Minimum Group-Size Filters and Multiple-Testing Adjustments Focus Outreach on Projected Coverage Losses
We filter small-group swings across modeled county subgroup benchmarks so outreach funds target the largest plan drops.
County Medi-Cal teams track monthly enrollment across dozens of language, age, and demographic cells to spot renewal drops. In small language groups with fewer than 60 members, a few moving families can create a large percentage swing. Combining a minimum group-size floor with net headcount losses directs outreach staff to the largest coverage drops.
In a multilingual county like Santa Clara (where the 62-cell DHCS grid spans both 47-member and 138,215-member language groups), pairing a modeled 62-group screen with total headcount focuses outreach on Spanish-speaking members (-24,880 net drop, or 41.3% of the county drop) and Hispanic adults ages 19 to 44 (-17,380).[1][2] Sorting those 62 groups by raw percent drop puts tiny groups first, like Samoan (-34.04%, 47 to 31 members) and French (-25.00%, 56 to 42), where a few moving homes shift the rate.[1][3] Setting a size floor (100+ members) with modeled drops past the county rate guides staff.[2][3]
Table 1. Checking a 62-Cell County Medi-Cal Grid: Small-Group Swings vs. Large Drops
We compare observed June 2025 to June 2026 Medi-Cal net snapshot changes with size floors and benchmark gaps.
| 62-Group Check Measure |
Rank #1 Raw %
Samoan Language Cell
|
Ranks #2 & #5 Raw %
French & Thai Cells
|
41.3% of County Drop
Spanish Language Cell
|
High-Loss Adult Group
Hispanic & Male 19-44
|
|---|---|---|---|---|
| Count & Net % Drop | -34.04%47 to 31; -16 people | -25.00% (-14) & -20.75% (-11) | -18.00%138,215 to 113,335 | -20.11% (Hisp) & -20.19% (Male) |
| Raw % Rank vs. Volume Rank | Rank #1 by %Rank #49 by vol | Rank #2 & #5 by %#50 & #52 vol | Rank #11 by %Rank #1 by vol | Rank #6 & #7 by %#2 & #3 vol |
| June 2025 Enrolled Group Size* | 47 memberstiny cell; 3-4 homes | 56 (French) & 53 (Thai) | 138,21529.3% of county | 86,421 (Hisp) & 76,734 (Male) |
| Minimum Group-Size Check* | Under 100-Member CutoffPartner List | Under 100-Member CutoffSmall Cell | Clears 100+ Member & Volume Check | Clears 100+ Member & Volume Check |
| Net Drop Past -12.77% Rate* | -10 past rate-16 vs -6.0 exp | -6.8 (French) & -4.2 (Thai) | -7,230 past rate-24,880 net | -6,344 (Hisp) & -5,697 (Male) |
| Share of County Net Drop (-60,292) | 0.027% of drop16 / 60,292 | 0.041% combined25 / 60,292 | 41.27% of drop24,880 | 28.83% (Hisp)25.70% (Male) |
Table 1 Notes & * Modeled Estimates
Expand notes
Table 1 Notes (Observed DHCS Snapshot Counts & Mode A Screening Rules): Observed June 2025 (472,014) and June 2026 (411,722) counts come from scc_dhcs_medical_enrollment.csv (county net change -60,292 or -12.773%, rounded to -12.77%). Rows 3-5 show Mode A operational rules: a minimum group-size floor (n ≥ 100) and the net drop past the -12.77% county benchmark (Observed_Net_Change - Base_2025 × (-0.1277)). Because monthly DHCS snapshots record net stock changes (entries minus exits), overlap across age/sex/language tables, and cluster within families, exact binomial exit p-values are not valid on net snapshot counts; see Table 2 Notes for verified replicate-weight, Holm/FDR, and permutation tests.
