Select Two (Multi-Select) Questions

Rely solely on color to convey information

Why it fails: Colour-only encoding excludes users with colour vision deficiency and anyone viewing greyscale prints, so meaning is lost. It is tempting because colour separates categories cleanly on screen, yet accessible design pairs colour with labels, patterns or shapes so information survives without hue perception.

Use 3D effects to make charts visually appealing

Why it fails: 3D effects distort lengths and angles, so values are misread, and they add visual clutter that impedes screen readers and low-vision users. It is tempting because perspective appears engaging in presentations, but accessible charts use flat, two-dimensional encodings with clear axes and labels.

Include animated transitions between views

Why it fails: Animated transitions add motion that can trigger vestibular discomfort and cannot be paused by screen-reader users, so they undermine accessibility rather than support it. Animation is tempting for demonstrating state changes or drawing attention in presentations, where motion aids engagement for sighted audiences without vestibular or cognitive sensitivities.

Both groups must contain at least 30,000 users to satisfy the central limit theorem.

Why it fails: There is no fixed 30,000-user requirement. The central limit theorem concerns the sampling distribution of the mean, and for many watch-time distributions a few hundred observations per group already yields an approximately normal sampling distribution. Imposing an arbitrary large threshold would make most realistic experiments impossible and is not a stated assumption of the two-sample t-test.

Watch time must be measured on a ratio scale with a true zero and no upper bound.

Why it fails: The t-test requires an approximately interval-scaled continuous outcome, not strictly a ratio scale with an unbounded maximum. Minutes of watch time per week are continuous and have a meaningful zero, but even an artificial ceiling in the measurement would not by itself invalidate the test. Measurement scale is therefore not the binding condition the analyst must verify.

The control and treatment groups must have identical means before the experiment begins.

Why it fails: Requiring identical pre-experiment means is neither possible nor necessary. Random assignment makes groups probabilistically comparable, and any baseline difference is handled by the hypothesis test itself. Demanding equal starting means would only be relevant to a paired or repeated-measures design, not to the independent-groups comparison the analyst is running here.

Data can be categorized into groups

Why it fails: Categorization is a property of nominal/ordinal data, not specific to ratio.

Allows negative values

Why it fails: True zero implies no negative values (e.g., weight cannot be negative).

Data is processed in micro-batches

Why it fails: Micro-batch is a style, but not a defining characteristic of streaming.

Data is stored permanently before processing

Why it fails: Streaming data is typically processed on arrival.

Retain data indefinitely for analysis

Why it fails: Indefinite retention breaches the GDPR storage-limitation principle, which requires personal data be kept no longer than necessary for its stated purpose. It is tempting because long retention supports historical trend analysis, and it would be acceptable only where a documented legal or regulatory obligation specifies that retention period.

Include customer names for context

Why it fails: Including customer names exceeds the data-minimisation principle: satisfaction reporting needs only the scores, not directly identifying personal data. It is tempting because names give context and aid interpretation, and it would be justified only where identity is genuinely necessary and a lawful basis plus appropriate safeguards exist.

Select every 10th transaction from the entire dataset

Why it fails: Selecting every 10th transaction is systematic sampling, which draws from the whole dataset without partitioning by the category column. Stratified sampling requires splitting into strata first, then sampling within each. Systematic sampling suits ordered lists where even coverage across a sequence is wanted, not proportional representation across groups.

Combine all categories into a single group and perform simple random sampling

Why it fails: Pooling categories into one group and sampling randomly destroys the strata, so proportional representation by category cannot be guaranteed. Stratified sampling keeps each category as a separate stratum and samples within it. Simple random sampling over a homogeneous population is the right choice when subgroup proportions do not matter.

Use simple random sampling on the whole population

Why it fails: Simple random sampling draws from the entire population with equal probability, so no subgroups are formed and proportional representation is not enforced. It is tempting because it is the baseline probability method, but stratified sampling requires first dividing the population into strata, then sampling randomly within each one.

Randomly select entire clusters of customers

Why it fails: Cluster sampling randomly selects whole groups and then samples every member within them, so strata are never defined or sampled proportionally. It is tempting because both methods partition a population into groups, but clustering serves convenience and cost reduction, whereas stratification requires splitting by a shared characteristic before random selection within each stratum.

Select every nth customer from a list

Why it fails: Systematic sampling picks every nth record from one ordered list, which produces a single sample rather than independent random draws inside defined strata. It is tempting because it is a probability method that is quick to apply, yet it cannot guarantee proportional representation of each subgroup the way stratified sampling demands.

