Data Scientist

Technology, Institutions, Public Policy

Saving Billions in U.S. Healthcare: Clues from Medicare Drug Prescription Data

Drugs prescribed under the U.S. Medicare Part D program in 2013 amounted to $103 billion, contributing about 31% of the $330 billion spent on prescription medicines in the country. The government and its agencies have undertaken a number of initiatives to reduce healthcare expenditure on drugs, and an important focus area has been to promote greater use of generic drugs in place of equivalent and more expensive brand-name drugs. To encourage transparency and further research, in April 2015, the U.S. Centers for Medicare and Medicaid Services (CMS) released a dataset with utilization and costs for all prescription drugs that individual physicians and other health care providers prescribed in 2013 under the Medicare Part D Prescription Drug Program.

In the following, we analyze the Medicare Part D dataset in combination with other publicly available data, and identify opportunities for significant cost savings.

1. The Nuggets Upfront

A summary of the key results is as follows:

2. Big picture: Drug cost trends

We start by looking at some high level trends that will provide clues for where to focus attention on for savings.

2.1. Drug costs are top heavy

The total cost for each drug is computed and the list of drugs in descending order of total cost is obtained. The plot below shows the cumulative expenditure distribution for this sorted list. It is observed that, out of more than 2700 drugs featuring in the data, the top 50 drugs in terms of cost contribute about 51% of the total cost, the top 100 drugs 67%, and the top 200 drugs 81% of total cost.

Key takeaway: In exploring opportunities for cost savings, it is important to focus on the top few hundred drugs. By exploiting this, we will arrive at valuable insights quickly.

Figure 1. Cumulative cost plot with drugs sorted by contribution to total cost. Out of 2700+ drugs, top 50 account for 51% of costs, top 100 for 67%, and top 200 for 81%.

2.2. What makes them big - cost per unit and volume

There are two contributing factors in total cost for each drug - cost per unit and number of prescriptions. To identify the primary driving factor for individual drugs, we generate a scatter plot of total cost vs. the total number of prescriptions for the top 50 drugs. The drug names are color coded by the class of drugs they belong to, with the classes being listed in the legend.

The high cost, lower volume drugs on the left side of the plot are all brand-name drugs, while the high volume, high cost drugs on the right are generics. Some drug classes (eg. metabolic agents) are dominating the top 50 list. For respiratory and psychotherapeutic agents, there are no generic 'challengers' for expensive brand-names in the top 50.

The plot can be better explored by using the zoom and reset tools in the left-top corner of the plot. Also, information for individual drugs can be seen more clearly by hovering over the corresponding circle/name.

Key takeaway: Trends for drug classes need to looked into separately.

Figure 2. Scatter plot of total cost vs number of prescriptions for top 50 drugs in terms of cost. Hover over circles and use zoom/reset tools in panel.

2.3. Drugs have families.

To understand trends for high level drug classes, we manually collect class information for drugs in the dataset by searching on and Google. The drug class information provided in the CMS dataset is at a fine grained level, and not sufficient for our purposes.

Estimated relative contributions to total cost by each class of drugs are shown below. The class identification was carried out manually for the top 200 cost contributing drugs which account for about 81% of the total costs. The results shown are therefore indicative and may be off by a few percentage points.

Key takeaway: The top 9 classes contribute about 90% of the cost.

Figure 3. Percentage contribution to total cost by drug class.

2.4. Mapping cost per enrollee.

To understand geographical trends, we begin by mapping mean cost per enrollee by state. Enrollment data for Medicare Part D is obtained from the CMS website. The data is provided on a monthly basis, and we use the numbers for the middle of the year, July 2013, for computing the average cost per enrollee. Again, the results are to be treated as indicative rather than exact.

The plot below shows that there is significant variation in the average cost per enrollee by state, with the highest for Hawaii ($5024) being more than three times the lowest for Wyoming ($1636).

Key takeaway: The causes for large variations between states need investigation.

Figure 4. State-wise map of average cost per prescription. The four color shades indicate the states in four quartiles, the cost ranges for which are as shown in the legend. The maximum value is more than 3 times the minimum.

