Top 30 Intermediate Data Science & Analytics Interview Questions and Answers

30 questions. Pick an answer, or open “Show the answer” to read it.

Answers are graded in your browser. Nothing is saved, and no XP or streak is earned here. The app keeps score.

1. Question 1 of 30

   A monthly subscriber cancels mid-cycle but retains access until month-end. Which approach reflects a rigorous churn definition for this user?

   • aWaiting until the billing period ends or a reactivation window passes
   • bCounting churn the instant the user clicks cancel
   • cFlagging churn immediately when the next payment fails
   • dLabeling churn after 30 days of inactivity while billing continues
   Show the answer

   Answer: a · Waiting until the billing period ends or a reactivation window passes

   A rigorous definition ties churn to the actual end of access or the close of a defined renewal window, not just the intent signal. Counting the cancellation click immediately is the naive error highlighted in the card, because the user may still revert and the revenue is still active through the billing period.

   Read the full bite: How do you define churn for a subscription service?

2. Question 2 of 30

   A checkout A/B test shows significantly higher conversion but slightly lower AOV. What is the strongest basis for a launch recommendation?

   • aLaunch because conversion is up and automatically outweighs any decrease in basket size
   • bRequire AOV to also achieve statistical significance as a co-success metric
   • cAdjust p-values for multiple comparisons and launch only if both metrics remain significant
   • dEstimate net revenue and ensure the AOV decline is within a pre-specified non-inferiority margin
   Show the answer

   Answer: d · Estimate net revenue and ensure the AOV decline is within a pre-specified non-inferiority margin

   The right framework classifies conversion as a success metric and AOV as a guardrail requiring non-inferiority, then uses net revenue to judge the business outcome. Treating AOV as a co-success metric that must significantly increase is a tempting error that causes unnecessary conservatism and missed wins.

   Read the full bite: How would you recommend launching a checkout flow with mixed A/B metrics?

3. Question 3 of 30

   A marketing team debates predicting exact spend versus High/Medium/Low tiers. What most strongly determines whether regression or classification is the better framing?

   • aWhether the downstream campaign action requires a ranked list and continuous segmentation or a hard gate into fixed segments.
   • bThe preference of business stakeholders for yes-or-no answers over probabilistic scores.
   • cThe algorithm that achieves the highest cross-validated R-squared or F1 score on historical data.
   • dThe degree of class imbalance in the historical customer base, since severe imbalance favors regression.
   Show the answer

   Answer: a · Whether the downstream campaign action requires a ranked list and continuous segmentation or a hard gate into fixed segments.

   The campaign action dictates whether you need a ranked continuous score or a discrete gate, making B correct. D is tempting but dangerous because choosing classification solely for yes/no simplicity ignores the information loss from binarizing a continuous signal and may mismatch flexible budget allocation.

   Read the full bite: How do you frame high-value customer identification as classification versus regression?

4. Question 4 of 30

   What justifies using a z-test for a population mean when the underlying data are heavily skewed?

   • aThe sampling distribution of the sample mean becomes approximately normal for large n
   • bThe sample mean converges to the population mean as sample size increases
   • cThe raw observations in large samples themselves become normally distributed
   • dThe population variance is known and the sample size exceeds thirty
   Show the answer

   Answer: a · The sampling distribution of the sample mean becomes approximately normal for large n

   The CLT states that the sampling distribution of the sample mean approaches normality as n grows, which justifies using z-tests even when the population is skewed. Option B describes the Law of Large Numbers, a common look-alike that explains convergence to a single value rather than the bell-curve shape required for inference.

   Read the full bite: Explain the Central Limit Theorem and its importance for hypothesis testing

5. Question 5 of 30

   Which statement correctly describes how increasing model complexity affects bias and variance?

   • aBias decreases while variance stays low until the model passes an overfitting threshold and variance spikes.
   • bBias decreases but variance increases because the fit depends more on the specific training sample.
   • cBias decreases while variance stays unchanged, because variance depends only on training set size.
   • dBoth bias and variance decrease as the model fits the training data more closely.
   Show the answer

   Answer: b · Bias decreases but variance increases because the fit depends more on the specific training sample.

   The card explains that increasing complexity reduces bias by fitting training data more closely but simultaneously increases variance by making predictions dependent on the specific sample. Option D is the most tempting distractor because it reflects the common red-flag misconception that adding parameters improves both error sources.

