# APAC Tertiary-Institution Quality Index — Full Roster Results (2026-06) 32 clusters, 50 anchor institutions, all four pillars. Supersedes the 5-cluster provisional map. Method per `apac-institution-quality-index-spec.md` v0.1: base-effect log-damping → tier-relative z-score (locked) + a global-z view for cross-tier comparison; retention = additive level pillar. ## Cross-tier map (global z; origin = roster average) ``` TRAJECTORY (momentum) ^ [global z; origin = roster avg] Jk | Dk Ta | Hy | Vn| Mn Ch | Dl HK | | Az | Mu | Kg ----------------Ul----------+-Bk-------------------------------- LEVEL -> We Ak | Sg Bg Mb | | Se Br | | Ts | Po Pe | Ng Cb | | Tk | Kh | Dj ``` Codes: Sg Singapore · Se Seoul · Dj Daejeon · Po Pohang · Tk Tokyo · Kh Keihanshin · Ng Nagoya · Ts Tsukuba · Bg Bengaluru · Dl Delhi · Mu Mumbai · Ch Chennai · Hy Hyderabad · Kg Kharagpur · KL Klang Valley · Jk Jakarta · Bk Bangkok · Mn Manila · Vn Vietnam · Dk Dhaka · Mb Melbourne · Sy Sydney · Cb Canberra · Br Brisbane · Pe Perth · Ak Auckland · We Wellington · Cc Christchurch · Az Astana–Almaty · Ta Tashkent · Ul Ulaanbaatar · HK Hong Kong ## Quadrant reading - **Accelerating leaders** (high level + high momentum): **Hyderabad, Chennai, Delhi** (the IIT catch-up cohort) — the standout cell. - **Rising stars** (low level, high momentum): Jakarta, Tashkent, Dhaka, Vietnam, Hong Kong. - **Established / plateauing** (high level, low momentum): **Pohang, Daejeon, Seoul, Tokyo, Keihanshin, Nagoya, Tsukuba** — the Korea/Japan frontier: top captured quality, low growth. - **Lagging** (low level, low momentum): the Australia/NZ clusters (Canberra, Perth, Brisbane, Wellington, Christchurch, Auckland) and Ulaanbaatar. ## Global-z composite (cross-tier comparable), sorted by Level | Cluster | Tier | Level | Traj | |---|---|---|---| | Pohang | frontier | 1.60 | −0.71 | | Daejeon | frontier | 1.10 | −1.52 | | Seoul | frontier | 0.74 | −0.41 | | Tokyo | frontier | 0.67 | −1.08 | | Keihanshin | frontier | 0.64 | −1.36 | | Hyderabad | catchup | 0.55 | **1.22** | | Delhi | catchup | 0.49 | 0.64 | | Chennai | catchup | 0.47 | 0.81 | | Nagoya | frontier | 0.44 | −0.87 | | Tsukuba | frontier | 0.44 | −0.55 | | Klang Valley | catchup | 0.43 | −0.10 | | Bengaluru | catchup | 0.34 | −0.04 | | Kharagpur | catchup | 0.33 | 0.17 | | Mumbai | catchup | 0.16 | 0.35 | | Astana–Almaty | emerging | 0.11 | 0.37 | | Bangkok | catchup | 0.11 | 0.09 | | Singapore | frontier | 0.11 | −0.13 | | Manila | emerging | −0.04 | 0.85 | | Vietnam | emerging | −0.12 | 0.96 | | Tashkent | emerging | −0.23 | 1.16 | | Perth | frontier | −0.36 | −0.81 | | Jakarta | catchup | −0.36 | **1.29** | | Sydney | frontier | −0.42 | −0.47 | | Melbourne | frontier | −0.45 | −0.22 | | Auckland | frontier | −0.45 | −0.15 | | Christchurch | frontier | −0.45 | −0.06 | | Brisbane | frontier | −0.50 | −0.46 | | Hong Kong | frontier | −0.53 | 0.57 | | Ulaanbaatar | emerging | −0.64 | 0.04 | | Canberra | frontier | −0.65 | −1.05 | | Wellington | frontier | −0.67 | −0.13 | | Dhaka | emerging | −1.28 | 1.14 | ## Tier-relative momentum (who's rising *within* their peer group) - **Frontier** rising-stars: Hong Kong, Christchurch, Auckland, Wellington, Melbourne, Sydney (Anglo + HK gaining on the Asian frontier); established: Pohang, Daejeon, Tokyo, Keihanshin, Nagoya. - **Catch-up**: Jakarta (rising-star); Hyderabad, Chennai, Delhi (accel-leaders); Bengaluru, Klang Valley, Kharagpur (established); Bangkok, Mumbai (lagging). - **Emerging**: Tashkent, Dhaka (rising-star); Vietnam (accel-leader); Manila, Astana–Almaty (established); Ulaanbaatar (lagging). ## Key findings 1. **Korea + Japan dominate captured quality** (Pohang, Daejeon, Seoul, Tokyo, Keihanshin top the Level axis) — driven by high domestic-firm linkage (Pohang 75%, Tokyo 68%) and retention (~80–87%). But they are **plateauing** (most negative trajectory) — the mature-frontier signature. 2. **The IIT catch-up cohort (Hyderabad, Chennai, Delhi) is the genuine accelerating-leader cell** — already above roster-average on captured quality *and* fastest-growing. The clearest "rising in a way that should feed competitiveness" story. 