I85PREP
⌘ K
Dell Data Science Associate study materials
Dell Data Science Associate study resources

Dell Data Science Associate practice questions

One question at a time. Build your Dell Data Science Associate understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Data Science Fundamentals

In a Dell data science workflow, a column stores customer ages and a histogram of that column is strongly right-skewed. Which measure of central tendency best represents a typical value for this column?

  1. The mean, because it is the most widely reported statistic
  2. The median, because it is resistant to extreme high values Correct answer.
  3. The mode, because it always equals the most frequent age
  4. The standard deviation, because it describes the center

Explanation

A right-skewed distribution has a long tail of large values that pulls the mean upward, so the mean overstates the typical value. The median is a positional measure that stays stable under extreme observations, making it the better summary of center here.

Question 2Data Science Fundamentals

A data scientist at Dell classifies support tickets into categories such as hardware, software, and network. Which measurement scale do these category labels use?

  1. Ordinal
  2. Interval
  3. Ratio
  4. Nominal Correct answer.

Explanation

Nominal scales label categories with no inherent order or distance between them. Ticket categories such as hardware, software, and network differ in kind but cannot be ranked, so they are nominal.

Question 3Data Science Fundamentals

A dataset of laptop battery tests records runtime in hours. Which variable type is runtime?

  1. Categorical nominal
  2. Discrete quantitative
  3. Ordinal categorical
  4. Continuous quantitative Correct answer.

Explanation

Runtime can take any value within a range, including fractional hours, so it is a continuous quantitative variable. Categorical variables describe groups, and discrete variables are countable whole units.

Question 4Data Science Fundamentals

A model predicting whether a server will fail within 24 hours outputs 0 or 1. Which task type is this?

  1. Dimensionality reduction
  2. Clustering
  3. Binary classification Correct answer.
  4. Regression

Explanation

Predicting one of two discrete labels is binary classification. Regression predicts continuous values, clustering groups unlabeled data, and dimensionality reduction compresses features rather than predicting labels.

Question 5Data Science Fundamentals

A team splits labeled data into a training set and a test set. Which statement describes the purpose of the test set?

  1. It supplies the gradients used to update model weights
  2. It stores the hyperparameter values chosen by the team
  3. It replaces the need for cross-validation entirely
  4. It estimates performance on data the model did not learn from Correct answer.

Explanation

The test set is held out so that performance estimates reflect generalization to unseen data. Training uses the training set to fit parameters, while the test set provides an unbiased check.

Question 6Data Science Fundamentals

Which statistic quantifies the strength and direction of a linear relationship between two numeric variables?

  1. Variance
  2. Pearson correlation coefficient Correct answer.
  3. Interquartile range
  4. Median absolute deviation

Explanation

The Pearson correlation coefficient ranges from -1 to 1 and captures both the direction and strength of a linear association. The other options all measure spread of a single variable.

Question 7Data Science Fundamentals

A dataset has 10,000 rows and one feature contains 40 percent missing values. Which action is a reasonable first step before modeling?

  1. Increase the number of rows by duplicating existing records
  2. Investigate the missingness pattern and choose an imputation or exclusion strategy Correct answer.
  3. Delete all 10,000 rows to avoid any bias
  4. Silently fill the gaps with the value 999

Explanation

Missing data handling depends on why values are missing and how much information is lost. Examining the pattern first allows a defensible choice such as median imputation, model-based imputation, or dropping rows or columns.

Question 8Data Science Fundamentals

A confusion matrix for a fraud detector shows 90 true positives, 10 false negatives, 30 false positives, and 870 true negatives. What is the recall?

  1. 0.75
  2. 0.10
  3. 0.90 Correct answer.
  4. 0.03

Explanation

Recall equals true positives divided by the sum of true positives and false negatives, so 90 divided by 100 is 0.90. It measures the share of actual positive cases the model caught.

Question 9Data Science Fundamentals

A retail analytics team wants to group customers by purchasing behavior without any predefined segments. Which approach fits this goal?

  1. A paired t-test on two store means
  2. Logistic regression on churn labels
  3. K-means clustering on behavioral features Correct answer.
  4. Linear regression on purchase amount

Explanation

Clustering is an unsupervised method that discovers natural groupings in feature space without labels. Regression and logistic regression require a target variable, and a t-test compares group means rather than forming segments.

Question 10Data Science Fundamentals

A model has high accuracy on training data but much lower accuracy on held-out data. Which term best describes this gap?

