Descriptive Statistics

Classification of Data and Variables in Anaesthetic Research

Correct statistical analysis begins with the measurement scale of the variable, not with the software menu. In anaesthetic research, many familiar observations look numerical but are not necessarily quantitative in the mathematical sense. The key distinction is between qualitative variables, which classify patients or events, and quantitative variables, which measure magnitude on a defined numerical scale.

Mathematical taxonomy of variables

Scale	Defining property	Permissible summaries	Anaesthetic examples
Nominal	Categories with no intrinsic order	Counts, proportions, mode, risk ratio, odds ratio	Airway device: tracheal tube, supraglottic airway, facemask; volatile agent: sevoflurane, desflurane, isoflurane
Binary nominal	Two unordered categories	Proportion, sensitivity, specificity, likelihood ratios	Difficult intubation: yes/no; postoperative nausea and vomiting within 24 h: yes/no; mortality: alive/dead
Ordinal	Ordered categories, but unequal or unknown spacing	Median, IQR, cumulative proportions, rank-based or ordinal models	ASA physical status I–VI; Mallampati class I–IV; Ramsay sedation score 1–6; Bromage motor block 0–3
Discrete quantitative	Numerical counts, usually integers, true zero	Mean if distribution appropriate, rate, Poisson/negative binomial models	Number of laryngoscopy attempts; vasopressor boluses; PACU vomiting episodes; red-cell units transfused
Continuous interval	Equal intervals, arbitrary zero; ratios not meaningful	Mean, SD, regression if assumptions met	Temperature in degrees Celsius; calendar time; some transformed scale scores
Continuous ratio	Equal intervals and non-arbitrary zero; ratios meaningful	Mean, SD, coefficient of variation, log-transformed models	Cardiac output in L min−1; plasma propofol concentration in mg L−1; arterial pressure in mmHg; MAC fraction; temperature in kelvin

Clinical translation in anaesthetic measurement

ASA physical status is ordinal, despite being labelled I to VI. ASA I denotes a normal healthy patient; ASA II mild systemic disease; ASA III severe systemic disease; ASA IV severe disease that is a constant threat to life; ASA V moribund; ASA VI brain-dead organ donor, with emergency status denoted by suffix E. The numerical distance between ASA II and III is not equivalent to that between IV and V, and ASA IV is not “twice” ASA II. It should not be averaged as if it were a haemoglobin concentration.

Mallampati class is similarly ordinal. Classes I–IV reflect progressive obscuration of faucial pillars, uvula and soft palate and are associated with difficult laryngoscopy, but diagnostic performance is modest when used alone; meta-analyses report sensitivities commonly around 35–60% with specificities around 70–85%, depending on the definition of difficult airway. Its scale encodes ordered anatomical visibility, not linear airway difficulty.

MAC is a continuous ratio pharmacodynamic measure when expressed as alveolar partial pressure or volume percent at 1 atmosphere. MAC is the end-tidal concentration preventing movement to surgical incision in 50% of subjects: approximately sevoflurane 2.0%, isoflurane 1.15%, desflurane 6.0% and nitrous oxide 104% in a healthy 40-year-old adult. MAC decreases by roughly 6% per decade after age 40 and is reduced by opioids, benzodiazepines, pregnancy and hypothermia. Because MAC is a concentration with a meaningful zero, 1.0 MAC is twice 0.5 MAC as a dose metric, although the clinical response is sigmoid rather than linear.

Cardiac output, normally about 4–8 L min⁻¹ in adults with a cardiac index of approximately 2.5–4.0 L min⁻¹ m⁻², is a continuous ratio variable: zero flow is physiologically meaningful and ratios are interpretable. Temperature is more subtle. Core temperature is an intensive thermodynamic state variable. Measured in degrees Celsius, it is interval scale: 38°C is 1°C above 37°C, but not “twice as hot” as 19°C because zero Celsius is arbitrary. In kelvin, temperature is ratio scale because absolute zero is 0 K. Perioperative hypothermia is commonly defined as core temperature <36.0°C, consistent with NICE guidance and major perioperative warming recommendations.

