Data Analysis practice questions

Concatenate

Why it fails: Concatenate can be ambiguous; in pandas, it can combine along rows or columns, but append is more specific.

Merge

Why it fails: Merge is for combining on a common key, not simply adding rows.

Aggregate

Why it fails: Aggregate summarizes data, not combine rows.

Bar chart

Why it fails: Bar chart typically shows aggregate values, not distribution.

Histogram

Why it fails: Histogram shows distribution of one continuous variable without categories.

Scatter plot

Why it fails: Scatter plot is for two continuous variables.

No multicollinearity

Why it fails: Incorrect. No multicollinearity is an assumption for multiple regression, not for simple linear regression, and even in multiple regression it is not always considered a 'common' assumption in the same sense.

Normality of the dependent variable

Why it fails: Incorrect. Linear regression assumes normality of residuals, not the dependent variable.

Homoscedasticity

Why it fails: Incorrect. Although homoscedasticity is an important assumption for unbiased standard errors, it is not selected here as it is often not considered one of the most common assumptions for basic linear regression.

Durbin-Watson test for autocorrelation

Why it fails: Autocorrelation is not a core assumption for linear regression.

Variance inflation factor (VIF) for multicollinearity

Why it fails: Multicollinearity is a concern but not a core assumption.

Cook's distance to identify influential points

Why it fails: Influential points are diagnostics but not assumptions.

Duplicate customer IDs

Why it fails: All IDs appear unique.

Missing values in total_charges

Why it fails: No missing values are present.

Outliers in monthly_charges

Why it fails: Values like 99.65 are not extreme outliers.

Data Steward: Implements technical security controls.

Why it fails: Incorrect — implementing technical security controls is the role of a Data Custodian, not a Data Steward.

Data Owner: Ensures data quality and manages metadata.

Why it fails: Incorrect — ensuring data quality and managing metadata is the role of a Data Steward, not a Data Owner.

Min-max normalization

Why it fails: Min-max scales to [0,1] but can be skewed by outliers, and does not guarantee equal contribution.

Log transformation

Why it fails: Log transformation is used to reduce skewness, not for scaling features.

Box-Cox transformation

Why it fails: Box-Cox is for making data more normal, not for equal weighting.

Increase minimum samples split

Why it fails: Raising minimum samples split constrains further splits, which reduces variance only mildly; the stem asks for the most effective adjustment. Pruning depth or raising minimum samples per leaf limits tree complexity directly, so this option's effect is weaker than the intended answer.

Increase number of leaves

Why it fails: Increasing the number of leaves grows the tree, adding splits that fit training noise and raise variance further. Leaf count controls model complexity, so more leaves moves in the opposite direction; reducing leaves or limiting depth is what constrains variance.

Use a smaller dataset

Why it fails: Shrinking the dataset discards training information without changing the tree's capacity to memorise, so variance need not fall. Hyperparameters govern complexity; resampling addresses bias-variance only indirectly and is not the tuning adjustment the question targets.

Uniqueness

Why it fails: Uniqueness measures duplicate records, not the exhibit's defect. It is tempting because deduplication is a common data quality concern, and uniqueness would be the correct dimension if the same customer or transaction appeared twice in the dataset.

Timeliness

Why it fails: Timeliness measures whether data is current enough for its purpose, not the exhibit's defect. It is tempting because stale records are a frequent data quality problem, and timeliness would be the correct dimension if the data arrived too late to be useful.

Completeness

Why it fails: Completeness concerns missing values or records, not the dimension actually breached in the exhibit. It is tempting because absent data is a common quality issue, but the exhibit's specific defect must be matched to its own dimension, such as accuracy, consistency or timeliness.

Non‑linear decision boundary

Why it fails: Logistic regression assumes a linear decision boundary; if the boundary is non‑linear, overall accuracy would drop, but precision for the minority class could still be higher if the model captures it.

Multicollinearity among predictor variables

Why it fails: Multicollinearity inflates coefficient variance and destabilises estimates, but it does not drive a model to predict the majority class almost exclusively. It tempts because it is a genuine logistic regression diagnostic, and it would be the answer if coefficients showed inflated standard errors or implausible signs rather than the class-imbalance symptom described.

Overfitting due to too many features

Why it fails: Overfitting with 500 churners would typically depress test performance broadly, not produce 98% accuracy with 15% churn precision. It tempts because 500 predictors is a large feature count, and overfitting is the correct diagnosis when training scores far exceed validation scores on a balanced dataset.

Lasso regression (L1)

Why it fails: Lasso applies L1 regularisation to shrink coefficients, not to select polynomial degree; cross-validation is the technique for choosing degree. Lasso suits high-dimensional linear models with irrelevant predictors, a different problem from overfitting control in polynomial fitting.

Principal component analysis (PCA)

Why it fails: PCA reduces dimensionality by projecting onto principal components; it does not evaluate polynomial degree or detect overfitting. PCA suits correlated feature sets before modelling, whereas degree selection requires cross-validation comparing validation error across candidate degrees.

Stepwise selection

Why it fails: Stepwise selection adds or removes predictors by an information criterion; it does not tune polynomial degree, which is set by cross-validation error across candidate degrees. It is tempting because stepwise selection genuinely suits choosing among many candidate input variables in multiple regression.

Line chart

Why it fails: Line charts are best for showing trends over time, not distributions across categories.

Scatter plot

Why it fails: Scatter plots are used to show relationships between two continuous variables, not distributions across regions.

