Heterogeneous Responses in Prescriptions to Medicare Part D: A Case Study on Physician Decision-Making and Antibiotics

Heterogeneous Responses in Prescriptions to Medicare Part D: A Case Study on Physician Decision-Making and

Antibiotics

Master thesis in Economics Trent Tsun-Kang Chiang

Nationalekonomiska institutionen Uppsala Universitet

VT 2015

Heterogeneous Responses in Prescriptions to Medicare Part D: A Case Study on Physician Decision-Making and Antibiotics

Trent TsunKang Chiang Faculty Advisor: Prof. Rita Ginja

Chiang, T., 2015: Heterogeneous Responses in Prescriptions to Medicare Part D: A Case Study on Physician Decision-Making and Antibiotics. Master thesis in Economics at Uppsala University, 2015, 38pp, 30 ECTS/hp

Abstract: To study the decision-making model behind how physicians making prescribing decisions, we studied the effects of the introduction of Medicare Part D in 2006 on numbers and characteristics of medications prescribed by physicians. We identified a significant increase in overall number of medications prescribed due to Medicare Part D but did not find any effects on the number of antibiotics. The result suggests there exist factors distinguishing antibiotics from other medications that led to a change in incentives to prescribe antibiotics, such as costs of antibiotics resistances. . We also identified the heterogeneity responses to Medicare Part D with respect to physician’s employment status, primary care relationship and patient’s gender and diagnostic categories.

JEL Classification: I13, I18. L65, I31

Keywords: Prescriptions, Physician Decision-Making, Antibiotics, Medicare Part D, Healthcare Reform

Trent Tsun-Kang Chiang, Department of Economics, Uppsala University, Kyrkogårdsgatan 10 B, 4th floor, SE- 751 20 Uppsala, Sweden. tsunkang@gmail.com

1

I want to thank Professor Rita Ginja, my primary advisor, and Professor Erik Grönqvist

during the entire thesis process for their consultation, wisdom and guidance. I also want

to thank Professor Mikael Elinder and Per Engström for their leadership in Master of

Economics Program at Uppsala University as well as all my friends and peers who

worked on Master’s thesis during Spring term of 2015 (VT2015). Lastly, the thesis will

not be possible without the scholarship and sponsorship from Swedish Institute’s

Scholarship Awards Program.

1. Introduction

Prescription medicine has been the dominant form of treatments chosen by physicians in the United States (Mott, 2001). With healthcare and pharmaceutical costs playing a crucial role in cost-effectiveness and cost-benefits studies for healthcare industry, prescription drug costs play an increasing important role on policy decisions made by either government agencies or health insurance organizations (Hart, 1997). Different from other common goods, demand for prescription medicine is mostly driven not by the consumers (patients) but physicians, who issue prescriptions to the patients (Carrera, 2013). While there have been proposed models suggesting that physicians take patient input and suggestions into account in clinical scenarios, sometimes as a result of direct- to-consumer advertising, there have been no definitive studies showing the size of effects of patient request in clinical decision making process (Carrera, 2013; Armantier, 2003).

Therefore, a model for decision-making process of physicians is an important component to understand and make informed policy decisions regarding health care policies in health insurances, payment schemes, and cost-controls.

Many studies in the past have found that physicians are not perfect agents of patients in prescribing medicines. Besides patient’s clinical and financial benefits, such as the insurance status and clinical advantages, physicians were also found be influenced by financial benefits for themselves, advertising to consumers, advertising to physicians and probability of non-compliance in prescribing drugs (Liu, 2009; Armantier, 2003). Besides observing effects on the expenditure on drugs or number of drugs prescribed, some previous studies have also used generic substitution of brand name drugs in understanding physicians’ decision-making models (Liu, 2009; Godman, 2013).

However, there exist few studies that examined the heterogeneity of physicians

prescription behavior change in response to changes in the financial status of patients

with different categories of drugs. Of particular public interests are antibiotics, which

may results in negative externalities in the form of possible antibiotics resistances with

every prescription. Using the implementation of Medicare Part D in 2006 in the United

States, we investigate the heterogeneous effects of the policy change on change in

physicians prescribing behaviors between antibiotics and other drugs to determine if

physicians take the unique negative externality of antibiotics prescription into account during the prescription decision-making process.

This paper is organized as follows: in section 2, we first introduce the relevant background policies and institutional models on relevant issues. Then, we introduce a model on physician prescribing decision. In section 4, we discuss the empirical strategy used while section 5 presents data. Section 6 discusses the results and section 7 concludes with relevant policy implications.

2. Background

A. The choice of focus on antibiotics

Antibiotics are a clinical class of compounds that is effective in treating common bacterial infections. They are one of the most widely-known healthcare intervention in the public. Healthcare providers provided 258 million courses of antibiotics in the US in 2010 (833 prescriptions per 1000 persons) (Hicks, 2013). Because of its effectiveness and popularity, patients often request antibiotics even for mild conditions that may not be bacterial infections. Prescription for antibiotics is high, especially to persons younger than 10 years old or older than 65 years old. However, antibiotics prescriptions are often unnecessary despite medical best practices suggest to only prescribe antibiotics if confirmed bacterial infections. Doctors often feel pressured by patients to prescribe unnecessary antibiotics (Bennett, 2010). One qualitative study actually recorded a physician indicating that “You can’t just say ‘It’s viral, you don’t need antibiotics, go away,’ because [patients] feel they’re being fobbed off. They feel that their illness is not being taken seriously.” (Butler, 1998). Some studies have suggested that as much as 50%

of the antibiotics to outpatients in the United States may be unnecessary (Hicks, 2013).

Antibiotics uses contribute significantly to the development of antibiotics resistances around the world (Hicks, 2013). Antibiotics select mutated bacteria with resistances to antibiotics to survive and eliminate the competing non-resistant bacteria in patients.

People then share resistant bacteria within the population with any subsequent

interactions with other people. With decades of antibiotics usage, resistance to

erythromycin, a common antibiotic, is 28.3% in the US and higher overseas (72.4% in

Hong Kong) (Bennett, 2010).

Therefore, it is of interest to examine if there exists heterogeneous effects on antibiotics prescription relative to other treatments, which may result in the negative externalites in making clinical decisions for patients. With physicians already found to prescribe medicines in higher quantity and more expensive drugs to patients with prescription drug insurance such as Medicare part D, it is crucial to understand how the change in patient’s payment status affects prescription of antibiotics relative to other medicines (Hu, 2014).

