Research design · Sample size

How to calculate sample size for quantitative research: how many respondents are enough?

Your committee asks “what is this sample size based on?” and you are not sure. Here are the most-cited formulas, when to use each one, and how to write it up in your thesis.

⏱ ~7 min readSources: Hair et al. · Tabachnick & Fidell · YamaneUpdated:

You have a model and a questionnaire. Now, how many responses do you actually need? Too few and your EFA falls apart, your coefficients wobble and real effects fail to reach significance. Too many and you waste weeks. And your committee will almost certainly ask: “What is this sample size based on?” This guide covers the most-cited sample size formulas, when to use each one, worked examples and a paragraph you can adapt for your thesis.

Table of sample size formulas for EFA, regression, SEM, PLS-SEM, Yamane and Cochran with examples
Sample size formulas by method, with worked examples.

1Why sample size decides your results

Sample size is not just a box to tick in your methods chapter. It determines whether your tests can detect a real relationship in the data, known as statistical power. When a sample is too small, three problems tend to arrive together:

  • EFA cannot recover the structure: KMO is low and items scatter across factors.
  • Coefficients are unstable: a handful of different responses changes the size or even the sign of β.
  • Hypotheses miss significance even when the relationship is real, because standard errors are too large.
🎯 The golden rule

Size your sample from the analysis you will run. A study using both EFA and regression has to satisfy both, so take the larger number.

2Which sample size formula fits your study?

Answer two questions: do you know the size of the population, and what is your main analysis?

SituationFormula to use
Running EFA in SPSSHair et al.: n ≥ 5 × observed items
Multiple linear regressionTabachnick & Fidell: n ≥ 50 + 8m
CFA or SEM in AMOSBy model complexity (Hair et al., 2019)
PLS-SEM in SmartPLSInverse square root (Kock & Hadaya, 2018)
Known population N (say, 2,000 students in a faculty)Yamane / Slovin
Very large or unknown populationCochran

3Sample size for exploratory factor analysis

According to Hair, Black, Babin & Anderson (2019), factor analysis needs at least 50 observations, preferably 100 or more, with a minimum ratio of 5 observations per variable and preferably 10.

Example

A model with 6 constructs of 4 items each has 24 items.

  • 5:1 ratio → n ≥ 24 × 5 = 120
  • 10:1 ratio → n ≥ 24 × 10 = 240
✦ Tip

Count every item going into the EFA, including the dependent variable if you run them together. Demographic questions do not count.

📊
Read next
Exploratory factor analysis in SPSS
KMO, Bartlett’s test, variance explained and cleaning up the rotated matrix.
→

4Sample size for multiple regression

Tabachnick & Fidell (2013) give two formulas, where m is the number of predictors:

  • n ≥ 50 + 8m to test the overall model (R²).
  • n ≥ 104 + m to test individual coefficients.
Example

With 5 predictors: 50 + 8 × 5 = 90 and 104 + 5 = 109, so aim for at least 109 responses.

5Sample size for CFA and SEM in AMOS

CB-SEM (AMOS) is hungrier for data than regression because it estimates many parameters at once. Hair et al. (2019) tie the minimum to the complexity of the model:

Model characteristicsMinimum sample
Up to 5 constructs, more than 3 items each, high communalities (≥ 0.6)100
Up to 7 constructs, modest communalities (around 0.5)150
Up to 7 constructs, low communalities (< 0.45) or constructs under 3 items300
Many constructs with low communalities500

In practice, SEM studies usually aim for 200 responses or more so that fit indices behave.

6Sample size for PLS-SEM and SmartPLS

PLS-SEM copes better with smaller samples, but flexible does not mean anything goes.

The 10-times rule

Sample size ≥ 10 × the largest number of arrows pointing at any one construct (Barclay et al., 1995). It is easy to apply, but Hair et al. (2022) warn that it tends to underestimate the sample you need.

