2 4578 13.76% 3 1935 5.82% 4 670 2.01% 5 247 0.74% 6 70 0.21% 7 33 0.10% 8 9 0.03% 9 6 0.02% 10 1 0.00% 11 1 0.00% 12 1 0.00% 14 1 0.00%
of 10. Since RA is also included in the Elixhauser score this implies that most of the patients in the data do not have another classified disease. An Elixhauser score of 0 could occur for patients with different forms of Arthritis than RA.
The Elixhauser score is calculated by LOGEX based on the diseases as observed from the data so that diseases that have been cured in a different hospital or have been cured outside a hospital’s observable time span are not taken into account. The Elixhauser score as used in this study is thus just a proxy of a patient’s actual comorbidity. The social classes are constructed by LOGEX as well and based on patients’ zip-codes. These are linked to the average income as established by ”The Netherlands Institute for Social Research”. Three classes were constructed based on the average income’s tertiles. The average social class observed in the data is the middle class.
Age groups were composed by the quantiles of the observed ages resulting in 4 age groups: younger than 50, between 50 and 60, between 60 and 70 and older than 70. Patients are followed over time which creates the possibility that a patient falls within multiple age groups while being on the same drug. The regression analysis requires one age group per treatment line so that a transformation was needed. Therefore, the median age group was taken for each patient’s treatment line. The same situation occurs for a patient’s social class which is based on a patient’s zip-code and the Elixhauser score. These variables were thus transformed in the same manner as age group. Looking at the overall switch behaviour, there are 7585 out of 33311 (22.77%) patients with at least one switch. There turned out to be 33 patients with multiple switches on one day. Inspecting the particular patients, it was observed that they were constantly issued the same combination of drugs which gives rise to the interpretation that they involved in some sort of combination therapy. Keeping the analysis simple and straightforward, these patients were dropped. I thus implicitly assumed that patients can only involve in a mono therapy of expensive drugs. The percentage of patients and their number of treatment lines can be found in Table 2. Over 77% of all patients have only one treatment line and were thus issued the same drug over their entire observable time span. The maximum number of treatment lines is 14. The correctness of the established treatment lines was checked by randomly selecting patients with over 5 treatment lines and investigating their drug issues. Only the altering drug issues of the patient with 14 treatment lines seem to could have been caused by combination therapy rather than switches, but this could not be stated with certainty.
5
Results
The set-up of this section will follow the method section and will thus consist out of three parts. As explained in the method, every subsection’s analysis is based on a different selection of the baseline data. Therefore, every analysis will start with a brief description of the data being used.
5.1
Switchers
Table 3: Characteristics of the subsets taken from the baseline data being used in section 5.1 and 5.2. Percentages are taken with respect to the baseline data.
Subset section 5.1 Subset section 5.2
Drug Number Percentage (%) Number Percentage (%)
A Drug issues 20055 13.23 149052 98.20 Patients 3244 22.00 13668 91.18 Treatment lines 3319 22.05 13928 91.14 B Drug issues 13193 12.06 105161 96.60 Patients 2469 19.78 10768 85.98 Treatment lines 2533 19.93 10952 85.96 C Drug issues 6174 18.35 31459 94.44 Patients 944 33.00 2366 81.87 Treatment lines 958 32.74 2418 82.19 D Drug issues 3528 11.00 31487 98.40 Patients 418 21.82 1762 91.15 treatment lines 427 21.64 1814 91.43 E Drug issues 2442 16.21 14656 97.00 Patients 484 27.50 1563 87.86 Treatment lines 492 27.35 1599 88.00
of the hospitals the share of patients is rather similar in both sets. Investigation of the Elixhauser score, age, sex and social class in the subset showed that there was little to no difference in the distribution between the baseline data and the subset.
The histograms of the number of days a patient has been treated with a drug before switching to a subsequent one is shown in figure 2. The evaluation periods as stated by the FK have been drawn in the figure by a vertical line. The frequencies of the histograms are normalized over the number of observations to make comparisons between drugs possible.