  1. Underfitting
  2. Class imbalance
  3. Overfitting Correct answer.
  4. Data leakage

Explanation

A large gap between training and held-out performance indicates the model memorized training noise rather than general patterns, which is overfitting. Underfitting would show poor performance on both sets.

Question 11Data Science Fundamentals

A feature has a heavy right skew and the team applies a log transformation before training a linear model. What is the main intended effect?

  1. It converts the feature into a categorical variable
  2. It compresses large values and makes the distribution more symmetric Correct answer.
  3. It guarantees the target becomes normally distributed
  4. It removes all outliers from the dataset

Explanation

A log transform reduces the influence of large values by compressing the upper tail, often making a skewed feature closer to symmetric. This can help models that assume roughly linear, well-behaved relationships.

Question 12Data Science Fundamentals

Two numeric features are standardized to mean zero and unit variance before a distance-based algorithm runs. Why does this matter for k-nearest neighbors?

  1. It prevents features with larger scales from dominating the distance Correct answer.
  2. It removes the need to choose a value of k
  3. It guarantees perfect accuracy on the test set
  4. It converts the problem into a supervised regression task

Explanation

Distance-based methods sum squared differences across features, so a feature measured in large units can swamp the others. Standardizing puts features on a comparable scale so each contributes fairly to the distance.

Question 13Data Science Fundamentals

A binary classifier is evaluated with a precision-recall curve on a rare-event dataset. Why is this often preferred over an ROC curve here?

  1. Precision-recall curves always produce higher area values
  2. Precision-recall curves focus attention on performance for the minority positive class Correct answer.
  3. ROC curves cannot be drawn for binary problems
  4. Precision-recall curves ignore the positive class entirely

Explanation

With severe class imbalance, the large number of true negatives can make ROC curves look optimistic. Precision-recall curves emphasize the positive class, so they reveal how well the model finds rare events.

Question 14Data Science Fundamentals

A team wants to estimate how a model will perform on new data while using limited labeled data efficiently. Which method repeatedly splits the training data into folds and averages results?

  1. One-hot encoding
  2. Principal component analysis
  3. k-fold cross-validation Correct answer.
  4. Stratified sampling of the test set

Explanation

k-fold cross-validation rotates which fold serves as validation, training on the rest, and averages the results. This gives a more stable performance estimate than a single split.

Question 15Data Science Fundamentals

In a linear regression output, a predictor has a coefficient of 2.5 with a p-value of 0.40 at the usual 0.05 threshold. What is the most defensible interpretation?

  1. The model has perfect fit for this predictor
  2. The coefficient proves the predictor causes the outcome
  3. The predictor has a strong and statistically significant effect
  4. There is insufficient evidence that the predictor's effect differs from zero Correct answer.

Explanation

A p-value of 0.40 is well above 0.05, so the data do not provide sufficient evidence to reject the null hypothesis of no effect. The estimated coefficient exists but is not statistically distinguishable from zero at that threshold.

Question 16Data Science Fundamentals

A categorical feature has 12 distinct values with no natural order. Which encoding is most appropriate for a linear model?

  1. Dropping the feature because it is categorical
  2. One-hot encoding into 12 binary columns Correct answer.
  3. Replacing each value with its frequency rank
  4. Label encoding with integers 1 through 12

Explanation

One-hot encoding represents each unordered category as its own binary indicator, avoiding any implied numeric ordering. Arbitrary integer labels would impose a false ordinal relationship on the model.

Question 17Data Science Fundamentals

A dataset records monthly sales for 36 months. The team wants to forecast the next quarter. Which characteristic makes this a time series problem rather than a standard regression problem?

  1. The data were collected by a large organization
  2. The target variable is numeric rather than categorical
  3. The observations are ordered in time and are likely autocorrelated Correct answer.
  4. The dataset contains more than one feature

Explanation

Time series data have a temporal order, and observations close in time tend to be correlated. Standard random train-test splits violate this structure, so forecasting requires time-aware validation and models that account for autocorrelation.

Question 18Data Science Fundamentals

A recruiter asks why a model rejected a loan application. Which concept does this request concern?

  1. Hyperparameter tuning
  2. Batch size selection
  3. Model interpretability Correct answer.
  4. Data normalization

Explanation

Explaining why a model produced a specific decision is the domain of interpretability, often using feature attributions or surrogate models. The other options concern preprocessing or training configuration.

Question 19Data Science Fundamentals

A team trains a decision tree that perfectly fits the training data but generalizes poorly. Which change is most likely to reduce this overfitting?

