Statistical analysis and Artificial Intelligence

Artificial Intelligence (AI)-related publications are becoming more frequent in the journal. Although this was not a systematic review, and was limited only to this journal, from January 2026 until July 2026 there were six publications, in press or paginated, identified by a text search of the journal website for ‘AI’ in the title. These publications covered AI for teaching in medical education [1], clinical reasoning [2], generating multiple-choice questions in medical education [3], for competency-based medical education [4], generating scientific abstracts [5], and generating peer-review reports of scientific papers [6]. Large language models and related forms of AI may revolutionize knowledge development, dissemination, and related scientific processes. One of these processes is the statistical analysis of clinical and associated data. Functions of statistical analysis include, but are not limited to, description, estimation, prediction, inference about causality, process control, and hypothesis generation. By data is meant a collection of measured values, variables, that can be ordered by individual or a hierarchy of grouping structures of individuals. Data can be conventionally described as categories (nominal), ordered categories (ordinal), interval data where the distance between values is meaningful, and ratio data where the ratio of two numbers has a meaning. This typology of data is from Stevens’. It was originally used to describe mathematical properties of these data types called ‘invariance’. By this is meant that the information associated with each variable doesn’t change when the variable undergoes a transformation. For example, for a categorical variable originally described as ‘Yes’ or ‘No’, the same information is present if it is described as ‘Red’ and ‘Green’. For ratio data, such as height, the same information is present if it is expressed by a simple transformation, e.g. height expressed as centimetres or inches. The scale of measurement is usually associated with generally recognized appropriate analysis approaches that respect this invariance. For example, in linear regression it is generally not appropriate to compare nominal variables because the coefficients reflect a change in response per unit change in predictor but there is no sense of a unit of change for categories. As an example, in a linear regression where, say, weight is predicted by height, the response variable could be expressed in kilograms or pounds. The regression coefficient will be change in weight in kilograms per unit change in predictor or change in weight in pounds per unit change in predictor, and the values of the regression coefficients will be different. However, the P-value for the regression and the amount of variability explained by the regression, the R-squared, will be the same. The same simple transformation, of 2.2 pounds per kilogram, will, when multiplied out for the regression coefficients, give the identical result. Similarly, the continuous predictor here could be measured in centimetres or inches with a simple transformation converting one to another also applying to the regression coefficients. This is an example where the statistical method respects invariance to simple transformations from one scale to another. Respect for data types, and this invariance to simple transformations, is often an inbuilt feature of statistical software. Namely, an error is reported when attempting to perform linear regression with a categorical response variable. It is possible, however, to bypass these data analysis checks. A categorical response variable can be recoded as a simple number with only two values, and linear regression will then ‘work’. For example, a response variable could be male or female sex, but coded as two simple numbers, but there is no sense of a unit change in sex, per unit change in predictor; and statistical assumptions regarding linear regression, such as normality or the error function (residuals), are unlikely to be met. The regression coefficients will not be meaningful or rendered meaningful by a simple transformation.