Micro-Segmentation of Doctors: A Practical Guide

Why decile targeting breaks

Worth being precise about, because deciles are not stupid — they are a reasonable answer to a question that has since changed.

Decile targeting ranks prescribers one to ten by total prescriptions in a therapy area over a rolling period, and concentrates effort on the top bands. It is the oldest framework in commercial pharma and still the most common. It works because past volume is genuinely predictive of future volume, and because it is simple enough for a field force to act on.
The three failures, in the US where the data exists

The consequence is measurable. Poorly targeted programmes have been shown to waste a meaningful share of spend on misidentified prescribers, while well-executed targeting produces new-to-brand lift in the range of 5 to 15 percentage points over unexposed controls. That range is a US benchmark on US data, and it is quoted here as a direction of travel rather than as something an Indian team should expect to reproduce.

Why it breaks completely in India

In India, the inputs that make decile targeting possible largely do not exist at prescriber level. The dominant commercial data infrastructure — secondary sales audit, stockist and retail panel data from providers such as AIOCD AWACS — operates at distributor, stockist and chemist level, not at doctor level. Sales are visible by territory and by outlet, not by prescriber. So a decile built in India is usually not built from that doctor's prescriptions at all; it is built from a representative's estimate of that doctor's prescriptions, which is a different instrument with different failure modes.
This is the single most important thing to understand before designing an Indian segmentation, and it is absent from essentially all published guidance. The practical consequences are worth stating plainly.

• Rep-rated potential is subjective and gameable. A rating that influences territory targets and call plans is a rating the rater has an interest in. Ratings tend to correlate with relationship strength and with effort already invested, not with untapped opportunity.
• It is systematically conservative about doctors you do not yet know. A rep cannot rate a doctor they have never met, so unknown doctors default to low ratings, and low ratings mean no calls. The segmentation quietly encodes and perpetuates the existing coverage pattern.
• It cannot detect movement. The NBRx signal that makes US behavioural segmentation powerful — a doctor who just started prescribing — has no direct Indian equivalent in most therapy areas.
• It is not auditable. When a target list is questioned, "the rep said so" is not a defensible basis, and under UCPMP scrutiny that matters more than it used to.

None of this means Indian segmentation is hopeless. It means the potential dimension must be modelled from proxies rather than measured from scripts — which is a solvable problem, and the subject of the next section.

What you actually have — the Indian signal inventory

Before choosing a technique, inventory the signals honestly. Most Indian brand teams have more than they use and less than the literature assumes.

The two signals worth building on first
Digital engagement and patient support programme data.

Both are first-party, objective, timestamped and already in your systems. Neither depends on a rep's judgement. Neither can be gamed by the person being measured. And in most Indian brand teams both are collected diligently and used for almost nothing beyond campaign reporting.

A segmentation that combines modelled potential with observed digital behaviour is dramatically better than a rep-rated list, and it can be built from data you already own without licensing anything new. That is the fastest available improvement in Indian HCP targeting, and it is sitting unused in the marketing platform.

How to segment doctors for pharma marketing — the six steps

1. Start from the decision, not the data. Write down what will change as a result. "Which doctors get a rep visit versus digital only", "which get the switch message versus the initiation message", "where does the launch effort go first". A segmentation without a named decision behind it will not be used, because nobody knows what to do differently on Monday.
2. Inventory the signals you hold against the table above, and mark each as reliable, indicative or absent. This determines technique. A team with strong digital engagement and no PSP data should build a different model from one with the reverse.
3. Model potential from proxies. In the absence of prescriber-level scripts, build a potential estimate from practice characteristics, sub-specialty fit, catchment, chemist-proximity signal and PSP evidence. Validate it against something — territory sales, or a sample of doctors where you have better information.
4. Layer behaviour on top. Current engagement, direction of travel, channel responsiveness. This is where micro-segmentation separates from ranking: two doctors with identical potential and opposite engagement behaviour need different treatment.
5. Cut four to eight segments and name them in plain language. "High potential, digitally engaged, not yet prescribing" is a segment a brand manager and a rep both understand. "Cluster 4" is not.
6. Assign one differentiated action per segment, and write it down. Different message, different channel, different frequency, different owner. If two segments get the same action, they are one segment.
   Step six is the test that most segmentation projects fail. It is worth applying it early — if you cannot articulate a distinct action for a proposed segment, do not build it.

The five dimensions of micro-segmentation

Deciles use one dimension. Micro-segmentation uses the intersection of several — which is what makes segments small, specific and actionable rather than merely numerous.

The intersection is the segment. A doctor who is high potential, never engaged, unaware, digital-led and connected is a completely different commercial problem from one who is high potential, loyal, advocating, rep-led and independent — and a decile model puts both in band nine and sends both the same detail aid.

The dimension almost everyone omits

Channel preference — dimension four.

It is the only dimension that determines whether any of the others can be acted on. A high-potential, high-influence, early-adopting doctor who does not accept rep visits and does not open email is, operationally, a low-yield target until you find a channel that works.

It is also the easiest dimension to build honestly, because it is pure first-party observation — response rate by channel over two quarters, no modelling required. Add it before you add anything sophisticated, and expect it to reorder your target list more than any other single change.

How do I segment physicians beyond decile targeting?

Replace the single ranked list with a matrix of potential against current behaviour, then qualify each cell by channel preference. That gives roughly six to nine cells, of which four to eight will be commercially meaningful. Each cell gets its own action, not its own call frequency. The shift from deciles is not that you have more granularity — it is that you stop treating engagement as a volume problem and start treating it as several different problems that happen to share a field force.