  1. Adding more raw features without selection
  2. Limiting tree depth and requiring a minimum number of samples per leaf Correct answer.
  3. Removing the validation set
  4. Increasing the maximum tree depth

Explanation

Constraining depth and requiring larger leaves are pruning-style controls that limit how finely the tree can carve the training data. Simpler trees usually generalize better when the fully grown tree is memorizing noise.

Question 20Data Science Fundamentals

A marketing team observes that ice cream sales and drowning incidents rise together across the year. Which explanation is most consistent with sound data science reasoning?

  1. Drowning incidents cause ice cream sales
  2. Both are driven by a third variable, warm weather Correct answer.
  3. The correlation proves a direct causal mechanism
  4. Ice cream consumption causes drowning

Explanation

Correlation does not establish causation; a confounding variable such as seasonal warm weather can drive both series. Without controlling for confounders, a causal claim is unsupported.

Question 21Data Science Fundamentals

A dataset of sensor readings has a mean of 50 and a standard deviation of 8. Using the empirical rule for an approximately normal distribution, roughly what percentage of readings falls between 34 and 66?

  1. About 95 percent Correct answer.
  2. About 50 percent
  3. About 99.7 percent
  4. About 68 percent

Explanation

The interval from 34 to 66 spans two standard deviations on each side of the mean, since 50 minus 16 equals 34 and 50 plus 16 equals 66. For an approximately normal distribution, about 95 percent of values fall within two standard deviations.

Question 22Data Science Fundamentals

A data scientist compares two models with 5-fold cross-validation and reports mean accuracy of 0.81 and 0.83 with fold-to-fold standard deviations near 0.04. What is the most reasonable conclusion?

  1. The second model is clearly superior and should be deployed immediately
  2. The difference is small relative to variability, so more evidence is needed before choosing Correct answer.
  3. Cross-validation results cannot be compared between models
  4. Both models are unusable because accuracy is below 0.90

Explanation

A two-point mean difference is small compared with fold-to-fold variation of roughly four points, so the observed gap may be noise. Additional folds, repeated cross-validation, or a paired statistical comparison would strengthen the decision.

Question 23Data Science Fundamentals

A team builds a churn model where only 2 percent of customers churn. A model that predicts no churn for everyone achieves 98 percent accuracy. Why is this metric misleading?

  1. Accuracy is undefined for binary targets
  2. The dataset is too small to compute accuracy
  3. Accuracy rewards the majority class and hides failure to detect the minority class Correct answer.
  4. Precision is always equal to accuracy in this setting

Explanation

When one class dominates, a trivial majority-class predictor scores high accuracy while catching zero churners. Metrics such as recall, precision, F1, or area under the precision-recall curve better reflect performance on the rare class.

Question 24Data Science Fundamentals

A model is trained on data where a feature was recorded only for customers who had already been approved for a credit card. Predictions on new applicants are unreliable. Which problem does this illustrate?

  1. Selection bias in the training sample Correct answer.
  2. Multicollinearity among predictors
  3. Insufficient learning rate
  4. Random measurement noise

Explanation

The training data represent only previously approved customers, a non-random subset of the population the model will score. This selection bias means learned relationships may not transfer to the full applicant pool.

Question 25Data Science Fundamentals

A team wants to reduce 200 correlated numeric features to a smaller set of uncorrelated components before modeling. Which method directly serves this goal?

  1. Median imputation
  2. Chi-square test of independence
  3. Simple random sampling
  4. Principal component analysis Correct answer.

Explanation

Principal component analysis constructs orthogonal linear combinations of the original features ordered by explained variance. It produces uncorrelated components and can reduce dimensionality while retaining most variability.

Question 26Data Science Fundamentals

A data science team stores customer transaction records in a table where each row is one purchase and one column is the customer identifier. Which data structure best describes this arrangement?

  1. A graph structure in which customers are nodes and purchases are edges
  2. A hierarchical JSON document with nested purchase arrays
  3. A tidy tabular dataset in which each row is an observation and each column is a variable Correct answer.
  4. An unstructured text corpus requiring tokenization

Explanation

Tidy tabular data follows the convention that every row represents a single observation and every column represents a single measured variable, which is exactly the described layout. This rectangular form is the default input for most analytics and modeling libraries. Graph, nested-document, and text-corpus representations describe other storage patterns with different semantics.

Question 27Data Science Fundamentals

During exploratory analysis a data scientist notices that a numeric column contains values of vastly different magnitudes across records. Which descriptive statistic pair best summarizes the center and spread while remaining sensitive to those extreme values?

