STATISTICS


typical performance, though the mean may still be relevant for understanding resource burden across all cases. A third seting is reference intervals. Although the present article is not a full guide to reference interval establishment, the reference interval field illustrates the core issues very clearly: laboratory data are often non-Gaussian, percentiles are frequently more useful than mean-based limits, and outlier handling decisions can alter the final interval substantially. More generally, these habits are


relevant whenever a laboratory professional is faced with a dataset and needs to decide what it is saying before reaching for formal comparative methods. That is the place of Part 1 in this series: it is about how to read data before you decide how to test hypotheses about them.


Core concepts and statistical rationale


A distribution is the patern formed by the values in a dataset. In practical terms, it describes how often values occur across the measurement range and whether they cluster symmetrically, pile up at one end, extend into a long tail, or separate into more than one group. The reason distributions mater is that they determine whether particular summaries are informative or misleading. A single average can conceal a lot. Figure 1 illustrates three common distribution paterns that may be encountered in laboratory datasets. Some laboratory datasets are approximately symmetric. In those cases, the mean often does a reasonable job of describing the centre of the data, and the standard deviation gives a useful sense of the spread around that centre. But many datasets in laboratory medicine are right-skewed. In such distributions, most observations lie toward the lower end, while a minority of larger values extend the upper tail. When this happens, the mean is pulled upwards and may no longer represent a typical value well. The median, which marks the middle ordered observation, is often more resistant to the influence of a small number of extreme


Possible explanation Data entry or transcription error


Sample or patient identification problem Pre-analytical or analytical problem Hidden subgroup or mixed population


values and may therefore be more informative.


This is not a mater of declaring the


mean bad and the median good. The two statistics answer slightly different descriptive questions. The most useful summary statistics depend on both the shape of the data and the question being asked (Table 1). The mean is sensitive to all observations and can therefore be useful if the purpose is to reflect the total burden or average magnitude across all cases, including extremes. The median is often beter when the purpose is to describe what is typical in an asymmetric dataset. The same logic extends to measures of spread. Standard deviation works naturally with the mean in roughly symmetric data, whereas the interquartile range is often a more robust summary in skewed data because it focuses on the middle half of the observations rather than the tails. Visual inspection is therefore not optional. It is one of the most important descriptive steps. A histogram, dot plot, or boxplot can reveal skewness, possible bimodality, floor or ceiling effects, gaps suggesting mixed populations, or isolated extreme observations. A formal normality test can sometimes be helpful, but it should not dominate judgement. In small samples, such tests may have too litle power to detect important departures from normality. In large samples, they may identify very slight deviations that mater litle in practice. Ploting the data and asking why they look as they do is often far more informative than treating a normality P value as a pass–fail certificate. A dataset that combines distinct


subgroups may show an apparently odd or broad distribution not because the analyte is inherently strange, but because the laboratory has pooled populations that should have been considered separately. A mixed inpatient and outpatient dataset, or a combined adult and paediatric dataset, may look more variable or less symmetric than either subgroup on its own. In such cases the problem is not that the data have ‘failed’ a normality assumption. The


Typical clue Implausible value or wrong scale Result inconsistent with patient/context


problem is that the wrong data grouping has been used. Transformation is sometimes helpful when the raw scale obscures structure or makes a particular analysis difficult. The most familiar example is log transformation for right-skewed data. A logarithmic scale can compress the upper tail and sometimes produce a more symmetric patern, making summaries or further analysis easier. But transformation should not be used automatically and should not be treated as a repair mechanism for every awkward dataset. It is a tool. If it improves interpretability and aligns with the purpose of the analysis, it may be useful. If it merely produces more convenient software output while making the results harder to explain, it is not helping. Confidence intervals belong naturally in this descriptive framework because they move the discussion from “what estimate did I get?” to “how certain am I about that estimate?” A point estimate, such as a mean, median, proportion, or percentile, is derived from a particular sample. If a different sample had been drawn from the same underlying population, the estimate would usually differ somewhat. A confidence interval reflects that sampling uncertainty. A narrow interval suggests the estimate is relatively precise; a wide interval suggests more uncertainty. The width of the interval depends mainly on two things: how much variability there is in the data and how much information the sample provides, which is strongly influenced by sample size. That logic is especially important


in laboratory setings where local datasets are often modest in size. Two laboratories may report the same median or the same percentile but with very different precision. One may have many observations and a fairly stable estimate; the other may have a smaller, more variable dataset and a much wider interval. Without the interval, those two situations look identical when they are not. Confidence intervals therefore add essential context, even when no hypothesis test is being performed.


Practical response


Check source record and correct or exclude if confirmed Review identification and specimen traceability


Unusual result with specimen or assay concerns Review sample quality, handling, QC, and analyser status Extreme result fits a different subgroup


Consider subgrouping rather than exclusion


Genuine rare biological/pathological value Unusual but clinically and analytically plausible Usually retain and interpret in context No clear cause after review


Value remains unusual but unexplained Table 2. Potential causes of outlier data, and how to spot and deal with them. 24 WWW.PATHOLOGYINPRACTICE.COM May 2026 Do not remove silently; report handling transparently


Page 1  |  Page 2  |  Page 3  |  Page 4  |  Page 5  |  Page 6  |  Page 7  |  Page 8  |  Page 9  |  Page 10  |  Page 11  |  Page 12  |  Page 13  |  Page 14  |  Page 15  |  Page 16  |  Page 17  |  Page 18  |  Page 19  |  Page 20  |  Page 21  |  Page 22  |  Page 23  |  Page 24  |  Page 25  |  Page 26  |  Page 27  |  Page 28  |  Page 29  |  Page 30  |  Page 31  |  Page 32  |  Page 33  |  Page 34  |  Page 35  |  Page 36  |  Page 37  |  Page 38  |  Page 39  |  Page 40  |  Page 41  |  Page 42  |  Page 43  |  Page 44  |  Page 45  |  Page 46  |  Page 47  |  Page 48  |  Page 49  |  Page 50  |  Page 51  |  Page 52