Starting from the Decision

The first discipline

Analysis begins with a decision, not with a dataset. Before opening anything, establish:

• The question — stated specifically enough that an answer would be recognisable. “Why are costs up?” is not yet a question; “which cost categories increased more than five per cent against last year, and in which sites?” is.
• The decision it informs and who makes it.
• The actions available — if every answer leads to the same action, the analysis is optional.
• The threshold that matters — how large a difference would change the decision. This determines how precise the analysis needs to be, and usually it needs to be far less precise than people assume.
• The deadline — a rough answer in time beats a precise one afterwards.

Descriptive, diagnostic, predictive

• Descriptive — what happened. Most business questions are answered here.
• Diagnostic — why. Segmentation, comparison, and testing alternative explanations.
• Predictive — what is likely next.
• Prescriptive — what to do, which requires evidence that the proposed action works.

Types of data

• Categorical (nominal) — unordered labels: site, product, department. Count them; do not average them.
• Ordinal — ordered but with unequal intervals: satisfaction scales, severity ratings. Medians are safer than means.
• Interval and ratio — numeric with meaningful differences; ratio data has a true zero, so ratios make sense.
• Time series — ordered observations over time, with seasonality and trend.
• Text and unstructured — comments, reports, logs.

Applying the wrong operation to a type is a standard error: averaging a satisfaction score of one to five treats the gap between one and two as identical to the gap between four and five, which it is not.

Populations and samples

Where you have all the data — every transaction, every employee — you have a population, and differences are simply facts about it; no significance test is needed to state that one site is higher than another. Where you have a sample, the question is whether the pattern would hold in the whole, and that depends far more on how the sample was obtained than on its size. A large self-selected sample is worse than a small random one, because size does nothing to remove selection bias.