Cluster Analysis of PrescriberLevel Messaging Systems Using PatientLevel Data

1
Cluster Analysis of Prescriber-LevelMessaging
Systems Using Patient-Level Data

• Author Jody Fisher
• Senior Director, Product Management
• Verispan, LLC
• May 24, 2004

2
Concepts

• This presentation will demonstrate a unique and
  progressive method for segmenting physicians.
• The proposed method for changing messaging to
  physicians utilizes patient-centric data to
  cluster and evaluate physician types.
• Proposed concept serves to place reps in more
  optimal place with more optimal message.
• Recent information suggest physicians require
  more customized message.

3
Physicians Requiring Specialized Message
According to a recent survey conducted by
Accenture, physicians want more detailed,
comparative and customized information from
pharmaceutical sales reps. Primary care
physicians reported that customized content,
clinical evidence and comparative analyses of
medicines rank at the top of the list of
information theyd like to receive from
reps. www.eyeforpharma.com, August 4, 2003
4
Current Sources vs. Patient-Centric Data

• Current targeting typically use Rx share and
  volume baselines to evaluate physician value
• Over 80 of NRx are not really true new
  patients.

5
The patient-centric world
New patient is put on Brand X
New Patient Rx
Patient on Brand X for one year runs out of
refills, gets new Rx from Dr.
Continuing Patient Rx
Patient on Brand Y switches to Brand X
Switch Patient Rx
Cholesterol Market 11M TRx
Cholesterol Market 11M TRx
6
Typical Patient-Centric Definitions

• New Patient Prescription filled no prior
  market therapy in lookback period.
• Continuing Patient Prescription filled
  evidence patient has taken same brand having been
  taken in lookback period.
• Switch/Add Prescription filled first time
  taking therapy evidence patient has taken
  different brand within market.
• Requires market knowledge temporal relationship
  construction.
• but once applied per transaction can be
  evaluated at any geographic level
  nationalthrough physician.

7
Case Study Physician Segmentation

• Study will demonstrate that patient-centric
  principles can be applied to evaluate appropriate
  message to provide to physicians.
• Data source used longitudinal prescription
  database used
• 2B prescriptions received every year from all
  pay-types across nearly 100 of stores.
• Data can be attributed to distinct physician
• Every claim is classified as New, Continuing, or
  Switch/Add based on longitudinal classification
• Data is projected at transaction level and
  aggregated to physician level.

8
Case Study Background

• Market 1B retail sale market focus on 3
  competitive drugs.
• Two drugs newer than one other drug (Product C)
• Drugs used equally by primary care physicians and
  specialists.
• Estimated 6,700 sales representatives detailing 3
  competitive products during 1Q - 2004.
• Estimated 150M in detailing, event, journal, and
  ePromotion for 3 products during 1Q 2004.

9
Cluster Analysis Segmentation

• Cluster Analysis is commonly used to link
  entities with like-attributes together.
• K-means Cluster Analysis forces K number of
  clusters to be created clustering an art not
  science
• Cluster physicians not on Rx types but on
  patient-centric measures.
• Patient-centric Assumptions
• New Continuing Switch/Add TRx
• 6 Month Lookback to determine type

10
K-Means Cluster Solution

• Model evaluates physicians on following criteria
• Total Rx Volume high writers vs. low writers
• New Patients probability to start patients on
  first-line therapy
• Switch Patients probability to switch
  patients of overall writing.
• Product C TRx Share probability that
  physician writes existing product
• Product Switch Away from Product C Of
  switches
• Ran specialist data for 1Q 2004 used
  physicians with xgt10 TRx 7,300 specialists
  segmented
• Produced K 2 through K 6 cluster solutions

11
Solution Results
3 cluster solution and 5 cluster solutions
yield most significantly distinct clustering
solutions
12
3 Cluster Solution Results
Forcing 3 clusters yields the following physician
groups Cluster 1 (Converted) High volume, low
Product C Share, lowest new start rate.
Cluster 2 (Loyalist) Average volume, high
Product C Share, new starts, no
switching. Cluster 3 (Aggressors) Lowest
volume but starts greater percentage of new
patients and by far most willing to switch
patients.
13
5 Cluster Solution Results
Forcing 5 clusters yields the following physician
groups Cluster 1 (Converted) High volume, low
Product C Share, lowest new start rate lower
new start rate. Cluster 2 (Switchers) Low
volume but switching most patients away from
Product C. Cluster 3 (Vanilla) No outstanding
attributes Cluster 4 (New Starters) Low volume
but percent new starts the highest virtually
no switching. Cluster 5 (Loyalists) High
Product C share low new starts, low switching
14
Message Potentials Switch v. Share
High Share Lo Switching Potential Message
Less Competitive Focus on Dosing and
Indications Goal Drive 1st Line Therapy
High Share Hi Switching Potential Message All
messages but rotate! Goal Dont bore doctor.
Switch Business
Lo Share Lo Switching Potential Message More
Competitive Focus on Dosing and
Indications Goal Drive more product C writing
Lo Share Hi Switching Potential Message
Competitively focused Long-term benefits Goal
Provide less incentive to continue switching
Product C Share
15
Graphically3 clusters
3,738 Docs 84.1 Share 1.15 Switch
573 Docs 62.8 Share 11.4 Switch
Switch Business
3,039 Docs 40.6 Share 1.54 Switch
Product C Share
16
Graphically5 clusters
Difficult to represent all axes Clusters 4 and
5 really differ on probability to start new
patients tweak message accordingly
2,372 Docs 88.5 Share 0.65 Switch
929 Docs 76.4 Share 0.71 Switch
1,135 Docs 64.6 Share 6.31 Switch
Switch Business
130 Docs 59.4 Share 21.2 Switch
2,784 Docs 40.1 Share 1.23 Switch
Product C Share
17
Next Steps

• Test other statistical clustering/grouping
  methods to try to better classify groups
• Work to remove multi-co linearity/improve
  variance explanation
• Factor Analysis
• Latent Class Analysis
• Test other temporal relationships for patient
  model
• Other categories (restarts/switch vs. adds)
• Different amount of patient histories
• Test many other factors for cluster inputs
• Method of payment factors
• Promotional indices
• Indication specific influences.
• Test validity on markets with less competition
  different goals.

18
Recommendations

• Recommend providing limited metrics at a
  prescriber level to sales reps.
• Recommend that recommended message solution be
  calculated in home office and provided to
  representative.
• Recommend frequent adjustment of
  physician/message combinations quarterly?
• Recommend that targeting approach be reviewed
  against compensation structure to ensure maximum
  compliance.

19
Sample Report
Explain the new indication and ask if they have
any questions.
20
Thank You For Your Attention

• Any Questions?