Use data compression

Why it fails: Compression shrinks file size but retains every record, so it reduces storage cost without enforcing deletion or expiry of data past its legal retention window. It is tempting because it lowers footprint cheaply, and it suits long-lived archives that must be kept, but retention policy requires lifecycle rules and defensible disposal.

Increase primary storage capacity

Why it fails: Increasing primary storage capacity raises cost and retains data indefinitely, directly contradicting the retention policy's aim of reducing storage costs. It is tempting because capacity expansion addresses storage pressure, and would suit scenarios where data must be preserved long-term for active access rather than aged out under legal schedules.

Implement data deduplication

Why it fails: Deduplication removes duplicate blocks but keeps all retained records, so it does not implement expiry or deletion against legal retention periods. It is tempting because it cuts storage substantially, and it suits backup repositories holding repeated copies, but a retention policy needs lifecycle rules that actually purge data on schedule.

A foreign key must have the same name as the primary key it references

Why it fails: Foreign keys can have different names.

A foreign key can reference a column that is not a primary key

Why it fails: Foreign keys typically reference a primary key or unique constraint.

A table can have multiple primary keys

Why it fails: A table can have only one primary key, but it can be composite.

Create a dashboard for stakeholders

Why it fails: A dashboard presents aggregated results to stakeholders, producing no metadata about null counts, key uniqueness or data types across the source files. Dashboards are the right choice once data is loaded and modelled, not during acquisition, when the analyst must first understand each file's structure and quality.

Build a linear regression model

Why it fails: Regression modelling predicts a continuous target from historical data, so it cannot assess source-file completeness, duplicates or value distributions before loading. It is tempting because regression is a core analytics technique, but it belongs after profiling, when relationships between cleaned variables are being quantified.

Perform cluster analysis

Why it fails: Cluster analysis groups records by similarity, which requires already-loaded, cleansed and scaled data; it reveals nothing about column-level completeness, duplicates or format inconsistencies. Clustering is correct later for segmentation or anomaly discovery, but profiling must precede any such modelling.

Avoid mentioning uncertainty to maintain confidence

Why it fails: Omitting uncertainty misrepresents statistical results; a knowledgeable audience expects confidence intervals and significance levels to judge reliability. It is tempting because confident delivery appears authoritative, and would be correct for non-technical executives needing a decisive headline rather than statistical rigour.

Use basic visualizations like pie charts

Why it fails: Pie charts cannot represent distributions, correlations or multivariate relationships central to complex statistical analysis. They are tempting because they are familiar and easy to read, and would be correct for showing simple part-to-whole proportions to a general audience.

Simplify the findings to avoid confusion

Why it fails: Simplifying findings discards the statistical detail this expert audience needs to evaluate methodology and results. It is tempting because simplification aids comprehension generally, and would be correct when presenting to business stakeholders lacking statistical background who require only headline conclusions.

Assume the anomaly is real and report it

Why it fails: Treating an anomaly as genuine without verification risks publishing fabricated findings built on corrupt or duplicated records. It is tempting because anomalies sometimes reveal real events, and reporting it would be correct once validated through source checks or stakeholder confirmation, but assumption bypasses the investigation the stem requires before reporting.

Immediately alert all stakeholders

Why it fails: Alerting all stakeholders before validating the anomaly risks spreading false alarms and erodes trust if the data proves correct. Notification is tempting because incident communication is a genuine duty, but it belongs after verification and impact assessment, not as a first investigative step.

Remove the anomaly without documentation

Why it fails: Deleting an anomaly silently destroys the audit trail and prevents reviewers from validating whether it reflects a data-entry error, fraud or a genuine event. It is tempting because it produces a clean dataset quickly, and it would be acceptable only after documented investigation and approval, but undocumented removal breaches analytical integrity and reproducibility.

The data is paired between the two locations

Why it fails: Paired data would require a paired t-test, not a two-sample t-test.

The sample sizes are equal

Why it fails: Equal sample sizes are not a required condition for the two-sample t-test.

The variances of the two populations are equal

Why it fails: Equal variances are required only for the pooled t-test; Welch's t-test does not require this assumption.

Random color scheme for every chart

Why it fails: Arbitrary colours add visual noise and imply false groupings, obscuring the decline the narrative must convey. A narrative arc requires consistent, purposeful encoding such as a single highlight colour for the falling quarter. Random palettes suit exploratory charting where no message is being communicated, not a structured story.

