The Diversity of Concentrated Prescribing Behavior: An Application to Antipsychotics

Anna Levine Taub, Anton D. Kolotilin, Robert S. Gibbons and Ernst R. Berndt

Special thanks to the IMS Health Services Research Network for data support

The usual disclaimers apply. This research has not been sponsored.

Scenario: Physician sees patient…

• Patient has confirmed diagnosis for which alternative FDA approved treatments are available, but patient response (efficacy, side effects) is idiosyncratic and unpredictable

• Little evidentiary literature available for physician to base ex ante treatment choice

• How best can physician learn regarding efficacy and tolerability of alternative drugs?

• Concentrate or diversify on drug treatment? What complementary actions can be utilized?

Issues We Address

• Trade-off of exploiting or exploring, concentrating or diversifying, prescribing “ready-to-wear” vs. “custom-made” treatments (Frank-Zeckhauser 2007)

• How does physician’s location along the treatment diversity continuum vary by physician’s specialty, patient volume, age, gender and training?

• If physicians concentrate, is there convergence and unanimity on choice of favorite drug, or is concentration non-uniform? If set of drugs prescribed is diverse, does physician’s distribution mimic regional or national shares? How to model?

Related Literature and Findings• Frank and Zeckhauser [2007]: 1372 PCPs – find that

share of most prescribed medicine to treat four acute and five chronic conditions is high – average 60% -- but 13% less for chronic conditions

• Patient clinical factors played a “startlingly minor role”, physicians practiced “ready-to-wear” treatment norms, sometimes “sensible” but in other cases “disturbing”

• We focus on antipsychotics that treat primarily chronic conditions, and examine prescriber behavior across a much larger number of prescribers and variety of specialties, but only have data on prescribers, not data on the patients they treat (unlike Frank-Zeckhauser)

Outline of Remaining Presentation

• Alternative antipsychotics for treatment of schizophrenia and related conditions (brief)

• Preliminary evidence on concentrated vs. diversified prescribing behavior (the setting)

• A model of prescriber learning and treatment behavior, and some predictions of the model

• Empirical findings on consistency of model predictions with observed prescriber behavior

• Limitations and discussion

Schizophrenia• Incurable mental illness, positive and negative

symptoms, 1-2% prevalence, strikes males in late teens and early 20s, females slightly later, unknown etiology (genetics?)

• Patients experience unemployment, lose family contact, become homeless, substantial portion experience incarceration for some time periods

• Goals of treatment are to reduce/eliminate symptoms, maximize quality of life and adaptive functioning, maintain recovery to maximum extent possible

• In US, Medicaid is the largest payer of medical and drug benefits to people with schizophrenia (Duggan)

Available Treatments for Schizophrenia

• From 1955 thru early 1990s, pharmacological treatments called typical, conventional, and/or neuroleptic antipsychotics – better for +ve than –ve symptoms, but for some patients have lasting extrapyramidal side effects (e.g., tardive dyskinesia) for which there’s no treatment

• Clozaril (clozapine) introduced in 1989, still considered most effective antipsychotic drug, but for 1-2% side effect of agranulocytosis (can be lethal), so use requires frequent white blood cell count monitoring

• Clozaril dubbed first generation atypical drug – “FGA”

Second Generation Atypicals (SGAs)

• Five SGAs introduced in US 1993-2002– Risperdal (risperidone) – 1993– Zyprexa (olanzapine) – 1996– Seroquel (quetiapine) – 1997– Geodon (ziprasidone) – 2001– Abillify (aripiprazole) -- 2002

• Initially SGAs perceived as similar to typicals for +ve but better for –ve symptoms, without typicals’ extrapyramical side effects, but since 2001-2 concern has grown over SGA metabolic syndrome and weight gain side effects

• Typicals and clozapine now off patent, SGAs very costly• CATIE [2005] and related studies questioned relative

efficacy and cost-effectiveness of atypicals vs. typicals

Typical and Atypical Prescriptions Annually 1996-2007

Total Prescriptions by Year

0

500,000

1,000,000

1,500,000

2,000,000

2,500,000

3,000,000

1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007

Atypical Antipsychotics Typical Antipsychotics

Preliminary Evidence on Concentration

• By 2007, five years after launch of last SGA, among those writing at least one antipsychotic Rx, mean share of prescriber’s favorite drug was 66% (greater than in Frank-Zeckhauser [2007]).

