Arete Healthcare LLC

Independent Practices | Shared Infrastructure | Clinician-Led Governance

Key Findings

In Pictures

The figures follow the analytic sequence of the paper.

Figure 1 shows that reliability is measure-specific and depends on denominator volume: at the same number of patients, visits, or spinal CMT units, the share of “profiles” reliable enough for quality-improvement feedback or for consequential decisions such as “tiering” differs widely by measure.

Figure 2 shows how much empirical uncertainty remains around “profile” values at common denominator volumes.

Figures 3a and 3b show which observed patient, benefit, diagnosis, lifecycle, concentration, and measurement factors are associated with differences in the measures: Figure 3a compares different chiropractors’ “profiles” in the same measurement period, and Figure 3b follows the same “profile” across sequential measurement periods.

Figure 4 plots every provider-insurer “profile” by its own reliability and precision, showing that these are properties of each “profile” rather than of a sample size.

Figure 5 brings these together at the kind of line a “tiering” program draws: near hypothetical thresholds for visits and cost per patient, very few “profiles” were both reliable and precise, and even a reliable, precise “profile” above a line can reflect the factors shown in Figures 3a and 3b.

Figure 1. Reliability differs by measure and by denominator volume. From upper left: visits/patient, CMS-standardized cost/patient, units/visit, and 98941 share of spinal CMT. Each chart shows the share of provider-insurer profiles reaching 0.70 reliability (purple; Arete considers this acceptable for aggregate quality-improvement feedback) and 0.90 reliability (orange; appropriate for consequential individual decisions such as tiering) when every profile is measured at the same denominator. The horizontal axis matches Figure 2: patients for visits/patient and cost/patient, visits for units/visit, and spinal CMT units for 98941 share. At each tier, only profiles with at least that denominator are included, and each is resampled to exactly that denominator with the patient-cluster bootstrap. At 25 patients (or visits or spinal CMT units), 51% of visits/patient, 74% of cost/patient, 94% of units/visit, and 56% of 98941 profiles reached 0.70, and 9%, 46%, 64%, and 43% reached 0.90. Fewer profiles qualify at each higher tier (for example, 257 of 463 at 25 patients and 39 at 100 patients; Table 2A), so each point describes a different set of profiles, and the 200-patient points rest on only 4 profiles (shown dashed and labeled “small n”; the apparent cost/patient decline from 150 to 200 patients reflects this rather than a loss of reliability). For 98941 share, the share reaching 0.90 stayed near 50% from 50 to 200 spinal CMT units, so added volume alone did not make most of these profiles reliable enough for tiering. Reliability indicates whether a measure distinguishes providers; it does not mean an individual profile is precise relative to a fixed benchmark (Figures 2 and 4). 98941 share is on the percentage scale. Lines connect tested denominators only. Visualization was human generated.

Figure 2. Residual profile uncertainty was also measure-specific and remained material at common denominator volumes. From upper left: visits/patient, CMS-standardized cost/patient, units/visit, and 98941 share of spinal CMT. Each chart shows 90% and 95% empirical bootstrap error intervals pooled across 1,000 replicates per qualifying provider-insurer profile and denominator tier, with patients resampled as clusters. At 25 patients, a visits/patient point estimate of 7.0 visits per patient maps to an empirical 90% uncertainty range of approximately 5.42–8.75 and a 95% range of 5.04–9.29; a $500 cost/patient estimate maps to approximately $342–$666 and $287–$737. At 50 visits, 3.0 units/visit maps to approximately 2.60–3.40 and 2.45–3.53; at 50 spinal CMT units, 65% 98941 share maps to 45.8%–82.0% and 38.8%–87.4%. These worked examples apply the pooled interval to an illustrative value; an individual profile’s range depends on its own patient mix, and because most profiles have a 98941 share near 100%, the pooled 98941 interval is asymmetric. Meeting an administrative or study-defined minimum denominator should not be interpreted as eliminating uncertainty or making interval width irrelevant, particularly when a point estimate is near a tiering decision benchmark. These are empirical bootstrap uncertainty ranges around the observed profile, not universal population confidence intervals. High reliability (Figure 1) and wide uncertainty ranges are not contradictory: reliability describes whether a measure distinguishes providers from one another, while these ranges describe how precisely an individual provider’s value is known relative to a fixed benchmark. The ranges are wide largely because patients differ — repeated visits by the same patient are not independent — so which patients a provider happens to see can move the profile. Boxes beneath each chart show the share of the 463 profiles retained at two example volumes (for example, 55% at 25 patients and 8% at 100 patients). The 200-patient points rest on only 4 profiles and are shown dashed and labeled “small n.” Visualization was human generated.

Figure 3a. Comparison of different provider-insurer “profiles” in the same measurement period: how much can patient mix alone move a chiropractor’s “profile”? Each bar shows how much a “profile’s” value changes when its patients are matched to the network average on that factor, within the same 12-month window (September 2025–August 2026; 463 provider-insurer “profiles”). One in 10 “profiles” changed at least this much. Effects compare patients of the same chiropractor and insurer, so they reflect who the patients are rather than how the chiropractor practices. Patient characteristics (age, sex, HCC risk score), insurance product, and expected copay describe the patient and benefit; factors marked † (entry timing, diagnosis mix, and new patient proportion) also partly reflect practice patterns and coding. Factors overlap, so the combined bar is not the sum of the others. The top 5% bar shows the effect of removing the roughly 5% of patients with the greatest upward influence (“profiles” with at least 20 patients) and is not included in the combined bar. For 98941 share, changes are relative (for example, 12% of an 83% share is about 10 percentage points). Expected copay showed no measurable effect on visits per patient in this window. Values are in Table 6B. These are associations, not causal effects. Visualization was human generated.

Figure 3b. Comparison of the same provider-insurer “profile” during sequential measurement periods: observed contributors to profile movement were measure-specific. From upper left: visits/patient, CMS-standardized cost/patient, units/visit, and 98941 share. Structural contrasts include removal of the approximately 5% of patients with greatest upward influence, the sequential rolling-12-month range, the maximum adjacent-window change, and first-quarter versus fourth-quarter new-patient timing; the first three are shown at their 90th-percentile (P90) values across profiles. Covariate contrasts show average adjusted estimates, from two-way fixed-effects models, for specified increments in patient mix, entry timing, diagnosis prevalence, or expected copay (per $10 for copay). Because P90 structural values and average covariate effects are not like-for-like, the chart overstates the gap between them; however, median rolling ranges (Table 3) also exceeded every single adjusted covariate increment for all four measures. Rolling-range, adjacent-change, and timing values are centered around zero solely to compare magnitude; top-5% removal is one-sided. No single measured factor set explained profile movement consistently across outcomes: model-explained within-profile variation ranged from 0.8% for 98941 share to 13.7% for 98943/visit. Top-5% removal and first-quarter versus fourth-quarter timing are same-period contrasts, retained here for reference; Figure 3a shows them alongside the other same-period factors. Associations are descriptive, not causal. Visualization was human generated.

Figure 4. Reliability and precision are properties of each provider-insurer profile, not of its sample size. From upper left: visits/patient, CMS-standardized cost/patient, units/visit, and 98941 share of spinal CMT. Each dot is one latest-window provider-insurer profile (463; 461 for 98941 share), sized by its number of commercially insured patients (10 to 428). The horizontal axis is the profile’s reliability: how well the measure distinguishes this provider from others within the same insurer and state. The vertical axis is precision: the half-width of the profile’s own 90% patient-cluster bootstrap range, expressed as plus or minus a percentage of the profile’s value and plotted on a log scale, so lower dots are more precise. The vertical lines mark 0.70 and 0.90 reliability. Arete considers reliability below 0.70 unreliable, 0.70–0.89 acceptable for aggregate quality-improvement feedback, and 0.90 or higher appropriate for consequential individual decisions such as tiering (see Methods and [3]). The horizontal line marks ±10%, a study-chosen cutoff that reflects judgment rather than an established standard. Percentages give the share of profiles in each of the six regions. Only 2 visits/patient profiles and 2 cost/patient profiles (0.4% each) were both reliable at 0.90 and precise within ±10%, compared with 67.4% for units/visit and 54.9% for 98941 share; with a ±20% cutoff (sensitivity analysis) the shares were 14.0%, 19.9%, 90.1%, and 62.5%. Profiles with similar patient counts appear in several regions (Table 10). Because most profiles bill 98941 for nearly all spinal CMT, a small relative range for 98941 share partly reflects that ceiling. Nine units/visit profiles whose bootstrap range had zero width are plotted at the 0.1% axis floor. “Precise” describes only how tightly a profile’s own value is estimated. It does not mean the profile is accurate, and a profile inside the ±10% range cannot be read as showing that its comparison with an intermediary network “benchmark,” or its “tier” assignment, is accurate. Benchmarks carry their own uncertainty and were not evaluated here. Benchmark reliability and uncertainty, and the comparison of an individual provider’s performance with a benchmark, form a separate analysis that will be addressed in a subsequent Arete Original Research paper. Visualization was human generated.

Figure 5. How reliable is a “profile” near a “tier” line? Provider-insurer “profiles” in the latest window (463 for each measure) are split at a hypothetical threshold for each measure commonly used in intermediary “profiles” and “tiering” programs, in Arete’s experience: 6.0 visits per patient, $400 CMS-standardized cost per patient, 2.4 units per visit, and 60% 98941 share. The thresholds are illustrative and are not any intermediary’s criteria. Bars show the share of “profiles” on each side that reached 0.90 reliability (purple) and whose 90% uncertainty range was within ±10% of their own value (orange); the ±10% cutoff is a study choice. Above the visits and cost thresholds, no “profile” was both reliable and precise: none of the 95 “profiles” above 6.0 visits reached 0.90 reliability, and for 66% of them the 90% range included 6.0. Units per visit and 98941 share were generally reliable and precise, except for the 70 “profiles” at or below 60% 98941 share. “Profiles” above a utilization threshold tend to have more patient-to-patient variation, which lowers reliability. Even a reliable and precise “profile” above a threshold does not by itself show how the chiropractor practices: its position can also reflect when patients started care, diagnosis mix, new-patient proportion, insurance product, a few influential patients, and which 12-month window is measured (Figures 3a and 3b). Values are in Table 9A. Visualization was human generated.

Arete's Take

At a Glance

Table 1. Study Design and Commercial Insurance Analytic Cohort

The study evaluates measurement behavior and associations; it does not reproduce or validate any specific intermediary’s proprietary profiling or tiering algorithm.

Abstract

Provider profiling can affect network tiering, prior-authorization requirements, and practice operations, yet annual profile measures may reflect patient mix, benefit design, patient lifecycle, influential patients, denominator size, and measurement timing in addition to provider practice patterns. We evaluated seven chiropractic profile measures for commercially insured patients, using patient-de-identified Arete Healthcare EHR/RCM data from Colorado, Minnesota, and Texas. The unit of analysis was provider × insurer × monthly rolling 12-month window; profiles required at least 10 unique patients. We quantified sequential-window variability, influential-patient sensitivity, patient lifecycle and self-pay overlap, multivariable associations, and denominator-specific uncertainty using 1,000-replicate nonparametric bootstrap analyses with patients as the resampling unit. Empirical 90% and 95% signed-error intervals were calculated relative to each observed full profile, and classification reversal was measured for outcomes with analytic reference values. We also estimated signal-to-noise reliability for each profile and the model-based probability of wrong-tier placement at hypothetical tiering thresholds, relative to 0.70 and 0.90 reliability benchmarks.