Pain scores illustrate the commonest error. A 0–10 numerical rating scale or 0–100 mm visual analogue scale is bounded and anchored by “no pain” and “worst imaginable pain”, but equal spacing is not guaranteed, and 8/10 is not necessarily twice the pain intensity of 4/10. The same applies to sedation scores such as Ramsay 1–6, Richmond Agitation-Sedation Scale −5 to +4, or modified Observer’s Assessment of Alertness/Sedation. These are ordered clinical states generated by neurobehavioural thresholds, not continuous measurements of cortical arousal.

Why misclassification matters

Treating ordinal outcomes as continuous interval data permits convenient means, standard deviations, t tests and linear regression, but it imposes assumptions that are usually untrue: equal category spacing, normally distributed residuals, homoscedasticity and linear treatment effects. A trial reporting mean Ramsay sedation score 3.7 versus 3.2 implies fractional and equally spaced sedation states, which have no clinical existence. More defensible approaches include medians with interquartile ranges, cumulative category plots, Mann–Whitney U tests for simple comparisons, and ordinal logistic regression when modelling covariates. Dichotomisation, such as pain ≥4/10 versus <4/10, may aid clinical interpretation but discards information and should be prespecified. For Primary FRCA statistics, the examination point is precise: the numerical appearance of a variable does not determine its statistical scale; the measurement process does.

Mathematical Descriptors of Central Tendency and Dispersion

Descriptive statistics reduce a sample x₁, x₂, …, x_n to interpretable summaries. In anaesthetic research and audit, the choice of descriptor is not cosmetic: it determines whether reported haemodynamic responses, drug concentrations, length of stay, or recovery times are clinically intelligible and mathematically defensible.

Central tendency

Descriptor	Definition	Key property	Typical clinical use
Arithmetic mean	x̄ = Σxi/n	Minimises Σ(xi − a)2	Approximately Gaussian variables: haemoglobin, PaCO2, systolic BP in stable cohorts
Median	Middle ordered value; for even n, average of two central values	Minimises Σ|xi − a|	Skewed outcomes: ICU length of stay, blood loss, time to extubation
Mode	Most frequent value or class	Applicable to nominal or discrete data; may be multimodal	ASA physical status class, airway grade, pain score category
Geometric mean	GM = (Πxi)1/n = exp[Σln(xi)/n], for xi > 0	Mean on the logarithmic scale; appropriate for multiplicative variation	Drug concentrations, biological half-lives, ratios, titres

The arithmetic mean has an important least-squares proof. Define S(a) = Σ(x_i − a)². Differentiating gives dS/da = −2Σ(x_i − a). Setting this to zero gives a = Σx_i/n = x̄. Thus the mean is the unique centre that minimises squared error, explaining its central role in linear models, ANOVA, and normally distributed measurement error.

The median is a rank-based estimator. For absolute loss L(a) = Σ|x_i − a|, moving a below the median leaves more observations above than below and reduces L by increasing a; moving above the median does the opposite. Any median therefore minimises absolute deviation. This robustness is why median and interquartile range are preferred for non-normal peri-operative data such as postoperative opioid consumption, where a few patients receiving very large morphine-equivalent doses may distort the mean.

The geometric mean is crucial when data are log-normal. If X is log-normal, ln(X) ~ N(μ, σ²), then the geometric mean estimates exp(μ), the median of X, whereas the arithmetic mean is exp(μ + σ²/2) and is pulled upward by right skew. Biological half-lives and plasma concentrations often vary multiplicatively because clearance, volume of distribution, enzyme activity, organ blood flow, and protein binding combine as ratios. For example, morphine elimination half-life is commonly about 2–3 h in healthy adults but is prolonged in renal failure through active metabolite accumulation; propofol has a short distribution half-life of approximately 2–8 min but a terminal half-life of several hours; remifentanil has an effect-site equilibration half-time around 1–1.5 min and a context-sensitive half-time of about 3–5 min. When comparing concentrations or half-lives across patients, log transformation followed by geometric mean reporting is usually more defensible than using an arithmetic mean.