   Read the full bite: Describe the bias-variance tradeoff and how model complexity affects bias and variance

6. Question 6 of 30

   A medical model reports 99% accuracy on a disease with 1% prevalence. Which observation best shows why accuracy is misleading?

   • aThe model could predict every patient is healthy and still score 99%.
   • bPrecision and recall become identical when the positive class is rare.
   • cThe dataset is too small to yield a reliable accuracy estimate.
   • dHigh accuracy proves the model has a low false positive rate.
   Show the answer

   Answer: a · The model could predict every patient is healthy and still score 99%.

   When prevalence is 1%, an all-negative classifier automatically achieves 99% accuracy while detecting zero actual cases, proving accuracy can mask total failure. Distractor A is a common misconception the card explicitly warns against, as the problem is class imbalance rather than sample size.

   Read the full bite: Why is 99% accuracy misleading with 1% disease prevalence?

7. Question 7 of 30

   What happens when the learning rate in gradient descent is set too high?

   • aSteps overshoot the minimum and the loss may oscillate or diverge
   • bThe gradient is computed in the wrong direction
   • cThe model converges to the global minimum guaranteed
   • dTraining always reaches the minimum faster with no downside
   Show the answer

   Answer: a · Steps overshoot the minimum and the loss may oscillate or diverge

   Too large a step size overshoots the minimum, causing oscillation or divergence instead of convergence. A high rate does not guarantee faster or correct convergence, and it does not change how the gradient is computed.

   Read the full bite: How gradient descent and the learning rate work

8. Question 8 of 30

   You need to combine customers and transactions on customer_id to compute total spend per customer, including those with zero transactions. Which approach correctly uses pandas?

   • aUse pd.concat to align the tables on customer_id, then group by customer and sum amounts
   • bUse pd.merge with how='left' on customer_id, then group by customer, sum amounts, and fill NaN totals with zero
   • cUse pd.merge with how='inner' on customer_id, then group by customer and sum amounts
   • dUse pd.merge with how='right' on customer_id, then group by customer and sum amounts
   Show the answer

   Answer: b · Use pd.merge with how='left' on customer_id, then group by customer, sum amounts, and fill NaN totals with zero

   A left join preserves every customer and shows NaN for missing transactions, which you fill with zero after grouping; an inner join is tempting because it is the pandas default, but it silently drops customers with no matching transactions.

   Read the full bite: How would you combine customer and transaction DataFrames and describe join types?

9. Question 9 of 30

   When squaring a million-element NumPy array, why is np.square(arr) orders of magnitude faster than a Python for-loop?

   • aIt automatically distributes the computation across multiple CPU threads.
   • bIt delegates to pre-compiled C loops on contiguous memory, avoiding Python interpreter overhead per element.
   • cIt removes explicit loops by expressing the entire operation in a single Python statement.
   • dIt uses Python list comprehensions that the interpreter optimizes for numerical arrays.
   Show the answer

   Answer: b · It delegates to pre-compiled C loops on contiguous memory, avoiding Python interpreter overhead per element.

   NumPy vectorization dispatches operations to C ufuncs on contiguous buffers, eliminating per-element Python interpreter dispatch and object boxing. The most tempting distractor confuses vectorization with parallel processing, yet the card explicitly states that standard NumPy vectorization is single-threaded unless an external multithreaded library is used.

   Read the full bite: What is vectorization in NumPy and pandas?

10. Question 10 of 30

    Which single expression best computes total and average Sales_Amount per Region in pandas?

    • aApply a lambda that returns a Series with sum and mean for each group
    • bSelect Sales_Amount after groupby and use agg with named aggregations
    • cChain two separate groupby operations for sum and mean, then join the results
    • dLoop through Region groups with a for-loop and accumulate sums and counts manually
    Show the answer

    Answer: b · Select Sales_Amount after groupby and use agg with named aggregations

    Selecting the column and using named aggregation computes both statistics in a single vectorized pass with clean column names. Option C is a common anti-pattern that runs two separate groupby passes and requires manual merging, which is slower and harder to maintain.

    Read the full bite: Calculate total and average sales per region in pandas

11. Question 11 of 30

    Which code correctly and efficiently extracts weekday names like 'Monday' from a pandas DataFrame column of timestamp strings?