• However, concentration is diverse – when we limit 2007 sample further to highly concentrated prescribers (> 75% favorite ), 55% chose Seroquel as favorite drug, 28% Risperdal, 13% Zyprexa, 3% Abilify, 2% Geodon and 0.4% clozapine.

• 2007 respective national atypical market shares were 36%, 27%, 13%, 14%, 7% and 2%.

Regional Geographic Variations as in Dartmouth Studies? No!!!

• Compute Herfindahl measure of concentration HHI = Σjsj

2, 0 ≤ sj ≤ 100, so HHI ranges up to 10,000 for a monopoly (if ten drugs had equal shares, HHI would be 10*(10)2 = 1000)

• Calculate HHI at various levels -- individual prescriber, county, HRR (hospital referral region, Dartmouth Atlas Project), state and nation

• Also compute coefficient of variation (COV), with Phelps [1992] saying COV in 0.1 – 0.2 range reveals “low variability”, and > 0.4 “highly variable” practice of medicine

Antipsychotic HHIs in 2007 at Alternative Geographic AggregatesMeans, Standard Deviations and Coefficients of Variation

Geographical Aggregate Mean HHI Std. Dev. Coef. Var. NIndividual Prescriber 4946 2499 0.505 19537County 3234 1773 0.548 1904Hospital Referral Region 1989 359 0.180 306State (plus District of Columbia) 1859 16 0.008 51Nation 1825 na na 1

Since the prescriber data indicate that there’s little variability in HHI at the geographic HHR level and higher (unlike Wennberg and colleagues), whereas there is very substantial heterogeneity at the level of the individual prescriber, here we will focus on factors affecting prescribing behavior of individual prescriber. Will look at county “spillovers” at future date.

How can we explain the wide variance in the way physicians respond to patients?

• Perception (Don’t think we’re behind.)– Differing mental models / categories / priors– Example: Two physicians read different articles arriving at different

conclusions• Motivation (Don’t want to do it.)

– Weak competition, agency issues– Example: MD - Pharma contacts are influential and heterogeneous

• Administration (Don’t know to get there.)– Know where you want to go but can’t get there due to financial or

administrative constraints– Limited formulary coverage

• Inspiration (Don’t know what to do.)– Bandit problem, learning model– Don’t know about the efficacy of the drug, complex treatment

regime and need to learn how to use the drug effectively

How can we explain the wide variance in the way physicians respond to patients?

• Perception (Don’t think we’re behind.)– Differing mental models / categories / priors– Ex. Two doctors read different articles arriving at different

conclusions• Motivation (Don’t want to do it.)

– Weak competition, agency issues– Ex. Phy - Pharma contacts are influential and heterogeneous

• Administration (Don’t know to get there.)– Know where you want to go but can’t get there due to financial or

administrative constraints– Limited formulary coverage

• Inspiration (Don’t know what to do.)– Bandit problem, learning model– Don’t know about the efficacy of the drug, complex treatment

regime and need to learn how to use the drug effectively

How can we explain the wide variance in the way physicians respond to patients?

• Perception (Don’t think we’re behind.)– Differing mental models / categories / priors– Ex. Two doctors read different articles arriving at different

conclusions• Motivation (Don’t want to do it.)

– Weak competition, agency issues– Ex. Phy - Pharma contacts are influential and heterogeneous

• Administration (Don’t know to get there.)– Know where you want to go but can’t get there due to financial or

administrative constraints– Limited formulary coverage

• Inspiration (Don’t know what to do.)– Bandit problem, learning models– Don’t know about the efficacy of the drug, complex treatment

regime and need to learn how to use the drug effectively

How can we explain the wide variance in the way physicians respond to patients?

• Perception (Don’t think we’re behind.)– Differing mental models / categories / priors– Ex. Two doctors read different articles arriving at different

conclusions• Motivation (Don’t want to do it.)

