NEVVI Medicare utilization intelligence
Data & method

Methods & Sources

Every number on Nevvi is public CMS data, presented as-is. Here is exactly where it comes from and what it does — and doesn't — mean.

Provenance — where every number comes from

Every count shown is a direct, unmodified read of CMS public-use files (Medicare Part B FFS Physician & Other Practitioners PUF; prescribing volumes from the Medicare Part D Prescribers file; provider affiliations from the CMS Doctors & Clinicians files). Service counts reflect services billed to and paid by Medicare fee-for-service only. They exclude Medicare Advantage, commercial, and Medicaid volume, and exclude provider×code rows CMS suppresses below 11 beneficiaries; provider-level figures are therefore lower bounds on fee-for-service activity, not estimates of total procedures performed. Prescribing counts are the exception: they are Medicare Part D claims, which include Medicare Advantage drug plans and cover no commercial or cash purchases, and CMS leaves out any prescriber-and-medication pair with 10 or fewer claims (see Prescribing volumes). The observed record is never modeled, projected, or fractionally attributed — no count, charge, or provider total is grossed up or guessed. Market-level totals carry one disclosed exception: a labeled all-Medicare estimate, scaled from that observed fee-for-service foundation by each state's FFS enrollment share and shown beside the observed count — still Medicare only, an estimate, never an observation. Counts of beneficiaries served, on code and market surfaces, are beneficiary-episodes within a code and year, not unique patients. One figure on a clinician’s page is different, and is labeled differently: “Beneficiaries billed for” is CMS’s own deduplicated count of the beneficiaries that clinician billed in the year (see Beneficiaries billed). Any number on this page can be traced to its source file and row — if a count looks wrong, you can check it against the public record. We encourage it.

Hospital inpatient figures come from CMS's Medicare Inpatient Hospitals — by Provider and Service file (IPPS, MS-DRG grain). The data year is the calendar year, as CMS has published it since its April 2023 restatement; each calendar year straddles two MS-DRG definition versions, because CMS changes them every October 1. Inpatient figures are never combined with the Part B figures elsewhere on this site — a different file at a different grain: stays on a CCN, not professional claims on an NPI. CMS excludes any hospital×DRG cell derived from 10 or fewer discharges, so hospital rows are a floor, not a census; shares are computed only against the state and national totals CMS publishes in the companion by-Geography file, which are not restricted to disclosed rows. A discharge is a counted stay, not a procedure. Average covered charge is the hospital's submitted list amount per discharge; average Medicare payment is what Medicare paid. Neither is a negotiated rate. Maryland hospitals are paid under that state's all-payer waiver, so their charge and payment figures do not compare with the rest.

The sources

A note on the code descriptions. Procedure codes come in three families. Most letter-led codes (HCPCS Level II) are created and published by CMS; the dental codes among them (a D and four digits, CDT) are the American Dental Association's, and Nevvi prints no description of them. The digit-led codes are CPT, maintained by the American Medical Association — their descriptions reach Nevvi inside CMS's public files, and they remain the AMA's work, not CMS's and not ours. CPT copyright 2025 American Medical Association. All rights reserved. Fee schedules, relative value units, conversion factors and/or related components are not assigned by the AMA, are not part of CPT, and the AMA is not recommending their use. The AMA does not directly or indirectly practice medicine or dispense medical services. The AMA assumes no liability for data contained or not contained herein. CPT is a registered trademark of the American Medical Association.

Coverage

The most recent CMS annual release (CY2024) spans all 50 states plus territories, 6,269 CPT/HCPCS codes, 69,825 physician groups, and 1,528,560 providers. The trend layer covers every calendar year from CY2013 to CY2024.

What the numbers mean — and their limits

Reading Medicare data well means knowing its edges. These apply everywhere on the site:

How volume is attributed to groups

CMS reports volume per clinician; Nevvi rolls clinicians up into the physician groups they're registered with. Some clinicians are registered with more than one group, and their volume cannot be split between those groups without guessing. Nevvi doesn't guess:

How city figures work

City filtering places volume by each clinician's Part B billing city — the city CMS publishes with the clinician's billing record. That is the only key that divides a group's volume across its offices exactly, and it is why the same group can appear under several cities with different volumes in a city view.

How specialty figures work

Each group carries one specialty label: the most common medical specialty among its member clinicians in CMS's Doctors and Clinicians register (the "modal" specialty). Advanced-practice roles — physician assistant, nurse practitioner, nurse anesthetist, and the like — are left out of that count. Advanced-practice clinicians often outnumber physicians in a specialty practice, so counting them would make the label describe a group's staffing rather than its work: a large orthopedic group would read as a physician-assistant practice. A group whose members are all advanced-practice clinicians — or whose other members carry no specialty in the register — keeps that label, because there it is the accurate description.

A multi-specialty group — a health system's medical group, a multidisciplinary clinic — is labeled by its largest specialty, so specialty filters compare groups by their dominant clinical identity, not by every service line they contain.

How market-structure figures work

The market-structure panel describes one procedure market — a code in a state, for a year — with four families of figures. All of them are computed over disclosed Medicare fee-for-service volume: CMS suppresses any provider×code row under 11 beneficiaries before publication, and that suppressed volume is absent from both the tops and bottoms of every ratio here. These figures describe the disclosed market, not the total market; comparisons across markets of very different size carry that bias.

How Atlas shows a market

An Atlas market is a set of procedure codes, a geography, and a year. Geographies are the United States and the fifty states plus the District of Columbia. There is no county or metro grain in this data.

Billing state. Volume is placed where the clinician bills from, which is not always where the patient lives. A national billing operation concentrates its volume in one state, so a state whose rate is a multiple of the median is a booking pattern before it is a demand pattern.

The rate. Services per 1,000 Part B FFS beneficiaries. Medicare Advantage enrollees file no fee-for-service claims and are in neither the top nor the bottom of this figure. Where a state-year has no enrollment vintage, no rate is shown.

Growth. The figure on the market page is year over year — this year against the year before, for the same market and geography. It is not shown when any code in the market has no published row for the prior year, including a code’s own first published year, because a partial first year manufactures growth.

Services by year. The chart shows each year's published services for the market in the geography, up to six years ending at the selected year. A year with no published volume is a gap, not a zero. A hollow point marks a year to read with care: 2020 and 2021 for suppressed and rebounding utilization, and the first year one of the market's procedures was billed, which is partial. Part B enrollment is the denominator of the rate beside the market's figures and is not a count of the market's patients.

Suppression. CMS suppresses any provider×code row under 11 beneficiaries before publication. That volume is absent from both the tops and the bottoms of every ratio here. It cannot be estimated from this file, and Nevvi does not estimate it.

