Data Analyst Interview Prep · Final test · 25 questions

Data Analyst Interview Prep final test: 25 free questions

Untimed here · timed and scored in the app

A free Data Analyst Interview Prep practice test with 25 questions drawn from the whole course. Answer at your pace and read why each option is right or wrong.

Swipe, scroll or use ← →

Q1 / 25

The brief allows any method. A simple monthly comparison answers the question; a time series model would answer it too. Which do you submit and what do you say?
1. AThe simple one, without mentioning that a model was ever considered.
2. BBoth, so the reviewer can see the simple result and the advanced one.
3. CThe model, because it demonstrates the technical range the role asks for.
4. DThe simple one, saying what it answers and where it would fall short.
Show the answer
The strongest signal is choosing the method that fits the question and being able to say why, including its limits. Reaching for the heavier tool to show range invites the question of why the question needed it. Submitting both doubles the reading with no gain, and hiding the alternative throws away the part that shows judgement.

Next → 1 / 25
Q2 / 25

An analyst must set amount to 0 wherever status is cancelled, keep every row and change the original frame. Their code sometimes leaves the frame untouched. What repairs it?
1. AA single assignment through loc, selecting rows and column at once.
2. BSelecting the cancelled rows first, then assigning to that result.
3. CReplacing the assignment with a filter that drops those rows.
4. DCopying the frame and assigning on the copy instead.
Show the answer
Assigning through the result of a first selection means writing into an object that may be a copy, so the write sometimes lands nowhere. One loc call that names the rows and the column at the same time always writes into the frame you meant. Copying keeps the original untouched by design, and deleting the rows changes the answer instead of the value.

Next → 2 / 25
Q3 / 25

A query has a FROM clause, a WHERE clause, a GROUP BY and a select list. Which rows does GROUP BY actually receive?
1. AThose left once the select list has been computed for each row.
2. BEvery row of the tables named in FROM, before any filter runs.
3. CThose left by FROM and then WHERE.
4. DThose left by WHERE and by HAVING working together.
Show the answer
FROM derives a virtual table, WHERE discards rows from it, and only the survivors reach GROUP BY. That pipeline is the reason a row condition belongs in WHERE: by the time HAVING runs the groups already exist, so it can drop a whole group but never a row inside one.

Next → 3 / 25
Q4 / 25

Two minutes are left. A business manager has to decide today and an analyst wants to know why one data check failed. You cannot serve both in full. What do you say?
1. AMove the detail offline and use the time to repeat the main finding.
2. BExplain the failed check in full, since nobody should decide on weak evidence.
3. CGive the decision and its limit in plain words, then the detail on request.
4. DGive the headline result and leave the failed check for the written follow-up.
Show the answer
Different stakeholders need the same conclusion in different forms, so the decision and the limit that touches it go first, and the mechanism goes to whoever asks for it. Choosing one listener, hiding the failed check or repeating the headline each leave somebody in the room unable to act.

Next → 4 / 25
Q5 / 25

Three analyses are on the table and you have time for one before the decision meeting. Which do you pick?
1. AThe one whose result would change what the team does.
2. BThe one that reuses the pipeline you built for the last case study.
3. CThe one that the most people in the meeting have already asked to see.
4. DThe one that covers the largest share of the revenue in the business.
Show the answer
Analysis earns its time by moving a decision, so the test is whether either possible outcome would lead to a different action. Popularity, reuse and revenue coverage are proxies that can all point at work whose answer changes nothing: a big number that everyone already agrees on is still an expensive confirmation.

Next → 5 / 25
Q6 / 25

You fit a regression and the coefficient is significant, but the residuals fan out as the fitted values grow. What does that pattern cost you?
1. AThe model can still be reported as it is, because the p-value is small.
2. BThe conclusions drawn from the fit may no longer hold.
3. CThe coefficient estimate is reversed in sign and must be recomputed.
4. DThe fit has to be discarded, since residual patterns cannot be repaired.
Show the answer
Inference from a model rests on assumptions about its errors, so a systematic pattern in the residuals means the standard errors and the p-values they feed can no longer be taken at face value. That is a warning about the conclusions, not a proof that the coefficient has the wrong sign, and it is usually addressed by changing the model rather than abandoning it.