The inverse square root method (recommended)

From Kock & Hadaya (2018), recommended by Hair et al. (2022), at the 5% significance level:

📐 Formula

nmin > (2.486 / |βmin|)², where βmin is the smallest path coefficient you expect to be significant, taken from earlier studies.

Example
  • βmin = 0.2 → (2.486 / 0.2)² = 154.5 → 155 responses
  • βmin = 0.15 → (2.486 / 0.15)² = 274.7 → 275 responses

You can also run a power analysis in G*Power for multiple regression, using the largest number of predictors pointing at one construct, power of 0.80 and an effect size of f² = 0.15.

7Known population: the Yamane (Slovin) formula

When you can count the population (every student in a faculty, every employee of a company), use the Yamane (1967) formula, which some sources call the Slovin formula:

📐 Formula

n = N / (1 + N × e²), where N is the population size and e the margin of error (usually 0.05).

Example

For 2,000 students at e = 0.05: n = 2,000 / (1 + 2,000 × 0.0025) = 2,000 / 6 = 333.3 → 334 students.

8Unknown population: the Cochran formula

For a very large or uncountable population (say, “online shoppers in a city”), use the Cochran (1977) formula:

📐 Formula

n = Z² × p × (1 − p) / e², with Z = 1.96 (95% confidence), p = 0.5 (the safest assumption) and e = 0.05.

Example

n = 1.96² × 0.5 × 0.5 / 0.05² = 384.16 → 385 responses.

9Settling on a number and writing it up

Three steps:

  1. Calculate for each method you will use.
  2. Take the largest number as your minimum.
  3. Add a 10–20% reserve for responses you will discard during cleaning (incomplete, screened out, straight-lined).
Worked example

5 independent constructs plus 1 dependent, 22 items in total, running EFA and regression:

  • EFA (10:1): 22 × 10 = 220
  • Regression: max(50 + 8 × 5; 104 + 5) = 109
  • → Minimum sample of 220; with a 15% reserve, send out about 255 questionnaires.
Bar chart: EFA needs 220, regression needs 109, so the minimum is 220 and about 255 questionnaires go out
Take the largest number, then add 10–20% for responses you will discard.
✍️ A paragraph you can adapt for your methods chapter

“According to Hair et al. (2019), the sample for exploratory factor analysis should have at least 5 observations per measured variable, and preferably 10. With 22 observed items, the minimum sample is 220. Tabachnick and Fidell (2013) indicate that multiple regression with 5 predictors requires at least 109 responses. To allow for invalid responses, 255 questionnaires were distributed.”

⚠️ Reach your target with real data

If you are short of responses, widen your channels or collect for longer. Fabricating or padding data is a breach of research integrity and can get a thesis invalidated.

Get ready faster with the right tools

🎯 In short
  • EFA: n ≥ 5–10 × observed items · Regression: n ≥ 50 + 8m and 104 + m.
  • AMOS: by model complexity, usually ≥ 200 · SmartPLS: (2.486 / βmin)².
  • Known population → Yamane; unknown → Cochran (about 385).
  • Take the largest number, add 10–20% and state your reasoning in the thesis.

Frequently asked questions

What is the minimum sample size for a thesis?

There is no single number. Work it out for each analysis you plan to run (EFA, regression, SEM or PLS-SEM), take the largest result, then add a reserve for responses you will have to discard.

Are the Slovin and Yamane formulas different?

They are the same formula, n = N / (1 + N·e²). Some sources credit Yamane (1967), others Slovin. Use it only when you know the population size N.

How do I calculate sample size when the population is unknown?

Use Cochran’s formula with p = 0.5, 95% confidence and a 5% margin of error, which gives about 385 responses. Or size the sample from the number of items your analysis requires.

Should I collect more responses than the calculated sample size?

Yes. Incomplete, screened-out and straight-lined responses get removed during cleaning, so collect about 10–20% extra to still have enough afterwards.

How do I calculate sample size for SmartPLS?

The 10-times rule is easy but widely criticised as too lenient. Hair et al. (2022) recommend the inverse square root method of Kock & Hadaya (2018) or a power analysis in G*Power.