3. Digging in: Generic vs. brand-name drugs

The results above have identified two key factors for further investigation: variations by drug class and state. We now add another factor to the mix: generic vs. brand-name drugs. A primary objective of our analysis is to identify and establish compelling evidence for cases in which expensive brand-name drugs have been prescribed at an unreasonably high rate. A key challenge is that it is difficult to establish a "correct" rate of brand-name prescription. Arguments about high prices, direct-to-consumer marketing, and counter-arguments about effectiveness, safety, discounts and innovation costs are legion.

In this section, we will arrive at some compelling opportunities for cost savings by increasing the use of less expensive generics. To begin with, it is useful to take stock of initiatives over the years that have made progress on the same.

3.1. Existing measures for encouraging use of generics

These efforts date back to the Hatch-Waxman Act of 1984, which has been partly reponsible for shifting overall generic utilization rates from a very low 19% in 1984 to around 86% in 2015. In recent years, the U.S. Congressional Budget Office (2010) and Government Accountability Office (2012) have produced illuminating reports on the cost impacts of using generics. These findings were incorporated in the Affordable Care Act, and as detailed in the Medicare Budget Overview for 2015, CMS has introduced several measures to increase generic drug prescriptions under Part D by altering copayments for generic and brand-name drugs, preventing collusion between brand and generic drug manufacturers, and suspension of questionable prescriptions.

3.2. Asking different questions

So, how can we find further opportunities for cost savings? Instead of (only) asking generalized questions such as "Are brand-name drugs being overprescribed?" or "Do branded drugs provide better outcomes than generic drugs?" or "Is the higher price of brand-name drugs justified?", we exploit the Part D data and assembled drug class information to ask more specifically:

3.3. Identifying prominent generics and brand-names by drug class

To answer the questions laid out, we first identify the most prominently prescribed generics and brand-names across all states for each drug class, as listed in the table below.

Drug Class Prominent Generic Drugs Prominent Branded Drugs
Metabolic Atorvastatin Calcium, Simvastatin, Metformin HCL, Pravastatin Sodium, Alendronate Sodium Crestor, Januvia, Lyrica, Lantus Solostar, Lantus, Zetia
Central nervous system Hydrocodone-Acetaminophen, Gabapentin, Tramadol HCL, Donepezil HCL, Oxycodone HCL Namenda, Celebrex, Lyrica, Oxycontin, Fentanyl, Exelon
Psychotherapeutic Quetiapine Fumarate, Sertraline HCL, Citalopram HBR, Risperidone, Olanzapine, Escitalopram Oxalate, Venlafaxine HCL ER Abilify, Seroquel XR, Invega Sustenna
Cardiovascular Metoprolol Succinate, Lisinopril, Amlodipine Besylate, Losartan Potassium, Tamsulosin HCL Diovan, Ranexa, Tracleer, Letairis
Respiratory Fluticasone Propionate, Montelukast Sodium Proair HFA, Advair Diskus, Spiriva, Symbicort, Ventolin HFA, Flovent HFA
Gastrointestinal Omeprazole, Pantoprazole Sodium, Lansoprazole Nexium
Antineoplastic Methotrexate Copaxone, Revlimid, Tarveca, Gleevec, Zytiga
Immunologic Restasis, Zostavax Enbrel, Humira, Rebif, Gilenya, Avonex
Antiinfective Doxyxycline Hyclate Norvir, Truvada, Isentress, Atripla, Prezista, Reyataz

3.4. Where the new, big savings are: It varies by drug class

Having identified the prominent generics and brand-names in each drug class, we can now add up their numbers and costs for each state and look at the resulting trends. A few illustrative examples are shown below, and the trends for all classes can be explored with this interactive tool.

a. Metabolic agents

The plot below shows the statewise scatterplot for the percentage of a selection of most commonly prescribed generic drugs vs. the percentage for a selection of branded drugs prescribed for metabolic ailments. The size of each bubble is a measure of the mean amount per claim - larger bubbles indicate that, on average, more expensive drugs are being prescribed.

  • As can be expected, with some variations, the higher the percentage of branded drugs, the greater the cost per claim, and hence larger bubbles.
  • There is significant variation in the generic vs. brand-name prescription rates across states. The highest mean per prescription cost is $137 for Delaware, and the lowest is $64 for Rhode Island.
  • Delaware, New Jersey, West Virginia, New York, and Connecticut have mean per prescription cost higher than $100.
  • Rhode Island, Oregon, and Colorado have mean per prescription cost lower than $70.
  • By region, states in the West have the lowest costs, with Utah ($86) being the highest regionally, and Oregon ($66) the lowest.

b. Central nervous system agents

The plot below shows the statewise scatterplot for central nervous system agents.

