Categorical Predictors & Dummy Variables - Business Statistics

1. A regression model predicts monthly sales (in thousands) for a retail chain using $Sales = 45 + 12 \cdot Store\_Type\_A + 8 \cdot Store\_Type\_B + 0.3 \cdot Advertising$, where $Store\_Type\_A$ and $Store\_Type\_B$ are dummy variables for store types A and B respectively, and $Advertising$ is in hundreds of dollars. If a Type C store spends $2,000 on advertising, what is the predicted monthly sales?
   Answer: $51,000 For a Type C store, both dummy variables equal 0 since Type C is the reference category. With $2,000 advertising = 20 hundreds: Sales = 45 + 12(0) + 8(0) + 0.3(20) = 45 + 6 = 51 thousand = $51,000. Choice B incorrectly uses Type A coefficient, Choice C uses Type B coefficient, Choice D uses both coefficients.
2. A retailer models daily sales using $Sales = 2500 + 800 \cdot Weekend + 400 \cdot Holiday + 200 \cdot Weekend \times Holiday$. The $Weekend \times Holiday$ term represents an interaction between weekend and holiday dummy variables. What is the predicted sales increase when comparing a holiday that falls on a weekend versus a regular weekday?
   Answer: $1,400 increase when accounting for the additional interaction effect beyond main effects When you encounter regression models with interaction terms, you need to carefully compare the total predicted values rather than just adding individual coefficients. The interaction term captures how the combined effect of two variables differs from simply adding their separate effects. To find the predicted sales increase, compare two scenarios using the given model: $Sales = 2500 + 800 \cdot Weekend + 400 \cdot Holiday + 200 \cdot Weekend \times Holiday$ For a regular weekday: $Weekend = 0$, $Holiday = 0$, so $Sales = 2500 + 0 + 0 + 0 = 2500$ For a holiday weekend: $Weekend = 1$, $Holiday = 1$, so $Sales = 2500 + 800(1) + 400(1) + 200(1 \times 1) = 2500 + 800 + 400 + 200 = 3900$ The increase is $3900 - 2500 = 1400$. This includes the weekend effect ($800), the holiday effect ($400), plus the interaction effect ($200) that occurs when both conditions are present simultaneously. Answer A ($1,200) incorrectly adds only the main effects ($800 + $400) while ignoring the interaction term entirely. Answer B ($600) makes no logical sense given the coefficient values. Answer C ($1,000) appears to misinterpret how interaction effects work, possibly confusing proportional adjustments with additive effects. The correct answer is D because it accounts for all components: both main effects plus the crucial interaction term. Study tip: Always substitute actual values into regression equations when dealing with dummy variables and interactions—don't just try to mentally combine coefficients, as you'll miss interaction effects.
3. According to this model, what is the predicted price difference between a 2,000 sq. ft. house in an urban area and a 2,000 sq. ft. house in a rural area?
   Answer: A difference of $35,000, with the urban house being more expensive. The predicted price for an urban house is $40 + 0.15(\text{SqFt}) + 25$. The predicted price for a rural house is $40 + 0.15(\text{SqFt}) - 10$. The difference is $(40 + 0.15(\text{SqFt}) + 25) - (40 + 0.15(\text{SqFt}) - 10) = 25 - (-10) = 35$. Since the units are in thousands, the difference is $35,000.
4. Based on the model, what is the estimated impact on Clicks for each additional $100 of advertising spending on a Sunday?
   Answer: An increase of 18.5 clicks. The slope for AdSpend depends on whether it is a weekend. For a weekday (D_Weekend = 0), the slope is 3.5. For a weekend (D_Weekend = 1), the slope is $3.5 + 15 = 18.5$. Therefore, on a Sunday, each additional $100 in AdSpend is associated with an increase of 18.5 clicks.
5. What is the correct interpretation of the coefficient for the Masters dummy variable?
   Answer: Holding experience constant, employees with a Master's degree earn a statistically significant average of $8,000 more than employees with a Bachelor's degree. When interpreting dummy variable coefficients in multiple regression, remember that each dummy coefficient represents the difference between that category and the reference category, holding all other variables constant. In this regression, 'Bachelors' is the reference category, so the Masters coefficient of 8,000 tells you exactly how much more (or less) someone with a Master's degree earns compared to someone with a Bachelor's degree, assuming they have the same years of experience. Since the p-value (0.04) is less than the significance level (0.05), this difference is statistically significant. Therefore, answer D correctly states that employees with Master's degrees earn a statistically significant $8,000 more than those with Bachelor's degrees, holding experience constant. Answer A is wrong because $8,000 isn't the total average salary—it's just the premium above the Bachelor's baseline. Answer B incorrectly compares Masters to PhD degrees, but dummy coefficients don't directly compare non-reference categories to each other. To find that difference, you'd subtract the coefficients: $15,000 - $8,000 = $7,000. Answer C makes this calculation error but then incorrectly claims the difference is non-significant, when actually both education coefficients are individually significant. Study tip: Always identify the reference category first in dummy variable problems. Each coefficient tells you the difference from that baseline, not comparisons between other categories or absolute values. Watch for p-values to determine statistical significance at your chosen alpha level.
