Mean

Mean: Concept Architecture

The mean is useful because it turns a distribution of individual outcomes into a quantity that can be compared, combined, or used in a decision model. This page starts with the ordinary arithmetic mean, then shows how weighting, uncertainty, skewed outcomes, and incomplete follow-up change its interpretation. The intended population, outcome, time horizon, and method of calculation must always accompany a reported mean.

What the arithmetic mean measures

For $n$ observed values $x_1,\ldots,x_n$ of the same quantitative variable, the sample arithmetic mean is the total divided by the number of observations. It is a descriptive statistic for the observations actually included. If those observations are used to learn about a target population, the population mean is a distinct quantity to be estimated.

$$\bar{x}=\frac{1}{n}\sum_{i=1}^{n}x_i,\qquad n>0;\qquad \mu=E[X]$$

Here $\bar{x}$ is the sample mean and $\mu$ is the expectation of a population random variable $X$, when that expectation exists. The mean retains the units of $X$: the mean of costs measured in pounds is in pounds, and the mean of years lived is in years. It also obeys $\sum_i(x_i-\bar{x})=0$, a useful calculation check, but this balance does not make the mean a typical individual outcome.

For a discrete population distribution with possible values $x_k$ and probabilities $p_k$, the population mean is $E[X]=\sum_k p_kx_k$, provided the sum is defined. For a continuous variable with density $f(x)$, it is $E[X]=\int x f(x)\,dx$ when the integral exists. These expressions define an expected value over a specified population or model; an observed average is an estimate of that target only when selection, measurement, and missingness support the intended inference.

Why a mean can differ from a typical person's value

The arithmetic mean uses every recorded value and is affected by extreme observations. The median instead splits the ordered distribution in half; for a skewed distribution the two often differ. Report both or show the distribution when the difference helps readers understand individual experience and aggregate resource use.

Consider annual costs of £100, £100, £100, £100, and £1,100 for five people. The sum is £1,500 and the mean is £300, while the median is £100. The mean correctly represents total cost per person across this five-person group; calling £300 the cost of a typical person would hide the skew. For a cohort of 1,000 people with the same outcome distribution, an illustrative aggregate cost calculation is 1,000 times £300, or £300,000; this projection additionally assumes the cohort has the same relevant distribution and the same cost definition.

An outlying value may be a genuine high-cost patient or a data error. Check the source and influence of such observations before making a justified correction; removing genuine costs solely to reduce the mean changes the estimand. Even a mathematically correct mean can mislead if costs use different price years, currencies, perspectives, follow-up periods, or accounting rules.

When observations need weights

A weighted mean gives different contributions to observations or groups. The weights must have an identified meaning, such as frequency, a survey design, or a defined standard population, and must be nonnegative with a positive total. When weights are proportional to counts, the weighted mean equals the mean that would be obtained from the correctly represented combined observations.

$$\bar{x}_w=\frac{\sum_{i=1}^{n}w_ix_i}{\sum_{i=1}^{n}w_i},\qquad w_i\geq0,\quad\sum_{i=1}^{n}w_i>0.$$

For instance, if a clinic has 20 patients with mean cost £200 and another has 80 with mean cost £500, the combined mean is $(20\times200+80\times500)/100=£440$. The unweighted average of the two clinic means is £350 and instead gives equal weight to the clinics. Either can be meaningful for a different question, so specify the unit of analysis and weighting target. Survey weights can also correct for unequal sampling probabilities, but valid uncertainty estimates must reflect the survey design; the naive unweighted standard-error formula is not automatically appropriate.

Age standardisation provides another weighted calculation: age-specific rates are combined using the same standard age distribution to aid comparison across populations with different age compositions. The result is tied to that chosen standard and is not automatically the crude rate of either population. A weighted mean cannot repair missing strata, incompatible measurements, or unrepresentative outcomes merely by choosing convenient weights.

What averaging does in health economic comparisons

Economic evaluations often compare mean costs and mean health outcomes for intervention alternatives in a defined population. If $C_A,C_B$ denote per-person total costs and $Q_A,Q_B$ per-person QALYs under alternatives $A,B$, the incremental mean outcomes are $\Delta C=E[C_A]-E[C_B]$ and $\Delta Q=E[Q_A]-E[Q_B]$. These are differences of population expectations, not differences in a selected patient's outcomes.

For example, if mean cost is £12,000 under $A$ and £10,000 under $B$, then $\Delta C=£2{,}000$ per person. If mean QALYs are 6.2 and 6.0, then $\Delta Q=0.2$ QALYs per person. The two treatment groups need comparable populations, outcome definitions, horizons, and adjustments before these numbers support a causal comparison. A decision model may use expectations over uncertain pathways; observed trial means and modeled lifetime means answer different time-horizon questions.

