The time and price effect on the probability of the choosing a generic versus branded drug version

(1)

Faculty of Economics and Business

The time and price effect on the probability of choosing

a generic versus a branded drug version

Bachelor Thesis

Name: Denitsa Danailova

Student number: 6087388

(2)

2

Abstract

This paper investigates the different factors influencing the probability that generic medicines are chosen over branded versions of the same medicine. The data sample is collected from a doctor‟s practice in Bulgaria over the period of ten years and the study is focused on the particular drug class of statins, widely-used bad cholesterol diminishing medications. Results show, consistent with some of the previous findings, that there are two main aspects to consider when predicting the probability of choosing a generic- a financial incentive and a time adjustment period. Generics are less probable to be preferred as the relative generic-to-branded drug price increases. In contrast, the more time passes since a patent on a branded drug expires, the higher the probability to choose a generic. The probability varies between the different statin classes (simva-, atorva-, and rosuva-statin) and implies drug individual characteristics. Patient characteristics, such as age, although found significant previously, prove to be insignificant in this particular study. Implications from the study can have large potential benefits for public authorities aiming to deliver affordable quality medication to its citizens. Further research on a larger cross-country scale including more drug groups for the same time period should be conducted to draw more robust conclusions about generic preference patterns.

Keywords:

generic drugs, medical cost reductions, relative drug prices, patent expiry

(3)

3

The pharmaceutical industry is one of the most profit-generating industries, one of the few resilient to economic fluctuations. The fair inelastic demand for pharmaceuticals allowed it to be only marginally affected by the financial crisis of the last decade and with remaining healthy profits as the report of the former Booz&co (2009), now Strategy&, shows. Even if the overall prospects for the sector are positive, however, countries hit by the crisis tend to initiate medical budget cuts. These cuts, Behner, Vallerien, Ehrhardt, and Rollman argue, will result in pressure on drug prices and policies stimulating increase in the use of generics.

A distinction exists between drugs manufactured by the so-called Big Pharma, branded

medicines and their cheap substitutes, generic medicines. The Hatch-Waxman Act, known as the Drug Price Competition and Patent Term Restoration Act from 1984 defines the drug approval and patent protection regulations for both generic and branded drugs. On the one hand the Food and Drug

Administration ensures generic bioequivalence upon generic approval. Generics do not have to undergo clinical trials to be granted an approval as long as they “perform in the same manner as branded drugs” and “deliver the same amount of an active ingredient to a patient‟s bloodstream in the same amount of time as the innovator drug”, in other words they have to be bioequivalent to the branded drug. On the other hand, generics cost up to 85% less than an original drug. The discrepancy in price between generics and branded medicines with the same performance is explained by the large R&D, clinical trials, and marketing costs incurred by the innovator Big Pharma companies.

One of the main characteristics of the pharmaceutical industry is that it is heavily R&D based. Each year billions are given worldwide for further research and development of new more effective drugs, according to Statista (2016) in 2014 alone 142billion U.S. dollars, or 14.4% of total industry revenues for the year, were invested. Patents are used to protect drugs that take years and millions to be developed, therefore restricting market entry and keeping competition low and prices high for as long as possible (United Nations, 2015). Patents protect the intellectual property of a Big Pharma company so that competing companies, and above all other generic companies, do not replicate the unique chemical compound for a period of a couple years. In this way, the developer company benefits exclusivity on the market and can reap higher returns, in part to cover the initial investment in R&D (Nelson, Slordal, and Spigset, 2009). Generics, in contrast, once the patent has fallen, can afford to sell the same medical compound at a substantially lower price since they have only production but no

(5)

5 development costs.

Because of their bioequivalence and affordability, generics can provide the solution for accessible quality medication under the pressure of ageing population, raising pharmaceutical expenditures, and budget constraints. Therefore, Recommendation 4 from G10 High Level Group of Innovation and Provision of Medicines in the EU from back in 2002 proposes that the key towards improving the healthcare provision in the EU is “to secure the development of a competitive generic market, generic prescribing and dispensing “. Since adoption of generics can benefit public well-being it is important to understand the factors facilitating the adoption. The consumer preference, expressed in a patient-doctor prescription process, shapes the demand for both generic and branded medicines. A higher preference for generics intuitively facilitates higher adoption. It is of interest if the two versions of a drug are equally effective whether the generic one will be preferred and which factors will change the preference.

In a study conducted on the medical group of drugs cephalosporins, Ellison, Cockburn, Griliches and Hausman (1997) develop a multi-stage demand decision model where first the doctor chooses the right treatment and then on a second-stage the doctor and the patient choose a generic or branded version of the needed active ingredient. By definition generics have the same or very close chemical compounds to the branded ones and can be substituted and used for the same conditions. The results of this study are that indeed there is high cross-price elasticity between chemically identical substances as the scientists argue that price of the drug and its alternative is the main determinant when preferring a generic over a trade-name version. In contrast, Pechlivanoglou and van der Veen (2011) conduct a study for the Netherlands, in which they consider price to be irrelevant because of the local full reimbursement medical framework. Instead, they prove other factors, such as patient characteristics and time passed since patent expiry, important. Iizuka (2012) introduces a new angle to the decision making process, where the price and time are still valid predictors but where their magnitude depends on the physician who is a non-perfect agent for the patient because he might be led by financial incentives when prescribing a particular drug. Lundin (2000) has already raised the moral hazard problem in drug prescription as the consumer is not the one who is making the ultimate decision. He shows that the doctor who chooses instead considers the utilities of both the patient and the insurer and once again price seems to be not a deciding factor, unlike a switch in the reimbursement policy.

In this paper, factors influencing the likelihood that a generic is chosen over a branded drug are investigated, in particular the time and price effect. Hence, the research question addressed is:

(6)

6

What is the effect of price difference and time after expiry on the likelihood of choosing a generic over a branded drug?