6. A researcher is studying crop yield. The model includes a categorical variable for FertilizerType with five levels (A, B, C, D, and a control 'None'). The goal is to determine if each fertilizer type has a significantly different effect on yield compared to using no fertilizer at all. How many dummy variables are needed and what is the appropriate reference category?
   Answer: Four dummy variables are needed, with the 'None' category serving as the reference. The research question is focused on comparing each fertilizer to the control condition ('None'). Therefore, 'None' is the most logical reference category. For a categorical variable with $k=5$ levels, $k-1=4$ dummy variables are required to avoid multicollinearity. The coefficients of these four dummies will represent the difference in yield for fertilizers A, B, C, and D relative to no fertilizer.
7. Holding income constant, what is the model's predicted difference in approved loan amount between an applicant with an 'Excellent' credit rating and one with a 'Fair' credit rating?
   Answer: $20,000 The coefficient for D_Excellent ($25,000) represents the difference between 'Excellent' and the 'Poor' reference. The coefficient for D_Fair ($5,000) represents the difference between 'Fair' and 'Poor'. To find the difference between 'Excellent' and 'Fair', we find the difference between their respective coefficients: $\$25,000 - \$5,000 = \$20,000$.
8. A study examines house prices using $Price = \beta_0 + \beta_1 \cdot Size + \beta_2 \cdot Urban + \beta_3 \cdot Suburban$, where $Urban = 1$ for urban locations and $Suburban = 1$ for suburban locations. After fitting the model, $\beta_2 = 15000$ and $\beta_3 = 8000$. What is the expected price difference between an urban house and a rural house of the same size?
   Answer: $15,000 higher for urban properties since rural is the reference category Rural is the reference category (both dummies = 0). Urban houses have Urban = 1, Suburban = 0, so the price difference from rural is exactly β₂ = $15,000. Choice A confuses this with urban vs suburban difference, Choice B misunderstands the dummy structure, Choice D incorrectly adds both coefficients.
9. A model is built to predict the monthly Cost of a mobile phone plan. The predictors include DataUsed (in GB) and the Carrier (A, B, C, D). Carrier D is set as the reference category. The coefficient for the dummy variable for Carrier A, D_A, is -5.50. How is this coefficient correctly interpreted?
   Answer: Holding data usage constant, the average monthly cost for Carrier A is $5.50 less than the average cost for Carrier D. When analyzing regression models with categorical predictors, understanding dummy variable coefficients is crucial. These coefficients represent differences between categories, not absolute values. In this model, Carrier D serves as the reference category, meaning all other carrier coefficients show how they differ from Carrier D specifically. The coefficient for D_A of -5.50 tells you that when data usage is held constant, Carrier A plans cost $5.50 less on average than Carrier D plans. Answer A correctly captures this interpretation. The phrase "holding data usage constant" acknowledges that we're controlling for the other predictor variable, and the comparison is explicitly between Carrier A and the reference category (Carrier D). Answer B misinterprets the coefficient as an absolute cost rather than a difference. The -5.50 represents a relative change from the reference category, not a standalone price. Answer C confuses the interpretation by suggesting this difference applies "for every GB of data used." This would be correct if we were interpreting an interaction term between carrier and data usage, but D_A is a simple dummy variable coefficient representing a constant difference between carriers. Answer D incorrectly compares Carrier A to the overall average across all carriers. Dummy variable coefficients always compare to the specific reference category, not to the grand mean of all categories. Remember: dummy variable coefficients always represent the difference between that category and the reference category, holding other variables constant. Never interpret them as absolute values or comparisons to overall averages.
10. In a linear regression model that includes a nominal categorical predictor with $k$ levels, which statement most accurately describes the effect of changing the reference category?
    Answer: It changes the coefficients of the dummy variables as well as the model intercept, but not the predicted values. Changing the reference category re-parameterizes the model but does not change its fundamental fit or predictive outcomes. The intercept will change to reflect the mean of the new reference group (at X=0), and the dummy coefficients will change to reflect differences relative to this new baseline. However, the final predicted value ($\hat{y}$) for any given observation will remain exactly the same.
