PRESS RELEASE — ANNEXURE A

Methodology, Sampling & Transparency Disclosure

West Bengal Legislative Assembly Election 2026 — People’s InsightFinal Exit Poll Projection

An exit poll is, finally, an act of public trust. As Chief Executive of Encuesta People’s Insight Private Limited, I take personal ownership of every number we publish. This Annexure is being released alongside our headline projection so that journalists, political stakeholders, and members of the public can interrogate our methodology in detail. We are putting on record what we measured, how we measured it, and where we ourselves believe the inference is fragile.

1. Headline Projection (Recap)

The all-India election ecosystem expects exit polls to publish numbers, not narratives. We do both — but we lead with the numbers, and the numbers for West Bengal 2026 are below.

Editorial reading:  AITC+ leads on vote share but BJP is projected to lead on seats — a classic FPTP compression where a marginally lower vote share converts into a higher seat count due to higher tactical efficiency in concentrated geographies. The seat bands overlap, which is why we describe the outcome as a hung-leaning verdict rather than a clear majority for either side. Final reconciliation will be done by the Election Commission of India.

2. Survey Design — A Five-Month Longitudinal Build

The 8,66,514-respondent footprint was not a single sweep. It was assembled across four deliberately staggered phases between 29 November 2025 and 29 April 2026, designed to capture how voter intent evolved as the campaign matured — from pre-candidature mood, through ticket distribution, through final-week consolidation, into post-vote confirmation.

Why this matters:  A pure exit-day poll captures emotion. A longitudinal panel captures trajectory. By the time we entered Phase 4, we already had a stable behavioural model for all 294 of the 294 ACs — the exit-day sample was used to confirm or correct that model, not to build it from scratch.

3. Mode of Data Collection

The 8,66,514 figure represents valid responses received through Interactive Voice Response (IVR) calls. Calls were placed against a pre-defined, pre-cleaned database of mobile numbers stratified by AC. Respondents could opt out at any stage; no telephonic interviewer was inserted into the loop.

• Zero in-person field surveyors — eliminates surveyor-induced bias, prompting effects, and the “preferred-answer” problem documented in WB rural belts.
• Pre-recorded audio in Bengali only — the mother tongue of the West Bengal electorate — standardises the question wording across every respondent and ensures cultural and linguistic authenticity at the point of response.
• Numbers were drawn from a stratified database mirroring AC-level demographic distribution, not from open-call lists.
• Only completed responses where the respondent navigated the full questionnaire are counted in the 8,66,514 sample. Partial, abandoned, or non-responsive calls are excluded from the analytical universe.

4. Voter-Base Scrubbing — The SIR-Aware Universe

Before any sampling weight was applied, the Research Cell scraped and reconciled the entire ECI voter list for West Bengal, including all Special Intensive Revision (SIR) deletions notified up to 28 February 2026. Each AC was then re-profiled along five demographic axes — pre-SIR and post-SIR — to isolate the impact of roll changes on the addressable electorate.

• Age cohort distribution at AC level — pre-SIR vs post-SIR.
• Gender composition — pre-SIR vs post-SIR.
• Settlement classification — rural, semi-rural, semi-urban, urban.
• Religion-cohort distribution — Hindu, Muslim, others — at AC level.
• Net additions and deletions caused by SIR, mapped seat-by-seat.

Net effect:  We are projecting against the actual ECI universe that voted on polling day, not against an outdated 2024 voter list. This is the single largest structural correction we have applied versus our LSE-2024 model.

5. Structural Bias Removal — Best, Worst, Optimum

Every survey instrument carries a latent tilt — and in IVR polling, the single largest source of latent tilt is prompt order. The party named first to a respondent has a measurable advantage over the party named second, third, or fifth. To neutralise this, our questionnaire is built on a multi-preference rotation design that captures every respondent’s full ranking and rotates which party is queried first across the sample pool.

The respondent-side architecture.  Every respondent is asked to rank five to six preferences in order — not a single forced binary choice. This produces a complete preference vector per individual: P1 (most preferred), P2, P3, P4, P5, P6 (least preferred / actively disliked). The full ranking is what feeds the analytical engine; we never throw away the lower preferences.