Across 528 provider-insurer entities with at least three eligible windows, 30.5% crossed an analytic 6-visits-per-patient reference across rolling periods, 9.8% crossed 2.4 units/visit, 12.0% crossed 60% 98941 share, and 29.0% crossed a $400 CMS-standardized cost reference. Median rolling ranges were 2.07 visits/patient, 0.286 units/visit, 7.0 percentage points for 98941 share, and $148 standardized cost/patient. At 20 patients, the bootstrap 90% error interval for visits/patient was −1.71 to +1.93 and the 95% interval was −2.13 to +2.55; 13.8% of bootstrap replicates reversed classification relative to the 6-visit reference. At 50 visits, units/visit had a 90% interval of −0.40 to +0.40 and a 95% interval of −0.55 to +0.53; at 100 visits, 60.7% of profiles remained eligible and the 90% interval narrowed only to −0.30 to +0.30, and no visit- or CMT-based measure met its study-defined error tolerance through 300 visits or spinal CMT units. Removing the approximately 5% of patients with greatest upward influence reduced visits/patient by a median of 0.59 and standardized cost/patient by a median of $39. New-patient share, patient-entry timing, diagnosis mix, and expected copay were associated with selected measures in adjusted models. At each profile’s observed volume, 23.5% of profiles reached 0.90 reliability for visits/patient, 54.4% for cost/patient, 64.1% for 98941 share, and 94.2% for units/visit; with a 50-patient minimum, the corresponding shares were 44.2%, 62.8%, 85.9%, and 99.2%. Because a minimum removes profiles as it raises reliability, the share of all profiles both retained and reliable at 0.90 fell as the minimum rose; at 25 patients it was 16.8% for visits/patient, 33.0% for cost/patient, 38.5% for 98941 share, and 54.2% for units/visit, and 11.7% of profiles met 0.90 on all four measures. At a standardized 25 patients, the mean modeled probability of wrong-tier placement was 14.1% for visits/patient and 7.7% for cost/patient. At each profile’s observed volume, only 0.4% of visits/patient and cost/patient profiles were both reliable at 0.90 and had a 90% uncertainty range within ±10% of their own value, compared with 67.4% for units/visit and 54.9% for 98941 share.

These findings indicate that a provider profile is a time-bounded estimate for which the denominator required to achieve selected precision, the residual uncertainty around the point estimate, and the observed factors associated with movement all differ by measure. The analysis does not establish causal effects or invalidate profiling; it identifies measurement properties, including measure-specific reliability, that should be considered when profiles are used for consequential decisions.

Methods

Background and Prior Work

How provider profiles behave statistically has been studied extensively in medicine, though not, as far as we could find, for chiropractic. Physician-level “report cards” for a chronic disease were found to be unreliable at typical physician panel sizes more than two decades ago [1], and hospital comparisons built on small case volumes face the same problem [2]. The standard way to judge whether a profile can distinguish providers is its reliability: the share of observed differences between providers that reflects true differences rather than chance [3]. Applied to commercial physician cost profiles, that framework showed that low reliability translates into a meaningful risk of placing physicians in the wrong cost tier [4]; those profiles were built from large, multi-plan commercial claims data, so scale alone did not resolve low reliability. Patient caseload [5], patient case-mix and physician-level clustering [6], and whether profiles are built for individual physicians or physician groups [7] all affect how trustworthy a profile is, and funnel plots were developed specifically to account for denominator size when comparing performance [8]. In the same multi-plan data used in [4], cost profiles for physician groups were considerably more reliable than those for individual physicians, but a group’s classification did not predict how its individual physicians performed [7]. Public programs have adopted minimum case counts as a partial safeguard; Medicare’s Merit-based Incentive Payment System (MIPS) cost measures, for example, apply case minimums of 10 or 20 depending on the measure [9]. For health plan physician-measurement programs, the National Committee for Quality Assurance (NCQA) proposed 2026 Physician and Hospital Quality (PHQ) certification standards that, in the draft released for public comment, would require comparisons on cost, resource-use, or utilization measures to rest on a 90% confidence interval or a measure reliability of at least 0.70, and would not permit programs to be designed to act on cost, resource use, or utilization alone, except where standardized quality measures are unavailable for a specialty or a physician’s data are insufficient, in which case the program must disclose prominently that quality performance is not known [10, 11]. PHQ certification applies to organizations that seek it, and the standards are written for physician measurement, so they serve here as a reference point rather than as requirements for chiropractic networks. The final standards were scheduled for release in November 2025 and may differ from the public-comment draft [11].

In chiropractic, visit frequency is known to vary with insurance coverage and chiropractor characteristics [12] and is more sensitive to patient cost sharing than general medical care [13], and outpatient rehabilitation clinics have been benchmarked on functional outcomes with explicit attention to case-mix [14]. In a targeted, non-systematic search, we did not identify a peer-reviewed study of the statistical properties of the utilization measures intermediary networks use to profile chiropractors. This study applies the reliability framework above to chiropractic profile measures and complements it with measure-specific bootstrap uncertainty, rolling-window variation, and influential-patient sensitivity.

Study Design and Research Questions

This retrospective observational study evaluated chiropractic provider profiles for commercially insured patients across sequential rolling 12-month measurement periods. The analysis addressed five questions: (1) how much each profile measure varies for the same provider and insurer across adjacent annual windows; (2) what observation volume is required for empirical stability; (3) how much a small number of influential patients can move each measure; (4) which observed patient, benefit, diagnosis, and lifecycle characteristics are associated with each measure; and (5) whether selected interactions among these factors improve interpretation beyond single-factor associations. Evaluation of specific intermediary network profiles, benchmarks, or tier criteria was out of scope.

Data Sources and Observation Period

The source consisted of patient-de-identified Athena EHR/RCM claim-line extracts for Arete network chiropractors in Colorado, Minnesota, and Texas, drawn from the Arete MSO/CIN EHR/RCM platform and delivered as 12 claim line-item CSV files [15]. At the time of the study, the Athena EHR, and standardized appointment types, intake, workflows and data definitions, were adopted by all practices in the Arete MSO/CIN. Colorado and Minnesota data begin in January 2024; Texas begins in July 2024. The analysis used service history through September 2026; the 20 rolling windows run from February 2024–January 2025 through September 2025–August 2026. Utilization and payment were derived only from the no-copay extracts to preserve the original claim-line grain. Separate expected-copay extracts were used solely as attribute sources after visit-level reconciliation.

Commercial Insurance Population and Analytic Grain

Only providers with a chiropractic specialty in the source system were included; physical therapy and massage therapy providers present in the extracts were excluded. The study population was commercially insured patients, defined by combining the Athena source payer categories “Commercial” and “Blue” (Blue Cross Blue Shield plans), while the individual insurer remained the profiling grain. These categories also include Medicare Advantage (MA) plans administered by those insurers, identified by an Athena product type containing “Medicare.” MA patients were retained because, in Arete’s experience, some intermediary networks combine commercial and MA patients in a single insurer profile. In the latest window, MA patients accounted for 8.3% of patient-profile observations and 8.8% of visits (Table 12). Throughout this report, “commercially insured” and “commercial insurance” refer to this combined population; a sensitivity analysis excluding MA is reported in Results. The same categories also include Medicaid managed care plans (Athena product type “Medicaid HMO”) administered by three insurers, all in Minnesota: 326 latest-window patient-profile observations (1.6%) and 1.3% of visits, in 43 of 463 profiles (median 5.9% of an affected profile’s patients; maximum 18.2%). These patients were also retained, because intermediary networks do not disclose whether Medicaid managed care patients are included in their profiles. Excluding them changed the share of profiles at or above 0.90 reliability by 0.9 percentage points or less for every measure, and no profile fell below the 10-patient floor, so their inclusion is not material to the findings. Each analytic observation represented one provider, one insurer, and one rolling 12-month window. Windows advanced monthly. A provider-insurer-window entered the primary analysis when at least 10 unique commercially insured patients were observed. The 10-patient rule was an analytic eligibility floor aligned with the lower end of operational thresholds encountered by Arete; it was not presumed to represent statistical sufficiency. Provider identity was tracked using the source supervising-provider identifier; NPI was retained as a reconciliation field. Two NPIs mapped to more than one source provider identifier during the observation period, a limitation discussed below.

Visit and Service Construction

A visit was defined as a unique provider-patient-service-date-payer combination. Procedure modifiers were not used to create separate visits. Service units were summed from the source charge-unit field after preserving the no-copay extract as the authoritative utilization dataset. Spinal CMT comprised CPT 98940, 98941, and 98942. 98943, 97110, and 97140 were analyzed as units per visit. Within the Arete MSO/CIN, these treatment services are commonly utilized by chiropractors and in Arete’s experience are commonly the focus of intermediary network “profiles” and “tiering” programs. Patient counts were unique within each provider-insurer-window.

Primary Profile Measures

Seven measures were pre-specified: visits per patient; units per visit; 98941 units as a proportion of all spinal CMT units; CMS-standardized dollars per patient; 98943 units per visit; 97110 units per visit; and 97140 units per visit. The first four correspond to commonly discussed domains of chiropractic provider profiling — frequency, service intensity, spinal CMT mix, and economic utilization — while the latter three describe additional service mix.

CMS-Standardized Economic Measure

To separate utilization from payer- and provider-specific reimbursement, each priced CPT unit was valued using the 2025 CMS Physician Fee Schedule average across 47 localities [16]. CPT 98943 was absent from the supplied fee schedule and was assigned an analytic value equal to 90% of the 98940 amount, or $19.53. Relevant reference values included $21.70 for 98940, $33.24 for 98941, $44.77 for 98942, $29.48 for 97110, and $27.81 for 97140. No geographic adjustment, MPPR, bundling rule, or payer coverage rule was applied. The resulting value is a standardized utilization weight, not expected Medicare reimbursement. Unpriced service units were retained in QA measures but contributed no standardized dollar value.

Expected Copay Reconciliation

Athena exposed three expected-copay fields through separate source relationships. Joining all three directly to claims created one-to-many row multiplication, so copay was processed independently and then merged as a lookup attribute. The claim-payer expected-copay field (Source 3) was prospectively designated primary after development work in Minnesota and Texas. At the provider-patient-payer-date level, positive Source 3 values overrode same-visit zeros; zero-only observations were classified as recorded $0 expected copay; all-null observations were unknown; and multiple positive values were summarized by the median and flagged. Patient-level median expected copay was calculated before provider-insurer aggregation. Sources 1 and 2 were retained for validation. Colorado Source 1 versus Source 3 showed 94.3% exact agreement when both were known and 93.2% exact agreement among positive values, supporting retention of the pre-specified rule.

Patient Attributes, Diagnoses, and Lifecycle Measures

Patient attributes included age, sex, and HCC RAF. Diagnosis-mix variables were constructed from all available ICD-10 positions and included the proportion of patients/visits with M25 and M53 families, with additional degenerative/disc and injury indicators used in sensitivity work. A new-patient proxy was defined by any same-provider CPT 99201–99205 encounter during the measurement window. New-patient timing was the first such encounter date within the window. Overall patient entry timing was the first observed same-provider service date within the window, expressed as days from window start.

Window Completeness and Self-Pay Overlap

For windows with complete observation two full months before and after the 12-month period, continuity variables identified same-provider care in the prior two months, subsequent two months, both, or either. The two-month interval is an observation window, not a clinical episode definition. For commercially insured patients, same-provider self-pay services were identified during the 12-month window and within the bounded ±2-month interval. A paired sensitivity analysis compared insurer-visible profiles with profiles that added observed self-pay services for the same insured patients.