Column Scope. Samoan Language Cell. 47 to 31 Members (-16).. French & Thai Cells. 56 to 42 (-14) | 53 to 42 (-11).. Spanish Language Cell. 138,215 to 113,335 (-24,880).. Hispanic & Male 19-44. 86,421 to 69,041 (-17,380)..
Two-Ruler Benchmark Summary. Observed 62-Cell Santa Clara County DHCS Medi-Cal Snapshots (472,014 to 411,722, -12.77% Net Change) & Mode A Screening Rules. Ruler 1: Raw Percent Rank Across 62 Cells (No Size Floor): Rank #1 Samoan -34.04% (-16) | Rank #2 French -25.00% (-14). Mode A: scanning 62 independent tests at raw α = 0.05 gives 3.10 expected false leads (0.05 / 62 = 0.00081 Bonferroni cutoff). Observed: the top percent drops are tiny cells like Samoan (47 to 31, -16 people), French (56 to 42, -14 people), and Thai (53 to 42, -11 people). Ruler 2: Size Floor (n ≥ 100) Paired With Net Drop Past County Rate: Spanish -24,880 (-18.0%) | Young Adults -17,380 (-20.1%). Requiring cells to clear a size floor (n ≥ 100) and ranking by net drop past the -12.77% county rate finds the main targets: Spanish (-24,880 net drop; -7,230 past county rate) and Hispanic adults 19-44 (-17,380 net drop; -6,344 past county rate).
Table 2. Balancing Multi-Group Screens and Headcount Scale Across County Subgroups
We set dashboard rules that filter out small-group swings and pair multi-group gates with net headcount drops.
| Choice Factor | Samoan Language Cell | French & Thai Cells | Spanish Language Cell | Hispanic & Male 19-44 |
|---|---|---|---|---|
| What Drives the Rate in This Cell | With n = 47 members, three families moving out of county shifts the cell rate by -34.04% (-16 people). | Across 62 cells, seeing -25.00% in French or -20.75% in Thai fits small-group swings. | Across 138,215 Spanish members, an -18.00% drop gives a -7,230 gap. | Young adults 19-44 see -20.11% and -20.19% drops from moves and forms. |
| Limits of Raw Percent Ranks | Sorting a dashboard by raw percent drop puts a 16-person shift ahead of 24,880 Spanish-speaking drops. | Flagging every group with raw single tests across 62 cells gives 3.10 expected false leads by chance. | Spanish ranks #11 by raw percent (-18.00%) behind ten tiny cells, yet holds 41.3% of all county drops. | Treating 62 overlapping age, sex, and language slices as separate wins scatters outreach staff. |
| Budget Risk if Unchecked | Building a media drive around Rank #1 percent change (-34.04%) reaches just 0.03% of county drops. | Shifting translation funds toward small-group swings pulls staff hours from groups with thousands of drops. | Under-funding Spanish outreach because -18.00% sits near the -12.77% county rate leaves a -7,230 gap. | Skipping focused outreach for adults 19-44 leaves out over 17,000 members who move apartments often. |
| Dashboard & Outreach Rule | Set a Size Floor (n ≥ 100). Send small cells (n < 100) to local partner lists instead of top alerts. | Adjust for Multiple Tests. Use multi-group checks on valid study tests before shifting funds. | Rank by Net Gap (-7,230). Sort large groups by net drop past the county rate. | Send Texts & Address Checks. Reach adults 19-44 (-20.1% drop) with SMS texts and postal address matches. |
Table 2 Notes: Snapshot Accounting, Multi-Test Benchmarks & Provenance
Expand notes
• Observed 62-Cell DHCS Snapshot Math & Why Binomial Exit Tests Do Not Apply: The file scc_dhcs_medical_enrollment.csv tracks 62 age, aid, and language cells in Santa Clara County from June 2025 (472,014 members) to June 2026 (411,722 members, net change -60,292 or -12.773%). Comparing each group's net change against the rounded -12.77% county rate (Observed_Drop - Base_2025 × 0.1277) gives a net gap of -7,230 for Spanish (-24,880 vs. -17,650 expected; -7,225 at unrounded -12.773%), -6,344 for Hispanic 19-44 (-6,341 unrounded), -5,697 for Male 19-44 (-5,700 unrounded), -10 for Samoan (47 to 31), -6.8 for French (56 to 42), and -4.2 for Thai (53 to 42). Because monthly DHCS snapshots measure net stock differences (exits minus new entries), overlap across age/sex/language tables, and cluster within households, individual member exits are not observed and binomial exit p-values are not valid on these snapshot counts.