B. Institutional Model for Prescribing Medicines

Formulary and Insurances

In the US, health plans can influence the usage of prescription drugs by adjusting the level of cost sharing of the cost and changing the procedures for obtaining prescription drugs. During late 1990s and early 2000s, many private insurers in the US started to cut costs on drug expenditures by implementing stringent cost-sharing models, such as a tiered or incentive-based formularies of benefit design (Carrera, 2013). Governments or insurers also use formularies and treatment guidelines to limit the usage of prescription drugs in other countries with different payment systems, such as in Sweden and Germany (Persson, 2012). The formulary is typically controlled by the health organizations or contracted pharmacy benefit manager, which provides cost information, such as tiers or generic substitution information for a specific drug, via computer software to prescribing physicians (Mott, 2001). Two other common strategies to control drug expenditures in some European countries, reference pricing and price cap regulations, are less common in the US, because US government lacks to power to regulate the prices of drugs directly, except through limited influences from Medicare and Medicaid (Brekke, 2009).

Procedures to obtain drugs may also adjusted in order to discourage or encourage drug usage. By requiring physicians to obtain prior authorizations from insurances before prescriptions or requiring the usage of low-cost generic drugs before brand-name drugs, insurers can also lower the expenditures by reducing the usage of expensive drugs (Carrera, 2013).

Previous literature has focused on if physicians prescribe differently for patients with

different insurances systems, and the result is affirmative. Glied et al. (2002) found that

physicians’ prescription pattern responds to the insurance status (if belongs to a Health Management Organization, HMO or traditional-fee-for-service plans) of the majority of their patients but less to the status of individual patients.

Physicians

Physicians issue prescription to patients they treat or see. Prescription denotes a specific molecule and dosage, either by brand name or molecular (generic) names, for the specified patient (Carrera, 2013). In issuing the prescription, physicians generally take patient’s symptoms, medical information into account in finding the appropriate medication to prescribe. However, there also exist a number of other factors that physicians may consider in prescription decision-making process other than medical knowledge and patient’s symptoms.

Substantive amount of literature has detailed the effects of pharmaceutical firms’

marketing efforts on physician’s prescribing choices. It is commonly recognized that physicians’ prescribing decisions are affected by pharmaceutical detailing, sampling or other marketing efforts (e.g. sponsored academic conferences) (Fischer, 2010; Campo, 2005; Epstein, 2014). There is also a large amount of literature showing that physicians take patient’s payment methods, such as health insurance status into account in prescribing treatments. Physicians are more likely to prescribe more expensive, brand name pharmaceuticals when patients insured, relative to the cheaper, generic equivalents of the drug (Lundin, 2000). In a setting where patients face no marginal costs for prescribing more medicines, physicians were found to prescribe more expensive medicines to elderly patients in Japan (Iizuka, 2007). Physicians who also have direct financial incentives themselves in dispensing drugs were found to prescribe more drugs in Taiwan (Liu, 2009). However, using qualitative evidences, Campo (2005) concluded that physicians generally do not pay large attention to patient’s financial status, to a higher degree when large portions of patient’s costs are covered by insurances. Hart (1997), on the other hand, concluded that drug costs can be an important factor in physician’s prescribing decision.

In the American context, physicians generally have no financial incentives themselves in

prescribing medicines. However, many studies have indicated that physicians still take

patients’ financial status into account when prescribing medicines. Inpatient physicians were found to prescribe more in response to drugs costs in a simulated survey (Hart, 1997). Epstein et al. (2014) found that patient’s formulary information plays a more important role in prescribing decision when physicians have access to information technology platforms that provides an easier access to formulary information. Without the use of information technology, formulary information plays a smaller role in physicians’ decision-making process. Hu, Decker and Chou (2014) found that the expansion of Medicare part D to include prescription drug medicines in 2006 resulted in a statistically significant 35% increase in prescription medicines after the reform.

Pharmacists

Patients with prescriptions are required to go to a pharmacist in order to have the prescription filled. While pharmacists cannot change prescriptions, pharmacists can suggest a generic substitution of brand-name drugs to patients without prescribing physicians’ approval, which is also appreciably accepted by physicians (Godman, 2013).

In fact, both physicians and pharmacists believe that pharmacists are responsible for reviewing a patient’s health plan and its formulary in order to choose cost-saving alternatives (Carrera, 2013). Pharmacists can also substitute the prescribed drug with similar and less expensive, but not molecularly identical, drugs to patients, with the approval from the prescribing physician (“therapeutic interchange”). Pharmacists are also likely the source of patient’s drug price information, besides price references or physicians (Mott, 2011). However, generic substitution, in which pharmacist supply a generic version of a prescribed multi-source drug molecule, does not require physician approvals.

Medicare and Medicare Part D

Medicare is a national social insurance managed by Centers for Medicare and Medicaid

Services, a part of US Federal government, for elderly citizens in the US that are more

than 65 years old. Before 2006, Medicare had only 2 traditional fee-for-service parts,

part A and B, and a managed care component, part C. Medicare Part A is the “hospital

insurance” that covers mostly inpatient hospital services ranging from lab tests to doctors visits to hospice cares. Medicare Part B is a supplemental program that covers services that are not covered by part A, ranging from costs associated with outpatient services to ambulance costs to preventive care.

Both part A and B do not have prescription drug coverage and only covers inpatient and outpatient healthcare services except in extremely limiting circumstances in part B (Hu, 2014). Thus, Medicare patients had to obtain prescription drug insurances from other sources, such as the employer, Medicare Part C, or state programs prior to 2006. On January 1, 2006, Medicare part D was introduced to cover prescription drugs. Two types of private insurance plans of part D were introduced for patients to voluntarily enroll: a Prescription Drug Plan (PDP) and a Medicare Advantage-Prescription Drug Plan (MA- PD) that covers both healthcare services and prescription drugs.

The introduction of part D decreased the number of Medicare beneficiaries without any drug coverage from 19% in 2002 to 10% in July 2006 (The Henry J. Kaiser Family Foundation, 2010). Medicare Part D increased the number of annual prescriptions by 30% and the expenditure for prescription drugs by 40% for both normal elderly population and elderly population in poor health. (Kaestner and Khan, 2012). Other studies have similarly concluded that the introduction of Medicare Part D increased the total monthly drug spending among enrollees by $13-41, depending on the number of previously drug spending (Zhang, 2009). Yin (2008) concluded that Medicare part D resulted in a modest increase in drug usage and reduced the average out-of-pocket drug expenditures among Medicare beneficiaries.