As expected, all histograms show a small number of switchers around zero days. Furthermore, the distribution of each histogram was expected to show a drop in the number of switchers just after the evaluation period plus some time correction for possible situations such as where an appointment is postponed due to circumstances. As can be observed in figure 2, most of the drugs (A, B and E), show an increase in the number of switchers up until the evaluation period and a gradual decrease of the number of switchers after. However, drug D shows an almost constant number of switchers over the first 300 days. Drug C shows a peek of switching patients around 90 days implying that a lot of patients do not proceed treatment with the drug until the evaluation period as proposed by the FK, but make a prior switch. The distribution of the treatment period before a switch of drug D shows a drop in the normalized frequency from 0.0025 to 0.0015 just after the evaluation period and the number of switchers remains rather constant after that. The histograms of the other drugs do not show a clear drop in the number of switchers right after the evaluation period.
5.2
Risk magnitude
For this subsection’s analysis the definition of unobservable start and end dates of a treatment line could be relaxed. All treatment lines that are observable for a time span greater than the threshold could be included. For drug A-E, the number of drug issues, patients and treatment lines in the data subset resulting from this selection process can be found in table 3. The subset of the data for this part of the research for each drug comprises over 90% of its drug issues and over 80% of each drug’s patients and treatment lines compared to the baseline data.
The histograms of figure 3 show the percentage of the refundable patients per hospital, with the size of the hospital indicated by colour. The axes of the histograms differ per drug since the differences among drugs are so large that identical axes would have resulted in uninterpretable figures. The histograms show that for each drug there are hospitals with approximately zero refundable patients. However, the majority of these hospitals are categorized in the groups with the smallest amount of total patients. For these hospitals, an almost zero percent of refundable patients is thus not surprising. The histograms of drug A, D and E show that there are also hospitals in the highest number of total patients category with almost zero percentage refundable patients. Especially remarkable is the case of drug A where the highest total number of patients class comprises hospitals with more than 269 patients. The right-hand side outliers, i.e. hospitals with a great percentage of refundable patients, are most commonly caused by hospitals in the groups with the least number of total patients. The outliers of drug C even go up to almost 100 percent. Investigating the 2 right-hand side outliers in this drug’s histogram, the 100 percent refund rate was caused by a hospital with only one patient and the 66 percent refund rate by a hospital with 3 patients. Hence, these outliers are not likely to be based on robust information. For drug D however, both outliers are caused by hospitals with a great amount of patients. Figure 4 shows the sensitivity of the percentage of refundable patients per hospital to different thresholds. The first quantile, median and third quantile of the percentage of the refundable patients per hospital, as determined for drug A to E, are plotted for different thresholds. In all plots it can be observed that the median is increasing in the threshold which is as expected. After all, when the threshold increases the number of patients that switch before the threshold is very likely to increase as well. Also the spread between the first and third quantile increases in the threshold. Implying that the differences in the percentages of refundable patients increase as the threshold is expanded.
5.3
Risk sources
The regression analysis is based on the same subset as the one described in subsection 5.2. Hence, the inclusion of treatment lines depends on the observable time period and the threshold per drug. The only difference lies in the exclusion of patients with unknown social class, age, sex, or Elixhauser score. The inclusion of these patients for previous analysis was harmless since the information was not being used.
Presenting the results of this subsection’s analysis is limited to drug A and B for space consider-ations and the selection of these two particular drugs is based on the magnitude of their subsets. Correlation matrices and VIF scores of all patient specific variables for drug A and B are shown in table 4. The correlations and VIF scores are similar for drug A and B. The largest correlation emerges between a patient’s Elixhauser score and age group, 0.179 for drug A and 0.159 for drug B. Age group also has the greatest VIF score of 1.042 for drug A and 1.037 for drug B.
Figure 4: Sensitivity of the percentage of refundable patients per hospital to different time thresholds. The 25%, 50% and 75% quantiles of the total distribution of the percentages of refundable patients per hospital are shown on the y-axis for different thresholds, displayed on the x-axis.
for a one unit increase in the independent variable.
Table 4: Correlation matrices for drug A and B of the patient specific variables included in the regression analysis. VIFs shown in the last column representing the extent to which the particular variable could be explained by the others.
