Remits vs. MRF Stoploss Scoring

Overview

This page documents our methodology for connecting Komodo remits payment patterns to MRF-extracted stoploss provisions. The core idea: if remits payment patterns consistently match MRF stoploss terms for known payer-provider combinations, we can apply those same rules to combinations where we don't have MRF data.

We use a 6-signal scoring system that evaluates each provider-payer combo on a 0-100 scale, producing a confidence tier from "Confirmed" (Tier 1) to "Unlikely" (Tier 5).

---

Data Description

Source: remits_mrf_first_dollar.csv — remit lines for inpatient claims >$200K total billed, matched to MRF stoploss provisions.

Metric	Value
Total rows	24,725
Unique provider-payer combos	224
Unique provisions (combo × MRF rate)	237
Provisions with MRF data	131
Provisions scored (MRF + 5+ lines)	104
Multi-provision combos	13
Unique providers	186
Unique payers	21
Stoploss type	First dollar (all rows)

What each row represents: A group of remit lines for a specific provider × payer × network × revenue code combination, all paid at the same allowed/billed percentage. The lines_paid_at_this_percentage field counts how many individual lines share that rate.

Filters applied:

• Total billed >$200K (high-acuity inpatient)
• First dollar stoploss type only
• Revenue codes grouped at the 4-digit level

Known data limitations

• $200K threshold filter — This is an arbitrary cutoff. Claims below $200K but above the MRF dollar threshold are excluded, reducing signal.
• Multi-plan mixing — Remits are NOT unique on plan. A single payer-provider combo contains lines from multiple plans (some stoploss, some DRG/case rate). The stoploss signal may be a minority cluster.
• Network duplication — The same remit lines appear under multiple network names (e.g., Cigna PPO, Cigna OAP, unmapped). We deduplicate to provider × payer before scoring.
• MRF vintage — MRF rates may be from a different contract year than the remits, causing slight rate mismatches.

---

Methodology: 6-Signal Scoring Framework

Design Principle: Cluster Analysis, Not Modal Rate

A naive approach would check if the most common payment rate matches the MRF stoploss percentage. This fails because of multi-plan mixing: a combo where 25% of lines pay at the MRF rate and 75% pay at DRG rates still contains stoploss evidence — it just means one plan has stoploss and others don't.

Our approach looks for any significant cluster near the MRF rate, regardless of whether it's the dominant rate. This is critical for detecting hidden stoploss in mixed-plan data.

Per-Provision Scoring

A single provider-payer combination can have multiple stoploss provisions at different rates — for example, different networks or tiered thresholds. Previously, the system took the first MRF rate encountered and discarded the rest. Now, each distinct percentage_reimbursement in the MRF data becomes its own scored row.

Example: Dartmouth-Hitchcock / Anthem has provisions at 45% (unmapped network) and 69.28% (NH Open Access network). Under the old system, only the 45% provision was scored, producing a Tier 4 result. With per-provision scoring, the 69.28% provision is scored independently and achieves Tier 2 — correctly identifying a real stoploss arrangement that was previously missed.

Signal 1: MRF Cluster Strength — weight: 30%

"What fraction of remit lines cluster near the MRF stoploss %?"

Sums all lines paid within 5 percentage points of the MRF rate and calculates the cluster fraction.

Scoring thresholds

Cluster %	Score
75%+	100
50-74%	80
25-49%	60
10-24%	40
5-9%	20
<5%	0

Rationale: This is the strongest single indicator. Even a 10-25% cluster is meaningful — it suggests at least one plan under this payer has stoploss at the MRF rate.

Example: Virginia Mason / UHC — 52 of 62 lines fall within 5pp of the MRF rate (60%), giving a cluster fraction of 83.9% → score 100.

Signal 2: Best Rate Proximity — weight: 25%

"How close is the nearest high-volume rate to the MRF %?"

Finds the single payment rate with the most lines that falls within 10pp of the MRF rate, then measures the distance.

Scoring thresholds

Distance from MRF	Score
0pp	100
1-2pp	85
3-5pp	65
6-10pp	35
No rate within 10pp	0

Rationale: An exact or near-exact rate match is strong evidence. Small deviations (1-5pp) are expected from threshold dilution — charges below the stoploss threshold are included in totals, pulling the observed rate below the contractual rate.

