YouthPOWER District Score: Technical Methodology Annex

This annex sits behind the public methodology note and gives the full specification: the universe, the principles that decide which measures are scored, the scoring formulas, the weights, the treatment of missing data, the robustness tests with their results, and the complete list of scored measures with each one’s target, the basis of that target, and the issuing ministry and year. It is meant to be read by anyone who wants to check the score rather than simply use it.

A1. Scope and universe

The score covers every district in the country, around eight hundred in all, on a single release. It reads around one hundred and eighty measures, drawn from twenty-seven government databases, in five areas: Education, Skilling, Opportunity, Work, and Participation. Fifty-four of those measures carry the score directly; the remainder are shown alongside it as supporting evidence and do not move the score. Young people are taken as those aged 15 to 29, with some measures using a wider window up to 35, and the school-age band of 15 to 21 where a measure concerns being in education. Every measure, scored or shown, is attributed on its card to the issuing ministry and the year the data describes.

A2. Which measures are scored

Every measure, scored or shown, comes from official government data. Around one hundred and eighty are in scope; fifty-four carry the score and the rest are shown as evidence. Five written principles decide which role a measure takes.

1. Does it carry the pillar’s central question? Each pillar makes a claim it must answer: Education, whether young people are in school and learning; Work, whether they are in decent jobs. A measure that helps answer that question is a candidate to be scored; one that is relevant but peripheral is kept as context.

2. Is it available at the district level, for enough districts? A measure is scored only where it is measured for the district itself, or imputed to it by an explicit method. Measures that exist only at the state or national level, such as district GDP or per-capita income, are kept out of the score, as are measures whose district coverage is too thin to be reliable; these may still be shown as evidence.

3. Is it clean, or confounded? This disqualifies the most candidates. Measures attached to the wrong place (credit structurally concentrated in the metros), measures that mislead in the Indian context (a low open-unemployment rate usually reflects distress self-employment rather than strength), measures whose direction is ambiguous (counts of enterprises or schools can signal either depth or a surfeit of tiny units), and political representation, which cannot fairly be reduced to a score, are all shown where they inform but kept out of the score.

4. Does it add information, or echo another measure? Where several measures move together and carry the same signal, one is scored and the rest become evidence: labour-force participation is scored on the prime-age 22-35 cut, with the other cuts shown. Where several items express one idea, the composite is scored and its parts are shown beneath it.

5. Does it vary across districts? A measure that is near-identical in almost every district cannot tell districts apart and is shown as evidence rather than scored.

Every scored measure is then placed on the common 0 to 100 scale described next, so districts are compared on like terms.

A3. Normalization

Each scored measure is placed on a 0 to 100 scale in one of two ways.

Distance to a target. Where a measure has a fixed target, the score is the distance from a floor (the value scoring zero) to that target:

score = 100 × (value − floor) ÷ (target − floor), bounded to the range 0 to 100,

oriented so that higher always means better. For measures where lower is better, such as child stunting or trainer vacancy, the orientation is reversed so that approaching the target still raises the score. Because the floor and target are fixed, the same value yields the same score in every district and in every release.

Position in the national spread. A minority of measures have no natural fixed target, for example the number of bank branches per hundred thousand people. These are winsorised at the first and ninety-ninth percentiles to limit the influence of extreme values, then mapped linearly so the lowest maps to 0 and the highest to 100. A logarithmic scale is used where a measure is strongly skewed. These measures are labelled as scored on their spread.

A4. The basis of every target

Every target is one of five kinds. Of the fifty-four scored measures (the two poorest-quintile measures are scored as a single composite, so they share one target):

The three FOIF-calibrated targets are the only ones that rest on our own judgement rather than an external reference. The basis of every target is recorded for all fifty-four scored measures and set out in the indicator table in A9.

A5. Aggregation and weights

Measures are combined from the bottom up. Within a family of related measures the weights are equal. Families combine into a pillar, and the five pillars combine into the overall score.

Foundation and outcome. Within the areas that have both, foundation measures (the conditions in place) and outcome measures (the results achieved) are blended, with more weight on outcomes, and a bounded penalty applies where foundations run well ahead of outcomes:

pillar = 0.40 × Foundation + 0.60 × Outcome − minimum of [ 0.20 × (Foundation − Outcome, if positive) , 10 ].

Pillar weights. The five pillars carry fixed, recorded weights:

Opportunity and Work carry the most weight as the outcomes the score leans toward; Education and Skilling, upstream of those outcomes, carry less.

Arithmetic aggregation. Pillars are combined by a weighted average. A weighted average is compensatory: the same overall score can be reached by different combinations of pillar scores, so the overall figure measures a district’s total standing against the standards rather than its profile. The profile is not lost; it is preserved and published in the pillar and indicator breakdown, which is exactly where two districts with the same overall score differ. The overall score is therefore an absolute, comparable measure of level, and the breakdown is the measure of shape. We checked the aggregation choice against a geometric mean, which penalises uneven districts more sharply (see A8), and found it changes the ordering very little, so the simpler and more transparent average is retained and the breakdown carries the imbalance.