Dispersion

Dispersion describes spread around a centre. The range is max(x) − min(x). It is intuitive but extremely sensitive to outliers and sample size. The interquartile range is Q₃ − Q₁, the width of the middle 50% of ordered observations. It is robust and is conventionally paired with the median: for example, time to discharge from PACU might be reported as median 75 min, IQR 55–110 min, rather than mean ± SD if skewed.

The population variance is σ² = Σ(x_i − μ)²/N. The sample variance is s² = Σ(x_i − x̄)²/(n − 1). The denominator n − 1 is Bessel’s correction: because x̄ is estimated from the same data, the deviations must sum to zero, leaving only n − 1 independent degrees of freedom. Dividing by n would systematically underestimate the population variance.

Squared deviations are used for several reasons. First, simple deviations sum to zero around the mean. Secondly, squaring gives non-negative contributions and penalises larger errors more heavily, which is appropriate for many measurement-error models. Thirdly, the squared-error function is differentiable, enabling tractable estimation. Fourthly, variances of independent variables add: Var(X + Y) = Var(X) + Var(Y) when Cov(X,Y) = 0. This additivity underpins standard errors, confidence intervals, meta-analysis, and power calculations.

The standard deviation is s = √s². Its key clinical advantage is dimensional homogeneity: if arterial pressure is measured in mmHg, variance has units mmHg², which are physiologically awkward, whereas SD is again in mmHg. Thus one may report mean arterial pressure 78 ± 12 mmHg, haemoglobin 132 ± 14 g/L, or PaCO₂ 5.3 ± 0.5 kPa. For normally distributed data, approximately 68% lie within 1 SD, 95% within 1.96 SD, and 99.7% within 3 SD of the mean. These empirical rules are invalid for markedly skewed data, where median, IQR, and possibly geometric mean should replace mean ± SD.

The Normal Distribution, Central Limit Theorem, and Sampling Theory

The Gaussian probability density function

The normal, or Gaussian, distribution is the central continuous probability model used in inferential statistics. A variable X is normally distributed if its probability density function is:

f(x) = 1 / (σ√2π) × exp[-(x - μ)² / 2σ²]

It is fully defined by two parameters: the mean μ, which determines location, and the variance σ², whose square root is the standard deviation σ. The curve is unimodal, bell-shaped, and exactly symmetrical about μ; therefore, for a true normal distribution, mean = median = mode. Its tails approach but never touch the x-axis, hence the distribution is asymptotic. This is important in examinations: although extreme values are increasingly improbable, they are not impossible under a normal model.

The standard normal distribution has μ = 0 and σ = 1. Any normally distributed value can be transformed to a standard normal deviate, or z-score:

z = (x - μ) / σ

This expresses how many standard deviations an observation lies from the mean. For example, if adult haemoglobin is modelled as normally distributed with mean 150 g/L and SD 15 g/L, a value of 120 g/L has z = -2.0.

The empirical rule

The familiar 68-95-99.7 rule is not arbitrary; it is obtained by integrating the standard normal density between z-boundaries. The area under the curve represents probability, with total area equal to 1. Thus:

Interval	Probability within interval	Probability outside interval	Approximate clinical interpretation
μ ± 1 SD	68.27%	31.73%	Common central spread of individual observations
μ ± 1.96 SD	95.00%	5.00%	Two-sided 5% reference or test boundary
μ ± 2 SD	95.45%	4.55%	Often used as an approximate 95% range
μ ± 3 SD	99.73%	0.27%	Extreme observations; approximately 3 per 1000 outside

For a normally distributed biological variable, an approximate 95% reference interval is therefore mean ± 1.96 SD, not mean ± 1.96 SEM. This distinction is a common Primary FRCA viva trap.