Example: WakeMed / BCBS NC — best matching rate is 19% (4 lines) vs MRF rate 22.7%, a gap of 3.7pp → score 65.

Signal 3: Rate Dispersion Context — weight: 15%

"Is the rate distribution concentrated (flat %) or dispersed (mixed pricing)?"

Counts distinct rounded payment rates for the combo.

Scoring thresholds

Distinct rates	Score
1-3	100
4-6	70
7-12	40
13-20	20
>20	10

Rationale: Fewer distinct rates suggests a flat percentage contract. However, high dispersion is scored at 10 rather than 0 — a combo with 30+ rates but a strong MRF cluster is still valid evidence of multi-plan mixing where one plan has stoploss.

Example: WakeMed / BCBS NC — 26 distinct rates across all lines → score 10. The high dispersion reflects heavy multi-plan mixing.

Signal 4: Clinical Code Quality — weight: 10%

"When we exclude drugs/implants/supplies, does the stoploss signal strengthen?"

Recalculates cluster strength using only non-excluded revenue codes. Drugs (025x, 0636-0638) and implants/supplies (027x, 0621-0624) are known stoploss carveouts.

Scoring thresholds

Comparison	Score
Clinical cluster stronger (>5pp improvement)	100
About the same (within 5pp)	50
Clinical cluster weaker	25
Insufficient clinical data	50

Rationale: If excluding carveout codes improves the match, it confirms both that stoploss exists AND that the carveout structure matches expectations.

Example: WakeMed / BCBS NC — all-code cluster 11.9% vs clinical-only 21.5%, an improvement of +9.6pp → score 100. Removing carveout codes nearly doubled the cluster strength.

Signal 5: Line Volume — weight: 10%

"More claims = more statistical confidence in the pattern."

Scoring thresholds

Lines	Score
100+	100
50-99	80
20-49	60
10-19	40
5-9	20
<5	0

Example: Holston Valley / Anthem — 40 remit lines → score 60 (20-49 bucket).

Signal 6: Directional Consistency — weight: 10%

"Is the best matching rate at or below the MRF %?"

For first dollar stoploss, the observed remit rate should be at or below the MRF rate due to threshold dilution (sub-threshold charges dilute the average).

Scoring thresholds

Direction	Score
Remit ≤ MRF (expected)	100
Remit 1-5pp above MRF	60
Remit 6-10pp above MRF	30
Remit >10pp above MRF	0

Example: Virginia Mason / UHC — best matching rate 60% equals MRF rate 60%, remit ≤ MRF → score 100.

---

Confidence Tiers

The weighted total of all 6 signals produces a 0-100 score, mapped to tiers:

Score	Tier	Label	Meaning
80-100	1	Confirmed	Remits strongly confirm MRF stoploss
60-79	2	High confidence	Consistent pattern, minor noise or multi-plan dilution
40-59	3	Moderate	Signal present but diluted by multi-plan mixing or limited data
20-39	4	Candidate	Weak signal, needs more data or manual review
0-19	5	Unlikely	No stoploss pattern detected in remits

---

Findings

Tier Distribution

Of 104 scored provisions (MRF data + 5 or more remit lines, from 92 unique provider-payer combos):

Tier	Label	Count	%
1	Confirmed	9	9%
2	High confidence	25	24%
3	Moderate	37	36%
4	Candidate	22	21%
5	Unlikely	11	11%

33% of provisions (Tiers 1-2) show strong remits-MRF alignment. Another 36% (Tier 3) show moderate signal, often diluted by multi-plan mixing. Multi-provision scoring added 12 rows from combos with multiple stoploss rates — several of these new provisions scored higher than the original single-provision result.

Score Statistics

• Mean: 50.5
• Median: 53.8
• Min: 12.0
• Max: 91.0

All Scored Provisions

The full dataset of 104 scored provisions across 92 provider-payer combinations. Combos with multiple provisions (n_provisions > 1) have separate rows for each MRF rate. Use column headers to sort and filter.

---

Example Walkthroughs

Each walkthrough shows the raw rate distribution from remits, how each signal is calculated step-by-step, and the final score. All data is from the actual scored dataset.