    • apd.to_datetime(df['ts']).day_name()
    • bpd.to_datetime(df['ts']).dt.day_name()
    • cdf['ts'].apply(lambda x: pd.to_datetime(x).day_name())
    • ddf['ts'].day_name()
    Show the answer

    Answer: b · pd.to_datetime(df['ts']).dt.day_name()

    pd.to_datetime vectorizes parsing to datetime64 without Python loops, and the .dt accessor exposes day_name(); wrapping conversion in apply processes rows one by one and destroys performance, while calling day_name directly on a string or Series raises an AttributeError.

    Read the full bite: Convert string timestamps to datetime and extract day of week

12. Question 12 of 30

    Which client-side strategy best maximizes throughput for a 100 req/min API without triggering excessive 429 errors?

    • aLimit concurrency to four workers but send requests at full speed, using only reactive exponential backoff and ignoring rate-limit headers.
    • bBound concurrency to a small worker pool, proactively pace requests using rate-limit headers, and apply exponential backoff with jitter on 429s.
    • cKeep an unbounded number of workers and rely solely on reactive exponential backoff after receiving 429 responses.
    • dInsert a fixed one-second delay after every request and retry 429s with linear backoff while ignoring rate-limit headers.
    Show the answer

    Answer: b · Bound concurrency to a small worker pool, proactively pace requests using rate-limit headers, and apply exponential backoff with jitter on 429s.

    The correct answer combines proactive throttling, header-aware dynamic pacing, bounded concurrency, and resilient retries as described in the card. Option A is tempting because it limits concurrency, but it still creates burst traffic and ignores headers, relying on the server to punish the client rather than preventing 429s proactively.

    Read the full bite: Design a rate-limited REST API data collection script

13. Question 13 of 30

    When scraping a React site where the initial HTML lacks the target data, which tool and technique combination is most reliable?

    • aSelenium with a fixed ten-second sleep after loading the page before reading the source
    • brequests paired with BeautifulSoup, retrying the HTTP request until the element is found
    • cA headless browser that extracts the page source immediately after the navigation event fires
    • dPlaywright using an explicit wait for the target element to appear in the rendered DOM
    Show the answer

    Answer: d · Playwright using an explicit wait for the target element to appear in the rendered DOM

    Dynamic JavaScript frameworks inject content after the initial document loads, so only a real browser engine with an explicit wait can reliably capture it. While Selenium is a valid browser tool, using a fixed sleep makes the scraper slow and flaky under variable network conditions rather than waiting for the element to actually appear.

    Read the full bite: How would you scrape a page with dynamically loaded JavaScript content?

14. Question 14 of 30

    When designing a pipeline to ingest millions of semi-structured mobile events per second, why is NoSQL typically chosen for the initial landing zone?

    • aIt requires multi-row ACID transactions to guarantee exactly-once ingestion of immutable logs.
    • bIt performs relational joins and complex aggregations faster than a data warehouse on raw JSON.
    • cIt enforces strict schemas upfront to prevent bad data from entering the pipeline.
    • dIt allows schema flexibility and horizontal write scaling without blocking ingestion during field additions.
    Show the answer

    Answer: d · It allows schema flexibility and horizontal write scaling without blocking ingestion during field additions.

    NoSQL is preferred because document stores handle new fields instantly and scale writes horizontally, avoiding the locking migrations a relational schema would require. The tempting ACID distractor is wrong because immutable append-only event logs do not need multi-row transactions at ingestion time.

    Read the full bite: SQL or NoSQL for high-volume semi-structured event ingestion?

15. Question 15 of 30

    A Python daemon runs nightly without user interaction to pull API data on its own behalf. Which OAuth 2.0 approach should it use to obtain an access token?

    • aUse the resource owner password credentials grant, exchanging a service account username and password at the token endpoint.
    • bUse the client credentials grant, authenticating directly to the token endpoint with its client_id and client_secret.
    • cUse the authorization code grant, redirecting a user to the authorization endpoint and storing a refresh token for later runs.
    • dUse the implicit flow, receiving an access token directly from the authorization endpoint via browser redirect.
    Show the answer

    Answer: b · Use the client credentials grant, authenticating directly to the token endpoint with its client_id and client_secret.

