A ranking should show its working.

1. Define the population and geography

The development cohort is the 63 active cities currently covered by BestCityToLive. Coverage does not imply a representative sample of every city worldwide. Scored observations require compatible harmonized urban-area boundaries. Municipal, census-metropolitan and police-metropolitan observations are separately labelled until a documented crosswalk supports comparability. Country averages are excluded.

2. Publish the weights before calculating results

Means receives 25% because household resources constrain everyday choices. Infrastructure, trust and care, and environment each receive 20% as core living conditions. Opportunity receives 15%. These are editorial policy choices, not statistically discovered truth. The same rules apply to every city; ranks are never capped or manually promoted.

3. Normalize against explicit policy anchors

Each indicator has a unit, direction, lower anchor and upper anchor. For a beneficial indicator: 100 × clamp((value − lower) / (upper − lower), 0, 1). For a burden indicator, subtract that result from 100. Anchors are provisional model choices, frozen in v1.0.0 and available in the dictionary below. They need external review before an official release.

Fixed anchors avoid a city’s score changing simply because another city joins the sample. We will review correlated measures before accepting inputs so the same underlying condition is not counted repeatedly.

4. Require evidence instead of filling gaps invisibly

Each scored observation needs an HTTPS source URL, exact unit, observation end date, compatible urban geography and explicit review status. Missing, future-dated, stale, non-finite and incompatible records are rejected. Maximum ages range from one to three years by indicator. Latest accepted observations are selected per indicator.

A city needs at least 90% weighted indicator coverage overall and 80% in every pillar. Each pillar has six equally weighted indicators. Missing accepted indicators are reweighted within that pillar only. No pillar is removed to improve a city’s eligibility. Sparse cities receive no score. A common-indicator and geography audit remains required before release.

5. Reward balance and expose uncertainty

A pillar score is the arithmetic mean of its accepted indicators. The composite is a weighted geometric mean: exp(Σ weight × ln(max(1, pillar score))). The one-point floor permits logarithms for zero scores and is part of the published model. This aggregation limits how easily a strong pillar compensates for a weak one.

We rank by unrounded scores and use city slug for exact ties. Displayed scores will use one decimal; differences below 0.2 points should be treated as near ties. Sensitivity checks vary each pillar’s weight by ±20% and omit each pillar in turn, re-normalizing the remaining weights. The resulting rank range is a model-sensitivity range, not a statistical confidence interval.

6. Freeze quarterly releases; check sources weekly

Weekly collection records source revisions without inventing changes. Official editions require an editorial review, a fixed data cutoff, an eligible cohort, archived observations and a model version. The current edition is research only. A failed source check retains the prior snapshot and records the failure. Source checks, underlying data changes and results publications have different dates.

The 30-indicator dictionary

All indicators await reviewed, compatible observations. Anchors are design parameters, not measured city values.