Share and its denominator. Shares in Atlas divide by the market's own attributable volume, so they sum to 100% of the named-group market. That is not the same denominator as the share column on a group's own page, which divides by all disclosed volume in the state. Atlas links to that page rather than restating its figure, so the two never sit in one table.

Concentration. The top-5 share and its four bands, ours and not a regulator's — see How market-structure figures work above. The band word is printed only where named groups cover at least half of the code's disclosed volume; below that Atlas prints the top-5 share without it.

HHI. The Herfindahl–Hirschman Index, the sum of every named group's squared percentage share, over the named-group market. Atlas prints it as a number and never as a word: it answers whether any single group is dominant, while the band above answers how much the largest five hold together, and a label drawn from one placed beside the other is how a market with 54% in five hands once carried the wrong band word. It carries the same floors as the band — fewer than 11 named groups, or under a quarter of the market attributable, and no figure is published.

Position. Rank and percentile are computed on services per 1,000 Part B fee-for-service beneficiaries, never on raw size, across the state markets that bill any procedure in the market and have an enrollment vintage, out of the fifty states and the District of Columbia. They are not published below twelve of them, because a percentile over a thinner set manufactures a position.

Procedures absent from a record. A procedure is listed as absent from a state's record when at least twelve state markets have published volume for it and that state has published none. An absent row means the volume is either not there or not disclosed, and this file cannot tell those apart. Nevvi does not estimate what is missing.

What a dash means. A dash (“—”) marks a figure that is not shown, and it is never bare on Atlas. In a group's volume, it marks a clinician on more than one group's roster, whose volume cannot be attributed to a single group. Where a figure has no endpoint to compute it from, such as a change with no prior year, it marks that the figure could not be computed. Where CMS withholds a figure under its disclosure floor, it marks the withholding. Where a figure is not yet available because the data is refreshing, it marks that. In a group's procedures listing, the States column shows a dash where the code was billed only in rows CMS reports without a state. Nevvi counts markets by billing state, so it leaves those rows unattributed rather than guess. On Atlas the mark always carries its reason beside it.

The method version. Every Atlas page and export carries a method version, a revision number, and eight characters. The eight characters are computed from the floors, the bands, and the wording rules the method depends on, so they change whenever any of those changes. Two exports with the same version were built the same way. Two with different versions were not, and the export's method rows say which floors each used.

How the What-if simulator places a practice’s charges

The What-if simulator’s first question is where one practice’s submitted charges sit against its market. A submitted charge is the practice’s list price. Medicare pays its fee schedule whatever the charge says, so nothing here is a Medicare payment figure.

One group, one unit. Each group’s figure for a code is its average submitted charge per service: its total submitted charges on the code divided by its services, never an average of averages. In the state comparison that is the group’s volume billed from the practice’s billing state; in the national comparison it is the group’s volume summed across every state it bills from, plus any volume CMS published with no state. A group whose record is missing a charge or an allowed amount on the code is left out, because a partial sum is a wrong number.

One setting at a time. A group counts toward the office setting when at least 90% of its services on the code were billed in the office, and toward the facility setting when at least 90% were billed in a facility. A group billing the code in both settings counts toward neither, and a practice billing a code in both is not compared on it.

The market. For each setting, the median, the middle half (the 25th to the 75th percentile) and the median of each group’s charge divided by its Medicare allowed amount. With fewer than 30 groups in a setting only the count is shown.

The position. A practice’s percentile is the share of groups in the same setting charging less, plus half of those charging the same, with the practice itself counted, shown as a whole number from 1st to 99th. It needs at least 30 groups. The distance from the median is per service and carries no sign; it is never multiplied into a total.

Which codes. A practice’s five largest codes are its five codes with the most Medicare payment in the year, across its billing states, among codes where CMS published both a charge and an allowed amount. Its state is the billing state carrying most of its Medicare payment.

Suppression. CMS suppresses any provider-by-code row under 11 beneficiaries before publication, so that volume is absent from every group’s figure and from the count of groups.

The What-if simulator’s second question takes a service some practices billing one of this practice’s services also bill, where this practice has no disclosed volume, and asks what it would come to if their pattern held here. It is arithmetic on the record, stated as a condition. A practice may simply not do that work: it may lack the equipment, have a different case mix, or send that work elsewhere, and the claims record cannot tell that apart from any other reason.

One practice set. A practice counts toward a service when it billed at least 50 of it in the year, summed across every state it bills from. The share (practices billing the starting service that also bill the related one) and the pattern (the related service’s volume per starting service) are both read over the practices that cleared 50 on both, so the two figures describe the same practices. Each practice counts once, with its own ratio, so no one practice weighs more than another.

Which services. A starting service is one this practice billed at least 50 of, never a routine office or hospital visit (CMS’s evaluation and management category); the page considers the 25 of these it billed most. A related service is one billed alongside a starting service at least twice as often as chance would give, by at least 11 practices, with no disclosed volume for this practice. Each related service takes the starting service it is billed alongside most strongly, measured against chance, among those with at least 30 practices billing both, or among those with at least 11 when none has 30; figures are never added up across services. A service in a CMS service family this practice already bills is not offered, and a family is shown once. At most three are shown, the ones with the most practices billing both.

The range. The middle half is the 25th to the 75th percentile of those practices’ ratios, with the median marked. It needs at least 30 practices billing both; with 11 to 29, only the share is shown. Applied to this practice’s own volume of the starting service, it gives a range of services a year, and at Medicare’s rate a range of gross Medicare payments, traditional Medicare only.

The rate. Medicare paid per service for the related service in this practice’s state, in the setting (office or facility) where most of its starting service was billed: every clinician’s Medicare payment for the service there, divided by their services.

Before a Deep Dive. The share, the pattern, and the rate are open for any practice. This practice’s own volume of the starting service, and so the range applied to it, fills in with a Deep Dive on the practice for the year.

How the Screener ranks markets, groups, and clinicians

The Screener has two modes. Procedure markets ranks every Medicare Part B procedure market nationally. Groups and clinicians lists the medical groups or clinicians that bill Medicare in one state.

The ranks. Procedure markets rank on one measure at a time: services in the latest year, Medicare payments in the latest year, the 1-year change per 1,000 fee-for-service beneficiaries, or the 5-year growth per 1,000 fee-for-service beneficiaries.

The rate. Growth is measured per 1,000 Part B fee-for-service beneficiaries, so a shrinking fee-for-service pool does not read as a shrinking market. Five-year growth excludes each code’s first observed year, and a code with under three years of history shows a dash.

