WGU C207 DATA-DRIVEN DECISION-MAKING
FINAL EXAM () REVISED QUESTIONS
AND HIGHLIGHTED ANSWERS, 100% PASS.. .
1. A logistics company is analyzing delivery delays. The data includes delivery time (continuous),
distance (continuous), traffic index (categorical: low/medium/high), and driver experience (ordinal:
<1 year, 1-3 years, >3 years). Which regression model would be most appropriate to predict
delivery time while accounting for the non-linear effects of distance and the ordinal nature of
driver experience?
A. Multiple linear regression with distance squared term and driver experience as a continuous variable
B. Generalized additive model with a smooth term for distance and driver experience as an ordered factor
C. Poisson regression with distance as offset and driver experience as a factor
D. Ridge regression with all predictors standardized and interaction terms
Answer: B
Rationale: Generalized additive models (GAMs) allow non-linear smooth terms for continuous predictors
like distance, and can handle ordinal predictors appropriately via ordered factors. Linear regression
with a squared term (A) may not capture complex non-linearities. Poisson regression (C) is for count
data, not continuous delivery time. Ridge regression (D) is for regularization, not for modeling
non-linear effects or ordinal nature.
2. A retailer uses a decision tree to classify customers as high-value or low-value. The tree has a
depth of 5 and uses Gini impurity. After training, the tree achieves 95% accuracy on training data
but only 70% on validation data. Which of the following interventions is most likely to improve
validation accuracy while maintaining interpretability?
A. Increase the tree depth to 10 to capture more complex patterns
B. Prune the tree by setting a minimum number of samples per leaf to 50
C. Switch to a random forest with 100 trees and no pruning
D. Use a support vector machine with a radial basis function kernel
Answer: B
Rationale: The large gap between training and validation accuracy indicates overfitting. Pruning (B)
reduces overfitting by simplifying the tree, improving generalization while keeping interpretability.
Increasing depth (A) worsens overfitting. Random forest (C) may improve accuracy but reduces
interpretability. SVM (D) also sacrifices interpretability and may not address the overfitting issue
directly.
3. In a hypothesis test for the mean delivery time of two courier services, a p-value of 0.03 is
obtained at ±=0.05. Which of the following interpretations is correct?
A. There is a 3% probability that the null hypothesis is true
Page 1
,B. If the null hypothesis were true, the probability of observing a test statistic as extreme as the one observed is
3%
C. The probability of making a Type I error is 3%
D. The observed effect size is large because the p-value is small
Answer: B
Rationale: The p-value is the probability of observing the data (or more extreme) given that the null
hypothesis is true. It is not the probability that the null is true (A). The Type I error rate is set by ±, not
the p-value (C). A small p-value does not indicate effect size (D); it only indicates statistical
significance.
4. A data scientist is building a logistic regression model to predict customer churn. The dataset
has 10,000 observations with 500 churners. Which of the following evaluation metrics would be
most appropriate if the cost of missing a churner is 10 times higher than the cost of falsely
predicting churn?
A. Accuracy
B. Area under the ROC curve (AUC)
C. F1-score
D. Weighted accuracy with a cost matrix
Answer: D
Rationale: When misclassification costs are unequal, using a cost matrix to weight errors is most
appropriate, as it directly incorporates the cost ratio. Accuracy (A) assumes equal costs. AUC (B)
measures ranking ability but not absolute performance under cost. F1-score (C) balances precision and
recall but does not incorporate specific costs.
5. A pharmaceutical company conducts a randomized controlled trial for a new drug. The primary
outcome is binary (success/failure). The trial is designed with 80% power to detect a 10% absolute
risk reduction at ±=0.05. The observed risk reduction is 8% with a p-value of 0.07. Which of the
following statements is correct?
A. The study failed to reject the null hypothesis, but the result is clinically meaningful and warrants further
investigation
B. The study proves that the drug has no effect because the p-value is above 0.05
C. The observed effect is smaller than the minimum clinically important difference, so the trial is inconclusive
D. The study had insufficient power to detect the observed effect, so the results are invalid
Answer: A
Rationale: A p-value > 0.05 means failure to reject the null, but the observed effect (8% reduction) is
close to the clinically meaningful 10% reduction. The result is not statistically significant but may still
be clinically important, warranting further study. Option B is incorrect because non-significance does
not prove no effect. Option C is incorrect because the observed effect is clinically meaningful. Option D
is incorrect because power is about detecting a specified effect, not about validity.
6. A company uses a linear programming model to optimize its product mix. The model has two
decision variables (x1, x2) and three constraints. The feasible region is bounded. Which of the
following statements about the optimal solution is always true?
