Summary measures in statistics

1. High-Yield Summary

• Variable-Driven Summary Selection: Descriptive summary measures must strictly align with variable scale and distribution geometry. Normally distributed continuous data mandate the arithmetic mean and standard deviation (SD), whereas skewed continuous data and ordinal scoring systems require the median and interquartile range (IQR).
• Descriptive vs. Inferential Distinction: The standard deviation (SD) quantifies biological and clinical dispersion within an observed sample. The standard error of the mean (SEM = SD / √n) quantifies the sampling variability and precision of the estimated population mean. Substituting SEM for SD in baseline demographic tables or clinical outcomes to artificially diminish perceived variability is a serious methodological error.
• The Altman-Bland Skewness Rule: For non-negative continuous surgical variables (e.g., intraoperative blood loss, length of hospital stay, postoperative morphine consumption), if the mean is less than twice the standard deviation (Mean < 2 × SD), the underlying distribution cannot be Gaussian and is inevitably right-skewed, requiring non-parametric presentation.
• Orthopaedic PROMs Distributional Characteristics: Validated patient-reported outcome measures (e.g., DASH, WOMAC, KOOS, Visual Analog Scale) frequently demonstrate bounded scales with prominent ceiling or floor effects. Presenting these metrics as parametric means with standard deviations distorts actual treatment efficacy and obscures clinical non-responders.
• Consensus Reporting Requirements: Following the international SAMPL (Statistical Analyses and Methods in the Published Literature) guidelines, summary measures must always be reported in paired combinations (Mean ± SD or Median with 25th–75th percentiles) accompanied by sample size (n), precise numerator/denominator fractions for categorical rates, and 95% confidence intervals for comparative effect sizes.

2. Conceptual Foundation & Variable Taxonomy

Biomedical data generated in orthopaedic surgery and musculoskeletal research originate across diverse measurement hierarchies. Selecting appropriate summary measures requires precise categorization of the underlying variable type, because mathematical operations applicable to continuous metrics introduce substantial bias when applied to categorical or rank-ordered observations.

Measurement Scales in Clinical Research

• Nominal Scale: Qualitative categories without inherent numerical ordering or magnitude. Examples include surgical approach (anterior vs. posterior), fixation construct (intramedullary nail vs. plate and screws), or surgical site infection status (present vs. absent). These data can only be summarized by frequencies, proportions, percentages, and the mode.
• Ordinal Scale: Categorical classifications with a logical order or rank, but lacking equal or uniform intervals between consecutive tiers. Examples include Kellgren-Lawrence radiographic osteoarthritis grading (Grades 0–4), Gustilo-Anderson open fracture classification (Types I, II, IIIA, IIIB, IIIC), and Likert satisfaction scores. Calculating an arithmetic mean on ordinal categories is mathematically invalid; the median and interquartile range provide the only valid summary measures of central tendency and dispersion.
• Interval Scale: Continuous numerical data with equal distance between consecutive increments, but lacking a true physical absolute zero. A classical biological example is temperature in Celsius. Arithmetic addition and subtraction are valid, but ratios cannot be formed.
• Ratio Scale: Continuous numerical data featuring both constant intervals and an absolute physical zero point representing total absence of the measured property. Examples include operative time (minutes), intraoperative blood loss (milliliters), active range of motion (degrees), bone mineral density (g/cm²), and serum inflammatory markers (C-reactive protein in mg/L). All mathematical operations, including ratios, geometric means, and coefficient of variation calculations, are fully permissible.

Discrete versus Continuous Variables

Variables are further subdivided by their structural granularity. Discrete variables assume distinct, countable integer values (e.g., number of revision surgeries, number of anchored sutures, or surgical ward hospital days). Continuous variables can theoretically assume an infinite number of intermediate fractional values between any two points within an interval (e.g., peak joint contact force in newtons, femoral head translation in millimeters). Continuous measures demand formal distributional assessment prior to summarization.