3. **Australia/NZ score low on *captured* quality** despite research strength — because the index weights domestic-firm linkage (~10–13%, a resource/MNC economy), STEM share (AUS 17%), and retention. This is a value-capture reading, not a research-output reading; interpret accordingly (§11). 4. **Hubs (Hong Kong, Singapore) are pulled down by the retention pillar** (HK 20.5%, SG 39%) — the known foreign-trainee-return confound, now additive (not multiplicative) so the drag is bounded. Both still show positive momentum. 5. **Emerging risers** (Dhaka, Tashkent, Jakarta, Vietnam) cluster top-left: explosive momentum off low bases. (Manila, now correctly on UP Diliman, sits mid-pack — established, not a riser.) ## Caveats & accepted limitations *(Acknowledged as known limitations of this build, not blocking — to be revisited only if the index goes to external/formal use.)* - **Pillar A imputation:** Japan and Hong Kong have no World Bank STEM-share data → imputed with tier means. Affects the 4 Japan clusters + HK on Level. - **Australia enrolment CAGR (−4.1%)** likely reflects the COVID international-student collapse, not a real capacity decline — A-trajectory for AUS is suspect. - **Small samples:** Tashkent (domestic-firm n=4, retention n=5), Astana–Almaty (n≈21–35), Manila (n=41) — low-output clusters have noisy Pillar B/D proxies. - ~~Manila = UP Manila~~ **fixed (2026-06):** Manila now resolves to UP Diliman (flagship) + Ateneo; recomputed — Manila moved from "rising star" to "established". - **Retention is bibliometric proxy** — valid for non-hubs; hub values (SG/HK, partly AUS/NZ) need migration/labour data to be fair. - **Domestic-firm & retention sampled at the primary anchor only** per cluster (not all anchors). ## Data lineage (in pilot_data/) ids.tsv (roster+IDs) · stage2.tsv (Pillar C/B counts, 50 inst) · stage3_pillarA.tsv (World Bank, 16 countries) · stage4.tsv (domestic-firm + retention, 32 anchors) · stage5_composite.tsv (final Level/Traj) · resolve.py/stage2-5.py (reproducible pipeline). --- # §11 Per-Economy Interpretation Layer (built 2026-06) **Method.** The two value-capture metrics — domestic-firm linkage (Pillar B level) and retention (Pillar D) — are scored as **deviation from the cluster's own economic-model baseline** (residual ÷ global residual spread), instead of against a universal mean. Performance metrics (research C, human-capital A) stay tier-relative as locked. This stops the index from penalizing economies whose *structurally* low value-capture reflects their economic model, not university weakness. **Economic-model taxonomy** (judgment call — documented as such): - **national-champion** (KOR, JPN): Seoul, Daejeon, Pohang, Tokyo, Keihanshin, Nagoya, Tsukuba - **conglomerate** (IND): Bengaluru, Delhi, Mumbai, Chennai, Hyderabad, Kharagpur - **advanced-open** (SGP, HKG, AUS, NZL): Singapore, Hong Kong, Melbourne, Sydney, Canberra, Brisbane, Perth, Auckland, Wellington, Christchurch - **emerging-FDI** (MYS, VNM, THA, IDN): Klang Valley, Vietnam, Bangkok, Jakarta - **emerging-thin** (BGD, PHL, KAZ, UZB, MNG): Dhaka, Manila, Astana–Almaty, Tashkent, Ulaanbaatar ## Two lenses — read them together The naive and §11 Levels answer **different questions**, and the *gap* between them is the insight: - **Naive Level = absolute local value-capture** ("how much value does the economy actually retain/feed locally?") — the competitiveness-relevant quantity; Korea/Japan lead because they genuinely capture more. - **§11 Level = university quality net of economic structure** ("how good is the system, controlling for its economy?") — fairly credits AUS/HK; stops conflating university quality with economic model. | Pattern | Clusters | Meaning | |---|---|---| | High naive, **low §11** | Seoul, Daejeon, Tokyo, Keihanshin, Tsukuba | Strong capture is largely the *economy's* structure; universities good-not-exceptional among champion peers | | **Low naive, high §11** | Australia/NZ, Hong Kong | Strong universities **held back by