Data Cleaning Exercises

How to Talk About Data Cleaning in a Data Analyst Interview

How to describe data cleaning in a data analyst interview: one story with real numbers, a STAR structure, clear trade-offs, and phrases to avoid.

By Tnas · Published · 8 min read

Dark title card reading How to Talk About Data Cleaning in an Interview beside a grid of table cells with a few flagged cells

To talk about data cleaning in a data analyst interview, describe it as a series of decisions backed by evidence: what you checked, which rule you chose and why, how you proved nothing was lost, and what changed in the final number. "I cleaned the data" is easy to say and tells the interviewer very little. What separates candidates is whether they can explain one cleaning decision clearly enough that the interviewer could repeat it.

This guide shows how to build that answer: one story with real numbers, a structure that keeps it under two minutes, plain-language translations of technical steps, and the phrases that undersell the work.

Key Takeaways

  • Prepare one cleaning story with numbers: rows in, problems found, rows flagged or fixed, and the effect on the result.
  • Use the STAR structure, and spend most of the time on the actions: the checks you ran and the rules you chose.
  • Name a trade-off you made (flag versus delete, fill versus leave missing) and why you chose that side.
  • Translate steps into decisions: say "I kept refunds as negative amounts after confirming them with finance", not "I fixed the negatives."
  • If you lack job experience, a practice project told honestly works, as long as the numbers are real.

Why interviewers ask about data cleaning

Cleaning is part of the job description, not a warm-up topic. The US Department of Labor's O*NET profile for data scientists lists "Clean and manipulate raw data using statistical software" among the occupation's tasks, alongside "Prepare data for analysis" as a work activity (O*NET OnLine, 15-2051.00). Analyst roles involve much of the same work, so cleaning questions are a natural way to test three things:

  • Judgment: do you look before you change, and can you tell an error from an unusual but real value?
  • Honesty with data: do you keep a record of what you removed, or does data quietly disappear?
  • Communication: can you explain a technical choice to someone who will rely on the number?

The rest of this guide is about showing all three in one answer.

Step 1: Choose one story and collect its numbers

Pick a single project where cleaning changed the outcome. Before the interview, write down these numbers, because vague stories sound invented and specific ones sound lived:

Number to prepare Example of what it sounds like
Size of the input "about 48,000 order lines from three monthly exports"
Problems found, by type "1,900 duplicate deliveries, 310 lines with no customer id"
What you did with each "removed the duplicates by event id, sent the 310 to an exceptions list"
Reconciliation "rows in equalled rows kept plus rows set aside, and revenue matched finance's total"
Effect on the answer "the monthly figure dropped by 4% once double-counted lines were removed"

The figures in the right column are an illustration of the level of detail, not a benchmark. Use your own, and only numbers you can explain if asked how you got them. If you no longer have the exact values, say "roughly" and give the order of magnitude.

Step 2: Structure the answer with STAR

STAR is a four-part answer structure: situation, task, action and result. The University of Washington's career center suggests giving the situation about 20% of the answer, the task about 10%, the action about 60% and the result about 10% (UW Career & Internship Center). For a cleaning story, the action section is where you describe checks and rules, so the 60% weighting fits well.

Part Share of the answer What it covers in a cleaning story
Situation about 20% The data and why it mattered
Task about 10% The question the cleaned data had to answer
Action about 60% What you profiled, the rules you chose, how you checked them
Result about 10% Reconciled numbers and the decision they changed

Inside the action section, keep a fixed order so you never ramble: profiled, found, decided, verified. "I profiled the file, found three problems, decided how to handle each with the owner, and verified the totals." That order also mirrors how a careful cleaning job runs, which makes the story more believable.

Step 3: Translate technical steps into decisions

Interviewers for analyst roles are often managers who will use your numbers, not engineers who will review your code. Describe each step as a decision with a reason, and mention the function only if asked.

Instead of saying Say
"I removed duplicates." "I defined a duplicate as the same event id delivered twice and kept the first delivery, so reruns give the same result."
"I dropped the nulls." "Lines with no customer id went to an exceptions list with a reason, so finance could chase them instead of losing the revenue."
"I fixed the dates." "Two sources wrote dates day-first and one month-first, so I parsed each with its declared format and reported the 0.3% that failed."
"I removed outliers." "I flagged orders over $10,000 rather than deleting them; most were genuine bulk orders."
"I cleaned up the categories." "I mapped 14 spellings of channel to 5 approved values and listed the 2 that nobody recognized."

