Sample Size Determination

export const frontmatter = { title: "Sample Size Determination", description: "Comprehensive guide to determining appropriate sample sizes for research including formulas, software, and practical guidance", keywords: ["sample size", "power analysis", "research design", "statistical power", "sampling"] };

Sample size determination is one of the most critical—yet frequently mishandled—aspects of research design. Too few participants and your study lacks statistical power to detect real effects; too many wastes resources and may raise ethical concerns about unnecessary participant burden. The right sample size balances statistical requirements with practical constraints.

Why Sample Size Matters

Underpowered studies (too small):

• Cannot detect effects that genuinely exist (Type II error)
• Produce unstable, unreplicable results
• Waste participants' time and trust on studies that cannot yield conclusions
• Are ethically questionable (collecting data you cannot meaningfully analyze)

Overpowered studies (too large):

• Detect trivially small effects as "significant" (statistical but not practical significance)
• Waste resources that could fund other research
• May unnecessarily expose additional participants to research procedures
• Often unnecessary for achieving research objectives

Factors Affecting Required Sample Size

1. Desired Statistical Power (1 - β)

The probability of detecting a real effect when it exists. Convention: 0.80 (80% chance of finding a real effect).

Increasing power from 0.80 to 0.95 substantially increases required n—only do this for high-stakes research.

2. Significance Level (α)

The acceptable probability of a false positive. Convention: 0.05.

Using α = 0.01 (more stringent) requires larger samples to maintain the same power.

3. Expected Effect Size

The magnitude of the effect you expect to find. This is the MOST important and most difficult parameter to estimate.

Cohen's d (for t-tests): Small = 0.2, Medium = 0.5, Large = 0.8 Cohen's f² (for regression): Small = 0.02, Medium = 0.15, Large = 0.35 Pearson's r (for correlations): Small = 0.10, Medium = 0.30, Large = 0.50

Where to get effect size estimates:

• Previous studies on the same topic (most reliable)
• Pilot study data
• Meta-analyses in your field
• Theoretical reasoning about expected magnitude
• When completely unknown, use a medium effect as default

4. Number of Groups/Predictors

More groups (in ANOVA) or more predictors (in regression) require larger total samples.

5. Measurement Reliability

Instruments with lower reliability require larger samples because measurement error increases random variability.

Sample Size Formulas

For Comparing Two Group Means (Independent t-test)

Formula: n per group = 2 × [(Zα/2 + Zβ) / d]²

Where:

• Zα/2 = 1.96 (for α = 0.05, two-tailed)
• Zβ = 0.84 (for power = 0.80)
• d = expected Cohen's d

Quick reference table:

For Estimating a Population Mean

Formula: n = (Z × σ / E)²

Where E = desired margin of error, σ = population SD.

For Estimating a Population Proportion

Formula: n = Z² × p(1-p) / E²

For maximum (most conservative): use p = 0.5

• With E = 0.05 (±5%), α = 0.05: n = 384
• With E = 0.03 (±3%), α = 0.05: n = 1,068

For Multiple Regression

Rule of thumb (minimum): n ≥ 50 + 8k (where k = number of predictors)

• 5 predictors: n ≥ 90
• 10 predictors: n ≥ 130

Better approach (power analysis): Use G*Power software

• For R² = 0.15, 5 predictors, power = 0.80: n = 92
• For R² = 0.10, 8 predictors, power = 0.80: n = 160

For Chi-Square Tests

Minimum expected cell frequency should be ≥ 5. For a 2×2 table with moderate effect:

• Small effect (w=0.1): n = 785
• Medium effect (w=0.3): n = 88
• Large effect (w=0.5): n = 32

For Structural Equation Modeling (SEM)

Rules of thumb:

• Minimum: 200 (regardless of model complexity)
• Preferred: 10-20 cases per estimated parameter
• Complex models: 300-500+

Using G*Power Software

G*Power is the standard free tool for sample size calculations:

1. Select test family (t-tests, F-tests, χ², correlation)
2. Select specific test (independent t-test, one-way ANOVA, linear regression)
3. Select type of power analysis (A priori = determine n)
4. Input parameters: effect size, α, power, number of groups/predictors
5. Calculate: Software reports required total sample size

**Example in G*Power:**

• Test: Linear multiple regression, R² deviation from zero
• Effect size f² = 0.15 (medium)
• α = 0.05
• Power = 0.80
• Number of predictors = 6
• Result: n = 98

Adjustments to Calculated Sample Size

For Expected Non-Response

If you expect 30% non-response: Adjusted n = Required n / (1 - expected non-response) Example: Need 200, expect 30% non-response: 200 / 0.70 = 286 invitations needed

For Cluster Sampling (Design Effect)

Multiply by design effect: n_adjusted = n × DEFF Where DEFF = 1 + (cluster size - 1) × ICC

For Finite Population

If population is small: n_adjusted = n / (1 + (n-1)/N) Example: Need 384 but population is only 2,000: 384 / 1.192 = 322

Common Sample Size Mistakes

1. Using rules of thumb without justification — "I used 200 because that's a common number" is not acceptable
2. Not reporting power analysis in the methodology — Every quantitative study should document how n was determined
3. Choosing effect size after seeing results — Effect size for power analysis must be specified BEFORE data collection
4. Ignoring attrition/non-response — Always inflate your target to account for data loss
5. Confusing statistical significance with adequate power — A significant result from an underpowered study may not replicate

Qualitative Sample Size

Qualitative research uses different criteria:

• Data saturation: Continue until no new themes emerge
• Information power: Fewer participants needed when the study aim is narrow, participants have specific experience, theory guides analysis, dialogue quality is strong, and analysis strategy is targeted

Typical ranges:

Conclusion

Sample size determination is not a formality—it is a critical design decision that directly affects whether your study can answer its research questions. Always conduct a formal power analysis, document it in your methodology, account for expected attrition, and justify your chosen parameters (especially effect size). A well-justified sample size demonstrates methodological competence and protects against producing inconclusive results.