Quantitative Methodology & Biostatistics
How to Calculate Sample Size and Statistical Power Using G*Power
An authoritative step-by-step protocol for graduate researchers, dissertation committees, and journal authors on conducting a priori power analysis in G*Power 3.1: specifying Cohen’s effect sizes (d, f, f², w, r), applying participant attrition inflation, and drafting turnkey APA 7 methodology justifications.
6 Non-Negotiable Sample Size Checkpoints Before Data Collection
Before collecting empirical data or submitting your research methodology to an Institutional Review Board (IRB) or thesis defense committee, confirm that your sampling strategy satisfies these six foundational standards:
- Formulate your research hypothesis to determine the exact inferential test family (t-test, ANOVA, multiple regression, chi-square, or bivariate correlation)
- Establish nominal significance (alpha α = .05) and target statistical power (1 - β = .80 as the scientific floor, or .95 for high-stakes/clinical designs)
- Justify your effect size parameter using published meta-analyses, prior empirical literature, or Smallest Effect Size of Interest (SESOI) rather than arbitrary guessing
- Execute an A Priori power analysis in G*Power 3.1 prior to data collection to calculate the minimum clean analyzable sample size (N_clean)
- Apply the mathematical attrition inflation formula [N_recruitment = ⌈N_clean / (1 - d)⌉] to safeguard against survey abandonment, failed attention checks, and missing data
- Adhere strictly to APA 7th edition Chapter 3 (JARS-Quant) reporting standards, eliminating leading zeros for bounded metrics (p, r, α, β) and refuting post-hoc observed power
Master Parameter Decision Matrix for Common Research Designs
The table below summarizes the exact computational outputs generated by the G*Power 3.1.9.7 algorithm under standard scientific conditions (two-tailed α = .05, statistical power 1 - β = .80, balanced group allocation, and Cohen's conventional medium effect size benchmarks), alongside the adjusted recruitment target accounting for standard 15% participant attrition:
| Research Design & Test | G*Power Test Family & Specific Test | Effect Size Metric | Significance & Power | Clean Sample (N_clean) | Adjusted Recruitment Target (15% Attrition) |
|---|---|---|---|---|---|
| Independent Samples t-Test (Standard) N₂/N₁ = 1 (Two-tailed) | t tests → Means: Difference between two independent means | d = 0.50 (Medium) | α = .05, 1 - β = .80 | Total N = 128 (n₁ = 64, n₂ = 64) | N_rec = 151 (~76 per group) |
| Independent Samples t-Test (High Power) N₂/N₁ = 1 (Two-tailed) | t tests → Means: Difference between two independent means | d = 0.50 (Medium) | α = .05, 1 - β = .95 | Total N = 210 (n₁ = 105, n₂ = 105) | N_rec = 248 (~124 per group) |
| Paired / Dependent Samples t-Test Single cohort (Pre-test vs. Post-test) | t tests → Means: Difference between two dependent means (matched pairs) | d_z = 0.50 (Medium) | α = .05, 1 - β = .80 | Total N = 34 pairs (34 subjects) | N_rec = 40 pairs |
| One-Way ANOVA (3 Groups) Number of groups = 3 (df = 2) | F tests → ANOVA: Fixed effects, omnibus, one-way | f = 0.25 (Medium; η² ≈ .059) | α = .05, 1 - β = .80 | Total N = 159 (n = 53 per group) | N_rec = 188 (~63 per group) |
| One-Way ANOVA (4 Groups) Number of groups = 4 (df = 3) | F tests → ANOVA: Fixed effects, omnibus, one-way | f = 0.25 (Medium; η² ≈ .059) | α = .05, 1 - β = .80 | Total N = 180 (n = 45 per group) | N_rec = 212 (~53 per group) |
| Multiple Linear Regression (5 Predictors) Number of predictors = 5 | F tests → Linear multiple regression: Fixed model, R² dev from zero | f² = 0.15 (Medium; R² ≈ .130) | α = .05, 1 - β = .80 | Total N = 92 | N_rec = 109 |
| Chi-Square Contingency (2 × 2 Table) Degrees of freedom df = (2-1)(2-1) = 1 | χ² tests → Goodness-of-fit tests: Contingency tables | w = 0.30 (Medium) | α = .05, 1 - β = .80 | Total N = 88 | N_rec = 104 |
| Pearson Bivariate Correlation Two-tailed (Null ρ = .00) | Exact tests → Correlation: Bivariate normal model | r = 0.30 (Medium) | α = .05, 1 - β = .80 | Total N = 82 (or 84 asymptotic) | N_rec = 97 |
The 6-Stage A Priori Power Analysis Protocol
Hypothesis Formulation and Statistical Model Mapping
In graduate dissertations, empirical grant applications, and peer-reviewed journal manuscripts, calculating sample size is not a cosmetic administrative chore—it is an ontological requirement of statistical inference. An underpowered study wastes institutional resources, exposes participants to uninformative burdens, and frequently produces false-negative findings (Type II error) where genuine phenomena remain undetected. Conversely, an oversized sample consumes excessive funding and risks detecting trivial differences that possess statistical significance without clinical or practical importance.