The volume floor. A market is ranked only when its services reach the floor chosen: at least 10,000 services, at least 100,000 services, or no floor. The floor and the growth columns stay on the fee-for-service observed basis.

The basis. Services and payments can be read as an all-Medicare estimate or as fee-for-service observed. The estimate scales each state’s fee-for-service volume by that state’s fee-for-service enrollment share and then sums the states; it is a scaled estimate, not an observation, and it is labeled as one wherever it prints. Billing-state buckets without payer-mix data are excluded from the estimate. The basis changes counts and dollars only; the per-1,000 rates are the same on either.

Groups and clinicians. A reader picks a Product (or all codes), a state or their own saved items, and groups or clinicians. Each column is a measure the group’s own page computes, on its own scale: growth against the market, standing among groups its size, who bills the work, and first-year cohort. Groups sort on one measure at a time, and nothing adds, weights, or blends two measures. A group with no figure for the sorted measure follows the divider with no number. Clinicians carry no rank: they list by name, or by first-year cohort.

First-year cohort. A first year is the first year a count cleared CMS’s floor of 11 beneficiaries, which is not the year a clinician first used a code. That is why the list orders by first-year cohort and never by the bare year, and why a first year only just above the floor carries a near-the-floor mark.

A cell with no figure says why: below the floor the measure needs, withheld because CMS does not publish counts under 11, no first year, needs a Product, or not measured for clinicians.

The limits that apply to every Part B figure are under What the numbers mean — and their limits.

How the Clinician network draws a line

The Clinician network starts from one clinician or one group and draws the clinicians who billed Medicare through the same group, in the same specialty, in overlapping years, from the CMS rosters of 2019 on. Each line names the group and the years.

A line. Two clinicians share a line when, in one roster year, both are on the same group’s CMS roster and both carry the same specialty that year. A clinician’s specialty in a year is the most common Medicare Part B provider type they billed under that year. A year with no specialty on file draws no line, so roster years after the latest published Part B year draw none.

A group is what CMS calls an organization: a CMS-enrolled group (PAC ID). A health system enrolled as several groups counts as several groups, so two of its clinicians who are never on the same group’s roster in the same year draw no line.

One line per pair. A pair of clinicians has one line, which lists each group they share and the years. A line is heavier the more distinct years it covers, from one year to five or more. A year is counted once even when a health system bills under two groups, and a year in between with no shared roster is never filled in.

The size rule. Where more than 50 clinicians in one specialty were on one group’s roster in a year, that group draws no lines for that year: at that size, a shared roster says little about any two people. Those groups are listed with their counts rather than dropped.

The second step. It shows the other groups each clinician billed through, and how many clinicians in the same specialty overlapped with them there. Those clinicians are counted, not named. It shows at most 3 other groups for each clinician, and 40 in all, most years first. The size rule applies there too.

A shared roster is a billing record. It shows that two clinicians billed Medicare through the same group in the same years; it is not evidence that they met. A large system may bill for sites in several states under one group, so two clinicians on its roster may never have practiced in the same place. Each roster year is one mid-year snapshot, so a clinician who joined and left between snapshots does not appear.

How Compare puts objects side by side

Compare puts two to six objects of one kind side by side. The kinds are procedure codes, medical groups, clinicians, state markets, MS-DRGs, hospitals, and medications. Each figure keeps the unit and the denominator named below, and a figure the record does not carry is a dash or a dashed thread, never a zero.

The Difference. For each row where every object has a figure, Compare prints one Difference, always on one raw measure. For counts, dollars, and rates it is the ratio of the largest to the smallest. For shares it is the difference in percentage points, but only where every column has the same denominator: a hospital’s share of its own state’s total, and the share of a state’s services credited to a group, are not compared. Standings are not compared: each is a percentile in its own cohort. A row with a figure missing says not compared and why. The answer line at the top of the sheet states the widest difference on the sheet’s size measure in words, over the objects that carry it, and names any object left out.

What Compare never does. It never sums the objects into one figure or blends measures into an index. It never ranks on anything but one raw measure at a time, and never on a percentile. It never puts a code beside a group, a group beside a clinician, or a hospital beside an MS-DRG. It never turns a low bar into a claim about care or need. And it never drops a scope silently: a panel that can’t take it says so.

Trend figures come in two honestly different depths. The 12-year series is market-level only — a code's total volume in a state (or nationally), never a group's. Group-level history exists only for years with an archived group roster (2019 onward), so group trend depth is a five-year window and is labeled as such; we never present twelve years of history for a group. Within that window, each year's volume is attributed through that year's roster (single-membership clinicians only — the same rule as everywhere else on the platform), and years where the group's attributable member set changed from the prior year are marked "roster changed": a volume step that coincides with roster churn can be a composition change, not growth.

Growth rates are computed conservatively. A code's partial introduction year is never an endpoint — a partial first year fabricates growth. The compound annual growth rate (CAGR) comes in two views a reader can switch between: the all-years default keeps every observed year, and an ex-pandemic option additionally drops 2020 and 2021, whose utilization cliff and rebound distort a growth base. The window is stated beside every figure, and the two views are never mixed on one page. Rate views are denominated per 1,000 Part B fee-for-service beneficiaries in the market's geography, using CMS enrollment counts for that state and year; years without an enrollment vintage show no rate rather than a fabricated one. The denominator is the Part B FFS population because the numerator is Part B claims — a beneficiary with Part A only cannot generate a Part B claim. (Methods note, July 2026: rates were re-baselined from the broader Original Medicare count to the Part B FFS count — a roughly 8–10% upward shift — so app rates and Nevvi's published reports share one denominator.) Every year-over-year market delta is shown beside the same year's FFS enrollment delta, because a shrinking market inside a shrinking fee-for-service pool is an enrollment story, not a procedure story. All deltas are statements about disclosed Medicare fee-for-service volume — never about care delivered or unmet need.