economic structure** (deep-tech-industry gap, not a university gap) | | High on **both** | **Pohang, Hyderabad, Chennai** | Genuinely exceptional on their own terms | | Low on **both** | Dhaka, Ulaanbaatar | Thin across the board | ## §11 model-adjusted map ``` TRAJECTORY ^ [§11 model-adjusted Level; origin=roster avg] | Jk | Dk Ta| Hy | Vn | Mn Ch | Dl HK | | Az Mu | Kg -----------------Ul--------+---Bk------------------------------- We Ak Bg KL Mb Se | Bry Ts | | Po NgPe Cb | | Tk | |Kh | Dj | MODEL-ADJUSTED LEVEL (quality net of economic model) -> ``` ## §11 effect (Level change vs naive), selected | Cluster | Model | Naive | §11 | Δ | |---|---|---|---|---| | Auckland | advanced-open | −0.45 | −0.01 | +0.44 | | Melbourne | advanced-open | −0.45 | −0.05 | +0.40 | | Hong Kong | advanced-open | −0.53 | −0.35 | +0.18 | | Singapore | advanced-open | 0.10 | 0.18 | +0.08 | | Pohang | national-champion | 1.60 | 1.70 | +0.10 | | Tokyo | national-champion | 0.67 | 0.09 | −0.58 | | Daejeon | national-champion | 1.10 | 0.17 | −0.93 | | Seoul | national-champion | 0.75 | −0.57 | −1.31 | **Recommendation:** report naive + §11 side by side. Naive answers the competitiveness hypothesis directly (absolute capture); §11 isolates university quality. Neither alone is complete. ## §11 caveats - **Model assignment is a judgment call** (e.g. Vietnam→emerging-FDI, Central Asia→emerging-thin); sensitivity to regrouping not yet tested. - **Astana–Almaty +0.49** is inflated by a tiny-sample domestic-firm value (34% on n=35) sitting far above the emerging-thin baseline — small-sample artifact, not a real signal. - A finer build would use **separate model groupings for industry-capture vs talent-capture** (e.g. SG/HK are transit-hubs for talent but MNC-hubs for industry; Vietnam is FDI-industry but brain-drain for talent) — collapsed here into one taxonomy. - Data: `stage6_peconomy.tsv`, `stage6_peconomy.py`, `map_s11_ascii.txt` in pilot_data/. --- # Sensitivity analysis (built 2026-06, stage7_sensitivity.py) Tests whether the conclusions are robust to the two judgment-laden choices: **pillar weights** and **§11 model groupings**. ## 1. Weight robustness — the index is stable - **Equal-weights vs base tier-weights: Spearman ρ = 0.990.** The whole weighting scheme barely moves the ranking — results are *not* an artifact of the chosen weights. - **Monte-Carlo (3000 Dirichlet draws around base weights):** top and bottom are pinned — Pohang (rank 1→1), Daejeon (2→2), Dhaka (32→32) never move; the Korea/Japan champions, the IIT cohort, and Australia/NZ all swing ≤5 ranks. - **Genuinely weight-sensitive clusters** (rank swing ≥10, p5–p95): **Singapore (7–20), Tashkent (14–29), Klang Valley (3–15), Hong Kong (21–31), Nagoya (6–16).** These are exactly the model-contested / small-sample clusters already flagged — sensitivity localizes the uncertainty rather than undermining the headline. ## 2. Leave-one-pillar-out (ρ with base ranking) | Pillar dropped | ρ | Biggest mover | Reading | |---|---|---|---| | A human capital | 0.899 | Tashkent 20→32 | Tashkent's rank was propped by high STEM share (UZB 34.5%) | | B innovation | 0.883 | Keihanshin 5→16 | Japan's standing leans on domestic-firm linkage (national-champion) | | C research | 0.949 | Ulaanbaatar 29→20 | weak-research clusters rise when C is removed | | D retention | 0.937 | **Hong Kong 28→17** | **quantifies the retention drag: HK rises 11 ranks without it** | All ρ ≥ 0.88 → no single pillar drives the result; each contributes without dominating. ## 3. §11 model-grouping robustness - **ALT1** (merge the two emerging groups): ρ = 0.978 — robust; Vietnam moves most (22→13). - **ALT2** (split advanced-open into MNC-hub SG/HK vs resource AUS/NZ): ρ = 0.958 — robust; **Hong Kong 26→15, Singapore 10→5** (a 2-member hub group flatters them further). Grouping choice matters most for the hubs — exactly where flagged — but overall §11 ranking is stable. ## Verdict The headline conclusions are **robust**: Korea/Japan lead captured quality (and plateau); the IIT cohort (Hyderabad/Chennai/Delhi) are the