Back to Projects
Statistical Analysis

Statistical Hypothesis Testing Framework

A rigorous statistical analysis framework demonstrating the application of parametric and non-parametric hypothesis tests to real organisational data, with particular focus on the critical distinction between correlation and causation.

Type
Statistical Analysis
Domain
Organisational Data
Methods
Parametric & Non-Parametric
Status
Completed

The Challenge

Organisations routinely misinterpret data by conflating correlation with causation, leading to flawed strategic decisions. A marketing team might attribute a sales increase to a recent campaign when seasonal effects or external factors were the actual drivers.

Without rigorous statistical validation, data-driven decisions are built on sand. Hypothesis testing provides the framework to distinguish genuine effects from noise, correlation from causation, and significant findings from statistical artefacts.

Approach

01
Problem Framing
Defined clear null and alternative hypotheses for each organisational data scenario, ensuring the statistical questions were well-specified before any analysis began.
02
Test Selection
Selected appropriate statistical tests based on data characteristics: normality, sample size, variance homogeneity, and measurement scale. Applied both parametric (t-tests, ANOVA) and non-parametric alternatives (Mann-Whitney, Kruskal-Wallis).
03
Analysis and Interpretation
Executed tests with proper significance levels, effect size calculations, and power analysis. Interpreted results in context, distinguishing statistical significance from practical significance.
04
Reporting
Produced clear, accessible reports translating statistical findings into business language, with explicit discussion of limitations and the correlation-causation distinction.
HYPOTHESIS TESTING
p < 0.05
H₀

Results

6 Hypotheses
Tested across distinct organisational scenarios — commission, pricing, productivity, social media behaviour, quality control, and revenue
Effect Size
Cohen's d applied — statistical vs practical significance distinguished
2 Non-parametric
Normality violations detected via Shapiro-Wilk (p < 0.001) — Mann-Whitney and Wilcoxon applied in place of parametric alternatives

Applied across 6 distinct organisational hypotheses, the framework correctly identified cases where apparent relationships did not survive rigorous testing — including a pricing comparison where the null was retained (p = 0.15) and a marketing engagement analysis where Shapiro-Wilk tests (p < 0.001 for both groups) ruled out parametric methods entirely, triggering Mann-Whitney U and Wilcoxon signed-rank alternatives. Where nulls were rejected, effect sizes were calculated to assess practical significance alongside statistical significance.

Power analysis was conducted prospectively for the key tests, confirming adequate sample sizes to detect effects of the minimum practically meaningful magnitude. This matters because underpowered tests that fail to reject the null prove nothing — they are as misleading as spurious significant results, and far more common in practice.

Technology Stack

Python SciPy Statsmodels Pandas Matplotlib Seaborn
Interested in this work or something similar?