Interaction Terms - Business Statistics

In a month with an active in-store promotion and an online ad budget of $30,000, what is the estimated increase in sales compared to a month with the same ad budget but no in-store promotion?
Answer: $24,000 The difference in sales between a promotion month and a non-promotion month is found by comparing the predicted sales. Non-promo month: $20 + 0.8(\text{OnlineAds})$ . Promo month: $20 + 0.8(\text{OnlineAds}) + 15 + 0.3(\text{OnlineAds}) = 35 + 1.1(\text{OnlineAds})$ . The difference is $(35 + 1.1(\text{OnlineAds})) - (20 + 0.8(\text{OnlineAds})) = 15 + 0.3(\text{OnlineAds})$ . Plugging in OnlineAds = 30 (since the unit is $1000s), the difference is $15 + 0.3(30) = 15 + 9 = 24$ . Since sales are in $1000s, the difference is $24,000.
In a regression model predicting house prices ( $Y$ ) with square footage ( $X_1$ ), number of bedrooms ( $X_2$ ), and their interaction ( $X_1X_2$ ), the coefficient on the interaction term is -0.8. If we compare two houses that differ by 500 square feet, how does the difference in their predicted prices vary as the number of bedrooms increases from 2 to 4?
Answer: The price difference decreases by $800 when bedrooms increase from 2 to 4 The interaction coefficient tells us how the effect of square footage changes as bedrooms increase. For a 500 sq ft difference, the change in price difference per additional bedroom is 500 × (-0.8) = -$400. When bedrooms increase from 2 to 4 (a change of 2), the total change in price difference is 2 × (-$400) = -$800. Choice B has the wrong sign. Choice C doubles the correct answer. Choice D uses only one bedroom change instead of two.
According to this model, what is the estimated additional annual salary an employee with an advanced degree earns for each additional year of experience?
Answer: $2,900 The effect of an additional year of experience on salary is given by the marginal effect: $\frac{\partial \text{Salary}}{\partial \text{Experience}} = 2.1 + 0.8(\text{AdvancedDegree})$ . For an employee with an advanced degree, AdvancedDegree = 1. So, the effect is $2.1 + 0.8(1) = 2.9$ . Since salary is in thousands of dollars, this corresponds to $2,900.
How should the coefficient $\gamma_1 = 800$ for Age_c be interpreted?
Answer: For a person with 14 years of education, each additional year of age is associated with an $800 increase in income. When variables are centered, the main effect coefficient for one variable is interpreted as its effect when the other centered variable is equal to zero. Education_c is zero when Education is at its mean, which is 14 years. Therefore, $\gamma_1 = 800$ is the effect of a one-year increase in age on income for someone with the mean level of education (14 years).
To model the stock price, what is the most direct way to incorporate this hypothesis into a regression framework where StockPrice is the dependent variable, EPS_News measures the earnings surprise, and UncertaintyIndex is a measure of market uncertainty?
Answer: Include an interaction term between EPS_News and UncertaintyIndex. The hypothesis is that the 'sensitivity' (i.e., the slope or coefficient) of StockPrice with respect to EPS_News changes depending on the level of UncertaintyIndex. This is precisely the definition of an interaction effect. Including an interaction term (EPS_News × UncertaintyIndex) directly models this changing sensitivity within a single equation.
A researcher studying crop yield ( $Y$ ) includes fertilizer amount ( $X_1$ ) and rainfall ( $X_2$ ) with interaction: $\hat{Y} = 100 + 5X_1 + 2X_2 + 0.1X_1X_2$ . The researcher claims that "fertilizer is most effective when rainfall is high." Based on the model, when is this claim supported?