3. Model

Hu, Decker and Chou (2014) described a model for physician decision-making:

(1)

in which physician maximize his/her utility function in treating patient i as described

above. D

represents the drug treatment the patient received while T

represents other

non-pharmaceutical patient received. P

and p

is the unit price for a unit of drug and

other treatment, respectively. k is the fraction of out-of-pocket price for patients for drugs

and other treatments, respectively, after the insurance or other payment discounts. A

is the effort by physician on treating the patient, which may include things such as lifestyle recommendations. Physician’s efforts on patients’ health A

are typically not observable by patients. C(A

) thus represent the cost to physician to make such efforts. Finally, F(.) is a “health production function”, in which patient produced health with a combination of drug, other treatments and physician efforts. The unit health of patients is worth m for the physician.

To account other facets of the physician’s decision to prescribe medicine found in previous literatures, we decided to modify the above model and propose the following:

(2)

in which physician prescribed patient i non-antibiotics drug D

,antibiotics B

and other treatment T

Physician has a negative utility with prescribed antibiotics B

since antibiotics prescriptions help develop global resistance to antibiotics with a probability q

and unit cost of resistance c

. In addition, physicians also take patient’s preference I

of drugs, antibiotics and treatments into account. I

can be negative when physician prescribe in disagreement with patient’s preference of treatments and drugs, or positive when both patients and physicians agree on the treatment, antibiotics and medications prescribed. I

can also be zero if patient do not express specific preferences to patients. Notably, the patients can request unnecessary antibiotics from the physician and if physician refuses, it will produce a negative I

while, if physician comply, I

would be positive. Physicians receive positive utility when they agree to prescribe such medicines, in which they receive a unit “agreement” worth of n. In addition, physicians usually have imperfect                                                                                                                

2

We recognize that this is a simplified model where health is treated as a static stock in the model in one time period. Alternatively, we can write the dynamic model as follows:

in which s is the current time period and F

is the health stock from the previous period, and r

is the discount factor for the effectiveness of current antibiotics due to current antibiotic resistance. However, we do not currently understand the detailed mechanism of antibiotics resistances. For simplicity, we choose to use our static model in equation 2.

information on specific financial information of patients, k

and k

. q

and c

are estimated by individual physicians from current scientific literatures without definitive magnitudes, but the sign of q

and c

should definitively be larger than zero, as the positive link between antibiotics usage and resistances is scientifically sound.

Thus, we can also conclude that physicians are imperfect agent for patients, as they do not have perfect information on patients’ financial/insurance status. The optimal level of physicians’ effort level A

happens when the marginal benefit of added efforts equals to zero.

If we examine the first order derivative of the above model with respect to non-antibiotics D

and antibiotics B

, we can write the following by taking the first-order derivative to solve for physician’s welfare maximization problem:

(3)

(4) Using equation 3 and 4, we can solve conditions in which physicians maximize their welfare:

(5) In equation 5, ∂F/∂D and ∂F/∂B reflect the marginal health benefits of non-antibiotic drugs and antibiotics, respectively. ∂I/∂D and ∂I/∂B are the marginal “preference” of patients on an additional unit of non-antibiotic drugs and antibiotics, respectively.

Using equation 5, we examine a specific scenario: when patients express no preferences over the treatment or medication prescribed; that is, I(D

B

,T

)=0.

(6)

Thus, the differences between marginal health benefits of non-antibiotics and antibiotics must equate to the unit cost of antibiotic resistances. Because q

c

is determined by current public health conditions and not likely to change when drug insurance policy                                                                                                                

3

About 50% of antibiotics prescription in US is estimated unnecessary and antibiotics

prescription is an important factor in growing antibiotics resistances (Hicks, 2013; US

Department of Health and Human Services, 2014; Mott, 2011)

changes (k

increases or decreases), the physician, who considers the potential cost of antibiotic resistances would not increase antibiotics prescription in the event of a policy change to decrease k

, such as the introduction of Medicare Part D.

Lastly, we must note the constraint of the limited amount of financial resources available for the patient’s to pay for drug out-of-pocket expenditures.

While patient’s detailed financial status beyond insurance status is generally observable by the physician, patients may change the preferred combination of drugs, antibiotics, and treatments because of their own financial constraints. These preferences are illustrated through the preference function I

in our model. Notably, the introduction of Medicare Part D resulted in both an increase in medications prescribed and a reduction in out-of-pocket drug expenditure (Yin, 2008). Thus, the lower out-of-pocket cost after policy implementation may lessen the magnitude of I

in the model after 2006.

4. Empirical Strategy

In this study, we aim to see if physicians behave differently when deciding prescribing antibiotics against all other drugs, which is similar to antibiotics in other aspects but will not result in the negative externality of antibiotics resistance. To examine the hypothesis, we must be carefully in preventing selection bias in which difference was a result from the unique quality of antibiotics other than antibiotics resistance. Thus, we utilize the implementation of Medicare Part D as the exogenous policy shock and examine if the degree in increase in drug prescription were different for antibiotics compared to other treatments (a difference-in-difference approach coupled with regression-discontinuity DD-RD). This approach was used by Hu et al (2014) and they found a 35% increase in drug prescriptions with Medicare Part D. However, they did not examine any specific prescriptions, such as antibiotics.

4

Patient’s financial constraint can be written as following, but is not typically observable by the prescribing physician:

in which p

is the price level for all other goods, C

is the consumption of all other goods,

and M

is the budget of patient i, assuming patients won’t borrow for out of pocket

healthcare expenditures.  

Difference-in-Difference (DiD or DD) design is a common strategy to test for policy treatment effects in economics. By estimating the differences in changes of outcomes, DD can estimate the treatment effects of policy, assuming that both 60-64 and 65-69 year-old patients share similar prescribing patterns prior to 2006. However, the assumption may not hold if patients’ age within the range of 60-69 affects prescribing patterns, and if the prescription patterns change discontinuously at age 65, for example, due to retirement age or any other motives unrelated to life cycle. Thus, we decided to employ the combined RD-DD model in order to reduce the weaknesses of DD. RD-DD model compares the difference in changes immediately before and after age 65 before and after the policy changes. RD-DD model allows us to address the policy confounding issue in which patients become qualified for all Medicare Programs at age 65 by assuming that Medicare Part A and B remains time-invariant over the 2006 threshold of Medicare Part D implementation.