  1. Mode and frequency count
  2. Minimum and maximum only
  3. Median and interquartile range Correct answer.
  4. Mean and standard deviation

Explanation

The median and interquartile range are based on rank position, so a few extreme values do not distort them. Mean and standard deviation both use every raw value and are pulled strongly by outliers. This makes the median and IQR the preferred summary when magnitude varies widely.

Question 28Data Science Fundamentals

A data engineer must guarantee that each row in a customer table has a unique, non-null value in the customer identifier column. Which constraint enforces this requirement?

  1. A default constraint supplying a fallback value
  2. A primary key constraint on the identifier column Correct answer.
  3. A foreign key constraint on the identifier column
  4. A check constraint limiting values to a numeric range

Explanation

A primary key enforces both uniqueness and non-nullness on the designated column, which is precisely the stated requirement. Foreign keys enforce referential links to another table, check constraints validate value conditions, and defaults fill missing entries without guaranteeing uniqueness.

Question 29Data Science Fundamentals

A team wants to compare the distribution of order values across three regions on a single chart. Which visualization most directly supports this comparison?

  1. Side-by-side box plots, one per region Correct answer.
  2. A single pie chart of total revenue by region
  3. A heat map of correlation among numeric columns
  4. A scatter plot of order value against order index

Explanation

Box plots display median, quartiles, and outliers, so placing them side by side lets viewers compare distribution shape and spread across groups. Pie charts show part-to-whole totals rather than distributions, and scatter or heat maps answer different questions about relationships.

Question 30Data Science Fundamentals

A data scientist wants to report the typical value of household income in a city where a small number of very wealthy households exist. Which measure of central tendency best represents the typical household?

  1. The median household income Correct answer.
  2. The mean household income
  3. The maximum household income
  4. The variance of household income

Explanation

Income distributions are right-skewed, and the mean is inflated by a handful of very large values. The median, being the middle rank, better reflects what a typical household earns. Maximum and variance describe extremes or spread rather than central tendency.

Question 31Data Science Fundamentals

Before fitting a model, a data scientist splits records into training and test sets. Which property must the test set satisfy to give an unbiased estimate of generalization performance?

  1. It must be larger than the training set
  2. It must contain the same rows as the training set in a different order
  3. It must include only records with the most common label
  4. It must contain records not used in any way during training or tuning Correct answer.

Explanation

A test set estimates generalization only if it is untouched during training and model selection, so its records must never influence the fitted model or hyperparameters. Size relative to training and label composition are secondary considerations. Reusing training rows would leak information and inflate the estimate.

Question 32Data Science Fundamentals

A team computes a 95 percent confidence interval for average session duration from a sample. Which interpretation of that interval is correct?

  1. The sample average is wrong 5 percent of the time
  2. There is a 95 percent probability that the true average lies inside this specific interval
  3. 95 percent of individual sessions fall inside the interval
  4. If the study were repeated many times, about 95 percent of intervals built this way would contain the true average Correct answer.

Explanation

Frequentist confidence intervals have a coverage interpretation: across repeated sampling, the stated percentage of intervals constructed by the procedure capture the true parameter. Once computed, a specific interval either contains the parameter or does not. The interval also concerns the parameter, not the spread of individual observations.

Question 33Data Science Fundamentals

A data scientist computes a Pearson correlation of about 0.02 between two features but observes a clear U-shaped pattern in a scatter plot. What conclusion is most appropriate?

  1. The features are independent because the correlation is near zero
  2. The correlation should be recomputed after removing the middle of the range
  3. The scatter plot must be mislabeled because correlation captures all dependence
  4. A strong nonlinear relationship may exist even though the linear correlation is near zero Correct answer.

Explanation

Pearson correlation measures only the strength of a linear association, so a symmetric U-shape can average out to roughly zero. A near-zero coefficient therefore does not establish independence. Visual inspection is needed to detect such nonlinear dependence.

Question 34Data Science Fundamentals

A dataset contains a numeric column with a small number of missing entries. A data scientist wants to preserve the overall distribution and avoid discarding rows. Which imputation strategy best fits this goal?

  1. Replace missing values with zero regardless of scale
  2. Drop every row containing a missing value
  3. Encode missing values as a new category label
  4. Fill missing values with the column median Correct answer.

Explanation

Median imputation retains all rows and is robust to skew, so it preserves the central tendency of the column without large distortion. Dropping rows loses information, and filling with zero injects an artificial value that can shift the distribution. Category encoding applies to categorical, not numeric, features.