The first operational stage requires translating your theoretical research questions into an exact inferential test family. G*Power organizes calculations into distinct families based on the underlying test distribution: t-tests (means, regressions slopes), F-tests (ANOVA, ANCOVA, multiple regression models), chi-square tests (goodness-of-fit, contingency tables), and exact tests (bivariate correlations, Fisher's exact). You must never open G*Power without first finalizing your primary independent variables, dependent variables, and observational structure (between-subjects versus within-subjects).
Execution Verification Checklist:
- Explicitly articulated whether the primary research hypothesis evaluates group mean differences, directional prediction, or bivariate association
- Mapped the target model to its foundational G*Power test family (t-tests, F-tests, χ²-tests, or Exact tests)
- Designated the operational dependent variable scale (continuous interval/ratio versus categorical counts)
Critical Methodological Note: Never use sample size formulas designed for population proportion surveys (e.g., Taro Yamane's formula or Cochran's 1977 formula) for experimental or regression studies. Survey formulas ignore effect sizes, group variance, and inferential models entirely.
Specifying Significance Level (α) and Statistical Power (1 - β)
Statistical inference navigates two fundamental error probabilities: Type I error (α), the probability of rejecting the null hypothesis when it is true (a false positive), and Type II error (β), the probability of failing to reject the null hypothesis when a genuine population effect exists (a false negative). Statistical power is mathematically defined as 1 - β: the probability of correctly rejecting a false null hypothesis.
By scientific convention established by Jacob Cohen (1988), the nominal significance threshold is set at α = .05 (a 5% false-positive risk), and nominal statistical power is set at 1 - β = .80 (an 80% detection probability, with β = .20). This implies a conventional 4:1 weighting between Type I and Type II errors. However, for confirmatory doctoral dissertations, multi-site clinical trials, or high-stakes intervention studies, setting 1 - β = .90 or .95 is strongly recommended by committee chairs and journal editors to minimize the risk of uninformative null outcomes.
Furthermore, researchers must select two-tailed tests unless an directional hypothesis is non-negotiable and strictly pre-registered. Opting for a one-tailed test merely to cut the required sample size by approximately 20% is widely regarded by methodological reviewers as questionable research practice (QRP).
Execution Verification Checklist:
- Specified nominal Type I error rate (α = .05; or Bonferroni-adjusted α if testing multiple primary endpoints)
- Set target statistical power to at least 1 - β = .80 (or .95 for confirmatory clinical or high-stakes experimental designs)
- Selected two-tailed testing directionality to safeguard against unanticipated reversal effects in empirical data
Critical Methodological Note: If your study tests three primary outcome variables simultaneously without an omnibus multivariate model, apply a Bonferroni correction: enter α = .05 / 3 = .0167 into G*Power.
Estimating Effect Sizes: SESOI, Prior Literature vs. Cohen Conventions
The effect size is the standardized magnitude of the experimental phenomenon or relationship in the population. Specifying the effect size represents the most challenging and scrutinized input in any power calculation. G*Power requires specific standardized metrics depending on the chosen design: Cohen's d for t-tests, Cohen's f for ANOVA, Cohen's f² for multiple regression, Cohen's w for chi-square tests, and Pearson's r for bivariate correlation.
