Is it necessary to transform nutrient variables prior to statistical analyses?

Helen Millns, Mark Woodward, Caroline Bolton-smith

Research output: Contribution to journalArticlepeer-review

25 Scopus citations


The distributions of the intakes of many nutrients are skewed, yet this is often overlooked when standard statistical analyses are applied to nutrient data. The nutrient intakes of 5,123 men and 5,236 women, recorded by food frequency questionnaire in the Scottish Heart Health Study, were transformed to achieve approximately symmetric distributions. Power transformations were chosen using letter value analyses. A letter value analysis uses selected order statistics and their position around the median to assess symmetry. The effect that each transformation had on a comparison of nutrient intakes between those with and without prevalent coronary heart disease was determined from t tests on the untransformed and transformed variable. The effects of the logarithm and square root transformation and of the optimum Box-Cox transformation were also determined, and the results were compared with the nonparametric Mann-Whitney test. The conclusion of whether or not to reject the null hypotheses often varied, depending on the transformation and test used. The nonparametric test usually gave a conclusion similar to that of the t test on the letter value-transformed data, the Box-Cox-transformed variable, and after either the logarithm or square root transformation of the data, but not always both. The results from the untransformed variable were sometimes very different. Failure to account for skewness in nutrient variables may thus lead to spurious conclusions.

Original languageEnglish (US)
Pages (from-to)251-262
Number of pages12
JournalAmerican journal of epidemiology
Issue number3
StatePublished - Feb 1 1995
Externally publishedYes


  • Models
  • Nutrition assessment
  • Statistical

ASJC Scopus subject areas

  • Epidemiology


Dive into the research topics of 'Is it necessary to transform nutrient variables prior to statistical analyses?'. Together they form a unique fingerprint.

Cite this