Sample Size Calculation: How Many Survey Responses Do You Really Need?

Sample Size Calculation: How Many Survey Responses Do You Really Need?

Sample size calculation determines how many responses a survey needs to produce reliable results, based on three inputs: confidence level, margin of error, and expected proportion. For the standard benchmark, 95% confidence and a ±5% margin of error, the answer is 385 responses, and counterintuitively, that number works whether your population is a city or a country.

Quick Answer

Sample size in 20 seconds:

• The benchmark: 385 responses = 95% confidence, ±5% margin of error
• The formula (Cochran's): n = Z² × p × (1-p) ÷ e²
• The three inputs: Confidence level (how sure), margin of error (how precise), expected proportion (use 0.5 when unsure)
• The counterintuitive truth: Population size barely matters above ~20,000: 385 works for Bengaluru and for India
• The rule most guides miss: Size the smallest subgroup you need to read (100+ per cell), not the total: subgroups, not totals, drive real sample sizes

Introduction

"How many responses do we need?" is the most asked and worst answered question in research. The textbook answer buries you in Greek letters; the vendor answer is suspiciously often "whatever your budget allows"; and the office answer, "a thousand sounds credible", is how studies end up simultaneously overpowered on the total and useless on every subgroup that mattered.

This guide gives you the real answer in three layers: the formula and a ready-reference table you can use in thirty seconds, the intuitions behind it (including why population size barely matters), and the practitioner layer competitors skip: sizing for subgroups, correcting for response rates, and knowing when the entire margin-of-error framework does and does not legitimately apply.

Why This Topic Matters for Brands

Sample size is where research budgets are silently won and lost:

• Undersized studies decide nothing: A finding at ±9% precision cannot separate 42% from 51%: which is usually the exact difference the decision needed
• Oversized studies burn the budget: Moving from ±3% to ±2% precision costs more than double the sample (1,068 → 2,401 responses) for precision most decisions never use: the diminishing-returns curve is brutal and rarely shown
• The subgroup trap catches almost everyone: A 500-response study reads beautifully on the total and collapses the moment someone asks "and what do Tier-2 women under 30 think?": 27 respondents, and a confident chart built on them
• Credibility theatre is expensive: Big samples with bad selection are precision applied to bias: n = 10,000 from a skewed source is worse than n = 400 done right, and the selection half of that equation lives in probability vs non probability sampling
• It is a negotiation skill: Knowing that ±5% needs 385, not 1,000, is the difference between funding one study and funding two

What Is Sample Size Calculation?

Sample size calculation is the process of determining the minimum number of responses needed for survey results to reflect the target population within a chosen precision and confidence. Three inputs drive it:

• Confidence level: How sure you want to be that the true value falls inside your range. 95% is the research standard; 90% is acceptable for directional business decisions; 99% for high-stakes claims
• Margin of error (e): The precision window around your result. ±5% means a 46% finding represents a true value between 41% and 51%
• Expected proportion (p): Your best guess of the result. When unknown, use 0.5: it is the mathematically conservative choice that produces the largest required sample

And one input that famously does NOT drive it: population size. Once a population exceeds roughly 20,000, required sample size flattens: the same 385 responses serve a mid-size city and a nation of a billion. Precision comes from the absolute number of responses, not the fraction of the population sampled: the single most counterintuitive and useful fact in survey math.

The Formula and the Ready-Reference Table

Cochran's Formula

n = Z² × p × (1-p) ÷ e²

Where Z is the confidence level's z-score (90% = 1.645, 95% = 1.96, 99% = 2.576), p is the expected proportion, and e is the margin of error as a decimal.

The benchmark worked through: 95% confidence, p = 0.5, e = 0.05: n = (1.96² × 0.5 × 0.5) ÷ 0.05² = 0.9604 ÷ 0.0025 = 385 responses

The Ready-Reference Table

The Small-Population Correction

For populations under ~20,000, apply the finite population correction: n_adjusted = n ÷ (1 + (n - 1) ÷ N). Example: surveying a customer base of 1,000 at the ±5%/95% benchmark needs only n = 385 ÷ (1 + 384/1,000) ≈ 278 responses, not 385. Small populations are the one place population size genuinely helps.

The Practitioner Layer: What the Formula Doesn't Tell You

Rule 1: Size the Cells, Not the Total

The formula sizes your headline number; decisions usually live in the breakdowns. The working rule: every subgroup you need to read independently needs 100+ respondents (30 is the absolute statistical floor; treat sub-100 cells as directional only).

• Need to compare 4 age bands × 2 genders? That is 8 cells × 100 = a real minimum of 800, structured by quota: not 385
• This is why real studies are sized backwards: list the comparisons the decision needs, size the smallest cell, and let the total follow

Rule 2: Budget for the Response Funnel

Required completes ≠ required invitations. Invitations = n ÷ (response rate × qualification rate): 385 completes at a 20% response rate and 50% qualification means ~3,850 invites. Panels with pre-verified, behaviourally qualified members compress this funnel dramatically: which is a sampling-cost lever most buyers never inspect.

Rule 3: Know When the Math Legitimately Applies

Margins of error formally belong to probability samples. For the quota-designed non-probability samples that dominate commercial research, the same n-sizes remain the right planning discipline, but report precision as indicative, not as formal confidence intervals: the vocabulary rule covered on the probability vs non probability sampling page applies to every number this page produces.

Rule 4: Qualitative Sizes by Saturation, Not Formulas

Interviews and focus groups size differently: 12-20 in-depth interviews per distinct segment typically reaches saturation (the point where new interviews stop producing new themes). Applying survey math to qualitative work, or qualitative sizes to survey claims, are twin errors: the method families in consumer behaviour research methods each carry their own sizing logic.