    A daemon acting on its own behalf should use the client credentials grant and authenticate directly to the token endpoint with its client_id and client_secret. The authorization code flow with a stored refresh token is designed for scripts acting on behalf of a user, not for non-interactive service-to-service calls.

    Read the full bite: Implement OAuth 2.0 flow to get an access token for API requests

16. Question 16 of 30

    When you fit a StandardScaler on the full dataset before cross-validation, what is the specific leakage problem?

    • aTarget values from the validation fold leak into the model's loss function
    • bDuplicate rows between folds cause the scaler to compute identical z-scores
    • cValidation fold statistics influence the training phase through the global mean and standard deviation
    • dThe scaler overfits to the training fold because it memorizes every sample value
    Show the answer

    Answer: c · Validation fold statistics influence the training phase through the global mean and standard deviation

    Fitting the scaler on the entire dataset incorporates validation fold statistics into the global mean and standard deviation, constituting feature leakage. Distractor C describes target leakage, which involves the target variable, whereas the card emphasizes feature leakage through preprocessing steps like scaling.

    Read the full bite: What is data leakage in preprocessing and cross-validation?

17. Question 17 of 30

    A decision tree's split on a feature remains unchanged if that feature is multiplied by 100 because...

    • aDecision trees optimize Euclidean distances between samples, which are scale-invariant under linear transformations
    • bInformation gain and Gini impurity depend only on the relative ordering of feature values, not their absolute scale
    • cFeature magnitude is absorbed by the regularization term in tree-based loss functions
    • dThe tree internally standardizes features at each node before evaluating splits
    Show the answer

    Answer: b · Information gain and Gini impurity depend only on the relative ordering of feature values, not their absolute scale

    Decision trees split based on rank order, so rescaling a feature leaves information gain and Gini impurity unchanged. The distractor about Euclidean distances confuses combinatorial tree splitting with the geometric distance objectives that make scaling critical for SVMs and K-Means.

    Read the full bite: Why is scaling unnecessary for trees but critical for SVM or K-Means?

18. Question 18 of 30

    Which practice best ensures safe recovery from a mid-job crash in an idempotent daily batch pipeline loading API data into a partitioned data lake?

    • aUse an idempotency key per partition, checkpoint progress, and atomically overwrite only failed partitions on retry
    • bAppend each run's output with a unique run ID and have downstream queries select the most recent batch
    • cRerun the complete job after any failure and trust the pipeline to handle duplicates automatically
    • dDelete all data in the target partition before reloading to ensure a clean slate
    Show the answer

    Answer: a · Use an idempotency key per partition, checkpoint progress, and atomically overwrite only failed partitions on retry

    Checkpointing and atomic overwrites let retries resume exactly where they left off without corrupting data or exposing partial states. Deleting the partition first seems clean but creates a window where downstream consumers see missing data.

    Read the full bite: What is data pipeline idempotency and how do you design for it?

19. Question 19 of 30

    Analysts query an events table with date-range filters and event_type filters. Which layout best minimizes bytes scanned?

    • aPartition by event_date without clustering to reduce metadata
    • bPartition by event_date and cluster by event_type and user_id
    • cPartition by user_id and cluster by event_date
    • dPartition by event_type and cluster by event_timestamp
    Show the answer

    Answer: b · Partition by event_date and cluster by event_type and user_id

    Partitioning by event_date prunes irrelevant time ranges, while clustering by event_type and user_id speeds up filters within each day. Partitioning by user_id is a classic anti-pattern that explodes metadata with millions of tiny partitions and does not help time-based queries.

    Read the full bite: How would you partition a massive user events table?

20. Question 20 of 30

    Which pairing best illustrates when streaming is genuinely justified over batch, alongside a critical infrastructure challenge that arises specifically in streaming?

    • aMonthly financial reporting; handling duplicate data during ingestion
    • bReal-time ad bidding; scheduling nightly downtime for operational maintenance
    • cFraud detection during payment authorization; managing backpressure during traffic spikes
    • dEnd-of-day inventory reconciliation; checkpointing stateful operations for recovery
    Show the answer

    Answer: c · Fraud detection during payment authorization; managing backpressure during traffic spikes

    Fraud detection requires sub-minute reaction time that batch cannot provide, and backpressure is a core operational challenge specific to streaming. Distractor D correctly identifies a low-latency use case but incorrectly assumes streaming pipelines have nightly downtime windows, whereas they actually impose a 24/7 operational burden with no batch-style maintenance window.