Sequential Rolling-Window Variability

For each provider-insurer entity with at least three eligible rolling windows, we calculated the minimum, maximum, absolute range, median adjacent-window change, and largest adjacent-window change for each measure. Adjacent changes were calculated between consecutive eligible windows; in 20 of 528 entities an intervening window fell below the eligibility floor, so a small number of adjacent pairs span more than one month. For visits/patient, units/visit, 98941 share, and CMS-standardized cost/patient, we also recorded whether the entity crossed analytic reference values of 6 visits/patient, 2.4 units/visit, 60% 98941 share, and $400 standardized cost/patient. These references describe measurement stability and should not be interpreted as reproducing any specific intermediary’s proprietary profiling criteria.

Influential-Patient Sensitivity

In the latest common complete window (September 1, 2025–August 31, 2026), profiles with at least 20 patients were evaluated for patient-level influence. For each measure, patients were ranked by their upward influence on the baseline ratio or mean. We then removed floor(5% × patient count), with a minimum of one patient, and recomputed the profile. This counterfactual quantifies concentration; it does not propose excluding patients from profiling. As a comparator, the same number of patients was removed at random 50 times per profile.

Administrative Eligibility Versus Empirical Stability

Arete reports that intermediary profiling programs it encounters commonly require approximately 10–25 unique patients in a 12-month period before a provider-insurer profile is considered administratively eligible for profiling or tiering. This operational threshold preserves applicability across provider-insurer combinations with relatively low patient volume. It should not be interpreted as evidence that 10–25 patients provides comparable statistical precision for every measure.

The present study therefore separates two concepts: administrative eligibility, which determines whether a profile can be acted on, and empirical stability, which describes how much patient-level sampling variation remains at a given denominator. The analysis does not presume that one should determine the other; instead, it quantifies the inherent tradeoff between broader program reach and more stable measurement. Adams and colleagues reached a similar conclusion for physician cost profiles, recommending that reliability be assessed directly rather than inferred from a minimum sample size; in their data, a 30-episode minimum raised median reliability in 18 of 28 specialties but reduced the number of physicians who could be profiled from 12,789 to 8,689 [4].

Empirical Denominator Stability and Applicability Tradeoff

Denominator uncertainty was evaluated in the latest complete window using a nonparametric bootstrap with replacement as the primary analysis [17]. For each qualifying provider-insurer profile and candidate denominator tier, 1,000 bootstrap replicates were drawn with patients as the resampling unit: each resampled patient contributed all of his or her visits and services, so that repeated visits by the same patient were not treated as independent. For visits/patient and CMS-standardized cost/patient, the tier was a number of patients; for units/visit, 98943/visit, 97110/visit, and 97140/visit, patients were drawn until the tier’s visit count was reached, and for 98941 share until the tier’s spinal CMT unit count was reached (because whole patients were drawn, the realized count slightly exceeded the tier, for example a median of 53 visits at the 50-visit tier). Versions of this report before v1.4 resampled individual visits or CMT units for these measures, which understated their uncertainty (documented in the internal AI Assistance and Reproducibility Record). Profiles contributed to a denominator tier only when the observed denominator was at least that large. Each bootstrap estimate was compared with the full observed profile. The empirical 90% signed-error interval was defined by the 5th and 95th percentiles of bootstrap error pooled across qualifying profiles, and the 95% interval by the 2.5th and 97.5th percentiles. For visits/patient, units/visit, 98941 share, and CMS-standardized cost/patient, classification reversal was the proportion of bootstrap replicates that fell on the opposite side of the analytic reference from the full observed profile. Earlier without-replacement subsampling analyses were retained as sensitivity analyses because finite-panel error narrows mechanically as the sampled denominator approaches the observed panel size. The study also retained its previously specified P90 absolute-error tolerances (±0.5 visits/patient, ±0.15 units/visit, ±5 percentage points for 98941 share, ±$50 standardized cost/patient, and ±0.05 units/visit for 98943, 97110, and 97140) as descriptive precision benchmarks rather than validity standards.

Reliability and Tier Misclassification

Reliability was estimated for each latest-window provider-insurer profile as the ratio of between-profile signal variance to signal plus that profile’s sampling (noise) variance [3, 4]. Signal variance was estimated for each measure by restricted maximum likelihood (REML) in a hierarchical model that included insurer and state fixed effects, so that reliability describes a profile’s ability to distinguish providers within the same insurer and state. Each profile’s sampling variance was estimated from the patient-cluster bootstrap described above. All measures, including 98941 share, were modeled on their original (percentage or rate) scales, the scales on which profiles report them. For the 103 of 462 profiles in which all or none of the spinal CMT units were 98941, the bootstrap variance is zero, so a conservative binomial variance floor (Jeffreys-adjusted and multiplied by the measure’s median patient-cluster design effect of 3.5) was applied. A logit-scale analysis of 98941 share is reported as a sensitivity analysis. The reliability method follows the approach RAND describes for cost and other continuous measures: a hierarchical model fit by REML using each profile’s own score and sampling variance, with the sampling variance computed by bootstrap where no formula applies [3]. Reliability was summarized three ways: at each profile’s observed volume; under a single administrative patient minimum (10, 20, 25, 50, 75, 100, or 150 patients), in which retained profiles contribute all of their visits and spinal CMT units; and at standardized denominators matching Figure 2 (10 to 200 patients for visits/patient and cost/patient; 10 to 300 visits or spinal CMT units for units/visit and 98941 share), for which each profile with at least that denominator was resampled to exactly that denominator with the patient-cluster bootstrap (Figure 1). A patient split-half analysis (100 random splits among profiles with at least 20 patients, Spearman–Brown corrected) was used as a sensitivity check. To describe the combined effect of an administrative minimum, the share of all latest-window profiles that were both retained at each patient minimum and at or above each benchmark was also calculated (the share retained multiplied by the share of retained profiles meeting the benchmark), along with the number of profiles meeting 0.90 on all four primary measures simultaneously.

The RAND provider-profiling tutorial notes that different reliability levels have been advocated for different purposes, citing a psychometric rule of thumb of 0.90 for drawing conclusions about individuals and 0.70–0.80 for conclusions about groups [3]; even at 0.90, it shows that misclassification can remain material. Other sources place the line differently: Hofer and colleagues describe 0.80 as often considered the minimum for decisions about individual physicians [1]; Adams and colleagues note that no consensus exists and describe 0.70 as a commonly used marker of suboptimal reliability [4]; and NCQA’s physician-measurement standards use 0.70 as a minimum [10]. The study reports results against both 0.70 and 0.90, consistent with this use-dependent interpretation (Figures 1 and 4), and does not treat either as a universal standard. Tier misclassification was estimated as the probability that a profile measured at a given denominator would fall on the opposite side of a tiering threshold from its estimated true value. The estimated true value was an empirical-Bayes score (the insurer-and-state expected value plus the reliability-weighted observed deviation), which pulls less reliable profiles toward the average; prior work showed this adjustment is needed because unadjusted profiles overstate differences between physicians [1]; the profile’s own patient-cluster bootstrap errors at that denominator were added to it, and threshold crossings were counted. Profiles with no measurable variation (for example, no 97140 use, or all spinal CMT billed as 98941) received a conservative variance floor. Hypothetical tiering thresholds were 6.0 visits/patient, $400 CMS-standardized cost/patient, 2.4 units/visit, and 60% 98941 share. The Author selected these as a reasonable blend of intermediary tiering criteria; they do not reproduce any specific intermediary’s criteria.

Profile-level reliability and precision (Figure 4; Table 10) were described for the four primary measures in the latest window. For each profile, a separate 2,000-replicate patient-cluster bootstrap at the profile’s own observed volume gave a 90% percentile range. Precision was expressed as the half-width of that range divided by the profile’s value (the relative 90% half-width, reported as plus or minus a percentage of the profile value), so that measures in different units could be compared. Precision complements reliability: reliability describes how well a measure separates providers from one another, while precision describes how closely a single profile’s value is known, which is what matters when that value is compared with a fixed benchmark. Nyweide and colleagues took a precision-based approach (statistical power) for this reason, judging it better aligned with detecting differences from national benchmarks [5]. Each profile was placed in one of six regions defined by its reliability band (below 0.70, 0.70–0.89, or 0.90 and above) and by whether its relative half-width was within ±10%. The ±10% cutoff is a study-chosen judgment, not an established standard, although a national study of physician caseloads also treated a 10% relative difference as meaningful [5]; ±20% is reported as a sensitivity analysis. For the 103 profiles billing all or none of their spinal CMT as 98941, the 98941 range used the same conservative variance floor applied to reliability. One 98941 profile with a 0% share was excluded because a relative range is undefined, and a second had no spinal CMT. Median and interquartile patient counts are not reported for regions with fewer than 11 profiles. Relative precision describes how tightly a profile’s own value is estimated. It does not evaluate any benchmark or tier assignment, which was out of scope.

Because between-provider variation in a larger or more varied network may exceed that in Arete’s network, the sensitivity of reliability to greater between-provider (signal) variance was examined by multiplying each measure’s estimated signal variance by 1.5, 2, and 3 while holding each profile’s own sampling variance fixed. To inform the plausible range, Arete reviewed a small sample of provider profiles and tier notifications from several of the primary intermediary chiropractic networks. The sample was used only to compare measure definitions and reported network statistics. No intermediary network, health plan, or provider is identified, and no intermediary’s methods or criteria were evaluated.

To test whether including MA patients affected the conclusions, the latest-window reliability and precision analysis was repeated from the claim-line data twice with the same methods and a common random seed: once for the study population as reported, and once with MA claims removed (profiles again required at least 10 patients). A third run added each profile’s MA share as a covariate in the model for between-provider variance, to estimate how much of that variance reflected product mix rather than practice.

Association and Multivariable Models

Primary multivariable models were linear two-way fixed-effects regressions including provider-insurer entity fixed effects and common window-end fixed effects. Standard errors were clustered by provider-insurer entity [18]. Core predictors were new-patient share, median patient expected copay, mean age, female share, mean HCC RAF, overall patient entry timing, M25 prevalence, M53 prevalence, and insured-patient self-pay overlap. Models were limited to provider-insurer-windows with known expected copay for at least half of patients, and entities contributing a single window were dropped (7,070 observations; 484 entities). A bounded-window model added prior- and post-window continuity. A separate new-patient timing specification replaced overall entry timing with mean new-patient entry day because the two timing variables were strongly correlated. Effects were scaled to interpretable increments: 10 percentage points for shares, $10 copay, 10 years age, 0.10 RAF, 30 days timing, and 5 percentage points self-pay overlap. For relative importance, each coefficient was also expressed as a within-entity standardized beta. Partial R-squared quantified the incremental within-profile variance explained by the covariate block beyond provider-insurer and common-time fixed effects. Because adjacent annual windows overlap by 11 months, p-values and confidence intervals were treated as descriptive inferential aids rather than evidence of independent repeated experiments.