– Weak competition, agency issues– Ex. Phy - Pharma contacts are influential and heterogeneous

• Administration (Don’t know to get there.)– Know where you want to go but can’t get there due to financial or

administrative constraints– Limited formulary coverage

• Inspiration (Don’t know what to do.)– bandit problem, learning models– Don’t know about the efficacy of the drug, complex treatment

regime and need to learn how to use the drug effectively

Learning to Concentrate• Patients p = 1, 2, … arrive sequentially• Symptoms s = 1, …, S randomly drawn• Drugs d = 1, …, D d*(s)• Time btwn patients w Arrival at time 0, w, 2w, …• Discount rate r Continuous time

• Maximum potential benefit of drug d for symptom s = Bsd o Example: Bsd*(s) > Bsd for all d ≠ d*(s)

• Therapy = drug (d) and unobserved complementary actions (a) o Ideal effectiveness requires ideal complementary actions (uncertain)

o Ex. Dosage of the drug (titrating), actions that effect adherence (Management and Communication about possible side effects and their duration)

• Realized effectiveness = bsdp = Bsd[1 – (a – xdp)2]o xdp = θd + εdp where θd and εdp are independently normal for all d,p

• Provider observes xdp , learns about θd17

Learning to Concentrate: Intuition• The physician learns by combining different complementary

actions a when prescribing drug d, and observes how patients respond xdp.

• The best action that the physician can potentially learn to make, θd, depends only on the drug prescribed

• Symptoms determine which drug has the highest potential for giving a patient the best outcomes, d*(s).

• Speed of learning the complementary action θd for each drug d depends only on how often the physician prescribes drug d

18

Learning to Concentrate

• The physician’s dynamic strategy is to choose a drug, dp(s, hp-1), and complementary actions, ap(d, hp-1), for each patient p with symptom s and each history hp-1, where history is determined by the physician’s strategy implementation in prior patient encounters.

• Assuming that the physician cares only about maximizing the expected discounted patient benefit of all patients s/he expects to see over time in the future, the physician’s optimization problem, provided that e-rw>0, is to solve:

]|[[ 1sdp1()

max

pp

rwp

d

hbEeEp

Learning to Concentrate: Intuition Cont’d• Suppose first that w is large, i.e., patients are infrequently seen by the

physician (the physician is a small volume prescriber): – In this case over time the physician will concentrate on only a subset of drugs.– Number and identity of the drugs for which the physician concentrates depend on the

initial history of symptom presentation to the physician -- idiosyncratic.

• As w is decreased (i.e., the volume of patients seen by the physician increases):

– Physicians will have a larger incentive to invest in learning how to use new or different drugs effectively, future patients will benefit.

– The set of drugs a physician uses will still depend on the initial history of symptoms of the patients the physician has seen; however, this dependence will be weaker.

• As w decreases to zero (the physician sees patients very frequently – almost continuously):

– The physician will eventually learn a great deal about optimal complementary actions θd for each drug d in D and prescribe d*(s) for every s.

– To the extent the symptom set s presented to the physician is heterogeneous, given patient volume so too will be the range of drugs prescribed.

Predictions From This Model

• Concentration decreases with volume• Variability of treatment regimes from national

norms is greater among low volume prescribers (greater dependence on idiosyncratic random history)

• Prescribers with higher discount rates have more concentrated prescribing behavior (those near retirement or likely to leave the labor force?)

Limitations of This Model

• Physicians know Bsd, maximum potential benefit of each drug d given symptoms s, but don’t know optimal complementary actions, and learn only from their own experience with drug d treating symptoms s

• Don’t learn about drugs not prescribed, nor infer anything from national market shares – richer model would include costs and benefits of learning about a drug from other sources

Physician Prescribing Data

• IMS Health Xponent™ monthly data tracks prescriptions dispensed at retail and mail order by NDC code, links to prescriber ID, and links to American Medical Association directory of physicians

• We take 10% random sample of all prescribers who wrote at least one antipsychotic Rx in 1996, refresh each year with 10% random sample of “new” prescribers, but here we only utilize 2007 data – five years since introduction of last new antipsychotic

