Study guide
11.1 Inferential Statistics, Correlation, Regression, ANOVA, and Effect Sizes
NSCA Certified Performance and Sport Scientist CPSS
Track your progress as you read
Finished sections get a check mark.
1 Introduction: CPSS Exam Overview, Eligibility, and Case-Study Strategy
1.1 CPSS Exam Format, Eligibility Pathways, and the 2026 Blueprint 1.2 Mastering Athlete Case Studies and Research Review Scenarios
2 Chapter 1: Scientific Process, Adaptation Theory, and Periodization (Domain 1)
2.1 Training Principles, Adaptation Models, and Injury-Risk Mitigation 2.2 Evaluating Performance Programs Using the Scientific Process 2.3 Periodization Paradigms, Fitness-Fatigue Models, and Tapering
3 Chapter 2: Scientific Disciplines in Athlete Preparation (Domain 1)
3.1 Exercise Physiology, Bioenergetics, and Systemic Adaptations 3.2 Applied Biomechanics, Force-Time Kinetics, and Force-Velocity Profiling 3.3 Motor Learning, Skill Acquisition, and Constraints-Led Practice Design 3.4 Performance Nutrition, Recovery Modalities, and Sport Psychology
4 Chapter 3: Technology Review, Research Appraisal, and Knowledge Translation (Domain 1)
4.1 Critically Reviewing Sport Performance Products, Technologies, and Trends 4.2 Evaluating Research Studies: Design, Validity, and Practical Relevance 4.3 Communicating Research Findings to Support Evidence-Based Practice
5 Chapter 4: Sport Demands Analysis, KPIs, and Resource Strategy (Domain 2)
5.1 Researching Biomechanical, Tactical, and Technical Sport Demands 5.2 Physiological Demands, Injury Epidemiology, and Environmental Constraints 5.3 Qualitative and Mixed-Methods Research for Sport Demands Analysis 5.4 Collaborating with Stakeholders to Establish Key Performance Indicators 5.5 Identifying Resources and Strategies for Short- and Long-Term Goals
6 Chapter 5: Profiling Benchmarks and Needs Analysis Communication (Domain 2)
6.1 Developing Normative Benchmarks: Z-Scores, T-Scores, and Percentiles 6.2 Multidimensional Athlete Profiling and Return-to-Performance Milestones 6.3 Communicating Needs Analysis Results to Multidisciplinary Stakeholders
7 Chapter 6: Assessment Ethics, Safety Standards, and Tool Psychometrics (Domain 3)
7.1 Ethical, Legal, and Privacy Considerations in Athlete Data Collection 7.2 Evaluating Assessment Environments, Equipment, and Safety Standards 7.3 Validity, Reliability (ICC, CV%, TE), and Smallest Worthwhile Change
8 Chapter 7: Applied Assessment and Load-Monitoring Systems (Domain 3)
8.1 Movement Tracking Systems: GNSS/GPS, LPS, IMUs, and Optical Tracking 8.2 Neuromuscular Testing: Force Platforms, Isometric Pulls/Pushes, and VBT 8.3 Internal Load, Autonomic Readiness (HRV), Biomarkers, and PROMs
9 Chapter 8: Quality Assurance, Protocols, and Assessment Efficacy (Domain 3)
9.1 Applying Quality Assurance Techniques in Sport Science Data Collection 9.2 Developing, Documenting, and Standardizing Data Collection Protocols 9.3 Systematically Reevaluating Assessments, Protocols, and Interventions
10 Chapter 9: Data Governance, Storage Architecture, and Cleaning (Domain 4)
10.1 Data Privacy, Security Protocols, and Protecting Organizational Information 10.2 Designing Plans for Data Collection, Storage, Analysis, and AMS Workflows 10.3 Evaluating Data Quality, Signal Integrity, Outliers, and Cleaning Procedures
11 Chapter 10: Statistical Modeling, Feedback Delivery, and Analytical Limitations (Domain 4)
11.1 Inferential Statistics, Correlation, Regression, ANOVA, and Effect Sizes 11.2 Longitudinal Load Modeling: ACWR, EWMA, Monotony, Strain, and Change Detection 11.3 Data Visualization, Dashboard Design, and Timely Stakeholder Feedback 11.4 Communicating Analytical Limitations, Bias, and Uncertainty
11.1 Inferential Statistics, Correlation, Regression, ANOVA, and Effect Sizes
Key Takeaways Explain This Topic AI
Parametric hypothesis tests require verification of normality (Shapiro-Wilk p > 0.05 and Q-Q plot inspection), homogeneity of variance (Levene's test p > 0.05), and—in repeated-measures ANOVA—sphericity (Mauchly's test p > 0.05, using Greenhouse-Geisser or Huynh-Feldt epsilon corrections when violated).