Similarly to prior research (e.g., Iizuka, 2012 and Lundin, 2000) actual prescription data is collected. The study focuses on Bulgaria and the therapeutic class of statins and aims to give a more thorough understanding of the decision process of drug prescription in a country where such a study has not been conducted yet. A total of 4,471 observations, over a ten year period from 2005 to 2014, from a physician practice in Sofia are divided into three groups, each containing a different active ingredient.

Here are several important reasons why the group of statins (a group diminishing the level of bad cholesterol, which is one of the reasons for high blood pressure and cardiovascular diseases) is highly suitable for analysis. First, this is a widely prescribed group of drugs which constitutes a large segment of the pharmaceutical market in most countries, and as such there are large sets of data available. Second, it is possible to analyze therapeutic substitution in this setting due to similarities among drugs. This means that differences, if any, will not be influenced by effectiveness of the drug but by other factors, such as price, image, etc. Third and very important, a lot of the branded statins have lost their patent protection which allows us to study generic substitution. A binary logit regression is run to determine the relationships between factors, such as price, reimbursements, time after patent expiry, as well as individual characteristics, and the probability that a generic is chosen. Results are to an extent consistent with previous literature, showing that price difference and time after patent expiry increase the chances that a generic is chosen.

The rest of this paper is structured as follows. The introduction will be followed by a literature review section where relevant research methods and results on the topic of generic and brand-name substitution prior to this paper are presented. Institutional background and data collection and

explanation will come next. A regression analysis and corresponding discussion of the results will form the main body of the thesis. Finally, the paper will finish with conclusions drawn on generic preference, combining the results of the regression with previous research.

2. Literature review

A well-known trade-off in healthcare exists between static and dynamic efficiency, between the opportunity to reduce costs and to foster innovation (Iizuka, 2012). Governments struggle to provide

(7)

7 mass, quality, and cheap medication for their citizens. Though medicines constitute only a small

proportion of healthcare expenditure compared to appliances, or stuff for instance, they are among the fastest growing segments of medical expenditures (Iizuka, 2012) and the greatest cost resulting from the ageing population (Clarke and Fitzgerald, 2010). Several past studies in a couple of countries illustrate the magnitude of cost reductions, should higher generic market share be fostered. Simoens and de Coster (2006) argue that generic medicines penetration indeed creates major savings for

healthcare providers and stimulates innovation. Taking into account that generics make up almost a half of the volume of all drug sales but only a fraction of the total cost, Simoens and de Coster have

estimated the cost of the ten most used active substances in the Netherlands assuming that 95 per cent of them are provided as generic substances. Results show an approximate 41 per cent decrease in public expenditure on these drugs compared to the case when only brand-names are dispensed. The Netherlands already advocates a strict reimbursement policy, which promotes one of the highest generic adoption rates of above 50%. Even in this advanced state, Gumbs (2007) has predicted a 2.4 million euros further annual cost cut if generic statins are substituted for branded among existing statin users and 87 million euros if all prescribed medicines are substituted for generics. Australia as well can benefit from generic substitution. Clarke and Fitzgerald assert that substantial savings can be made by allowing healthy competition and fostering cheaper generic drug versions. Strikingly, if Australia followed the pricing strategy of the UK of reimbursing only the cheapest drug and generics were prescribed in place of branded 50% of the time, for a decade, Australia could save $3.53 billion only from statins, Clarke and Fitzgerald (2010) estimate. Following the saving projections, Simoens and de Coster advocate the idea that the EU should foster and sustain the European generic industry to guarantee the competitiveness of the European pharmaceutical industry as a whole which will lead to both more competitive prices and create a stimulus to develop new more effective products. Identifying pharmaceutical costs is moreover easy, thus many countries, especially in Central and Eastern Europe, already focus their expenditure reduction policies on pharmaceuticals (Simoens, 2009).

For the above reasons, the switch towards generics is so crucial for the sound development of the pharmaceutical industry. Recommendation 4, mentioned before, suggests that the transfer can only be procured with policy measures from both the supply and the demand sides, united in a coherent general policy. Assuming that the supply side is more difficult to change and reacts to developments in the demand side, it is essential to investigate the demand side problem.

In the model of Ellison, Cockburn, Griliches, and Hausman (1997) the demand side problem is a two stage decision problem, where on the first stage the doctor makes a decision of the optimal

(8)

8 chemical entity needed for the patient depending on his condition and reactions with other substances the patient is currently taking. Then, on the second stage, the patient has the choice given the doctor‟s prescription to choose between a generic or original version of the chemical entity. It should be accounted that once a particular drug is chosen, only relative prices matter.

On the one hand, almost insignificant cross-price elasticity of demand between pairs of

therapeutic substitutes is observed as these will have different effects on the patient‟s condition and one pair will always be more suitable than the other. On the other hand, very high cross-price elasticity is characteristic within a pair of the therapeutic class, mainly patients are very price sensitive and will switch easily from a branded to a generic or between generic versions of the same drug. Similarly, we will expect high cross-price elasticity and high likelihood that generics are preferred if they are cheaper once generics and branded statins with the same active ingredient are grouped together.

Pechlivanoglou and van de Veen (2011) concentrate their research on an area in the north of the Netherlands, Groningen, where they use the linked data of all doctor practices and pharmacies in the region to analyze the likelihood of either generic or branded substitution with the aim to seek cost cutting strategies. They use observations for the four most prescribed therapeutic classes in the Netherlands with patents expiring during the observed period, among which the group of the statins. Each class is represented by one brand-name drug and one or more corresponding generic versions. Price, the scholars claim, has no explanatory power as none of the drugs studied requires patient co-payment. Difference in health insurance policies is also disregarded as everyone is obliged to be insured and all prescriptions were anyways fully reimbursed. The heterogeneity among patients and pharmacists and the resulting potential differences in probabilities, in other words the individual-specific effects, called for model random effects use. The result is that generic substitution occurred in 25% of branded subscription and that likelihood is higher when the patient is older, more experienced drug user and when more time has passed since a patent has expired and the highest among statins. As can be concluded, in this case, there is a strong relation between generic substitution and patient experience and timing.