Insurance product mix was examined separately because product type varies far more between provider-insurer entities than within an entity over time (standard deviation of POS/EPO share across entities 20.1 percentage points, versus 3.6 within an entity across windows), so the two-way fixed-effects models above cannot detect it; adding product shares to those models produced no distinguishable associations. In the latest window, each patient-profile observation was assigned the patient’s most frequent Athena product type in that profile: PPO, HMO, POS/EPO, or not specified (Athena “OTHER”). Medicare Advantage and Medicaid managed care observations were excluded from this contrast only. Patient-level regressions with provider and insurer fixed effects and state compared patients of the same chiropractor and insurer by product type, with standard errors clustered by provider (visit-weighted for units/visit and spinal-CMT-weighted for 98941 share). Differences were also scaled to a 10-percentage-point shift in product share, matching the covariate increments above, and each profile’s expected shift from its own product mix, relative to the network’s average mix, was summarized at the 90th percentile across profiles.

Figure 3a compares different provider-insurer profiles in the same measurement period (the latest window). For each profile, we estimated how much its value would change if its patients matched the network average on each factor, holding the chiropractor’s own pattern fixed (indirect case-mix standardization). Patient-level regressions included all factors jointly with profile fixed effects, so each factor’s effect compares patients of the same chiropractor and insurer: patient characteristics (age band, sex, HCC RAF quartile); insurance product (PPO, HMO, POS/EPO, not specified, Medicare Advantage, Medicaid managed care); expected copay (Source 3 rule above; $0, $1–$40, $41–$100, more than $100); diagnosis mix (any M25, any M53); new patient (any same-provider 99201–99205 encounter in the window); and entry timing (quarter of first same-provider service in the window). Patients with no recorded copay (25.7%) were identified by a separate indicator that was excluded from every adjustment, because unrecorded copay is mechanically associated with fewer visits. Per-visit measures were weighted by visits and 98941 share by spinal CMT units. A profile’s shift for each factor is the average fitted contribution of that factor across its patients minus the network average; Figure 3a and Table 6B report the 90th percentile across profiles of the absolute shift, as a percentage of the profile’s value. Entry timing, diagnosis mix, and new-patient proportion are reported separately from patient and benefit factors because they may partly reflect practice patterns and coding. The share of variation among profiles explained was the reduction in the variance of profile values after adjustment, allocated across factors by Shapley averaging so that items sum to the total, computed across all 463 profiles and among profiles of the same insurer and state. Sampling noise was estimated as the mean patient-cluster bootstrap variance divided by the observed variance. Figure 3b compares the same provider-insurer profile across sequential measurement periods, using the rolling-window and two-way fixed-effects analyses described above.

Selected Interaction Terms

A multivariable model is useful because profile factors co-vary and single-factor associations can attenuate or reverse after simultaneous adjustment. However, an exhaustive all-pairs interaction search was not used: nine core predictors would generate dozens of interaction terms, increasing multicollinearity, multiple-testing burden, and interpretive instability. Three clinically or operationally motivated interactions were pre-selected for visits/patient: new-patient share × new-patient timing, new-patient share × expected copay, and M25 × M53 diagnosis prevalence. Interactions were treated as exploratory.

Analyses are documented in the internal reproducibility workbook [19]; reliability, patient-cluster bootstrap, and tier-misclassification analyses are documented in a separate reproducibility workbook [20]; and an independent re-execution of the profile construction, patient-cluster bootstrap, and reliability estimates from the claim-line extracts, including the percentage-scale 98941 analysis, is documented in [21].

Independent recomputation of the reported statistics, and the reconciliation of discrepancies against earlier drafts, are documented in the internal AI Assistance and Reproducibility Record maintained under the Arete Research Library Governance Framework.

Results

Observation Volume, Administrative Eligibility, and Measure-Level Stability

The rolling dataset contained 8,225 eligible provider-insurer-window observations across 20 monthly endpoints. The latest common complete window included 463 provider-insurer profiles meeting the study’s ≥10-patient eligibility floor and 90 unique NPIs, representing 19,904 patient-profile observations, 97,234 visits, 279,453 service units, and 90,025 spinal CMT units. Table 2 summarizes bootstrap uncertainty at common lower-volume conditions, Table 2A shows the operational loss of profile applicability as denominator thresholds rise, and the accompanying patient- and visit-volume results describe the joint tradeoff among profile reach, numerical uncertainty, and classification stability.

Table 2. Measure-Specific Bootstrap Uncertainty at Common Lower-Volume Conditions

Intervals are signed bootstrap error relative to the observed full provider-insurer profile. The 90% interval is the 5th–95th percentile and the 95% interval is the 2.5th–97.5th percentile across pooled profile-level bootstrap replicates. Reference reversal is reported only for measures with an analytic comparator. Patients were resampled as clusters for all measures; for visit- and CMT-based measures the realized denominator slightly exceeds the tier because whole patients are drawn. Patient-based rows retain the v0.5a workbook values, which the independent patient-cluster rerun closely reproduced (differences of about 0.02 visits/patient and $2 or less). These are empirical bootstrap uncertainty intervals, not universal population confidence intervals or validity thresholds.

Table 2A. Administrative Eligibility Versus Profile Applicability

At the latest common window, increasing the minimum patient count from 10 to 25, 50, 100, and 150 retained 55.5%, 27.9%, 8.4%, and 2.6% of provider-insurer profiles, respectively. The bootstrap intervals show that uncertainty narrows as denominators increase, but the reduction in eligible profiles is steep. For visits/patient, the 90% error interval narrowed from −2.30 to +2.66 visits at 10 patients to −0.67 to +0.70 at 150 patients; no tested denominator through 200 met the study-defined ±0.50 P90 absolute-error tolerance. For units/visit, thresholds of 25, 50, 100, and 150 visits retained 97.6%, 86.4%, 60.7%, and 44.3% of profiles. With patients resampled as clusters, the units/visit 90% error interval narrowed from −0.745 to +0.724 at 10 visits to −0.213 to +0.217 at 300 visits, which only 20.5% of profiles reached, and never met the study-defined ±0.15 P90 tolerance. No other visit- or CMT-based measure met its tolerance through 300 visits or spinal CMT units either; P90 errors at 300 were 8.9 percentage points for 98941 share, 0.091 for 98943/visit, 0.083 for 97110/visit, and 0.058 for 97140/visit. Treating each visit as an independent observation, as earlier versions of this analysis did, understated the variance of these measures by a median factor of approximately 2.9 to 3.7 (documented in the internal AI Assistance and Reproducibility Record). Standardized cost/patient first met the study-defined ±$50 P90 tolerance at 150 patients, but only 12 profiles qualified, and at 200 patients (4 profiles) the P90 error ($56) again exceeded the tolerance; this crossing should therefore not be treated as robust. Administrative eligibility and statistical uncertainty therefore create an inherent, measure-specific tradeoff rather than one universal minimum-volume rule.

Sequential Rolling 12-Month Variation

Profile values varied materially for the same provider and insurer as the annual measurement window advanced by one month. Table 3 separates the total within-entity rolling range from adjacent-window movement. The P90 maximum adjacent change captures the 90th percentile across entities of each entity’s single largest absolute change between adjacent eligible rolling windows, which ordinarily share 11 of 12 months.

Table 3. Sequential Rolling-Window Variation

Analytic references: 6 visits/patient, 2.4 units/visit, 60% 98941 share, and $400 CMS-standardized cost/patient. These are used to show whether timing can change a binary characterization, not to replicate a proprietary tiering algorithm. Range = maximum minus minimum across an entity’s eligible windows.

Influence of the Top 5% of Patients

In 321 latest-window provider-insurer profiles with at least 20 patients, targeted removal of the approximately 5% of patients with greatest upward influence produced larger median changes than equal-sized random patient removal for every measure except 97140/visit, for which the median change was zero under both approaches. Median reductions were 0.59 visits/patient, 0.069 units/visit, 1.53 percentage points of 98941 share, and $39 standardized cost/patient (12.8%, 2.8%, 1.7%, and 12.5% of the median profile’s own value, respectively); corresponding P90 reductions were 1.11, 0.162, 6.58 percentage points, and $119. Among profiles initially at or above the 6-visits/patient analytic reference, 29 of 64 (45.3%) fell below it after counterfactual removal. Earlier research found a similar pattern in physician profiles for diabetes care: physicians with the worst glycemic-control profiles could substantially improve them by no longer treating the 1 to 3 patients with the highest prior-year HbA1c, and detailed case-mix adjustment did not prevent this [1]. When a few patients can move a profile this much, profiling can create an incentive to avoid complex or high-need patients [1].

Table 4. Counterfactual Influence of the Top 5% of Patients

*97140 is highly zero-inflated; the random-removal median often equals zero, making a ratio unstable. Reference crossing counts profiles at or above the reference before removal. Targeted removal is a sensitivity analysis and does not imply that high-impact patients should be excluded from profiling.

Patient Lifecycle and New-Patient Timing

New and ongoing patient populations had different utilization patterns. New-patient proxies averaged fewer visits/patient than ongoing patients (4.61 versus 5.16) but greater service intensity (3.21 versus 2.77 units/visit), higher CMS-standardized cost/patient ($443 versus $407), more 98943 and 97110 use, and a lower 98941 share. Timing within the fixed annual window also mattered. Among commercially insured new-patient proxies, Q1 entrants averaged 5.82 visits/patient and $555 standardized cost/patient versus 3.10 visits and $308 for Q4 entrants, while units/visit remained nearly unchanged. Table 5A reports the Q1–Q4 contrast for all seven measures; these are descriptive timing contrasts, not adjusted causal effects.

Table 5. New-Patient Proxy Versus Ongoing-Patient Utilization

Table 5A. New-Patient Timing: Q1 Versus Q4 Entry Within the Measurement Year

New-patient status is a claims proxy based on a same-provider 99201–99205 encounter within the window; it is not a definitive episode-of-care classification. Q1 n=1,663 and Q4 n=1,438 new-patient proxies.

Window Completeness and Self-Pay Visibility

Among fully observable commercial insurance windows with the ±2-month boundary, 36.1% of patients had same-provider care immediately before or after the nominal 12-month period. Same-provider self-pay services were observed for 4.7% of insured patients during the measurement year and 5.6% when the ±2-month interval was included. In paired profiles that added same-provider self-pay services for the same insured patients, median changes in the seven profile measures were generally modest; however, the analysis demonstrates that insurer-visible claims do not always contain the complete observed care history.

Adjusted Multivariable Associations

Table 6 summarizes statistically distinguishable associations from the core multivariable fixed-effects models. The same provider-insurer entity was compared with itself over time while controlling for common window-end effects. Standardized within-entity weights show the relative magnitude of covariate movement on each outcome. Observable factors explained substantially different amounts of within-profile variation by measure: 13.7% for 98943/visit, 10.5% for visits/patient, 10.1% for units/visit, 7.9% for standardized cost/patient, 5.2% for 97110/visit, 3.5% for 97140/visit, and only 0.8% for 98941 share.

Table 6. Significant Adjusted Associations in Core Multivariable Models

Core models included new-patient share, median expected copay, mean age, female share, mean HCC RAF, overall patient entry timing, M25 prevalence, M53 prevalence, and insured-patient self-pay overlap (7,070 provider-window observations; 484 entities; 98941 share 7,067 and 483). Full coefficients, including non-significant covariates, are reported in Appendix A. Standardized beta is based on within-provider-insurer variation. The positive within-entity association of new-patient share with visits/patient is conditional on entry timing and should not be read against the unadjusted new-versus-ongoing comparison in Table 5.