How the size-peer volume read works

On each physician group's page, Nevvi compares that group only with clinics like it — the same modal specialty, a similar size band (solo, small, mid, or large by clinicians billing in the state), and the same state — and shows its billed Medicare fee-for-service volume for the searched code as a percentage of what that size-peer cohort predicts (its observed÷expected level; see Expected volume below for the construction, suppression, and shrinkage rules). This is a rules-based structural comparison, not a model or a prediction, and it is a statement about billed volume among peers — never about care, patient need, or what a group "should" do. (Methods amendment, August 2026 — comparison floors and the geography ladder; chosen parameters, published here.) A comparison forms with as few as 2 peers, but what is shown depends on the peer count: at 11 or more peers the group's percentile and the peer benchmark ("peers typically N") are shown — the same 11 threshold CMS applies as a disclosure floor at the source; under 11 peers the comparison is shown only as a named rank ("2nd of 4 similar practices"), with no percentile, median, or benchmark figure — a benchmark computed from a handful of practices is unstable and can point back at an individual practice's numbers. Where a group's state has fewer than 2 peers, its comparison is drawn instead from its CMS region — the same specialty and size band across the region's states, labeled on the row ("Region 4 (Southeast)") — and where the region also has fewer than 2, no comparison is shown at all. A regional comparison counts state-market presences, the same unit as every figure on this site: a group billing in two of the region's states holds two entries, each sized by its own in-state footprint, so a regional peer count is a count of in-state practices, not of distinct parent organizations. The basis (state or region) is printed on every row. (Methods amendment, August 2026 — incidental footprints.) A group whose national roster is 20 or more clinicians and at least 5× its footprint in the state is treated as an incidental footprint of a larger organization — a token out-of-state presence, not a practice of that size in that market — and is excluded from cohort formation entirely: it is never counted or named as anyone's peer, and its own page shows no size-peer read for that market rather than a misleading one. This read is a property of the group, so it lives on the group's page, not on the code-market results view. (Methods amendment, September 2026 — the national fallback; chosen parameters, published here.) Where a code clears neither the state nor the region rung, Nevvi compares it against a national cohort of practices of the same specialty and size band instead of showing nothing. The national rung reuses the size-peer read's cohort definition at national scope, but it answers a narrower and coarser question than the state and region reads above: never a percentile, a rank, or a median comparison — only whether the group's volume for that code falls above, below, or inside the cohort's middle half (its 25th–75th percentile range), with the range's own values printed alongside so the width can be judged. A call within a small margin of either edge is shown as indistinguishable from the edge rather than forced one way. This is a different, coarser kind of statement on purpose: a national cohort spans practices in different rate environments, so it does not support a claim as precise as "14th nationally" the way a same-market cohort can, and Nevvi does not publish one. The national rung has its own, higher floor — 50 or more practices nationally billing the code, versus 2 for the state and region reads above — and its own prevalence requirement: at least 25% of the practices eligible to form the cohort must actually bill the code, so a code billed by only a small share of a large national population does not read as broadly adopted. A code that clears neither floor renders nothing, the same honest absence as a state or region cohort that never forms. Whichever rung a comparison is drawn from — state, region, or national — is printed on the row every time, and a national read is never combined with a state or region read into a single figure, sentence, or export.

How the co-billed codes read works

The Co-billed tab answers one question: which other codes do the same organizations bill? Every figure there is an organization-level billing fact — the same group billed both codes at supported volume in the same period — and never a statement about clinical sequence, about any patient, or about what is performed together. Co-billing reflects practice scope and equipment.

What each row reports. Every row leads with the ranking bar. A reader with full access also gets the share behind it, in plain words and with the counts folded in — a percentage never appears on its own. A reader with full access sees the partner side: the share of the partner code's billing organizations that also bill the code being looked at, with the number billing both and the partner's total (a partner billed by 963 organizations of which 851 also bill the searched code reads “88% … (851 of 963)”). A reader without access sees no figure at all: the bar shows where the row sits in the list, and the partner's code, share, and totals all stay withheld, because together they would identify it. Both readers see the same rows in the same order; one of them sees the quantity behind it. (Methods amendment, September 2026 — the list is now ordered by the share the row prints, a partner-cohort floor was added, and the reader without access sees position rather than a second, differently-directed share.)

How the list is ordered. Rows are ranked by the share the row prints — of the organizations billing the partner code, the proportion that also bill the code being looked at — highest first. The ordering is therefore something a reader can check against the row rather than take on trust, which is the property this ranking was chosen for. Ties are broken by the number of organizations billing both, then by code.

Until September 2026 the list was ordered by the overlap of the two codes' billing populations — organizations billing both, divided by the square root of the product of the counts billing each. That measure is sound and it is the geometric mean of the two directional shares, but a row can only print one of those two shares, and one factor of a product cannot explain the product's order. Readers reasonably read a printed 72% ranked below a printed 47% as a fault. The ordering now matches what is on the page.

Why the partner-cohort floor is part of the ranking. A share ranked on its own puts the smallest codes first: any code billed only by organizations that also bill the searched code scores 100%, whatever its size. On a myocardial perfusion imaging code the unfloored top eight were procedures billed by 11 to 41 organizations each — a reliable figure over a denominator too small to mean anything. Ranking instead by the raw count of organizations billing both fails in the opposite direction, filling the top with office visits that nearly every practice bills. The floor is what makes ranking on the printed share sound: it removes the codes in the first failure, and ranking on a share rather than a count keeps out the second. The two work together and neither is correct alone.

Why not "lift". Nevvi once ranked this list by lift — how much more often two codes are billed together than chance would predict. From a fixed searched code lift reduces to the partner's share alone, and it carries no floor, so it lands in exactly the first failure above. Lift is no longer shown on this tab.

Three published floors. An organization counts on a code only when its attributable volume for that code reaches 50 services within a single year, applied before any pairing — so the numerator and the shares underneath it describe one population. A pair is shown only at 11 or more shared organizations, the same 11 CMS applies as a disclosure floor at the source; below that a pair is withheld, both because overlap computed from a handful of organizations is unstable and because a handful of organizations sharing two uncommon codes can identify them. And a partner code is ranked only when 500 or more organizations bill it — the floor described above, without which the ranking would be led by codes whose share is reliable and whose denominator is not. Changing any floor changes the population every figure is computed over, so figures from different floors are never on the same basis. Absence of a pair is not evidence that the codes are never billed together.

The year ladder. (Methods amendment, August 2026 — the co-billing basis ladder; chosen parameters, published here.) The default basis is a single calendar year. At the floors above, most codes billed by a small number of organizations produce no qualifying pair in one year — their partners fall a few organizations short — and their tab would simply be empty. Where a code yields no qualifying pair in the year, Nevvi therefore draws its list from a wider window: the five calendar years ending in that year. On that basis an organization counts toward a pair when it billed both codes at supported volume in any single year of the window. Both floors are unchanged — the same 50 services within a year, the same 11 shared organizations. Volumes are never added across years to reach the support floor, and the 11-organization floor is never reduced: the window widens the pool of organizations that can clear the bar, it does not lower the bar. Because the wider basis has a different population, its figures do not line up with single-year figures, so the basis in use is printed on every list and the two are never combined in one list. Where neither basis produces a qualifying pair, no list is shown and the tab says so.