3. Measures of Central Tendency

Measures of central tendency identify the central, typical, or most representative value of an empirical dataset. The choice of metric fundamentally dictates the clinical impression conveyed to readers and operating surgeons.

The Arithmetic Mean

The arithmetic mean represents the mathematical average, computed by summing all individual observations and dividing by the total sample size (n). The mean utilizes all available quantitative information in the sample, rendering it mathematically efficient for subsequent inferential testing. However, the arithmetic mean is highly sensitive to extreme outliers and skewed tails. In an orthopaedic clinical trial with 50 primary total hip arthroplasties where 48 patients lose 200–350 mL of blood and 2 patients experience catastrophic vascular lacerations with 4,000 mL blood loss, the mean blood loss shifts dramatically upward, failing to represent the clinical trajectory of the vast majority of patients.

The Median

The median represents the 50th percentile of an ordered distribution, effectively dividing the dataset into two equal halves such that 50% of observations fall at or below the median and 50% fall at or above it. The median is completely non-parametric and robust against severe outliers and distribution asymmetry. It serves as the primary measure of central tendency for skewed continuous surgical data and ordinal clinical outcome scores.

The Mode

The mode identifies the most frequently occurring value in a dataset. While rarely informative as a standalone summary for continuous variables, the mode is the primary measure of central tendency for nominal variables (e.g., the modal organism isolated in periprosthetic joint infection is Staphylococcus aureus) and bimodal distributions where two distinct subpopulations exist (e.g., bimodal age distribution in high-energy trauma).

The Geometric Mean

The geometric mean is defined as the n-th root of the product of n observations, or equivalently, the antilogarithm of the mean of log-transformed values. The geometric mean is indicated for continuous variables that demonstrate exponential growth, multiplicative relationships, or log-normal distributions, such as serum cytokine titers, bacterial colony counts in biofilm sonication assays, or systemic pharmacokinetic drug concentrations.

4. Measures of Dispersion and Variability

Reporting a measure of central tendency without an accompanying measure of dispersion is methodologically unacceptable. Dispersion quantifies the scatter, spread, and biological heterogeneity of the data around its central point.

Range and Interquartile Range (IQR)

The total range spans the minimum to maximum observed values (Max − Min). While intuitive, the extreme range is exceptionally sensitive to sample size and isolated outliers. A far superior metric is the interquartile range (IQR), defined as the difference between the 75th percentile (third quartile, Q3) and the 25th percentile (first quartile, Q1), capturing the central 50% of the distribution. Modern reporting standards require presenting the specific quartile boundaries (e.g., “Median: 42 mm; IQR: 31–56 mm”) rather than a single subtracted difference.

Variance and Standard Deviation (SD)

Sample variance (s²) is the sum of squared deviations from the sample mean divided by the degrees of freedom (n − 1). Because variance is expressed in squared units of measurement, its direct clinical interpretation is limited. The sample standard deviation (s or SD) is the positive square root of the variance, restoring the metric to the original units of measurement. In a Gaussian distribution, the empirical 68-95-99.7 rule dictates that approximately 68.3% of observations fall within Mean ± 1 SD, 95.4% fall within Mean ± 2 SD, and 99.7% fall within Mean ± 3 SD.

Coefficient of Variation (CV)

The coefficient of variation expresses the standard deviation as a percentage of the mean (CV = [SD / Mean] × 100%). It provides a dimensionless, normalized index of relative variability, permitting direct comparison of measurement precision across disparate units or magnitudes, such as comparing the intraoperative reproducibility of robotic navigation angles versus fluoroscopic measurements.

5. Data Distribution & Detecting Skewness

Evaluating distribution geometry is the mandatory prerequisite before compiling summary tables in orthopaedic literature. A normal (Gaussian) distribution is symmetrical, unimodal, and bell-shaped, wherein the mean, median, and mode coincide at the exact center.