Toshiaki Iizuka (2012) considers both timing and patient cost but adds the doctor agency when deciding to prescribe either a generic or a branded medicine. Iizuka uses rich 360,000 micro-panel data from Japan for 40 drugs and builds a model around the assumption that, physicians in Japan act to maximize their utility since they can earn profits when making a choice. He explains that in his country, more than half of the physicians are vertically integrated because they both prescribe and dispense pharmaceuticals. Health insurance covers all prescription medicines and all patients are using

(9)

9 some kind of insurance. In Japan, unlike the Netherlands, Iizuka claims further, doctors have little incentive to act on behalf of the insurer who has in most cases little influence. Therefore, the scientist focuses his research on the doctor-patient relationship instead. Out of the 40 drugs in his data, 6 drugs make it into his dynamic probit model on the criteria that these drugs generic version entered the market during the period covered by his data (1998 -2005) and their generic share is above 5%.

Iizuka‟s results show that, as expected, for integrated doctors mark-ups earned on generics have substantial positive effect on the probability of a generic choice, while for non-integrated doctors no effect can be concluded. In a similar fashion, patient cost is not significant on its own but only if the doctor is virtually integrated, indicating that non-integrated doctors do not take into account the patient costs. This can be partially explained with information advantage of those doctors who also purchase to dispense medicines. Iizuka‟s conclusion is that, in line with Simoens and De Coster (2006),

governments aiming to promote generic prescription should raise their mark-ups on generics even if for the minority doctors who only prescribe but do not dispense drugs this incentive will be irrelevant. Since doctors are not perfect agents, the patient cost, in comparison, has limited impact in their choice. Finally, some doctors will never adopt generics simply because of personal preference and the freedom to do so.

Furthermore, to shed light on why even in the presence of more affordable generic versions branded medicines remain market leaders, Lundin (2000) concludes that other unobserved

characteristics such as perceived not necessarily actual quality and even brand loyalty rather than price can be determinants. Most importantly, though, Lundin reasons that this price-insensitivity must be caused by the entering of financial incentives through reimbursements. If the reimbursement rates are high, then there would be less incentive to prescribe a generic even if its end price is cheaper. Using a microdata set from Sweden, Lundin chooses seven drugs whose patent has expired and for which generic versions are available, and builds a binary choice model, where the outcome is either a generic or a branded drug is chosen.

Lundin finds out that indeed after introduction of a reference price system in Sweden, from 1992 to 1993, which reimburses only 10% over the price of the most inexpensive drug, the market share of brand-name drugs has declined with 13% on average. The decline differs across drug classes and a drug priced further above the reference level has suffered a greater decline in market share. The scientist claims that the change brought by the new system is triggered by making patients more cost-savvy and in Sweden the patient is of main importance. Similar to Iizuka, Lundin thinks of the drug prescribing decision as weighting up the utility of two agents, this time the patient and the insurer

(10)

10 rather than the doctor. The Swedish doctor‟s remuneration is not related to the prescriptions he makes and there is no government initiative to keep costs down. At the same time, it is fairly easy for patients to switch practitioners, making doctors more patient- rather than insurer-oriented. The researcher states further that apart from the cost, patients prefer a given version only because of habit since there is no difference in the therapeutic effects, so past prescriptions will be indicative for the present preference. Doctor‟s heterogeneity in preference, in contrast, is exerted only when the cost is of no importance to the patient. Overall, both patients and doctors have significant acquired preferences.

3. Background and the data

Statins are a class of pharmaceutical drugs developed to fight, combined with a suitable diet and exercise, cardiovascular diseases and reduce the level of low-density lipoprotein (LDL) cholesterol, known as “bad cholesterol” (Consumer report, 2012). In July 1987, lovastatin became the first commercial statin released on the US market by Merck. The substantial revenues encouraged the consequent development of simvastatin and pravastatin. Atorvastatin by Pfizer known as Lipitor surpassed the success of its predecessors in 1996 and since then has become and remained the best-selling medical drug in history generating sales of over 125billion USD. The National Health and Nutrition and Examination Survey in the USA (NHNES) concludes that nearly 22 per cent of all Americans, aged 45 or older, currently take some form of a statin, making the class of statins one of the most widely prescribed class of medicines in the country. In 2009, the world market revenue for statins accounted for 27bln USD, claims IMS Health. These figures have earned statins the reputation of a gold mine.

Statins, as already mentioned are chosen because they constitute a vast market and conclusions drawn about them will be indicative for the pharmaceutical industry and because enough drugs within the class are no longer under the protection of a patent, with simvastatin patent expiring first in 2005 and rosuvastatin with patent just about to expire in June 2016.

To investigate the likelihood that a generic statin is prescribed, data is retrieved from a

physician‟s practice of six professionals in total based at one of the largest hospitals in Bulgaria. Out of a total of about 116, 000 entries, 5,699 were identified as statin prescriptions. By mistake, 461 entries of the fenofibrate Lipanthyl have been added to the subset of data and later removed, as fenofibrate, although used for reducing low-density lipoprotein, it is a complement not substitute of statins. Contrary to expectations for a very widely prescribed drug created from other countries, in Bulgaria

(11)

11 statins constitute merely less than 5% of all sample prescriptions. ). In the analysis, only simva-, atorva-and rosuvastatins are included because they are represented in the dataset with both generic atorva-and

branded variations, so a comparison can be made. Lova-, fluva-, and pravastatins are left out because either only a generic or only a branded version has been prescribed; therefore no substitution pattern conclusions can be made. Each drug from the data is placed within its statin group (simva-, atorva-, or rosuvastatin) and the groups all include non-perfectly substitutable active ingredients. Hence,

preference is investigated between generic and branded statins within the class itself but not cross-categorically, following Ellison, Cockburn, et al. (1997). A table of all brand names, the active ingredient, the manufacturer and whether the medicine is branded or generic (Table 1), is provided in the Appendix.