Tier 1: Virginia Mason Medical Center / UnitedHealthcare

MRF stoploss: 60% of charges, $250,000 threshold Networks in data: Choice Plus, HMO, Select EPO, unmapped (deduplicated to single combo)

Sample source data (5 of 148 raw rows)

Each row = one revenue code group within one network. The same rev code can appear under multiple networks (deduplication collapses these).

Network	Rev Code	Lines at Rate	Total Lines for RC	Rate	MRF %
Choice Plus	0637	1	1	60%	60%
HMO	0272	1	1	60%	60%
Select EPO	0120	1	2	45%	60%
unmapped	(blank)	2	35	26%	60%
Choice Plus	0200	1	2	62%	60%

Note: Carveout codes (0637 pharmacy, 0272 supplies) pay at 60% here — this payer apparently does NOT carve out drugs/supplies from the stoploss rate.

Raw rate distribution (62 lines, 12 distinct rates)

Rate	Lines	% of Total
60%	49	79.0%	████████████████████████████████████████
59%	2	3.2%	██
26%	2	3.2%	██
62%	1	1.6%	█
52%	1	1.6%	█
45%	1	1.6%	█
32%	1	1.6%	█
30%	1	1.6%	█
25%	1	1.6%	█
22%	1	1.6%	█
18%	1	1.6%	█
17%	1	1.6%	█

The 60% rate dominates — 49 of 62 lines pay at exactly the MRF-stated rate. The remaining 13 lines are scattered across 11 other rates, likely representing non-stoploss plans or different service categories.

Revenue code breakdown

Rev Code	Lines	Top Rates	Notes
(blank)	35	60% (25), 26% (2), 59% (1)	Unclassified — most still at 60%
0200	2	60% (1), 62% (1)	General room & board
0120	2	59% (1), 45% (1)	Room & board - semi-private
0637	1	60% (1)	Pharmacy — carveout code, still pays at stoploss rate
0272	1	60% (1)	Supplies — carveout code, still pays at stoploss rate
0361, 0483, 0390, 0921, 0324	5	All at 60%	Various clinical codes

Step-by-step signal scoring

Signal 1 — MRF Cluster Strength: Lines within 5pp of 60% = 49 (at 60%) + 2 (at 59%) + 1 (at 62%) = 52 lines. Cluster fraction = 52/62 = 83.9%. That's ≥75%, so score = 100.

Signal 2 — Best Rate Proximity: Highest-volume rate within 10pp of 60% is 60% itself (49 lines). Distance = 0pp, so score = 100.

Signal 3 — Rate Dispersion: 12 distinct rates. Falls in the 7-12 bucket, so score = 40. (The non-stoploss lines are spread thin across many rates — expected in multi-plan data.)

Signal 4 — Clinical Code Quality: All-code cluster = 83.9%. Clinical-only cluster (excluding 0637, 0272) = 48/58 = 82.8%. Difference is <5pp (about the same), so score = 50.

Signal 5 — Line Volume: 62 lines, falls in the 50-99 bucket, so score = 80.

Signal 6 — Directional Consistency: Best matching rate (60%) = MRF rate (60%). Remit ≤ MRF, so score = 100.

Final Score

Signal	Weight	Score	Weighted
S1: Cluster Strength	30%	100	30.0
S2: Rate Proximity	25%	100	25.0
S3: Rate Dispersion	15%	40	6.0
S4: Clinical Code Quality	10%	50	5.0
S5: Line Volume	10%	80	8.0
S6: Directional Consistency	10%	100	10.0
Total			84.0 — Tier 1 (Confirmed)

Takeaway: Clean, textbook stoploss confirmation. The MRF says 60% and 79% of lines pay at exactly 60%. The only reason this isn't 90+ is the rate dispersion penalty from the 11 outlier rates (likely non-stoploss plans).

---

Tier 3: WakeMed Cary Hospital / BCBS of North Carolina

MRF stoploss: 22.7% of charges, $111,417 threshold Networks in data: HMO, Preferred Provider Network, unmapped (deduplicated)

Sample source data (5 of 120 raw rows)

Each row = one revenue code group within one network.

