How we compute things

This page renders our internal methodology document verbatim. When the method changes, the changelog says what changed and whether old numbers remain comparable.

v1.4 (2026-07-15; v1.0→v1.1 ACE-0 F1 re-noun, v1.2 residents interim, v1.3 households primary, see CONVENTIONS_CHANGELOG for each). The public methodology page is generated from this file (generator strips internal comments and normalizes dashes per brand floor).

What we rank

Municipalities (169 CT towns; ranked set = towns with sufficient data, see floors). Never USPS localities, never counties presented as towns (level disambiguation is load-bearing).

County-grain rankings (v1.4, the grain the interval data honestly supports today)

Counties are ranked per COMPLETE CALENDAR YEAR on customer-hours per 1,000 households (households = ACS 2021). A year is rankable only when (a) its raw interval data passes the month-grain coverage checks (edge checks are blind to mid-year holes; 2019's May-Aug gap is the proof) and (b) all 8 counties are present. Single-year county rankings are STORM-DRIVEN BY DESIGN: the #1 position typically belongs to whichever county the year's biggest storm hit, and every ranking surface must carry the worst-3-day concentration column so readers see that mechanism instead of inferring chronic difference. Ranks always render with their values. Multi-year averages publish only when >=3 rankable years exist (not yet).

The primary ranking metric (v1.3)

Customer-hours of outage per 1,000 households per year (households = ACS 5-year 2021 table B11001, the last vintage on legacy-county geography; 2020 Census population ships alongside as the secondary normalizer and the cross-check), computed per town from EAGLE-I public data: customer-hours = sum(customers_out × interval) over county intervals, apportioned to towns via the town-level snapshot where available and county scaling elsewhere (every row carries apportionment: measured | county-scaled, the disclosure key). Once live-era observation matures (≥12 months) AND the Phase-4 legal gate opens, our observed incident history joins as a second, separately-labeled series.

• Why households, not "metered customers" (ACE-0 F1): per-town meter counts are not public; EIA-861 customer counts exist only at utility grain. Calling Census households "customers" would violate our own noun discipline (conventions §3). The denominator is named exactly what it is; utility-grain EIA customer counts appear on the methodology page for scale context only. Households ≈ metered residential accounts far more closely than population does (multi-person homes share one meter), which is why households is primary and population secondary.
• It is a RATE (normalized by size) with the N attached, not a raw count that crowns big towns.
• "Hours without power" is the reader's own unit, extractable, comparable, honest.
• Inputs trace to the published datasets on the open-data page: the DOE EAGLE-I derived town outage-history dataset (customer-hours), the town denominators dataset (households, population), and the as-filed reliability dataset (the comparison series).

Presented alongside (never merged into one score)

1. Frequency: reported outage events per year (feed + definition stated per §2 of conventions).
2. Duration: median and p90 estimated restoration time (quantized-to-poll caveat).
3. As-filed vs as-observed: the utility's SAIDI-style figures (major event days EXCLUDED by IEEE 1366) next to all-days observations, the disagreement is the story, presented as method.
4. The trend: is the town improving or worsening (window-over-window, only when both windows are complete per the completeness frame).

Floors and refusals (what we do NOT rank)

• Minimum-N floor: a town enters the ranking only with ≥ 12 months of healthy-feed coverage AND ≥ 5 attributed events or ≥ 1,000 Census households (the apportionment base). Below floor → "insufficient data to rank" (an honest badge, not a zero, not an omission).
• Eastern CT (no deep history): shown with live data + "tracking since June 2026", excluded from historical rankings until 12 months accrue.
• major_per_yr = 0 (no-household-match sentinel) → NULL → below floor. Never rank on it.
• No ranking on a window containing a documented feed gap unless the metric is gap-insensitive (stated per metric in the completeness ledger).

Ties, stability, cadence

• Ranks carry their value: "#3 (41 outage customer-hours per 1,000 households per year)", never a bare ordinal.
• Ties within the value's precision are shown as ties (=3), not broken cosmetically.
• Recompute cadence: monthly for historical rankings (stability > freshness for evergreen pages); live status is separate and never re-orders the historical rank between recomputes.

Every ranking page states

1. Metric, window, feed(s), and N (the denominator and its unit).
2. The scaled-estimate disclosure where county apportionment is used.
3. Major-event-day handling (both views).
4. Date of computation + methodology version.
5. What would change the number ("one large storm can move a small town several places, that is the nature of the data, and why we show the N").

The bias checks (run per recompute, results kept)

• Direction audit: does this recompute make the grid look worse than the last one? If yes, verify harder (it favors us). If a methodology change ONLY ever worsens grid optics, it fails the barium test and does not ship.
• Mirror audit: did any town's number improbably improve (the reassuring-bug direction, HAA5 precedent)?
• Spot-verify the top-3 and bottom-3 towns against raw feed data by hand, every recompute.

Data sources and conventions on ranking surfaces

Outage records: DOE EAGLE-I (county grain; raw 15-minute intervals for customer-hours). Filed reliability: EIA-861. Denominators: Census (2020 base; ACS 2021 households). Town-grain measured rows: the February 23, 2026 blizzard snapshot, Eversource and United Illuminating counts as published by Patch.com via its live Datawrapper embed (14:16 EST Eversource, 14:12 EST UI; a mid-afternoon snapshot, not the event peak, which came that morning). Zeros are kept: a measured zero is a fact about that moment. Towns served mostly by municipal utilities (Bozrah; the Norwich and Wallingford munis; Groton Utilities' territory) are not covered by that source and their pages say so.

Change control

Metric changes = version bump + changelog + side-by-side old/new publication for one cycle (publish the loosenings too, rankings that only ever get scarier are marketing, not measurement).