Skip to content
Avanti Docs

Seat Accessibility Model

What this answers: of India’s engineering and medical (MBBS + BDS) seat capacity, how much is actually reachable by what fraction of students — once you gate each seat twice: once on merit (exam percentile) and once on money (fee vs family budget). Those two gates are deliberately anti-correlated in the Indian seat design — cheap seats have high merit bars; low-merit seats are expensive — so “how many seats exist” is a wildly misleading measure of access. This model computes the honest one.

Built for a policy / advocacy paper. The findings write-up is in paper.md; this README is how it’s built and how to reproduce it.

1. Worked example — one college, end to end

Section titled “1. Worked example — one college, end to end”

Take a private-state MBBS college with 150 sanctioned seats. The model does this:

  1. Split into regulatory buckets (rules in sources/assumptions.yaml → medical_split):
    • pvt_govt_rate = 75 seats (50%) — the NMC-2022 rule: half of a private college’s seats are charged at the government rate, filled merit-first.
    • management = 53 seats (35%) — FRC-capped but ₹10–25 L/yr.
    • nri = 22 seats (15%) — ₹25–50 L/yr, effectively affluent-only.
    • Check enforced in code: 75 + 53 + 22 = 150 = total intake.
  2. Attach a fee band and a merit floor to each bucket (from the config): pvt_govt_rate → ~₹1.2 L/yr, merit floor ~90th percentile; management → ~₹18 L/yr, floor ~50th (near qualifying); nri → ~₹40 L/yr, floor ~qualifying.
  3. A student reaches a bucket only if BOTH gates pass: merit_percentile ≥ floor AND fundable_budget ≥ fee.
    • A 90th-percentile student from a ₹1 L/yr-budget family reaches the 75 cheap merit seats — but nothing else in this college.
    • A 55th-percentile student from a ₹20 L/yr family reaches the 53 management seats (merit is a non-issue) but is locked out of the cheap ones (filled by higher scorers).
    • A 75th-percentile student from a ₹1 L/yr family — the squeezed middle — reaches zero seats here: too low for the merit-gated cheap seats, too poor for the management seats.
  4. Aggregate this over every college and every student cell → the reachability heatmap and headline statistics.

That single college shows the whole point: 150 “available” seats, but for a high-merit poor student only 75 are reachable, and for the squeezed middle, none.

2. What the model produces

Section titled “2. What the model produces”
Output (on GCS) What it is
exploded_buckets.csv one row per (college × seat-bucket): seats, fee band, merit floor, domicile scope. The model’s core table.
reachability.csv R[merit_band × budget_tier] = reachable seats, at every calibration (fee low/high × cutoff tight/loose × affordability conservative/generous).
headline_stats.csv the paper’s punchline numbers per calibration.
fig_heatmap.png reachable seats across merit percentile × budget tier.
fig_access_curves.png cumulative reachable seats vs merit, one line per budget tier.
fig_lorenz.png inequality-of-access curve (cumulative capacity vs cumulative students).
fig_capacity_trend.png capacity by college founding year (medical) + AICTE govt-vs-private intake trend (engineering).
bucket_summary.csv, fee_spotcheck.csv per-bucket table; Gujarat real-fee validation.

3. Provenance map — where every number comes from

Section titled “3. Provenance map — where every number comes from”

Pulled from BigQuery (avantifellows.external_data_sources):

  • nmc_fact_mbbs_seats — MBBS capacity: 780 colleges, 118,190 seats, management_category, year_of_inception. The MBBS split + trend anchor.
  • nirf_fact_aggregate — engineering capacity: first_year_intake per NIRF-ranked institute (latest academic_year per institute). This is the “where we have NIRF” engineering scope (~260 ranked institutes, ~252 k seats).
  • aicte_fact_intake — engineering govt-vs-private intake trend, 2012–13→2022–23.
  • plfs_fact_households — MPCE / household consumption distribution, used to ground the affordability budget-tier shares (weighted, calendar_2025).

Custom / built here (not in BQ):

  • Bucket split — sanctioned intake exploded into regulatory buckets (govt AIQ/state, private govt-rate/management/NRI, deemed management/NRI) via the rules in sources/assumptions.yaml. Encodes NMC-2022 (50% govt-rate), the 15% AIQ, and the deemed “no-government-quota” reality.
  • Two-gate reachability — the merit×money computation in build_model.py.
  • Management → archetype normalisation — NMC has no “Deemed” label (Government/Trust/Society/Private only); engineering archetype derived from College DB college_type.

Non-BQ inputs — fetched from GCS (College DB origin):

  • bds_all_colleges_2025-26.csv — BDS capacity (330 colleges; BQ has no BDS seat matrix). From Avanti’s College DB medical-national-ranks pipeline.
  • college_master.csv — engineering management archetype via nirf_id join.
  • gujarat_acpc_2024.csv — real per-course fees, used only to spot-check the indicative fee bands (Gujarat).

All modelling choices live in ONE file: sources/assumptions.yaml — split shares, fee bands, merit floors, budget tiers, qualifier counts. To change an assumption you edit that file and rerun; you never touch the code.

4. Reproduce it

Section titled “4. Reproduce it”
Terminal window
# 1. fetch the non-BQ inputs into sources/ (BQ + these = full reproduction)
gcloud storage cp \
  "gs://avantifellows-external-data/seat-accessibility/inputs/*" \
  analysis/neet-inequity-paper/sources/


# 2. build the model (reads BQ + sources/, writes output/)
python3 build_model.py


# 3. render figures + tables
python3 analyze.py

Requires bq authenticated (see repo docs/BQ_SETUP.md) and pandas, matplotlib, pyyaml.

5. Where the estimate is soft (read before citing a number)

Section titled “5. Where the estimate is soft (read before citing a number)”
  • Fees are indicative, not sourced per-college. No systematic Indian fee dataset exists. Bands come from published FRC caps / university notifications and are swept low/high to bracket. Gujarat ACPC real fees validate the accessible end (merit medical ~₹96 k/yr, engg ~₹75 k/yr — both inside the bands); management/NRI fees can’t be validated this way.
  • Management “cutoffs” are demand-clearing prices, not fixed floors — swept tight/loose.
  • Affordability is a modelling assumption. PLFS gives the consumption distribution (≈0.1 % of households consume ≥₹10 L/yr); fundable outlay lifts that with loans + an aspirant-skew factor. Reported at two calibrations. This input drives the answer as much as the seat data — never quote a single point.
  • Deemed medical colleges are not separately identified. NMC labels them Trust/Society/Private, so they get the private-state 50%-govt-rate bucket. If genuine deemed universities (Manipal, DY Patil, …) don’t honour that rule, the true government-rate share is lower than 40.8 % (nearer ~30 %) and the affluent-only share higher. Directional caveat; a deemed-identification join is future work.
  • Engineering scope = NIRF-ranked only. Undercounts the private long tail (AICTE shows ~1 M private engg seats vs the ~250 k modelled here) — so the engineering affluent-only share is understated.
  • Domicile: state-quota buckets carry scope=state; a reachable_seats_national column isolates what an out-of-state student can reach. Full state-by-state fragmentation is not modelled in v1.
  • Trend caveat: medical trend attributes current seat counts to each college’s founding year (survivorship + current-size proxy), not historical seat counts.

6. Replication data

Section titled “6. Replication data”

De-identified model outputs (college-level aggregates, no PII) on GCS: