Chapter 60 · Case Study
Survey Demographics: the Chi-Square Test 📋
Are two categorical variables related, or independent? The chi-square test of independence compares the counts we observe against the counts we would expect by chance. Subscription tier strongly drives satisfaction, but region does not.
Statistics, Data Science and AI: A Visual Handbook · John Fisher · 2026
⚙️ Setup & data¶
In [1]:
import numpy as np, pandas as pd, matplotlib.pyplot as plt
from scipy import stats
AMBER="#d97706"; TEAL="#0d9488"; INK="#1a2138"; GRID="#e6e9f2"; PINK="#db2777"
plt.rcParams.update({"figure.facecolor":"white","axes.facecolor":"white","figure.dpi":110,"font.size":11,
"axes.edgecolor":GRID,"axes.grid":True,"grid.color":GRID,"axes.axisbelow":True,"axes.spines.top":False,
"axes.spines.right":False,"axes.titlesize":12,"axes.titleweight":"bold","legend.frameon":False})
BASE="https://raw.githubusercontent.com/johnfisher-ai/Statistics-Data-Science-AI-Visual-Book/main/data/"
try: df = pd.read_csv("../../data/ab_survey_demographics.csv")
except FileNotFoundError: df = pd.read_csv(BASE+"ab_survey_demographics.csv")
print("loaded:", df.shape)
df.head()
loaded: (1000, 4)
Out[1]:
| respondent_id | demographic_region | subscription_tier | satisfaction_level | |
|---|---|---|---|---|
| 0 | RESP_50000 | EU | Premium | High |
| 1 | RESP_50001 | APAC | Free | Low |
| 2 | RESP_50002 | NaN | Premium | High |
| 3 | RESP_50003 | EU | Free | Medium |
| 4 | RESP_50004 | EU | Free | Low |
BEAT 1 · THE CONTINGENCY TABLE
Cross-tabulate tier against satisfaction
The chi-square test works on counts, not averages. We cross-tabulate every respondent's subscription_tier against their satisfaction_level to get a contingency table, the raw material of the test.
In [2]:
ct = pd.crosstab(df.subscription_tier, df.satisfaction_level)
ct = ct[["Low","Medium","High"]] # order the satisfaction levels
print("observed counts:"); print(ct)
print(f"\ntotal respondents tabulated: {ct.values.sum()}")
observed counts: satisfaction_level Low Medium High subscription_tier Free 202 198 101 Premium 17 53 130 Standard 60 157 82 total respondents tabulated: 1000
In [3]:
fig,ax=plt.subplots(figsize=(7,3.2))
ct.plot(kind="bar",ax=ax,color=["#f59e0b","#fcd34d","#0d9488"],edgecolor="white")
ax.set_xlabel("subscription tier"); ax.set_ylabel("respondents"); ax.set_title("Satisfaction by tier: Premium skews High, Free skews Low")
ax.tick_params(axis="x",rotation=0); ax.legend(title="satisfaction")
plt.tight_layout(); plt.show()
Eyeballing the bars, Premium users pile up in High satisfaction while Free users pile up in Low. The pattern looks real, but is it more than sampling noise? The chi-square test answers exactly that.
BEAT 2 · OBSERVED vs EXPECTED
What would independence look like?
If tier and satisfaction were independent, each cell's count would be (row total times column total) / grand total. Chi-square sums the squared gaps between observed and expected, scaled by expected.
In [4]:
chi2, pval, dof, expected = stats.chi2_contingency(ct)
exp = pd.DataFrame(expected, index=ct.index, columns=ct.columns).round(1)
print("expected counts under independence:"); print(exp)
print(f"\nchi-square = {chi2:.2f}, df = (r-1)(c-1) = {dof}, p-value = {pval:.2e}")
cramers_v = np.sqrt(chi2/(ct.values.sum()*(min(ct.shape)-1)))
print(f"Cramers V (effect size) = {cramers_v:.3f}")
print("-> reject independence: tier and satisfaction ARE related" if pval<0.05 else "-> independent")
expected counts under independence: satisfaction_level Low Medium High subscription_tier Free 139.8 204.4 156.8 Premium 55.8 81.6 62.6 Standard 83.4 122.0 93.6 chi-square = 175.39, df = (r-1)(c-1) = 4, p-value = 7.28e-37 Cramers V (effect size) = 0.296 -> reject independence: tier and satisfaction ARE related
The gaps are stark: Premium / Low has only 17 respondents where independence predicts 56, and Free / Low has 202 where independence predicts 140. The chi-square statistic of 175 with df = 4 gives p ≈ 10−37, so satisfaction is overwhelmingly tied to tier. Cramer's V ≈ 0.30 says the association is moderate in strength, not just statistically detectable.
BEAT 3 · THE CONTROL: REGION
A variable that IS independent
Significance is not automatic, the test can also confirm independence. We run the same test on demographic_region against satisfaction, where we expect no real relationship.
In [5]:
ct_r = pd.crosstab(df.demographic_region, df.satisfaction_level)[["Low","Medium","High"]]
chi2r, pvalr, dofr, _ = stats.chi2_contingency(ct_r)
print("region vs satisfaction:"); print(ct_r)
print(f"\nchi-square = {chi2r:.2f}, df = {dofr}, p-value = {pvalr:.3f}")
print("-> reject independence" if pvalr<0.05 else "-> fail to reject: region and satisfaction look INDEPENDENT")
region vs satisfaction: satisfaction_level Low Medium High demographic_region APAC 73 102 75 EU 69 103 91 LATAM 72 109 63 chi-square = 4.80, df = 4, p-value = 0.309 -> fail to reject: region and satisfaction look INDEPENDENT
Here the chi-square is a tiny 4.8 with p ≈ 0.31, far above 0.05, so we fail to reject independence: satisfaction does not vary by region. This is the discipline of the test, it flags the real association (tier) and clears the spurious one (region), instead of finding a pattern in everything.
✅ Case closed
The chi-square test of independence compares observed cell counts against the counts expected if two categorical variables were unrelated. Subscription tier and satisfaction are strongly dependent (χ² = 175, p ≈ 10−37, Cramer's V ≈ 0.30); region and satisfaction are independent (χ² = 4.8, p ≈ 0.31). Counts, not means, and a chi-square that grows with every gap between what we see and what chance predicts.
Statistics, Data Science and AI: A Visual Handbook · © 2026 John Fisher