Question 35Data Science Fundamentals

A team wants to detect whether the mean delivery time differs across four warehouses. Which statistical approach is designed for comparing means of more than two groups simultaneously?

  1. A one-way analysis of variance (ANOVA) Correct answer.
  2. A chi-square test of independence
  3. A single-sample z-test
  4. A paired two-sample t-test

Explanation

One-way ANOVA tests whether group means are equal across two or more groups by comparing between-group and within-group variance. Paired t-tests handle two related samples, chi-square tests address categorical association, and single-sample tests compare one mean to a reference value.

Question 36Data Science Fundamentals

A data scientist observes that a numeric feature has values with a wide range and a long right tail. Which transformation is commonly applied to reduce the influence of extreme values before modeling?

  1. A one-hot encoding
  2. A principal component rotation
  3. A train-test split
  4. A logarithmic transformation Correct answer.

Explanation

A logarithmic transform compresses large values and pulls in a long right tail, making the distribution more symmetric. One-hot encoding applies to categorical variables, splitting partitions rows, and PCA is a dimensionality reduction step rather than a value transform.

Question 37Data Science Fundamentals

A dataset records customers who were sent a promotional email, and the target is whether each customer purchased. Which framing of this problem is correct for building a predictive model?

  1. A time-series forecasting task predicting future demand
  2. A clustering task grouping customers by similarity
  3. A binary classification task predicting purchase versus no purchase Correct answer.
  4. A regression task predicting the exact purchase amount

Explanation

The target has two discrete outcomes, purchase or no purchase, which makes this a binary classification problem. Regression would require a continuous numeric target, clustering is unsupervised, and forecasting predicts future values over time.

Question 38Data Science Fundamentals

A data scientist wants to summarize how spread out a set of commute times is while reporting the answer in the same units as the data. Which measure is appropriate?

  1. The standard deviation of the commute times Correct answer.
  2. The skewness of the commute times
  3. The kurtosis of the commute times
  4. The variance of the commute times

Explanation

Standard deviation is the square root of variance and is therefore expressed in the original units, making it directly interpretable as spread. Variance is in squared units, and skewness and kurtosis are dimensionless shape measures.

Question 39Data Science Fundamentals

A data scientist joins two tables on a customer identifier, but the output contains more rows than the left table. Which cause best explains this result?

  1. The right table contains duplicate values of the join key Correct answer.
  2. The join used an inner join instead of a left join
  3. The left table contains a primary key on the identifier
  4. The join key was stored as an integer in both tables

Explanation

In a join, each row on the left matches every row on the right with the same key, so duplicate keys on the right multiply the output rows. A primary key on the left guarantees uniqueness there, and join type or matching key types do not create duplicates.

Question 40Data Science Fundamentals

A team is designing a dashboard for executives who need to monitor weekly revenue trends. Which design choice best supports quick comprehension?

  1. A scatter plot of revenue against transaction identifier
  2. A line chart with a clear time axis and a single highlighted metric Correct answer.
  3. A 3D pie chart with many small slices
  4. A dense table listing every transaction individually

Explanation

A line chart maps time to the horizontal axis and emphasizes trend, which suits monitoring revenue over weeks. Detailed transaction tables and 3D pie charts overload viewers, and plotting revenue against an identifier lacks meaningful interpretation.

Question 41Data Science Fundamentals

A data scientist notices that a model's performance drops sharply when the same model is applied to data from a different region collected a year later. Which concept does this illustrate?

  1. Distribution shift between training and deployment data Correct answer.
  2. Multicollinearity among input features
  3. Data leakage from the test set into training
  4. Overfitting caused by too many training epochs

Explanation

When the data-generating process changes across region or time, the deployment distribution differs from the training distribution, producing distribution shift. Overfitting, leakage, and multicollinearity describe other failure modes that do not require a change in the underlying population.

Question 42Data Science Fundamentals

A data scientist wants to estimate the proportion of defective units in a large shipment by sampling. Which sampling method gives every unit an equal and independent chance of selection?

  1. Convenience sampling of the first units on the loading dock
  2. Stratified sampling that overrepresents one production line
  3. Quota sampling that fills fixed counts per category
  4. Simple random sampling without replacement from the shipment Correct answer.

Explanation

Simple random sampling gives each unit an equal probability of selection and avoids systematic bias, supporting valid inference to the shipment. Convenience, quota, and deliberately unbalanced stratified samples introduce selection bias that distorts the estimate.