Correlation matrix Drug A
Sex Age group Elix. Score Social Class First drug VIF
Sex 1.000 1.003
Age group 0.019 1.000 1.042
Elix. Score -0.025 0.179 1.000 1.037
Social Class -0.009 0.033 0.059 1.000 1.004
First drug 0.032 -0.087 -0.004 0.005 1.000 1.009
Correlation matrix Drug B
Sex Age group Elix. Score Social Class First drug VIF
Sex 1.000 1.005
Age group 0.035 1.000 1.037
Elix. Score -0.024 0.159 1.000 1.023
Social Class -0.005 0.022 0.057 1.000 1.003
First drug 0.048 -0.092 -0.015 0.008 1.000 1.011
0.18, the impact thus increases.
For the investigation of the predictive power, 500 different training and test sets were drawn (Q=500) from the data subset. Subsequently, for each test set 100 random values have been drawn from a Bernoulli distribution (B=100) with probability dSWid to compare observed and predicted
switches/ non-switches before threshold. The sensitivity and specificity for each model can be found in table 5. Under the first model for drug A, only 4.5 percent of all switches before threshold could be well predicted. For drug B this is 4.4 percent. The no information rate describes the percentage of accurately predicted values when a prediction was always set as not switching before the thresh-old and shows that the data contains a great unbalanced amount of switchers and non-switchers before threshold.
Figure 5 gives a visual representation of the predictive power for the percentage of refundable pa-tients per hospital. As can be observed in the top 2 histograms, for both drugs the model with case-mix correction predicts on average 4% refundable patients per hospital with a very small variance. The spread of the percentage refundable patients between hospitals can thus not be ex-plained by the type of patients (as described by the incorporated person characteristics) that are under treatment in the particular hospital.
The second model adds hospital dummies to the model, being 1 if the patient was treated in the particular hospital. The coefficients of this model are shown in the second and fifth column of table 5. For drug A, all case-mix coefficients decrease in impact compared to the model without hospital dummies. For drug B, however the impact of sex, age and the highest social class increase in impact compared to the model without hospital dummies.
Table 5: Logit regression output, with coefficients representing the change in log odds of a patient switching before the threshold and standard errors shown in parenthesis
Drug A B
Model 1 Model 2 Model 3 Model 1 Model 2 Model 3
(Intercept) −3.37∗∗∗ −3.32∗∗∗ −4.54∗∗∗ −3.66∗∗∗ −3.13∗∗∗ −4.23∗∗∗ (0.14) (0.37) (0.19) (0.15) (0.34) (0.18) Woman 0.44∗∗∗ 0.41∗∗∗ 0.35∗∗∗ 0.38∗∗∗ 0.38∗∗∗ 0.29∗∗ (0.10) (0.10) (0.10) (0.11) (0.11) (0.11) Age 50-60 −0.19 −0.19 −0.20 0.05 0.08 0.15 (0.11) (0.11) (0.11) (0.12) (0.12) (0.12) Age 60-70 −0.44∗∗∗ −0.42∗∗∗ −0.27∗ −0.32∗ −0.31∗ −0.17 (0.12) (0.12) (0.12) (0.14) (0.14) (0.14) Age > 70 −0.78∗∗∗ −0.73∗∗∗ −0.48∗∗∗ −0.58∗∗∗ −0.60∗∗∗ −0.34∗ (0.14) (0.14) (0.15) (0.17) (0.17) (0.17) Elixhauser score 0.10 0.04 0.10 0.23∗∗ 0.22∗∗ 0.25∗∗∗ (0.07) (0.07) (0.07) (0.07) (0.08) (0.08)
Social class middle 0.14 0.12 0.15 0.14 0.12 0.16
(0.11) (0.11) (0.11) (0.12) (0.12) (0.12)
Social class high 0.11 0.11 0.11 0.18 0.23 0.20
(0.11) (0.12) (0.12) (0.12) (0.13) (0.13)
Average number of treated patients per year 0.00 0.00
out of 1652 test were significant under the 5% level, which is 6.0 percent of all tests. Investigating the allocation of these significant tests showed that most of the significant tests were caused by four hospitals. The particular hospitals were examined, but based on the available data I could not establish the link between the four.