The final number of observations used is 4471 and they include characteristics, such as the market name of the drug, the active ingredient of the drug, information whether the drug is generic or branded, the average end price of the drug in the pharmacies as well as the reimbursement from the total price that the National Health Fund contributes to the patient, information about the packaging of the drug in number of tablets in a package and their strength in milligrams, characteristics of the patient such as age and gender, and finally the date on which the prescription has taken place.

All information about the end price and the reimbursements from the National Health Insurance Fund (NHIF) is extracted from the historic data of the doctor‟s practice medical software and double-checked with the official Positive Drug List published on the NHIF website which is updated every 1st and 16th of the month. Prices in the data are quoted as price per package. However, due to the

availability of various dosages in a package for the same drug, such as 30 tablets of 10mg, 28 tablets of 20mg and so on, prices need to be transformed in a uniform form before use. A standard price per 100 mg of a drug is therefore calculated. 1 Finally, an adjustment for inflation is made using the HICP from the Bulgarian National Statistics Office with the beginning of our data 2005 as a base year. All prices are in Bulgarian Lev, which is pegged with the Euro at 1.95583 Leva for 1 Euro. On the basis of the reimbursement in absolute terms and the end price, indicators such as reimbursement rate and patient co-payment are derived. These characteristics may be of more value to the patient as he is interested in the end price he pays as well as the per cent cost savings after NHIF reimburses him.

The information on the date of prescription is not in itself of interest but it provides a reference to the number of months passed since the patent expiry for a branded drug in a particular group. Zocor, branded simvastatin, has lost its patent on 1st June 2006. Sortis, branded atorvastatin, has lost its patent

1_{There is a slight inconsistency here as price per tablet is sometimes lower for a greater dosage in mg explained by the less}

(12)

12 on 1st November 2011. Finally Crestor, branded rosuvastatin as the most recently developed drug is yet to lose its patent 1st June 2016. What is interesting is that observations about rosuvastatin generic prescription exist as early as 1st December 2010 in the data at hand. It should be accounted that patent expiry dates are in respect to the FDA dates in the USA. In Europe and Central and Eastern Europe in particular patents are not so strictly defined and generic locally produced versions are available on the market before USA patent expiry even if the entry of generics is still impeded to a certain extent. Moreover, the period under investigation 2005-2014 encompasses the entry of Bulgaria in the EU in 2007 and regulations are likely to be under convergence. In a similar situation was Poland when, following 2004, it was given a 15-year transitional period to implement the harmonized EU data exclusivity legislation (Simoens, 2009). Alternatively, it might be the case that the generics present before patent expiry are authorized generics. These are drugs which “may be produced and sold by the brand manufacturer (through a subsidiary) or through a licensing agreement with another company yet are marketed and sold as generics”, as Berndt, Mortimer, Bhattacharjya, Parece, and Tuttle (2007) define. Therefore, as a fictional date of patent expiry for rosuvastatin, is counted the date of first generic prescription. Duration is as a result the created independent variable measuring the number of months passed between the patent expiry and the prescription date.

Outliers are identified among the relative price and the patient payment observations. In the case of relative price, all outliers are kept as valid data. Two data errors for values above 20 leva were corrected in the patient payment‟s observations. The rest of the outliers are kept as they are early prescriptions for rosuvastatin, which was very expensive at the time.

Framework-wise, in Bulgaria health insurance is state-owned and compulsory for all citizens. Therefore, difference in insurance policies should not be a leading factor for our analysis. Switching a family doctor is easy, so similarly to Sweden, the doctor takes into consideration above all the patient. Under these circumstances the patient has power in the prescription choice and patients in Bulgaria are generally considered to be cost savvy, as a result they could influence the doctor to prescribe the cheaper drug. A subject of interest is whether indeed and how price influences this choice in our sample. Further, if and how relevant is time; is it true that as more generics enter the market at a lower price, demand for them would rise in so called adjustment after the patent expiry.

4. Empirical model and results

To further evaluate the determinants of preference for generic medicines and in particular the medical class of statins, a regression mapping the link between the time that generics become available

(13)

13 on the market and their affordability and the probability that a statin is generic is built. A binary

logistic regression is conducted where the dependent variable contains information on whether the observed drug is generic or branded statin and the independent variables include financial incentives and cost savings from the patient side. Results are accounted for the three different active substance classes of statin and a categorical independent variable gensub represents accordingly the categories of simvastatin, atorvastatin, and rosuvastatin. Generic is a dichotomist variable taking the value 1 when the observed drug is a generic statin and 0 otherwise, when a branded statin.

It is useful to start the analysis with pre-estimation of the model and summary statistics of the data to understand the relationships between them. As can be seen from Table 2 the most substantial

Table 2 Distribution of statin observations across classes and type

proportion of generics is observed where the patent has expired first, namely the most time has passed since then, among the simvastatins, in line with Pechlivanoglou‟s (2011) expectations. Simvastatins are also the most prescribed class in the sample among the three as they are probably the most well-known and prices for both generics and branded have declined over time. Summary statistics on the means of price per 100mg per statin class indeed supports this assumption as average price for simvastatin is the lowest (1.95lv) and for rosuvastatin is the highest (10.07lv) and duration and pr100 are highly negative correlated. In the case of rosuvastatin, where the patent is yet about to expire but the market is partially penetrated by generic versions, generic and branded versions are equally observed and the branded is still slightly more frequent.

To examine the relationship between the candidate dependent variable and the main

independent variable relpr, a scatter plot is run with a fitted line. The scatter plot depicts the probability that a statin is generic (generic=1) or branded (generic=0) as the relative price increases. Graph 1 shows that higher relative prices make it more probable that the observed chosen drug is branded as the incentive for choosing a generic decreases.