• Mode A Multiple Testing Illustration & Verified CAPHE Inference Benchmarks (Mode B): Across m = 62 independent hypothesis tests at raw α = 0.05, a dashboard yields 62 × 0.05 = 3.10 expected false leads (1 - 0.95^62 = 95.8% chance of at least one false lead, with Bonferroni cutoff 0.05 / 62 = 0.00081). In CAPHE empirical studies where valid sampling distributions exist, three checks prevent Small-Group Swings: (1) across the top 3 Santa Clara County PUMAs for procedural Medi-Cal losses (scc_puma_pairwise_tests.json, 80 ACS replicate weights), South County (08519, 1,641 ± 186), East San Jose (08521, 1,590 ± 167), and Central/East San Jose (08510, 1,521 ± 166) have pairwise z = 0.35 to 0.72 (p = 0.469 to 0.730), proving the top-3 rank order is within survey noise; (2) across m = 16 lag horizons in Value_of_Public_Health/paper/ajph_submission.tex (Table 3), a 4-year first-difference result with raw p = 0.0049 shifts to p = 0.0788 under Holm and q = 0.0788 under Benjamini-Hochberg FDR; and (3) across 1,000 random label shuffles in Mcaid/05_fraud_detection/paper/output/results.json, the max null AUC is 0.593 versus 0.724 observed (p = 0.001).
• Extended Cell Context. What Drives the Rate in This Cell (French & Thai Cells). Across 62 cells, seeing -25.00% in French (n = 56) or -20.75% in Thai (n = 53) fits small-group swings. · What Drives the Rate in This Cell (Spanish Language Cell). Across 138,215 Spanish members, an -18.00% drop (-5.23 pp past the -12.77% county rate) gives a -7,230 gap. · What Drives the Rate in This Cell (Hispanic & Male 19-44). Young adults 19-44 see -20.11% (Hispanic, -17,380) and -20.19% (Male, -15,496) drops from moves and forms.
Why Health Teams Pair Size Floors and Multi-Test Checks With Net Headcount Scale
When a county health team tracks Medi-Cal renewal drops across 62 age, sex, and language groups (472,014 to 411,722 members in Santa Clara County; -60,292 or -12.77% net change), two common dashboard habits pull staff off course. First, sorting groups by raw percent drop puts tiny groups at the top. Samoan (-34.04%, 47 to 31 members; -16 people), French (-25.00%, 56 to 42; -14 people), and Thai (-20.75%, 53 to 42; -11 people) rank above Spanish (-18.00%, Rank #11 by percent). Yet Spanish-speaking homes account for 24,880 net lost members, or 41.3% of the entire county drop. Second, running 62 unadjusted tests at α = 0.05 produces 3.10 expected false leads by chance alone, and treating net monthly snapshot counts as independent member exits misstates uncertainty.
Pairing a size floor (n ≥ 100), valid multiple-testing checks (such as replicate-weight pairwise tests, Holm step-down, or Benjamini-Hochberg FDR where sampling variance is observed), and net headcount drop past the -12.77% county rate links statistical care with real scale. Once small cells go to partner outreach and large groups are ranked by net drop past the county rate, two major priorities stand out. First is Spanish-speaking homes (-24,880 net drop; -7,230 past the county rate). Second is young adults ages 19 to 44 (Hispanic 19-44 at -20.11%, -17,380 net drop and -6,344 past the county rate; Male 19-44 at -20.19%, -15,496 net drop and -5,697 past the county rate).