Therefore, we simply estimate the following equation as our main specification:

Outcome

can be the number of total medications prescribed, antibiotics prescribed to

patient i by physician j or the share of antibiotics in the total number of medications

prescribed. Charlsonindex is a dummy variable that takes value of 1 if the Charlson

comorbidity index, which predicts the 10-year morbidity calculated by the diagnosis, is

larger than 0, and 0 otherwise. Elderly is an indicator variable that takes value of 1 ifa

patient is over age 65 at the time of the visit and thus qualified for Medicare Part D and 0

otherwise. After2006 is whether the visit happens on or after 2006, when Medicare Part D

was available. AgeYears are the years away or from age 65. AgeYears can be modeled in

either a quadratic or a cubic structure. X

are the control variables for patients such as

race, ethnicity, gender, and major diagnostic category associated with the visit. In some

specifications, we also used ϕ

, which are physician fixed effect as we can track if visits

were treated by the same physician during the same survey year. We focused on

individual’s age between 60-69 at the time of the visit for the band of DD-RD designs

since people who are either too young or too old have different observable and unobservable variables compared to patients’ age between 60 and 69. To further ensure the homogeneity among observations, we also limit our samples to those who are seen by primary care physicians, i.e. those physicians who specialized either in family medicines or internal medicines. In addition to the main specification, we also investigated the role of physician’s employment status; if owning a practice affects prescribing behaviors. For all specifications, we used simple ordinary least square (OLS) to estimate the coefficients for variables. To capture the uncertainty in using patient age in years as supposed to days, which is a continuous variable, we employed age-clustered robust standard error in all regressions as detailed in Lee and Cards (2008). The conventional standard errors fail to capture the effect of having the running variable in clustered format and would produce a smaller standard error. By assuming the group (“clustered”) structure behind the running variable (age), the estimation in this study have the same coefficient but a different standard error compared to the conventional standard error. The age-clustered standard errors also accounted the for the heteroscedasticity and should be appropriate to our model the conventional heteroscedasticity-consistent robust standard errors.

5. Data

We use the National Ambulatory Medical Care Survey (NAMCS) from the National

Center for Health Statistics, a US federal government agency. The survey has been

conducted annually since 1989 and data are available until 2010. We used the data from

2002 to 2004 and from 2006 to 2010 to eliminate the possibility of anticipatory effects in

2005 (Table 1). We used data from 2002 because NAMCS underwent a significant

reform in 2002 and made several changes in its data collection techniques as well as

items collected. The nationally representative sample of non-federally-employed office-

based physicians provided visit-level data in which physicians provided information on

each visit by a single patient during a one-week period. The variables in the dataset

included the four geographical regions of the physician’s practice, if the physician is

located in a metropolitan-statistical-area (MSA), patient’s basic demographic

information, drugs prescribed and patient’s insurance and payment methods. It also

cataloged the patient’s diagnosis code in ICD-9-CM and symptoms. Due to the use of the

restriction of public-use file, we were not able to pinpoint the exact location or the birthday of the patient, which requires us to use the age in years instead to control for eligibility for Medicare Part D. We looked at patients with age from 60-69 on the date of the visit in NAMCS. To define antibiotics, we use the list of antibiotics drug codes from US Department of Health and Human Services (HHS) (2014) to generate the count for numbers of antibiotics prescribed during each visit, as listed in Appendix Table A1.

Table 1 presents the description of the variables used in the analysis, which are all observations in NAMCS with patient age between 60 and 60, occurred between year 2002-2010 (excluding 2005) that are seen by a physician specialized in either family medicine or internal medicine.

Table 1. Descriptive Statistics for the Estimation Samples

Variables Obs Mean Std.

Dev.

Min Max

Number of total Medications Prescribed

7035 3.36 2.63 0 8

Number of Antibiotics Prescribed

7035 0.12 0.35 0 3

Age 7035 64.23 2.83 60 69

Charlson Index 7035 0.30 0.53 0 2

Male (%) 7035 43.14%

Older than 65 Years Old

7035 46.35%

Pay with Medicare

5295 32.60%

Race

White 4911 69.81%

African American 592 8.42%

Asian 180 2.56%

Native Hawaiian or Pacific Islander

16 0.23%

Native American 112 1.59%

Blank 1215 17.27%

Diagnostic Categories*

Respiratory System 1111 15.79%

Infectious and Parasitic 214 3.04%

Diseases

Neoplasms 169 2.40%

Endocrine, Nutritional and Metabolic Diseases

2380 33.83%

Diseases of the Blood and Blood-Forming Organs

105 1.49%

Mental Disorders 519 7.38%

Nervous System 478 6.79%

Circulatory System 2424 34.46%

Digestive System 558 7.93%

Genitourinary System 491 6.98%

Skin-Related 372 5.29%

Musculoskeletal and Connective Tissue

1306 18.56%

Congenital Anomalies 16 0.23%

Injury 347 4.93%

* A patient can have up to three diagnoses recorded in NAMCS during a single visit to physician’s office.

6. Results

A. Regression Discontinuity in Number of Medications and Antibiotics

We first replicated the results in previous literature indicating an increase in drug usage

after the implementation of Medicare part D. We used a simple regression discontinuity

(RD) design with local linear regression (triangle kernel) to graph any discontinuity in

four different outcomes: 1) numbers of total medications, 2) number of antibiotics, and 3)

share of antibiotics as part of total number of medications for patients with age between

60 and 69 (Figure 1-8). For the graphs below, we have included data from 2005 as well

as data from physicians specialized in all specialties in NAMCS Datasets to maximize the

number of observations available. In each graph, we listed the local linear regression

estimator for the discontinuity and if the local linear estimate at the age cutoff is

statistically significant in the caption. The solid lines are the local linear regression

results after the introductions of Medicare in 2006 while dashed lines represent the period

from 2002 to 2005. Solid filled circles are the averages of the outcome variable post-2006

while pluses are prior to 2006. To allow easier visual inspections on the figures, the

graphs below are the representation of local linear regressions run independently on both

sides of the threshold. The RD graphs used to generate the bandwidth can be found in the appendix Figure A1 to Figure A6. In these figures, the estimate of the effects of the implementation of Medicare Part D is given by

(

where α

and α

are the sizes of the discontinuity at age 65 after and before 2006, respectively.