The parametric vs. non-parametric test matrix pairs Independent t-tests with Mann-Whitney U, Paired t-tests with Wilcoxon Signed-Rank, One-Way ANOVA with Kruskal-Wallis H, and Repeated-Measures ANOVA with Friedman's test.
Linear Mixed-Effects Models (LMMs) outperform classical repeated-measures ANOVA in longitudinal athlete monitoring because LMMs handle missing data without dropping entire athletes and simultaneously model fixed group effects alongside random athlete intercepts and slopes.
In correlation and regression modeling, the coefficient of determination (R^2 = r^2) quantifies the proportion of explained variance (e.g., r = 0.60 explains 36% of variance, leaving 64% unexplained), while Variance Inflation Factors (VIF > 5 to 10) diagnose multicollinearity among predictors.
Binary logistic regression models the log-odds of dichotomous outcomes (such as injured vs. uninjured) and outputs Odds Ratios (OR = e^beta), whereas standardized effect sizes (Cohen's d, Hedges' g, and Partial Eta-Squared) quantify practical magnitude independent of sample size.
Last updated: October 2026
11.1 Inferential Statistics, Correlation, Regression, ANOVA, and Effect Sizes
Quick Answer: Choosing the valid statistical model in sport science depends on four structural factors: the measurement scale and distribution (parametric continuous vs. non-parametric ordinal/skewed), the sampling design (independent groups vs. repeated within-athlete measures), the number of conditions ( 2 2 2 vs. ≥ 3 \ge 3 ≥ 3 ), and whether longitudinal observations contain missing data and nested random effects (requiring Linear Mixed-Effects Models [LMMs] ). Because p -values p\text{-values} p -values confound effect magnitude with sample size, inferential tests must always be paired with standardized effect sizes ( Cohen's d d d , Hedges' g g g , Partial Eta-Squared [ η p 2 \eta_p^2 η p 2 ] , Correlation [ r r r ] , or Odds Ratios [ OR \text{OR} OR ] ) and confidence intervals.
In high-performance sport science, sample sizes are naturally constrained by roster limits ( N = 15 – 60 N = 15\text{–}60 N = 15 – 60 athletes), and datasets frequently exhibit skewness, outliers, or missing observations due to injury and international travel. Mastering assumption diagnostics, test selection, regression modeling, and effect-size scaling ensures that performance conclusions rest on sound mathematical foundations.
Evaluating Parametric Assumptions: Normality, Homoscedasticity, and Sphericity
Before applying parametric models that rely on sample means and standard deviations, the sport scientist must audit three distributional assumptions:
Normality of Residuals / Distributions:
Formal Hypothesis Test: The Shapiro-Wilk test is preferred for typical squad sample sizes ( n < 50 n < 50 n < 50 ). The null hypothesis ( H 0 H_0 H 0 ) states that the data follow a normal Gaussian distribution. Therefore, p > 0.05 p > 0.05 p > 0.05 satisfies the normality assumption , whereas p ≤ 0.05 p \le 0.05 p ≤ 0.05 indicates a statistically significant departure from normality (skewness or heavy tails).
Visual & Descriptive Diagnostics: Because the Shapiro-Wilk test can be underpowered in tiny samples ( n < 12 n < 12 n < 12 ) and overly sensitive in huge wearables datasets ( n > 1,000 n > 1{,}000 n > 1 , 000 ), practitioners always inspect Quantile-Quantile (Q-Q) plots (checking that standardized residuals adhere closely to the 45 ∘ 45^\circ 4 5 ∘ diagonal line), histograms, and standardized skewness and kurtosis ratios.