Methodological literature (Lakens, 2022; APA JARS-Quant) establishes a strict hierarchy for justifying effect sizes: (1) Meta-analytic pooled effect sizes from systematic reviews; (2) Effect sizes computed from prior empirical literature or well-designed pilot studies using G*Power's built-in 'Determine =>' calculation drawer; (3) Smallest Effect Size of Interest (SESOI)—the minimum threshold that holds practical, educational, or clinical utility; and (4) Cohen's (1988) classical rule-of-thumb conventions (small, medium, large) as a recognized fallback when no empirical data exists.
Researchers frequently make the fatal mistake of equating partial eta-squared (η²_p) directly with Cohen's f in ANOVA. The mathematical relationship is f = √[η²_p / (1 - η²_p)]. If a prior study reports η²_p = .06 (a medium effect in ANOVA), Cohen's f is approximately √[.06 / .94] = 0.2529. In regression, f² = R² / (1 - R²); thus, an expected R² = .13 translates to f² = .13 / .87 = 0.1494 ≈ 0.15.
Execution Verification Checklist:
- Documented the empirical or theoretical rationale for the selected effect size parameter
- Correctly converted published metric variance (η²_p to f, or R² to f²) before entering values into G*Power
- Avoided artificially inflating expected effect sizes (e.g., claiming d = 0.80) solely to rationalize an underpowered convenience sample
Critical Methodological Note: Assuming an unrealistically large effect size (d = 0.80 or f = 0.40) without prior evidence is the most common reason dissertation committees reject quantitative methodologies.
Navigating G*Power 3.1 & Computing A Priori Sample Size
G*Power 3.1.9.7 (developed by Franz Faul and colleagues at Heinrich-Heine-Universität Düsseldorf) is the international gold standard software for statistical power analysis across behavioral, social, biomedical, and physical sciences. Once parameters are established, executing the calculation follows a standardized procedural sequence.
Within the G*Power interface: (1) Open the 'Test family' dropdown and choose the model (e.g., t tests); (2) Select the exact inferential test from 'Statistical test' (e.g., Means: Difference between two independent means); (3) In 'Type of Power Analysis', select 'A priori (Compute required sample size - given α, power, and effect size)'; (4) Enter Input Parameters: Tails (two), Effect size d (0.50), α err prob (0.05), Power 1 - β (0.80), and Allocation ratio N₂/N₁ (1.0 for equal groups); (5) Click 'Calculate'.
The 'Output Parameters' panel immediately displays the noncentrality parameter (δ = 2.8284), critical value (t = 1.9791), degrees of freedom (df = 126), sample size for Group 1 (n₁ = 64), sample size for Group 2 (n₂ = 64), total required sample size (Total N = 128), and actual achieved statistical power (Actual power = .8014). This total represents the clean, complete sample size required at the moment of inferential analysis.
Execution Verification Checklist:
- Selected 'A priori' power analysis mode rather than post-hoc, compromise, or criterion modes
- Verified that the Allocation Ratio reflects real-world recruitment capabilities (N₂/N₁ = 1 for balanced designs)
- Recorded all computational outputs: Total N, group n allocations, critical threshold, and noncentrality parameter
Critical Methodological Note: Always record the exact software version number (G*Power 3.1.9.7) and cite both foundational papers (Faul et al., 2007, 2009) in your manuscript references.
Applying the Participant Attrition & Drop-Out Adjustment Formula
A universal pitfall in quantitative research is treating G*Power's calculated Total N as the target recruitment figure. G*Power calculates the minimum number of clean, complete, and analyzable observations required in your statistical software dataset. It does not account for participant drop-out, survey abandonment, failed attention checks, missing values, withdrawal of consent, or multivariate outlier exclusions.
If G*Power indicates N_clean = 128, and a researcher collects exactly 128 survey responses, standard attrition (typically 10% to 20%) will leave only 102 to 115 usable cases. The study will enter data analysis underpowered, directly violating the a priori protocol approved by the IRB or thesis committee.
To safeguard empirical power, researchers must inflate the recruitment target before launching data collection using the standard attrition formula: N_recruitment = ⌈N_clean / (1 - d)⌉, where d is the anticipated proportion of unusable responses (0.10 ≤ d ≤ 0.20). For N_clean = 128 with d = 0.15: N_recruitment = ⌈128 / (1 - 0.15)⌉ = ⌈128 / 0.85⌉ = ⌈150.59⌉ = 151 participants (76 per group).