Note. Total Number of Medications Coded, age 60-69. (2002-2005 Estimate: -.074 (0.159); 2006-

2010 Estimate: 0.264 (0.156)*)

Note. Total Number of Antibiotics, age 60-69. (2002-2005 Estimate:-.0136(0.00967); 2006-2010 Estimate: 0.00734 (0.00837))

Note. Share of Antibiotics, age 60-69. (2002-2005 Estimate:-0.00378(0.00517); 2006-2010

Estimate: 0.00141(0.00390))

Examining Figure 1, we can clearly identify a discontinuity in total number of medications prescribed after Medicare Part D in or after 2006 but not in the samples before 2006, which is consistent with prior literature (Hu, 2014). However, the number of antibiotics showed similar prescribing patterns before and after 2006 as indicated in Figure 2. Using the share of antibiotics in the total number of medications prescribed, we can also see that the prescribing pattern remains similar prior and after the implementation of Medicare Part D in 2006 on Figure 3. However, it is worth noting that there exist decrease in shares for both before and after 2006 groups at age 65 in Figure 3, which may be a result from the combination of an increase in total number of drugs prescribed and the constant number of antibiotics prescribed.

Importantly, RD design assumes that the assignment to either side of discontinuity threshold is as good as in a random experiment. In this study, RD suffers from a confounding policy discontinuity at age 65: besides becoming eligible for Medicare Part D, patients who turn 65 would also be qualified for Medicare Parts A and B, which covers inpatient and outpatient services. Thus, we cannot infer causal relationships from Figures 1-6 and must look for other strategies in order to identify the effects of Medicare Part D expansion.

B. RD-DD Design and Results

On Table 2 , Table 3 and Table 4, we present the results from the simple OLS regression

with outcome variable being the number of medications prescribed, number of antibiotics

prescribed and share of antibiotics in the medications prescribed, respectively. Across all

three tables, specification 1 is our main specification without controlling for diagnostic

categories from the model described above. Specification 2 added 14 diagnostic category

dummies as controls to the specification 1. Specification 3 added the physician fixed

effects since NAMCS survey was conducted in one physician’s office to record all visits

to the office during that period and allowed us to identify records of visits to the same

office during single survey year. We changed the age variable structures from the cubic

structure, used in previous three specifications, to quadratic structures in specification 4.

Specification 5 is the regression with only age controls and without any other covariates, such as sex, gender, or if physicians are in solo practice.

From Table 2, we can clearly observe that the introduction of Medicare Part D created an increase in total number of medications prescribed in all specifications. The largest magnitude of the variable of interest (Elderly*After2006), which indicated the effects of the implementation of Medicare Part D on the outcome variable, was observed in specification 2 on Table 2. Due to the many control variables omitted due to multicollinearity in specification 3 (physician fixed effect), the reduction in significance can be attributed to the larger standard errors caused by lack of control variables. It is worth noting that the results for the total number of medications prescribed remain significant across all 5 specifications, even without any control variables in specification 5. In addition, the magnitude of the effect remains relatively stable ranging from 0.3-0.35 additional medications per visit due to Medicare part D.

From Table 3, we observed that the introduction of Medicare Part D did not result in an increase in prescriptions of antibiotics. Since numbers of antibiotics are strictly less than the total number of medications, we can see that the coefficients for antibiotics are significantly smaller than those on Table 2. However, across all specifications on Table 3, none of them showed a statistically significant effect. Moreover, in specification 3 on Table 3, we can see that the magnitude actually became negative controlling with physician fixed effects. Thus, we can conclude from Table 2 that the number of antibiotics prescribed, in general, did not increase with the introduction of Medicare Part D in 2006.

The results from the share of antibiotics in total medications prescribed are presented in Table 4. Similar to Table 3, Table 4 shows no statistically significant effect at age 65 before or after 2006

. These results are consistent with the observation from Part A’s RD graphical analysis, which indicated that antibiotics did not have a jump in usage after the introduction of Medicare Part D in 2006, either in absolute terms or in relative terms to other medications. Furthermore, the lack of discontinuity for the number (and share) of                                                                                                                

5

Besides number and share of antibiotics prescribed, we also tested using a dummy

variable indicating any antibiotics were prescribed and reached similar conclusions as in

Table 3 and 4.

antibiotics prescribed after 2006 is consistent with our model’s prediction that physicians have different decision-making process for prescribing antibiotics and non-antibiotic medications, such as considering antibiotics resistances. Since the unit cost of antibiotic resistance was not changed when Medicare Part D was introduced, physicians are not more likely to prescribe antibiotics due to a change in patient’s financial status, which decreased k

, the percent of out-of-pocket costs of prescription medicines for patients with insurances.

Table 2. Results for Total Number of Medications NAMCS 2002-2004, 2006-2010 Number of

Medicines

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

Elderly*After2006 0.350** 0.385** 0.298* 0.341** 0.322**

(0.132) (0.129) (0.154) (0.130) (0.112)

Age Cubic Cubic Cubic Quadratic Cubic

Covariates# Yes Yes Yes Yes No

Diagnostic Categories

No Yes Yes No No

Physician Fixed Effect

No No Yes No No

Observations 5,373 5,371 5,371 5,371 5,371

Robust standard errors, clustered by age of the individuals, in parentheses

*** p<0.01, ** p<0.05, * p<0.1

# Covariates includes (Visit Month, year, race, ethnicity, sex, Charlson Index, e-medical record and solo practice, months from Dec.31,2001)

Table 3. Results for Total Number of Antibiotics, NAMCS 2002-2004, 2006-2010

Number of Antibiotics (1) (2) (3) (4) (5)

Elderly*After2006 0.009 0.006 -0.001 0.009 0.009

(0.016) (0.014) (0.020) (0.016) (0.017)

Age Cubic Cubic Cubic Quadratic Cubic

Covariates# Yes Yes Yes Yes No

Diagnostic Categories No Yes Yes No No

Physician Fixed Effect

No No Yes No No

Observations 5,373 5,371 5,371 5,371 5,371

Robust standard errors, clustered by age of the individuals, in parentheses

*** p<0.01, ** p<0.05, * p<0.1

# Covariates includes (Visit Month, year, race, ethnicity, sex, Charlson Index, e-medical record and solo

practice, months from Dec.31,2001)

Table 4. Results for Share of Antibiotics, NAMCS 2002-2004, 2006-2010

Share of Antibiotics (1) (2) (3) (4) (5)

Elderly*After2006 -0.004 0.005 -0.009 -0.004 -0.003

(0.007) (0.005) (0.007) (0.007) (0.007)

Age Cubic Cubic Cubic Quadratic Cubic

Covariates# Yes Yes Yes Yes No

Diagnostic Categories

No Yes Yes No No

Physician Fixed Effect

No No Yes No No

Observations 5,373 5,371 5,371 5,371 5,371

Robust standard errors, clustered by age of the individuals, in parentheses

*** p<0.01, ** p<0.05, * p<0.1

#Covariates includes (Visit Month, year, race, ethnicity, sex, Charlson Index, e-medical record and solo practice, months from Dec.31,2001)

C. Robustness Checks

To verify the robustness of our results above, we check the results for varying bandwidths for the DD-RD designs. We first verify the robustness for the RD-DD bandwidth selections within close range to age 65. Table 5 presents results from varying bandwidths with the identical specifications from specifications 1 and 2 on Table 2.