Question 43Data Science Fundamentals

A team has a small labeled dataset and a large unlabeled dataset. They want to improve a classifier by letting it learn from the unlabeled records as well. Which paradigm matches this goal?

  1. Reinforcement learning
  2. Supervised regression
  3. Unsupervised clustering
  4. Semi-supervised learning Correct answer.

Explanation

Semi-supervised learning combines a small labeled set with a large unlabeled set to improve model learning. Reinforcement learning learns from reward signals through interaction, clustering ignores labels entirely, and regression predicts a continuous target.

Question 44Data Science Fundamentals

A data scientist must store a dataset where records have inconsistent fields, some nested arrays, and varying schemas. Which storage format is most flexible for this requirement?

  1. A single-column CSV of concatenated values
  2. A fixed-schema relational table with strict columns
  3. A comma-separated file with a fixed header row
  4. A document-oriented NoSQL store such as JSON documents Correct answer.

Explanation

Document stores accommodate varying and nested schemas without forcing every record into identical columns. Fixed relational tables and CSV files require a uniform structure, and concatenating values into one column destroys field structure.

Question 45Data Science Fundamentals

A data scientist wants to reduce the number of input variables by keeping only those that carry the most information about the target. Which technique directly performs this selection?

  1. Normalizing each feature to unit variance
  2. Applying k-means clustering to the feature matrix
  3. Filtering features by their mutual information with the target Correct answer.
  4. Splitting the data into training and validation sets

Explanation

Mutual information quantifies how much a feature tells us about the target, so ranking and filtering features by it removes uninformative variables. Clustering, normalization, and data splitting are preprocessing or evaluation steps that do not select features by target relevance.

Question 46Data Science Fundamentals

A data scientist wants to assess whether the proportion of customers choosing a premium plan differs between two independent regions. Which test is appropriate?

  1. A two-proportion z-test Correct answer.
  2. A one-sample t-test of the premium rate
  3. An ANOVA comparing premium prices
  4. A paired t-test on the two regions

Explanation

Comparing proportions of a binary outcome across two independent groups calls for a two-proportion z-test. Paired t-tests require dependent numeric samples, a one-sample test compares one proportion to a reference, and ANOVA compares numeric means across groups.

Question 47Data Science Fundamentals

A data scientist computes the variance of a sample and divides by n minus one rather than by n. What is the purpose of this adjustment?

  1. It guarantees the sample is normally distributed
  2. It corrects the estimator so that it is unbiased for the population variance Correct answer.
  3. It removes the influence of outliers from the estimate
  4. It makes the variance equal to the standard deviation

Explanation

Dividing by n minus one, known as Bessel's correction, compensates for the sample mean being estimated from the same data, yielding an unbiased variance estimator. It does not change units, affect normality, or neutralize outliers.

Question 48Data Science Fundamentals

A model predicts a continuous target, and the residuals show a clear funnel shape, widening as the predicted value increases. Which assumption is most clearly violated?

  1. Constant variance of the errors, or homoscedasticity Correct answer.
  2. Normality of the feature distributions
  3. Independence of the target variable from the features
  4. Linearity of the model's output scale

Explanation

A funnel pattern in residuals indicates that error variance changes with the predicted value, violating the homoscedasticity assumption. Residual diagnostics target the errors, not the marginal distribution of features, and the pattern concerns variance rather than the model's output scale or target independence.

Question 49Data Science Fundamentals

A data scientist trains a model on historical loan decisions and finds it reproduces past approval patterns, including disparities across demographic groups. Which concern does this raise?

  1. The model cannot be evaluated without a confusion matrix
  2. The model's accuracy is guaranteed to improve over time
  3. The model will necessarily be underfit because it copies the labels
  4. The model may inherit and amplify historical bias present in the training labels Correct answer.

Explanation

When labels encode past human decisions, a model trained on them learns those patterns, including any historical disparities, and can perpetuate them. This is a fairness and bias concern rather than a fitting, evaluation, or improvement guarantee issue.

Question 50Data Science Fundamentals

A team reports that a feature has a high correlation with the target, but the correlation is computed only on records where a later outcome was observed. Which problem does this raise?

  1. Multicollinearity, because the target correlates with the feature
  2. Overfitting, because the model saw too many features
  3. Selection bias, because the analyzed records are not representative of the full population Correct answer.
  4. Heteroscedasticity, because the residuals vary in spread

Explanation

Restricting analysis to records with an observed outcome conditions on a post-treatment variable, so the sample may differ systematically from the full population. This is selection bias, distinct from overfitting, multicollinearity, or residual variance issues.