Insurance Product Mix

Within the same chiropractor and insurer, POS/EPO patients averaged 0.97 more visits and $102 more CMS-standardized cost than PPO patients (Table 6A). For a profile, a 10-percentage-point shift from PPO to POS/EPO patients corresponds to about +0.10 visits/patient and +$10 cost/patient: smaller per increment than new-patient share (+0.246 visits; +$22.4) and comparable to M25 prevalence (+0.151 visits; +$21.9) in Table 6. HMO patients did not differ from PPO patients on visits or cost but had slightly more units per visit (+0.18). No product contrast was distinguishable for 98941 share. Because product mix differs mainly between profiles rather than over time, it acts as a fixed offset: at the 90th percentile, a profile’s own product mix shifted its value by 0.42 visits/patient (7.3%) and $43 cost/patient (10.4%) relative to the network’s average mix, and by less than 1% for units/visit and about 1 percentage point for 98941 share. In Arete’s experience, intermediary networks do not segment profiles by product or disclose product mix, so two chiropractors with the same practice pattern could receive different visits/patient and cost/patient values because of the plans their patients hold.

Table 6A. Insurance Product Mix: Differences From PPO Patients of the Same Chiropractor and Insurer

Latest window; 17,938 commercial-product patient-profile observations in 454 profiles (PPO 28.3%, HMO 2.2%, POS/EPO 8.8%, not specified 60.7%); Medicare Advantage and Medicaid managed care observations excluded from this contrast. Patient-level regressions with provider and insurer fixed effects and state; standard errors clustered by provider. P90 profile mix shift is the 90th percentile across profiles of the absolute difference between a profile’s expected value at its own product mix and at the network’s average mix (percent of the profile’s value in parentheses). Product type is from Athena and was not verified against insurer enrollment. Data: Figure3_ProductMix_Data.xlsx.

Different Profiles in the Same Measurement Period (Figure 3a)

Comparing different chiropractors in the same window, patient mix moved individual profiles substantially (Table 6B). For 1 in 10 profiles, all measured factors combined changed the value by at least 41.5% for visits/patient, 60.7% for cost/patient, 13.5% for units/visit, and 5.3% for 98941 share. Entry timing was the largest single factor for visits/patient (33.4%) and cost/patient (46.0%): patients whose first visit fell in the fourth quarter of the window averaged about 4 fewer visits than patients of the same chiropractor and insurer who started in the first quarter, so a profile weighted toward late-year entrants appears low because of measurement timing. Diagnosis mix was next (16.2% and 32.6%). Insurance product and patient characteristics moved profiles by about 6% to 10%, and expected copay had little effect. Removing the 5% most influential patients changed 1 in 10 profiles by 20.5%, 19.5%, 5.9%, and 12.3%, and was the largest factor for 98941 share. After adjustment for all factors, 31 of 463 profiles crossed the hypothetical visits/patient threshold of 6.0 (18 moved above and 13 below), 32 crossed the $400 cost threshold, 10 crossed 2.4 units/visit, and 8 crossed 60% 98941 share; counting only patient and benefit factors, 10, 4, 5, and 2 profiles crossed.

Although measured patient mix moved some profiles substantially, it explained a modest share of the overall variation among profiles (Table 6C): 6.3%, 19.2%, 13.7%, and 3.5% across all profiles, and 1.1%, 7.5%, 8.4%, and 2.9% among profiles of the same insurer and state. Mix effects run in both directions, so adjustment reorders profiles more than it narrows their spread. Among profiles of the same insurer and state, sampling noise accounted for 30.5% of the variation in visits/patient, 19.2% in cost/patient, 3.8% in units/visit, and 17.2% in 98941 share. The remaining variation reflects provider practice together with patient factors not measured here, such as severity and treatment goals; it is an upper bound on, not an estimate of, the contribution of practice patterns.

Table 6B. How Much Patient Mix Moves a Profile: Different Profiles in the Same Measurement Period

Latest window; 463 profiles; 19,904 patient-profile observations. Percentages are the 90th percentile across profiles of the absolute change in a profile’s value when its patients are matched to the network average on that factor (percent of the profile’s value; relative for 98941 share). † Factors that may partly reflect practice patterns and coding. Top 5% removal uses profiles with at least 20 patients and is not part of the combined adjustment. Hypothetical thresholds (Table 9): 6.0 visits/patient, $400, 2.4 units/visit, 60% 98941 share. Single model run; crossing counts do not include sampling uncertainty. Data: Figure3a_SameWindow_Data.xlsx.

Table 6C. Share of Variation Among Profiles Explained by Measured Patient Mix and by Sampling Noise

Measured patient mix = reduction in the variance of profile values after adjustment, allocated across factors by Shapley averaging (factor-level values in the data workbook). Same insurer and state = variation around each insurer-state mean (36 insurer-state groups). Sampling noise = mean patient-cluster bootstrap variance divided by the observed variance; mix and noise are treated as non-overlapping, so the decomposition is approximate. The remainder is an upper bound on the contribution of practice patterns, because it also contains patient factors not measured here.

Additional Bounded-Window and Copay Findings

In the bounded ±2-month model, a 10-percentage-point higher share of patients with prior-period care was associated with 0.078 fewer visits/patient and $8.69 lower standardized cost/patient. A 10-percentage-point higher post-window continuity share was associated with a 0.37-percentage-point higher 98941 share. These associations quantify measurable lifecycle effects but should not be interpreted as causal.

At the patient level, within provider-payer groups and adjusting for age, RAF, sex, and diagnosis mix, expected copay showed an inverse relationship with annual visit frequency in calendar year 2025: patients with a recorded $41–$100 expected copay averaged approximately 0.47 fewer annual visits than patients with a recorded $0 expected copay, and in a sensitivity analysis limited to expected copays ≤$100, each additional $10 was associated with approximately 0.06 fewer visits. The calendar-2025 association was statistically distinguishable only in Texas (694 patients), and it did not hold in the latest measurement window (September 2025–August 2026), where patients with a $41–$100 copay averaged about 0.1 more visits than patients with a $0 copay (not significant). The copay association is therefore window-dependent and should not be treated as a stable effect. In the fully adjusted provider-profile model, the median copay coefficient for visits/patient attenuated and was not independently significant, suggesting that copay is relevant to patient-level utilization but overlaps with other profile characteristics at the provider-window level. The calendar-2025 association is consistent with earlier evidence that chiropractic use is sensitive to cost sharing [13].

Selected Interactions

Three pre-specified interactions were tested in the new-patient timing model for visits/patient. Higher new-patient share amplified the downward association between later new-patient entry and visits/patient. An interaction between new-patient share and expected copay was also statistically distinguishable from zero but smaller and should be considered exploratory. No evidence of an M25 × M53 interaction was observed.

Table 7. Selected Interaction Terms for Visits per Patient

Scaled coefficient units correspond to the model increments used for each main effect (6,683 observations; 466 entities). Interactions were pre-selected; no exhaustive pairwise interaction search was performed.

Reliability and Tier Misclassification

Reliability differed sharply by measure (Table 8). At each profile’s observed volume, median reliability was 0.79 for visits/patient, 0.92 for CMS-standardized cost/patient, 0.95 for 98941 share, and 0.98 for units/visit. The share of profiles reaching 0.90 was 23.5%, 54.4%, 64.1%, and 94.2%, respectively; the share reaching 0.70 was 69.3%, 81.2%, 91.6%, and 99.6%. Applying a single administrative patient minimum raised reliability for the profiles that remained, but unevenly: at a 25-patient minimum (257 of 463 profiles retained; median 50 patients, 247 visits, and 231 spinal CMT units), 30.4% of visits/patient profiles, 59.5% of cost/patient profiles, 69.5% of 98941 profiles, and 97.7% of units/visit profiles reached 0.90. At a 50-patient minimum (129 profiles), the shares were 44.2%, 62.8%, 85.9%, and 99.2%. On the logit scale (sensitivity analysis), 98941 share reached 0.90 in 23.2% of profiles at observed volume, 40.2% at a 25-patient minimum, and 48.4% at 50. The same patient-count eligibility rule therefore produced materially different reliability across the measures in a profile.

Measured at the same denominator, reliability rose with volume but at very different rates by measure (Figure 1). At 10 patients (or visits or spinal CMT units), the share of profiles reaching 0.90 was 6.0% for visits/patient, 30.5% for cost/patient, 34.1% for units/visit, and 32.0% for 98941 share; at 25 it was 8.6%, 45.9%, 63.9%, and 43.3%; and at 50 it was 14.0%, 53.5%, 80.8%, and 47.9%. Visits/patient did not reach 0.90 in a majority of profiles until 150 patients (75.0% of 12 profiles), a volume that only 2.6% of profiles reached. Units/visit reached 0.90 in a majority of profiles at 20 visits. For 98941 share, the share reaching 0.90 changed little between 50 and 200 spinal CMT units (47.9% to 57.1%). The profiles qualifying at each tier differ, and the highest patient tiers rest on few profiles (12 at 150 and 4 at 200 patients).

Because a patient minimum removes profiles as it raises reliability, the share of all profiles that were both retained and reliable fell steadily as the minimum rose (Table 8A; Figure 1). At 0.90, it fell from 23.5% at the 10-patient study floor to 16.8% at 25 patients, 12.3% at 50, and 2.4% at 150 for visits/patient, and from 54.4% to 33.0%, 17.5%, and 2.4% for cost/patient. A minimum excludes unreliable profiles; it does not make them reliable. Only 5 of the 28 measure-minimum combinations in Table 8A exceeded 50%, all at minimums of 25 patients or fewer and three of them for units/visit. Only 69 profiles (14.9%) reached 0.90 on all four primary measures simultaneously at the study floor, and 54 (11.7%) at a 25-patient minimum.

Table 8A. Share of All Profiles Both Retained and Reliable at 0.90, by Minimum Patients per Profile

Percentages are of all 463 latest-window provider-insurer profiles (462 for 98941 share); counts in parentheses. Retained profiles contribute all of their visits and spinal CMT units. “All four measures” counts profiles at or above 0.90 on visits/patient, cost/patient, units/visit, and 98941 share simultaneously. † 98941 share on the percentage scale; on the logit scale (sensitivity analysis), the 98941 values were 23.2%, 22.9%, 22.3%, 13.4%, 9.7%, 5.4%, and 1.9%. Values at 150 patients rest on 12 profiles.

Units/visit illustrates why reliability and absolute precision are separate properties. Its reliability was high because provider profiles differ substantially from one another on this measure, yet its absolute bootstrap uncertainty did not meet the ±0.15 study tolerance through 300 visits (Table 2). Reliability describes how well a profile ranks providers; absolute uncertainty describes how close a single value is to a fixed benchmark. 98941 share shows the same pattern. Most 98941 profiles reached 0.90 reliability because providers differed widely on this measure (most near 90%–100%, some far lower), yet at 50 spinal CMT units the 90% uncertainty range around an illustrative 65% value was 45.8%–82.0% (Table 2). For a tier decision at a fixed 60% threshold, what matters is that absolute range: the typical profile, far above 60%, was rarely at risk, while the 10% of profiles most at risk each carried a wrong-tier probability of at least about 34%–36% (Table 9). The width reflects differences among patients — some patients receive 98941 at nearly every visit and others rarely — so a provider’s measured share depends substantially on which patients were seen in the measurement period. When profiles were projected to a standardized 25 patients (or 25 visits or spinal CMT units), median reliability was 0.71 for visits/patient, 0.88 for cost/patient, 0.80 for 98941 share, and 0.93 for units/visit, and the share reaching 0.90 was 8.6%, 45.9%, 43.3%, and 63.9%.