  • Again, there are significant differences in the relative rates of prescription of branded and generuc drugs across states.
  • New Jersey ($110), Hawaii ($103), Connecticut ($99), and New York ($95) have the highest rates of branded drug prescriptions and correspondingly high mean cost per prescription.
  • Missisippi ($56), Arkansas ($59), West Virginia ($60), Washington ($60), Nevada ($60), and Oregon ($61) have the lowest mean cost per prescription.
  • Again, with the significant exception of Hawaii, states in the West have lower rates of branded drug prescription.

c. Antiinfectives

The plot below shows the statewise scatterplot for antiinfectives. Strikingly, brand-name drugs are prescribed at a rate close to 90% in Washington D.C., compared to 50% in Massachusetts, and only 15% in South Dakota. Since `Doxyxycline Hyclate` is the only prominent generic we have identified in this class, it also shows that this drug is prescribed at a rate close to 3% in D.C and at 80% in South Dakota.

4. Potential savings

We can now estimate potential savings from using the insights obtained above. The first targets will be states with the highest brand-name prescriptions for each drug class. For simplicity, we consider the following hypothetical programs:
  • Program 1: Top 5 states for each drug class decrease brand-name prescriptions to achieve average per prescription cost of the 6th highest.
  • Program 2: Top 10 states for each drug class decrease brand-name prescriptions to achieve average per prescription cost of the 11th highest.
  • Program 3: Top half of states for each drug class decrease brand-name prescriptions to achieve median per prescription cost.
  • Program 4: States decrease brand-name prescriptions to achieve the lowest per prescription cost for each class.

The savings are estimated as follows, with target mean prescription costs being decided by the program of choice:

\begin{align} per\_prescription\_saving(drug\_class, state, program) = &\text{ } mean\_prescription\_cost(drug\_class, state)\\ &- target\_mean\_prescription\_cost(drug\_class, program)\\ savings(drug\_class, state, program) = &\text{ } per\_prescription\_saving(drug\_class, state)\\& *num\_prescriptions(drug\_class, state)\\ total\_savings(program) = & \sum_{drug\_classes} \sum_{target\_states} savings(drug\_class, state, program) \end{align}

The table below shows the estimated savings for each drug class and program. Note that the total savings shown are only from prescriptions for drugs in the top 9 classes among the top 180 drugs, which account for 80% of the cost, and hence provide conservative estimates for potential savings. Also, since the Part D dataset does not include data for cases with 10 or fewer prescriptions, it accounts for about $81 billion of the $103 billion total cost. If we assume that the proportion of drugs is similar in the missing prescriptions, we can use an adjustment factor to estimate the overall savings ('Adjusted total' in the table).

Estimated savings (in $ millions)
Drug class Program 1
Program 2 Program 3 Program 4
Metabolic 93 328 976 3935
Central nervous system 71 188 531 1872
Cardiovascular 87 220 567 1773
Respiratory 15 37 221 1115
Psychotherapeutic 10 91 367 1317
Antineoplastic 23 135 498 2009
Gastrointestinal 84 183 530 1671
Immunologic 7 27 119 1874
Antiinfective 90 144 485 1403
(out of $81 billion)
$0.48 billion $1.35 billion $4.27 billion $16.97 billion
Adjusted total
(for $103 billion)
$0.60 billion $1.69 billion $5.34 billion $21.21 billion

We see that significant cost savings can be achieved even with the modest targets for Program 1 and 2. The higher savings from Program 3 can be considered to be a reasonable medium term target, and savings from Program 4, if not realistically achievable, serve to provide a rough upper bound on potential savings from this approach.

As a reference point, the CMS budget estimates for expected cost savings from specific measures over a ten year period are as shown in the table below.