• Only data on prescriber, no data on patients

Five Prescriber Groups

• Primary care physicians (PCP) – internal medicine, family medicine and practice, pediatrics, and general practice prescribers

• Psychiatrists (PSY) – general, child-adolescent and geriatric psychiatry

• Neurologists (NEU) – general, geriatric and child neurologists

• Non-physicians (NP) – primarily nurse practitioners and physician assistants

• Other (OTH) – all other prescribers

Mean Values of Characteristics of 2007 Prescriber Sample, by Prescriber Specialty

Specialty Group

Number of Prescribers

Antipsychotic Annual Rx

Atypical Annual Rx

No. Distinct Antipsychotics

No. Distinct

Atypicals

Antips-ychotic

HHIAtypical

HHI

PSY 3463 609.56 551.37 7.46 4.71 3,245 3,661

NEU 728 97.64 82.30 3.60 2.33 5,657 7,025

PCP 9544 68.03 52.82 4.31 2.73 4,612 5,915

OTH 4161 54.24 29.53 2.76 1.65 6,912 7,081

NP 1641 174.85 155.30 4.29 2.88 5,181 5,633

Prescriber Data

• AMA directory of physicians does not have any data on NPs

• For MDs/DOs, have data on age (divide sample into quartiles), specialty, hospital or office based, zip code of practice, group or solo practice, population size of county in which physician practices, gender, and whether physician allows prescribing behavior data to be used by pharmaceutical or other for-profit organizations

Additional Measure of Diversity

• Theory suggests we not only look at concentration, but also diversity of concentration (should decrease with volume)

• Physician i prescribing drug j in region r – share of Rx’s for j is sijr, whereas regional share for j is mjr. Our measure of deviance Dijr is:

Dijr = Σj (sijr – mjr)2

Alternative Dependent Variables

• Various dependent variables at level of individual prescriber – number of distinct antipsychotic drug molecules prescribed (Poisson), number of distinct atypical molecules prescribed (Poisson), share of all antipsychotic Rxs that are atypicals (Tobit), log antipsychotic HHI (Tobit), log atypical HHI (Tobit), and log deviance (OLS)

• Drop NPs from sample, yields n = 17,652

Common Set of Explanatory Variables in Regression Equations

• Various dependent variables are function of (where bold denotes reference case) – Age quartiles (< 43, 43-50, 51-58, 59+ )– Volume interacted with specialty (OTH, PCP, PSY

and NEU)– County population (< 150K, 150K-500K, 500K-1M,

and > 1m)– Male/Female, Group/Solo practice,

Office/Hospital base, MD/DO flag and Physician Rx Data Can Be Used/Opt Out

Summary StatisticsVariable Obs Mean Std. Dev. Minimum MaximumDeviation of Physician's Antipsychotic prescribing from HRR Shares 17,652 2,660 2,441 5 10,321Deviation of Physician's Antipsychotic prescribing from National Market Shares

17,652 2,735 2,499 30 10,051

HHI of Individual Physician's Antipsychotic Prescribing 17,652 4,920 2,484 1,196 10,000HHI of Individual Physician's Atypical Prescribing 16,262 5,708 2,498 1,701 10,000% of Prescriptions for Antipsychotics that were for Atypicals 17,652 71.46 32.60 0 100Number of Different Antipsychotics Prescribed 17,652 4.54 2.70 1 17Number of Different Atypicals Prescribed 17,652 2.85 1.64 0 6Total Yearly Antipsychotic Prescriptions 17,652 171.80 431.35 12 7,186Total Yearly Atypical Antipsychotic Prescriptions 17,652 145.92 388.51 0 6,780Prescriber Age 17,652 50.37 10.80 26 92PCP 17,652 0.54 0.50 0 1PSY 17,652 0.19 0.40 0 1NEU 17,652 0.04 0.20 0 1OTH 17,652 0.23 0.42 0 1Solo Practice 17,652 0.20 0.40 0 1Population (county) 17,652 1,065,738 1,810,008 1,299 9,734,701Female 17,652 0.26 0.44 0 1Hospital Based Physician 17,652 0.08 0.27 0 1DO Flag 17,652 0.09 0.28 0 1Physician Opt Out 17,652 0.03 0.18 0 1

Results: Variety of Drugs Prescribed• Number distinct antipsychotics – Given volume,

PSY>PCP>NEU>OTH; effect of volume +ve for OTH, +ve but very small for PSY, with PCP and NEU in between; effect of age +ve for two middle quartiles, else zero; female –ve as is population, but flat after 500K; solo, hospital and MD opt out insignificant, but DO +ve

• Number distinct atypicals – same results except –ve for oldest age quartile (cohort effect?)