Table 8. Share of Provider-Insurer Profiles Meeting 0.70 and 0.90 Reliability

Profiles: 463 at observed volume (462 for 98941 share); 257 at the 25-patient minimum (256 for 98941); 129 at the 50-patient minimum (128 for 98941). Reliability = signal variance ÷ (signal variance + the profile’s patient-cluster sampling variance), with signal variance estimated by REML within insurer and state. † 98941 share on the percentage scale, with a variance floor for the 103 profiles billing all or none of their spinal CMT as 98941. On the logit scale (sensitivity analysis), median reliability was 0.79 and the shares were 65.6% / 23.2%, 76.2% / 40.2%, and 83.6% / 48.4%. The patient split-half estimate (median Spearman–Brown 0.91) is consistent with the percentage-scale results. *97140/visit values near 1.00 reflect the many profiles with no 97140 use (a variance floor) and are not interpreted.

Modeled wrong-tier probabilities followed a similar pattern (Table 9). At a standardized 25 patients, the mean probability that a profile was on the wrong side of the hypothetical threshold was 14.1% for visits/patient at 6.0 and 7.7% for cost/patient at $400; at a standardized 25 visits or spinal CMT units, it was 5.3% for units/visit at 2.4 and 10.3% for 98941 share at 60%. The risk was concentrated in profiles close to the threshold: for every measure, the median profile had a much lower probability than the mean, while the 10% of profiles most at risk each carried a wrong-tier probability of at least roughly 21% to 37%, depending on the measure. For 98941 share, whose profiles mostly sit well above 60%, misclassification was more often false-low. For visits/patient at 25 patients, misclassification was more often false-high (a profile whose estimated true value was below the threshold measuring above it) than false-low. In the patient split-half sensitivity analysis (321 profiles with at least 20 patients), median Spearman–Brown reliability was 0.77 for visits/patient, 0.86 for cost/patient, 0.97 for units/visit, and 0.91 for 98941 share.

Table 9. Modeled Probability of Wrong-Tier Placement at Hypothetical Tiering Thresholds

Thresholds are hypothetical and were selected by the Author as a reasonable blend of intermediary tiering criteria; they do not reproduce any specific intermediary’s criteria. Wrong-tier probability combines each profile’s empirical-Bayes estimated true value with its own patient-cluster bootstrap errors and depends on the estimated signal variance. False-high: estimated true value below the threshold, measured value above it; false-low: the reverse. Profiles are those whose observed denominator reached the standardized level. 98941 rows use the percentage-scale model (documented in the internal AI Assistance and Reproducibility Record).

Figure 5 and Table 9A apply the same hypothetical thresholds to each profile at its observed volume, separating profiles above and below each line. For visits/patient and cost/patient, profiles above the threshold were the least reliable: none of the 95 profiles above 6.0 visits/patient reached 0.90 reliability (median 0.57), compared with 30.2% of the 368 profiles at or below it, and 10.7% of the 196 profiles above $400 compared with 86.5% of those at or below it. No profile on either side was precise within ±10% for either measure (at most 0.7%). For 66.3% of profiles above 6.0 visits/patient and 33.7% of profiles above $400, the profile’s own 90% range included the threshold, so its side of the line was not established at that level of confidence. Units/visit and 98941 share were generally reliable and precise on both sides, except for the 70 profiles at or below 60% 98941 share, of which 20.0% reached 0.90 and none was precise within ±10%. Reliability and precision describe how well a profile’s value is measured, not why it is high or low; a reliable and precise profile above a threshold can still reflect the patient-mix, timing, and influential-patient factors in Figures 3a and 3b.

Table 9A. Reliability and Precision Above and Below Hypothetical Thresholds

Latest window; 463 profiles per measure, each at its observed volume. Above = strictly above the threshold. Reliability from the same re-run as Table 13 (REML with insurer and state fixed effects; 98941 on the percentage scale). Precision = half-width of the profile’s 90% patient-cluster bootstrap range, as a percentage of its value. Thresholds as in Table 9; they are hypothetical and do not reproduce any intermediary’s criteria. 98941 share includes one profile with a single spinal CMT unit that other 98941 analyses omit (documented in the internal AI Assistance and Reproducibility Record).

Profile-Level Reliability and Precision

Figure 4 and Table 10 place every latest-window profile by its own reliability and precision. Precision differed sharply by measure: the median profile’s 90% range was about ±26% of its value for visits/patient and for cost/patient, compared with about ±8% for units/visit and for 98941 share. Wide ranges for cost are consistent with prior work: a national study of Medicare primary care estimated that detecting a 10% difference in ambulatory costs would require about 2,500 patients, because individual patients’ costs vary widely even after risk adjustment [5]. Only 2 visits/patient profiles and 2 cost/patient profiles (0.4% each) were both reliable at 0.90 and precise within ±10%. A further 23.1% of visits/patient profiles and 54.0% of cost/patient profiles were reliable at 0.90 but less precise than ±10%. For units/visit and 98941 share, 67.4% and 54.9% of profiles fell in the reliable-and-precise region. Apart from 2 units/visit profiles in the 0.70–0.89 band, no profile below 0.90 reliability was precise within ±10% for any measure.

Reliability and precision behaved as properties of each profile rather than of its sample size. Patient count was only moderately related to reliability (Spearman rank correlation 0.27 to 0.48 across the four measures). Among the 107 visits/patient profiles with 20 to 30 patients, reliability ranged from 0.14 to 0.99, and 11.2% reached 0.90. This is consistent with RAND’s observations that reliability varies from provider to provider even at equal sample sizes and that minimum sample-size rules alone may not ensure reliable comparisons [3], and with physician cost profiles, in which reliability varied widely even among physicians with many episodes (for example, 100) [4]. Patient counts overlapped across regions (Table 10). For visits/patient, the middle half of profiles reliable at 0.90 had 22 to 90 patients, while the middle half of profiles below 0.70 had 13 to 27. For 98941 share, profiles that were reliable but imprecise had more patients (median 83) than profiles that were reliable and precise (median 27). Precision tracked patient count more closely for visits/patient and cost/patient (rank correlations −0.77 and −0.74) than for units/visit and 98941 share (−0.44 and −0.21), yet even the largest visits/patient and cost/patient profiles rarely reached ±10%. The same provider could also fall on different sides of 0.90 reliability with different insurers. Among 84 providers with profiles for two or more insurers, 57.1% had visits/patient profiles on both sides of 0.90 reliability (cost/patient 46.4%, 98941 share 60.7%, units/visit 22.6%). The median within-provider range of visits/patient reliability across insurers was 0.31. A single patient minimum applied uniformly across a provider’s insurers therefore does not produce uniform reliability or precision.

Table 10. Provider-Insurer Profiles by Reliability Band and Precision (±10%)

Percentages are of 463 latest-window profiles (461 for 98941 share). Precision is the half-width of the profile’s own 90% patient-cluster bootstrap range as a percentage of its value; ±10% is a study-chosen cutoff. Patient medians and interquartile ranges (IQR) are shown for the precise region only where it held 11 or more profiles. *Fewer than 11 profiles; patient counts not reported. † Percentage scale, with the variance floor for profiles billing all or none of their spinal CMT as 98941. Reliability bands follow Arete’s use-dependent interpretation (Methods); they are not universal standards.

Sensitivity to Greater Between-Provider Variation

The reviewed intermediary profiles used measures similar to those studied here: visits per patient, services per visit, 98941 share (in one network combined with 98942), and an allowed or weighted cost per patient. One network reported the network-wide standard deviation for each measure. For visits per patient, the reported standard deviation was 1.90 to 2.44 across three health-plan networks of 106 to 1,120 providers, compared with 1.95 in this study. A reported standard deviation includes both true differences between providers and chance variation, so it sets an upper limit on true between-provider variation. If all of the reported variation reflected true differences, between-provider variance would be 1.9 to 3.1 times that estimated here; if the reported and study networks carry a similar share of chance variation, it would be 1.0 to 1.6 times as large. For 98941 share combined with 98942, the corresponding upper limit was 1.7 to 1.8 times. Cost and services measures were not comparable because of differences in definition.

Table 11 shows how reliability would change. With between-provider variance 1.5 times that estimated here, the share of visits/patient profiles reaching 0.90 would rise from 23.5% to 34.3% at actual volume and from 8.6% to 15.2% at a standardized 25 patients; even at the upper limit (3 times), it would reach 60.9% and 38.1%. Precision would not change under any of these scenarios, because each profile’s uncertainty range depends only on that provider’s own patients. Some reviewed profiles reported values computed from fewer than 10 patients with a single health plan, below the smallest denominator evaluated in this study.

Table 11. Share of Profiles Reaching 0.70 and 0.90 Reliability Under Greater Between-Provider Variation

Each profile’s own sampling variance is held fixed; only the between-provider (signal) variance is multiplied. Standardized 25 = 25 patients for visits/patient and cost/patient, 25 visits for units/visit, and 25 spinal CMT units for 98941 share (Figure 1). Multipliers are informed by network standard deviations reported in a sample of intermediary profiles (text); for visits/patient, about 1.0–1.6× is plausible and 3× is an upper limit. † Percentage scale. Precision (Figures 2 and 4) does not change under these scenarios.

Medicare Advantage Patients and Sensitivity to Their Exclusion

MA patients were present in 181 of the 463 latest-window profiles; 122 profiles were at least 10% MA, 50 at least 25%, and 15 at least 50%. MA patients differed from other patients mainly on service-mix measures (Table 12). Their units per visit were about half those of other patients (1.46 versus 3.01), their 97110, 97140, and 98943 use was much lower, and their CMS-standardized cost per patient was lower ($207 versus $423), consistent with narrower Medicare coverage of chiropractic services. Visits per patient and 98941 share were similar.

Table 12. Medicare Advantage and Other Commercially Insured Patients in Latest-Window Profiles

Latest common window (September 2025–August 2026), 463 profiles with at least 10 patients. MA identified by an Athena product type containing “Medicare” within the Commercial and Blue categories. Patient-level rates are pooled across profiles (total visits, units, or cost divided by total patient-profile observations or visits). Nineteen patients had both MA and other commercial claims within the same profile and are counted in both columns.

Excluding MA patients did not change the study’s conclusions (Table 13). The share of profiles reaching 0.90 reliability changed by 2 percentage points or less for every measure, in both directions, which is within about one standard error of those shares (approximately 2 percentage points at roughly 450 profiles). Median precision changed by less than 1 percentage point. The largest change was for units per visit, where the share of profiles both reliable at 0.90 and precise within ±10% rose from 66.7% to 71.3% without MA, consistent with MA patients adding variation to that measure. Adding MA share to the model reduced the estimated between-provider variance by about 4% for units per visit and by less elsewhere, so product mix accounted for little of the differences between providers. Twenty-one profiles fell below the 10-patient floor when MA was removed. For the 32 profiles that were at least 25% MA, removing MA did change profile values materially: the median change was +11.5% for units per visit, −8.7% for visits per patient, and +6.6% for cost per patient. Product mix therefore matters for individual profiles with many MA patients even though it does not change the network-level findings.

Table 13. Reliability and Precision With and Without Medicare Advantage Patients

Both rows for each measure come from the same re-run (2,000 patient-cluster bootstrap replicates; REML with insurer and state fixed effects; 98941 on the percentage scale with the variance floor described in Methods). The “as reported” re-run reproduces the published values within Monte Carlo error (published: ≥0.90 for 23.5%, 54.4%, 94.2%, and 64.1%; reliable and precise for 0.4%, 0.4%, 67.4%, and 54.9%). Median precision is the median relative 90% half-width at each profile’s observed volume.