Homogeneity of Variance (Homoscedasticity):
Tested across independent groups via Levene's test (or Bartlett's test ). A result of p > 0.05 p > 0.05 p > 0.05 confirms equal variances across position groups or cohorts. If Levene's test is violated ( p ≤ 0.05 p \le 0.05 p ≤ 0.05 ) in a two-group comparison, the practitioner substitutes Welch's unpooled t -test t\text{-test} t -test , which adjusts the degrees of freedom via the Welch-Satterthwaite equation to prevent Type I error inflation.
Sphericity in Repeated-Measures Designs:
In repeated-measures Analysis of Variance ( ANOVA \text{ANOVA} ANOVA ) across ≥ 3 \ge 3 ≥ 3 timepoints, sphericity requires that the variances of the differences between all possible pairs of within-subject conditions are equal ( σ T 1 − T 2 2 ≈ σ T 1 − T 3 2 ≈ σ T 2 − T 3 2 \sigma_{T1-T2}^2 \approx \sigma_{T1-T3}^2 \approx \sigma_{T2-T3}^2 σ T 1 − T 2 2 ≈ σ T 1 − T 3 2 ≈ σ T 2 − T 3 2 ).
Mauchly's Test of Sphericity: Evaluates whether sphericity holds ( p > 0.05 p > 0.05 p > 0.05 ). When Mauchly's test is significant ( p ≤ 0.05 p \le 0.05 p ≤ 0.05 , indicating sphericity is violated), the standard unadjusted F -test F\text{-test} F -test overestimates degrees of freedom and inflates the Type I (false positive) error rate.
Epsilon ( ϵ \epsilon ϵ ) Degrees-of-Freedom Corrections: When sphericity is violated, practitioners multiply both numerator and denominator degrees of freedom by an epsilon adjustment factor ( 0 < ϵ ≤ 1 0 < \epsilon \le 1 0 < ϵ ≤ 1 ), which raises the critical F -value F\text{-value} F -value required for significance: the Greenhouse-Geisser correction (more conservative; recommended when ϵ < 0.75 \epsilon < 0.75 ϵ < 0.75 ) or the Huynh-Feldt correction (less conservative; preferred when ϵ ≥ 0.75 \epsilon \ge 0.75 ϵ ≥ 0.75 ).
Parametric vs. Non-Parametric Statistical Test Selection Matrix
When data are ordinal (such as raw 1 – 5 1\text{–}5 1 – 5 Likert soreness scores) or remain severely skewed after logarithmic transformation, non-parametric rank-based equivalents must be selected:
Analytical Research Design Sport Science Scenario Example Parametric Test (Normal, Interval/Ratio) Non-Parametric Equivalent (Skewed or Ordinal) Post-Hoc / Error Control & Key Statistic
2 Independent Groups Starters vs. Non-starters 10 m 10\text{ m} 10 m sprint time Independent-Samples t -Test t\text{-Test} t -Test (or Welch's t -test t\text{-test} t -test if unequal variance) Mann-Whitney U U U Test (Wilcoxon Rank-Sum Test) t -statistic t\text{-statistic} t -statistic or U -statistic U\text{-statistic} U -statistic ; Cohen's d d d / Hedges' g g g
2 Paired / Dependent Measures Pre-camp vs. Post-camp countermovement jump ( CMJ \text{CMJ} CMJ ) on same squad Paired-Samples (Dependent) t -Test t\text{-Test} t -Test Wilcoxon Signed-Rank Test Paired t -statistic t\text{-statistic} t -statistic or W / Z W/Z W / Z ; standardized mean change d z d_z d z
≥ 3 \ge 3 ≥ 3 Independent Groups Forwards vs. Midfielders vs. Defenders VO 2 max \text{VO}_{2\max} VO 2 m a x One-Way Between-Groups ANOVA \text{ANOVA} ANOVA ( F = MS between / MS within F = \text{MS}_{\text{between}} / \text{MS}_{\text{within}} F = MS between / MS within ) Kruskal-Wallis H H H Test Tukey's HSD , Bonferroni , or Holm-Bonferroni (Dunn's test for Kruskal-Wallis)