Pie chart

Why it fails: Pie charts show proportions of a whole, but when comparing multiple regions, they become cluttered and hard to interpret.

Ignore missing values and proceed with analysis

Why it fails: Deleting or ignoring incomplete records biases results and shrinks the sample, so it cannot be a general technique here. It is tempting because listwise deletion is valid when missingness is completely at random and affects only a tiny fraction of rows, but the stem asks for handling several columns.

Always replace missing values with zero

Why it fails: Zero-filling distorts distributions and correlations unless zero is the true value, so it is not universally appropriate. It tempts because zero imputation suits count or sparse data where absence genuinely means none, but applying it unconditionally across arbitrary columns fabricates signal.

P-value

Why it fails: P-value is used in hypothesis testing, an inferential statistic.

Regression coefficient

Why it fails: A regression coefficient quantifies the relationship between variables for prediction, placing it in inferential statistics. Descriptive statistics summarise the observed dataset itself, such as mean, median, mode, range or standard deviation. Regression coefficients would be the correct selection when the question asked about modelling or inference instead.

P001 only

Why it fails: P001's average quantity does not exceed the threshold of 8, so HAVING filters it out. It tempts because P001 appears in the sales exhibit and may have individual rows above 8, but HAVING evaluates the grouped AVG, not individual row values — the aggregate for P001 falls at or below 8.

P002 only

Why it fails: P002's grouped average quantity is not greater than 8, so it fails the HAVING predicate. It tempts because P002 is present in the exhibit and some of its rows may exceed 8, but the filter applies to AVG(quantity) per product, and P002's computed average does not clear the threshold.

P003 only

Why it fails: P003's average quantity does not exceed 8, so HAVING excludes it from the result set. It tempts because P003 appears in the sales exhibit with quantities that look high on individual rows, yet the predicate tests the grouped AVG, which for P003 is at or below 8.

The mean decreased significantly

Why it fails: Imputing median may shift the mean slightly, but not significantly, especially if only 5% are missing.

The standard deviation increased

Why it fails: Imputing 50 missing salaries with the median (52000) replaces extreme values with a central value, which compresses the spread and typically reduces, not increases, the standard deviation. It tempts because imputation adds data points, which some assume inflates variability, yet median substitution is specifically chosen to avoid that.

The minimum value decreased

Why it fails: Filling missing entries with the median 52000 cannot lower the minimum unless 52000 is below the existing minimum, which is implausible for salaries. It tempts because imputation changes the column's values, so candidates assume any statistic may shift, but median substitution leaves the minimum untouched.

Allows both read and write access to the bucket

Why it fails: Only s3:GetObject is allowed; no write actions are included.

Allows only specific users to read objects

Why it fails: Restricting reads to named principals narrows access below the stated goal of granting read access across the entire bucket. It is tempting because least-privilege designs often name specific users, and that would be correct if the requirement were limited access, but the stem asks for bucket-wide read permission.

Allows read access to a specific folder within the bucket

Why it fails: A bucket-level policy applies to every object in 'data-lake', not to a single prefix, so scoping read access to one folder contradicts the stated intent. It is tempting because folder-scoped grants are common when teams need segmented access, but that scenario would use a prefix condition rather than a whole-bucket policy.

SARIMA

Why it fails: SARIMA models time series by differencing to achieve stationarity, effectively removing trends and seasonality before applying AR and MA components. This makes it tempting for data exhibiting both characteristics. However, for a clear upward trend, SARIMA does not explicitly model or forecast the trend component as a separate, evolving factor. Instead, it removes the trend through differencing, which can sometimes oversimplify the underlying trend dynamics compared to models that explicitly track and project the trend and seasonal components.

Simple exponential smoothing

Why it fails: Simple exponential smoothing does not handle trend or seasonality.

ARIMA

Why it fails: ARIMA handles trend but not seasonality.

Impute missing values with mean, then use PCA for dimensionality reduction

Why it fails: Mean imputation is sensitive to outliers, PCA may not be needed initially.

Drop all rows with missing values, then apply min-max scaling

Why it fails: Dropping every row containing a missing value discards substantial data across 100 variables, biasing the model and reducing sample size; min-max scaling also fails to tame the extreme outliers. Imputation with robust scaling addresses both issues.

Remove outliers using Z-score, then apply standard scaling

Why it fails: Z-score removal deletes outliers rather than limiting their influence, discarding genuine extreme observations that may carry signal; standard scaling then remains sensitive to any retained extremes. Winsorising or robust scaling preserves records while curbing leverage.

Apply min-max scaling to range [0,1]

Why it fails: Min-max scaling rescales to [0,1] but preserves the relative spread, so a value in thousands still dominates one in hundreds; it does not equalise differing units. Z-score standardisation centres each feature to mean zero and unit variance. Min-max suits bounded inputs like pixel intensities where the range is already meaningful.

Apply log transformation

Why it fails: Log transformation compresses skew and multiplicative spread but does not produce a common scale; values in thousands remain far larger than those in hundreds. Z-score standardisation gives mean zero and unit variance. Log suits right-skewed monetary data when the model assumes multiplicative effects, not unit harmonisation.

Remove the column

Why it fails: Dropping 'Income' discards a predictive feature entirely, leaving the model without that information; standardisation exists precisely to retain such columns. Removal is correct only when a column is irrelevant, leaky or overwhelmingly missing. The scenario requires rescaling, not elimination.