accelerating leaders; emerging risers and the AUS/NZ "captured-quality gap" all survive perturbation. The **only genuinely fragile positions are Singapore, Hong Kong, Tashkent, and Klang Valley** — the model-contested and small-sample clusters — which should be read with the §11 second lens and sample caveats, not as point estimates. --- # Refinements round 2 (2026-06) ## R1. Dual model groupings (industry vs talent) The §11 single taxonomy is split: **domestic-firm linkage** is benchmarked against an *industry-structure* model, **retention** against a *talent-flow* model — because they diverge. (E.g. Australia/NZ are resource-services for industry but education-hub-transit for talent; Vietnam is MNC-dependent for industry but brain-drain for talent.) - Industry models: national-champion (KOR/JPN), conglomerate (IND), MNC-dependent (SGP/HKG/MYS/VNM/THA/IDN), resource-services (AUS/NZL/KAZ/MNG), thin-industrial (BGD/PHL/UZB). - Talent models: strong-retention (KOR/JPN), education-hub-transit (SGP/HKG/AUS/NZL), large-domestic (IND), brain-drain-emerging (rest). - Effect: Hong Kong (+0.33) and Singapore (+0.21) rise more cleanly than under the single taxonomy; **Astana–Almaty's small-sample spike is tempered (+0.21 vs +0.49)**. Data: stage8_dualmodel.tsv / .py. ## R2. Real STEM data for Japan & Hong Kong (replaces imputation) - **Japan = 20.0%** (OECD Education at a Glance 2025: 20% of bachelor's graduates in STEM, vs OECD avg 23%). Slightly *below* the prior tier-mean imputation → Japan clusters ease down marginally (correct direction). - **Hong Kong = 33.0%** (estimate from UGC broad-academic-category enrolment: Sciences + Engineering/Technology ≈ a third of UGC-funded students; documented as approximate, not a UIS figure). *Above* the prior imputation. - Combined with R1, Hong Kong moves from −0.53 (naive, double-penalised by low-imputed STEM + hub retention) to **+0.11** (dual-§11) — now near roster-average, far fairer. ## R3. Real migration data for hub retention — attempted; validates §11 - **Data dead-ends:** the canonical World Bank indicator `SM.EMI.TERT.ZS` (emigration rate of tertiary educated) exists in the catalogue but has **no API time-series**; the underlying Docquier/WB-Factbook brain-drain data is static (~2000–2010 vintage) and the source PDFs would not parse here. A full implementation needs OECD DIOC or KNOMAD bilateral micro-data. - **What the evidence shows (and it validates the §11 dual-model rather than requiring new inputs):** - **Singapore** is a *net skilled-talent magnet* — its low bibliometric retention (39%) is a transit confound; true national retention is high. → §11 education-hub-transit benchmarking already corrects this (Singapore +0.21). ✓ - **Hong Kong** has *genuine* skilled emigration (large historical diaspora to CA/AU/US **plus** a post-2020 National-Security-Law wave; MigrationPolicy.org, East Asia Forum) — so its low retention is **partly real, not pure transit**. Caveat: do **not** over-rescue HK beyond the dual-model correction. - **New Zealand** leaks skilled labour to Australia (real outflow) — its moderate retention is partly genuine. - **Korea/Japan** have low skilled emigration → genuine high retention (confirms the bibliometric reading). - The bibliometric proxy already ranks **Singapore (39%) > Hong Kong (20.5%)**, matching the migration reality (magnet vs. real drain) — so the *relative* ordering is sound; only the *absolute* hub level was confounded, which §11 fixes. - **Conclusion:** no model change beyond §11 dual-model is warranted on current evidence; migration data confirms the structural fix and adds the "don't over-correct HK" caveat. Replacing the proxy with real bilateral data remains the only true fix for formal/external use. ## Final map (Level = dual-model §11; Japan/HK STEM sourced) ``` TRAJECTORY ^ [Level = dual-model §11; Japan/HK STEM sourced] | Jk| Dk Ta | Hy | |Vn | Mn Ch | Dl | HK | Az Mu | Kg ------------Ul-------------+-----Bk----------------------------- We Ak BgSgKL Mb Se | Bry Ts | | Po Ng|Pe Cb | |Tk | Kh | Dj | MODEL-ADJUSTED LEVEL -> ```