Total 1,213 3,258 4,471 rosuvastatin 578 563 1,141 atorvastatin 309 1,038 1,347 simvastatin 326 1,657 1,983 substance branded generic Total generic type

(14)

14

a) with outliers b) with dropped outliers Graph 1 Linear relationship between the relative price and the probability of being a generic or branded statin

Relpr is the relative price of an observed statin compared to an equivalent branded statin with

the same active ingredient, e.g. simvastatin, atorvastatin, or rosuvastatin, available at the same time and with a similar packaging. The relative price has been chosen over the initial proposed predictor variable

pr100 with Ellison and Cockburn‟s (1997) argument that once it has been decided that a patient needs a

particular class statin in his medication, the absolute price of the drug itself seems to be out of the question as a purchase needs to take place. The real decision is not whether to take statin but whether to choose for a generic or branded version and the patient will likely be cost efficient comparing his options. Relpr has been generated from inflation adjusted pr100 as a rate that normally should fall between 0 and 1 for generic and equal 1 for branded observations. However, in 5% of the observations

relpr is actually above 1 indicating cases of market inefficiencies when the generic version is more

expensive than the branded version. In Graph 1 b) the outlier values above 1 have been dropped and the slope of the fitted line has become steeper, as the correlation between generic and relpr rises from -0.55 to the very high value of -0.80. Because the outlier values are not errors in the data but legitimate observations, however, simply dropping them is not acceptable. Investigation into transforming the outliers is not conducted and the outliers are left as they are in the model.

Although the direction of the relationship is as expected, that lower relative generic-to-branded prices make it more probable that a generic is chosen, the plot refers to the misspecification of using a linear model. As can be seen, the probabilities can go below or above 0 and 1 which is not in line with the binary dependent variable of interest, as pointed out by Stock and Watson (2012). The linear relationship with a binary dependent variable as specified in the linear probability model is bound to bias for the following reasons. It assumes a continuous dependent variable and the regressors are not

-1 .5 -1 -. 5 0 .5 1 0 1 2 3 4 relative price Fitted values type

0 .5 1 1 .5 0 .2 .4 .6 .8 1 relative price Fitted values type

(15)

15 restricted from 0 to 1 and freely range from minus to plus infinity. If OLS regression is still run, the magnitudes of the estimators may severely be over- or underinflated, the hypothesis tests may not be justified, and in general the estimates will be highly sensitive to values outside of the range observed. As a conclusion, OLS assumptions of linearity are inordinate and instead another probability regression model must be explored.

Such a model is the binary logistic model. Under it, the relationship between relative price and probability of being a generic as plotted in Graph 2 is closer to reality as the probability varies between 0 and 1 in the characteristic for the logistic regression S-like curve. Logit is preferred to probit

suggested by Iizuka(2012) and Lundin(2000) since results produced by the two empirical methods are very close but logit results are easier to interpret.

Graph 2 Logistic relationship between the relative price and the probability of being a generic or branded statin

The assumptions for logistic regression should be taken into account before further analysis proceeds. The normality of the independent variables is not among the assumptions. However, the regressors must be transformed if the transformation fits the relationship logically. Relpr has a non-normal distribution with a right-wing skew but its log transformation has a skewness of -0.24 and a kurtosis of 1.93, closer to a normal distribution. This transformation is not necessary per se but only if it makes sense in reality. The logistic regression also requires a large sample size because maximum likelihood estimators are less powerful than ordinary least squares. The sample size of above 4000 provides enough data. Next, the dependent variable should be binary, with the value 1 indicating a success, as is the case with defining generic. Multicollinearity should not be present but including interaction terms remains an option. This assumption will be tested in the process. Linear relation between the independent variables and the log odds is the last assumption.

0 .2 .4 .6 .8 1 Pr(g e n e ri c) 0 1 2 3 4 relative price

(16)

16 To start with, only two key explanatory variables are included in the regression and gradually more potential regressors will be added and tested for significance. The initial specification consists of

duration, the months that have passed since the patent of the branded statin has expired, and relpr, the

relative price of an observed statin compared to an equivalent branded statin. All output results can be seen for reference in Table 2 and 3.

1) Pr(Generic=1|relpr,duration)= F(β0 +β1relpr+β2duration)

The output of the binary logistic regression with two independent variables returns satisfactory test statistics. The chi-statistic as the square of the improvement of the model after including the two

(1) (2) (3) (4) (5)

VARIABLES Model 1 Model 2 Model 3 Model 4 Model 5

relpr -5.116*** -4.191*** (0.161) (0.158) duration 0.0243*** 0.0335*** -0.0596*** -0.0509*** -0.0574*** (0.00127) (0.00170) (0.00313) (0.00329) (0.00321) rembrat 16.19*** (0.917) price_a -1.397*** 3.100*** (0.0557) (0.305) patpay100_a -5.278*** -1.598*** (0.403) (0.0671) Constant 4.379*** 1.386*** 7.560*** 7.262*** 7.399*** (0.134) (0.152) (0.282) (0.312) (0.290) Observations 4,471 4,471 4,471 4,471 4,471 Pseudo R2 0.3987 0.3996 0.6431 0.5683 0.6261 Logistic Chi-2 2083 2088 3361 2970.34 3272.73

Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 2 Regression output results

independent variables over a model with only a constant is high at 2038 with 0.000 p-value, which indicates that the model has a good fit. The pseudo-R2 indicates that duration and relpr explain 39.87% of the variance in the binary dependent variable generic. Both independent variables are very highly significant at 1% level with p-values of 0.000. At this point precise conclusions about the strength of the predictors cannot be made but the signage is in line with the literature results from Pechlivanoglou (2011) and Ellison and Cockburn (1997). To be concrete, a unit increase in duration will increase the odds that the prescribed drug is indeed generic as more months after patent expiry allow deeper market penetration as well as better doctor and consumer awareness of generic options. For relpr, the effect is

(17)

17 negative, meaning that if the price of the available generic approach the original branded version, it will be less likely that a generic is preferred as the price incentive for that decreases.