Network	Rev Code	Lines at Rate	Total Lines for RC	Rate	MRF %
HMO	(blank)	3	28	22%	22.7%
Preferred Provider Network	(blank)	1	28	23%	22.7%
unmapped	0636	2	32	11%	22.7%
HMO	0278	3	6	46%	22.7%
Preferred Provider Network	(blank)	2	28	49%	22.7%

Note: The 22% and 23% rows (near MRF) come from different networks, showing stoploss signal is consistent across plans. The 0636 pharmacy code at 11% is a carveout priced far below the MRF rate.

Raw rate distribution (118 lines, 26 distinct rates)

Rate	Lines	% of Total
46%	42	35.6%	██████████████████
50%	15	12.7%	██████
42%	10	8.5%	████
45%	6	5.1%	███
47%	6	5.1%	███
19%	4	3.4%	██
30%	4	3.4%	██
44%	4	3.4%	██
22%	3	2.5%	█
24%	3	2.5%	█
13%	2	1.7%	█
34%	2	1.7%	█
11%	2	1.7%	█
49%	2	1.7%	█
55%	2	1.7%	█
23%	1	0.8%	▏
21%	1	0.8%	▏
20%	1	0.8%	▏
27%	1	0.8%	▏
28%	1	0.8%	▏
15%	1	0.8%	▏
33%	1	0.8%	▏
12%	1	0.8%	▏
40%	1	0.8%	▏
48%	1	0.8%	▏
67%	1	0.8%	▏

At first glance, this looks nothing like 22.7% stoploss. The dominant cluster is 42-50% (83 lines, 70% of total) — probably a case-rate or DRG-based plan. But look at the 18-24% range: 19% (4 lines), 22% (3), 24% (3), 23% (1), 21% (1), 20% (1) = 13 lines near the MRF rate. That's the stoploss signal, hiding in the long tail.

Revenue code breakdown

Rev Code	Lines	Top Rates	Notes
0636 (carveout)	32	46% (24), 45% (6), 11% (2)	Pharmacy — pays at 46%, NOT at stoploss rate
(blank)	28	30% (4), 22% (3), 24% (3)	Unclassified — contains most of the stoploss-range rates
0278 (carveout)	6	46% (3), 44% (2), 12% (1)	Supplies — mostly at case-rate
0272 (carveout)	6	46% (6)	Supplies — all at 46%
0481	6	47% (6)	Cardiology — all at 47%
0305	6	50% (6)	Lab — all at 50%
0301	6	50% (6)	Lab — all at 50%
0250 (carveout)	4	42% (2), 46% (2)	Pharmacy
0324	3	19% (3)	Radiology — all 3 lines at 19% (near MRF)
0370	3	42% (3)	Anesthesia

This is where the story becomes clear. The carveout codes (0636, 0278, 0272, 0250 = 48 lines) overwhelmingly pay at 42-46% — these are priced under a different methodology (likely case-rate). Removing them concentrates the stoploss signal.

Step-by-step signal scoring

Signal 1 — MRF Cluster Strength: Lines within 5pp of 22.7% = 19% (4) + 22% (3) + 24% (3) + 23% (1) + 21% (1) + 20% (1) + 27% (1) = 14 lines. Cluster fraction = 14/118 = 11.9%. That's in the 10-24% bucket, so score = 40.

Signal 2 — Best Rate Proximity: Highest-volume rate within 10pp of 22.7% is 19% (4 lines). Distance = 3.7pp, falls in 3-5pp bucket, so score = 65.

Signal 3 — Rate Dispersion: 26 distinct rates, falls in the >20 bucket, so score = 10.

Signal 4 — Clinical Code Quality: All-code cluster = 11.9%. Clinical-only (excluding 48 carveout lines): 14/65 = 21.5%. Improvement of +9.6pp (>5pp), so score = 100. Removing carveouts nearly doubled the cluster strength — the carveout codes were almost entirely paying at non-stoploss rates, diluting the signal.

Signal 5 — Line Volume: 118 lines, falls in the 100+ bucket, so score = 100.

Signal 6 — Directional Consistency: Best matching rate (19%) < MRF rate (22.7%). Remit below MRF, consistent with threshold dilution, so score = 100.

