BAR · Business Analysis · 6 practice questions
Adjusted R-Squared When Adding Regression Predictors
Adding a predictor can raise R-squared while lowering adjusted R-squared because the adjustment penalizes extra predictors. Below: a worked comparison showing why a higher R-squared does not automatically mean a better model.
Try one first
Hint
OLS assumptions: (1) Linearity (Y is linear function of X), (2) Homoscedasticity (constant error variance), (3) Independence (uncorrelated errors), (4) Normality (errors normally distributed). Test each via diagnostic plots and tests; address violations before using model.
Answer A. ORDINARY LEAST SQUARES (OLS) LINEAR REGRESSION rests on FOUR KEY ASSUMPTIONS, sometimes called 'LINE' or 'GAUSS-MARKOV' assumptions. Violation of any can compromise the regression's validity for inference and prediction. (1) LINEARITY: the relationship between Y and X is LINEAR over the observation range. Y = α + βX + ε approximately holds. Diagnostic: scatter plot of X vs Y; residual plot (residuals should show no systematic pattern). Remedies for non-linearity: transform variables (log, sqrt, square), use non-linear regression, or restrict the analysis to a linear subrange. (2) HOMOSCEDASTICITY: errors have CONSTANT VARIANCE across all values of X. The 'spread' of residuals around the regression line is uniform. Diagnostic: residual plot; Breusch-Pagan test. Remedies for heteroscedasticity: weighted least squares, robust standard errors, or transformation of Y. (3) INDEPENDENCE: errors are NOT CORRELATED across observations. Each residual is independent of others. Critical violation: AUTOCORRELATION in time-series data (residual at t correlated with residual at t-1). Diagnostic: Durbin-Watson statistic. Remedies: include time-series controls (lagged variables, AR/MA terms), use time-series methods. (4) NORMALITY: errors are NORMALLY DISTRIBUTED. Most important for hypothesis testing and confidence intervals on coefficients. Diagnostic: histogram of residuals; Q-Q plot; Shapiro-Wilk test. Remedies: transform Y; use bootstrap methods; rely on Central Limit Theorem with large samples. ADDITIONAL CONCERNS (often listed alongside): (a) NO PERFECT MULTICOLLINEARITY, variables not perfectly correlated; (b) NO ENDOGENEITY, error term uncorrelated with X; (c) CORRECT MODEL SPECIFICATION, relevant variables included; (d) MEASUREMENT ACCURACY, variables measured without significant error. INTERNAL AUDIT FOCUS: testing these assumptions is part of model validation. Material violations should be addressed before relying on the model for forecasting or pricing decisions. Answer B is incorrect: OLS doesn't require Y or X to be financial vs physical, or 'deterministic'; the WHOLE POINT of regression is modeling stochastic relationships. Answer C is incorrect: there's no 'exactly 30' sample size requirement (general guideline: 10+ per variable). Precision uniformity is irrelevant. Answer D is incorrect: OLS does NOT require X normality (X is treated as fixed/observed). The NORMALITY assumption is for residuals/errors, not raw variables.
Step by step
- Set up the comparison
A manufacturer uses 25 monthly observations to predict overhead from labor hours and machine hours, producing R² of 0.700. Adding setups as a third predictor raises R² to 0.710 using the same observations.
- Identify the formula inputs
Adjusted R² = 1 - (1 - R²)(n - 1)/(n - k - 1), where n is the observation count and k is the predictor count. Do not count the intercept as a predictor.
- Update the predictor count
Adding setups changes k from 2 to 3 while n stays 25, so the new adjusted R² is 1 - (0.290 × 24/21) = 0.6686, or approximately 0.669. Incorrectly keeping k at 2 gives 1 - (0.290 × 24/22) = 0.6836, or approximately 0.684.
- Calculate the original adjusted fit
For the original model, adjusted R² = 1 - (0.300 × 24/22) = 0.6727, or approximately 0.673. Compare this with the expanded model's adjusted R² of approximately 0.669.
- Judge the added predictor
Ordinary R² rises, but adjusted R² falls because the added predictor's improvement does not offset its penalty. On adjusted R² alone, the expanded model is not an improvement.