2) Pr(Generic=1|relpr,duration,rembrat)=F(β0 +β1relpr+β2duration+β3rembrat)

Duration and relpr are unlikely to be the only predictors and another factor that might play a

role in choosing a generic is the reimbursement policy of the National Health Insurance Fund. The relative price is important but if there is a difference between generics and branded statins in the proportion of the price that is reimbursed and this difference influences the prescription decision, as Lundin (2000) proves, then the model will be subject to omitted variable bias. The variable rembrat is added to the model to correct for that. The pseudo-R2 jumps from 39 to 56%, so rembrat most likely predicts well the dependent variable. The Chi-statistic also improves and increases to 2970, suggesting that including a third independent variable improves the fit of the model. Rembrat is significant at the 1% level and has a positive effect on the odds that the statin chosen is generic. If the patient has to contribute only a fraction of the price for a generic, associated with a higher reimbursement rate, it is more likely that he perceives it as a better decision. The standard errors of the initial independent variables duration and relpr remain more or less unchanged. However, the standard error for rembrat is significantly higher and the beta is also exceptionally high. This high log odd ratio may be due to wrong scaling. Rembrat is a rate variable and to be easier to interpret, it can be multiplied by 100. The resulting output shows the log odd ratio for a 1% change instead of a 1 unit change. Both the log odd ratio and the standard error fall accordingly to 0.162 and 0.00917. Attention to the high log odd ratios when rembrat is included will be given further in the research.

3) Pr(Generic=1|relpr, duration,price_a)=F(β0 +β1relpr+β2duration+β3price_a)

Earlier the price was excluded from the model as the argument was that the relative price has more meaning as an indicator as in Ellison and Cockburn (1997). However, price is unlikely to have no importance at all. Not surprisingly, price_a (adjusted for inflation pr100) enters the model significantly at the 1% level. A unit increase in the price for 100 mg will decrease the probability that the drug is generic. All previously included variables remain highly significant, although the coefficients decrease, especially the one of relpr which suggests that previously the variance in the dependent variable caused by pr100 has been partially captured by relpr. Excluding price_a will cause bias in the estimators. The behavior of duration is note-worthy. The regressor remains close in magnitude but the signage flips and

(18)

18 the effect turns from positive to negative. A diagnostic includes running a linear regression with the same variables. Since multicollinearity is a problem of the variables and not of the model, this is plausible. Estimating the variance inflation factors of the predictors we obtain values all below 2 where the agreed critical value is 10. It can be concluded that multicollinearity cannot be an issue in this specification. Since the magnitude of duration is very small in the initial model with only two predictors, remains small with adding of price_a and is close to zero, even a small change of the variance can cause an estimator to flip sides and in this case the standard error duration indeed increases even if slightly from 0.0012 to 0.0032. Further, though they are not multicollinear, the predictor variables price_a and duration have a significant intervariable correlation of -0.5278. . A ridged or penalized logistic regression can be run to investigate in detail this relation and remedy the

duration estimator accuracy but that remains out of the scope of this thesis.

4) Pr(Generic=1|relpr, duration, price_a,patpay100_a)= F(β0 +β1relpr+β2duration+β3price_a+

β4patpay100_a)

If the inflation-adjusted patient payment patpay100_a is added as a third variable explaining variation from financial incentives, because patient co-payment matters according to Lundin (2000) and Iizuka (2012), the estimated inflation factors increase above 50 for both price_a and patpay100_a. The suggested multicollinearity biases the estimators and one of these independent variables should be dropped. Patpay100_a predicts the variation better as it produces a higher adjusted R2 and therefore

patpay100_a is preferred over price_a.

5) Pr(Generic=1|relpr, duration, patpay100_a)= F(β0 +β1relpr+β2duration+β3patpay100_a)

6) Pr(Generic=1|relpr, duration, rembrat, atorvastatin, rosuvastatin)= F(β0 +β1relpr+β2duration+β3rembrat+ β4atorvastatin+β5rosuvastatin)

As age and price_a were discarded as candidate regressors, the caterorical variable gensub, differentiating between the three classes of statin- simva-, atorva-, and rosuvastatin, is added to control for the effect across groups of statins. The previously recognized problem of high log odd ratios when

rembrat is used as a predictor was only partially resolved by rescaling rembrat. Odd ratios remain high

for the intercept (as well as for the catergorical variable gensub and rosuvastatin). One explanation can be that the dependent variable is more in the positive class (more generic than branded observations

(19)

19

(1) (2) (3) (4) (5)

VARIABLES Model 6 Model 7 Model 8 Model 9 Model 10

log odds relpr -5.130*** -3.055*** 1.341*** -5.124*** (0.162) (0.156) (0.313) (0.162) duration 0.0244*** 0.0503*** -0.0718*** 0.0154*** 0.00933*** (0.00127) (0.00277) (0.00667) (0.00194) (0.00214) rembrat 25.30*** 5.595** (1.491) (2.316) atorvastatin 2.906*** 3.477*** -0.868*** -2.103*** (0.251) (0.660) (0.199) (0.236) rosuvastatin -0.622*** 16.70*** -2.218*** -3.092*** (0.193) (1.339) (0.138) (0.174) 2.age 0.158 (0.127) 3.age -0.0616 (0.121) patpay100_a -4.403*** (0.261) lrelpr -4.875*** (0.161) Constant 4.361*** -1.405*** 10.19*** 5.389*** 0.395*** (0.162) (0.291) (0.786) (0.192) (0.126) Observations 4,471 4,471 4,471 4,471 4,471 Pseudo R2 0.3996 0.6261 0.8267 0.4636 0.6113 Logistic Chi2 2088 3272 4321 2423 3195

Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 3 Regression output results

are present). However, a possible problem of overfitting should be tested. That would mean that the data, though perfectly explaining the variance in the sample, may fail to make accurate predictions for future observations outside of the sample. The area under the receiver operating characteristics (AUC) is above 95% which indeed suggests overfitting. For every model there is a tradeoff between its bias and its variance. As the bias is very small in the case of overfitting and the model fits almost exactly the observations, an inherent bias toward predictive probabilities comes along with it and the variance would be large. The same high AUC is observed when patpay100_a is used instead of rembrat, suggesting persistence of the problem.