Data Sources, Model Inputs, Notes & Endnotes
(Numbered endnotes [1]–[5], state data sources & math rules)
Expand details
Case Study Focus: Study Credibility & Measurement · Multiple-Group Checks (Observed 62-Cell DHCS Grid & Mode A Rules)
Numbered Endnotes & Main Sources
- [1]California Department of Health Care Services (DHCS): Santa Clara County Medi-Cal Certified Eligibles across 62 Demographic & Language Cells (June 2025 to June 2026 snapshots, 472,014 to 411,722 enrollees).
- [2]Benjamini, Y., & Hochberg, Y. (1995): "Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing", JRSS-B, 57(1), 289–300.
- [3]U.S. Census Bureau: American Community Survey (ACS) 5-Year Language Spoken at Home & Residential Mobility Tables.
- [4]CMS Medicaid & CHIP Enrollment Trends
- [5]County of Santa Clara Social Services Agency & Health System
Observed 62-Cell DHCS Grid, Mode A Multiple Testing Rules & Mode B Benchmarks: All 62 group counts in June 2025 (472,014 total) and June 2026 (411,722 total), net headcount changes (-60,292 total; -24,880 Spanish; -17,380 Hispanic 19-44; -15,496 Male 19-44; -16 Samoan; -14 French; -11 Thai), and percent shifts come from scc_hr1_medicaid/scc_dhcs_medical_enrollment.csv. Expected false leads (62 × 0.05 = 3.10), Bonferroni cutoff (0.05 / 62 = 0.00081), and net drops past the -12.77% county rate are Mode A calculations. Replicate-weight PUMA pairwise tests (p = 0.469 to 0.730 in scc_puma_pairwise_tests.json), 16-horizon Holm/FDR adjustments (raw p = 0.0049 to Holm p = 0.0788 in Value_of_Public_Health/paper/ajph_submission.tex), and the 1,000-shuffle permutation test (null max AUC = 0.593, p = 0.001 in Mcaid/05_fraud_detection/paper/output/results.json) are Mode B benchmarks.
Main Sources: We cite Benjamini, Y., & Hochberg, Y. (1995), "Controlling the False Discovery Rate", JRSS-B, 57(1), 289-300 and California DHCS Medi-Cal Certified Eligibles & Redetermination Dashboards. We also use CMS Medicaid & CHIP Enrollment Trends and County of Santa Clara Social Services Agency & Health System.
Common Questions & Method Notes
- Why do small language cells like Samoan (-34.04%) and French (-25.00%) sit at the top of raw percent tables?
- In a group with just 47 members (Samoan) or 56 members (French), a net shift of 14 to 16 people (often three or four families moving or changing jobs) swings the cell rate by 25% to 34%. Without a size floor (such as n ≥ 100) and headcount weights, raw percent lists mostly show small-sample swings.
- Why can't we run simple binomial exit p-values on monthly DHCS enrollment snapshots, and how do Holm or FDR gates work on valid tests?
- Monthly DHCS snapshots show net stock changes (exits minus entries), overlap across age and language tables, and cluster by family, so individual exit counts are not observed. When valid sampling errors exist, such as the 80-replicate-weight PUMA tests (pairwise p = 0.469 to 0.730) or the 16-horizon public health spending check (where raw p = 0.0049 shifts to Holm p = 0.0788 and FDR q = 0.0788), Bonferroni, Holm, and Benjamini-Hochberg FDR prevent false leads.
- How is the Net Drop Past the County Rate calculated and why should dashboards show it alongside size floors?
- It compares a group's observed net change against what would happen at the county net rate (-12.77%). For Spanish-speaking members (138,215 base), a -12.77% shift equals -17,650 members. The observed net drop of -24,880 leaves a -7,230 gap past the county rate (-7,225 unrounded). That shows where bilingual outreach staff can help the most families keep care.