≥ 3 \ge 3 ≥ 3 Repeated Measures (Complete Cases) Pre-, Mid-, and Post-season DXA \text{DXA} DXA lean mass in same 20 20 20 athletes One-Way Repeated-Measures ANOVA \text{ANOVA} ANOVA Friedman's Two-Way ANOVA \text{ANOVA} ANOVA by Ranks Mauchly's test + Greenhouse-Geisser / Huynh-Feldt ϵ \epsilon ϵ ; Bonferroni pairwise comparisons
Longitudinal / Hierarchical with Missing Data Daily GNSS \text{GNSS} GNSS high-speed distance across 38 38 38 matches with rotation & injuries Linear Mixed-Effects Model ( LMM \text{LMM} LMM ) Generalized Linear Mixed Model ( GLMM \text{GLMM} GLMM ) Fixed Effects ( β \beta β ) + Random Intercepts ( u 0 j u_{0j} u 0 j ) & Random Slopes ( u 1 j u_{1j} u 1 j ) via REML \text{REML} REML
Categorical / Contingency Proportions Playing Surface (Turf vs. Grass) × \times × Non-Contact Injury (Yes vs. No) Chi-Square ( χ 2 \chi^2 χ 2 ) Test of Independence (all expected cells ≥ 5 \ge 5 ≥ 5 ) Fisher's Exact Test (any expected cell frequency < 5 < 5 < 5 ) χ 2 = ∑ ( O − E ) 2 E \chi^2 = \sum \frac{(O - E)^2}{E} χ 2 = ∑ E ( O − E ) 2 ; Cramer's V V V or Odds Ratio ( OR \text{OR} OR )
ANOVA Familywise Error Control and Post-Hoc Testing
In a One-Way ANOVA \text{ANOVA} ANOVA , the omnibus F -ratio F\text{-ratio} F -ratio divides systematic between-group variance by unsystematic within-group error variance: F = MS between MS within F = \frac{\text{MS}_{\text{between}}}{\text{MS}_{\text{within}}} F = MS within MS between
A significant omnibus F -test F\text{-test} F -test ( p < 0.05 p < 0.05 p < 0.05 ) indicates only that at least one group mean differs from another, without specifying which pair differs. Running multiple unadjusted t -tests t\text{-tests} t -tests across k k k pairs inflates the Familywise Type I Error Rate ( α FW = 1 − ( 1 − α ) c \alpha_{\text{FW}} = 1 - (1 - \alpha)^c α FW = 1 − ( 1 − α ) c , rising to 14.3 % 14.3\% 14.3% across 3 3 3 comparisons and 40.1 % 40.1\% 40.1% across 10 10 10 comparisons at α = 0.05 \alpha = 0.05 α = 0.05 ). Practitioners control α FW \alpha_{\text{FW}} α FW using:
Tukey's Honestly Significant Difference ( HSD \text{HSD} HSD ): Optimal for all-pairwise comparisons among balanced independent groups.
Bonferroni Correction: Divides the alpha threshold by the number of comparisons ( α adjusted = α / c \alpha_{\text{adjusted}} = \alpha / c α adjusted = α / c , e.g., 0.05 / 6 = 0.0083 0.05 / 6 = 0.0083 0.05/6 = 0.0083 ). Highly conservative, increasing Type II (false negative) error risk when c c c is large.
Holm-Bonferroni Step-Down Procedure: Ranks p -values p\text{-values} p -values from smallest to largest and applies sequentially less restrictive denominators ( α / c , α / [ c − 1 ] , … \alpha / c, \alpha / [c-1], \dots α / c , α / [ c − 1 ] , … ), preserving greater statistical power than classical Bonferroni while strictly controlling α FW \alpha_{\text{FW}} α FW .
Why Linear Mixed-Effects Models (LMMs) Have Replaced Repeated-Measures ANOVA
Classical repeated-measures ANOVA \text{ANOVA} ANOVA suffers from a fatal flaw in applied sport settings: listwise deletion . If an athlete misses even a single testing timepoint out of 10 10 10 due to national team call-up or minor illness, repeated-measures ANOVA \text{ANOVA} ANOVA discards that athlete's entire longitudinal record. Linear Mixed-Effects Models ( LMMs \text{LMMs} LMMs ) overcome this via Restricted Maximum Likelihood ( REML \text{REML} REML ) estimation:
Retaining Unbalanced & Missing Data: LMMs \text{LMMs} LMMs incorporate every valid observation from every athlete under Missing at Random ( MAR \text{MAR} MAR ) conditions without listwise deletion.