Central Limit Theorem and sampling theory

The Central Limit Theorem states that, for sufficiently large independent random samples drawn from a population with finite mean and variance, the distribution of the sample means tends towards a normal distribution, irrespective of the shape of the parent population. This distribution is the sampling distribution of the mean. Its mean equals the population mean μ, and its standard deviation is the standard error of the mean:

SEM = SD / √n

The CLT is the reason parametric inference is often robust even when individual patient data are not perfectly normal, provided observations are independent and the sample size is adequate. A rough examination rule is that n ≥ 30 often gives a reasonably normal sampling distribution, but this is not a law: highly skewed data, heavy-tailed distributions, clustering, repeated measures, and small samples may invalidate simple normal approximations.

SD versus SEM: the examination-critical distinction

The standard deviation describes variability among individual observations in a sample. It answers: how dispersed are the patients? The standard error of the mean describes uncertainty in the sample mean as an estimate of the population mean. It answers: how precisely have we estimated the mean? SEM decreases as sample size increases, whereas SD does not necessarily decrease with larger samples because biological variability remains.

Statistic	Formula	Meaning	Use in reporting
SD	√[Σ(x - x̄)2 / (n - 1)]	Spread of individual observations	Describe baseline age, weight, haemodynamic values, PaCO2, drug concentrations
SEM	SD / √n	Precision of the estimated mean	Construct confidence intervals; compare estimated means
95% CI for mean	x̄ ± 1.96 SEM, approximately	Range likely to contain the true population mean in repeated sampling	Inferential reporting of treatment effect or mean difference

Example: postoperative morphine consumption after major abdominal surgery is 30 mg with SD 12 mg in 64 patients. The SEM is 12/√64 = 1.5 mg. The 95% confidence interval for the mean is approximately 30 ± 1.96 × 1.5 = 27.1 to 32.9 mg. Reporting 30 ± 1.5 mg as if it described patient variability is misleading; most individual patients are not clustered between 28.5 and 31.5 mg. The appropriate descriptive summary for normally distributed individual consumption is mean 30 mg, SD 12 mg.

Clinically, report SD when describing variability among patients: baseline systolic blood pressure 132 ± 18 mmHg, age 64 ± 9 years, or plasma propofol concentration 3.2 ± 0.8 µg/mL. Report SEM or, preferably, 95% CI when expressing precision of an estimate: mean reduction in systolic blood pressure 12 mmHg, 95% CI 7 to 17 mmHg. In modern trial reporting, confidence intervals are preferred to SEM because they communicate both precision and clinical compatibility of effect sizes.

Non-Normal Distributions, Skewness, Kurtosis, and Transformations

Many clinically important anaesthetic and critical care variables are not Gaussian. Examples include ICU length of stay, perioperative blood loss, plasma cytokine concentrations, time to extubation, vasopressor dose, and hospital cost. These data are often bounded at zero and have a long tail of unusually large observations. For the Primary FRCA, the key consequence is that the mean and standard deviation may be misleading; the median, interquartile range, centiles, or a suitable transformation may better describe the sample.

Skewness

Skewness describes asymmetry around the centre of a distribution. In a perfectly normal distribution, skewness is 0 and the mean, median, and mode coincide. In skewed distributions, these measures separate in predictable ways because the mean is pulled towards the tail by extreme values.