(20)

20 7) Pr(Generic=1|relpr, duration, rembrat, atorvastatin, rosuvastatin, patpay100_a)=

F (β0 +β1relpr+β2duration+β3rembrat+ β4atorvastatin+β5rosuvastatin+ β 6patpay100_a)

To double-check this peculiarity, if patpay100_a and rembrat are added at the same time, the bias gets more obvious. The log odds ratios for both relpr and duration change. Relpr becomes positive and duration becomes negative, implying that time after patent expiry decreases the likelihood to choose a generic, contrary to theory and prior results. To remedy this overfitting issue, the number of predictors should be shrunk to a ratio of at least 1 to 10 to the number of observations for good practice with MLE (maximum likelihood estimators), meaning that for the 4471 entries, there should be roughly and no more than 4 predictors. All meaningful variables should be included but only the most

meaningful variables will not lead to overfitting. The predictors left are relpr, duration and gensub.

8) Pr(Generic=1| relpr, duration, rembrat, atorvastatin, rosuvastatin )= F(β0 +β1relpr+β2duration+ β3atorvastatin+β4rosuvastatin)

The model specification above is derived from the most meaningful regressors. Standard errors are low, p-values are zeros and the relationships between the predictors and the dependent variable

generic take the expected signage. In addition, no extra-large log odds ratios result. Although

pseudo-R2 is lower and the chosen independent variables explain 46% of the variance of generic, Chi-statistic remains reasonably high 2423.50 and final decision should be made based on best judgment. For visibility, as it is often hard to understand the precise magnitude of change in a logistic regression, the following Graph 3 depicts the likelihood of a statin being a generic as the relative price increases across all three classes of statin. Indeed, the line for rosuvastatin lies below those for atorvastatin and

simvastatin, as with keeping relpr constant, the likelihood is the lowest for a rosuvastatin drug which is still under patent and it is highest for simvastatin, for which the patent has expired first.

As coefficient results are in log odds ratios, the coefficient for relpr indicates that, holding

duration and gensub at a fixed value, for a one unit increase in the relative price e(-5.124)=0.00595, a

99.4% decline in the odds of a chosen statin being a generic results. Holding relpr and duration at a fixed value, the odds of being a generic for a chosen atorvastatin (gensub=atorvastatin=2) over the odds for a simvastatin are e(-0.867)=0.42 or 58% lower. Finally, holding relpr and gensub constant, the odds of being a generic are 1.55% higher with one month increase in duration since e(0.0154)=1.0155.

These percentage changes in odds can be directly seen from running a logistic instead of logit command. In Table 4 the regression coefficients in different transformations can be seen. The

(21)

21

Graph 3 Logistic relationship between the relative price and the probability of the drug chosen being a generic or a trade-name for the three distinct groups of statins

column „b‟ lists the logistic regression coefficients, the log odds ratios. The column „e^b‟ presents the exponentiated coefficients which are the odds ratios. The last two columns return the odds ratios for a one standard deviation instead of one unit change in the independent variable and the standard

deviations themselves. The columns „z‟ and „p‟ give the Wald and p=values for the z-test.

Finally, a post-estimation check is done. A specification test which considers whether the “link” is appropriate and if all explanatory variables have been included, the linktest, regresses the dependent variable on the predicted values and their squares. The squares of the predicted values will have no power should the model be correctly specified but this is not the case, as can be seen in Table 5 in the Appendix. Earlier the number of the predictors was reduced on purpose to fit MLE conditions, so surely linktest indicates that no explanatory variables are included. An option to improve the model is to attempt to catch some of the residuals variance with an explanatory variable already in the model. A log transformation of relpr seems reasonable. Perhaps generic is linearly related to the log odds ratios of a transformation of relpr, not of relpr itself.

Table 4 Coefficient estimates

0 .2 .4 .6 .8 1 0 1 2 3 4 relpr

gensub = simvastatin gensub = atorvastatin

(22)

22 9) Pr(Generic=1|lrelp, duration, atorvastatin, rosuvastatin)=F(β0 +β1lrelpr+β2duration+

β3atorvastatin+β4rosuvastatin)

Immediately pseudo- R2 increases to 61.13% and Chi-statistic to 3195 and log odds ratios for the intercept decrease. Fit statistics provides evidence that indeed the log transformation of the relative price provides a better fit for the model. The regression coefficients again are presented below.

Table 6 Coefficient estimates

The results from the regression run coincide with the outcomes from Pechlivanoglou (2011) that longer period after a patent expiry causes higher preference for generics and further support Ellison and Cockburn‟s (1997) conclusion that prices, and in particular relative prices are significant determinants of generic over branded drug choice, as higher relative generic-to-branded prices make it less probable that a generic statin version is chosen. Distinction was made between statin classes similar to Cockburn and Ellison and again showed that preference patterns differ for different active substances pairs of generic and brand-name statins.

5. Summary and Conclusion

Governments around the world face a dilemma between keeping medical costs low and providing access to quality medication. As pharmaceutical patents worldwide expire, there is a high expenditure reduction potential by switching to less expensive but equally effective versions of the drugs. The purpose of this study was to research and evaluate the validity of factors shaping generic preference as proposed in previous studies discussed in the literature review section of this thesis, in a slightly different institutional environment where such a study has not been previously conducted, in

(23)

23 Bulgaria.

Researchers contradict as much as they complement each other‟s hypothesis about the most relevant and significant determinants of generic adoption. An empirical study concentrated on the segment of statins aimed to investigate relationships influencing the probability for choosing a generic over a branded drug for one of the most prescribed medical groups and concluded that all researchers were right to an extent.

Evidence to support the financial incentive as a leading one was found. Apparently, in Bulgaria patients are cost aware, consider the relative price of statins and since they have some power over the doctor‟s decision, the doctor is inclined to take the patient‟s costs into account. Therefore, lower relative prices will encourage generic prescription and higher will diminish this incentive. Time passed after generics become available also seem to shape positively the preference for generics. Nevertheless, the physician moral hazard problem, evident in other countries, is negligible in Bulgaria, where the doctor acts in the patient‟s interest. Patient‟s characteristics, such as age, turn to be insignificant contrary to one of the studies.

