Simple Random Sampling in India: How to Draw It, Cost It and Defend It

The definition is precise and unforgiving: every possible sample of size n has an equal chance of being the one you get. That is stronger than the loose version people usually have in mind, which is merely 'every individual has an equal chance'. The stronger property is what makes independence hold, and independence is what makes the variance formulas valid. It implies you must be able to reach every unit on the list, and that no unit's inclusion changes another's odds — which rules out most convenient shortcuts, including grabbing an entire office, an entire housing society, or an entire WhatsApp group.

Sampling without replacement is the norm, and it is why the finite population correction exists. Once a person is selected, you do not put them back in the pot — nobody wants the same respondent twice. That makes draws very slightly dependent, and the correction term accounts for it. In practice the effect is negligible unless your sample is more than about 5% of the population, so a 2,000-person national study ignores it entirely, while a 300-person draw from a 1,500-employee company absolutely should not. Sampling with replacement mostly appears in bootstrap resampling during analysis rather than in field design.

Random does not mean haphazard, and this is the mistake that quietly destroys studies. Standing at a metro gate and 'randomly' picking people who look approachable is not randomisation — it is convenience sampling with a randomness costume, and human selection instinct reliably favours the young, the unhurried and the friendly. Genuine randomisation requires a mechanical device: a pseudo-random number generator, a lottery draw, a random start on an ordered list. If a human eye chose the respondent, the design is non-probability regardless of intent. The taxonomy is laid out on sampling methods India.

Simple random sampling is unbiased in expectation, which is not the same as accurate in your one sample. Averaged over infinitely many draws, the estimator lands on the truth. Your single actual sample can still come out with 68% women, or three respondents from the entire South zone, purely by chance — the design promises no balance on any particular variable. This is precisely the weakness that stratification was invented to remove, by forcing balance rather than hoping for it. See stratified random sampling India.

Step one is the frame, and the frame is the whole battle. You need a list where every unit appears exactly once, with no omissions and no duplicates. Panels, CRM databases, employee master lists, loan books, transaction logs, student rolls and app user tables can all qualify. Voter rolls and phone number ranges look tempting and disappoint quickly — stale entries, duplicate registrations, disconnected numbers and systematic omission of migrants make coverage error the dominant problem long before sampling error shows up. Every unit missing from the frame has a selection probability of zero, and no amount of randomisation repairs that.

Step two is deduplication, because a frame with duplicates is not a frame. If one person appears three times — three phone numbers, three email addresses, three panel accounts — their selection probability is triple everyone else's, and the sample is silently weighted toward the kind of person who accumulates multiple accounts. This is not a hypothetical problem in Indian online panels; it is the single most common defect. ZK verification solves it structurally by proving each identity is unique without exposing personal data, which is why the frame underneath matters as much as the draw. See survey data verification India.

Step three is the mechanical draw, and you should record the seed. Number the frame from 1 to N, generate n distinct pseudo-random integers in that range using a fixed seed, and select the matching rows. The seed is the unglamorous detail that makes the whole thing auditable: with the seed and the frame snapshot, anyone can regenerate the identical sample and verify you did not quietly re-roll until the numbers looked better. Reproducibility is the difference between a documented method and a story about a method.

Step four is handling non-response without breaking the design. If 30% of selected units never respond, substituting them with whoever is available converts your probability sample into a convenience sample at exactly the point it mattered. The disciplined options are to over-draw at the start using known incidence and response rates, or to draw formal replacement units by the same random mechanism from the same frame, and to document response rates by stratum. Bias from non-response does not shrink as n grows — see sampling bias India.

For a proportion, the standard error is the square root of p times one minus p, divided by n. Multiply by 1.96 for a 95% confidence interval. At p = 0.5, which is the most conservative case, that gives roughly ±4.9 percentage points at n = 400, ±3.1 at n = 1,000, ±2.2 at n = 2,000 and ±1.5 at n = 4,000. Two things follow immediately. Precision improves with the square root of n, so quadrupling the sample only halves the error. And extreme proportions are measured more precisely — an 8% incidence at n = 1,000 carries about ±1.7 points, not ±3.1.

For a mean, the standard error is the standard deviation divided by the square root of n. This is why pilots matter: you cannot size a sample for a continuous measure such as monthly category spend without an estimate of its variability, and Indian consumption variables are often wildly dispersed with long right tails. A category where spend ranges from ₹80 to ₹8,000 a month needs a substantially larger sample for the same absolute precision than one clustered between ₹200 and ₹400. Reporting medians and trimmed means alongside the average is usually more honest for such distributions.