Large multi-specialty organizations. A group billing hundreds of distinct codes shares many pairs it has no particular relationship to. Nevvi discloses this rather than adjusting for it: an organization billing 100 or more distinct codes at the support floor in the period counts as high-breadth, and a pair whose shared organizations are mostly high-breadth is labeled as such on its row. The figures themselves are never reweighted.

What this list is not. It is a statement about which organizations bill which codes, and nothing more. It does not say the two codes are performed together, performed on the same patient, or clinically related; two codes overlap here when the same practices bill both, which usually reflects the scope of what a practice does and the equipment it has. A high overlap is not a recommendation, and a low one is not a gap.

Market position and expected volume

Two reads sit on top of the volumes above (methodology v1, 2026-07). Both are statements about billed Medicare fee-for-service volume — where a market sits, and what peers predict — never about care, patient need, or what anyone "should" do.

The comparison universe. Comparative statistics rank the 50 states and DC; territory markets are shown but not ranked; non-geographic billing codes count in national totals only.

Ranking basis (2026-07-21). Size rankings re-based to the all-Medicare scaled estimate (2026-07-21); the FFS-observed basis remains disclosed on every surface. Where any state in a comparison lacks a computable enrollment share, the whole comparison renders on the observed FFS basis and says so. Within a re-based comparison, the 5-year-change and CAGR figures follow the same basis as the sizes they rank beside — both endpoints scaled or both observed, never mixed. Rate-based growth rankings (per 1,000 Part B FFS beneficiaries) and every per-1,000-FFS rate are basis-independent and unchanged.

Market position. For the searched code set, every state market is placed on a volume × growth grid and ranked by 5-year change and CAGR. Volume bands are terciles among all state markets for the search; growth bands are the market's own 5-year change (±2%). CAGR excludes the code's introduction year. Neither the card nor the ranking is shown when fewer than 12 state markets bill the searched codes — position claims over a thinner universe are noise.

Expected volume. A group's expected volume is what its peers — same specialty, size band, and state — predict for a group its size: the cohort's per-provider rate applied to the group's own clinician count (indirect standardization, the same observed-vs-expected construction CDC and CMS use). A group's size band counts every clinician of record billing through it in that state; expected volume uses the narrower count of clinicians billed through it alone there — both are the group's in-state footprint, never its national roster, so a multi-state system is compared as the size it actually is in that market. Incidental footprints — groups whose national roster is 20+ clinicians and at least 5× their in-state footprint (August 2026 amendment above) — are excluded from cohort formation before any of this is computed. Two published rules protect it:

Suppression. No expected volume or observed÷expected ratio is shown when a cohort has fewer than 11 peers, or when the expected service count itself is under 11.

Shrinkage. Observed÷expected ratios are shrunk toward the cohort mean with an Empirical-Bayes prior weighted by cohort cell size, so small groups do not top rankings on noise.

The Standing tab’s notable band (August 2026 amendment; published parameters, not discovered constants). On a group’s Standing tab, a code is called out only when its shrunk observed÷expected ratio clears the notable band: 1.5× or more above expected (“well above”) or 0.5× or less (“well below”). Everything between renders in-range, without a verdict. The band is deliberately wider than a naive “above/below the benchmark” cut: expected volume is a mean-based expectation and volume within a cohort is right-skewed, so the typical group sits somewhat below its cohort’s expected level by construction — sitting somewhat below expected is ordinary, and the Standing tab says so rather than treating it as remarkable. Standings group by the code’s RBCS family on screen. All positions are billed-volume positions among structural peers, never statements about care.

Standing comparison floors and the geography ladder (August 2026 amendment; published parameters). A Standing comparison needs at least 2 peers, drawn from the group’s state cohort where possible and otherwise from its CMS-region cohort (same specialty and size band at region grain), labeled per row. Peer benchmark figures (percentile, median, “peers typically”) print only at 11 or more peers; at 2–10 peers a position renders as a rank — “3rd of 7 similar practices” — with no benchmark figure of any kind. Expected volume and the ratio remain state-basis reads under the suppression rule above.

How “Lookalikes” for a code works

A code’s Lookalikes are other codes whose published Medicare rate is close to it — within 20% either way. The rate is the national fee-schedule amount (the geographic adjustment set to 1.000), so it is the same figure everywhere in the country rather than any one locality’s. The candidates are codes in the same RBCS family and in sibling families under the same RBCS subcategory; the band is a published constant, not something tuned per code, and it is not adjustable. Codes CMS has not assigned to a procedure family are not candidates: the unassigned set is a residual bucket, not a family, so a code inside it is neither a family member nor a sibling of anything.

This is a statement about payment, and only about payment. Two codes landing in the same band means Medicare pays about the same for them. It does not mean they are clinically alike, interchangeable, or substitutes for one another, and nothing here suggests billing one instead of another.

The tab also shows, for each code, the average submitted charge, the average amount Medicare actually paid, and the ratio between those two. Those three are observed figures from named billing groups — a different basis from the fee-schedule rate, which is why the published rate and the observed payment differ and are labeled separately. The observed payment is lower than the published rate for reasons that have nothing to do with the code: the deductible, the 20% coinsurance a patient or their secondary insurer pays, local geographic adjustment, and the mix of office versus facility settings. We do not print a ratio between the published rate and what was observed — it would read as Medicare paying less than its own fee schedule, which is not what the gap means.

Where Medicare publishes no national rate for a code — contractor-priced codes, where the amount is set locally — there is no band to compute, and the tab says so rather than showing an empty table.

Lookalike practices: structural similarity (August 2026 amendment; published parameters, not discovered constants). A group’s Lookalikes are other PRACTICES built like it, and the comparison is a size comparison and nothing more. Three measures, each taken across the practice’s whole published footprint for the year: its clinician count, its attributed services, and its attributed Medicare payment. A practice qualifies only when all three are within a factor of two of this one — that is, the ratio on each measure lies between 0.5× and 2×. Qualifying practices are then ranked by the average of the three ratios, measured in logs so a gap counts the same at any scale: for measures a and b, the distance is (|ln(ca/cb)| + |ln(va/vb)| + |ln(ra/rb)|) ÷ 3, and the “about N% apart on size” a row prints is edistance − 1. The three measures carry equal weight; no weight is fitted, learned, or tuned to an outcome. Up to 8 practices are shown. Specialty ranks, it does not filter. A practice in a different specialty carries a fixed penalty of 0.10 in the same units — 10.5% — so it must be that much closer on size to rank ahead of a same-specialty practice. Nothing is excluded for its specialty, no list is shortened by it, and at this weight most shown practices are still a different specialty. The penalty is a chosen constant, like the factor of two; it is not fitted to anything. The “% apart on size” a row prints is the size figure alone and never includes this penalty — it says what it measures. Geography is a filter, not a measure, and it walks the same ladder Standing does: the practice’s own state first, and where fewer than 2 practices there fall inside the window, its CMS region — labeled on every row, so the basis of a comparison is never inferred. Where neither reaches 2, no lookalikes are shown and the group’s own page names the reason to subscribers and trials. Note what this comparison is NOT: it does not match on specialty, and practices of a similar size may do entirely different work. It is a claim about scale in the public record, never about patients, quality, or care.