Among the hospitals that issue drug B there was one hospital with only two observable patients. When splitting the data into training and test sets, this resulted in training sets with no observable patients for the particular hospital. As a consequence the particular hospital dummy coefficient could not be estimated and the prediction of the test set could not be made. Therefore this hospital was excluded for the analysis of the predictive power of model 2, drug B. The specificity of model 2 for drug A is 5.3 percent and 4.8 percent for drug B. Both show an increase in the individual predictive power compared to the first model. The hospital dummy coefficients are based on the switch behaviour in the particular hospital. These coefficients thus describe the differences among hospitals in the number of patients switching before the threshold after case-mix correction. On patient level these coefficients show to have little predictive power by the specificity. However, collecting patients’ predictions again over the hospitals shows an almost perfect match as can be observed by the histograms on the second row of figure 5. This shows that the percentage of refund-able patients is almost entirely determined by the hospital dummy and stresses the little predictive power of the case-mix variables.
Model 1 and 2 indicate that the observed spread in the percentage of refundable patients per hospi-tal cannot be explained by a case-mix correction. As first step in the investigation into other factors that cause the spread the final model replaces the hospital dummies with a variable indicating the average number of patients treated with the specific drug in a hospital per year and a dummy stating whether the drug is issued as a patient’s first drug. The results of this logit regression can be found again in table 5, model 3. The implementation of these two variables causes the intercept of drug A to decrease from -3.37 in model 1 to -4.54 in model 3. For drug B this is a decrease from -3.66 to -4.23. For both drugs the negative impact on the log odds of switching before the threshold of being a woman compared to being a man declines with respect to the first model. The impact of age, as captured in the age groups, decreases for both drugs relative to the first model. Patients older than 70 now have a log odds of switching before the threshold of -0.48 for drug A and -0.34 for drug B compared to the patients younger than 50. Whether the drug under consideration was issued as a first drug shows to have an impact of 2.38 and 1.74 for drug A and B respectively, both showed to be significant with a p-value smaller than 0.001. This indicates that patients that receive the drug as a first drug have a greater probability of switching before the threshold than patients that receive the drug as a second, third etc. The average number of patients per year that is treated with the drug in a particular hospital shows to have no impact on the probability of switching before the threshold for neither of the drugs.
6
Conclusion
This research explored the magnitude and sources of risk for pharmaceutical companies engaging in a warranty risk-sharing agreement for bDMARDs. A pharmaceutical company was stated to refund drug expenses whenever a patient switched drugs within a certain time period after its first issue. The risk was defined as the uncertainty regarding the percentage of patients a pharmaceutical company should refund and was determined on the hospital level.
The investigation of patients’ treatment periods with bDMARD A-E until switching showed no evidence that the evaluation periods, as proposed by the FK, are being practiced. The histograms raise the question if there even exists a hard evaluation period which underlines the existing lit-erature [12] arguing the difficulties in the determination of the level of response and the set-up of risk-sharing agreements.
Comparing hospital’s refundable patients as a percentage of the total number of patients treated with the drug showed that there are great differences among hospitals. The range of the percent-age refundable patients was 0-11% for drug A and B, and even go up to 0-50% for drug C (only considering middle size hospitals). This implies that there is a spread in the percentage refundable patients on the hospital level, which differs per drug as well.
The regression models provide solely an indication of factors influencing the probability of switching before threshold and should be further investigated and expanded on before any stronger conclu-sion can be drawn. Nevertheless the case-mix correction shows that the spread in the percentage refundable patients per hospital cannot be explained by the person characteristics under consid-eration. The predictive power of the regression models including hospital dummies and hospital processes stress this further, showing that there are other factors captured within the hospitals that explain the spread. Comparison of the three regression models for drug A and B shows that there are some differences in the variables influencing the probability of a switch before the threshold. The differences in the variable coefficients for drug A and B show that identifying which factors influence switching before threshold is drug specific and could not simply be determined for the complete diagnose group.
This study shows that for RA drugs the risk as defined on the hospital level, differs per drug with a minimal range of 0-11% and a maximum of 0-50%. The spread can not be explained by basic case-mix characteristics implying that there are more factors influencing the risk than solely population selection uncertainty. Both pharmaceutical companies and payers should take this into account when exploring the opportunities of engaging in a risk-sharing agreement.
7
Discussion
7.1
Limitations
7.1.1 Data
The selected diagnosis group comprises RA patients as well as patients with other Arthritis type disorders. There has been no distinguishing between the different types of Arthritis although treatment processes could differ. Thereby, only expensive drug issues are observed from the data. This means that a patient switching from an expensive drug to a non-expensive drug is not observed. Patients changing hospitals could not be observed in the data either, resulting in a bias in the interpretation of patients starting with a first drug. This bias is further enlarged by the relative small observable time span of the data for a chronic disease like RA.