Next → 6 / 25
Q7 / 25

A chart shows revenue rising steeply. The vertical axis carries no label and no unit. Why is that a problem beyond tidiness?
1. AThe colour palette becomes harder to read without a legend.
2. BThe library will pick an ordering for the categories on its own.
3. CThe chart will not render correctly in the dashboard export.
4. DThe reader cannot tell which unit the revenue values are measured in.
Show the answer
The tick numbers may still show the scale, but without a unit the reader cannot tell dollars from thousands of dollars. So the steepness cannot be read as an amount. Export, ordering and palette are separate concerns that do not change what a value means.

Next → 7 / 25
Q8 / 25

A take-home brief says: investigate why sales fell. It names no decision and no audience. What do you settle before you start?
1. AWhich chart format the reviewer prefers for the final written report.
2. BWhich decision the analysis supports and how the findings will be used.
3. CWhich software package you are expected to use for the analysis.
4. DWhich of the supplied tables is the largest and most complete one.
Show the answer
A take-home is scored on judgement, and the first piece of judgement is naming the choice your work is meant to inform, because it decides what is worth analysing and what belongs in the summary. When the brief does not say, write your assumption in the first paragraph. Formats, tools and table sizes are all decided after that.

Next → 8 / 25
Q9 / 25

Two frames store customer_id as an ordinary column and their row indexes are unrelated. You need to bring the rows together by customer_id. Which operation matches them?
1. Aleft.merge(right, on='customer_id')
2. Bpandas.concat on the two frames with axis set to 1
3. Cleft.join(right) using the default arguments of the method
4. Dpandas.concat on the two frames with axis set to 0
Show the answer
merge matches rows using the key column you name. concat combines objects along one axis and aligns the labels on the other, so with axis=1 it lines the frames up by row index, which here carries no meaning, and with axis=0 it simply stacks them. join also works on the index by default, so it has the same problem until you tell it which column to use.

Next → 9 / 25
Q10 / 25

You are asked for the total paid amount per customer. Orders has one row per order and items one row per order line. What do you settle before writing the join?
1. AWhat one result row stands for, and which table can repeat that key.
2. BWhether to use an inner join, so that unmatched orders never reach the result.
3. CWhether the amount column should be added up with SUM or with COUNT.
4. DWhether the query returns the columns in the order the report expects.
Show the answer
A join emits one output row for every matching pair, so joining a table that repeats the customer key multiplies the rows you are about to add up. Deciding what one result row stands for tells you which key has to be unique, and that is the check that protects the total. Join type, aggregate and column order all come later and none of them stops a double count.

Next → 10 / 25
Halfway, at your pace

In the app the mock exam is timed and scored like the real thing.
Q11 / 25

A boundary case fails two minutes before the round ends. The interviewer asks what your result supports. Which answer is strongest?
1. AExplain that the case is an edge condition that is unlikely to occur in real data.
2. BName the failing case, limit the conclusion to what passed, and say what you would check next.
3. CReport what passed and mention the failure in the follow-up email.
4. DName the failing case and use the time left to try a fix.
Show the answer
Testing your own solution is part of the work, so an unfinished check is information the panel wants rather than a mark against you. Saying what failed, what the conclusion still covers and what comes next keeps the claim the size of the evidence. Rushing a fix, deferring the news or waving the case away all claim more than you verified.

Next → 11 / 25
Q12 / 25

Your query runs and the numbers look reasonable. Before you say you are done, what do you do?
1. AReread the code carefully, looking for mistakes in the logic you wrote.
2. BFormat the output so the interviewer can read it.
3. CTry a boundary case and check the total against something known.
4. DRerun it to confirm that the same numbers come back a second time.
Show the answer
Reasonable looking output is the most common way a wrong answer survives, so test it rather than inspect it: an empty group, a null, a duplicated key, and one total you can compare against a figure you trust. Rereading finds the errors you already know how to make, a second identical run proves nothing, and formatting tidies a result you have not checked.

Next → 12 / 25
Q13 / 25

You have to recommend on weak evidence and the decision is easy to undo. Which recommendation is strongest?
1. ARecommend waiting until the evidence is strong enough to be conclusive.
2. BRecommend acting and present the supporting evidence without the caveats.
3. CPresent the options with no recommendation, since the evidence is weak.
4. DAct now, say what the evidence does not cover, and set the point to review.
Show the answer
How reversible a decision is changes how much certainty it deserves: a cheap mistake that can be undone next week does not need the evidence a permanent one would. Naming the gap and the review point is what keeps that honest. Waiting has a cost too, hiding the caveats transfers a risk silently, and refusing to recommend leaves the judgement to whoever read least.