Final Score

Signal	Weight	Score	Weighted
S1: Cluster Strength	30%	40	12.0
S2: Rate Proximity	25%	65	16.3
S3: Rate Dispersion	15%	10	1.5
S4: Clinical Code Quality	10%	100	10.0
S5: Line Volume	10%	100	10.0
S6: Directional Consistency	10%	100	10.0
Total			59.8 — Tier 3 (Moderate)

Why This Matters

This combo demonstrates exactly why cluster analysis is necessary. A modal-rate approach would see "46% is the most common rate" and conclude no stoploss at 22.7% exists. But the data tells a richer story:

• The 42-50% cluster (83 lines) is likely a non-stoploss plan using case-rate or DRG pricing
• The 18-24% cluster (14 lines) is the stoploss signal from a different plan under the same payer
• Carveout codes (pharmacy, supplies) overwhelmingly pay at the non-stoploss rate, which is exactly what you'd expect if they're carved out of the stoploss arrangement

The 4pp gap between the best match (19%) and the MRF rate (22.7%) is consistent with threshold dilution: the $111K threshold means a significant portion of billed charges fall below it, pulling the observed allowed/billed ratio below the contractual 22.7%.

---

Tier 5: Holston Valley Medical Center / Anthem

MRF stoploss: 41.3% of charges, $185,572 threshold Networks in data: unmapped only

Sample source data (5 of 20 raw rows)

Each row = one revenue code group within one network.

Network	Rev Code	Lines at Rate	Total Lines for RC	Rate	MRF %
unmapped	(blank)	2	35	54%	41.3%
unmapped	(blank)	6	35	20%	41.3%
unmapped	0250	1	1	82%	41.3%
unmapped	(blank)	3	35	11%	41.3%
unmapped	0278	1	1	76%	41.3%

Note: All rows are from a single "unmapped" network. Rates range from 11% to 82% with no cluster near the MRF rate of 41.3%. Carveout codes (0250 pharmacy, 0278 supplies) pay at 76-82%, far above both the MRF rate and clinical code rates.

Raw rate distribution (40 lines, 18 distinct rates)

Rate	Lines	% of Total
20%	6	15.0%	████████
15%	6	15.0%	████████
14%	4	10.0%	█████
11%	3	7.5%	████
54%	2	5.0%	███
16%	2	5.0%	███
12%	2	5.0%	███
74%	2	5.0%	███
76%	2	5.0%	███
78%	2	5.0%	███
89%	2	5.0%	███
58%	1	2.5%	█
19%	1	2.5%	█
18%	1	2.5%	█
65%	1	2.5%	█
66%	1	2.5%	█
75%	1	2.5%	█
82%	1	2.5%	█

There's no cluster anywhere near 41%. The distribution has two vague groupings — a low cluster around 11-20% (25 lines) and a high cluster around 54-89% (15 lines) — neither of which overlaps with the MRF rate. This looks like DRG/case-rate pricing with wide variation across service types.

Revenue code breakdown

Rev Code	Lines	Top Rates	Notes
(blank)	35	20% (6), 15% (6), 14% (4)	Most lines at low rates — likely DRG pricing
0206	2	66% (1), 89% (1)	ICU — wide rate variation, not flat %
0278 (carveout)	1	76% (1)	Supplies
0250 (carveout)	1	82% (1)	Pharmacy
0360	1	89% (1)	OR services

No revenue code shows a consistent payment percentage. Even within the same code (0206), rates range from 66% to 89% — the opposite of what you'd expect from a flat stoploss arrangement.

Step-by-step signal scoring

Signal 1 — MRF Cluster Strength: Lines within 5pp of 41.3% = 0 lines. Cluster fraction = 0%. Score = 0.

Signal 2 — Best Rate Proximity: No rate falls within 10pp of 41.3% (nearest is 54% at 12.7pp away). Score = 0.

Signal 3 — Rate Dispersion: 18 distinct rates, falls in 13-20 bucket. Score = 20.

Signal 4 — Clinical Code Quality: Both all-code and clinical-only clusters are 0%. No difference. Score = 50 (neutral).

Signal 5 — Line Volume: 40 lines, falls in 20-49 bucket. Score = 60.

Signal 6 — Directional Consistency: No rate within 10pp of MRF to evaluate. Score = 0.