The mean of an individual ratio need not equal a ratio of means: in general, $E[C/Q]\ne E[C]/E[Q]$, and individual ratios may be undefined or unstable if a denominator is zero or near zero. Cost-effectiveness comparisons built from incremental mean costs and effects should therefore not be replaced by an average of patient-level cost-per-QALY ratios. Similarly, $E[C_A-C_B]=E[C_A]-E[C_B]$ by linearity when both expectations exist, but a paired-data analysis must preserve pairing in its variance and inference.

How uncertainty differs from variation among people

The standard deviation describes dispersion of individual observations around their mean, whereas a standard error describes sampling uncertainty in an estimated mean. For independent, identically distributed observations with finite variance, the familiar estimated standard error is $s/\sqrt{n}$, with $s$ calculated using the sample denominator $n-1$. It is an inferential calculation under assumptions, not a universal formula for clustered, weighted, paired, serially correlated, or selectively observed data.

$$s^2=\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2,\qquad \mathrm{SE}(\bar{x})\approx\frac{s}{\sqrt{n}},\qquad n>1.$$

Confidence intervals and sampling distributions require a design and method suited to the data. Skewed costs can have a well-defined mean, yet a small-sample normal approximation can be poor; defensible alternatives may include suitable bootstrap or model-based inference, with clustering and design features respected. A narrow interval for an observed sample mean does not remove bias from selective recruitment, incomplete capture, or an inadequate model. For a causal treatment comparison, randomization or another defensible design and correct handling of missing outcomes remain central.

Censoring, horizons, and transformations

Time-to-event outcomes require care because a right-censored person has survived at least to the censoring time, but their exact event time is unknown. Averaging only fully observed event times generally does not estimate population mean survival. Define a restricted horizon $\tau$ or specify a justified extrapolation model for a longer mean.

For nonnegative survival time $T$, the restricted mean survival time through $\tau$ equals the area under the survival curve: $\mathrm{RMST}(\tau)=E[\min(T,\tau)]=\int_0^\tau S(t)\,dt$. An unrestricted mean is $E[T]=\int_0^\infty S(t)\,dt$ only when that integral is finite; late survival may not be identifiable from short trial follow-up without assumptions. An apparent lifetime mean should state the extrapolation method, population, competing mortality, and time horizon.

Other transformations change the question. The geometric mean of positive observations, $\exp[n^{-1}\sum_i\log x_i]$, summarizes multiplicative scale and is generally not the arithmetic mean needed for an aggregate budget calculation. Back-transforming the mean of log costs gives a geometric mean, not generally $E[C]$. The harmonic mean may be relevant to certain averages of positive rates, but the numerator and denominator and exposure weights must match the underlying question. Specify the type of mean rather than saying only “average.”

Checks before reporting a mean

The following checks identify most interpretation errors without treating the arithmetic as the whole analysis. They are especially useful when a mean feeds a model or a comparison between alternatives. An analyst should document departures rather than silently substitute a different outcome.

1. Specify the estimand. Name the population, variable, unit, perspective where relevant, and observation or model horizon.
2. Inspect the data. Check missingness, censoring, duplicates, extreme values, and consistent currency and time units.
3. Choose the denominator and weights. Count the eligible observations; explain survey, frequency, or standard-population weights and the target they represent.
4. Match the design in uncertainty analysis. Account for pairing, clustering, unequal sampling, repeated observations, and uncertainty in modeled inputs as applicable.
5. Show the distribution when useful. Pair the mean with a median, spread, or plot if skew or heterogeneity changes interpretation.
6. Check the use in a decision. Keep mean totals and mean increments separate from individual experience and from ratios of individual values.

Sources and further study

The statistical references below support the definition, uncertainty calculations, and contrasts with other summaries. The institutional references explain why the chosen mean and time horizon matter when estimating costs, health outcomes, and survival for decisions.

1. NIST/SEMATECH. Measures of Location. Primary statistical reference for the arithmetic mean and comparison with other summaries.
2. NIST/SEMATECH. Confidence Limits for the Mean. Primary reference for sampling uncertainty and its assumptions.
3. NICE. Economic evaluation methodology checklist. Official guidance describing incremental comparisons of mean costs and outcomes.
4. NICE Decision Support Unit. Survival analysis for economic evaluations alongside clinical trials. Methods context for incomplete follow-up and extrapolated survival means.
5. U.S. National Center for Health Statistics. Age adjustment. Official explanation of standard-population weighting and interpretation.