11. A model predicting job performance uses $Performance = 75 + 8 \cdot College + 12 \cdot Graduate + 0.5 \cdot Experience$, where education levels are: High School (reference), College, and Graduate degree. Two employees have identical experience but different education: one College graduate and one Graduate degree holder. If the Graduate degree holder scores 4 points higher than predicted, what is the College graduate's predicted vs. actual performance difference?
    Answer: The College graduate's residual cannot be determined from the given Graduate degree information alone The Graduate degree holder's residual (actual - predicted = +4) provides no information about the College graduate's residual. Residuals are individual observation errors, not systematic across education levels. Choice A and B assume correlated residuals, Choice D wrongly suggests perfect prediction for the College graduate.
12. A researcher models the Weight of a product based on its Volume and Material (Plastic, Steel, Aluminum), with Plastic as the reference category. The resulting equation is \text{Weight} = 0.5 + 1.2(\text{Volume}) + 5.6(\text{D_Steel}) + 1.5(\text{D_Aluminum}). If the researcher decides to re-run the model using Steel as the reference category, what will the new coefficient for the D_Plastic dummy variable be?
    Answer: -5.6 The original model indicates that, on average, Steel products are 5.6 units heavier than Plastic products ($E[W|Steel] - E[W|Plastic] = 5.6$). When Steel becomes the reference category, the new coefficient for D_Plastic will represent the average weight difference of Plastic relative to Steel ($E[W|Plastic] - E[W|Steel]$). This is simply the negative of the original difference: $-5.6$.
13. An analyst is investigating the factors affecting flight delays. The model includes a categorical predictor Airline with 8 different carriers. The analyst groups the 3 smallest carriers into a single category called 'Other' and treats the remaining 5 major carriers as individual categories. To incorporate this new 6-level categorical variable into a regression model, how should the dummy variables be structured?
    Answer: Create 5 dummy variables, one for each major carrier, using the 'Other' category as the reference. The analyst has created a new categorical variable with $k=6$ levels (5 major carriers + 1 'Other' group). To include this in a regression model without falling into the dummy variable trap, $k-1=5$ dummy variables must be used. A logical choice for the reference category is the 'Other' group, so 5 dummies would be created for the 5 major carriers.
14. Which of the following statements is the most precise interpretation of the D_Mobile coefficient?
    Answer: After controlling for time spent on the site, the predicted likelihood to purchase score is 0.08 lower for mobile users compared to desktop users. When you encounter regression coefficients for dummy variables, you're interpreting the difference in the outcome variable between categories, holding all other variables constant. The key phrase here is "after controlling for" the other variables in the model. The coefficient of -0.08 for D_Mobile tells you that when comparing two users who spend the same amount of time on the site, the mobile user's predicted likelihood to purchase score is 0.08 points lower than the desktop user's score. This is exactly what answer A states - it correctly interprets the coefficient as a difference between groups while acknowledging that time on site is controlled for. Answer B misinterprets this as a slope coefficient related to time, but -0.08 is the coefficient for the device type dummy variable, not the interaction between device and time. Answer C incorrectly converts the 0.08 difference in likelihood score to "8% less likely" - these aren't the same thing. A difference of 0.08 in a 0-1 scale score doesn't directly translate to percentage likelihood. Answer D completely misunderstands the coefficient, thinking negative means impossible rather than recognizing it as a comparison showing mobile performs worse than the reference category (desktop). Remember that dummy variable coefficients always represent the difference from the reference category, holding other variables constant. Watch for the phrase "controlling for" or "holding constant" - this distinguishes proper interpretation from common misreadings that ignore the other variables in the model.
15. Given that the coefficient for the D_Month-to-Month dummy is positive (1.85), what does this imply about the churn behavior of customers on month-to-month contracts, holding other factors constant?
    Answer: They have a higher log-odds of churning compared to customers with two-year contracts. In logistic regression, the coefficients represent the change in the log-odds of the outcome for a one-unit change in the predictor. A positive coefficient of 1.85 for the D_Month-to-Month dummy means that customers in this category have higher log-odds of churning compared to the reference category ('Two Year'). While this does imply a higher probability, the coefficient itself is on the log-odds scale, not the probability scale.
16. An analyst creates dummy variables for product categories: Electronics, Clothing, and Books. She codes $D\_Electronics = 1$ for Electronics, $D\_Clothing = 1$ for Clothing, and $D\_Books = 1$ for Books, including all three in the regression model along with an intercept. What problem will she encounter?
    Answer: The regression software will automatically drop one variable to prevent computational singularities Including k dummy variables for k categories plus an intercept creates perfect multicollinearity (the dummy variables sum to a constant vector of 1s). Most software automatically detects and drops one dummy variable to solve this. Choice A wrongly suggests reference categories reduce explanatory power, Choice C and D describe incorrect consequences of the multicollinearity.