According to Lee and Lemieux (2010), regression discontinuity bandwidths need to balance the noise created by having too few observations and the heterogeneity in observations between the two ends of the selected bandwidth.

From Table 5, we can see that the statistically significant effects of the introduction of Medicare Part D remains significant when including larger or smaller bandwidths in specifications 1 to specification 5. The results also hold when controlling for diagnostic category dummies in specification 3-5. With a small age bandwidths, however, the significance of the effects of the policy was reduced in specification 5, which may be a result of smaller number of observations available in a more limited sample, which in term increased the possibility of been affected by noise in the sample. Lastly, we also use the number of antibiotics as outcome variable with various age bandwidths. To our surprise, we observed a marginally significant effect in specification 6 when we expand the RD sample bandwidth from 60-69 to 58-71. However, the significant result disappeared when we regress with 59-70 year-old patients in specification 7 on Table 5.

The result indicated that while Medicare Part D may also have an effect on the number of

antibiotics prescribed, it is marginally significant and relatively weaker than the effects

on the total number of medications prescribed. To further understand the reason behind the marginally significant results in specification 6 and 7, we examine the distribution of the observations between ages 58-71 in our dataset, with and without 2005 data as listed in appendix Table A2. We find that there are no outliers or aberrations in the numbers of observations across the age spectrum except the natural decline in number of observations as people age. Thus, the significant results in specification 6 can be a result of the inclusion of the high number (400) of observations at age 58 post-2006, relative to other age in the dataset.

On Figure 4, we can see that after expanding the RD bandwidth, there still exist little evidence in any discontinuity at age 65 post-2006.

Table 5. Regressions with Different RD Bandwidths

(1) (2) (3) (4) (5) (6) (7)

Outcome Variable Total Number of Medications Prescribed # of Antibiotics RD Bandwidth 58-71 61-68 58-71 61-68 62-67 58-71 59-70 Elder*After2006 0.358*** 0.428** 0.386*** 0.452** 0.307* 0.0255* 0.0200

(0.0906) (0.147) (0.0893) (0.148) (0.135) (0.0132) (0.0149)

Age Cubic Cubic Cubic Cubic Cubic Cubic Cubic

Diagnostic Categories

No No Yes Yes Yes Yes Yes

Observations 7,539 4,328 7,534 4,323 3,260 7,534 6,434

R-squared 0.081 0.077 0.126 0.118 0.159 0.171 0.164

Robust standard errors, clustered by age of the individuals, in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Figure 4. Antibiotics Prescription across age 65, individuals with age 58-71.

Table 6. Regressions Including 2005 Observations –Age 60 to 69

(1) (2) (3) (4) (5) (6)

Outcome Variable Total Number of Medications

Total Number of Antibiotics

Share of Antibiotics

Elderly*After2006 0.273** 0.201* -0.000256 0.00475 -0.00721 -0.00460 (0.107) (0.103) (0.0153) (0.0145) (0.00525) (0.00508) Diagnostic Categories

Control

Yes No Yes No Yes No

Observations 5,977 5,977 5,977 5,977 5,977 5,977

Robust standard errors, clustered by age of the individuals, in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Secondly, we includE the 2005 observations in our data and repeated the regressions with

varying RD bandwidth as presented in table 6. However, we observe surprising results as

the treatment effects of Medicare Part D decreased across all specifications. The

reduction in treatment effects after including 2005 observations is surprising as Hu et al.

(2014) showed an opposite direction of effects (increase in effects) after including 2005 observations in the sample, which they attributed to the anticipatory effects of Medicare Part D introduction by patients. Patients may conserve drug usage in 2005 in order to qualify for Medicare Part D on Jan.1, 2006 (Hu, 2014). The anticipatory effect would lead to the overestimation of Medicare Part D and resulted in a larger magnitude of treatment effect after including 2005 data. Alpert (2012) found that Medicare Part D introduction induced anticipatory effects, when patients delayed receiving chronic drug prescriptions in 2005 but not acute drugs. Nevertheless, the inclusion of 2005 data still indicated a positive jump in number of medications prescribed due to Medicare Part D.

One possible explanation for the decrease in magnitude of the coefficients may be due to our dataset for regressions contains only physicians specialized in either family or internal medicines in office visit (non-hospital) settings, which generally prescribe less specialized, higher-priced medications than specialized physicians in other fields (e.g.

cardiologists or dermatologists).

While we do not have a clear explanation for the reason behind the drop in magnitude, we still find that our results significant and valid despite the reduction in treatment effects after including 2005 data. Further research may be warranted in order to examine the effects of Medicare Part D on the prescribing decision-making process in 2005.

To further check the robustness of our regressions, we also generated a series of placebo

cutoffs on age and years of policy implementation as shown in Table 7. Using the main

specification similar to those of specification 1 on Table 2 with 2005 data, we can

conclude that our results are robust against placebo age cutoffs (age 64) from

specification 1 and 2 as well as placebo year of policy implementation (2007) from

specifications 3 and 4.

Table 7. Placebo Cutoffs for ages and Policy Year#

Panel A: Placebo for Treatment Age Number of Medications Prescribed

Number of Antibiotics Prescribed

(1) (2)

Age64*after2006 0.191 0.00431

(0.111) (0.0115)

Observations 5,977 5,977

R-squared 0.111 0.011

Panel B: Placebo for Years of Policy Implementation

(3) (4)

Elderly*After2007 0.141 0.00913

(0.111) (0.0168)

Observations 5,977 5,977

R-squared 0.111 0.011

Robust standard errors, clustered by age of the individuals, in parentheses

*** p<0.01, ** p<0.05, * p<0.1

#Physicians in the samples specialized in either family or internal medicine. .

D. Effects from the Employment Status of Physicians

To further understand the factors behind the reasons for the discrepancy with previous literature on the anticipatory effects of 2005 observations, we investigated the role of physician’s ownership of the practice in prescribing medicines. Sun (2006) reported that physicians who do are not the owner of their own practices prescribed 1.5 times more antibiotics in upper respiratory infection cases compared to those who do.

In Table 8, we can clearly identify that physicians who do not owns the practice are more likely to prescribe more medicines after the introduction of Medicare Part D. Physicians who do not own the practice may either be an employee or a contractor of the practice.

