Levels of Measurement

Levels of Measurement

The Four Levels

1. Nominal Scale

The weakest level. Data is placed into named categories with no meaningful order or distance between them.

Properties:

• Identity: each value belongs to a distinct category
• No order, no distance, no meaningful zero

Examples: Gender, blood type, nationality, color, product ID, political party

Valid statistics: Frequency, mode, chi-square test
Invalid: Mean, median, standard deviation

import pandas as pd
from scipy.stats import chi2_contingency

# Nominal: blood type distribution
blood_types = pd.Series(['A', 'O', 'B', 'AB', 'O', 'A', 'O', 'A', 'B', 'O'])
print("Mode:", blood_types.mode()[0])
print(blood_types.value_counts())
# Chi-square test of independence (nominal vs nominal)

2. Ordinal Scale

Categories have a meaningful order, but the intervals between categories are unknown or unequal.

Properties:

• Identity + Order
• No distance, no meaningful zero

Examples:

• Survey Likert scales (Strongly Disagree -> Strongly Agree)
• Education level (High School < Bachelor's < Master's < PhD)
• Race finishing position (1st, 2nd, 3rd)
• Socioeconomic status (Low, Middle, High)

Valid statistics: Median, IQR, percentiles, Spearman rank correlation, Mann-Whitney test
Invalid: Arithmetic mean (debated), standard deviation, Pearson r

import numpy as np
from scipy.stats import spearmanr

# Ordinal: race positions
team_a = [1, 3, 5, 7]   # positions team A finished
team_b = [2, 4, 6, 8]   # positions team B finished

# Spearman correlation (rank-based — appropriate for ordinal)
rho, p = spearmanr(team_a, team_b)
print(f"Spearman ρ = {rho:.3f}, p = {p:.4f}")

# Median is appropriate for ordinal
satisfaction = [3, 4, 2, 5, 4, 3, 4, 5, 2, 4]  # 1–5 scale
print(f"Median satisfaction: {np.median(satisfaction)}")

3. Interval Scale

Equal intervals between values, but no true zero — zero is arbitrary, not the absence of the quantity.

Properties:

• Identity + Order + Equal Intervals
• No true zero (ratios meaningless)

Examples:

• Temperature in Celsius or Fahrenheit (0°C ≠ "no temperature")
• IQ scores (IQ 0 doesn't mean no intelligence)
• Calendar years (Year 0 is arbitrary)
• Likert scales (when treated as interval — common in practice)

Valid statistics: Mean, standard deviation, Pearson r, t-tests, ANOVA
Invalid: Ratios ("twice as hot" is not meaningful in Celsius)

# Temperature conversion — shows why ratios fail for interval data
celsius_a = 20
celsius_b = 40

# It is NOT true that 40°C is "twice as hot" as 20°C
# Convert to Kelvin (ratio scale) to see why:
kelvin_a = celsius_a + 273.15  # 293.15 K
kelvin_b = celsius_b + 273.15  # 313.15 K

ratio_celsius = celsius_b / celsius_a        # 2.0 — misleading!
ratio_kelvin  = kelvin_b / kelvin_a          # 1.068 — true ratio

print(f"Celsius ratio: {ratio_celsius:.3f}  <- NOT meaningful")
print(f"Kelvin ratio:  {ratio_kelvin:.3f}  <- Meaningful thermodynamic ratio")

4. Ratio Scale

The strongest level. Has all properties of interval scale plus a true absolute zero (zero means absence of the attribute).

Properties:

• Identity + Order + Equal Intervals + True Zero

Examples:

• Height, weight, length (0 kg = no mass)
• Age, time duration
• Income (0 = no income)
• Temperature in Kelvin
• Number of items (count data)

Valid statistics: All statistics including geometric mean, coefficient of variation, and ratio comparisons.

import numpy as np

heights_m = np.array([1.65, 1.72, 1.80, 1.58, 1.90])

print(f"Mean: {np.mean(heights_m):.3f} m")
print(f"Ratio (tallest/shortest): {heights_m.max()/heights_m.min():.3f}")
print(f"Geometric mean: {np.exp(np.log(heights_m).mean()):.3f} m")
print(f"CV (coeff of variation): {(np.std(heights_m)/np.mean(heights_m)*100):.1f}%")
# All valid because height is ratio scale

Summary Table

Choosing the Right Statistical Test

def suggest_test(level_of_measurement, n_groups, paired=False):
    """Suggest appropriate statistical test based on measurement level."""
    if level_of_measurement == 'nominal':
        return "Chi-square test (categories) or Fisher's exact test (small samples)"
    elif level_of_measurement == 'ordinal':
        if n_groups == 2:
            return "Mann-Whitney U (independent) or Wilcoxon signed-rank (paired)"
        else:
            return "Kruskal-Wallis (independent) or Friedman (repeated measures)"
    elif level_of_measurement in ('interval', 'ratio'):
        if n_groups == 1:
            return "One-sample t-test"
        elif n_groups == 2:
            return "Independent t-test" if not paired else "Paired t-test"
        else:
            return "One-way ANOVA" if not paired else "Repeated measures ANOVA"

# Examples
print(suggest_test('nominal', 2))
print(suggest_test('ordinal', 2))
print(suggest_test('ratio', 2, paired=False))
print(suggest_test('ratio', 3))

Measurement Levels in Machine Learning & LLMs

Example — LLM Tokenization (Nominal Data):

# LLMs treat every token as a nominal category
# Each token gets an embedding vector (learned representation)

# Simple example of token embedding
import numpy as np

# Vocabulary: each word is a nominal category
vocab = {"the": 0, "cat": 1, "sat": 2, "on": 3, "mat": 4}

# Embedding layer maps nominal IDs to dense vectors
# (In practice, learned during training)
embedding_dim = 4
np.random.seed(42)
embeddings = np.random.randn(5, embedding_dim)  # 5 tokens × 4 dimensions

# "cat" has token ID 1
cat_embedding = embeddings[vocab["cat"]]
print(f"Token 'cat' (ID={vocab['cat']}):")
print(f"Embedding vector: {cat_embedding.round(3)}")
print(f"Vector dimension: {len(cat_embedding)}")

# Similarity between tokens (cosine similarity)
from numpy.linalg import norm
sim_cat_sat = np.dot(embeddings[1], embeddings[2]) / (norm(embeddings[1]) * norm(embeddings[2]))
print(f"\nSimilarity(cat, sat): {sim_cat_sat:.3f}")

Output:

Token 'cat' (ID=1):
Embedding vector: [ 0.497 -0.139  0.648  1.523]
Vector dimension: 4

Similarity(cat, sat): 0.234

Key Takeaways

"When in doubt, use conservative lower-level methods — they are always more robust than you think."

What to Learn Next