Morphology of Distribution Asymmetry

• Positive (Right) Skewness: The distribution tail extends toward higher positive values. In right-skewed data, the arithmetic mean is pulled substantially above the median by high-value outliers (Mean > Median > Mode). Typical orthopaedic examples include hospital length of stay, duration of mechanical ventilation, total surgical drainage output, and postoperative narcotic utilization.
• Negative (Left) Skewness: The distribution tail extends toward lower values, with the majority of observations clustered near the upper ceiling (Mean < Median). Typical orthopaedic examples include postoperative functional joint scores at long-term follow-up (e.g., Oxford Hip Score or Harris Hip Score in successful total joint replacement cohorts where most patients achieve near-maximal recovery).

The Altman-Bland Diagnostic Benchmark

In a seminal methodological contribution, Altman and Bland established an elegant mathematical rule of thumb for rapidly detecting skewness in published clinical summaries. In any continuous variable that cannot physically assume negative values (e.g., blood loss, surgical time, rehabilitation recovery days):

If the reported sample standard deviation is greater than half of the arithmetic mean (SD > Mean / 2), or conversely, if the mean is less than twice the standard deviation (Mean < 2 × SD), the data cannot follow a Gaussian distribution. If the data were truly normal under such parameters, Mean − 2 × SD would result in a negative value, implying that approximately 2.5% of the patient population had negative blood loss or negative operative duration, which is biologically impossible. Whenever this condition is identified, the data are unequivocally skewed, and the arithmetic mean should be replaced by the median and interquartile range.

6. Diagnostic Tests of Normality & Graphical Methods

Rigorous biostatistical workflows combine visual graphical assessment with formal inferential normality testing before selecting summary parameters.

Graphical Assessment Techniques

• Histograms and Density Plots: Direct inspection of bin distributions to observe unimodality, asymmetry, and tail thickness. Bin width selection must be calibrated to avoid artificial smoothing.
• Quantile-Quantile (Q-Q) Plots: Plotting the quantiles of empirical sample data against theoretical quantiles of a standard normal distribution. If the data are Gaussian, plotted points fall along a straight diagonal reference line. Upward or downward curvilinear departures immediately expose skewness and heavy tails.
• Box-and-Whisker Plots: Depicting the five-number summary (minimum, Q1, median, Q3, maximum). Asymmetry between the upper and lower halves of the central box or disproportionate whisker lengths provide instantaneous visual verification of non-normality.

Formal Hypothesis Testing for Normality

• Shapiro-Wilk Test: The most statistically powerful and preferred formal normality test for small-to-moderate clinical cohorts (n < 2,000). A p-value < 0.05 rejects the null hypothesis of normality.
• Kolmogorov-Smirnov Test with Lilliefors Correction: Evaluates cumulative distribution frequencies against theoretical distributions. Possesses lower statistical power than Shapiro-Wilk and should not be relied upon in small surgical case series.
• Sample Size Constraints in Normality Testing: In very small sample sizes (n 50,000), even clinically trivial deviations from symmetry achieve statistically significant p-values, making graphical inspection and effect size judgment paramount.

7. Classification of Summary Statistics & Reporting Standards

The following matrix summarizes the mandatory pairings of central tendency and dispersion metrics across biomedical data architectures, establishing compliance with international ICMJE and SAMPL reporting standards.

8. Inferential Precision: Standard Deviation vs. Standard Error vs. Confidence Intervals

One of the most persistent errors in published surgical literature is the confusion between standard deviation (SD) and standard error of the mean (SEM). While both metrics share identical mathematical origins and units of measurement, they address entirely distinct scientific questions.

Standard Deviation: Descriptive Parameter of the Sample

The standard deviation describes the natural biological variability and spread of individual observations within a specific clinical sample. As sample size (n) increases, the sample standard deviation does not systematically shrink; instead, it converges closer to the true population standard deviation (σ). SD must be used exclusively in descriptive statistics to describe patient cohorts (e.g., baseline age, body mass index, blood pressure).

Standard Error of the Mean: Inferential Metric of Estimate Precision

The standard error of the mean (SEM = SD / √n) is not a descriptive statistic. It is an inferential metric derived from the Central Limit Theorem that estimates the variability of hypothetical sample means if one were to repeatedly draw samples of size n from the underlying population. Because the denominator contains √n, the SEM becomes progressively smaller as sample size increases, regardless of how widely dispersed the individual patients are.