Peer codes: codes similar practices bill (August 2026) — method published, not currently shown on the site. We compute this read and do not currently surface it anywhere: the tab that showed it was removed in August 2026. It is documented here because the calculation still runs and the method may return. It identifies groups that MATCH a code's peer definition — same specialty, size band, and state as its disclosed billers — with no attributable disclosed volume of their own on the code. This is a claim about the public record, never about care or what a practice could do, and it distinguishes three states, not two: billed · not attributable (the group's clinicians bill the code through multiple groups, so no volume is assigned to any one of them) · no published volume. Its floors differ from the 11-peer rule above, FOR THIS READ ONLY, and each is a chosen parameter, not a discovered constant: a comparison needs at least 2 disclosed billers in the peer definition; at least 25% of the peer group must bill the code before absence is worth asserting; and the expected count at the peer rate must be 11+ (the same disclosure-cutoff mirror as above). That 25% is measured among the peers who do this kind of work, not among all of them (August 2026): the share is taken over the cohort members with billing in the same clinical family as the code — CMS’s own Restructured BETOS family, so the grouping is CMS’s and not ours — because a rarely-billed echocardiography code should be judged normal-or-not among the practices that do echocardiography, not among every cardiology practice. Judged against all members, a code that is routine for the practices that offer it at all can look marginal purely because most of the cohort works elsewhere in medicine. The 25% threshold itself did not change when the denominator did; we tested 50% on the same basis and it withholds more codes than the old rule did, so 25% stands. Below 11 peers the comparison is described as a named, counted set and no peer median or “typically bill” benchmark is ever printed. Solo practices never appear: below Medicare’s 11-beneficiary disclosure cutoff, zero and small are indistinguishable. Volume billed under a facility, another group ID, or by multi-group clinicians will not appear in this calculation.

Most codes produce no peer-code read at all, and that is the floors working rather than data missing: of the codes billed by groups we can form cohorts around in 2024, fewer than two in five clear all three — for most of the rest, no peer group anywhere reaches even two disclosed billers. Where a group’s peer-code read is withheld, the reason is one of these — its published footprint in each market is a single clinician, or nothing in the public record is attributable to it that year, or its presence in each market is a small share of a much larger national roster, or no peer group it matches reached 2 published billers, or the share of the practices doing that kind of work who bill a code never reached the 25% mark, or the expected count stayed under 11 for every code it does not already bill. It can also be that the group already bills every code that clears the floors. How close any of those came is not published, because those are precisely the figures the floors exist to withhold. The comparison is always made within a state: we tested pooling thin codes into CMS regions and nationally, and pooling qualifies FEWER codes, not more — a wider map adds peer practices more slowly than it adds practices to compare them against, so a rare code looks rarer the wider you look. There is no geographic fallback in this calculation — the region rung described above belongs to Standing and to lookalike practices, both of which ARE shown on the site, and where the thing being pooled is practices rather than a code's billers.

How “Non-billers” for a code works

Non-billers (August 2026; published parameters, not discovered constants). A code’s Non-billers are groups built like its billers, in the same state, with no published volume on the code in the public record. The label names the row set; the sentence is the claim, and it is a claim about the record only: a clinician below Medicare’s 11-beneficiary disclosure cutoff, volume billed under a facility, and a clinician attributed to another group all look identical to absence here, so the read never says a practice does not perform a service — only that nothing is published. The seed is the code’s billers in the year, state by state: the groups with attributed volume, plus the groups with a clinician on their roster who billed the code that year (so a biller whose volume is not attributable to any one group still seeds the comparison). “Built like” is the lookalike-practice window above, unchanged: a group qualifies when its clinician count, attributed services, and attributed Medicare payment are each within a factor of two of at least one biller in its state, and the “about N% apart on size” it prints is the same average log gap, from the same practice profile. The state is the biller’s billing state, and the compared group is one whose own primary market is that state; there is no regional fallback, because the question is “in the same states”. Three states, never two: a group with attributed volume on the code in any year, or with any roster clinician who has billed it, is excluded rather than listed — the list holds only groups with no published volume at all. Floors and the ceiling, each a chosen parameter: a state needs at least 2 billing groups before a comparison is drawn there; a code with more than 200 groups carrying attributed volume on it nationally in the year is not read at all, on any page (the read is for thin adoption, where a same-state set is a comparison rather than a census); the per-state counts are shown only above the floor; and solo practices are never listed — below Medicare’s 11-beneficiary cutoff, zero and small are indistinguishable, so a one-clinician group cannot be called absent. Two views of one list. Same specialty keeps groups whose main specialty matches a biller’s. Neighboring specialties keeps groups with at least 2 clinicians, or at least 10% of their roster, in a specialty that bills the code — where “a specialty that bills the code” is read from the record for each code: any clinician specialty holding at least 5% of the code’s billing clinicians, or at least 11 of them, with nurse practitioner, physician assistant, clinical nurse specialist, certified clinical nurse specialist, and the undefined physician type set aside as types rather than specialties. The second view includes the first. Specialty filters here; it does not rank. What this read is worth, measured. Across 936 thin codes, groups inside the window of a same-state biller went on to bill the code the following year at about 5× the rate of other groups in the same states (same specialty: about 30×). Absolute rates are small: most listed groups do not bill it the next year, and the list holds roughly one in six of those that do. That is a property of the read as a whole, stated once here; no row carries it, and nothing on the page predicts what any practice will do. Nothing here is a claim about patients, quality, or care.

When you type a question instead of a code, the question text is sent to our AI engine to translate it into the matching CPT/HCPCS codes — that translation is the only thing the AI does. Every count, charge, and ranking you then see is queried directly from our database of CMS data. No figure on Nevvi is generated, estimated, or inferred by an AI model.

Rankings

Groups are ordered by measured Medicare fee-for-service service volume for the code and geography you searched — attributed per the rule above — a direct read of the CMS data, not a rating or a score. Nevvi does not grade providers or judge quality of care; it reports what the public record shows.