• Interpretation: PSY training a substitute for volume experienced by other specialties?

Poisson Regressions: Number of Distinct Antipsychotic Molecules and Distinct Atypical Molecules prescribed in 2007

Number of Distinct Antipsychotics Number of Distinct Atypicals

Coefficient Standard

Error P>|z| Coefficient Standard Error P>|z|

Total Yearly Antipsychotic Prescriptions 0.001039 0.000048 <.001 0.001059 0.000061 <.001

PCP*Total Yearly Antipsychotic Prescriptions -0.000366 0.000051 <.001 -0.000405 0.000066 <.001

PSY*Total Yearly Antipsychotic Prescriptions -0.000792 0.000049 <.001 -0.000937 0.000062 <.001NEU*Total Yearly Antipsychotic Prescriptions

-0.000472 0.000071 <.001 -0.000478 0.000089 <.001

Age Quartile 43-50* 0.0374 0.0102 <.001 0.0113 0.0127 0.375Age Quartile 51-58* 0.0513 0.0101 <.001 0.0152 0.0126 0.230Age Quartile 59+* 0.0186 0.0107 0.081 -0.0423 0.0134 0.002PCP* 0.4520 0.0115 <.001 0.5093 0.0148 <.001PSY* 0.8965 0.0129 <.001 1.0558 0.0166 <.001NEU* 0.2608 0.0233 <.001 0.3502 0.0292 <.001Female* -0.0899 0.0084 <.001 -0.0308 0.0105 0.003Population 150,000-500,000 (county)* -0.0357 0.0099 <.001 -0.0460 0.0126 <.001Population 500,000-1,000,000 (county)* -0.0742 0.0105 <.001 -0.0900 0.0132 <.001Population more than 1,000,000 (county)* -0.0692 0.0101 <.001 -0.0868 0.0127 <.001Solo Practice* -0.0019 0.0091 0.831 -0.0086 0.0115 0.452Hospital Based Physician* 0.0085 0.0125 0.497 -0.0124 0.0160 0.437DO Flag* 0.0334 0.0129 0.010 0.0518 0.0161 0.001Physician Opt Out* 0.0120 0.0190 0.530 0.0348 0.0238 0.144Cons 0.9876 0.0141 <.001 0.4958 0.0179 <.001Number of Observations= 17,652 Pseudo R^2= 0.145 Pseudo R^2= 0.1045

Results: Old vs. New Antipsychotic Drugs

• Share of antipsychotic Rx’s that are atypical (new): Marginal effects evaluated at variable means – Given volume, PSY>NEU>PCP>OTH; effect of volume insignificant for OTH and NEU, +ve for PCP, and –ve but small for PSY; three older age quartiles –ve, especially oldest; female +ve and large, population monotonically -ve as is hospital based and MD opt out, but solo and DO flag not significant

• Greater use of older antipsychotics in more populous areas, particularly by PSYs – specialization in more urban areas serving heterogeneous populations?

Tobit Regression (Marginal Effects Estimated at Variable Means) on Percent of All Antipsychotic Prescriptions written for Atypicals in 2007