Measure-Level Synthesis

The seven measures did not behave as interchangeable indicators of a single latent construct. Under the primary patient-cluster bootstrap design, the study’s P90 absolute-error tolerance was not reached for visits/patient through the largest tested 200-patient denominator. Standardized cost/patient first met ±$50 at 150 patients, although only 12 profiles qualified; no visit- or CMT-based measure met its tolerance through 300 visits or spinal CMT units. Patient mix and lifecycle effects also differed by outcome: visits/patient and standardized cost/patient were sensitive to new-patient composition and entry timing; units/visit reflected diagnosis mix and entry timing; 98943/visit was strongly associated with M25 prevalence; 97110/visit was associated with overall patient entry timing; and measured covariates explained very little within-profile variation in 98941 share. Reliability followed a different order: units/visit and the service-mix measures were highly reliable, 98941 share and cost/patient were intermediate, and visits/patient was the least reliable. A single administrative patient minimum therefore cannot be assumed to confer comparable precision or reliability across all measures.

Key Limitations

• Observational three-state network study. Findings describe Arete provider-insurer profiles for commercially insured patients in Colorado, Minnesota, and Texas and may not generalize to other networks, specialties, benefit designs, or markets. Texas contributes 4 providers and 15 latest-window profiles.

• Applicability to proprietary profiling methods. This analysis does not assume that any intermediary network uses the exact analytic methods evaluated here, nor does it reproduce a proprietary profiling algorithm. Its relevance depends on whether the underlying measurement issues are addressed by the profiling process being used. A proprietary methodology that incorporates measure-specific reliability, provider-level precision, appropriate eligibility and attribution rules, case-mix or benefit-design adjustment where warranted, validated and periodically recalibrated benchmarks, and uncertainty-aware classification rules may address some or all of the concerns demonstrated in this study. Conversely, proprietary status or historical use alone does not establish that these measurement properties have been addressed. The appropriate comparison is therefore between the safeguards incorporated in a specific profiling methodology and the measurement issues quantified or described here.

• Overlapping rolling windows. Adjacent 12-month windows share 11 months of data. Provider-insurer clustered standard errors and common-time fixed effects address some dependence, but repeated windows are not independent experiments; p-values should be interpreted cautiously.

• Expected copay is incomplete benefit information. The field does not capture deductible status, coinsurance, quantitative visit limits, remaining benefits, or benefit exhaustion. Extreme expected-copay values may reflect source-system semantics broader than a conventional office copay.

• Absence of patient-level enrollment data. Limiting “profiles” to patients with continuous eligibility during the entire measurement period addresses multiple sources of confounding. Without this control, for measures like visits per patient and cost per patient a provider “profile” consisting of patients with longer periods of insurance coverage may exhibit higher values than a similar provider with patients with shorter periods of insurance coverage. This becomes particularly important in “profiles” with longer “look-back” periods (e.g., 3 years). Published cost-profiling research has applied this control; for example, one multi-plan study limited its analysis to patients continuously enrolled for two years [4].

• New-patient and continuity measures are proxies. CPT 99201–99205 is a claims-based new-patient proxy, and the ±2-month continuity analysis is an observation window rather than a complete episode definition.

• Claims diagnoses and HCC RAF are incomplete measures of clinical severity. Diagnosis coding can reflect documentation and billing practices, and HCC RAF is not a musculoskeletal-specific severity measure. Prior research on physician profiles found that age and sex adjustment alone was little better than no case-mix adjustment, and that even detailed adjustment left most variation between patients unexplained [1].

• Top-5% patient removal is counterfactual. It measures concentration and sensitivity; it is not a recommendation to exclude high-utilizing patients from profiling. Prior research found that profiles this sensitive to a few patients can reward avoiding complex patients, and that case-mix adjustment did not remove that incentive [1].

• Administrative eligibility and statistical uncertainty are distinct. Arete reports that many intermediary programs it encounters use minimum patient counts of approximately 10–25 to keep profiling operationally applicable; this operational context was not independently verified across all intermediary programs and is not presented as a statistical standard. The denominator required to meet a selected error tolerance differs by measure, so a single patient-count threshold should not be assumed to confer equivalent precision across all measures in a profile. Bootstrap intervals narrow as denominators increase, but the population of qualifying provider-insurer profiles changes simultaneously, and high-volume tiers contain progressively fewer profiles.

• CMS-standardized cost is an analytic utilization weight, not Medicare payment. CPT 98943 was imputed at 90% of 98940, 5.5% of latest-window units lacked a standardized price, and no MPPR, bundling, geographic, or payer coverage rules were applied.

• Self-pay overlap is incomplete. Only same-provider self-pay activity observable in Athena is captured; care purchased elsewhere is not observed.

• Multivariable models remain associative. Measured covariates can be correlated, omitted variables remain, and selected interaction results are exploratory. No causal effect of any patient or benefit factor is inferred.

• Reliability and misclassification estimates depend on modeling choices. Signal variance was estimated by REML within insurer and state; a simpler, unweighted method-of-moments estimate gives larger signal variance and therefore higher reliability (for example, 31.5% rather than 23.5% of visits/patient profiles and 64.1% rather than 54.4% of cost/patient profiles at or above 0.90 at observed volume). Visits/patient remained the measure least often reaching 0.90 under either estimator. Visits/patient and cost/patient were modeled on their original scales; on a log scale (delta-method sampling variance), the share of profiles at or above 0.90 was similar for visits/patient (22.9% versus 23.5%) and higher for cost/patient (62.2% versus 54.4%), and the share at or above 0.70 was higher for both (82.3% versus 69.3%; 98.5% versus 81.2%), so the size of the below-0.70 regions in Figure 4 depends on this modeling choice. 98941 share results depend on scale: on the percentage scale used here, 64.1% of profiles reached 0.90 at observed volume, versus 23.2% on the logit scale. The difference is driven largely by the 103 profiles (22%) billing all or none of their spinal CMT as 98941, whose sampling variance must be assumed rather than estimated. The patient split-half estimate (median 0.91) is consistent with the percentage-scale result. 97140/visit reliability is dominated by profiles with no 97140 use and is not interpreted. Wrong-tier probabilities are model-based, depend on the hypothetical Author-selected thresholds, and are not a replication of any intermediary’s classification. The reliability analysis was produced by a separate AI-assisted workflow; its summary tables and signal-variance estimates were independently reproduced from the profile-level outputs, but its patient-cluster bootstrap could not be re-executed without the claim-line files (documented in the internal AI Assistance and Reproducibility Record).

• Figure 4 regions depend on study choices and do not evaluate benchmarks. Region shares depend on the ±10% precision cutoff (±20% is reported as a sensitivity analysis), on the reliability model, and, for 98941 share, on the ceiling created when most spinal CMT is billed as 98941. Precision describes how tightly a profile’s own value is estimated, not whether it is accurate. No intermediary benchmark was evaluated, so the analysis cannot say whether any tier assignment is correct; that question is reserved for a subsequent paper.

• Network size and homogeneity. Arete’s network is relatively small and intentionally homogeneous. In a sample of profiles from larger intermediary networks, the reported spread in visits per patient was similar to that observed here, but those networks may have greater true between-provider variation, which would raise reliability (Table 11). Precision estimates depend only on each profile’s own patients and would not be expected to change. The reviewed profiles were a convenience sample, and their reported statistics could not be verified.

• Commercial and Medicare Advantage product mix. The study population includes Medicare Advantage plans administered by commercial and Blue insurers (8.3% of latest-window patient-profile observations). Medicare covers a narrower set of chiropractic services, so units per visit, 97110, 97140, 98943, and cost per patient differ by design between MA and other patients. Excluding MA did not change the study’s conclusions (Table 13), but values for individual profiles with many MA patients changed materially, and intermediary networks that profile commercial and MA patients separately, or together, may obtain different results for the same provider. Product type was taken from Athena and not verified against insurer enrollment.

• Medicaid managed care patients. The same insurer categories include a small number of Medicaid managed care patients (1.6% of latest-window patient-profile observations, in 43 Minnesota profiles). They were retained because intermediary networks do not disclose whether such patients are included in their profiles. Excluding them changed the share of profiles at or above 0.90 reliability by 0.9 percentage points or less, which is not material to the findings. Product type was taken from Athena and not verified against insurer enrollment.

• Insurance product type. Product type was recorded for only 39% of commercial patient-profile observations (61% not specified), only 391 observations were HMO, and product type was taken from Athena and not verified against insurer enrollment. Product-mix contrasts compare patients of the same chiropractor and insurer in one window; they are associations, not causal effects, and may partly reflect the employers and benefit designs behind each product rather than the product type itself.

• Same-period patient-mix adjustment. Figure 3a and Tables 6B–6C adjust each profile to the network’s average patient mix using effects estimated within profiles, which assumes those effects are similar across chiropractors. Entry timing, diagnosis mix, and new-patient proportion may partly reflect practice patterns and coding, so adjusting for them may remove some real practice differences; results are therefore also reported for patient and benefit factors alone. Patients with no recorded copay (25.7%) were not adjusted for copay. Factors overlap, values come from a single model run without sampling uncertainty, and the variance decomposition is approximate. Unmeasured patient factors, such as severity and treatment goals, could not be included.

• NCQA standards are cited from a public-comment draft. The NCQA PHQ requirements described in Background are from the 2026 standards proposed for public comment [10, 11]. The final standards may differ, and this report does not assess any organization’s compliance with them.

• Provider identifier reconciliation is imperfect. Two NPIs mapped to more than one source provider identifier during the observation period; the primary longitudinal entity used the source provider identifier to preserve source attribution.

• Bootstrap uncertainty and precision benchmarks depend on study choices. Results depend on the candidate denominator grid, bootstrap design, observed empirical distribution, analytic reference values, and the study-defined P90 absolute-error tolerances. Reported intervals are pooled across qualifying profiles; a specific profile’s interval depends on its own patient-level distribution. Bootstrap intervals characterize uncertainty around the observed profile under resampling from its empirical distribution; they do not establish universal minimum sample sizes or population-level confidence intervals. 97140/visit is sparse and zero-inflated, so its absolute-error results should be interpreted cautiously. Before v1.4, per-visit and CMT-based measures were resampled by visit or unit rather than by patient, which understated their uncertainty; the corrected patient-cluster results are reported here.

Disclosures and Publication Status

This work is classified as an analysis of de-identified Arete operational data under the Arete Research Library Governance Framework (§2); it is not PBRN-sponsored human-subjects research. This analysis was produced by Arete Healthcare without external funding; Arete supports independent outpatient chiropractic and physical therapy practices and has a commercial and professional interest in payer contracting, provider profiling, and network tiering methods. AI tools assisted with data review, calculation, and drafting. Factual assertions and calculations were subject to human review before publication, and AI reconciliation did not substitute for DAWG review and approval.

Data Source: Patient-de-identified Arete Healthcare operational EHR/RCM claim-line extracts for Colorado, Minnesota, and Texas, with separately extracted expected-copay attributes.

Economic Standardization: 2025 CMS Physician Fee Schedule average locality amounts supplied for this analysis; 98943 was assigned an analytic value equal to 90% of 98940.

AI Use Disclosure: AI-assisted data analysis, statistical modeling, table generation, and drafting were used to prepare this manuscript, and a separate AI system independently recomputed reported statistics from the reproducibility workbook. Arete’s Take was written by a human author, and all figures were generated by humans. Details are maintained in the internal AI Assistance and Reproducibility Record under the Arete Research Library Governance Framework. Human review was completed before release.

