Level of Measurement

Level of measurement refers to the classification of variables into one of four hierarchical scales: nominal, ordinal, interval, and ratio. This classification, originally proposed by Stanley Smith Stevens in 1946, determines which mathematical operations are meaningful for a variable and, consequently, which statistical analyses can be appropriately applied.

The level of measurement guides the range of statistical methods available for analysis. Applying a method designed for a higher level of measurement to data measured at a lower level (for example, calculating the mean of nominal categories) produces meaningless results. Conversely, treating data at a lower level than warranted (for example, dichotomizing a continuous variable) discards information and reduces statistical power. Correctly identifying the level of measurement is one of the first steps in any data analysis plan and influences decisions about descriptive statistics, visualizations, hypothesis tests, and regression models.

The four levels form a hierarchy of increasing mathematical properties. Nominal variables consist of categories with no inherent order. Examples include blood type (A, B, AB, O), country of residence, or diagnostic category. The only meaningful operations are counting frequencies and testing for differences in proportions (using chi square tests or Fisher's exact test).

Ordinal variables have categories with a meaningful order, but the intervals between categories are not necessarily equal. Examples include pain severity (mild, moderate, severe), education level (high school, bachelor's, master's, doctorate), and Likert scale responses (strongly disagree to strongly agree). Appropriate statistics include the median, percentiles, and nonparametric tests (Mann Whitney U, Kruskal Wallis, Spearman correlation).

Interval variables have equal intervals between values but no true zero point. The classic example is temperature in Celsius or Fahrenheit: the difference between 10 and 20 degrees is the same as between 20 and 30, but 0 degrees does not mean the absence of temperature. Calendar dates are another example. Means, standard deviations, and parametric tests (t tests, ANOVA, Pearson correlation) are appropriate, but ratios are not meaningful (40°C is not "twice as hot" as 20°C).

Ratio variables have all the properties of interval variables plus a true zero point, which makes ratios meaningful. Examples include height, weight, income, and reaction time. All arithmetic operations are meaningful, so ratio data allow the statistics used at the lower levels as well as those that need a true zero, such as ratios between values.

In practice, the boundaries between levels are sometimes debated. Likert scale items are technically ordinal, but composite Likert scores (averaging multiple items) are often treated as interval data, a practice that is widely accepted when the scale has a sufficient number of points and is approximately symmetric. This pragmatic approach is supported by simulation studies showing that parametric tests are reasonably accurate with Likert type data under many conditions.

Identify the level of measurement for every variable during the study design phase and use it to guide your choice of descriptive statistics and inferential methods. When uncertain about whether ordinal data can be treated as interval, consider running analyses under both assumptions and comparing results. The level of measurement also affects data visualization: bar charts and pie charts are appropriate for nominal data, ordered bar charts and box plots for ordinal data, and histograms, scatter plots, and line graphs for interval and ratio data.

A common error is applying the mean to ordinal data (for example, reporting the mean of a single Likert item) without acknowledging that the intervals between categories may not be equal. Another mistake is treating a numeric coding scheme as evidence of interval measurement; assigning the numbers 1, 2, and 3 to "mild," "moderate," and "severe" does not make the variable interval scale. Some researchers also dichotomize continuous variables (for example, converting BMI to "overweight" versus "not overweight") without recognizing the substantial loss of information and statistical power this entails. Finally, failing to distinguish between continuous and discrete variables within the ratio level can lead to the selection of inappropriate probability distributions for regression models.

In a study of patient satisfaction with hospital care, the researcher administers a questionnaire with five Likert items (each scored 1 to 5) and one open ended item asking patients to name their hospital. The Likert items are ordinal, but the composite satisfaction score (the sum of the five items, ranging from 5 to 25) is treated as interval for the purpose of regression analysis. The hospital name is nominal. Age (in years) and length of stay (in days) are ratio variables. The researcher uses Pearson correlation to examine the association between satisfaction score and length of stay, and ANOVA to compare satisfaction scores across hospitals.