Disease-burden context

State disease prevalence appears beside a ranked list of physician groups — for example, diagnosed-diabetes prevalence next to the continuous glucose monitoring market in the searched state. The figure is the CDC Behavioral Risk Factor Surveillance System (BRFSS) state series: the survey-native, age-adjusted prevalence estimate with its 95% confidence interval, taken from each measure's most recent survey year. Some BRFSS modules run in alternating years, so the survey year — always labeled on the figure — can differ by measure and lag the utilization year.

A search with no state filter shows the range of these age-adjusted state estimates across reporting states, with the highest-prevalence states linked for drill-down. CDC publishes no age-adjusted US aggregate in this series; where a single US figure appears, it is CDC's own published crude prevalence, quoted verbatim and labeled as such. Crude and age-adjusted estimates are not on the same basis, and Nevvi never computes its own national aggregate from the state series.

CDC BRFSS prevalence is survey data covering all-payer adults 18 and older; Nevvi utilization counts Medicare fee-for-service only, largely 65 and older. The two appear side by side as context — Nevvi never combines them into a score, rating, or ranking.

The Burden overlay puts this burden figure beside Medicare volume by state. It starts each condition's basket from Nevvi's signed map of the codes linked to that condition — for diagnosed diabetes, for example, continuous glucose monitoring and diabetes self-management training. A code a reader adds is their own selection and carries no claim of a link. Each code's figure is services per 1,000 Part B fee-for-service beneficiaries (for a drug, its billed units, as CMS counts them), counted by the state a clinician bills from, and the burden figure and the code figures are never combined.

Beside the survey figure, a state-filtered search may also show the share of Medicare fee-for-service beneficiaries in that state who carry the diagnosis, from CMS's Chronic Conditions Warehouse (CCW) as published through CMS's Mapping Medicare Disparities (MMD) tool. That figure is a claims-based diagnosis flag on the same fee-for-service denominator Nevvi's utilization counts use — Medicare Advantage enrollees excluded — published as an integer percent, both as the actual share and age-standardized to the national Medicare age distribution (the age-standardized figure appears in the chip's hover). Nevvi holds the 2022 and 2023 data years and shows the newest. Where CMS withholds a state × condition cell (denominator under 11), no figure appears — absence is suppression, never zero.

The survey figure is what adults reported; the CCW figure is a diagnosis recorded on a Medicare claim, and the two are adjusted to different populations, so they are not expected to match.

Prevalence appears only where Nevvi maintains a curated pairing between a procedure code and the condition that service directly manages (for example, dialysis and chronic kidney disease). No pairing, no context — a condition is never inferred from a code. And the context is display-only: Nevvi does not combine any population or prevalence figure with utilization into a score, index, or ranking.

How the Burden overlay sets burden beside billed volume

The Burden overlay puts one condition’s prevalence beside Medicare billing for a basket of codes, for each of the 51 state markets (the fifty states and the District of Columbia). Burden and each code are separate measures on separate scales.

The burden figure. It is the CDC Behavioral Risk Factor Surveillance System (BRFSS) state estimate: survey prevalence, age-adjusted, among adults 18 and older across all payers. It is taken from the measure’s most recent survey year, and the page labels the survey year, or the range of years where states differ, above the column. A state with no published estimate reads “not published”. The survey basis and the Medicare figure that can sit beside it are described under Disease-burden context.

The codes. A basket starts from Nevvi’s signed map of the codes linked to the condition. A code you add is your own selection, and we make no claim that it relates to the condition. The basket holds up to 25 codes.

Each code’s figure is services per 1,000 Part B fee-for-service beneficiaries (for a drug, its billed units, as CMS counts them), counted by the state a clinician bills from, for the calendar year shown. A code with no figure in a state was either not billed there or withheld by CMS below 11 beneficiaries; the record does not tell those apart. The table reads “none observed” where no row exists and “no enrollment figure” where the state has no enrollment vintage for the year.

The thresholds. You set two thresholds independently: a burden threshold (prevalence at or above a value) and a volume threshold on one chosen code (at or below a value per 1,000). A state is marked only when it meets both, and the page describes it in the two measures’ own words. The thresholds are a view setting: they ride the address’s # fragment, so no request, no query and no save carries them.

The caveat that travels with every view. CDC BRFSS prevalence is survey data covering all-payer adults 18 and older; Nevvi utilization counts Medicare fee-for-service only, largely 65 and older. The two appear side by side as context — Nevvi never combines them into a score, rating, or ranking.

What the overlay never does. It never sums the basket into one figure or blends it with the burden figure, and it never ranks on anything but one raw measure at a time. It never claims that a code you added relates to the condition, and it never turns a low bar into a claim about care or need.

Comparing Nevvi with all-payer estimates

Commercial healthcare datasets typically model total procedure volume across every payer, projecting from claims samples. Nevvi anchors its market totals on a labeled all-Medicare estimate scaled from the observed Medicare fee-for-service record — Medicare only, never all-payer. All-payer figures answer a different question, and they usually run two to four times the observed fee-for-service foundation for a code — with the same named providers and largely the same rank order. Neither number is "wrong"; they are different measurement constructs. Nevvi publishes the observed foundation beside every estimate and labels the estimate as one.

You can reconcile the two yourself, with whatever tool you license:

Beneficiaries billed

Population figures on a clinician's page come from CMS's Physician & Other Practitioners by Provider file for the calendar year shown: the Medicare fee-for-service beneficiaries a provider billed for in the year — not a patient panel, not a population the provider manages. Medicare Advantage enrollees are excluded.

Condition shares are integer percentages CMS computes from CCW diagnosis flags; CMS caps any share at 75% (shown as "≥75%") and withholds any share or count based on fewer than 11 beneficiaries (shown as "—"). Providers with fewer than 11 beneficiaries in the year are absent from the file entirely.

The CMS-HCC risk score is a Part A/B payment risk-adjustment average, sensitive to diagnosis-coding intensity; it is not a clinical severity measure. Scores carry the model version CMS used for that year and are never compared across years.

Percentiles rank a provider among providers of the same specialty with a disclosed profile for the calendar year shown, nationally, by the average risk score (or dual-eligible share) of the beneficiaries they billed for; the cohort floor is 11. The national specialty medians shown beside the panel are medians over the specialty's providers with a disclosed value; a median is withheld under 11 such values.

The beneficiaries a provider billed Medicare Part B for in the year — not a patient panel, not a population the provider manages. Medicare fee-for-service only; Medicare Advantage enrollees are excluded. Condition shares are CMS's CCW diagnosis flags; CMS caps any share at 75% (shown as ≥75%) and withholds any share or count based on fewer than 11 beneficiaries (shown as —).