Scope: This manuscript evaluates provider-profile measurement behavior for commercially insured patients. Medicare, Medicaid, PI/Auto, workers’ compensation, and self-pay populations are reserved for subsequent analyses, apart from the Medicare Advantage and Medicaid managed care patients included in the insurer categories studied (see Methods).

Publication Status: Final. Approved for release on October 5, 2026 by Data Analysis Work Group (DAWG) reviewers Alli Totzke and Kyle Prusso, following DAWG review of Version 1.39 and of the Figure 5 and Table 9A content added in Version 1.40. Arete’s Take reflects Arete’s perspective and was approved by Arete Executive Reviewers, independent of DAWG review of Findings. Report and dataset identifiers are registered in the Arete Suite Index.

Recommended Citation: Elton, D. What Shapes a Provider Profile? Arete Original Research ARETE-OR-2026-012, Version 1.40. Arete Healthcare LLC. Published October 5, 2026. https://arete.healthcare/research/what-shapes-a-provider-profile https://doi.org/10.13140/RG.2.2.17173.69606

References

1. Hofer TP, Hayward RA, Greenfield S, Wagner EH, Kaplan SH, Manning WG. The unreliability of individual physician “report cards” for assessing the costs and quality of care of a chronic disease. JAMA. 1999;281(22):2098-2105. doi:10.1001/jama.281.22.2098

2. Dimick JB, Welch HG, Birkmeyer JD. Surgical mortality as an indicator of hospital quality: the problem with small sample size. JAMA. 2004;292(7):847-851. doi:10.1001/jama.292.7.847

3. Adams JL. The Reliability of Provider Profiling: A Tutorial. Santa Monica, CA: RAND Corporation; 2009. Technical Report TR-653-NCQA. doi:10.7249/TR653

4. Adams JL, Mehrotra A, Thomas JW, McGlynn EA. Physician cost profiling — reliability and risk of misclassification. N Engl J Med. 2010;362(11):1014-1021. doi:10.1056/NEJMsa0906323

5. Nyweide DJ, Weeks WB, Gottlieb DJ, Casalino LP, Fisher ES. Relationship of primary care physicians’ patient caseload with measurement of quality and cost performance. JAMA. 2009;302(22):2444-2450. doi:10.1001/jama.2009.1810

6. Greenfield S, Kaplan SH, Kahn R, Ninomiya J, Griffith JL. Profiling care provided by different groups of physicians: effects of patient case-mix (bias) and physician-level clustering on quality assessment results. Ann Intern Med. 2002;136(2):111-121. doi:10.7326/0003-4819-136-2-200201150-00008

7. Mehrotra A, Adams JL, Thomas JW, McGlynn EA. Cost profiles: should the focus be on individual physicians or physician groups? Health Aff (Millwood). 2010;29(8):1532-1538. doi:10.1377/hlthaff.2009.1091

8. Spiegelhalter DJ. Funnel plots for comparing institutional performance. Stat Med. 2005;24(8):1185-1202. doi:10.1002/sim.1970

9. Centers for Medicare & Medicaid Services. 2026 Summary of Cost Measures. Quality Payment Program; December 2025. Accessed September 25, 2026. https://www.cms.gov/files/document/2026-mips-summary-cost-measures.pdf

10. National Committee for Quality Assurance. 2026 Physician and Hospital Quality Certification: Standards Proposed for Public Comment (PQ 1, Element D). Washington, DC: NCQA; 2025. Draft for public comment; final standards may differ. Accessed September 25, 2026. https://wpcdn.ncqa.org/www-prod/2026-Physician-and-Hospital-Quality-Certification_Standards-Proposed-for-Public-Comment.pdf

11. National Committee for Quality Assurance. Overview of Proposed Updates to 2026 Standards: Physician and Hospital Quality. Washington, DC: NCQA; August 4, 2025. Accessed September 25, 2026. https://wpcdn.ncqa.org/www-prod/Physician-and-Hospital-Quality-Certification-2026_Overview-Memo.pdf

12. Herman PM, Edgington SE, Hurwitz EL, Coulter ID. Predictors of visit frequency for patients using ongoing chiropractic care for chronic low back and chronic neck pain; analysis of observational data. BMC Musculoskelet Disord. 2020;21(1):298. doi:10.1186/s12891-020-03330-1

13. Shekelle PG, Rogers WH, Newhouse JP. The effect of cost sharing on the use of chiropractic services. Med Care. 1996;34(9):863-872.

14. Gozalo PL, Resnik LJ, Silver B. Benchmarking outpatient rehabilitation clinics using functional status outcomes. Health Serv Res. 2016;51(2):768-789. doi:10.1111/1475-6773.12344

15. Arete Healthcare. Athena EHR/RCM commercial insurance claim-line extracts (12 claim line-item CSV files) for Colorado, Minnesota, and Texas; service history through September 2026. Internal patient-de-identified analytic source; not publicly distributed.

16. Centers for Medicare & Medicaid Services. 2025 Medicare Physician Fee Schedule. Average locality fee schedule amounts as supplied in the study analytic file.

17. Efron B, Tibshirani RJ. An Introduction to the Bootstrap. New York, NY: Chapman & Hall; 1993.

18. Cameron AC, Miller DL. A practitioner’s guide to cluster-robust inference. J Hum Resour. 2015;50(2):317-372. doi:10.3368/jhr.50.2.317

19. Arete Healthcare. Commercial/Blue Provider Profile Analysis v0.5a (Helper Tabs Repaired). September 2026. Internal reproducibility workbook containing rolling profiles, influential-patient sensitivity, copay reconciliation, bootstrap 90%/95% uncertainty intervals, classification-reversal analyses, denominator sensitivity analyses, multivariable model outputs, and figure/helper tabs aligned to the primary bootstrap analysis.

20. Arete Healthcare. Provider Profile Reliability Estimates v0.2 (Patient-Cluster Corrected). September 2026. Internal reproducibility workbook containing the patient-cluster bootstrap, REML signal-variance models, profile-level reliability, reliability under administrative patient minimums, split-half sensitivity, and tier-misclassification estimates.

21. Arete Healthcare. Independent Re-execution of Profile Construction, Patient-Cluster Bootstrap, and Reliability Estimates from Claim-Line Extracts, v1.0. September 2026. Internal reproducibility file (profile reconciliation, sampling-variance and signal-variance comparisons, and percentage-scale 98941 reliability).

Glossary

Plain-language definitions of terms used in this report.

Appendix A. Supplemental Tables: Full Core Multivariable Model Results and Model Extensions

The tables below report every pre-specified core covariate for each outcome, including associations that were not statistically distinguishable from zero. Effects are scaled to the increments defined in Methods. Standardized beta represents within-provider-insurer relative weight.

A1. Visits per Patient

N=7,070 provider-window observations; 484 provider-insurer entities; partial R²=0.105.

A2. Units per Visit

N=7,070 provider-window observations; 484 provider-insurer entities; partial R²=0.101.

A3. 98941 % of Spinal CMT

N=7,067 provider-window observations; 483 provider-insurer entities; partial R²=0.008.

A4. CMS-Standardized $ per Patient

N=7,070 provider-window observations; 484 provider-insurer entities; partial R²=0.079.

A5. 98943 Units per Visit

N=7,070 provider-window observations; 484 provider-insurer entities; partial R²=0.137.

A6. 97110 Units per Visit

N=7,070 provider-window observations; 484 provider-insurer entities; partial R²=0.052.

A7. 97140 Units per Visit

N=7,070 provider-window observations; 484 provider-insurer entities; partial R²=0.035.

A8. Bounded-Window Continuity Extension

The ±2-month bounded model was limited to windows with complete observation in the two full months before and after the annual profile (5,832 observations; 466 entities). It added prior- and post-window continuity to the core covariate set. Selected additional associations were: prior-period care +10 percentage points → −0.078 visits/patient (p<0.001); prior-period care +10 percentage points → −$8.69 standardized cost/patient (p<0.001); and post-period care +10 percentage points → +0.37 percentage points of 98941 share (p=0.011). Core associations with new-patient share, entry timing, M25, and M53 were directionally similar.

A9. New-Patient Timing Specification

Because overall patient entry timing and new-patient entry timing were strongly correlated, they were not included together in the primary core model. In a separate specification (6,683 observations; 466 entities), 30 days later mean new-patient entry was associated with −0.037 visits/patient (p=0.028), +0.015 units/visit (p<0.001), and +0.0064 97110 units/visit (p<0.001). New-patient share and M25 prevalence remained independently associated with visits/patient. Interaction results from this specification are reported in Table 7.

A10. Collinearity and Interaction Strategy

Within-entity correlation was high between overall and new-patient entry timing (approximately 0.85) and between age and HCC RAF (approximately 0.79). Variance-inflation diagnostics were approximately 5 or below in the candidate multivariable set (highest for mean HCC RAF, approximately 5.1), and these relationships supported using separate timing specifications and limiting interaction tests to pre-specified hypotheses. Exhaustive pairwise interactions were not used because they would materially increase model dimensionality and multiple-testing risk without a clear clinical interpretation.

A11. Bootstrap Denominator Uncertainty and High-Low Visualization Construction

The primary denominator-uncertainty analysis uses nonparametric bootstrap sampling with replacement, with patients as the resampling unit for every measure. Patient-denominated outcomes resample the tier’s number of patients; per-visit outcomes and 98941 share resample whole patients until the tier’s visit or spinal CMT unit count is reached. One thousand replicates were generated per qualifying provider-insurer profile and denominator tier. The v0.5a workbook values for the patient-denominated outcomes (fixed seed 20260924) are retained; per-visit and 98941 values are from the patient-cluster rerun in the reliability workbook [20]. For per-visit and CMT-based measures, the median ratio of patient-cluster to visit-independent sampling variance (design effect) was 3.29 for units/visit, 3.65 for 98943/visit, 2.93 for 97110/visit, 3.15 for 97140/visit, and 3.50 for 98941 share. Signed bootstrap error was defined as the replicate estimate minus the observed full-profile estimate. Empirical 90% and 95% intervals were defined by the 5th–95th and 2.5th–97.5th percentiles of signed error. Classification reversal was calculated for measures with analytic references. Earlier without-replacement subsampling remains a sensitivity analysis. Figure-ready workbook tabs use a zero-centered display convention for structural sensitivity measures: top-5% removal is one-sided downward, P90 rolling max-min range is divided equally around zero, and P90 maximum adjacent-window change is shown symmetrically. These centering conventions are visualization devices and do not imply symmetric empirical distributions.

A12. Reproducibility Files

The internal reproducibility workbooks [19, 20] retain the rolling provider-insurer profiles, patient-cluster bootstrap results, profile-level reliability and misclassification estimates, precision curves, top-5% influence calculations, random-removal comparisons, patient-lifecycle summaries, copay QA, and multivariable model outputs needed to challenge or reproduce the reported results. Patient identifiers and proprietary source extracts are not included in the manuscript.

Appendix B. Questions to Ask About a Provider Profile

A measurement-transparency checklist for providers

1. What exactly was measured?

2. Is my result statistically reliable and precise?

3. How was the benchmark developed?

4. What patient, benefit, and measurement factors were considered?

5. Is the result stable?

6. How did the measurements become a profile-based decision or action?

7. What transparency, verification, correction, and appeal safeguards exist?

Suggested documentation request

Interpretive note: These questions are intended to determine whether the measurement and decision pathway can be understood, reproduced, and meaningfully evaluated; they do not presume that a provider profile or resulting decision is incorrect.