Over the years the competition on the drug market has grown, resulting in price decreases of particular drugs. This could have had an impact on the switch behaviour implying for example an earlier switch to new and cheaper drugs. However, in this study this was not taken into account and the considerations of switching drugs were assumed to be equivalent over time. The observations of the hospitals and patients were thus assumed to be independent of the time period.
7.1.2 Method and results
This study uses a switch between rheumatism drugs within a particular time period from the first drug issue as an estimate of the response to a drug. A different more objective measure would have been the change in DAS28 score. This score evaluates how many of 28 joints are painful and/or swollen and is the most common measure of disease activity in rheumatism (Gardiner et al. [11]). This data is however limitedly available and does not take into account a patient’s experience with the drug, the burden of possible side-effects and the relativity of the change in disease activity for each individual patient.
In the definition of switching drugs, biologicals and biosimilars have not been distinguished although there is an ongoing debate about the effectiveness of biosimilars (D.H. Yoo et al. [28]). Also, the assumption is made that patients can only involve in mono-therapy although in practice combination therapies do occur. In order to obtain a better insight in the risks of risk-sharing agreements this would have to be updated.
For the selection of the data subset being used for the histogram of the treatment period with a drug before switch, patients with an unobservable start date have been defined as patients with a drug issue in the same quarter as the first observable drug issue of their hospital. The average time between drug issues has not been investigated. Although this could have influenced the marking of treatment lines, I assumed the impact to be small.
The selection process for the analysis of the treatment period before switch resulted in a subset of 19 percent of the baseline data’s patients. This implies that especially for drugs that are less prescribed the number of treatment lines selected from the baseline data is only small. However, the results still give an insight in patients’ treatment periods. The selection process also might have caused bias in the sense that overall, mainly patients with a relatively great number of drug switches are incorporated. The fact that a patient is already on a high treatment line could imply that the patient is sensitive for side effects or has difficulties with rheumatism drugs in general and is thus more likely to have an early switch to different drugs. Both issues could be solved by expanding the data in order to be able to follow patients over a greater time span.
as a recommendation rather than a decision rule. This would imply that a threshold as proposed in this research is not suitable.
The sensitivity analysis of the distribution of the percentage refundable patients per hospital shows that the distribution is very sensitive to the choice of the threshold. The threshold used for part 2 and 3 of the analysis are based on the investigation of the treatment period before switch histograms of section 5.1 including some mark-up. The appropriate size of this mark-up could be further investigated by inspecting the lead time of the evaluation consult.
The Elixhauser score indicates the complications resulting from other diseases than RA. A variable describing the RA severity was not included in the model since such indications were not available. Ting et al. [23] propose a severity scoring system based on claim registries. However, for the approximation of the RA severity by this scoring method in a way that it is usable for the analyses it should be based on a patient’s claim information one year before starting treatment with a particular drug. The available data only provides a little number of patients who can be followed over a great time span and for who this RA severity approximation could be made. Therefore, for this research the RA severity was assumed to be partly captured by age.
The third regression model incorporates a variable stating whether the issued drug was a patient’s first. First of all, since only expensive drugs are observable from the data non-expensive drugs were not taken into account in the establishing of a patient’s first drug issue. Also, the drug issue was only marked as a patient’s first whenever the start date of the treatment period with the particular drug was certain and it was certain that there had not been a previous drug issues. The lack of information implies that there is likely to be a bias in this variable in the form of an underestimation of the number of patients on their first drug.
For both drugs the no information rate is 95.9% implying that only 4.1% of the patients show a switch before threshold. This unbalance in the number of switches and non-switches before threshold implies that the data only contains little information about switches before threshold. As a consequence the regression model estimations may not be fully capable to describe patterns in patients switching before the threshold.
7.2
Recommendations
The model used in this research is a first step and more research would be needed to further specify the risks of risk-sharing agreements. Nevertheless, the analyses performed in this research provide some quantification of the risk implied by a warranty risk-sharing agreement for RA drugs. This quantification contributes to the understanding of implications of risk-sharing agreements with the potential of broader usage of this type of agreements and reducing health care costs of expensive drugs.