    Read the full bite: When is streaming better than batch, and what are its infrastructure challenges?

21. Question 21 of 30

    You detect a multivariate outlier formed by a rare feature combination. What is the appropriate first step before choosing whether to keep or remove it?

    • aWinsorize all features to cap the observation's joint influence on the model
    • bInvestigate whether it results from measurement error or a legitimate extreme event
    • cApply univariate three-sigma rules to each feature to confirm the anomaly
    • dRun sensitivity analyses with and without the point to gauge its influence
    Show the answer

    Answer: b · Investigate whether it results from measurement error or a legitimate extreme event

    The card emphasizes starting with a root-cause check to distinguish errors from legitimate extremes before any removal or modeling decision. Sensitivity analysis is reserved for highly influential points only after this initial investigation, making it the most tempting distractor because it is valid but premature as a first response.

    Read the full bite: How do you detect and handle outliers in multivariate datasets?

22. Question 22 of 30

    You standardize 100 numerical features and find PCA's first two components capture 90% of variance with visually distinct clusters. What is the most appropriate next step?

    • aUse only the top two PCA components for downstream clustering to eliminate the remaining noise.
    • bRely primarily on PCA for visualization, as t-SNE adds little value when clusters are linearly separable with high variance capture.
    • cSet t-SNE perplexity proportional to the variance ratio to preserve the global structure in the embedding.
    • dRun t-SNE to confirm the clusters, since PCA's linear projection can miss true non-linear boundaries.
    Show the answer

    Answer: b · Rely primarily on PCA for visualization, as t-SNE adds little value when clusters are linearly separable with high variance capture.

    When PCA captures most variance and reveals clear clusters, the structure is globally linear, so t-SNE is unnecessary and can mislead. Option D is tempting but wrong because it assumes PCA's linear projection is insufficient and falsely requires t-SNE validation, losing the global variance context.

    Read the full bite: How do you visualize clusters in 100-dimensional numerical data?

23. Question 23 of 30

    EDA reveals a U-shaped relationship between user_age and monthly_spend. Which approach best enables a linear regression to capture this pattern?

    • aCenter age around its mean so the linear term models both directions
    • bApply a log transform to age to flatten the curve
    • cRemove age because its Pearson correlation with spend is near zero
    • dIntroduce an age-squared term or spline expansion
    Show the answer

    Answer: d · Introduce an age-squared term or spline expansion

    A linear model needs explicit non-linear terms such as age-squared or splines to fit a direction-reversing U-shape. Centering age alone cannot create curvature; it only helps reduce multicollinearity after the squared term is added.

    Read the full bite: How does a U-shaped EDA pattern influence feature engineering?

24. Question 24 of 30

    How do L2 regularization and shallow trees affect the bias-variance tradeoff according to the decomposition?

    • aThey reduce irreducible noise by preventing the model from fitting random training fluctuations.
    • bThey lower variance by accepting slightly higher bias to constrain model behavior.
    • cThey reduce bias by shrinking coefficients so rare features do not dominate predictions.
    • dThey minimize both bias and variance simultaneously as complexity is reduced.
    Show the answer

    Answer: b · They lower variance by accepting slightly higher bias to constrain model behavior.

    Regularization and depth limits constrain the hypothesis space, which reduces variance by preventing memorization of training noise while slightly increasing bias from missed interactions. The distractor describing coefficient shrinkage as bias reduction is wrong because these techniques intentionally increase bias to lower variance, not the other way around.

    Read the full bite: Explain bias-variance tradeoff and how regularization or tree depth manage it

25. Question 25 of 30

    When encoding a nominal categorical feature for logistic regression versus LightGBM, what is the key mechanical difference driving the choice?

    • aLogistic regression learns weights that imply direction and magnitude across a continuous space, so one-hot encoding is needed to avoid false ordinality.
    • bTree models assume linear relationships between ordinally encoded values, making them unsuitable for nominal data.
    • cLightGBM requires integer inputs to build histograms, so ordinal encoding is mandatory for all categorical features.
    • dOne-hot encoding is avoided in LightGBM solely because it increases memory usage, not because of model mechanics.
    Show the answer

    Answer: a · Logistic regression learns weights that imply direction and magnitude across a continuous space, so one-hot encoding is needed to avoid false ordinality.

