A survey systematically collects standardised information from a sample of people, using the same questions in the same way, so answers can be counted and compared. Its power is standardisation; its discipline is design.

Before commissioning a survey, check whether the answer is already in secondary data — the Census, NFHS, PLFS/NSS, or your own HMIS, UDISE+ and MGNREGA records. Re-collecting wastes budget and, worse, wastes respondents' time.

The first question is never 'what should we ask?' but 'what decision will this survey inform?' Work backwards from the decision to the findings you need, then to the indicators, then to the questions.

Sketch your dummy tables — the exact tables and charts you will produce — before collecting a single response. If a planned table needs a variable no question captures, you found the gap in time to fix it.

Pretesting sends you back to rewrite questions. Piloting sends you back to fix routing and sample. Analysis reveals what the next round should ask. A survey programme is a cycle of learning, not a single shot.

Most things surveys care about — empowerment, food security, trust, wellbeing — are concepts, not facts you can simply request. The questionnaire is the bridge from an abstract concept to a concrete, observable answer.

‘Is the household poor?’ only becomes measurable once you fix the definition — income below a line? deprived on a set of MPI indicators? — and write the exact questions that test it.

There is no underlying 'true survey' that a questionnaire merely records. The wording, order, options and translation are the instrument — change them and you change what you measure. Two questionnaires on the 'same' topic can yield different numbers, both correct for their wording.

Groves' Total Survey Error (TSE) framework organises every source of error in a survey into two families: errors of representation (the wrong people) and errors of measurement (the wrong answers). Good design manages both at once.

TSE's practical lesson: a fixed budget forces trade-offs. Spending everything on a huge sample (less sampling error) may starve enumerator training (more response error). Total error is what matters — not any one component.

A good survey question is understood the same way by every respondent and the question-writer. Three tests cover most failures: is it clear, is it specific, does it ask one thing?

Response options must cover every possible answer (exhaustive) and never overlap (mutually exclusive). A respondent should always find exactly one box that fits.

Example: Strongly disagree · Disagree · Neither · Agree · Strongly agree. Keep the steps balanced, label every point (not just the ends), and use the same scale throughout a block.

A genuine ‘Don't know’ or ‘Refused’ option prevents respondents from guessing or enumerators from inventing answers. But offered too readily, it becomes an easy escape that hollows out your data.

A filter (gateway) question routes each respondent to only the relevant items via skip logic — so a man is never asked about his last pregnancy, and a non-farmer skips the whole farming module.

On paper, complex skips invite enumerator error — missed branches, contradictory entries. On CAPI (computer-assisted personal interviewing), the device enforces routing automatically and can run real-time consistency checks.

Order is not cosmetic — it shapes answers and completion. A well-built instrument moves the respondent through a deliberate arc: welcome, easy questions, the core, the sensitive, then the close.

Asking about local crime, then overall life satisfaction, drags satisfaction down — the first question primes the second. Order effects are real and measurable.

The first questions should be easy, non-threatening and obviously relevant — they build rapport and signal that the survey is safe. Never open with income, caste or a long grid; you'll lose people at the door.

You rarely need to ask everyone. A well-drawn sample of a few thousand can describe a population of millions — the principle behind NFHS, PLFS and every credible poll. The magic is not size, it is representativeness.

If you need reliable estimates for a small subgroup — say, a tribal block or a minority community — a proportional sample may yield too few of them. Stratify and deliberately over-sample that group, then correct with weights.

When groups are over-sampled by design or respond at different rates, surveys apply weights so each respondent represents the right number of real people. Unit data from NFHS or PLFS gives wrong totals if you ignore the weights.

The mode — face-to-face, phone, web, self-completion — is not a neutral pipe. It shapes who you reach, how long they'll stay, how honest they are, and what you can ask. Choose it as carefully as the questions.

Most rigorous development surveys in South Asia still rely on face-to-face interviews, increasingly via CAPI — computer-assisted personal interviewing — on a tablet or phone running KoboToolbox, ODK or SurveyCTO.

Phone surveys and IVR (interactive voice response, automated calls) are cheap, fast and safe in a crisis — widely used during COVID-19. But they must be short, and they systematically miss people without phones.

Web surveys cost almost nothing to send to thousands and allow rich self-completion of sensitive items. But for general populations in South Asia, internet access is far from universal — coverage error is severe.

Mixed-mode designs (web first, phone follow-up, face-to-face for refusers) can boost response and cut cost. But they blend the modes' different biases, so disentangling real change from mode effects gets harder.

A survey in Bihar may need Hindi, Maithili, Bhojpuri and Urdu; one across India, a dozen languages and scripts. The instrument respondents actually hear is the translated one — so translation is instrument design, not an afterthought.

A literally correct translation can still measure something different. The aim is conceptual equivalence — the translated question evokes the same concept, with the same difficulty and connotations, as the original.

Back-translation catches literal errors but can miss awkwardness and cultural misfit. A committee or team approach — bilingual experts, translators and field staff reviewing together — produces more natural, usable wording.

‘In the last 12 months’ is abstract; ‘since last Diwali’ or ‘since the kharif harvest’ is vivid. A local events calendar built into the instrument sharpens recall and aligns reference periods across respondents.

The discipline is to adapt the wording, examples and units to each setting while holding the concept and the indicator identical. That is what keeps a multilingual survey both locally meaningful and nationally comparable.

Between ‘finished’ questionnaire and real fieldwork sit two distinct checks: pretesting (do the questions work?) and piloting (does the whole operation work?). Skipping them is the most expensive false economy in survey work.

However good the instrument, it is administered by enumerators. Standardised training — on the questions, the skips, neutral probing, consent and edge cases — is what makes 'the same question asked the same way' actually true in the field.

Supervisors re-contact a sample of respondents (back-checks) and observe live interviews (spot-checks) to confirm the interview happened, key answers match, and protocol was followed — catching error and fabrication early.

Non-response — refusals, not-at-homes, ineligibles — bites hardest when those who don't answer differ from those who do. The poorest, the busiest and the most private are often the hardest to reach, biasing the result.

In a survey, each row is someone who gave you their time and trust — often a poor, busy, over-surveyed person with little power over how their answers are used. Survey ethics is respect made operational, not a compliance form.

The Digital Personal Data Protection Act, 2023 sets duties for anyone handling personal digital data in India — including NGOs and researchers running CAPI surveys on tablets and phones.

Removing names is not enough. Village + age + caste + occupation can re-identify one person in a small area. Survey data needs deliberate de-identification, encryption and access control — especially with GPS coordinates attached.

A good instrument hands the analyst clean, well-named data. Build validation into the CAPI form, keep the raw export untouched, and ship a data dictionary so every variable, code and unit is self-explanatory months later.