The finite population correction multiplies your standard error by the square root of N minus n over N minus 1. For a national study it is irrelevant. For a 400-person draw from a 2,000-customer base it multiplies the error by about 0.89, a genuine 11% precision gain you are entitled to claim. For a 300-of-1,500 employee census it matters more. This is the one situation where the population size legitimately enters the calculation, and it always works in your favour — the formula library is on quantitative research methods India.

Every comparison you make consumes precision, and simple random sampling makes that arithmetic visible. The error on a difference between two subgroups is larger than the error on either one alone, so two cells of 200 each cannot reliably detect a five-point gap. Poseidon runs the appropriate test — two-proportion z, t-test, chi-square or ANOVA — checks assumptions, and applies multiple-comparison control when you crosstab aggressively, rather than letting you hunt until something turns green. See statistical significance testing India.

Run the sizing formula forwards rather than backwards, and the budget conversation gets much easier. To hit a target margin of error, required n is approximately 1.96 squared times p times one minus p, divided by the square of your target error expressed as a decimal. For ±3 points at the conservative p = 0.5, that is 3.84 × 0.25 ÷ 0.0009, which gives about 1,067 completes. Want ±2 points and the requirement jumps to roughly 2,401 — more than double the sample for one extra point of precision, which is exactly the trade-off worth putting in front of whoever approves the budget. If your key measure is a low-incidence behaviour at around 10%, the same ±3 points needs only about 384 completes, because p(1−p) is smaller. Sizing every study at the same round number ignores all of this. See survey methodology best practices.

The frame problem is structural, not solvable by effort. There is no maintained list of Indian consumers. Census enumeration blocks are not a contactable list of individuals. Voter rolls omit and duplicate. Mobile number ranges include disconnections and multi-SIM users. Any national frame you assemble will systematically under-cover the rural poor, migrant workers, women in low-smartphone-penetration districts and anyone off the formal grid — the exact groups whose behaviour often differs most. A perfect random draw from an incomplete frame is a perfect random draw from the wrong population.

Even with a good frame, chance imbalance is expensive in a heterogeneous market. Draw 1,000 people at random from a national panel and you might get 43 respondents from the entire East zone. That is not an error, it is variance behaving normally — but it means your regional read is unusable, and you discover this after fielding rather than before. In a market where Tier 1 and Tier 3 behaviour genuinely diverges on price sensitivity, pack size, language preference and channel, leaving composition to chance is a false economy. Stratification removes that risk by construction.

Dispersion drives cost when the mode is physical. A random national draw scatters respondents across hundreds of pincodes, which is fine for app-based fieldwork and ruinous for face-to-face work. This is exactly why household surveys use multistage cluster designs — the travel savings are enormous even after paying the design-effect penalty. If your study needs in-home observation or product placement, geography beats theory. See cluster sampling India.

Rare populations break the economics entirely. If your target is 2% of the population, a random draw of 1,000 yields about 20 qualifying respondents and 980 wasted incentives. Screening within a profiled frame, purposive sampling or referral chains are the rational answers, and each requires you to state the universe you actually sampled rather than implying the general public. Low-incidence recruitment on a profiled panel is covered on consumer panel India, with referral designs on snowball sampling India.

You start by defining the eligible pool, which is where the frame becomes concrete. From SuperJ's 20M+ verified users you apply your universe conditions — age band, gender, city tier, state, language, category behaviour — and the platform reports the eligible count before you spend anything. That number is the N in your formulas, and knowing it up front is what lets you compute incidence, feasibility and cost honestly rather than discovering mid-field that your universe is 40,000 people, not four million.

The draw itself is mechanical, seeded and logged. Hercules selects respondents by pseudo-random draw from the eligible pool with a recorded seed, so the identical sample can be regenerated for audit. Invitations go out in-app on SuperJ with real rewards attached, which is why completion sits at 60-90%+ rather than the low single digits typical of email panels — and high completion is what keeps non-response from becoming the dominant error term. Every response then passes attention, speeder, straight-lining and ZK-deduplication checks before entering the analysis set.

When a pure random draw is the wrong tool, the platform says so and offers the alternative. If your reporting plan needs a defensible Tier 3 or regional cut, you switch to stratified allocation with interlocking quotas rather than gambling on chance composition, using hard, soft or dynamic quota rules that rebalance as incidence data arrives. Cells are visible live during field, so a thin cell gets a booster on day one instead of an apology in the topline. Quota mechanics are documented on survey quota management India.

Analysis and documentation are automatic, and they show their working. Poseidon computes standard errors with the finite population correction where it applies, reports effective sample size, applies denominator correction so filtered-question bases are exact, and generates the methodology annexe — frame definition, eligible N, seed, achieved composition, exclusions, response rate. Results stream to a live dashboard while field is open. The full pipeline from brief to report is described on Hercules survey creation process India.