Four kinds of question
Data analytics is often presented as a list of techniques. It is more usefully understood as four questions, each harder than the last, and knowing which one you are being asked prevents most wasted effort.
| Type | Question it answers | Example |
|---|---|---|
| Descriptive | What happened? | sales fell 12% last quarter |
| Diagnostic | Why did it happen? | the fall came entirely from one region |
| Predictive | What is likely to happen? | sales will fall again unless something changes |
| Prescriptive | What should we do? | move the advertising budget to that region |
Most real work is descriptive, and that is fine
Students assume prediction is the goal and describe it as the point of analytics. In practice most valuable analysis is descriptive and diagnostic — organisations frequently do not know what happened last month, let alone why. A clear answer to "what happened" is worth more than a shaky prediction, and a question asking you to justify an approach usually rewards saying so.
The analytics pipeline
The stages are the same as in data science, and each one can invalidate everything after it.
Collection gathers the data. Cleaning removes duplicates, corrects errors and handles missing values — usually the longest stage. Transformation puts it into a usable shape: combining sources, converting units, deriving new columns. Analysis looks for the pattern. Visualisation and interpretation present it to someone who must act on it.
- Structured data fits neatly into rows and columns — a sales table, a student record.
- Unstructured data does not — photographs, free-text comments, audio. Most data generated today is unstructured.
- Semi-structured data has some organisation but no fixed table shape, such as JSON or XML.
- Metadata is data about data: when a file was created, by whom, in what format.
Describing a dataset honestly
Before any conclusion, summarise. The mean uses every value and is distorted by extremes; the median ignores them; the mode is the only one that works for categories. Spread matters as much as centre: the range is quick but depends only on the two extreme values, while the standard deviation uses all of them and stays in the original units.
Quoting a centre without a spread is how misleading summaries are made. "Average marks were 60" is compatible with everyone scoring 60 and with half the class scoring 20 and half scoring 100 — two situations demanding completely different responses.
Two branches each report average monthly sales of Rs 500 000. Branch A ranges from 480 000 to 520 000; Branch B from 100 000 to 900 000. What does this tell a manager?
- The means are identical, so the mean alone cannot distinguish the branches at all.This is exactly why a summary needs a measure of spread as well as a centre.
- Branch A has a range of 40 000 and Branch B a range of 800 000 — twenty times larger.The range is crude but the difference here is far too large to be noise.
- Branch A is predictable, so stock and staffing can be planned confidently.Low variability has real operational value beyond the average itself.
- Branch B is volatile: some months are excellent and some may not cover its costs.A month at 100 000 might be a loss, which the average conceals completely.
- The manager should investigate what drives Branch B's variation before treating the two as equivalent.That is a diagnostic question, and it is the one the data has raised.
Equal means but very different spread — Branch A is predictable, Branch B is volatile and needs investigation.
The ogive is the one worth studying here. Reading across at N/2 gives the median and at N/4 and 3N/4 the quartiles, so the middle half of the data can be read straight off the graph.
Correlation, causation and the ways analysis misleads
A correlation means two variables move together. It does not mean one causes the other, and the reasons it might not are worth knowing by name.
There may be a third factor causing both — ice cream sales and drownings both rise with temperature. The causation may run the other way — do police numbers cause crime, or does crime cause police numbers? Or it may be pure coincidence, which becomes more likely the more pairs of variables you test.
Three specific ways a chart lies
A truncated axis — starting the vertical axis at 95 rather than 0 turns a 2% change into a dramatic cliff. Cherry-picked ranges — choosing the start date that makes the trend look strongest. Missing context — a rise in reported cases may mean more disease or simply more testing. When a question asks you to criticise a presentation of data, these three cover most of what there is to say.
Turning analysis into a decision
Analysis that nobody acts on has produced nothing. The final stage is communication, and the same rules apply as in any technical writing: state the finding first, support it with the evidence, and be explicit about what the data cannot tell you.
The most valuable habit is stating the limitation yourself. "Sales rose 12% after the campaign, but a competitor also closed that month, so the campaign cannot be credited with all of it" is a stronger report than one claiming the full effect — because the person acting on it can now weigh it properly.
Before you leave this chapter
- Descriptive, diagnostic, predictive, prescriptive — know which question is being asked.
- Cleaning is the longest stage, and everything after it depends on it.
- Always report spread alongside centre. Equal means can hide completely different situations.
- Correlation may be a third factor, reversed causation, or coincidence.
- Truncated axes, cherry-picked ranges and missing context are how charts mislead.