Answer: The claim is always supported because the interaction coefficient is positive When you encounter regression models with interaction terms, focus on how the interaction coefficient affects the relationship between variables. In this crop yield model, the interaction term $0.1X_1X_2$ means that the effect of fertilizer depends on the level of rainfall, and vice versa. To evaluate the researcher's claim, examine how fertilizer's effectiveness changes with rainfall. Taking the partial derivative of yield with respect to fertilizer gives us: $\frac{\partial Y}{\partial X_1} = 5 + 0.1X_2$ . This shows that fertilizer's marginal effect starts at 5 (when rainfall = 0) and increases by 0.1 for each unit of rainfall. Since the interaction coefficient is positive, higher rainfall always makes fertilizer more effective. Answer D is correct because the positive interaction coefficient (0.1) ensures that fertilizer becomes increasingly effective as rainfall increases, supporting the claim universally. Answer A incorrectly suggests the main effect dominates, missing that even small interaction effects create the claimed relationship. Answer B wrongly implies there's a specific rainfall threshold needed, when actually any positive rainfall level makes fertilizer more effective than at zero rainfall. Answer C misinterprets the mechanics—it's not that rainfall "increases the coefficient to 5.1," but rather that the marginal effect of fertilizer becomes $5 + 0.1X_2$ , which grows continuously with rainfall. Study tip: In interaction models, always examine the sign of the interaction coefficient. Positive coefficients mean the variables enhance each other's effects, while negative coefficients mean they diminish each other's effects. Calculate marginal effects by taking partial derivatives to see these relationships clearly.
In building a regression model, what is the primary conceptual reason for including an interaction term between two independent variables, $X_1$ and $X_2$ ?
Answer: To test the hypothesis that the effect of $X_1$ on the dependent variable is conditional on the level of $X_2$ . The fundamental purpose of an interaction term is to model and test for situations where the relationship between one predictor and the outcome is not constant, but instead depends on the value of another predictor. This is also known as moderation or a conditional effect. The other choices describe different concepts: correlation is multicollinearity (A), transformations handle non-linearity (C), and residual properties are checked via diagnostics (D).
For a 20-year-old house, what is the estimated increase in price for an additional 100 square feet of space?
Answer: $16,000 First, find the marginal effect of SqFt on Price, which is $\frac{\partial \text{Price}}{\partial \text{SqFt}} = 0.12 + 0.002(\text{Age})$ . Second, substitute Age = 20 into this equation: $0.12 + 0.002(20) = 0.12 + 0.04 = 0.16$ . This means for a 20-year-old house, each additional square foot is associated with a $0.16 thousand (or $160) increase in price. Third, calculate the effect for 100 additional square feet: $100 \times \$160 = \$16,000$ .
Based on these results, which statement represents the most statistically sound decision?
Answer: Choose Model B because the interaction term provides a statistically significant improvement in model fit. The partial F-test directly compares the explanatory power of the nested models. A p-value of 0.015 is statistically significant at conventional levels (e.g., $\alpha = 0.05$ ). This indicates that adding the interaction term (moving from Model A to Model B) results in a significant improvement in the model's ability to explain the variance in PerformanceScore. Therefore, Model B is statistically preferred despite being less parsimonious.
Which model specification would allow the analyst to test if the entire quadratic relationship between speed and quality differs by machine type?
Answer: $\text{Quality} = \beta_0 + \beta_1 \text{Speed} + \beta_2 \text{Speed}^2 + \beta_3 \text{NewMachine} + \beta_4 (\text{Speed} \times \text{NewMachine}) + \beta_5 (\text{Speed}^2 \times \text{NewMachine}) + \epsilon$ To allow the entire quadratic relationship to differ, the model must allow both the linear and the quadratic terms of Speed to have different coefficients for each machine type. This requires interacting the dummy variable NewMachine with both Speed (the $\beta_4$ term) and Speed $^2$ (the $\beta_5$ term). This model allows for two completely different parabolas for the old and new machines.
The analyst generates a plot of the regression lines for both customer groups (premium and standard support). The plot shows two distinct, non-parallel lines, with the line for premium support customers having a less steep negative slope than the line for standard support customers. What does this visualization suggest?
Answer: An interaction term between Price and PremiumSupport is likely necessary in the model. When the relationship between a predictor variable (Price) and the outcome (Satisfaction) differs across levels of another predictor (PremiumSupport), this indicates an interaction effect. Graphically, this is represented by non-parallel regression lines. The fact that the slopes are different means the effect of price depends on the type of support, which is the definition of an interaction.
