Navigation: <-- Why Data Work Dominates | Part Index | Main Index | Datasets -->
Data Types and Measurement Scales
Requires: Why Data Work Dominates
Motivation: Before you can summarize, visualize, or model any column in your dataset, you need to know what kind of value it holds. The same column might look like a number and yet be meaningless to average. This nugget asks: what distinguishes a valid arithmetic operation on data from a meaningless one?
You will learn to classify any attribute by its measurement scale (nominal, ordinal, interval, ratio) and by whether it is discrete, continuous, or binary — knowledge that determines which statistics and encodings are valid for each column.
Table of Contents
Measurement Scales
Every column in your dataset represents an attribute: a property of the objects you are studying. Attributes come in four measurement scales: nominal, ordinal, interval, ratio. The scale describes what operations and comparisons are meaningful for that attribute.
The four scales form a strict hierarchy: each level inherits everything below it and adds exactly one new capability.
| Scale | What's added | Permissible Operations | Valid Statistics | Example Attributes |
|---|---|---|---|---|
| Nominal | Distinctness | =, ≠ | Mode, frequency, counts | Country, sensor type, loan status |
| Ordinal | Order | =, ≠, <, > | Median, percentiles, rank correlation | Survey ratings, severity levels, grades |
| Interval | Equal differences | +, − (+comparisons) | Mean, standard deviation, Pearson correlation | Temperature (°C/°F), calendar year |
| Ratio | True zero | +, −, ×, ÷ (+comparisons) | All statistics, geometric mean, % change | Temperature (K), length, mass, voltage |
Nominal scale supports only distinctness: you can tell whether two values are the same or different. Even if we assign numbers to nominal categories (e.g., 1 = "type A", 2 = "type B"), they are just labels. Assigned numbers do not carry mathematical meaning.
Ordinal scale adds order: you can rank values. However, the gaps between ranks are not guaranteed to be equal. Survey scales ("rate your agreement 1 to 5") are ordinal. The number 4 is higher than 3, but the psychological difference between them may not equal the difference between 2 and 3.
Interval scale adds the ability to measure differences. You can say one value is 5 units more than another, and that gap is meaningful and consistent. However, the zero point is arbitrary: it is just a reference point, not "nothing." Multiplying interval values produces nonsense: the year 1000 CE is not "twice as old" as 500 CE, because the year zero is a human choice, not true zero.
Ratio scale adds a true zero, making multiplication and division meaningful. Most physical measurement quantities are ratio-scale: length, mass, voltage, time duration, angular displacement.
Common errors: Applying an operation the scale does not support produces numbers that look real but mean nothing. The most common errors: computing the mean of nominal codes (the codes are just labels), treating ordinal gaps as equal intervals (e.g., averaging a 1–5 survey scale — technically wrong, but widely done and often acceptable for exploratory work), and computing ratios on interval data (e.g., claiming 20°C is "twice as warm" as 10°C).
The scale also controls which encoding choices make sense when preparing data for a model: nominal attributes need one-hot or dummy encoding; ordinal attributes can sometimes be passed as ordered integers; interval and ratio attributes can be fed directly to numeric models, typically after scaling. You will apply these rules in 🖝 Structural Cleaning and Encoding.
Discrete, Continuous, and Binary Attributes
The measurement scale describes what operations the data supports. A complementary question is how many values the attribute can take. This distinction matters for storage, algorithm choices, and how you interpret statistics.
There are essentially three answers to the "how many values"-question:
| Domain | Typical storage | Examples |
|---|---|---|
| Binary | bool / int | Loan approved, course attended, gender indicator |
| Discrete | int | Course registration count, social tie count, defect count |
| Continuous | float | Temperature, sensor output, age computed from a birth year |
A discrete attribute has a finite or countably infinite set of possible values. Discrete attributes can be categorical or numeric: the number of close social ties in the ESS (0, 1, 2, 3...) is discrete numeric, while a country code is discrete categorical. Values are typically stored as integers.
A continuous attribute takes real-number values within a range. Temperature, height, weight, and sensor readings are continuous. In practice, measurements are recorded with limited precision, but the underlying quantity is treated as continuous. Continuous attributes are stored as floats.
A binary attribute is a special case of discrete with exactly two values: true/false, on/off, or 0 and 1. The gender column in the ESS (0 = Female, 1 = Male) is binary. Binary attributes are usually stored as integers 0 and 1, or as Boolean values.
Note: Some attributes sit in a gray area. The ESS happiness rating (0–10) is technically discrete with 11 values, but researchers often treat it as quasi-continuous to compute means and standard deviations. Knowing you are making this simplification, and checking whether it changes your conclusions, is good analytical practice.
[Source: Tan et al. (2020), Introduction to Data Mining, Pearson]
Summary
- Every column in a dataset is an attribute. Its measurement scale — nominal, ordinal, interval, or ratio — determines what operations and statistics are valid.
- The four scales form a hierarchy: each level adds one capability, from distinctness only (nominal) to full arithmetic (ratio).
- Common errors: averaging nominal codes, treating ordinal gaps as equal, or computing ratios on interval data.
- Encoding choices for models follow from the scale: nominal needs one-hot encoding, ordinal can sometimes use ordered integers, numeric attributes need scaling.
- Separately, attributes are discrete, continuous, or binary — describing how many values they can take, not what operations apply.
- Identifying attribute types is one of the first steps in any EDA workflow.
As always: Happy learning, happy life! 🫶
Navigation: <-- Why Data Work Dominates | Part Index | Main Index | Datasets -->
Script v1.4 (2026-06-10) · FGN