Execution Verification Checklist:
- Identified anticipated participant attrition rate based on study modality (online survey: 15–25%; lab experiment: 5–10%; longitudinal: 20–35%)
- Applied the upward ceiling division formula [N_rec = ⌈N_clean / (1 - d)⌉] prior to IRB submission or participant recruitment
- Explicitly reported both N_clean and N_recruitment in Chapter 3 or the Methods section
Critical Methodological Note: Never reduce your sample size after data collection without documenting exact reasons for exclusion in a CONSORT or flow diagram.
Formulating the APA 7 Methodology Justification Section
The American Psychological Association (APA 7th Edition) Section 3.6 and Journal Article Reporting Standards for Quantitative Research (Chapter 3 JARS-Quant) mandate full transparency in sample size determination. Reviewers and committee members expect a dedicated subsection within the Methods section (typically under 'Participants' or 'Sample Size Justification') that articulates every input parameter and decision.
An APA 7 compliant power paragraph must state: (1) Software name and version with citations (G*Power 3.1.9.7; Faul et al., 2007, 2009); (2) Exact statistical test; (3) Type I error rate (α); (4) Desired statistical power (1 - β); (5) Anticipated effect size with supporting citation or SESOI justification; (6) Resulting minimum clean sample size (N_clean); and (7) Attrition rate and final recruitment target (N_recruitment).
In strict adherence to APA 7 Section 6.44 mathematical presentation rules, eliminate leading zeros for statistical metrics that cannot exceed 1.0 (such as α = .05, 1 - β = .80, r = .30, and p < .05). Retain leading zeros only for variables that can theoretically exceed 1.0 (such as Cohen's d = 0.50 and Cohen's f = 0.25).
Execution Verification Checklist:
- Formatted all statistical symbols in standard APA italics (N, n, d, f, r, p, α, β)
- Eliminated leading zeros for bounded probabilities (α = .05, power = .80, p < .05)
- Provided verifiable citations for both G*Power (Faul et al., 2007, 2009) and the effect size benchmark (Cohen, 1988)
Top 6 Methodological Pitfalls & Prevention Safeguards
The Post-Hoc Observed Power Fallacy
Taking an observed non-significant p-value (p > .05) from your study, feeding the sample effect size back into G*Power, and reporting 'retrospective power was low.'
Hoenig and Heisey (2001) proved that observed post-hoc power is merely a 1-to-1 mathematical transformation of the p-value; it is inherently circular and scientifically meaningless. If reviewers question statistical power after data collection, execute a Sensitivity Power Analysis in G*Power to report the minimum detectable effect size (MDES) your achieved sample was capable of detecting.
The Yamane / Survey Formula Misapplication
Using Taro Yamane's formula [n = N / (1 + Ne²)] or Cochran's proportion formula to determine sample sizes for experimental, ANOVA, or regression designs.
Yamane's formula only estimates single population proportions in descriptive surveys. It possesses zero parameters for variance, effect size, statistical power, or group comparisons. Using Yamane for an experimental thesis will result in immediate committee rejection.
P-Hacking and Optional Stopping
Collecting 30 participants, testing for significance, seeing p = .08, adding 10 more subjects, and repeating until p < .05.
Unplanned interim testing inflates the true Type I error rate from 5% up to 25–30%. Adhere strictly to the pre-calculated a priori sample size as a fixed stopping rule unless sequential analysis with adjusted alpha boundaries was pre-registered.
Confusing Eta-Squared with Cohen's f
Directly entering a published partial eta-squared (e.g., η²_p = .06) into G*Power's Cohen's f field for ANOVA.
Cohen's f is not eta-squared; it is the square root of the ratio of explained to unexplained variance: f = √[η²_p / (1 - η²_p)]. Entering .06 directly treats the study as testing a massive effect (f = .06 corresponds to η²_p ≈ .0036), drastically underestimating the required sample size.
Neglecting Participant Attrition
Recruiting exactly the clean Total N produced by G*Power, leaving the final analyzable dataset underpowered after standard 15% participant drop-out.