The analyses could be improved by further development of the used models. Person characteristics could be added in the regression analysis. Examples would be a patient’s disease duration, patient’s RA severity, or the number and type of a patient’s previous prescribed drugs. The investigation of risk in risk-sharing agreements for RA in this study is based on non-response measured as a patient switching drugs before a certain threshold. An interesting study would be to compare the risk based on this measure to the risk resulting from the response as measured by the DAS28 score. The analysis could also be performed on a data set following patients over a greater time span. This would provide better insight in RA treatment processes and increase the information contained in the data for the regression analysis resulting in more robust estimates.
other diagnose groups than RA. This would provide an insight into differences between diagnoses which could establish to which diagnoses this risk-sharing agreement has potential. Finally, the investigation of risk should be expanded to more types of risk-sharing agreements to create insight into the differences in implications.
Acknowledgements
During the process of writing this thesis I have received a lot of guidance and support. First of all, I would like to thank my supervisor Dr. C. Praagman. I am grateful for his advice and suggestions, as well as our many discussions about theoretical and non-theoretical topics.
I would also like to express my gratitude to LOGEX for giving me the opportunity of writing a thesis with them. It was very motivational and educational to write a thesis with practical implications and within a business setting. I would also like to thank my colleagues for their time, support and great encouragement over the past 6 months. In particular I would like to thank Jan van der Eijk for helping with the development of an interesting thesis subject and challenging my procedures. Hans Teunisse for his support on academic writing and data analyses. And finally Femke Oldenziel for all her guidance and critical notes throughout the whole process, keeping me on schedule and not letting me drift off to new ideas.
References
[1] Jakub Adamski, Brian Godman, Gabriella Ofierska-Sujkowska, Bogusawa Osiska, Harald Her-holz, Kamila Wendykowska, Ott Laius, Saira Jan, Catherine Sermet, Corrine Zara, Marija Kalaba, Roland Gustafsson, Kristina Garuolien`e, Alan Haycox, Silvio Garattini, and Lars L. Gustafsson. Risk sharing arrangements for pharmaceuticals: Potential considerations and rec-ommendations for European payers. BMC Health Services Research, 10, 2010.
[2] Maria L E Andersson, Stefan Bergman, and Maria K S¨oderlin. The Effect of Socioeconomic Class and Immigrant Status on Disease Activity in Rheumatoid Arthritis: Data from BAR-FOT, a Multi-Centre Study of Early RA. The Open Rheumatology Journal, (7):105–111, 2013. [3] Fernando Anto˜nanzas, Roberto Rodr´ıguez-Ibeas, and Carmelo A. Ju´arez-Castell´o. Personalized Medicine and Pay for Performance: Should Pharmaceutical Firms be Fully Penalized when Treatment Fails? PharmacoEconomics, 36(7):733–743, 7 2018.
[4] Pedro Pita Barros. The simple economics of risk-sharing agreements between the NHS and the pharmaceutical industry. Health Economics, 20(4):461–470, 4 2011.
[5] A. Carletto, A. Cicchetti, S. Coretti, V. Moramarco, and M. Ruggeri. Money back guarantee? A cost–benefit framework of performance-based agreements (PBAs) for the reimbursement of pharmaceuticals. Eurasian Business Review, 9(1):89–101, 3 2019.
[7] Mary E Charlson, Peter Pompei, Kathy L Ales, and C Ronald Mackenzie. A new method of classifying prognostic comorbidity in longitudinal studies: Development and validation. Journal of Chronic Diseases, 40(5):373–383, 1987.
[8] Jaime Esp´ın, Joan Rovira, and Leticia Garc´ıa. Experiences and Impact of European Risk-Sharing Schemes Focusing on Oncology Medicines. Technical report, 2014.
[9] Farmacotherapeutisch Kompas. Reumato¨ıde artritis, 2020.
[10] Ann Barry Flood, W Richard Scott, and Wayne Ewy. Does Practice Make Perfect? Part I: The Relation between Hospital Volume and Outcomes for Selected Diagnostic Categories. Medical Care, 22(2):98–114, 1984.