At what number of support tickets does an additional month of tenure cease to have a churn-reducing effect (i.e., the effect becomes zero or positive)?
Answer: 2 tickets First, find the marginal effect of Tenure on ChurnRate: $\frac{\partial \text{ChurnRate}}{\partial \text{Tenure}} = -0.001 + 0.0005(\text{Tickets})$ . The churn-reducing effect ceases when this marginal effect is no longer negative (i.e., is $\geq 0$ ). Set the expression to zero to find the turning point: $-0.001 + 0.0005(\text{Tickets}) = 0$ . Solving for Tickets: $0.0005(\text{Tickets}) = 0.001$ , which gives Tickets = 2.
A financial analyst models stock returns ( $Y$ ) using market volatility ( $X_1$ ) and firm size ( $X_2$ , logged): $\hat{Y} = 0.08 - 0.3X_1 + 0.02X_2 - 0.05X_1X_2$ . For a firm with log(size) = 4, if market volatility increases from 0.2 to 0.3, what is the predicted change in stock returns?
Answer: Returns decrease by 0.05 The marginal effect of volatility for a firm with log(size) = 4 is ∂Y/∂X₁ = -0.3 - 0.05(4) = -0.3 - 0.2 = -0.5. When volatility increases by 0.1 (from 0.2 to 0.3), the predicted change is -0.5 × 0.1 = -0.05. Choice A incorrectly uses -0.3 - 0.05 = -0.35 then multiplies by 0.1 and adds extra terms. Choice B uses only the main effect coefficient. Choice C makes an arithmetic error in calculating the interaction effect.
What is the most valid conclusion regarding the effect of training hours (X1) on productivity?
Answer: The effect of training hours on productivity is statistically significant, but only for certain levels of employee engagement. When an interaction term is significant, the main effects of the constituent variables cannot be interpreted in isolation. The effect of X1 is $\beta_1 + \beta_3 X_2$ . Even if $\beta_1$ is not significantly different from zero (p=0.35), the overall effect of X1 can be significant for values of X2 where the combined term $\beta_1 + \beta_3 X_2$ is significantly different from zero. The significant interaction (p=0.02) is the key finding.
A retail analyst estimates the model $\text{Sales} = \beta_0 + \beta_1 \text{Price} + \beta_2 \text{Quality} + \beta_3 \text{Price} \times \text{Quality} + \varepsilon$ . The estimated coefficients are $\hat{\beta_1} = -2.0$ , $\hat{\beta_2} = 1.5$ , and $\hat{\beta_3} = 0.3$ . At what quality level does price have zero effect on sales?
Answer: Quality = 6.67 The marginal effect of price on sales is ∂Sales/∂Price = -2.0 + 0.3(Quality). Setting this equal to zero: -2.0 + 0.3(Quality) = 0, which gives Quality = 2.0/0.3 = 6.67. Choice B incorrectly uses 1.5/0.3. Choice C uses 2.0 without dividing by 0.3. Choice D simply uses the β₂ coefficient value.
According to the model, how does the relationship between production line speed and the number of defects change as the ambient temperature rises?
Answer: As temperature rises, the positive relationship between speed and defects becomes even stronger. The marginal effect of Speed on Defects is given by the derivative with respect to Speed: $0.05 + 0.003(\text{Temp})$ . Since the coefficient of the interaction term (0.003) is positive, as Temp increases, the entire value of this expression increases. This means that the slope of the relationship between Speed and Defects becomes more positive (stronger) at higher temperatures.
In a model predicting customer satisfaction ( $Y$ ) from service quality ( $X_1$ ) and customer age ( $X_2$ ), researchers test whether to include the interaction term $X_1X_2$ . Without the interaction, $R^2 = 0.45$ . With the interaction, $R^2 = 0.47$ . The interaction coefficient has a p-value of 0.08. Given $\alpha = 0.05$ , what should the analyst conclude about including the interaction term?