The 95% Confidence Interval (95% CI)

In modern evidence-based orthopaedics, presenting the 95% confidence interval is strongly preferred over SEM for all inferential endpoints. Calculated approximately as Mean ± 1.96 × SEM, the 95% CI provides an intuitive range that contains the true population mean in 95% of repeated sampling experiments. When comparing operative outcomes between two surgical techniques (e.g., anterior vs. posterior total hip arthroplasty), authors must present the difference in means along with its 95% CI rather than isolated p-values, directly conveying both the statistical significance and clinical magnitude of the treatment effect.

9. Clinical Research Applications & PROMs Reporting

Patient-reported outcome measures (PROMs) serve as primary endpoints across orthopaedic subspecialties, including arthroplasty, sports medicine, spine, and trauma surgery. However, PROMs exhibit unique mathematical properties that challenge simplistic statistical summarization.

Ceiling and Floor Artifacts

Instruments such as the American Shoulder and Elbow Surgeons (ASES) score, the Disabilities of the Arm, Shoulder and Hand (DASH), the Knee Injury and Osteoarthritis Outcome Score (KOOS), and the Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) possess strict mathematical lower and upper boundaries (e.g., 0 to 100). Following successful reconstructive surgery, a substantial portion of the cohort achieves scores clustering at or near the perfect ceiling (or floor in pain scores). When scores cluster against a scale boundary, the distribution is truncated and severely skewed. Calculating a parametric mean and standard deviation falsely implies that the distribution is bell-shaped and symmetric, projecting theoretical scores that exceed the instrument’s absolute boundary (e.g., Mean 94 ± 12 implies patients score 106 on a 100-point scale).

Minimally Clinically Important Difference (MCID) Alignment

When reporting PROMs improvement, authors must correlate summary shifts with the established Minimally Clinically Important Difference (MCID). The proportion of individual patients achieving the MCID threshold (e.g., 78% of cohort exceeded the 12-point ASES MCID) provides vastly superior clinical utility compared to isolated group mean shifts, explicitly identifying non-responders who were obscured within an aggregated cohort mean.

10. Common Methodological Pitfalls & Statistical Misconceptions

Methodological audits of major orthopaedic and surgical journals consistently reveal recurring errors in data reporting and summary statistic selection. The following table highlights critical pitfalls, their mathematical consequences, and mandatory corrective actions.

11. Pearls and Pitfalls

• The Back-Calculation Verification Pearl: When reviewing published manuscripts where authors improperly reported “Mean ± SEM”, multiply the reported SEM value by the square root of the cohort size (√n) to reconstruct the true sample standard deviation and reveal the true clinical dispersion.
• Symmetric Format Avoidance for Skewed Metrics: Never report non-parametric dispersion using the plus-minus sign (e.g., writing “Median ± IQR” is mathematically incorrect). The interquartile range is inherently asymmetrical; report the median followed by bracketed quartiles: Median [Q1, Q3].
• Round-Number and Precision Discipline: Report summary measures to a clinically meaningful level of precision, typically one decimal place beyond the raw measurement instrument. Reporting operative duration as “84.329 minutes” or patient age as “62.451 years” represents false precision that distracts from clinical interpretation.
• Zero-Variance Floor Recognition: In surgical complications that are rare (e.g., deep periprosthetic infection in 1 out of 200 patients), calculating a standard deviation is uninformative. Report the exact absolute count, incidence proportion, and binomial 95% confidence interval.

12. Contemporary Controversies in Data Presentation

Modern clinical orthopaedics is undergoing a significant transition from binary hypothesis testing (p < 0.05 vs. p ≥ 0.05) toward estimation statistics emphasizing effect size, biological dispersion, and confidence intervals. Editorial leadership across leading surgical journals increasingly rejects manuscripts presenting isolated p-values devoid of clinical effect estimates.

