The Analyst Mindset

Tools are only as good as the person wielding them. A spreadsheet in the hands of a careful analyst can outperform Pandas in the hands of a careless one. This chapter is about the habits of mind that distinguish good analytical work — habits that apply no matter what tool you eventually pick up.

The five habits

1. Be skeptical of your data.
2. Explore before you conclude.
3. Prefer evidence over intuition (and intuition over guesses).
4. Make your reasoning legible.
5. Treat every analysis as if it will be audited.

We will walk through each.

1. Be skeptical of your data

The most dangerous assumption is that your data is correct. It almost never is — at least, not entirely. Before you trust a number, ask:

• Where did it come from? Which system collected it, when, and how?
• What was filtered out before I got it? Test users? Internal employees? Cancelled orders?
• What does "missing" mean? Was the field skipped, defaulted, or set to zero by a downstream process?
• What changed about the collection over time? Did the form add a new field in 2022? Did the logging switch from "click" to "tap" mid-year?

Every dataset is a photograph of reality, taken through an imperfect lens. Treat it that way.

2. Explore before you conclude

A common rookie mistake is to look up an answer (df["revenue"] .sum()) and immediately put it in a report. A more experienced analyst spends a few minutes first looking at the distribution:

• What does a histogram of revenue look like?
• Are there a few outlier orders that dominate the total?
• Are there negative or zero values (refunds, voids)?
• Does the count of orders look reasonable for the time period?

These checks take 30 seconds and routinely catch data bugs that would have made the eventual conclusion wrong.

The naive total includes a giant outlier and a refund — both of which deserve investigation before the number is reported.

3. Prefer evidence over intuition (and intuition over guesses)

Your gut is useful for generating hypotheses — "I think customers from New York spend more." Your gut is bad at confirming those hypotheses. Check.

But your intuition is still better than random guessing. If a result looks shocking, you are right to be suspicious. Most surprising results in data are data bugs, not breakthroughs.

The path through "surprises → check for bug" is travelled many times a week by every working analyst.

4. Make your reasoning legible

Two flavors of legibility:

• Code legibility. Use descriptive variable names. Break complex pipelines into small steps with intermediate variables. Add comments where intent is non-obvious.
• Communication legibility. When you present a number, also present its denominator, its time window, and any filters applied. "Revenue from US customers in Q1, excluding refunds: $4.2M"* is far more useful than *"$4.2M".

A handy heuristic: imagine your future self trying to extend the analysis in three months. Can they follow what you did?

5. Treat every analysis as if it will be audited

Even small companies eventually have an investor, a board member, a regulator, or an internal compliance officer who asks "how did you arrive at this number?" The analysts who can answer in two minutes — by re-running a script and walking through it — are the ones whose careers thrive.

This is why we will spend a whole chapter on Reproducible Analysis. It is not academic. It is professional self-defense.

Common cognitive traps

Even experienced analysts fall into these. Awareness helps.

Confirmation bias

You expect the answer to be X, so when the first analysis produces X, you stop looking. Counter: deliberately try to disprove your hypothesis.

Survivorship bias

Your dataset only contains the things that made it through some filter — users who completed onboarding, planes that landed, companies that did not go bankrupt. Conclusions about "users" or "companies" are really conclusions about the surviving subset.

Simpson's paradox

A pattern that holds in subgroups can reverse when the groups are aggregated. The most famous example: a 1973 lawsuit against UC Berkeley alleging gender bias in admissions found that men were admitted at a higher rate overall, but each department admitted women at an equal or higher rate than men. Women applied disproportionately to more competitive departments.

This is a real and constant risk. Whenever you compute an aggregate, ask whether breaking it down by an important sub-group changes the story.

Cherry-picking

Running fifty comparisons and reporting only the one that turned out interesting. The standard defense in scientific work is pre-registering your hypotheses — writing them down before looking at the data — but in business work the discipline is less common, so the temptation is constant.

A small skepticism drill

Below is a tiny dataset of customer satisfaction scores. Do not just take the mean. Look first.

The naive mean hides the fact that all the low scores came in through the phone channel. That is the headline. The headline mean is misleading. A skeptical analyst spots this in seconds.

When to slow down and when to speed up

There is also a counter-temptation: an analyst who is so careful that they never produce anything. Some questions deserve an hour, some deserve a week. A rough decision tree:

A decision about "which font should the homepage use" deserves a back-of-the-envelope check. A decision about "should we lay off 10% of the company based on this attrition model" deserves double-checking by someone other than you.

The mindset, distilled

If you internalize one thing from this chapter:

Every number you report is a contract with your reader. The contract says: "if you make a decision based on this number, I have done my due diligence so that the decision is well- founded."

That contract is what makes analysis a profession, rather than just a tool-using activity.

Check your understanding

QuestionSelect one

Why did the naive mean satisfaction score in the chapter's example mislead?

The mean function in Pandas is broken

There were too few rows

All the low scores came in through one channel; aggregating across channels hid the variation that mattered

The scores were on a wrong scale

QuestionSelect one

Which of these is Simpson's paradox?

The number of pandas dataframes you create doubles every six months

A function that returns different answers each time

A pattern that holds in every sub-group can reverse direction when the sub-groups are combined

A statistical test that always rejects the null

QuestionSelect one

Which habit does the chapter call "professional self-defense"?

Memorizing every Pandas function

Charting everything

Treating every analysis as if it will be audited — i.e., making it reproducible, traceable, and explainable

Drinking lots of coffee

QuestionSelect one

When a result is surprising, what is the chapter's recommended first instinct?

Publish it immediately

Assume the data is correct and report it

Suspect a data or code bug, and investigate before trusting it

Ignore it