Fixed Effects vs. Random Effects: LMMs \text{LMMs} LMMs simultaneously estimate population-level fixed effects (such as playing position, match location, or environmental temperature) alongside subject-specific random intercepts (accounting for each athlete's unique baseline physiological ceiling) and random slopes (modeling how individual athletes adapt or fatigue at different rates in response to the same training load).
- Correlation, Regression, and Effect Size Magnitude Scales
- Bivariate Correlation and the Coefficient of Determination ( R 2 R^2 R 2 )
Pearson Product-Moment Correlation ( r r r ): Quantifies the linear relationship between two continuous, normally distributed variables ( − 1.0 ≤ r ≤ + 1.0 -1.0 \le r \le +1.0 − 1.0 ≤ r ≤ + 1.0 ).
Spearman's Rank-Order Correlation ( ρ \rho ρ or r s r_s r s ): Quantifies the monotonic relationship between ranked, ordinal, or skewed variables, reducing distortion from extreme outliers.
Hopkins's Correlation Magnitude Scale:
- ∣ r ∣ < 0.10 |r| < 0.10 ∣ r ∣ < 0.10 : Trivial
- 0.10 – 0.29 0.10\text{–}0.29 0.10 – 0.29 : Small
- 0.30 – 0.49 0.30\text{–}0.49 0.30 – 0.49 : Moderate
- 0.50 – 0.69 0.50\text{–}0.69 0.50 – 0.69 : Large
- 0.70 – 0.89 0.70\text{–}0.89 0.70 – 0.89 : Very Large
- ∣ r ∣ ≥ 0.90 |r| \ge 0.90 ∣ r ∣ ≥ 0.90 : Nearly Perfect
Coefficient of Determination ( R 2 = r 2 R^2 = r^2 R 2 = r 2 ): Squaring the correlation coefficient reveals the exact proportion of variance in the outcome variable ( Y Y Y ) explained by the predictor ( X X X ) . For example, a "large" correlation of r = 0.60 r = 0.60 r = 0.60 between isometric mid-thigh pull peak force and sprint acceleration yields R 2 = 0.60 2 = 0.36 R^2 = 0.60^2 = 0.36 R 2 = 0.6 0 2 = 0.36 , meaning 36 % 36\% 36% of the variance in sprint time is explained by isometric force, while 64 % 64\% 64% ( 1 − R 2 1 - R^2 1 − R 2 ) remains unexplained (attributable to technique, rate of force development, and measurement error).
Ordinary Least Squares (OLS) Multiple Regression and Binary Logistic Regression
In Linear and Multiple Regression ( Y = β 0 + β 1 X 1 + β 2 X 2 + ⋯ + ϵ Y = \beta_0 + \beta_1 X_1 + \beta_2 X_2 + \dots + \epsilon Y = β 0 + β 1 X 1 + β 2 X 2 + ⋯ + ϵ ), practitioners evaluate three model diagnostics:
Adjusted R 2 R^2 R 2 : Unlike raw R 2 R^2 R 2 —which artificially increases every time any random variable is added to a model— Adjusted R 2 R^2 R 2 penalizes the model for adding predictors that do not meaningfully improve fit relative to the lost degrees of freedom.
Standard Error of the Estimate ( SEE \text{SEE} SEE ): Quantifies the standard deviation of the residuals around the regression line in the original measurement units of Y Y Y (e.g., ± 2.4 kg \pm 2.4\text{ kg} ± 2.4 kg for a 1 RM 1\text{RM} 1 RM prediction equation from barbell velocity).
Variance Inflation Factor ( VIF = 1 1 − R j 2 \text{VIF} = \frac{1}{1 - R_j^2} VIF = 1 − R j 2 1 ): Diagnoses multicollinearity (entering highly redundant predictors into the same regression model, such as including both Total Distance and Meters per Minute). A VIF > 5 to 10 \text{VIF} > 5\text{ to }10 VIF > 5 to 10 signals severe multicollinearity, which inflates standard errors and makes individual β \beta β coefficients unstable and uninterpretable.
