Note — N-01
How I work a data question
I keep this one around because the dataset is public and the answer is honest. It’s a Kaggle exploration of a fast-food chain’s promotion test, not client work. The point isn’t the result. It’s the process, which is the same whether the data is a toy set or a real one.
Write the business question first. Decide what "won" means before touching the data.
State the limits. Keep the line clear between what you measured and what you assumed.
Ask before you touch the data
Three promotions ran for a new menu item. The question sounds simple: which one won? It isn’t. “Won” could mean total sales, average weekly sales, or growth over the window. I wrote the business question down first, so the analysis had something to answer to.
Test for real, not by eye
The charts showed one promotion lagging. Charts lie. A visible gap can be random. So I ran the tests. A Chi-Square first, to confirm the promotions were fairly distributed across market sizes, so a difference couldn’t be an assignment artifact. Then ANOVA to check the three means differed at all: F was about 21.95, p under 0.0001. Then Tukey’s HSD to see which pairs. Promotions 1 and 3 both beat 2, and were statistically tied with each other.
Say what the data can’t
The tie is the interesting part. The honest recommendation was to keep both winners, not force a single one, because the data didn’t support picking between them. I also kept the dollar projections labeled as projections. The dataset can tell you which promotion performed. It can’t tell you the revenue that follows. Pretending otherwise is where analysis goes wrong.
The tools change. The discipline doesn’t. Ask the right question, validate before you believe, and stay clear about the line between what you measured and what you assumed.