[11] P V Gardiner, A L Bell, A J Taggart, G Wright, F Kee, A Smyth, R McKane, J Lee, M E Rooney, and E Whitehead. A potential pitfall in the use of the Disease Activity Score (DAS28) as the main response criterion in treatment guidelines for patients with rheumatoid arthritis. Annals of the Rheumatic Diseases, 64(3):506, 3 2005.
[12] Francisco R. Gon¸calves, Susana Santos, Catarina Silva, and Gabriela Sousa. Risk-sharing agreements, present and future. Ecancer Medical Science, 12, 4 2018.
[13] Kees Wessels and Ingrid Doude van Troostwijk. Zo werkt de geneesmiddelenzorg. De Argu-mentenfabriek, 2019.
[14] Kennis- en Exploitatiecentrum Offici¨ele Overheidspublicaties. Geneesmiddelen in de medisch-specialistische zorg. Technical report, Nederlandse Zorgautoriteit, 2019.
[15] Peter J. Neumann, James D. Chambers, Fran¸coise Simon, and Lisa M. Meckley. Risk-Saring Arrangements That Link Payment For Drugs To Health Outcomes Are Proving Hard To Im-plement. Health Affairs, 30(12):2329–2337, 2011.
[16] Trevor Jozef Piatkiewicz, Janine Marie Traulsen, and Tove Holm-Larsen. Risk-Sharing Agree-ments in the EU: A Systematic Review of Major Trends. PharmacoEconomics - Open, 2(2):109– 123, 6 2018.
[17] Helga Radner. Multimorbidity in rheumatic conditions. The Central European Journal of Medicine, 128(21-22):786–790, 11 2016.
[18] B. J. Radovits, W. Kievit, J. Fransen, M. A.F.J. Van De Laar, T. L. Jansen, P. L.C.M. Van Riel, and R. F.J.M. Laan. Influence of age on the outcome of antitumour necrosis factor alpha therapy in rheumatoid arthritis. Annals of the Rheumatic Diseases, 68(9):1470–1473, 9 2009.
[19] Rijksinstituut voor Volksgezondheid and en Milieu. Zorguitgaven. Hoe ontwikkelen zich de zorguitgaven in de toekomst? Technical report, 2018.
[20] MG Saggia. Risk Sharing Agreements (RSA): Exploratory Analysis of the Current Situation in Developed and Emerging Markets. Value in Health, 2016.
[22] Nicolas R. Thompson, Youran Fan, Jarrod E. Dalton, Lara Jehi, Benjamin P. Rosenbaum, Sumeet Vadera, and Sandra D. Griffith. A new elixhauser-based comorbidity summary measure to predict in-hospital mortality. Medical Care, 53(4):374–379, 3 2015.
[23] Gladys Ting, Sebastian Schneeweiss, Richard Scranton, Jeffrey N. Katz, Michael E. Weinblatt, Melissa Young, Jerry Avorn, and Daniel H. Solomon. Development of a health care utilisation data-based index for rheumatoid arthritis severity: A preliminary study. Arthritis Research and Therapy, 10(4), 8 2008.
[24] Adrian Towse and Louis P Garrison. Can’t Get No Satisfaction? Will Pay for Performance Help? Toward an Economic Framework for Understanding Performance-Based Risk-Sharing Agreements for Innovative Medical Products. Pharmacoeconomics, 28(2), 2010.
[25] Adrian Towse, Louis Garrison, and Ruth Puig-Peir´o. The Use of Pay-for-Performance for Drugs: Can It Improve Incentives for Innovation? Technical report, Office of Health Economics, 2012.
[26] Cornelia M Weyand, Dorle Schmidt, Ulf Wagner, Jorg J Goronzy, Cornelia M Wcyand, J Goronzy, Mayo Clinic, and Minnesota Rochester. The influence of sex on the phenotype of rheumatoid arthritis. Arthritis & Rheumatism, 41(5):817–822, 1998.
[27] John Wlodarczyk, Christopher M. Reid, and Glen Pater. Funding linked to ongoing research: Impact of the Bosentan patient registry on pricing in Australia. Value in Health, 14(6):961–963, 9 2011.
[28] Dae Hyun Yoo, Jung Yoon Choe, Seung Cheol Shim, and Chang Hee Suh. Switching to biosimilars in the treatment of rheumatic diseases. Expert Review of Clinical Immunology, 14(7):557–571, 7 2018.