Next → 13 / 25
Q14 / 25

You chose a weekly batch job over a streaming pipeline. The interviewer asks why. What makes the answer strong?
1. AThe fact that the team already knew the batch tooling and not the streaming one.
2. BThe fact that the batch job has run without incident since it was deployed.
3. CThe requirement it met, the option you rejected and what that cost.
4. DThe fact that streaming would have needed infrastructure nobody had approved.
Show the answer
A design choice is scored on the reasoning, which means the requirement that decided it, the alternative you weighed, and what you gave up by choosing as you did. Reliability since launch is an outcome rather than a reason, and familiarity and approvals are real constraints that only become an argument once the requirement is stated first.

Next → 14 / 25
Q15 / 25

Two teams compare the same two fitness programmes. One surveys people who already chose a programme; the other assigns people to a programme at random. Why can only the second team claim a cause?
1. ARandom assignment spreads the other influences across both groups.
2. BThe second team measured each participant twice rather than only once.
3. CThe second team used a population that represents the country better.
4. DThe second team worked with a larger sample than the first one did.
Show the answer
When people choose their own group, the groups differ in ways you did not record, and any of those differences could explain the outcome. Assigning at random leaves those influences spread across both groups, which is what licences a causal statement. Repeated measurement, sample size and representativeness all improve a study without removing that confounding.

Next → 15 / 25
Q16 / 25

A histogram of session length shows one smooth peak. With a narrower bin width, two separate peaks appear. Which reading is right?
1. ANeither view is automatic truth, so check the shape across several bin widths.
2. BThe narrow version is right, because more bins always preserve more detail.
3. CNeither can be used, so the distribution should be reported as a mean instead.
4. DThe smooth version is right, because the default bin width is chosen from the data.
Show the answer
Bin width is a parameter you chose, and the wrong one can hide a real feature or invent one out of random variation. Automatic defaults rest on assumptions about the data that may not hold here, and more bins eventually show noise as structure. Check more useful bin widths before treating either visible shape as a stable feature.

Next → 16 / 25
Q17 / 25

A reviewer must rerun your analysis and see exactly how the supplied data were changed. What do you submit?
1. AThe original inputs and the code that turns them into the analysis data.
2. BThe final tables and a notebook showing the charts that were produced.
3. CThe transformation code and the final tables, with the original inputs left out.
4. DThe cleaned data with a written summary of the edits you made by hand.
Show the answer
Reproducibility means somebody else can start where you started and arrive where you arrived, which takes both ends: the untouched inputs and the code that connects them. A prose summary of manual edits cannot be rerun, code without its inputs has nothing to run on, and final tables show the destination only.

Next → 17 / 25
Q18 / 25

Ten per cent of the delivery_time column is missing. A colleague proposes filling those rows with the column mean so no row is lost. When is that the wrong move?
1. AWhen more than five per cent of the column has no value.
2. BWhen the value is missing because the delivery never happened at all.
3. CWhen the column is stored as floating point rather than integer.
4. DWhen the mean is computed after dropping the missing rows.
Show the answer
Dropping and filling are both defensible, and what decides between them is what the absence means and what the analysis is for. A delivery that never happened has no delivery time, so inventing one adds a fact that never existed and pulls the average towards it. The storage type, how the mean was computed and a percentage threshold are not what makes the choice right or wrong.

Next → 18 / 25
Q19 / 25

The sales table holds one row per sale. You run a query that selects region and SUM(amount) grouped by region. What comes back?
1. AOne row per region, holding the amount of that region's largest sale.
2. BA single row holding the sum of every sale in the whole table.
3. COne row for each region, holding the total of that region's sales.
4. DOne row per sale, with the total for its region repeated on each row.
Show the answer
An aggregate reduces a set of input values to one value, and GROUP BY decides which sets those are, so the result has one row per region. Keeping every sale row next to its region total is what a window function does instead. Without GROUP BY you would get the single overall total, and the largest sale would need MAX.

Next → 19 / 25
Q20 / 25

You reach the correct result and present a clean final table, but you worked in silence. The panel says it cannot score your problem solving. What was missing?
1. AA chart that presents the final table more clearly to the panel.
2. BA second method that lands on the same number as the first.
3. CTalking through the approach and why you chose each tool.
4. DA faster solution that would have left time for questions.
Show the answer
A correct output shows where you arrived, not how you got there, and interviewers also assess how you approach the problem. Naming the tool and the reason for it makes that visible. A cross-check, a nicer chart or spare time all improve the delivery while leaving the reasoning hidden.