When the dependent outcome is dichotomous ( 0 = Uninjured , 1 = Injured 0 = \text{Uninjured}, 1 = \text{Injured} 0 = Uninjured , 1 = Injured ; or Win vs. Loss \text{Win vs. Loss} Win vs. Loss ), ordinary linear regression is invalid because predicted probabilities can fall below 0 0 0 or above 1 1 1 and residuals are heteroscedastic. Practitioners instead fit a Binary Logistic Regression modeling the natural log-odds ( logit \text{logit} logit ) of the event: ln ( P 1 − P ) = β 0 + β 1 X 1 + ⋯ + β k X k \ln\left(\frac{P}{1 - P}\right) = \beta_0 + \beta_1 X_1 + \dots + \beta_k X_k ln ( 1 − P P ) = β 0 + β 1 X 1 + ⋯ + β k X k
Exponentiating a predictor's regression slope ( β \beta β ) yields the Odds Ratio ( OR = e β \text{OR} = e^{\beta} OR = e β ) :
OR = 1.00 \text{OR} = 1.00 OR = 1.00 : Predictor has no association with the outcome odds.
OR > 1.00 \text{OR} > 1.00 OR > 1.00 : Each 1 -unit 1\text{-unit} 1 -unit increase in X X X multiplies the odds of the outcome (e.g., OR = 2.40 \text{OR} = 2.40 OR = 2.40 means 2.4 × 2.4\times 2.4 × higher odds, or + 140 % +140\% + 140% increased odds).
OR < 1.00 \text{OR} < 1.00 OR < 1.00 : Predictor is protective (e.g., OR = 0.55 \text{OR} = 0.55 OR = 0.55 indicates a 45 % 45\% 45% reduction in the odds of injury per unit increase in eccentric hamstring strength).
Standardized Effect Sizes for Group and ANOVA Comparisons
Cohen's d d d and Hedges' g g g : Standardize the difference between two means in standard deviation units ( d = X ˉ 1 − X ˉ 2 S D pooled d = \frac{\bar{X}_1 - \bar{X}_2}{SD_{\text{pooled}}} d = S D pooled X ˉ 1 − X ˉ 2 ). When sample sizes are small ( n < 20 n < 20 n < 20 per group), Cohen's d d d slightly overestimates the population effect; Hedges' g g g multiplies d d d by a degrees-of-freedom correction factor ( J ≈ 1 − 3 4 df − 1 J \approx 1 - \frac{3}{4\text{df} - 1} J ≈ 1 − 4 df − 1 3 ) to eliminate small-sample bias. On Hopkins's applied sport science scale, ∣ d ∣ < 0.20 |d| < 0.20 ∣ d ∣ < 0.20 is trivial, 0.20 – 0.59 0.20\text{–}0.59 0.20 – 0.59 is small, 0.60 – 1.19 0.60\text{–}1.19 0.60 – 1.19 is moderate, 1.20 – 1.99 1.20\text{–}1.99 1.20 – 1.99 is large, and ≥ 2.00 \ge 2.00 ≥ 2.00 is very large (whereas Cohen's classic behavioral thresholds are 0.20 0.20 0.20 small, 0.50 0.50 0.50 medium, and 0.80 0.80 0.80 large).
Eta-Squared ( η 2 \eta^2 η 2 ) and Partial Eta-Squared ( η p 2 \eta_p^2 η p 2 ): Quantify the proportion of variance explained by an ANOVA \text{ANOVA} ANOVA main effect or interaction: η p 2 = SS effect SS effect + SS error \eta_p^2 = \frac{\text{SS}_{\text{effect}}}{\text{SS}_{\text{effect}} + \text{SS}_{\text{error}}} η p 2 = SS effect + SS error SS effect
Conventional η p 2 \eta_p^2 η p 2 benchmarks classify 0.01 0.01 0.01 as a small effect, 0.06 0.06 0.06 as a medium effect, and ≥ 0.14 \ge 0.14 ≥ 0.14 as a large effect.