COLA computation methodology

This is the deep version. If you want the four-step summary, the plain-English walkthrough is here. What follows is everything needed to reproduce all 51 published adjustments from the raw index — including the parts that are not one rule but three.

1 · There is not one computation. There are three.

The rule people quote — third quarter against third quarter — applies from determination year 1984 onward. The first nine years used different quarters entirely, and any recomputation that applies the modern rule to them produces wrong answers for all of them.

2 · The base quarter is not simply the prior year

SSA states this rule in two places and states it differently in each. The series page says “the third quarter of the prior year”. The latest-determination page says “the last year in which a COLA became effective”. Those coincide in every year except the one following a zero — and in those years the prior-year reading is wrong.

The operative formulation is the second one. This site walks backward from the determination year to the most recent year whose published adjustment was greater than zero. On the published record that changes the answer for 3 payable years: 2011 and 2012 and 2017.

3 · The quarterly mean is rounded, and the precision changed in 2007

SSA rounds the arithmetic mean of the three months before computing anything from it. The precision is keyed to the determination year, not to the precision of the months themselves — which is the detail that catches recomputations out.

The clean demonstration: the same three months of the third quarter of 2006 — 199.200, 199.600, 198.400, summing to 597.200 — were averaged to 199.1 by the 2006 determination and to 199.067 by the 2007 determination. Same months, two answers, both correct in their own year.

4 · One adjustment in the record is not a computation at all

The 2000 adjustment was determined as 2.4% from the published index. Public Law 106-554 made it 2.5%. This site encodes that as a documented statutory override with its own test rather than as a fudge factor, and the page for that year says so plainly. The 2000 page.

5 · 6 adjustments cannot be reproduced from the index BLS serves today

This is a disclosure, not a caveat buried in a footnote. SSA computed every adjustment through the late 1980s from the CPI-W as published on the 1967=100 base. BLS rebased the index to 1982-84=100 in 1988, and the values now served for those months are one-decimal conversions.

The precision loss is the problem. An index near 300 carried to one decimal has roughly three times the relative precision of the same index near 100 carried to one decimal, so the rebasing discarded about two-thirds of the significant digits SSA actually worked from. For 6 determinations that is enough to flip the final tenth.

BLS’s own series metadata confirms only one base period is served, so the 1967=100 inputs are not publicly retrievable. This is a data-availability boundary rather than a disagreement about the rule: the statutory arithmetic is primary-sourced from the Federal Register determination notices and reproduces the other 44 determinations exactly. SSA’s published figure is what every page here states. The build gate allow-lists exactly these years and bounds each at one tenth of a percentage point, so the list cannot quietly grow.

6 · Exact arithmetic, and where rounding is allowed to happen

• Index values carry at most three decimals, so quarterly sums are computed in exact integer thousandths rather than in floating point.
• A published adjustment is carried as an integer number of tenths of a percent — 2.8% is the integer 28 — and applied as an exact integer operation on cents. The float route can land a hair below a whole dollar and floor to the wrong dollar, which is a real $1-a-month error on a page whose whole promise is exactness.
• Monthly benefit amounts floor to the whole dollar. The primary insurance amount floors to the next lower ten cents. Those are different steps and are never collapsed.
• The percentage is rounded half-up to the nearest tenth. Across all 51 determinations no year lands within five thousandths of a tie, so the tie-break rule is never exercised on the published record.

7 · A month that does not exist

The CPI-W series carries an explicit null for October 2025 — not a zero, not a missing key. Any function that averages a quarter containing it throws rather than skipping it. No adjustment computation touches that quarter, so nothing published here depends on it, but a naive month-by-month walk across the series would produce wrong answers in silence. What the gap does and does not affect.

8 · The gate

All 51 adjustments are recomputed from the committed CPI-W on every build, along with the Medicare tier reconstruction, the fact-sheet reconciliation and the full SSI derivation. If a single one disagrees, no page generates. The site-wide methodology covers the ingest, the sources and the scope limits.

• Site methodologysources and gates
• The plain versionfour steps
• The base quarterthe walk-back
• The CPI-W seriesthe input
• Every adjustment51 of them
• Sourcesevery dataset