17. An analyst studying housing prices creates dummy variables for neighborhood quality: $D\_Good$ and $D\_Excellent$, with Poor as the reference. After running the regression, she finds that $D\_Good = 25000$ and $D\_Excellent = 18000$. Confused by the results, she re-runs the analysis using Excellent as the reference category. What will be the new coefficient for the Good neighborhood dummy variable?
    Answer: $-7,000 since Good neighborhoods are worth $7,000 less than Excellent neighborhoods on average When working with dummy variables in regression, understanding how coefficients change with different reference categories is crucial for proper interpretation. In the original regression with Poor as the reference, $D_{Good} = 25000$ means Good neighborhoods are worth $25,000 more than Poor neighborhoods, while $D_{Excellent} = 18000$ means Excellent neighborhoods are worth $18,000 more than Poor neighborhoods. This implies Good neighborhoods are actually worth $7,000 more than Excellent neighborhoods (25,000 - 18,000 = 7,000). When you change the reference category to Excellent, all coefficients now represent differences relative to Excellent neighborhoods instead of Poor. Since Good neighborhoods are worth $7,000 more than Excellent neighborhoods, but we're now measuring relative to Excellent as the baseline (zero), the coefficient for Good becomes -$7,000. This negative sign indicates that when moving from Excellent (the new reference) to Good, housing prices decrease by $7,000. Answer A incorrectly states the direction—Good neighborhoods are worth more, not less, but misses that coefficients represent deviations from the reference. Answer B is wrong because dummy coefficients always change when you switch reference categories; only the underlying relationships remain constant. Answer C misunderstands how reference changes work—you don't add coefficients together. Remember this key principle: when you change reference categories in dummy variable regression, coefficients represent the same underlying relationships but measured from a different starting point. Always think about what the new baseline represents.
18. An analyst is modeling employee turnover based on a categorical variable for department: Sales, Marketing, Engineering, HR, and Finance. To avoid the dummy variable trap, the analyst uses k-1 dummy variables. They re-run the analysis several times, each time using a different department as the reference category. Which of the following model outputs will remain the same regardless of which department is chosen as the reference category?
    Answer: The overall model R-squared ($R^2$) and F-statistic. Changing the reference category changes the individual coefficients of the dummy variables (as they are measured relative to the baseline) and the model intercept. Consequently, the p-values for these specific coefficients will also change. However, the overall fit of the model to the data, as measured by $R^2$ and the F-statistic for the model's overall significance, does not change. The model is fundamentally explaining the same amount of variance, just parameterizing it differently.
19. In a regression model predicting AnnualBonus, the coefficient for a dummy variable D_TopPerformer (1 if in the top 10% of performance reviews, 0 otherwise) is $5,000. The 95% confidence interval for this coefficient is [-$500, $10,500]. At a significance level of $\alpha = 0.05$, what is the appropriate conclusion?
    Answer: We cannot conclude that there is a statistically significant difference in bonuses between top performers and others. Statistical significance at the $\alpha = 0.05$ level is determined by whether the 95% confidence interval for the coefficient contains zero. Since the interval [-$500, $10,500] includes 0, we cannot reject the null hypothesis that the true coefficient is zero. Therefore, the result is not statistically significant at the 0.05 level.
20. The model predicts a price of $77 for a wine from Region B. What was the age of this bottle of wine?
    Answer: 28 years When you encounter a regression problem with dummy variables, remember that you're working with categorical data (like regions) converted into numerical form for the model. Here, Region A is the reference category, so wines from Region A have both dummy variables equal to zero. To solve this, substitute the known values into the equation. For a wine from Region B with a predicted price of $77, you know that $D_B = 1$ (since it's from Region B) and $D_C = 0$ (since it's not from Region C). Plugging into the equation: $77 = 15 + 2.5(\text{Age}) - 8(1) + 12(0)$ Simplifying: $77 = 15 + 2.5(\text{Age}) - 8$, which becomes $77 = 7 + 2.5(\text{Age})$ Solving for Age: $70 = 2.5(\text{Age})$, so $\text{Age} = 28$ years. Answer choice A (20 years) would give a predicted price of $59, not $77. Choice B (24.8 years) appears to come from incorrectly adding the Region B coefficient instead of subtracting it. Choice C (30.8 years) likely results from forgetting to account for the dummy variable entirely and solving $77 = 15 + 2.5(\text{Age})$. The key strategy here is to carefully track which dummy variables equal 1 or 0 based on the category in question, then substitute all known values systematically. Always double-check your algebra when solving for the unknown variable, as the arithmetic can trip you up under time pressure.

Start this quiz