Final Score

Signal	Weight	Score	Weighted
S1: Cluster Strength	30%	0	0.0
S2: Rate Proximity	25%	0	0.0
S3: Rate Dispersion	15%	20	3.0
S4: Clinical Code Quality	10%	50	5.0
S5: Line Volume	10%	60	6.0
S6: Directional Consistency	10%	0	0.0
Total			14.0 — Tier 5 (Unlikely)

Takeaway: The only points come from neutral/volume signals (S3-S5). Every signal that actually measures stoploss alignment (S1, S2, S6) scores zero. This is likely a case where the MRF stoploss provision exists on paper but either (a) no active plan with claims in our data uses it, or (b) the $185K threshold is too close to our $200K billed filter — leaving almost no "excess" for the stoploss rate to apply to, so DRG/case-rate pricing dominates.

Threshold Proximity Problem

This combo illustrates a data limitation: when MRF threshold ($185K) is near our billed filter ($200K), claims barely exceed the stoploss trigger. If a $220K claim has a $185K threshold, only $35K is subject to the stoploss rate — but the allowed/billed ratio is calculated on the full $220K, washing out the signal. Future work item #1 (filter by MRF threshold instead of $200K) would help here.

---

Limitations and Caveats

Threshold Dilution

For first dollar stoploss, the allowed/billed ratio systematically undershoots the MRF rate because charges below the dollar threshold are included in totals. A 60% stoploss rate with a $250K threshold will appear as ~55-58% when averaged across all charges (including the first $250K reimbursed at different rates). Our 5pp cluster window and directional scoring account for this, but it remains a source of noise.

$200K Filter

Our data only includes claims with total billed >$200K. This is an arbitrary cutoff that:

• Misses stoploss activity on claims between the MRF threshold and $200K
• Over-represents the highest-acuity cases, which may have different coding patterns
• Creates a threshold proximity problem — combos where MRF threshold is near $200K (e.g., $185K) have very little "excess" for stoploss to apply to

MRF Vintage Mismatch

MRF rates are extracted from the most recently published files, while remits span a broader time period. Contract terms change at renewal — a combo scoring poorly may have had stoploss in a prior contract year, or vice versa.

Multi-Plan Noise

The single biggest challenge. Without plan-level identifiers in remits, we cannot separate stoploss plans from non-stoploss plans under the same payer. Our cluster analysis mitigates this but cannot eliminate it — a 10% cluster could be stoploss or could be a coincidental rate.

MRF Data Quality

Some MRF-extracted rates appear anomalous (e.g., 167%, 297%). These likely represent data extraction errors or different reimbursement structures that don't map cleanly to "percentage of charges." Combos with these rates score poorly, as expected.

---

Future Work

Near-Term Improvements (Next Data Pull)

1. Filter by MRF threshold, not $200K — Pull remit lines where total_billed > MRF_dollar_threshold instead of the arbitrary $200K cutoff. This directly addresses threshold dilution.

2. Above/below threshold split — Pull remit lines both above and below the MRF threshold for each combo. If the payment rate changes dramatically at the threshold, that's strong stoploss confirmation regardless of exact rate matching.

3. Threshold isolation — For combos without MRF data, look for a "breakpoint" in the allowed/billed ratio as a function of total billed amount. A sharp change in payment percentage at some dollar value likely indicates a stoploss threshold.

4. Plan-level separation — If possible, get plan-level identifiers in the remits data to separate plans and eliminate the multi-plan mixing problem entirely.

Medium-Term Expansions

5. Year-over-year trending — Compare 2024 remit rates to 2023. Stoploss terms tend to be stable year-over-year; volatile rates suggest non-stoploss pricing.

6. Health system inference — If Provider A / Payer X is confirmed stoploss at 52%, and Provider B (same health system) / Payer X shows a dominant rate at 52%, infer stoploss with bonus confidence.

7. Payer boilerplate detection — If a payer uses the same stoploss rate across multiple unrelated providers, it's likely a standard contract template. Weight these matches higher.

8. Revenue code anchoring — From validated combos, identify which specific rev codes are most predictive of stoploss and weight those codes more heavily in scoring.

Stoploss Type Expansion

9. Second dollar detection — Only excess above threshold is paid at stoploss %. Look for bimodal payment patterns (one cluster at DRG rates, one at stoploss %).

10. Per diem detection — Look for flat dollar amounts per day (requires length-of-stay data alongside remits).

11. % of charges above threshold — Similar to first dollar but applied only to the excess. Harder to detect from remit data alone.