Prescribing volumes (Medicare Part D)

Prescribing figures on a clinician's page come from CMS's Medicare Part D Prescribers by Provider and Drug file for the calendar year shown — Medicare Part D claims and drug cost only: no Part B drugs, no commercial, no cash. Drug identity is CMS's own brand/generic string pair, printed as published. Part D figures are never combined with Part B utilization and never printed beside an inpatient figure.

CMS excludes any (prescriber, drug) pair with 10 or fewer claims, and a prescriber none of whose medications clears that floor is absent altogether, however many claims they wrote in total, so every figure is a floor: an absent medication is suppressed at source or zero, never confirmed absent, and a withheld cell renders "—" — never a zero. A total on the page is the sum of the rows rendered above it and nothing else.

A claim is a dispensed fill. CMS counts original prescriptions and refills alike, and a fill may cover any number of days, so a 90-day fill and a 30-day fill each count once. A claim is not a prescription written — a script never filled, or paid another way, is not in the file — and it is not a beneficiary count. Cost divided by claims is therefore confounded by fill length and is not a price, and no such ratio is printed.

Part D drug cost is total spending, not what Medicare paid. CMS defines it as the aggregate drug cost paid for the associated claims, including ingredient cost, dispensing fee, sales tax, and any applicable vaccine administration fees, based on the amounts paid by the Part D plan, the beneficiary, government subsidies, and any other third-party payers. It is gross of manufacturer rebates, which CMS does not publish at this grain, and CMS states that total drug cost cannot be directly attributed to payments from the Medicare Trust Fund. It is not a price, a rate, or a negotiated amount.

A brand name is one string, not a molecule. CMS derives the brand and generic names from the dispensed National Drug Codes, and a medication sold in several presentations may carry several brand strings. A national row is that string's total; the molecule's total is the sum across its strings, which a clinician's page does not compute.

Drug-class labels are assigned from the FDA's published Established Pharmacologic Class terms, read through the National Library of Medicine's RxNorm and RxClass; the grouping of FDA terms into classes is Nevvi's own. A label is never a filter: unlabeled medications render identically, and any class figure is a floor over labeled rows. Drug-class terms courtesy of the U.S. National Library of Medicine (RxNorm and RxClass).

Drug markets (Medicare Part D)

A medication’s market figures come from two CMS files, both Medicare Part D and both for the calendar year shown. The totals (claims, 30-day fills, Part D drug cost, and the number of prescribers) come from the Medicare Part D Prescribers by Geography and Drug file, which CMS builds from all Part D claims, nationally and for each state. Prescriber rows come from the Medicare Part D Prescribers by Provider and Drug file, in which CMS withholds any prescriber with 10 or fewer claims for a brand row. So prescriber rows never add up to the totals, and they are not meant to: a prescriber with no row for a brand is not confirmed absent, and a sum of prescriber rows is a floor, never the market total. A share is taken only against the total CMS publishes for the place, never against a sum of prescriber rows. CMS publishes no state figure for a brand row with 10 or fewer claims in that state, so such a state shows no total rather than a zero.

A claim is one dispensed fill, refills included. CMS counts original prescriptions and refills alike. A claim is not a prescription written (a script never filled, or paid another way, is not in the file) and it is not a patient count. A fill may cover any number of days, so a 90-day supply is one claim, the same as a 30-day supply. The 30-day fill is CMS’s standardized measure of how much was dispensed: CMS derives it from the number of days supplied on each Part D claim divided by 30, and counts any one claim as at least one and at most twelve 30-day fills. A 90-day fill therefore counts as three. Where 90-day supplies are common, claims understate volume, and 30-day fills are the better guide to how much was dispensed.

Part D drug cost is gross. Part D drug cost is the gross amount paid at the pharmacy by plans, beneficiaries, and subsidies, before manufacturer rebates, which CMS does not publish. It is not a net price, a rate, or what Medicare paid. CMS defines it as the aggregate drug cost paid for all associated claims: ingredient cost, dispensing fee, sales tax, and any applicable vaccine administration fees, based on the amounts paid by the Part D plan, the beneficiary, government subsidies, and any other third-party payers. CMS states that these costs do not reflect manufacturer rebates, and that total drug cost cannot be directly attributed to payments from the Medicare Trust Fund. Cost divided by claims is confounded by fill length, and no such ratio is printed.

Prescriber counts belong to one brand name. CMS publishes a number of prescribers for each brand row, meaning each brand and generic name pair as CMS prints it, and a clinician who writes two brands of one medication is counted in both rows. Prescriber counts are therefore never added across brand names, and no prescriber total is shown for a medication. Claims, 30-day fills, and Part D drug cost do add across brand rows, so a medication’s totals are the sum of its brand rows. A medication’s page is keyed on CMS’s generic name, printed exactly as CMS prints it.

Calendar years. Every figure is for a calendar year. Market totals cover the five most recent years CMS publishes. The by Provider and Drug file’s prescriber rows cover the most recent year only, so no prescriber figure carries a trend. The procedure trendlines on this site mark 2020 and 2021 as pandemic years to read with care, and offer a view without them. That treatment is not applied to Part D: a medication’s claims are drawn for every year on one scale, with no year marked or set aside.

Where a claim is counted. State figures place each claim in the prescriber’s state as CMS files it, not where the pharmacy or the patient is. CMS fills that state in from the prescriber’s record in NPPES, the National Provider Identifier registry, when it builds the file. The state on a prescriber row is the state in the Part D file, and it may differ from the practice location on that clinician’s page, which comes from a different file and may be of a different date. No rate per 1,000 beneficiaries is shown, because this site holds no Part D enrollment count to divide by.

Scope. Medicare Part D only: stand-alone prescription drug plans and Medicare Advantage drug plans. No Part B drugs (those are billed as procedure codes in the Part B file), no commercial insurance and no cash purchases. These figures are never combined with Part B procedure volumes.

Updates

Nevvi refreshes when CMS publishes — annually for the Part B file, on each source's own cadence for the affiliation and enrollment files. The data footer on every page names the source vintage and load date currently in service.

Citing Nevvi

Please attribute as: "Nevvi (nevvi.app), from the CMS Medicare Physician & Other Practitioners PUF." Nevvi is not affiliated with, or endorsed by, CMS or HHS. For a one-page, versioned record you can drop into a data room, see the citable methods appendix — it pins every figure to the exact data build it came from.

Questions

Spot something that looks wrong, or want more detail on a method? Contact us — corrections are welcome.