CMS initiative 2015 savings 2015-2019 2015-2024
Encourage Use of Generic Drugs by Low Income Beneficiaries - $3.02 billion $8.49 billion
Increase Availability of Generic Drugs and Biologics $0.62 billion $4.33 billion $13.11 billion

5. Recommendations

What are the different ways in which the potential cost savings can be achieved?. Primarily, instead of a 'one size fits all' approach to generics vs. brand-names, it will be more effective to design customized initiatives for particular states for particular drug classes. We now present some recommendations for initiatives and interventions based on the insights obtained so far.
  1. Data tool/product development: Combine the insights presented with further analysis and yearly data to develop a comprehensive data tool/product that allows interactive exploration of trends and data visualizations. This will prove useful to policy makers, healthcare mangers, researchers, and the wider public, and can be basis of a number of the recommended engagements discussed below. See prototype.
  2. Law amendments: As mentioned in a 2012 report by the United States Government Accountability Office, state generic substitution laws have a major effect on the rates of generic vs. brand-name substitution. Policy makers in U.S. federal agencies (HHS/CMS) can further investigate any causal relationships of specific state regulations with the results presented here and engage with state law makers to achieve savings and improved outcomes.
  3. Behavioral nudges: As detailed by Cass Sunstein, 2009-2012 Chairman of the U.S. Office of Information and Regulatory Affairs, in his book 'Simpler: The Future of Government', insights from behavioral psychology and economics have been successfully applied to several regulatory processes of the federal government. These processes are designed to `nudge` people into making decisions that lead to better outcomes for both the individual and society at large. The following approaches can be considered:
    • Community level: Share regional trend discrepancies with regional medical bodies and physician associations to promote collective reflection of practices and encourage self-correcting behaviors.
    • Individual level: Send a report (only) to physicians who are prescribing branded drugs at a significantly higher rate than the median (for the same speciality/class of drugs across states, counties or cities) , which includes their rates of prescribing branded vs. generic drugs compared to the rates for others (median rate etc.). A similar technique has been successfully employed by companies and utilities in saving billions of dollars in energy usage, by helping people judge themselves against similar others ( eg. neighbors).
  4. Physician training: Various interventions in physician training can be considered. For instance, it is well known that patient demand is one of the driving factors for higher rates of brand-name drugs (which is in turn driven by direct-to-patient pharma advertising). In addition to being educated about the regional trends and encouraged for self-correction, physicians may need specific training for how to deal with patient demands for brand name drugs.
  5. Further research can be conducted to better understand causality in prescription rates, a few examples as follows:
    • Surveys of physicians in high brand prescription areas: do they believe brand drugs are more effective, there is a lack of equivalent generic alternatives, or are driven by patient demand.
    • Surveys of patients in high brand prescription areas: did they talk to your doctor about/ask for brand name drugs?

6. Looking ahead

The Part D dataset allows for several further analyses:
  1. Provider/Physician trends, comparison, anomalies
  2. City-wise, county-wise trends
CMS can consider the following data improvements:
  1. Include drug class information in future releases
  2. Provide data for several recent years to understand and track trends over time

7. References

  1. Medicare Provider Utilization and Payment Data: Part D Prescriber. United States Center for Medicare and Medicaid Services, April 2015.
  2. Fact Fact Sheet: CMS releases prescriber-level Medicare data for first time. United States Center for Medicare and Medicaid Services, April 2015.
  3. Medicare Fee-For Service Provider Utilization and Payment Data Part D Prescriber Public Use File: A Methodological Overview. United States Center for Medicare and Medicaid Services, April 2015.
  4. Medicare Advantage/Part D Contract and Enrollment Data. United States Center for Medicare and Medicaid Services.
  5. CMS Medicare 2015 Budget Overview. United States Office of Budget, June 2014.
  6. Medicines Use and Spending Shifts: A Review of the Use of Medicines in the U.S. in 2013. IMS Institute for Healthcare Analytics, April 2014.
  7. Medicines Use and Spending Shifts: A Review of the Use of Medicines in the U.S. in 2014. IMS Institute for Healthcare Analytics, April 2015.
  8. Effects of Using Generic Drugs on Medicare's Prescription Drug Spending, 2010. United States Congressional Budget Office. September 2010.
  9. Drug Pricing: Research on Savings from Generic Drug Use, 2012. United States Government Accountability Office, January 2012.
  10. A. Laskey. How behavioral science can lower your energy bills. TED Talk, February 2013.
  11. C. Sunstein. Simpler: The Future of Government. Simon and Schuster, April 2013.