The sample-side rotation.  Across the response pool, the party that is queried at the first preference slot (P1) is rotated. For one set of respondents, BJP is the first party prompted at P1; for another equal set, AITC+ is the first party prompted at P1; for further sub-pools, Left Front+, INC, ISF, AJUP and Others each take their turn at P1. By the time the sample is closed, every major party has been the first party named to a statistically equivalent share of respondents.

This rotation is what allows us to construct a clean best-case / worst-case bracket for each party:

• Best case for Party A — the responses where Party A was the first party prompted (queried at P1, before any opponent had been named). This is the upper bound: the cleanest read of Party A’s natural support, uncontaminated by opponent priming.
• Worst case for Party A — the responses where Party A’s principal opponent was the first party prompted (opponent queried at P1, with Party A queried only later in the sequence). This is the lower bound: the read of Party A’s support after the opponent has already anchored the respondent’s frame.
• Optimum scenario — the convergence point between the two bounds, computed in a straight line and anchored against the five-factor demographic normalisation (see §6). This is the number we publish.

The architecture is deliberately symmetric — the same rotation logic is applied to both AITC+ and BJP, so neither alliance is advantaged by question order. Constituencies where the optimum-case Winning Margin exceeds 10% (WM > 10%) on a single demographic profile are tagged Confident. Anything below this threshold is escalated to MRP stratification (see §7) before a final call is made.

6. Five-Factor Normalisation

Before the analytical layer, the raw response set is normalised across five demographic axes to scrub over- and under-sampling. This is not post-stratification dressing — it is a hard recalibration applied at the AC unit, not the state aggregate.

7. Modelling Layer — Swing, ML & MRP

People’s Insight ran the West Bengal 2024 Lok Sabha exit poll, and that calibration set is now our internal benchmark. Voter swing was studied between LSE-2024 and the current cycle, decomposed by community and region, and cross-referenced against community-specific seat clusters and publicly available historical results from earlier electoral cycles. Crucially, our own 2024 forecast errors at AC level were treated as labelled training data for the machine-learning correction layer.

• Voter swing decomposition — LSE-2024 → WB 2026 swing, decomposed by community and region.
• Community-specific seat clusters — Matua belt, Junglemahal, North Bengal Hills, Border-Muslim seats, Kolkata urban — modelled separately.
• Historical-result conditioning — publicly available base rates per AC over the last three cycles, used as priors only.
• Self-correction layer — past People’s Insight forecast errors fed back into the ML pipeline as ground-truth labels.

Once the straight-line definitive number for each party is locked, MRP (Multilevel Regression with Post-stratification) is applied to translate fixed-point estimates into probability bands. The seat range published in the headline table is the union of MRP percentiles around the central estimate, not a guess range.

8. Operational Integrity — Zero Human Intervention

From IVR despatch through normalisation, ML correction and MRP stratification, the operational pipeline runs without manual override. Analysts can audit but cannot edit. This design choice is non-negotiable — it removes the single largest historical source of polling failure in India: emotional human bias at the aggregation stage.

• Automated despatch and response capture.
• Automated normalisation and weighting.
• Automated swing modelling and ML correction.
• Automated MRP roll-up to state-level seat bands.
• Human role limited to audit, sign-off and public communication.

9. What We Will Not Claim

Polling is a discipline of estimation, not of prophecy. We will not claim certainty where the data does not provide it. The following caveats are placed on record so that no downstream commentary can attribute to People’s Insight a confidence we have not stated.

• Our seat bands for AITC+ and BJP overlap. The honest reading is that West Bengal 2026 is leaning toward a hung verdict, with BJP marginally ahead on seat count and AITC+ marginally ahead on vote share.
• ACs where our optimum-case WM is between 0 and 5% remain genuinely contestable on counting day; the seat band reflects this uncertainty.
• Booth-level dynamics that depend on polling-day administration (re-poll, EVM challenges, localised disruptions) are not modelled.

10. Disclaimer

People’s Insight is a polling agency. Our role is to interpret public sentiment using statistical instruments, and we can be wrong. Actual results are declared by the Election Commission of India, and the ECI declaration is the only authoritative outcome of the West Bengal Legislative Assembly Election 2026.