Next → 20 / 25
Q21 / 25

In a 45 minute round, a candidate codes for 40 minutes and presents an untested result. What did the plan get wrong?
1. ANo time was reserved to test the result before presenting it.
2. BThe candidate should have written the plan down before starting to code.
3. CThe candidate should have asked for the remaining minutes to be extended.
4. DThe candidate chose an approach that was too slow for the time available.
Show the answer
The candidate used the last five minutes to present a result that nobody had tested, and a solution that runs can still fail on a boundary case. Budget the minutes across planning, implementing, checking and explaining before you start. Writing the plan down helps, the approach may have been fine, and the clock is part of the exercise.

Next → 21 / 25
Q22 / 25

Five minutes are left and the data cannot settle whether the new page caused the drop. The interviewer asks what you would do next.
1. APropose a full experiment on checkout next quarter.
2. BName the smallest check that would separate the two explanations.
3. CPropose collecting more of the same data over a longer observation window.
4. DPropose reverting the page and watching whether the metric recovers.
Show the answer
Under a time box the answer that scores is the cheapest evidence that would change your mind, such as comparing exposed and unexposed users who are otherwise alike. A full experiment could test the same question but cannot report back inside the time box, more of the same data resolves nothing that is confounded, and reverting is an expensive experiment with no control.

Next → 22 / 25
Q23 / 25

You have told the situation, the task, your actions and the result. The interviewer waits. What closes the story?
1. AAn offer to go deeper into the technical part of the work.
2. BA question about how the team here handles similar problems.
3. CA summary of the result, stated once more in business terms.
4. DWhat you learned, and what you did differently the next time.
Show the answer
Reflection is the part the interviewer cannot infer from the events, and saying what changed in your later work is what shows the lesson was real rather than decorative. Repeating the result adds nothing, offering depth hands the initiative back, and a question about the team is a good question at the wrong moment.

Next → 23 / 25
Q24 / 25

You have two minutes to give a recommendation from an experiment whose result is small and uncertain. What does the recommendation have to contain?
1. AA clear ship or do not ship verdict, with the detail left for the written note.
2. BThe p-value and the sample size, which together summarise the evidence.
3. CThe method and the assumptions, so the audience can judge the analysis itself.
4. DThe size of the effect, its uncertainty, and what the decision needs.
Show the answer
A decision needs to know how big the effect is, how sure you are about it, and whether that is enough for the choice in front of the team. A p-value with a sample size leaves the magnitude out, a bare verdict hides the uncertainty the team is taking on, and the method matters but it is not what the two minutes are for.

Next → 24 / 25
Q25 / 25

Total signups are flat month on month, and the team concludes nothing changed. You have the same data split by acquisition channel. What do you check?
1. AWhether the monthly total was computed with the same date filter both times.
2. BWhether the flat total is within the normal week to week variation.
3. CWhether the signup event was renamed in the tracking during the month.
4. DWhether the channels moved in opposite directions and cancelled out.
Show the answer
A total is a sum, and a sum can stay still while its parts move a long way in opposite directions, so a flat total is not evidence that nothing happened. Splitting by a segment you already have is the cheapest way to find out. The other three are sensible data checks, but they ask whether the total is correct rather than whether it is hiding something.

Next → 25 / 25
That’s the whole mock exam

Every question you miss comes back exactly when you’re about to forget it.

Back to the course

How to use this mock exam

Sit all 25 questions in one go: the mix covers every domain in the same proportion as the exam, so a low score points at the domain you skipped rather than at bad luck.

Read the explanation under every question, including the ones you got right: the reason an option is wrong is usually the thing being tested.

Then retake it in the app, where the mock exam is timed and scored and the questions you miss come back on a schedule.

The whole course, on your phone

Lessons you can read, audio you can listen to on the way to work, and practice that remembers what you got wrong.

“This app is the ‘One Lake’ for preparing the certification exams. Really worth it.”

Abdelhak Y. Google Play review

Data Analyst Interview Prep final test: 25 free questions

Halfway, at your pace

That’s the whole mock exam

How to use this mock exam

The whole course, on your phone