Answer: Exclude the interaction because the p-value exceeds α, and the R² improvement is modest When evaluating whether to include additional terms in a regression model, you need to balance statistical significance with practical considerations. The key principle is that statistical significance (p-value < α) should be your primary criterion for inclusion, especially when the practical improvement is modest. Here, the interaction term has a p-value of 0.08, which exceeds the significance level of α = 0.05. This means you cannot reject the null hypothesis that the interaction coefficient equals zero - the evidence isn't strong enough to conclude the interaction effect exists in the population. While the $R^2$ did increase from 0.45 to 0.47, this 0.02 improvement represents only a 2% increase in explained variance, which is relatively small and could easily be due to random variation. Option A incorrectly prioritizes practical significance over statistical significance. You cannot claim practical significance when statistical significance is absent - the effect might not even exist. Option B makes the same error, assuming the $R^2$ improvement is meaningful despite the lack of statistical evidence. Option D overstates the case by suggesting both measures indicate lack of significance, when really the $R^2$ change is simply modest, not definitively insignificant. Option C correctly prioritizes the statistical test result while acknowledging that the $R^2$ improvement doesn't compensate for the lack of statistical significance. Study tip: In model selection problems, always check statistical significance first (p-value vs. α), then consider practical significance. Don't let modest improvements in fit statistics override the fundamental requirement for statistical evidence.
Based on this model, what is the correct interpretation of the interaction coefficient, -4?
Answer: The income gap between females and males is $4,000 larger in urban areas than it is in rural areas. The interaction coefficient $\beta_3 = -4$ represents how the effect of one variable changes based on the other. Let's calculate the gender gap in each location. Rural gender gap (IsUrban=0): (Rural Female) - (Rural Male) = $(\beta_0 + \beta_1) - \beta_0 = \beta_1 = -8$ . Urban gender gap (IsUrban=1): (Urban Female) - (Urban Male) = $(\beta_0 + \beta_1 + \beta_2 + \beta_3) - (\beta_0 + \beta_2) = \beta_1 + \beta_3 = -8 - 4 = -12$ . The urban gender gap (-$12k) is $4k larger (more negative) than the rural gender gap (-$8k).
A marketing analyst is modeling sales revenue ( $Y$ ) as a function of advertising spend ( $X_1$ ) and store location type ( $X_2$ , where $X_2 = 1$ for urban stores and $X_2 = 0$ for suburban stores). The fitted regression equation is $\hat{Y} = 50 + 3X_1 + 20X_2 + 1.5X_1X_2$ . If an urban store increases its advertising spend by $10,000, what is the predicted change in sales revenue?
Answer: $45,000 For an urban store, X₂ = 1. The marginal effect of advertising on sales is ∂Y/∂X₁ = 3 + 1.5X₂ = 3 + 1.5(1) = 4.5. Therefore, a $10,000 increase in advertising leads to a predicted change of 4.5 × 10 = $45,000 in sales revenue. Choice A ($30,000) incorrectly uses only the main effect coefficient (3 × 10). Choice C ($35,000) incorrectly adds the interaction coefficient to the base effect (3 × 10 + 5). Choice D ($65,000) incorrectly adds all coefficients except the intercept.
What is the most appropriate action for the analyst to take in response to the high VIF for the interaction term?
Answer: Disregard the high VIF for the interaction term, as it is expected and does not invalidate the model's coefficient estimates for the interaction. It is common and expected for an interaction term to be highly correlated with its constituent main effects, leading to a high VIF. This is often referred to as non-essential multicollinearity. As long as the VIFs for the main effects themselves are not excessively high, this situation is generally not considered problematic for the interpretation of the highest-order term (the interaction). The primary concern of multicollinearity is inflated standard errors, but significance tests on the interaction term are still considered valid. Mean-centering the variables before creating the interaction is a common technique to reduce this type of VIF, but disregarding it is also a valid viewpoint, especially when interpreting the interaction itself.

Start this quiz