What Data Analysis Is

People throw the phrase "data analysis" around as if everyone knows what it means. In practice, ask ten analysts and you will get ten subtly different answers. Let us pin it down.

A working definition

Data analysis is the disciplined process of turning raw recorded observations into useful, defensible answers to questions humans care about.

Three words in that sentence are doing all the work:

• Disciplined. It is not "looking at numbers until something feels right." It is a structured workflow with checkpoints, validation, and (often) peer review.
• Useful. The output is supposed to influence a decision — buy or sell, hire or fire, launch or kill, treat or wait.
• Defensible. If someone challenges your number, you can walk them through every step that produced it. Reproducibility again.

Notice what is not in the definition: nothing about Python, nothing about Pandas, nothing about machine learning. Those are tools. The definition is about the work, which existed centuries before any of those tools did.

The four kinds of analysis

A common framework — adapted from the consulting world — divides analytical work into four kinds, in order of increasing ambition.

Descriptive — What happened?

You count, you sum, you average, you slice. "Revenue last quarter was $12.4M, up 6% year over year." "We onboarded 312 new users in March." "The median time to resolve a support ticket is 4 hours." This is the bread and butter of every analyst's day and the focus of most of this course.

Diagnostic — Why did it happen?

You drill in. "Revenue grew because the West region grew 18%, while every other region was flat." "Support tickets are slower because two senior engineers left." Diagnostic analysis is still mostly Pandas — slicing, grouping, comparing — but it demands more domain knowledge and skepticism.

Predictive — What will happen?

You build a model. "If marketing spend stays flat, we expect 4,200 new users next month." This is where statistics and machine learning come in. We touch this lightly in the Hypothesis Intuition chapter; it is the subject of its own course.

Prescriptive — What should we do?

You recommend an action and quantify its expected impact. "We should reduce churn by improving onboarding; based on a 3% lift estimate that is worth $400k/year." This requires combining all the previous layers with business judgment.

This course lives almost entirely in the descriptive and diagnostic layers. Those are by far the most common kinds of data work and the foundation for everything above them.

What analysts actually do

Strip away the buzzwords and a working analyst spends their day doing some mix of:

1. Asking — translating a vague business question into a specific, answerable analytical question.
2. Acquiring — getting the right data into a tool you can work with.
3. Cleaning — fixing types, missing values, inconsistencies, bad rows.
4. Exploring — summary statistics, group breakdowns, visualizations.
5. Interpreting — what does the pattern mean? Could it be a data artifact? Could it be noise?
6. Communicating — distilling the result for a decision maker, in words, charts, and (sometimes) a recommendation.

Notice the dashed feedback arrow. Real analyses almost always generate new questions, sending you back to "Ask." A good analyst is comfortable with the loop and resists the temptation to stop at the first plausible answer.

The question-asking skill

The single most under-rated skill in data analysis is asking the right question.

A bad analytical question is vague:

"How is the marketing campaign doing?"

A good analytical question is specific:

"Of users who saw the new banner ad between March 1 and March 15, what fraction made a purchase within 7 days, and how does that compare with users who saw the old banner during the same window?"

The second question has a single, computable, defensible answer. The first does not.

An end-to-end example, slowly

Let us run the loop once at a relaxed pace. Question:

Among employees in our HR dataset, does monthly income differ between people who stayed versus people who left the company?

Here is the full process explicitly.

Step 1: Ask precisely

Define the comparison:

• "Monthly income" = the MonthlyIncome column.
• "Stayed" = Attrition == "No".
• "Left" = Attrition == "Yes".
• "Differ" = compare the median (less sensitive to outliers than the mean), and the 25th and 75th percentile (to see spread).

Step 2: Acquire

Step 3: Clean

Quick check: are there any missing values in the two columns we need?

If both are zero and Attrition only contains Yes/No, we are in good shape.

Step 4: Explore

You can see the quartiles for each group at a glance.

Step 5: Interpret

You will probably notice that the median income of people who stayed is higher than those who left. Does that mean low pay causes attrition? Not necessarily. It could mean:

• Lower-paid employees are also younger and more mobile.
• Lower-paid roles are in departments with higher turnover for other reasons.
• The company is shedding underperformers who happened to be lower-paid.

A descriptive comparison is a starting point, not a conclusion. This is why diagnostic analysis usually follows.

Step 6: Communicate

You would summarize this to a stakeholder as a short note:

Among the 1,470 employees in our dataset, the median monthly income of those who left was about $4,500, versus$5,300 for those who stayed — a ~15% gap. This is a correlation, not a proven cause; further analysis by department and job level is needed before recommending a pay-based retention program.

Notice how the language is hedged. Good analysts are honest about what their numbers can and cannot prove.

What data analysis is not

• It is not memorizing every method on the DataFrame.
• It is not the same as data engineering (building pipelines).
• It is not the same as machine learning (training models).
• It is not the same as business intelligence (building dashboards), though there is heavy overlap.
• It is not a substitute for domain expertise — it amplifies it.

A short pause for skepticism

Every dataset is a measurement of reality, not reality itself. Two analysts can look at the same dataset and reach different conclusions because they made different (and equally defensible) choices about cleaning, grouping, or comparison. This is uncomfortable, but it is fundamental to the work. We will keep coming back to it.

Check your understanding

QuestionSelect one

Which of the following best matches the chapter's working definition of data analysis?

Producing as many charts as possible

Memorizing Pandas function names

A disciplined process of turning raw recorded observations into useful, defensible answers to questions humans care about

Building machine-learning models

QuestionSelect one

Which of these is a descriptive analytical question (as opposed to diagnostic, predictive, or prescriptive)?

Why did revenue drop last quarter?

What should we do to reduce churn?

What will revenue be next quarter?

What was our revenue last quarter?

QuestionSelect one

Why does the chapter say "asking the right question" is an under-rated skill?

It saves keystrokes

It avoids using Pandas

Vague questions ("how is the campaign doing?") have many plausible answers, while specific questions ("of users who saw banner X between dates A and B, what fraction purchased within 7 days?") have one computable answer

It impresses your manager

QuestionSelect one

In the income-vs-attrition example, the chapter cautioned that the observed gap is "a correlation, not a proven cause." What is the closest reason given?

The dataset is too small

Pandas cannot compute causes

The income gap could be explained by many other factors (age, department, role), so a simple comparison cannot prove that pay drives attrition

Attrition is a categorical variable