Similar to Sun (2006), Medicare Part D only has statistically significant effects on physicians who do not own the practice in specification 1 and specification 2 on Table 8.

The fact is surprising since we controlled for if physician is a solo practitioner, the use of

electronic medical record and diagnostic categories as a proxy for the type of the primary

care physicians (all physicians in the sample specialized either in family or internal

medicine) in all specifications on Table 8.

Table 8. Physician’s Ownership and Prescribing Patterns#

Number of Medications Prescribed

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

Ownership Physician is not the Owner of the Practice

Physician is the Owner of the Practice

Elderly*After2006 0.401** 0.312* 0.240 0.335

(0.149) (0.170) (0.233) (0.228)

Diagnostic Categories Controls

No Yes No Yes

Observations 2,614 2,610 2,759 2,761

R-squared 0.122 0.163 0.120 0.157

Number of Antibiotics Prescribed

(5) (6) (7) (8)

Elderly*After2006 0.0283 0.0197 -0.00614 -0.00611

(0.0301) (0.0246) (0.0220) (0.0175) Diagnostic Categories

Controls

No Yes No Yes

Observations 2,614 2,610 2,759 2,761

R-squared 0.012 0.161 0.025 0.189

Robust standard errors, clustered by age of the individuals, in parentheses

*** p<0.01, ** p<0.05, * p<0.1

#Physicians in the samples specialized in either family or internal medicine.

Similar to our prior findings, Medicare Part D did not have a significant effect on number of antibiotics, regardless of physicians’ employment status, as shown in specification 5 to 8 on Table 8.

Previous literature has shown that physicians working for either private clinics or private hospitals are more likely to prescribe cheaper generic drugs compared to those who works for public sectors in Taiwan (Liu, 2009). Healthcare market competition has also driven up the number of prescribed antibiotics in Taiwan (Bennett, 2010). Table 8 shows that the introduction of Medicare Part D has only significant effects in increasing the total number of prescribed medications for primary care physicians who do not own the practices. However, physicians in the US, unlike those in Taiwan, are less likely to own a pharmacy and have less financial incentives tied to prescribed medications.

Past literatures have attributed the higher rate of prescription of physicians who do not

owns the practice to peer pressure, legal concerns or the physician’s desire to validate the

reason for an office visit (Sun, 2006). Our findings in Table 8 showed that physicians who do not owns the practice are also more likely to respond to the introduction of Medicare Part D than physicians who do not.

E. Physician being the Primary Care Physician (PCP) of the Patient

Sun (2006) also found that physicians who are also the primary care physician (PCP) of the patient visiting are more likely to prescribe antibiotics in upper respiratory infection cases. The generally long-term relationship between PCP and patient may result in patients’ higher willingness to request specific medicines or allows physicians greater knowledge regarding patient’s financial status. To see if PCPs are also more susceptible to the influence of drug insurance expansion, we presented the results on Table 9 below.

From specification 1 on Table 9, we observe that only physicians who are also PCP of the patients have statistically significant policy effect from Medicare Part D’s implementation. We need be cautious about interpreting the results when a physician is not PCP as the sample sizes were very small (658) and thus had large standard errors and may have inaccurate estimates in specification 2 and 4. However, Medicare Part D still did not have statistically significant effects on the number of antibiotics, even only looking at PCP physicians.

Table 9. Prescription Effect of Being the Primary Care Physician of the Patient#

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

Outcome Variable Total Number of Medications

Total Number of Antibiotics Primary Care Physician

of the patient?

Yes No Yes No

Elderly*After2006 0.337** 0.473 0.0121 -0.0356

(0.144) (0.398) (0.0191) (0.0500)

Observations 4,715 658 4,715 658

R-squared 0.151 0.203 0.155 0.303

Robust standard errors, clustered by age of the individuals, in parentheses

*** p<0.01, ** p<0.05, * p<0.1

# Sample includes only physicians specialized in family or internal medicine.

F. Prescribing Differences with Different Diagnosis categories

It is also possible that physicians treating patients with certain diagnoses prescribe differently in response to the implementation of Medicare Part D. To see the differences in prescribing behavior, we first return to specifications 2 on Table 2 and Table 3. By examining the statistically significant diagnostic category variables in Table 10, we identified two groups of diagnosis categories that are statistically significant in the total numbers of medications and antibiotics. For respiratory diseases, nervous system diseases and genitourinary system diseases (“Group A”), the dummies for these diagnoses are statistically significant and positive for both medications and antibiotics. However, for endocrine, nutritional and metabolic diseases, mental disorders, musculoskeletal system and connective tissue diseases, circulatory diseases and injuries (“Group B”), the dummies for these diagnoses are statistically significant and positive for total number of medications but negative for total number of antibiotics. Lastly, we will classify diagnostic categories that do not belongs to either group A or B into “group C”, which includes infectious diseases, neoplasms, blood-related diseases, digestive system diseases, skin-related diseases and congenital anomalies.

To further investigate the characteristics and effects of these groups of diagnostic categories, we examine the prescribing behavior change in patients that have a diagnosis in any of the diagnostic categories in all three groups in Table 11 below.

Table 10. Comparison Between Effect of Different Diagnostic Categories

(1) (2)

Outcome Variable Total Number of

Medications

Total Number of Antibiotics

Elderly*After2006 0.385** 0.00623

(0.129) (0.0138)

Group A: Significantly Positive in both Specifications

Respiratory System Diseases 0.609*** 0.310***

(0.124) (0.0190)

Nervous System Diseases 0.408** 0.0334*

(0.152) (0.0156)

Genitourinary System Diseases 0.316*** 0.179***

(0.0903) (0.0202)

Group B: Significantly Negative in

Antibiotics but Significantly Positive

in Number of Medications

Endocrine, Nutritional and Metabolic Diseases and Immunity Disorders (B)

0.567*** -0.0518***

(0.132) (0.0140)

Musculoskeletal System and Connective Tissue Diseases (B)

0.436*** -0.0528***

(0.131) (0.00791)

Mental Disorders (B) 0.629*** -0.0852***

(0.135) (0.0192)

Injury and Poisoning (B) 0.315* -0.0358*

(0.142) (0.0161)

Circulatory System Diseases (B) 0.829*** -0.0485***

(0.0655) (0.0101)

Group C: Insignificant in at least One of the Two Specifications

Digestive System Diseases (C) 0.645*** -0.00154

(0.141) (0.0148)

Blood and Blood Forming Organs (C) 0.527 -0.0126

(0.301) (0.0511)

Skin-Related Diseases (C) 0.286 0.102***

(0.202) (0.0228)