A6. State and national figures

A state or national figure combines its districts in two ways, depending on what is shown.

The state or national score is not a simple average of district scores. Each pillar score, and the overall, is a population-weighted mean (half the weight), the median (three tenths), and the average of the weakest quarter of districts (two tenths), so a state is credited for lifting its floor, not only its average.

A state or national indicator value, the underlying number shown on each card, is a population-weighted aggregate of its districts, presented together with the spread of district values, so the headline figure and the variation behind it are both visible.

A7. Missing data and reliability

Reweight to what is present. Where a measure is missing for a district, its weight is redistributed across the measures that are present, so a district is scored on what it has rather than penalised for a gap in the data.

The not-rated rule. If fewer than half of an area’s measures are available for a district, that area is left unrated rather than scored on too little. Unrated areas are shown as such and do not enter the overall score.

Thin-sample suppression. For survey-based measures, a district figure is suppressed at source where its sample falls below a minimum of about fifty observations, and the district instead carries the next reliable level of the cascade, a pooled regional or state figure.

How much is cascaded. Across the fifty-four scored measures, about 88 percent of district figures are measured directly in the district and the remaining 12 percent are cascaded from a pooled regional or state level. The cascading is concentrated in the survey-based work and agency measures; the gender wage-parity measure, a ratio of two cascaded quantities, is wholly derived.

A8. Robustness and uncertainty

Every test below was run on the current release and the results are published here.

Weighting. All fifteen reasonable alternative weightings were tested: each pillar’s weight moved up and down by five points, all pillars set equal, the foundation-outcome blend moved between 0.35 and 0.45, and the gap penalty moved between 0.10 and 0.30. Across every one, the rank correlation of district scores against the published version stays at or above 0.99, and at most two of around eight hundred districts move by more than one tenth of the distribution. The comparison between districts does not depend on the particular weights chosen.

Sampling uncertainty. For the survey-based measures, the score was recomputed five hundred times, each time perturbing every survey value within its sampling error, to read how much a district’s score moves from sampling alone. At a realistic survey design effect, a district’s overall score carries a ninety-percent band of about plus or minus 0.9 points, and never more than 2.6 points across all districts. Two districts are a real difference only when their scores differ by more than roughly 1.2 points; within that they should be read as equivalent. The precision varies by area:

Education and Skilling rest entirely on administrative records and carry no sampling error. Participation rests entirely on survey measures and carries the widest band, so Participation comparisons need a larger gap to be real.

Balance across areas. A geometric-mean recomputation, which prevents a high score in one area from offsetting a low score in another, leaves the ranking essentially unchanged (rank correlation 0.99), confirming that the ordinary average is not masking lopsided districts.

Internal consistency. Within each pillar the measures hold together to the degree expected of a deliberately broad composite. Internal-consistency coefficients run from about 0.33 to 0.62, which is appropriate here: these pillars combine complementary dimensions on purpose, and a very high coefficient would signal redundancy rather than quality.

Reweighting check. Scoring all districts on only the measures available everywhere, rather than reweighting to what each has, changes the typical district’s score by under half a point and correlates with completeness at near zero, so low-coverage districts are not flattered by the reweighting rule.

Outside comparison. The overall score was checked against independent district measures built by other bodies. Against NITI Aayog’s National Multidimensional Poverty Index, across about seven hundred districts, it is strongly and inversely correlated (rank correlation about −0.74): districts that score higher for young people are districts where multidimensional poverty is lower. It moves in step with independent state human-development assessments across roughly three hundred and sixty districts (rank correlation about +0.60, and +0.7 to +0.9 in the larger states). The relationship with a national list of districts already identified for focused development runs in the expected direction, and agreement with an independent measure of household consumption is strong at the state level.

Two cautions apply. The score is youth-specific and is built to track, not reproduce, a general development measure, so it agrees most on the material dimensions and keeps its own signal on the youth-civic ones, which is intended. And the poverty and human-development comparisons rest on older reference years than the current score (the human-development assessments especially, many a decade or more old, on district boundaries that have since changed), so the true present-day agreement is likely somewhat higher than these figures show.

A9. The scored measures

The fifty-four measures that carry the score, with each one’s layer, target, the basis of that target, and the issuing ministry and year.

A10. Vintages, refresh, and limitations

The index is refreshed as new official data is released, and each measure carries the year it describes. The measures do not all describe the same year, and we keep this visible. Work, economic, and education measures reflect the most recent year available. The household and family measures in Opportunity and Participation reflect the most recent survey of its kind, attributed by its year on each card; there is at present no newer source for these at the district level.

The score measures each district against fixed standards and is designed to compare like with like and to track a district against itself over time. Because many districts sit close together, the meaningful reading is the score and its distance to each target; differences smaller than the uncertainty band in A8 should be treated as equivalent rather than as a ranking position.

This annex is published in full so the score can be checked at the level of any single measure.