A worked version of the core matrix, with the action stated for each cell. This is the artefact to hand a brand team.

Two things this matrix does that a decile list cannot. It makes the "ignore, deliberately" decision explicit and recorded — most target lists never formally exclude anyone, so low-potential doctors quietly absorb rep time every cycle. And it separates acquire from grow from protect, which are three different messages that a frequency-based model collapses into one.

The adoption ladder

Dimension three deserves its own treatment, because it is the most useful concept in this article for launch brands and the most frequently misapplied.

The adoption ladder places each doctor on a journey — unaware, aware, trialling, adopting, advocating — and the operating principle is that you can only move a doctor one rung at a time. Messaging a doctor who has never heard of the brand as though they are deciding whether to expand usage does not accelerate them; it produces no response, which then reads as low potential.

For launches in specialty markets with a small defined universe, the adopter lifecycle framing — innovators, early adopters, early majority, late majority, laggards — is a useful overlay, because seeding evangelists early matters disproportionately when the total prescriber base is a few thousand rather than a few hundred thousand.

The common misapplication is treating the ladder as a segmentation on its own. It is one dimension. A doctor at the trialling rung with low potential and a doctor at the trialling rung with high potential and strong peer influence require very different investment, and the ladder alone cannot tell them apart.

AI-driven micro-segmentation examples in pharma

Concrete rather than aspirational, and honest about which of these need data an Indian team may not hold.

The honest framing on AI here

Four of those seven work on first-party data an Indian brand already owns. Movement alerts, look-alike modelling, behavioural clustering and natural-language segment building need no prescription panel at all.

That is worth stating clearly, because the usual conclusion drawn from the data gap is that sophisticated targeting is not possible in India until script data arrives. It is not true. What is not possible is copying the US method. The techniques that depend least on Rx data happen to be among the most useful, and they are unbuilt in most Indian brands not because the data is missing but because nobody has assembled the doctor 360 underneath them.

Which is the real prerequisite, and the subject of a separate article

Choosing the model — and why you should start with rules

A recurring failure is jumping to clustering because it sounds like the sophisticated answer. It usually is not the right first step.

The argument for starting with rules is not that models are worse. It is that a rules-based segmentation can be argued with. A brand manager who disagrees that sub-specialty should outrank practice size can say so, and the discussion improves the segmentation. A clustering output cannot be argued with in the same way, so it is either accepted uncritically or rejected entirely — and in our experience it is usually rejected, quietly, by a field force that carries on using its own list.

Ship rules in three weeks, learn what the segmentation is actually for, then earn the right to model.

How many segments, and how big?

Sizing rules that are boring and load-bearing.

• Four to eight segments. Below four you have grouped rather than segmented. Above eight, reps cannot hold them in mind and the segmentation is used by head office and ignored in the field.
• No segment below about 5% of the addressable universe unless it is a deliberate high-value micro-segment with its own dedicated programme. Small segments fragment execution and cannot be measured — you will never detect a lift on 40 doctors.
• No segment above about 40%. A segment containing nearly half the universe has not distinguished anything and will be treated as the default.
• Every segment needs a distinct action. The test stated earlier, repeated because it is the one that matters: if two segments receive the same action, they are one segment.
• Stability matters more than precision. A segmentation where 30% of doctors move segment every quarter destroys field trust and makes measurement impossible. Target under 10% quarterly churn, and if you exceed it, the boundaries are drawn on noise.

That last point deserves emphasis because it is rarely discussed. A more precise segmentation that reassigns doctors constantly is operationally worse than a slightly cruder one that holds still. Reps build relationships over quarters, not weeks, and a doctor who moves from grow to protect and back again three times a year teaches the field that the model is noise.

Where segmentation projects actually fail

Rarely in the analysis. Almost always in what happens next.

1. No differentiated action. Every segment receives the same detail aid at a different frequency. This is the dominant failure, and it means the entire project produced a call-planning input rather than a strategy.
2. Segments the field cannot recognise. If a rep cannot tell which segment a doctor is in without opening a system, the segmentation does not exist in the field. Names must describe observable characteristics.
3. Built on unverified doctor data. A segmentation on a database with duplicates and stale affiliations assigns the same doctor to two segments and mis-assigns others entirely. Clean first — this is why deduplication precedes segmentation rather than following it.
4. No holdout, so no proof. Without a control group, nobody can demonstrate the segmentation worked, so it gets replaced by the next brand manager's approach within eighteen months.
5. Refreshed too often or never. Quarterly is usually right. Monthly creates churn; annual means the segmentation is describing a market that has moved.
6. Owned by analytics, not by the brand. A segmentation the brand team did not help build is one they will not defend when it constrains something they want to do.

The test to apply before you build anything

Ask the brand manager: "What will a rep do differently on Monday for a doctor in segment three versus segment five?"

If the answer is a frequency — more calls, fewer calls — the segmentation is a call plan and does not need five dimensions and a model to produce it.

If the answer is a different conversation, a different channel, a different piece of evidence or a different owner, you have a real segmentation and it is worth the effort.

Ask it at the start, not at the review.

How to measure whether it worked

Four measures, and one design decision that determines whether any of them mean anything.

The design decision is the holdout. A randomly selected group of doctors within each segment who receive the previous approach. Without it, every result is confounded by market movement, competitor activity and seasonality, and you will spend the review arguing about attribution rather than about the segmentation.

Report field adoption first and separately. A segmentation with excellent differential response and 30% field adoption has not been tested — it has been tested on a third of the field, and the two-thirds who ignored it are the more interesting finding.

Frequently Asked Questions For Micro-Segmentation of Doctors