    Logistic regression computes dot products in a continuous input space where numeric distance between codes would falsely imply order and magnitude, making one-hot encoding necessary for nominal variables. Option C is a common red flag because LightGBM does not mandate ordinal encoding for every categorical feature and also offers native categorical support.

    Read the full bite: How would feature engineering for categoricals differ for logistic regression versus LightGBM?

26. Question 26 of 30

    What is the key reason gradient boosting can overfit more easily than a random forest?

    • aIt sequentially fits each tree to the prior ensemble's residual errors
    • bIt always uses fewer trees than a random forest
    • cIt uses bootstrap sampling of rows for every tree
    • dIt averages predictions from independent deep trees
    Show the answer

    Answer: a · It sequentially fits each tree to the prior ensemble's residual errors

    Boosting fits trees sequentially to remaining errors, so without regularization it can chase noise and overfit. A random forest averages independent trees, which reduces variance rather than aggressively fitting residuals.

    Read the full bite: Random Forest versus Gradient Boosting

27. Question 27 of 30

    What is the main reason k-fold cross-validation is more robust than a single train-test split for estimating generalization performance?

    • aIt acts as a regularization technique that prevents overfitting by exposing the model to diverse validation folds.
    • bIt averages multiple performance scores from different partitions, lowering estimate variance while using all data for both training and testing.
    • cIt guarantees an unbiased estimate of the true population error because every observation appears in a test fold exactly once.
    • dIt produces a single final model that has been trained on every fold, making it more robust than a model trained on just one training split.
    Show the answer

    Answer: b · It averages multiple performance scores from different partitions, lowering estimate variance while using all data for both training and testing.

    Cross-validation reduces variance in the performance estimate by averaging scores across k rotations where every observation serves in both training and validation. It does not create a single final model trained on all folds—its purpose is evaluation, not producing a production estimator.

    Read the full bite: What is cross-validation and why is it more robust than a holdout split?

28. Question 28 of 30

    You run a 30-day A/B test and peek daily, stopping early if p < 0.05. What is the principal statistical issue?

    • aStatistical power drops because the test was stopped before the planned sample size was reached
    • bThe overall false positive rate inflates because each look gives random noise another chance to appear significant
    • cThe Type I error rate is unaffected because each daily look still uses the same 5% alpha level
    • dEarly users differ systematically from later users, creating a selection bias that exaggerates the treatment effect
    Show the answer

    Answer: b · The overall false positive rate inflates because each look gives random noise another chance to appear significant

    Peeking gives random noise multiple chances to cross the significance threshold, compounding the false positive rate far above the nominal 5% level. Distractor D is tempting but wrong because a 5% alpha is only valid for a single pre-specified analysis, not for repeated uncorrected looks.

    Read the full bite: Why not stop an A/B test when it looks significant early?

29. Question 29 of 30

    When testing a social-network feature, why is cluster randomization along the social graph preferred over geographic clustering?

    • aGraph clusters remove the need for cluster-based regression adjustment during analysis
    • bGeography rarely maps to the interference mechanism, so cross-border spillover remains high
    • cGeographic clusters naturally capture most social-graph edges because friends tend to live nearby
    • dSocial-graph clusters fully contain all peer-to-peer interactions within each cluster
    Show the answer

    Answer: b · Geography rarely maps to the interference mechanism, so cross-border spillover remains high

    Social-graph clustering is preferred because geographic borders rarely capture the peer-to-peer edges that carry spillover, leaving control users contaminated. The intuition that friends live nearby is unreliable on global platforms, and graph clustering still does not eliminate all cross-cluster interference.

    Read the full bite: How do network effects violate A/B tests and how to mitigate them?

30. Question 30 of 30

    Why does heavy skew from whale users reduce the power of a standard t-test on revenue per user?

    • aIt biases the treatment assignment between groups
    • bIt makes the sample size effectively infinite
    • cIt violates the requirement that the metric be binary
    • dIt inflates variance, widening the standard error and confidence interval
    Show the answer

    Answer: d · It inflates variance, widening the standard error and confidence interval

    Extreme values raise the metric's variance, which enlarges the standard error and confidence interval, so real effects are harder to detect. It does not bias randomized assignment, which stays balanced in expectation.

    Read the full bite: Analyzing skewed revenue-per-user experiments