Data Types & Measurement Scales: Choose the Right Statistical Test in Interviews
A product team sees two signals after a checkout redesign: conversion improved, but customer rating slipped from 4 to 3.5. If they treat both numbers the same, they may approve a feature that earns revenue but damages trust - because conversion is a ratio measure, while a rating is usually ordinal.
- Data type decides the test. First ask: is the variable categorical or numerical?
- Measurement scale adds precision. Nominal means names, ordinal means order, interval means equal gaps, ratio means equal gaps plus a true zero.
- Nominal data usually leads to counts, proportions, chi-square tests or logistic models.
- Ordinal data should usually use medians, ranks, Mann-Whitney U, Wilcoxon or Spearman correlation.
- Interval and ratio data can often use means, standard deviation, t-tests, ANOVA, Pearson correlation and regression, if assumptions are reasonable.
- The dependent variable drives the test. The outcome variable matters more than the input variable.
- Most candidate mistakes happen with Likert scales. A single 1-5 rating is ordinal, not automatically interval.
The clean mental model is a ladder: as you move from nominal to ratio, you are allowed to do more mathematics. The mistake is not “forgetting a test name”; the mistake is choosing a test before respecting the scale of the outcome variable.
Core Explanation: Data Type First, Test Second
Data type tells you the broad nature of a variable. Categorical data puts observations into groups; numerical data measures quantity. Measurement scale then tells you what operations are legitimate.
For interviews, use this simple sentence: “I will identify the dependent variable’s measurement scale, then choose a test based on whether I am comparing groups, testing association, or predicting an outcome.”
The Test-Selection Flow You Should Memorise
The dependent variable is the outcome you are trying to explain, compare or predict. If the outcome is categorical, you are usually in the world of proportions and classification. If it is numerical, you are usually in the world of means, variance and regression.
Definitions You Can Say in One Breath
Stevens: “Measurement is the assignment of numerals to objects or events according to rules.”
- Nominal scale: Labels categories without order, such as city, brand, payment mode or customer segment.
- Ordinal scale: Orders categories, but differences between adjacent levels are not guaranteed equal.
- Interval scale: Has equal intervals, but zero is arbitrary and does not mean absence.
- Ratio scale: Has equal intervals and a true zero, so ratios like “twice as much” are meaningful.
- Parametric test: A test that assumes a specific population form or parameter, commonly normality of numerical outcomes.
- Non-parametric test: A test that relies less on distribution assumptions, often using ranks or counts.
Common Business Variables and the Test They Point To
Do not memorise 50 tests. Memorise the mapping below and you will handle most placement interview cases.
What to Report After Choosing the Test
A good analyst does not stop at “p-value is significant.” Report statistical strength and business relevance together.
Worked Example: Why AOV Needs a t-Test, Not Chi-Square
Suppose an e-commerce team tests a new checkout page. The outcome is average order value, measured in rupees. Because AOV has a meaningful zero and ratios are meaningful, it is a ratio-scale numerical variable.
The right instinct is a two-sample t-test, assuming independent users and no severe distribution issue. The approximate test statistic is:
t = (880 - 820) / √[(200² / 100) + (220² / 100)] = 60 / 29.73 ≈ 2.02
With roughly 198 degrees of freedom, this is around the 5% significance boundary. The interview-quality answer is not “launch immediately.” It is: “The AOV lift looks statistically borderline and commercially positive; I would also check conversion, margin, refund rate and whether the result holds across customer segments.”
Case Study: Ather Energy and the Measurement-Scale Mindset
Ather Energy shows how a connected product creates many data types at once - and why each one needs a different analytical treatment.
Situation: Electric scooters are not just vehicles; they are connected devices. Ather can learn from ride behaviour, charging patterns, service diagnostics, app interactions and customer feedback. But those signals are not statistically equal.
The move: The analytical advantage comes chiefly from connected vehicle telemetry - battery, distance, ride mode, charging and usage signals. Supporting drivers include app integration, service-centre diagnostics, software updates and structured customer feedback. Together, they allow the company to separate engineering questions from customer-experience questions.
Outcome or lesson: The business value is not “Ather has data.” The value is knowing which data can support means and regression, which data should be treated as counts or ranks, and which signals should be combined carefully. A wrong test can turn telemetry into false confidence; the right scale turns it into better product, service and customer decisions.
A Simple Visual Map: Scale to Statistic to Test
This is the revision map to keep in your head. The moment you classify the scale, the summary statistic and test options narrow naturally.
How AI Changes Data Types & Measurement Scales
AI does not remove the need to understand measurement scales. It makes the first pass faster - and makes undetected mistakes easier to scale.
- Automatic schema detection: Modern analytics tools can infer whether a column is categorical, date-time or numerical. The risk is misclassification: customer IDs often look numeric but are nominal labels.
- Natural-language analytics: GenBI tools can answer “Does AOV differ by campaign?” and suggest a test. You still need to verify whether AOV is skewed, whether samples are independent and whether the outcome is truly numerical.
- AI-assisted experimentation: AI can generate experiment plans, power calculations and segmentation hypotheses. The danger is treating ordinal survey responses as precise interval data without justification.
Use ChatGPT or Claude like a test-selection coach: paste a dataset dictionary and ask, “Classify each variable as nominal, ordinal, interval or ratio; identify the dependent variable; suggest valid tests and assumptions to check.” Then challenge every suggestion manually.
Interview Relevance
“You are analysing whether a new app design improved customer satisfaction and repeat purchase. What data types are involved, and which statistical tests would you use?”
If you are unsure between two tests, say your assumption out loud: “If I can justify treating the aggregated satisfaction score as approximately interval, I may use a t-test; otherwise I will use a rank-based test.” That sounds analytical, not confused.
Common Mistake
Treating every number as quantitative. PIN codes, customer IDs and jersey numbers are nominal labels; a 1-5 Likert item is ordinal. This costs candidates because it leads to wrong averages, wrong correlations and wrong tests. One-line fix: before calculating anything, ask, “Are equal gaps and true zero meaningful here?”
What to Revise Next
Once you can classify variables and choose tests, move to the data-quality layer. Revise these next as a journey from “What is this variable?” to “Can I trust this dataset?”