Always inflate your recruitment goal using N_rec = ⌈N_clean / (1 - d)⌉. Document both the clean inferential target and the recruitment total in your institutional review board (IRB) protocol.
Over-Optimistic Effect Size Assumptions
Assuming a large effect size (d = 0.80 or f = 0.40) without empirical justification solely to rationalize an inadequate convenience sample.
In social, psychological, educational, and biomedical sciences, true population effects are predominantly small to medium (d = 0.20 to 0.50). Overestimating effect sizes guarantees an underpowered study. Always base parameters on conservative meta-analytic estimates.
Turnkey APA 7 Methodology Reporting Paragraph Templates
Copy, adapt, and insert these rigorously formulated, peer-review-tested templates directly into your dissertation Chapter 3 or journal Methods section:
Template 1: Independent Samples t-Test (Two Experimental Groups)
Context: Comparing treatment vs. control conditions on a continuous outcome measure in dissertation Chapter 3 or journal methods.
An a priori statistical power analysis was conducted prior to participant recruitment using G*Power software (Version 3.1.9.7; Faul et al., 2007, 2009) to determine the sample size required to test the primary research hypothesis. For a two-tailed independent-samples t-test comparing two independent cohorts with equal group allocation (N₂/N₁ = 1.0), parameters were specified with a nominal significance criterion of α = .05, a target statistical power of 1 - β = .80, and a medium effect size benchmark of Cohen’s d = 0.50 (Cohen, 1988).
The calculation indicated that a minimum sample size of N = 128 participants (n₁ = 64, n₂ = 64) is required to detect a medium effect with adequate power (actual power = .801, critical t = 1.979, df = 126). To account for potential survey abandonment, incomplete responses, and failed attention checks estimated at approximately 15% based on prior empirical administration, the target recruitment goal was adjusted upward to N = 151 participants (approximately 76 per group) using the standard attrition formula [N_recruitment = ⌈N_clean / (1 - d)⌉].Template 2: One-Way Between-Subjects ANOVA (3 Groups)
Context: Testing mean differences across three intervention conditions or institutional cohorts.
Sample size adequacy for the between-subjects one-way analysis of variance (ANOVA) across three experimental groups was determined a priori using G*Power 3.1.9.7 (Faul et al., 2007). Testing parameters were established with an alpha level of α = .05, a statistical power of 1 - β = .80, and an omnibus medium effect size convention of Cohen’s f = 0.25 (corresponding to partial η² ≈ .059; Cohen, 1988).
G*Power computed a minimum required total sample size of N = 159 participants (n = 53 per group; numerator df = 2, denominator df = 156, critical F = 3.054, actual power = .806). Anticipating an attrition and data-screening exclusion rate of 15%, a total of 188 participants (approximately 63 per condition) will be recruited to ensure the final analyzable sample meets or exceeds the required threshold.Template 3: Multiple Linear Regression (5 Predictors)
Context: Evaluating variance explained (R²) by five simultaneous continuous and dummy-coded predictors.
An a priori power calculation for a multiple linear regression model containing five predictors was computed using G*Power 3.1.9.7 (Faul et al., 2009). The analysis was parameterized with a significance threshold of α = .05, statistical power of 1 - β = .80, and a conventional medium effect size of Cohen’s f² = 0.15 (corresponding to an explained variance of R² ≈ .130; Cohen, 1988).
The analysis demonstrated that a minimum clean sample size of N = 92 observations is required to detect an omnibus regression model deviating significantly from zero (critical F = 2.321, actual power = .804). Accounting for an anticipated 15% rate of incomplete submissions and multivariate outlier trimming, recruitment was expanded to N = 109 participants.Template 4: Sensitivity Power Analysis (Peer-Review Response to Underpower Claims)
Context: Refuting a reviewer's demand for post-hoc observed power after obtaining a non-significant result.
Rather than computing post-hoc observed power—which represents an uninformative and circular transformation of the observed p-value (Hoenig & Heisey, 2001)—a sensitivity power analysis was conducted in G*Power 3.1.9.7 (Faul et al., 2007) to determine the minimum detectable effect size (MDES) achievable with the final verified sample size (N = 120, n₁ = 60, n₂ = 60).
