What sample size does an Indian constituency survey actually need?

A vendor tells you 3,000 respondents and ±1.8% margin of error. Both may be true for the constituency as a whole. But you never make a decision at constituency level — you make it at booth cluster level, and 3,000 respondents spread across 250 booths is twelve interviews a booth. At that level your margin is ±28%, which is not analysis, it is astrology. Ask for the margin of error at the level of the smallest decision unit you will actually act on.

For an assembly seat of 2–2.5 lakh voters, we run 2,400–3,600 respondents grouped into 12–20 booth clusters, giving roughly ±5–7% per cluster and ±2% at seat level. For a Lok Sabha seat with 6–8 assembly segments, 8,000–25,000 with segment-level breakouts. Below 1,000 for an assembly seat, you are buying a mood reading.

An unstratified sample in India over-represents whoever is home and willing to talk at 11 AM on a weekday — which skews older, more female in some areas, and more available in others. Stratify by booth cluster, age band, gender and community against the electoral roll's own distribution, then weight the results back. If a vendor cannot show you their weighting frame, they do not have one.

One survey is a photograph, and campaigns over-interpret photographs. What actually changes decisions is the delta: a mid-campaign tracker at T-60 with the same instrument and the same clusters, and a rolling three-day tracker in the final fortnight. A smaller tracker sample measured consistently beats a larger one measured once.

Insist on GPS stamps and audio on a random 15% of interviews, and run agent-level distribution checks. Fabricated data has a signature: too few refusals, response distributions that are suspiciously smooth, and interviews clustered impossibly close in time. Every large field operation in India has some fabrication. The question is whether the agency looks for it.