Distribution	Tail direction	Typical clinical example	Relative position	Interpretation
Positive skew / right skew	Long tail to the right	ICU length of stay, blood loss, morphine consumption, postoperative troponin rise	Mode < median < mean	Few high values inflate the mean
Negative skew / left skew	Long tail to the left	Ceiling effects in high test scores, oxygen saturation in a well-oxygenated cohort	Mean < median < mode	Few low values pull the mean down
Symmetrical	Equal tails	Adult height, many measurement errors	Mean = median = mode	Gaussian methods often appropriate

A practical measure is Pearson’s coefficient of skewness. The first coefficient is:

Skewness = (mean − mode) / standard deviation

Because the mode is unstable in small samples, the second Pearson coefficient is commonly used:

Skewness = 3 × (mean − median) / standard deviation

Approximate interpretation is: 0 = no skew; ±0.5 = mild skew; ±1 = moderate skew; >±1 = substantial skew. These are descriptive conventions rather than biological thresholds. In exams, remember that a positively skewed dataset can have a clinically implausible mean if quoted alone: for example, most patients may leave ICU within 24–48 hours, but a few prolonged stays of 30–60 days can raise the arithmetic mean substantially above the typical patient experience.

Kurtosis

Kurtosis describes the weight of the tails and the sharpness of the peak relative to a normal distribution. It is particularly relevant to anaesthesia because rare but extreme events may dominate safety planning: massive haemorrhage, malignant hyperthermia, anaphylaxis, failed airway rescue, or prolonged ventilation. Modern statistical usage often refers to excess kurtosis, where the normal distribution has a value of 0; older usage assigns the normal distribution a kurtosis of 3.

Type	Excess kurtosis	Shape	Clinical implication
Mesokurtic	0	Normal-like tails	Extreme observations occur at Gaussian frequency
Leptokurtic	>0	Sharper peak, heavier tails	More outliers than expected; risk models must account for rare catastrophic events
Platykurtic	<0	Flatter peak, lighter tails	Fewer extreme values; observations more evenly dispersed around the centre

A leptokurtic distribution may have an apparently acceptable mean and variance yet still contain clinically important extremes. For example, transfusion requirement after major surgery is frequently zero or low for most patients but can include massive transfusion outliers. Descriptive statistics should therefore include tail-sensitive summaries such as the 90th or 95th centile when planning resources.

Transformations for skewed data

When parametric analysis is desired, skewed data may be transformed to approximate normality. The transformation is applied before analysis; results may then be back-transformed for clinical interpretation. This is not a cosmetic manoeuvre: it changes the scale on which differences are considered additive.

Transformation	Best suited to	Effect	Clinical examples
Logarithmic: y′ = log(y)	Positive, right-skewed data; multiplicative effects	Compresses large values more than small values	ICU stay, drug concentrations, C-reactive protein, cost
Square-root: y′ = √y	Counts and mild positive skew	Moderate compression of high values	Number of desaturation episodes, arrhythmia counts
Reciprocal: y′ = 1/y	Severe right skew; rates or inverse-time relationships	Strong compression; reverses order of magnitude	Time-based variables, clearance-type measures

Log transformation is the commonest in medical statistics. A log-normal variable has normally distributed logarithms; the appropriate summary is the geometric mean, not the arithmetic mean. If the mean of log-transformed values is back-transformed, it estimates the geometric mean. Ratios are then more natural than absolute differences: a treatment effect may be expressed as a 20% reduction rather than a 2-day reduction.

Important limitations apply. Zero or negative values cannot be directly log-transformed; analysts may add a constant, such as log(y + 1), for count data, but this must be pre-specified and reported. Reciprocal transformations can be difficult to interpret clinically and may exaggerate differences among small values. Transformations should be assessed using histograms, Q–Q plots, and residual plots rather than by mechanical reliance on a p value from a normality test.

Non-parametric description

If data remain non-normal, contain outliers, or are ordinal, it is often preferable to present median and interquartile range rather than mean and standard deviation. For instance, blood loss might be reported as median 350 ml, IQR 150–900 ml, rather than mean 820 ml with a very large SD driven by a few massive haemorrhages. ICU length of stay is similarly best summarised by median, IQR, and selected centiles. The corresponding inferential methods are non-parametric tests such as the Mann–Whitney U test, Wilcoxon signed-rank test, or Kruskal–Wallis test, depending on study design. The central exam principle is: describe the distribution before choosing the summary statistic.