Page 2
,A. The optimal solution lies at an extreme point (corner point) of the feasible region
B. The optimal solution is unique
C. The optimal solution will have at least one variable equal to zero
D. The optimal solution can be found by evaluating all integer points within the feasible region
Answer: A
Rationale: In linear programming with a bounded feasible region, the optimal solution (if it exists) will
occur at an extreme point (corner point) due to the linear objective function. Uniqueness (B) is not
guaranteed; there could be multiple optimal solutions along an edge. Option C is not always true; both
variables can be positive. Option D is incorrect because the feasible region is continuous, not integer.
7. A marketing team uses A/B testing to compare two webpage designs. They run a test for one
week, collecting 10,000 visitors per variant. The conversion rate for variant A is 5.2% and for
variant B is 5.8%. Which of the following additional statistics is most important to report
alongside the p-value?
A. The standard deviation of conversion rates within each variant
B. The confidence interval for the difference in conversion rates
C. The power of the test
D. The correlation between visitor characteristics and conversion
Answer: B
Rationale: A confidence interval for the difference provides an estimate of the effect size and its precision,
which is essential for practical decision-making beyond just statistical significance. Standard deviation
(A) is already used in the test. Power (C) is pre-determined, not a result. Correlation (D) is not directly
relevant to the A/B test outcome.
8. A data analyst is using k-means clustering to segment customers. After running the algorithm
with k=4, the silhouette score is 0.2. Which of the following actions is most likely to improve the
clustering quality?
A. Increase k to 10 to create more granular segments
B. Use hierarchical clustering instead of k-means
C. Standardize the variables before clustering
D. Remove outliers from the dataset
Answer: C
Rationale: A low silhouette score often indicates that clusters are not well-separated. Standardizing
variables ensures that all variables contribute equally to distance calculations, which can improve
clustering. Increasing k (A) may artificially increase scores but not necessarily quality. Hierarchical
clustering (B) may not address the underlying issue. Removing outliers (D) could help but is less
fundamental than standardization.
9. A time series of monthly sales data shows a clear seasonal pattern with a period of 12 months.
The analyst fits an ARIMA(1,1,1)(1,1,1)[12] model. Which component of this model accounts for
the seasonal differencing?
A. The first (1,1,1) part
B. The second (1,1,1) part with [12]
Page 3
, C. The order of differencing d=1
D. The seasonal moving average term
Answer: B
Rationale: In the notation ARIMA(p,d,q)(P,D,Q)[s], the seasonal differencing is represented by D, which
is the second number in the seasonal part. Here, D=1 in (1,1,1)[12]. Option A refers to the non-seasonal
part. Option C is non-seasonal differencing. Option D is the seasonal MA term, not differencing.
10. A company wants to predict employee turnover using a dataset with 20 features, including
salary, years at company, job satisfaction score, and commute distance. The target variable is
binary (left/stayed). The dataset has 5,000 records with 500 leavers. Which of the following
methods is most appropriate for handling class imbalance?
A. Use accuracy as the evaluation metric and train a logistic regression model
B. Use SMOTE to generate synthetic samples of the minority class
C. Remove 4,000 records of the majority class to balance the dataset
D. Use a decision tree with default parameters
Answer: B
Rationale: SMOTE (Synthetic Minority Over-sampling Technique) creates synthetic examples of the
minority class, which can help the model learn the minority class without losing information. Accuracy
(A) is misleading for imbalanced data. Removing majority records (C) discards valuable data. Decision
trees (D) may still be biased toward the majority class without addressing imbalance.
11. A retail chain with 500 stores is analyzing customer purchase data to optimize inventory. The
dataset includes transaction records from the past three years, with variables such as store
location, product category, price, discount rate, and customer zip code. The chain wants to predict
which products will be in high demand next month using a model that accounts for seasonal
patterns and regional preferences. Which modeling approach would be most appropriate,
considering the need for interpretability and the ability to handle complex interactions?
A. Fit a multiple linear regression with all main effects and two-way interactions, then use stepwise selection to
reduce overfitting.
B. Use a random forest model with 500 trees, and then apply SHAP values to interpret feature importance.
C. Implement a deep neural network with three hidden layers, using dropout regularization to prevent
overfitting.
D. Apply a support vector machine with a radial basis function kernel, tuning parameters via cross-validation.
Answer: B
Rationale: Random forests handle complex interactions and nonlinearities well, and SHAP values
provide a rigorous, interpretable measure of feature importance. Linear regression with stepwise
selection is prone to overfitting and cannot capture complex patterns easily. Deep neural networks lack
interpretability. SVMs with RBF kernels are less interpretable and may not scale well to large datasets.