In conclusion, it has become evident from theory, past research, and this empirical study that the choice for a generic depends on its relative price compared to a trade-name drug and time of

prescription and varies between drug classes. As this study was performed on the particular drug group of statins in Bulgaria, a suggestion for further investigation would be conducting a larger scale research where more drug groups are included and data is collected for the same time period for a couple of countries, so that cross-country effects can be compared and more robust to time frame, country, and medicine class characteristics conclusions can be drawn.

(24)

24

References

Behner, P., Vallerien, Dr. Sv., Ehrhardt, Dr. M., Rollmann, D. (2009). Pharmaceutical companies in the

economic storm Navigating from a position of strength. The United Kingdom: Booz&co. (Strategy&)

Berndt, E.R.., Mortimer, R., Bhattacharjya, A., Parece, A., Tuttle, E. (2007). Authorized generic drugs, price competition, and consumers‟ welfare. Health Affairs, 26(3), 790-799

Clarke, Ph.M., Fitzgerald, Ed. M. (2010). Expiry of patent protection on statins: effects on pharmaceutical expenditure in Australia. The Medical Journal of Australia, 192(11), 633-636

Consumers‟ Union of the United States Inc. (April 2012). Evaluating statin drugs to treat: High cholesterol and heart disease Comparing Effectiveness, Safety, and Price. Consumer Reports Health,

Best-buy drugs.

Ellison, S.F., Cockburn, I., Griliches, Z., and Hausman, J. (1997). Characteristics of Demand

Pharmaceutical Products: An Examination of Four Cephalosporins. The Rand Journal of Economics, 28(3), 426-446

Food And Drug Administration (1984). Drug Price Competition and Patent Term Restoration Act. , p.l. 98-417. Retrieved 14 June, 2016, from https://www.gpo.gov/fdsys/pkg/STATUTE-98/pdf/STATUTE-98-Pg1585.pdf

Global R&D expenditure for pharmaceuticals. (2006). Retrieved May 10, 2016, from

http://www.statista.com/statistics/309466/global-r-and-d-expenditure-for-pharmaceuticals/

Gumbs, P.D., Verschuren, M.W.M., Souverein, P.C., Mantel-Teeuwisse, AK., de Wit, G.A., de Boer, A., Klungel, O.H. (2007). Society already achieves economic benefits substitution but fails to do the same fir therapeutic substitution. British Journal of Clinical Pharmacology ,64(5), 680-685

Iizuka, T. (2012). Physician agency and adoption of generic pharmaceuticals. American Economic

(25)

25 Lundin, D. (2000). Moral hazard in physician prescription behavior. Journal of Health Economics, 19, 639-662

Nelson S., Slordal L., Spigset O. (2009). Generic drugs instead of branded drugs prescriptions-long overdue. Tidsskr Nor Laegeforen, 126(4), 441-443

Pechlivanoglou, P., van der Veen, W.J., Bos, J.H., and Postma, M. J. (2011). Analyzing generic and branded substitution patterns in the Netherlands using prescription data. BMC Health Services

Research, 87(11)

Simoens, St. (2009). Developing competitive and sustainable Polish generic medicines market.

Croatian Medical Journal, 50(5), 440-448

Simoens, St., de Coster, S. (2006). Sustaining Generic Medicines Markets in the EU. Leuven, Katholieke Universiteit Leuven: Research Centre for Pharmaceutical Care and Pharma-economics

Stock, J.H., Watson, M.M. (2012). Introduction to Econometrics. Global Edition. Essex: Pearson Education Limited, 425-435

United Nations Conference on Trade and Development, (July 2015). The role of competition in the

(26)

26 APPENDIX

name generic name copyright name holder type

ACTALIPID Atorvastatin Actavis generic

AMICOR Atorvastatin Actavis generic

ATORIS Atorvastatin KRKA generic

ATORVASTATIN Atorvastatin Genericon generic

ATORVA Atorvastatin Tchaikafarma generic

ATORVISTAT Atorvastatin Unipharm generic

CRESTOR Rosuvastatin AstraZeneca branded

NEOSIMVA Simvastatin Neobalkanika generic

ROSUCARD Rosuvastatin Zentiva generic

ROSWERA Rosuvastatin KRKA generic

SIMGAL Simvastatin TEVA generic

SIMVACOR Simvastatin Tchaikafarma generic

SORTIS Atorvastatin Pfizer branded

SUZASTOR Rosuvastatin Ssandoz generic

TINTAROS Rosuvastatin Actavis generic

TORVACARD Atorvastatin Zentiva generic

TULIP Atorvastatin Sandoz generic

VASILIP Simvastatin KRKA generic

ZARANTA Rosuvastatin Gedeeon Richter generic

ZEPLAN Simvastatin Gedeeon Richter generic

ZOCOR Simvastatin MSD branded

Table 1. All statins prescribed, name, generic name, copyright holder, type: generic or branded

Table 5 Post-estimation tests

Difference of 771.862 in BIC' provides very strong support for current model. BIC': -3161.738 -2389.876 -771.862 BIC: -35498.265 -34726.403 -771.862 AIC*n: 2043.700 2815.562 -771.862 AIC: 0.457 0.630 -0.173 Adj Count R2: 0.625 0.618 0.007 Count R2: 0.898 0.896 0.002 Variance of error: 3.290 3.290 0.000 Variance of y*: 15.375 8.501 6.874 Efron's R2: 0.645 0.584 0.061 McKelvey and Zavoina's R2: 0.786 0.613 0.173 Cragg & Uhler's R2: 1.000 1.000 0.000 Maximum Likelihood R2: 1.000 1.000 0.000 McFadden's Adj R2: 0.609 0.461 0.148 McFadden's R2: 0.611 0.464 0.148 Prob > LR: 0.000 0.000 0.000 LR: 3195.359(4) 2423.497(4) 771.862(0) D: 2031.700(4465) 2803.562(4465) -771.862(0) Log-Lik Full Model: -1015.850 -1401.781 385.931 Log-Lik Intercept Only: -2613.530 -2613.530 0.000 N: 4471 4471 0 Model: logit logit

Current Saved Difference Measures of Fit for logit of generic