Each right-hand answer contains a rule, a reason and evidence. The values are examples; replace them with yours. For more on phrasing a choice, see the lesson on explaining a trade-off.

Step 4: Name a trade-off you made

Most cleaning decisions have a cost on both sides, and interviewers often ask "what would have happened if you had done the opposite?" Prepare one trade-off from your story. Common ones:

  • Flag or delete. Deleting is simpler; flagging keeps the evidence and lets the analysis decide. The guide on quarantine, flag, fix or delete lays out when each is right.
  • Fill or leave missing. Filling makes charts look complete; leaving gaps keeps averages honest. Saying "I filled labels but left measurements missing" shows you know the difference.
  • Merge or keep separate. Merging two customer records simplifies counts but cannot easily be undone; keeping a review queue for uncertain matches is slower but safer.

A good trade-off answer ends with what you told the stakeholder, because that shows you did not make the choice in isolation.

Step 5: Show that you asked questions

Many cleaning problems are really questions about meaning: does a blank mean zero or unknown, are test accounts included, which record wins when two disagree? Mention one question you asked and how the answer changed your rule. It signals that you treat the data owner as the source of truth instead of guessing. The guide on clarifying questions to ask about messy data lists the ones that come up most, and the stakeholder interview questions page has practice prompts.

A worked example answer

The example below is a constructed illustration of the structure, not a real project. Read it for the shape, then build your own from real numbers.

Situation. A retail team wanted monthly revenue by sales channel from three CSV exports, around 48,000 order lines in total.

Task. My job was to produce a channel table finance could reconcile to their ledger.

Action. I profiled the files first and found three issues. Some orders appeared in two exports, because the exports overlapped by a few days; I removed repeats by order line id, keeping the earliest export. About 300 lines had no customer id; I asked finance, who said they were in-store sales, so I kept them under an "in-store" label rather than dropping them. The channel column had 14 spellings, so I agreed a mapping to five channels with the marketing lead and reported two unrecognized values instead of guessing.

Result. Rows in equalled rows kept plus duplicates removed, and total revenue matched finance's figure. The overlap fix lowered the reported total by about 4%, which explained a gap finance had been asking about, and finance switched its monthly report to the deduplicated figure.

Notice the proportions: two sentences of setup, most of the answer on actions, and a result that ends on a decision someone made.

If you do not have work experience yet

Use a practice project, and say it is one. "In a practice project with a public dataset, I..." is honest and still lets you show every step above. Choose a project with a clear question and real mess, keep your notes of what you found, and bring numbers. The guide on building a data cleaning portfolio project describes how to set one up. Graded tasks such as cleaning an order-line export or combining overlapping exports give you concrete problems to talk about.

What not to do: present a tutorial you followed as your own analysis, or claim results you did not measure. Dartmouth's Guarini School career center advises using "I" statements and truthful, specific stories rather than generalizations or exaggerations (Guarini Career Center). Interviewers ask follow-up questions, and invented details fall apart quickly.

In a live technical round: narrate as you go

When you clean data live, say each check before you run it: "First I'll count rows and distinct order ids to see the grain." If you run out of time, a narrated plan still earns credit, while silent typing gives the interviewer nothing to grade. Finish by stating what you would check next. The interview-ready learning path is built around explaining each decision as you make it.

Frequently Asked Questions

How long should a data cleaning story be in an interview?

Around 90 seconds to two minutes. Prepare a longer version in case the interviewer asks for detail, but lead with the short one and let them pull more out.

Should I mention specific tools like SQL or pandas?

Yes, briefly and where the job asks for them: "I did the deduplication in SQL with a window function." Then return to the decision. Tools show you can do the work; decisions show you understand it.

What if the interviewer says my approach was wrong?

Ask what they would expect, then discuss it. Many cleaning choices depend on business rules you were never told, so saying "if refunds should be excluded, I would filter them before the total" is a strong reply, not a defensive one.

Sources