Detailed description of each product's sales breakdown

Why it fails: Per-product breakdowns are supporting detail, not a narrative element; the arc needs a hook, tension and resolution or call to action. Granular product figures belong in an appendix or drill-down view. They would be the right inclusion when the audience needs diagnostic evidence behind the decline rather than a persuasive story.

Increase data freshness to real-time

Why it fails: Real-time freshness changes how current the data is, not how often the same analysis is requested; the recurring demand persists. It is tempting because latency is a common complaint, and streaming ingestion would be correct where decisions depend on sub-minute data, but here the requirement is eliminating repeat requests through reusable reporting.

Add more security to the data

Why it fails: Restricting data access does not reduce demand; analysts still receive the same requests, merely with narrower visibility. It tempts as a governance control, which is its actual purpose. Reducing ad hoc volume comes from publishing reusable dashboards or self-service reports that requesters can consult directly.

Ignore the requests until they become urgent

Why it fails: Ignoring requests does not reduce their volume; it simply delays delivery and damages stakeholder trust, leaving the underlying recurring need unaddressed. It is tempting as a triage tactic when capacity is exhausted, but the correct approach is to build a reusable self-service report or dashboard that answers the repeated question directly.

Employee turnover rate

Why it fails: Employee turnover rate is lagging: it records departures that have already happened, reflecting past workforce conditions rather than forecasting them. It is tempting because HR metrics are often discussed as predictive, and turnover is the correct indicator when assessing retention outcomes after the fact.

Net profit margin

Why it fails: Net profit margin is lagging: it is calculated from completed revenue and costs, so it summarises performance already realised. It is tempting because profitability is treated as a health signal, and it is the correct measure when the question asks for financial results reported after a period closes.

Monthly revenue

Why it fails: Monthly revenue is a lagging indicator: it reports income already earned, so it confirms past performance rather than predicting it. It is tempting because revenue growth feels forward-looking, and it is the correct measure when the question asks for outcomes used to evaluate results already achieved.

Supports only structured data

Why it fails: A data lake stores structured, semi-structured and unstructured data alike, so restricting it to structured data contradicts its purpose. It is tempting because structured-only storage is a genuine characteristic of a data warehouse, and would be correct when describing warehouse capabilities rather than lake capabilities.

Requires data transformation before loading

Why it fails: A data lake loads raw data without requiring prior transformation, applying schema on read instead. It is tempting because transformation before loading is the defining trait of a data warehouse ETL process, and would be correct when describing warehouse ingestion rather than lake ingestion.

Mean imputation

Why it fails: Mean imputation replaces missing values with the column average, altering the distribution and pulling extreme points toward the centre, which masks outliers rather than revealing them. It is tempting because imputation is a routine preprocessing step, and analysts often clean data before profiling it.

Pearson correlation coefficient

Why it fails: Pearson's coefficient measures linear association between two continuous variables, returning a value between -1 and 1; it cannot flag individual extreme values. It is tempting because correlation matrices are standard EDA output, and a coefficient near zero can hint at an outlier distorting the relationship.

Standard deviation alone

Why it fails: Standard deviation alone assumes a roughly normal distribution and a single symmetric spread; skewed or multimodal data produce thresholds that misclassify legitimate values. It is tempting because the three-sigma rule is widely taught, and it works acceptably when the feature is genuinely Gaussian.

A PDF report with free-form text

Why it fails: PDF text is often unstructured.

A JPEG image of a product

Why it fails: Image is unstructured.

A JSON file with nested key-value pairs

Why it fails: JSON is semi-structured.

Delete any rows with null values without review

Why it fails: Deleting rows with nulls unreviewed destroys valid records and hides systematic gaps, so the assessment never measures completeness or bias. It tempts because nulls often signal defects, yet removal belongs to cleaning after profiling; assessment first quantifies how many values are missing and why.

Immediately load all data into production

Why it fails: Loading unvalidated data into production propagates errors downstream before any profiling, so quality issues reach consumers unchecked. It tempts as the end goal of a pipeline, yet production loading is correct only once assessment has confirmed the dataset meets defined accuracy, completeness and consistency thresholds.

Transform data to match target system without verification

Why it fails: Transforming data before verifying it propagates errors and destroys the ability to trace discrepancies back to the source, so it defeats the assessment. Transformation is tempting because it moves the work toward the target system. It would be appropriate only after profiling and validation confirm the source data's accuracy and completeness.