Page 4
FINAL EXAM () REVISED QUESTIONS
AND HIGHLIGHTED ANSWERS, 100% PASS.. .
1. A logistics company is analyzing delivery delays. The data includes delivery time (continuous),
distance (continuous), traffic index (categorical: low/medium/high), and driver experience (ordinal:
<1 year, 1-3 years, >3 years). Which regression model would be most appropriate to predict
delivery time while accounting for the non-linear effects of distance and the ordinal nature of
driver experience?
A. Multiple linear regression with distance squared term and driver experience as a continuous variable
B. Generalized additive model with a smooth term for distance and driver experience as an ordered factor
C. Poisson regression with distance as offset and driver experience as a factor
D. Ridge regression with all predictors standardized and interaction terms
Answer: B
Rationale: Generalized additive models (GAMs) allow non-linear smooth terms for continuous predictors
like distance, and can handle ordinal predictors appropriately via ordered factors. Linear regression
with a squared term (A) may not capture complex non-linearities. Poisson regression (C) is for count
data, not continuous delivery time. Ridge regression (D) is for regularization, not for modeling
non-linear effects or ordinal nature.
2. A retailer uses a decision tree to classify customers as high-value or low-value. The tree has a
depth of 5 and uses Gini impurity. After training, the tree achieves 95% accuracy on training data
but only 70% on validation data. Which of the following interventions is most likely to improve
validation accuracy while maintaining interpretability?
A. Increase the tree depth to 10 to capture more complex patterns
B. Prune the tree by setting a minimum number of samples per leaf to 50
C. Switch to a random forest with 100 trees and no pruning
D. Use a support vector machine with a radial basis function kernel
Answer: B
Rationale: The large gap between training and validation accuracy indicates overfitting. Pruning (B)
reduces overfitting by simplifying the tree, improving generalization while keeping interpretability.
Increasing depth (A) worsens overfitting. Random forest (C) may improve accuracy but reduces
interpretability. SVM (D) also sacrifices interpretability and may not address the overfitting issue
directly.
3. In a hypothesis test for the mean delivery time of two courier services, a p-value of 0.03 is
obtained at ±=0.05. Which of the following interpretations is correct?
A. There is a 3% probability that the null hypothesis is true
Page 1
,B. If the null hypothesis were true, the probability of observing a test statistic as extreme as the one observed is
3%
C. The probability of making a Type I error is 3%
D. The observed effect size is large because the p-value is small
Answer: B
Rationale: The p-value is the probability of observing the data (or more extreme) given that the null
hypothesis is true. It is not the probability that the null is true (A). The Type I error rate is set by ±, not
the p-value (C). A small p-value does not indicate effect size (D); it only indicates statistical
significance.
4. A data scientist is building a logistic regression model to predict customer churn. The dataset
has 10,000 observations with 500 churners. Which of the following evaluation metrics would be
most appropriate if the cost of missing a churner is 10 times higher than the cost of falsely
predicting churn?
A. Accuracy
B. Area under the ROC curve (AUC)
C. F1-score
D. Weighted accuracy with a cost matrix
Answer: D
Rationale: When misclassification costs are unequal, using a cost matrix to weight errors is most
appropriate, as it directly incorporates the cost ratio. Accuracy (A) assumes equal costs. AUC (B)
measures ranking ability but not absolute performance under cost. F1-score (C) balances precision and
recall but does not incorporate specific costs.
5. A pharmaceutical company conducts a randomized controlled trial for a new drug. The primary
outcome is binary (success/failure). The trial is designed with 80% power to detect a 10% absolute
risk reduction at ±=0.05. The observed risk reduction is 8% with a p-value of 0.07. Which of the
following statements is correct?
A. The study failed to reject the null hypothesis, but the result is clinically meaningful and warrants further
investigation
B. The study proves that the drug has no effect because the p-value is above 0.05
C. The observed effect is smaller than the minimum clinically important difference, so the trial is inconclusive
D. The study had insufficient power to detect the observed effect, so the results are invalid
Answer: A
Rationale: A p-value > 0.05 means failure to reject the null, but the observed effect (8% reduction) is
close to the clinically meaningful 10% reduction. The result is not statistically significant but may still
be clinically important, warranting further study. Option B is incorrect because non-significance does
not prove no effect. Option C is incorrect because the observed effect is clinically meaningful. Option D
is incorrect because power is about detecting a specified effect, not about validity.
6. A company uses a linear programming model to optimize its product mix. The model has two
decision variables (x1, x2) and three constraints. The feasible region is bounded. Which of the
following statements about the optimal solution is always true?