dy/dx Standard Error P>|z| Mean Value

Total Yearly Antipsychotic Prescriptions 0.0108 0.0061 0.074 171.80

PCP*Total Yearly Antipsychotic Prescriptions 0.0154 0.0070 0.028 36.44

PSY*Total Yearly Antipsychotic Prescriptions -0.0155 0.0061 0.011 118.84

NEU*Total Yearly Antipsychotic Prescriptions 0.0032 0.0094 0.737 3.99

Age Quartile 43-50* -1.77 0.760 0.020 0.252

Age Quartile 51-58* -1.42 0.761 0.063 0.262

Age Quartile 59+* -2.00 0.810 0.014 0.222

PCP* 17.02 0.793 <.001 0.536

PSY* 44.72 1.035 <.001 0.194

NEU* 27.66 1.660 <.001 0.041

Female* 6.68 0.640 <.001 0.263

Population 150,000-500,000 (county)* -3.55 0.760 <.001 0.262

Population 500,000-1,000,000 (county)* -5.65 0.796 <.001 0.225

Population more than 1,000,000 (county)* -7.30 0.769 <.001 0.264

Solo Practice* 0.310 0.682 0.650 0.199

Hospital Based Physician* -2.67 0.995 0.007 0.081

DO Flag* 0.103 0.970 0.915 0.086

Physician Opt Out* 2.74 1.486 0.065 0.034

Number of Observations= 17,652Pseudo R^2= 0.017Left Censored=0 Right Censored=3,353* dy/dx is for a discrete change of a dummy variable from 0 to 1

Prescriber Concentration: HHI for All Antipsychotics and Just Atypicals

• For all antipsychotics, given volume, HHI is lowest for PSY<PCP<NEU <OTH; effect of volume is –ve for OTH, PCP and NEU, but essentially zero for PSY; age quartiles U-shaped; population effect +ve but peaks at 1m; female +ve, others not significant

• For atypicals, similar results except oldest age quartile has large +ve effect unlike younger quartiles (no effect), while DO, MD opt out -ve

Tobit Regressions (Marginal Effects Evaluated at Variable Means) on Log (Antipsychotic Prescription HHI for 2007) and Log

(Atypical Antipsychotic Prescription HHI for 2007)Antipsychotic HHI Atypical HHI

dy/dx Standard

Error P>|z| Mean Value dy/dx Standard

Error P>|z| Mean Value

Total Yearly Antipsychotic Prescriptions -0.000821 0.000080 <.001 171.80 -0.001026 0.000080 <.001 182.67PCP*Total Yearly Antipsychotic Prescriptions -0.000149 0.000090 0.089 36.44 -0.000051 0.000090 0.557 39.04PSY*Total Yearly Antipsychotic Prescriptions 0.000626 0.000080 <.001 118.84 0.000859 0.000080 <.001 128.99NEU*Total Yearly Antipsychotic Prescriptions 0.000359 0.000120 0.002 3.99 0.000384 0.000110 0.001 4.30Age Quartile 43-50* -0.029 0.010 0.002 0.252 0.002 0.009 0.869 0.252Age Quartile 51-58* -0.039 0.010 <.001 0.262 0.008 0.009 0.424 0.262Age Quartile 59+* -0.019 0.010 0.067 0.222 0.055 0.010 <.001 0.223PCP* -0.435 0.010 <.001 0.536 -0.206 0.011 <.001 0.563PSY* -0.761 0.013 <.001 0.194 -0.708 0.013 <.001 0.210NEU* -0.238 0.021 <.001 0.041 -0.011 0.021 0.592 0.043Female* 0.057 0.008 <.001 0.263 0.034 0.008 <.001 0.267Population 150,000-500,000 (county)* 0.020 0.010 0.037 0.262 0.013 0.009 0.159 0.261Population 500,000-1,000,000 (county)* 0.040 0.010 <.001 0.225 0.052 0.010 <.001 0.225Population more than 1,000,000 (county)* 0.032 0.010 0.001 0.264 0.047 0.010 <.001 0.262Solo Practice* -0.005 0.009 0.582 0.199 0.017 0.009 0.046 0.201Hospital Based Physician* 0.019 0.013 0.119 0.081 -0.003 0.013 0.805 0.079DO Flag* -0.023 0.012 0.056 0.086 -0.029 0.012 0.014 0.088Physician Opt Out* -0.035 0.019 0.056 0.034 -0.037 0.018 0.043 0.035

Number of Observations= 17,652 Number of Observations= 16,262Pseudo R^2= 0.218 Pseudo R^2= 0.22

Left Censored=0 Right Censored=1,571 Left Censored=0 Right Censored=2,606* dy/dx is for a discrete change of a dummy variable from 0 to 1