Dichotomization of Continuous Summary Measures

A contentious methodological practice involves artificially converting continuous surgical variables into dichotomous categories (e.g., dichotomizing operative blood loss into < 500 mL vs. ≥ 500 mL, or patient age into < 65 vs. ≥ 65 years). While simplifying tabular presentation, dichotomization discards substantial statistical information, reduces statistical power by up to 33%, and can create spurious statistical significance near arbitrary cutpoints. Methodological consensus strongly favors analyzing and summarizing continuous variables on their continuous scale, supplemented by non-linear modeling when threshold effects are hypothesized.

Data Transformation and Back-Transformation Challenges

When continuous data demonstrate marked positive skewness, investigators frequently apply logarithmic, square root, or Box-Cox transformations to achieve Gaussian symmetry, allowing parametric inferential testing. However, presenting statistical results on a log-transformed scale (e.g., “Mean log-serum IL-6 was 1.4 pg/mL”) prevents intuitive clinical interpretation by practicing surgeons. Investigators must back-transform log-scale means to the original scale (yielding the geometric mean) and present back-transformed confidence intervals, ensuring both biostatistical validity and clinical interpretability.

13. Evidence-Based Synthesis

Modern clinical decision-making is grounded in high-quality randomized controlled trials and prospective registry analyses, corroborated by seminal biomechanical and anatomical descriptions.

14. References

• Altman DG, Bland JM. Statistics Notes: Detecting skewness from summary information. BMJ. 1996. doi:10.1136/bmj.313.7066.1200 [Level V (Expert Biostatistical Standard): Methodological Investigation / Biostatistical Principle]
• Altman DG, Bland JM. Standard deviations and standard errors. BMJ. 2005. doi:10.1136/bmj.331.7521.903 [Level V (Expert Biostatistical Standard): Biostatistical Tutorial / Reporting Standards]
• Lang TA, Altman DG. Basic statistical reporting for articles published in Biomedical Journals: The ‘Statistical Analyses and Methods in the Published Literature’ or the SAMPL Guidelines. Int J Nurs Stud. 2015. doi:10.1016/j.ijnurstu.2014.09.006 [Level V (Methodological Consensus Guideline): Expert Consensus / SAMPL Reporting Guidelines]
• Strage KE, Parry JA, Mauffrey C. Standardizing statistics and data reporting in orthopaedic research. Eur J Orthop Surg Traumatol. 2021. doi:10.1007/s00590-020-02843-8 [Level V (Systematic Methodological Review): Systematic Methodological Review]
• Leopold SS, Porcher R. Editorial: Reporting Statistics in Abstracts in Clinical Orthopaedics and Related Research. Clin Orthop Relat Res. 2013. doi:10.1007/s11999-013-2984-9 [Level V (Journal Editorial Policy): Peer-Reviewed Editorial / Journal Policy Standard]
• Shi J, Luo D, Wan X et al. Detecting the skewness of data from the five-number summary and its application in meta-analysis. Stat Methods Med Res. 2023. doi:10.1177/09622802231172043 [Level V (Methodological Simulation & Analysis): Mathematical Simulation and Meta-Analytic Evaluation]

15. Master Class Pro-Tip

When evaluating clinical trials or preparing surgical manuscripts, treat the standard error of the mean (SEM) as an inferential metric of estimation precision, never a descriptive summary of patient variance. If an author presents baseline demographics or clinical range of motion as ‘Mean ± SEM’, recalculate the standard deviation by multiplying the SEM by the square root of the sample size. Furthermore, apply the Altman-Bland benchmark: for non-negative continuous clinical endpoints (such as operative blood loss, tourniquet time, or postoperative opioid consumption), if the reported mean is less than twice the standard deviation, the distribution is inevitably right-skewed. In such instances, parametric means distort clinical reality, and reporting the median with its 25th to 75th interquartile range is mandatory to convey true patient experience accurately.

Master Class Biostatistical Pearl: Rigorous Descriptive Reporting