Page 2
,A. The optimal solution lies at an extreme point (corner point) of the feasible region
B. The optimal solution is unique
C. The optimal solution will have at least one variable equal to zero
D. The optimal solution can be found by evaluating all integer points within the feasible region
Answer: A
Rationale: In linear programming with a bounded feasible region, the optimal solution (if it exists) will
occur at an extreme point (corner point) due to the linear objective function. Uniqueness (B) is not
guaranteed; there could be multiple optimal solutions along an edge. Option C is not always true; both
variables can be positive. Option D is incorrect because the feasible region is continuous, not integer.
7. A marketing team uses A/B testing to compare two webpage designs. They run a test for one
week, collecting 10,000 visitors per variant. The conversion rate for variant A is 5.2% and for
variant B is 5.8%. Which of the following additional statistics is most important to report
alongside the p-value?
A. The standard deviation of conversion rates within each variant
B. The confidence interval for the difference in conversion rates
C. The power of the test
D. The correlation between visitor characteristics and conversion
Answer: B
Rationale: A confidence interval for the difference provides an estimate of the effect size and its precision,
which is essential for practical decision-making beyond just statistical significance. Standard deviation
(A) is already used in the test. Power (C) is pre-determined, not a result. Correlation (D) is not directly
relevant to the A/B test outcome.
8. A data analyst is using k-means clustering to segment customers. After running the algorithm
with k=4, the silhouette score is 0.2. Which of the following actions is most likely to improve the
clustering quality?
A. Increase k to 10 to create more granular segments
B. Use hierarchical clustering instead of k-means
C. Standardize the variables before clustering
D. Remove outliers from the dataset
Answer: C
Rationale: A low silhouette score often indicates that clusters are not well-separated. Standardizing
variables ensures that all variables contribute equally to distance calculations, which can improve
clustering. Increasing k (A) may artificially increase scores but not necessarily quality. Hierarchical
clustering (B) may not address the underlying issue. Removing outliers (D) could help but is less
fundamental than standardization.
9. A time series of monthly sales data shows a clear seasonal pattern with a period of 12 months.
The analyst fits an ARIMA(1,1,1)(1,1,1)[12] model. Which component of this model accounts for
the seasonal differencing?
A. The first (1,1,1) part
B. The second (1,1,1) part with [12]
Page 3
, C. The order of differencing d=1
D. The seasonal moving average term
Answer: B
Rationale: In the notation ARIMA(p,d,q)(P,D,Q)[s], the seasonal differencing is represented by D, which
is the second number in the seasonal part. Here, D=1 in (1,1,1)[12]. Option A refers to the non-seasonal
part. Option C is non-seasonal differencing. Option D is the seasonal MA term, not differencing.
10. A company wants to predict employee turnover using a dataset with 20 features, including
salary, years at company, job satisfaction score, and commute distance. The target variable is
binary (left/stayed). The dataset has 5,000 records with 500 leavers. Which of the following
methods is most appropriate for handling class imbalance?
A. Use accuracy as the evaluation metric and train a logistic regression model
B. Use SMOTE to generate synthetic samples of the minority class
C. Remove 4,000 records of the majority class to balance the dataset
D. Use a decision tree with default parameters
Answer: B
Rationale: SMOTE (Synthetic Minority Over-sampling Technique) creates synthetic examples of the
minority class, which can help the model learn the minority class without losing information. Accuracy
(A) is misleading for imbalanced data. Removing majority records (C) discards valuable data. Decision
trees (D) may still be biased toward the majority class without addressing imbalance.
11. A retail chain with 500 stores is analyzing customer purchase data to optimize inventory. The
dataset includes transaction records from the past three years, with variables such as store
location, product category, price, discount rate, and customer zip code. The chain wants to predict
which products will be in high demand next month using a model that accounts for seasonal
patterns and regional preferences. Which modeling approach would be most appropriate,
considering the need for interpretability and the ability to handle complex interactions?
A. Fit a multiple linear regression with all main effects and two-way interactions, then use stepwise selection to
reduce overfitting.
B. Use a random forest model with 500 trees, and then apply SHAP values to interpret feature importance.
C. Implement a deep neural network with three hidden layers, using dropout regularization to prevent
overfitting.
D. Apply a support vector machine with a radial basis function kernel, tuning parameters via cross-validation.
Answer: B
Rationale: Random forests handle complex interactions and nonlinearities well, and SHAP values
provide a rigorous, interpretable measure of feature importance. Linear regression with stepwise
selection is prone to overfitting and cannot capture complex patterns easily. Deep neural networks lack
interpretability. SVMs with RBF kernels are less interpretable and may not scale well to large datasets.
Page 4
