Data Analysis

Data Analysis Patterns

Analysis Workflow

The Analysis Pipeline

1. Define Question — what decision does this answer?
2. Collect Data — identify sources, assess quality
3. Clean Data — handle nulls, outliers, types
4. Explore (EDA) — distributions, correlations, anomalies
5. Analyze — apply appropriate statistical method
6. Visualize — make findings legible
7. Interpret — connect back to business question
8. Communicate — executive summary + supporting detail

Descriptive Statistics

Key Metrics

EDA Checklist

import pandas as pd
import matplotlib.pyplot as plt

df = pd.read_csv('data.csv')

# Shape and types
print(df.shape)
print(df.dtypes)
print(df.head())

# Missing values
print(df.isnull().sum() / len(df) * 100)  # % missing

# Distributions
df.describe()  # count, mean, std, min, quartiles, max

# Correlations
df.corr()

# Categorical counts
df['category_col'].value_counts()

# Outliers (IQR method)
Q1 = df['value'].quantile(0.25)
Q3 = df['value'].quantile(0.75)
IQR = Q3 - Q1
outliers = df[(df['value'] < Q1 - 1.5*IQR) | (df['value'] > Q3 + 1.5*IQR)]

Hypothesis Testing

Choosing the Right Test

Question type                          → Test
─────────────────────────────────────────────────────
Compare means, 2 groups, normal        → t-test (independent)
Compare means, 2 groups, paired        → paired t-test
Compare means, 3+ groups               → ANOVA
Compare proportions, 2 groups          → z-test for proportions
Association between 2 categorical      → Chi-square test
Correlation, continuous                → Pearson (normal) / Spearman (non-normal)
Non-normal, 2 groups                   → Mann-Whitney U
Non-normal, 3+ groups                  → Kruskal-Wallis

Hypothesis Testing Framework

from scipy import stats

# 1. State hypotheses
# H0: no difference (null)
# H1: there is a difference (alternative)

# 2. Set significance level
alpha = 0.05  # 5% false positive rate

# 3. Run test
# Two-sample t-test
t_stat, p_value = stats.ttest_ind(group_a, group_b)

# 4. Interpret
if p_value < alpha:
    print(f"Reject H0: p={p_value:.4f} < alpha={alpha}")
    print("Statistically significant difference found")
else:
    print(f"Fail to reject H0: p={p_value:.4f} >= alpha={alpha}")
    print("No statistically significant difference")

# 5. Effect size (practical significance)
cohens_d = (group_a.mean() - group_b.mean()) / 
           ((group_a.std()**2 + group_b.std()**2) / 2)**0.5
# d < 0.2: small, 0.2-0.8: medium, > 0.8: large

A/B Test Analysis

Pre-Test Planning

from statsmodels.stats.power import NormalIndPower

# Calculate required sample size
analysis = NormalIndPower()
n = analysis.solve_power(
    effect_size=0.1,    # minimum detectable effect (10% relative lift)
    alpha=0.05,         # significance level
    power=0.8,          # 80% power (chance of detecting real effect)
    alternative='two-sided'
)
print(f"Required n per group: {n:.0f}")

A/B Test Results Template

Experiment: [Name]
Hypothesis: [Changing X will increase Y by Z%]
Start: [Date]  End: [Date]
Traffic split: 50/50

Results:
                Control    Treatment   Δ        p-value
Conversion rate  3.2%       3.8%      +18.75%   0.031
Revenue/visitor  $1.24      $1.41     +13.7%    0.047
Sample size      12,400     12,380    —         —

Statistical significance: YES (p < 0.05)
Practical significance: YES (18.75% lift exceeds 10% MDE)
Recommendation: SHIP treatment

A/B Test Pitfalls

• Peeking problem — don't stop early when you see significance; use sequential testing
• Multiple testing — each additional metric tested inflates false positives; use Bonferroni correction
• Novelty effect — new features get spike in engagement; run for 2+ weeks
• Sample ratio mismatch — if actual split differs from intended (e.g., 52/48 instead of 50/50), investigate before analyzing
• Segment imbalance — ensure groups are balanced on key covariates

Regression Analysis

Linear Regression

import statsmodels.api as sm

X = df[['feature1', 'feature2', 'feature3']]
y = df['target']

X = sm.add_constant(X)
model = sm.OLS(y, X).fit()

print(model.summary())
# Key outputs:
# R-squared: % variance explained
# Coef: effect of each feature on target
# p-value: significance of each feature
# Conf Int: uncertainty range

Logistic Regression (binary outcome)

from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report, roc_auc_score

model = LogisticRegression()
model.fit(X_train, y_train)

y_pred = model.predict(X_test)
print(classification_report(y_test, y_pred))
print(f"AUC-ROC: {roc_auc_score(y_test, model.predict_proba(X_test)[:,1]):.3f}")

Data Visualization Principles

Chart Selection Guide

Design Principles

• One message per chart — write the insight as the chart title
• Remove chartjunk — eliminate gridlines, borders, 3D effects
• Label directly — avoid legends that require eye movement
• Color sparingly — highlight only what matters; gray everything else
• Sort meaningfully — bar charts by value (not alphabetically) unless category order matters

import matplotlib.pyplot as plt
import seaborn as sns

# Clean chart template
fig, ax = plt.subplots(figsize=(10, 6))
sns.set_style("whitegrid")

# Bar chart with direct labels
bars = ax.barh(categories, values, color=['#2196F3' if v == max(values) else '#BDBDBD' for v in values])
ax.set_title("Conversion Rate by Channel", fontsize=14, fontweight='bold', loc='left')
ax.set_xlabel("Conversion Rate (%)")
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)

# Direct labels
for bar, val in zip(bars, values):
    ax.text(bar.get_width() + 0.1, bar.get_y() + bar.get_height()/2, 
            f'{val:.1f}%', va='center')

Communicating Findings

Insight Communication Framework (SCQA)

• Situation — what is the context?
• Complication — what changed or is the problem?
• Question — what question does this raise?
• Answer — your finding and recommendation

Executive Summary Formula

Finding: [X happened / X is true]
So what: [This means Y for the business]
Action: [We should do Z because of this]
Evidence: [p-value / % lift / confidence interval]