With α = .05 (two-tailed) and power fixed at 1 - β = .80, the sensitivity analysis indicated that the study possessed sufficient statistical sensitivity to detect any population effect of Cohen’s d ≥ 0.516 or larger. Consequently, while the non-significant outcome cannot definitively rule out minor effects (d < 0.50), the study was robustly powered to detect moderate to substantial empirical differences.Frequently Asked Questions
Why do journal reviewers and editors reject post-hoc 'observed power' calculations?
Post-hoc power (calculating power after data collection using the observed sample effect size and sample size) is mathematically flawed and scientifically circular. As proven by Hoenig and Heisey (2001) and Yuan and Maxwell (2005), observed power is a direct 1-to-1 monotonic transformation of the p-value. When p > .05, observed power is mathematically guaranteed to be low. Claiming that a null finding was caused by low power based on observed power provides zero new information. Instead, researchers should report an a priori power analysis or conduct a Sensitivity Power Analysis to specify the Minimum Detectable Effect Size (MDES).
How do I justify an effect size in G*Power if no pilot study or previous research exists?
When entering an entirely novel empirical domain without prior published data, researchers should: (1) Conduct a Smallest Effect Size of Interest (SESOI) analysis (Lakens, 2022) to establish the minimum magnitude that holds clinical, practical, or theoretical significance; (2) Draw from meta-analyses in closely related disciplines; or (3) Fall back on Cohen's (1988) conservative medium effect size conventions (d = 0.50 for t-tests, f = 0.25 for ANOVA, f² = 0.15 for regression, r = 0.30 for correlation), explicitly stating in the Methods section that conventional benchmarks were utilized in the absence of published literature.
Why can't I use Taro Yamane's formula for an experimental thesis or regression study?
Taro Yamane's (1967) formula [n = N / (1 + Ne²)] was derived exclusively for estimating a single population proportion in descriptive survey sampling. It contains zero parameters for group variance, experimental intervention effects, directional prediction, statistical power, or Type II error (β). Applying Yamane's formula to an ANOVA, t-test, or regression model is a fundamental methodological error that will result in immediate rejection by competent dissertation committees and peer-reviewed journals.
How does participant drop-out affect G*Power calculations?
G*Power computes the clean, complete, and analyzable sample size required at the exact moment of inferential testing (N_clean). Real-world data collection always incurs attrition, survey fatigue, missing values, and failed attention checks (typically 10% to 20%). If you recruit only N_clean, participant drop-out will leave your final dataset underpowered. You must inflate recruitment using the formula: N_recruitment = ⌈N_clean / (1 - d)⌉, where d is the expected attrition rate.
What is the difference between an A Priori power analysis and a Sensitivity power analysis in G*Power?
An A Priori power analysis is conducted before data collection to determine the required sample size (N) given a pre-specified alpha (α), target power (1 - β), and expected effect size. A Sensitivity power analysis is typically conducted when sample size is fixed or restricted (e.g., in rare clinical populations or post-data collection audits); it computes the minimum effect size that the study was capable of reliably detecting given the fixed N, α, and power.
Verified Primary Sources & Methodological Standards
- Faul, F., Erdfelder, E., Lang, A.-G., & Buchner, A. (2007)
G*Power 3: A flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behavior Research Methods, 39(2), 175–191.
- Faul, F., Erdfelder, E., Buchner, A., & Lang, A.-G. (2009)
Statistical power analyses using G*Power 3.1: Tests for correlation and regression analyses. Behavior Research Methods, 41(4), 1149–1160.
- Cohen, J. (1988)
Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Lawrence Erlbaum Associates. Definitive thresholds for d, f, f², w, and r.
- American Psychological Association (APA). (2020)
Publication Manual of the APA (7th ed.). Section 3.6 & Chapter 3 (JARS-Quant: Journal Article Reporting Standards - Sample Size, Power, and Precision).
- Hoenig, J. M., & Heisey, D. M. (2001)
The abuse of power: The pervasive fallacy of power calculations for data analysis. The American Statistician, 55(1), 19–24. Foundational paper refuting post-hoc observed power.
- Heinrich-Heine-Universität Düsseldorf G*Power Documentation
Official software manual, computational algorithms, and version distribution for G*Power 3.1.9.7.