Congenital Anomalies (C) -0.0466 -0.0944***

(0.642) (0.0272)

Infectious Diseases (C) 0.283 0.0684

(0.263) (0.0505)

Neoplasms (C) -0.196 -0.0507**

(0.246) (0.0208)

Observations 5,371 5,371

R-squared 0.154 0.166

Robust standard errors, clustered by age of the individuals, in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 11. Sub-Sample Analysis with Different Diagnostic Groups

(1) (2) (3) (4) (5) (6)

Outcome Variable Number of Medication

s

Number of Antibiotic

s

Number of Medication

s

Number of Antibiotic

s

Number of Medication

s

Number of Antibiotic

s

Group A A B B C C

Elderly*After200 6

0.251 -0.00665 0.131 0.0128 0.401* 0.0524**

(0.285) (0.0790) (0.177) (0.0153) (0.211) (0.0204)

Observations 1,534 1,534 3,768 3,768 1,062 1,062

R-squared 0.123 0.037 0.111 0.012 0.152 0.043

Robust standard errors, clustered by age of the individuals, in parentheses

*** p<0.01, ** p<0.05, * p<0.1

On Table 11, we observed no statistically significant effects in either group A and B for the total number of medications and antibiotics prescribed. However, we observed statistically significant effects for both outcome variables with patients been diagnosed with diseases in group C, which also had the a relatively low number of observations.

Due to the low number of observations available, it is impossible for us to further divide samples in group C to smaller groups in order to see which diagnostic groups are behind the statistically significant effect of Medicare Part D on total number of drugs and, specifically, antibiotics. We note that group C included infectious diseases, a group that may be more elastic to antibiotics and other medications than other types of diagnoses.

However, since only 160 out of the 1024 observations in group C were patients diagnosed with infectious diseases, we cannot conduct further analysis due to the low number of samples available. Further investigations may be necessary to understand if physicians treating patients with specific diagnostic categories are more likely to prescribe a higher amount of antibiotics due to the implementation of Medicare Part D.

G. Gender and Prescribing Behavior

Table 12. Gender and Prescribing Decision

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

Outcome Variable Total Number of Medications Total Number of Antibiotics

Gender Female Male Female Male

Elderly*After2006 0.415*** 0.425 0.0144 -0.00980

(0.100) (0.258) (0.0222) (0.0173)

Observations 3,053 2,318 3,053 2,318

R-squared 0.147 0.183 0.188 0.161

Robust standard errors, clustered by age of the individuals, in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Multiple studies have shown that women were more likely to have a higher amount of antibiotics prescription (Sun, 2006). Indeed, gender may play a role in respect to the tendency of patients to seek health care resources or indicate their preferences during the visit to physician’s office, which will lead to higher amount of drugs been prescribed after the implementation of Medicare Part D.

Table 12 shows the effects of Medicare Part D on the number of medications and

antibiotics on the female and male patients separately. We observe that the increase in

total numbers of medications prescribed after the implementation of Medicare Part D in 2006 was primarily driven by female patients with a statistically significant coefficient in specification 1 on Table 12

. In specification 2, male patients have the equal amount of magnitude but a large standard error, which is surprising due to the relatively large amount of male observations available. This may be a result from a higher variance in the number of medications prescribed to male patients than to female patients.

H. Do patients in the sample become sicker after the implementation of Medicare Part D?

Table 13. Charlson Index Before and After 2006

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

Outcome Variable Charlson Comorbidity Index Dummy (>0)

Elderly*After2006 0.00893 0.0173 -0.00686 0.0155

(0.0204) (0.0197) (0.0203) (0.0188)

Observations 5,373 5,371 5,977 5,977

R-squared 0.018 0.305 0.016 0.303

2005 Data No No Yes Yes

Diagnostic Category Controls

No Yes No Yes

Robust standard errors, clustered by age of the individuals, in parentheses

*** p<0.01, ** p<0.05, * p<0.1

One of the underlying assumptions in our DD-RD design is that the trend before and after the Medicare Part D’s implementation would stay the same in the absence of the policy intervention in 2006. The assumption has to hold in order for our findings regarding the increased number of total number of medications prescribed due to Medicare Part D to be valid, as well as our finding about the lack of a significant increase or decrease regarding antibiotics. However, Medicare Part D’s introduction may also induce a increase in tendency for more patients to seek health care in a physician’s office due to the lower costs and may change the demographics of patients in our sample. If sicker patients, which normally requires a higher amount of medications than less sick patients, were

6

We also conducted regressions in which we used the interaction of female and

elderly*after2006. However, given that the coefficient it’s not statistically significant for

both antibiotics, we cannot reject the null hypothesis that Medicare Part D introduction

has equal effect on men and women.

more likely to be included in our sample after 2006 than before, it could be an alternative explanation to our previous findings than the direct effect from Medicare Part D.

To see if patients became “sicker” and thus requires a higher amount of prescriptions after 2006, we used a dummy variable of positive Charlson comorbidity index as the outcome variable and the exact same specification from table 2 and 3 on Table 13, with the exception of using Charlson index as a control variable.

On Table 13, we observe that patients were not statistically sicker or less sick before and after the Medicare implementation in 2006 in a variety of specifications. The result ensured the robustness of our model and indicated that the indirect effects of Medicare Part D on patient’s behavior via patient’s finance are negligible. Medicare Part D implementation ‘s effect on total number of medications prescribed resulted primarily from its influences on physician’s prescribing decision-making process.

7. Discussion and Conclusion

In this study, we first seek to construct a model on physician’s decision-making model in

prescribing medications in an office-based setting with a physician specialized either in

either family medicine or internal medicine. We based our model on Hu et al. (2014) but

added patient preferences and antibiotics resistance costs into the decision-making

process. While there have been evidence showing that physicians do take patient’s

involvement and preferences into account, it is worth noting that patient’s input into

physician’s prescribing decisions are complicated by the existence of direct-to-consumer

(DTC) advertising, when pharmaceutical advertisements target patients and suggest

patients to ask their physicians for specific medications (Carrera, 2013). Campo (2005)

found that many physicians held negative views on DTC campaigns and rather appreciate

more patient inputs, may instead feel threatened by patient’s involvement in prescribing

decisions. Thus, our model is limited in interpreting individual physician’s variation in

the decision-making process but rather try to show a general model that can be used in

policy studies. We had special interests in antibiotics, a common class of drugs that have

societal negative externalities with every prescription in the form of the development of

antibiotics resistances. By incorporating antibiotics resistances into our model, we tried to