Worked Examples

• The concept test: Two concepts, national read plus metro/non-metro split. Smallest cell to read: non-metro, per concept. 100+ per cell × 2 cells × 2 concepts = 400+ completes, quota-structured: the subgroup rule, not the formula, set the number
• The tracker wave: Brand consideration tracked at ±4%/95%: 601 per wave, held constant across waves: consistency across waves matters more than any single wave's precision
• The customer-base pulse: 2,000 active customers, ±5%/95%: finite correction gives ~323 needed, and at a 30% response rate that means ~1,080 invites: the whole calculation in two lines
• The A/B price test: Detecting a 7-point conversion difference between price variants needs ~700 per variant: experiments size on detectable difference, not margin of error: when the stakes justify it, run the power calculation or use a platform that does
• The false-economy example: A team trims 385 to 200 to save budget: the margin widens to ±7%, and the 44% vs 51% concept race the study existed to call becomes statistically unreadable: the savings cost the answer

PulseAI Research Insight: Where Sample Size Meets Sample Depth

The formula tells you 385; the subgroup rule tells you why real studies need structured thousands; and both collide with the commercial constraint: filling deep cells, fast, with genuine category buyers.

That collision is what network scale solves. PulseAI Research fields from Smytten's network of 30M+ active Indian consumers, and the sizing implications show up directly in the work:

• Cells fill instead of collapsing: Tier-2/3 splits, age bands within categories, verified buyers of specific segments: the cells that turn a 27-respondent guess into a 100+ read: fillable at speed because the base is 30 million, not thirty thousand
• Qualification is behavioural, not claimed: The response-funnel math improves when eligibility is verified from real trial and purchase behaviour rather than screener claims: fewer invites wasted, cleaner cells
• Subgroup reads become findings: The Mattress? More Like "Mat-Stress" report's most decision-relevant numbers are cell-level reads: 89% early replacement among regular pain sufferers, channel-specific priorities (offline buyers weighting comfort at 77.5%), spend-tier feature demands: each one a subgroup sized well enough to trust
• Speed holds at depth: Research-grade output in 72 hours at full quota structure: the sizing discipline this page teaches, executed on sprint timelines

The sizing lesson in one line: the formula is about the total, but the value is in the cells, and cells are a scale problem.

How Brands Can Use This

1. Start from the decision's comparisons. List every split the readout must support, apply the 100-per-cell rule to the smallest, and let the total emerge: sizing backwards is the whole skill
2. Default to ±4-5% at 95%. Buy more precision only when the decision genuinely turns on small differences: check the reference table's cost curve before agreeing to ±2%
3. Do the invite math before the timeline promise. Completes ÷ (response × qualification rate) = invitations: the number that actually determines whether Friday is feasible
4. Protect cells during fieldwork. Monitor quota fill daily: a study that hits its total while a key cell sits at 40 respondents has failed quietly: this is where survey questions discipline meets fieldwork discipline
5. Report precision honestly. Formal margins for probability samples, indicative precision for quota samples, and never a margin of error on a self-selected poll: the vocabulary rule, applied to sizing
6. Feed sized data into structured analysis. Well-sized cells are what make the segment-level consumer behaviour analysis ladder trustworthy: descriptive to prescriptive, per subgroup, without the asterisks

Related Concepts

• Probability vs non probability sampling: The WHO half of sample design: selection method and what your numbers can claim
• Survey questions: The instrument half: what well-sized samples actually answer
• Consumer behaviour research methods: The method families, each with its own sizing logic
• Agile research sprint: Where sizing discipline meets the 5-day clock

FAQs

1.How do you calculate sample size for a survey?

Use Cochran's formula: n = Z² × p × (1-p) ÷ e², where Z is the z-score for your confidence level (1.96 for 95%), p is the expected proportion (0.5 when unknown), and e is the margin of error as a decimal. At 95% confidence and ±5% margin of error, this gives 385 responses.

2.What is a good sample size for market research?

For most commercial studies, 385 to 600 responses covers a headline read at 95% confidence with ±4-5% precision. The real driver is subgroups: every segment you need to read independently needs 100+ respondents, so studies with multiple splits typically require structured samples of 800 to 1,500.

3.Does population size affect sample size?

Barely, once the population exceeds roughly 20,000: the same 385 responses serve a city and a country, because precision depends on the absolute number of responses, not the fraction of the population sampled. For small populations (under ~20,000), the finite population correction reduces the requirement: a customer base of 1,000 needs only about 278 responses for a ±5% read.

4.What is the sample size formula?

Cochran's formula: n = Z² × p × (1-p) ÷ e². For small populations, adjust with the finite population correction: n_adjusted = n ÷ (1 + (n - 1) ÷ N), where N is the population size.

5.Is 100 a good sample size?

For a headline finding, 100 responses gives roughly a ±10% margin of error at 95% confidence: usable for directional reads, too imprecise for close calls. As a subgroup size, 100 is the working minimum for reading a cell independently, which is why the number matters more per segment than in total.

6.Why is 385 the standard sample size?

Because it is what Cochran's formula produces at the standard settings: 95% confidence, ±5% margin of error, and the conservative p = 0.5. It is the benchmark for a single-population read; studies needing subgroup comparisons or tighter precision size upward from there.

7.What sample size do you need for qualitative research?

Qualitative studies size by saturation rather than formulas: typically 12 to 20 in-depth interviews per distinct segment, the point where additional interviews stop surfacing new themes. Survey math does not apply to interview-based work, and qualitative sizes cannot support survey-style percentage claims.

8.How many people should I survey from a population of 1,000?

About 278 for a ±5% margin of error at 95% confidence, using the finite population correction (385 adjusted for N = 1,000). At a typical 30% response rate, plan roughly 930 invitations to reach that count.