Deviance from National and HRR Market Share Norms

• Similar results for national and HRR shares – only show national share results

• Given volume, deviance of PSY<PCP<NEU <OTH; effect of volume –ve for OTH and PCP, less so for NEU and especially PSY; effect of age quartiles U-shaped; population effect +ve but peaks at 1m; female effect +ve; other variables not significant

Linear Regression on Log (Deviance in Physician Antipsychotic Prescribing from National Market Shares)

Coefficient Standard Error P>|z|

Total Yearly Antipsychotic Prescriptions -0.00131 0.00014 <.001

PCP*Total Yearly Antipsychotic Prescriptions -0.00079 0.00016 <.001

PSY*Total Yearly Antipsychotic Prescriptions 0.00079 0.00014 <.001

NEU*Total Yearly Antipsychotic Prescriptions 0.00059 0.00022 0.008

Age Quartile 43-50* -0.013 0.018 0.456

Age Quartile 51-58* -0.007 0.018 0.674

Age Quartile 59+* 0.054 0.019 0.005

PCP* -0.661 0.019 <.001

PSY* -1.458 0.024 <.001

NEU* -0.499 0.039 <.001

Female* 0.045 0.015 0.003

Population 150,000-500,000 (county)* 0.021 0.018 0.228

Population 500,000-1,000,000 (county)* 0.101 0.019 <.001

Population more than 1,000,000 (county)* 0.085 0.018 <.001

Solo Practice* 0.012 0.016 0.469

Hospital Based Physician* 0.025 0.023 0.283

DO Flag* -0.036 0.023 0.110

Physician Opt Out* -0.052 0.035 0.134

Cons 8.204 0.023 <.001

Number of Observations= 17,652 R^2= 0.345* indicates dummy variable

Robustness Checks

• Results essentially unchanged when we add a quadratic term in volume, and when adding volume quadratic interacted with specialties – so volume-specialty findings are not due to simple nonlinearity

• If we restrict sample to 3K+ PSYs, still get same volume, age and gender results, but not as significant, and other variables not as significant

• When 305 HRR dummies are added, while jointly highly significant, increase in R2 de minimus -- <4% -- variability is at prescriber level, not at HRR level

Discussion: I• Our results largely complement and extend those reported by

Frank-Zeckhauser [2007] – here a chronic maintenance medication, a greater range and number of specialty prescribers – rational prescribing and “sensible use of norms”

• Our model entirely ignores learning from others, spillovers, and is inconsistent with herding behavior by physicians, as modeled and estimated in Chandra and Staiger [2007]

• Results different from large Dartmouth Atlas HRR and state level literature – pharmaceuticals differ from procedures (Zhang, Baicker, Newhouse [2010]); non-Medicare differs from Medicare (Rettenmeier- Saving [2009]), and under age 65 from 65+ (Gawande [2009] and Franzini, Mikhail and Skinner [2010]) in Medicare vs. private insurance – McAllen vs. El Paso

Discussion: II

• Higher volume physicians use wider variety of drugs, mimic more closely national norms – have most to learn and benefit from exploring vs. exploiting

• Oldest quartile and female MDs most concentrated – discount future at greater rate?

• Volume matters most for PCPs and those trained in specialties not typically prescribing antipsychotics in large volumes, and least for PSY – specialty training and volume prescribing are substitutes for learning about drug characteristics

Discussion: III• Deviance from national norms also decreases with

volume• Major limitation is that we don’t observe patient data

– but finding of dominant role of physician over patient is frequent (Frank-Zeckhauser [2007], Hellerstein [1997], Zhang, Baicker and Newhouse [2010], Schneeweis, Glynn, Avorn and Solomon [2005])

• How generalizable to other medication classes? Non-physician prescribers?

• Our 2007 analysis is a single cross-section

Discussion: IV

AtypicalMarket Share 2002 2008Seroquel 21% 37%Abilify 0 16Geodon 4 7Risperdal 35 26Zyprexa 34 12Other 6 2

Who switched most rapidly? High or low volume prescribers? PCPs or specialists? Old vs. young, male or female, group or solo practice MDs? Variation across HRRs? Who responds most/least to published clinical trial results, FDA warnings?