Key points
- Remember: a higher R² can hide a lower adjusted R² after adding a predictor.
- Count predictors, not the intercept, when determining k.
- Adding a predictor does not change n when both models use the same observations.
How the exam traps you
- Reporting 0.710 as the new adjusted R². That is ordinary R². Apply the adjustment to get approximately 0.669.
- Keeping k at 2 and getting approximately 0.684. Adding setups makes k equal 3. Update the denominator to 21.
Question 2
Hint
For one predictor in a multiple regression, focus on one-unit changes in that predictor while the other predictors stay fixed.
Answer C. In a multiple regression, a slope coefficient shows the change in the predicted dependent variable from a one-unit increase in that predictor, with the other predictors unchanged. The price coefficient is negative 12, so a $1 increase in price is associated with 12 fewer expected units sold, holding advertising and number of stores constant.
Why not A: This answer incorrectly converts the coefficient into a percentage effect. Because the dependent variable is in units sold, the coefficient represents a 12-unit change, not a 12% change.
Why not B: This answer treats the regression as guaranteeing actual results and ignores the role of the other predictors. The coefficient applies to expected units sold and is interpreted with advertising and stores held constant, not regardless of them.
Why not D: This answer reverses the direction of the relationship. The model predicts sales from price, so the coefficient tells how expected sales change when price changes, not how price changes when sales change.
Question 3
Hint
If two regressions look the same otherwise, consider which coefficient estimate is less likely to move around from sample to sample.
Answer A. When the model form and key output are otherwise the same, a larger sample generally makes the coefficient estimate more stable and less affected by random variation. That means the estimate from 120 observations is the most reliable for forecasting among the four choices.
Why not B: This answer reflects the mistaken view that any moderate sample is enough when the outputs look identical. Although better than 12 observations, 24 still provides less support for a stable estimate than the larger samples.
Why not C: This answer recognizes that more observations help, but it stops short of the strongest choice. With all else held constant, 120 observations provide a more reliable coefficient estimate than 60.
Why not D: This answer comes from ignoring sample size once the reported slope and R-squared match. With only 12 observations, the coefficient estimate is more vulnerable to sampling error and is less reliable.
Question 4
Hint
Multiple regression: Y = a + b₁X₁ + b₂X₂ + ... Use when cost is driven by multiple distinct factors. Cautions: multicollinearity (correlated predictors), overfitting (too many variables), data requirements (typically 10+ observations per variable).
Answer D. MULTIPLE REGRESSION extends simple regression by allowing more independent variables (Y = a + b₁X₁ + b₂X₂ + ... + bₖXₖ). It is preferred over simple regression when COST IS GENUINELY DRIVEN BY MULTIPLE FACTORS, and a single variable cannot capture the full cost-driver structure. WHEN MULTIPLE REGRESSION HELPS: (1) Different activities drive different portions of cost (e.g., labor hours drive labor costs; machine hours drive depreciation/utilities; setups drive setup-related supplies). (2) Higher R² (more variation explained), more accurate predictions. (3) Identifies which drivers matter most via t-statistics on each coefficient. (4) Reveals whether a presumed driver actually drives cost (statistically significant slope?). CAUTIONS WITH MULTIPLE REGRESSION: (1) MULTICOLLINEARITY, if independent variables are correlated with each other (e.g., labor hours ≈ machine hours), individual coefficients become unstable; remedy by combining or dropping correlated variables. (2) OVERFITTING, adding too many variables fits the noise; check ADJUSTED R² (penalizes # of variables) rather than raw R². (3) DATA REQUIREMENTS, typically need 10+ observations per independent variable for stable estimates. (4) INTERPRETATION, slope coefficients represent the effect of each X holding others constant; this requires the variables to vary independently in the data. WHEN SIMPLE REGRESSION IS FINE: (a) one obvious dominant cost driver; (b) limited data; (c) preliminary analysis; (d) explanatory simplicity is valuable. Answer A is incorrect: more variables = more computation, not less. Multiple regression is more complex than simple regression. Answer B is incorrect: industry doesn't determine the choice. Multiple regression applies broadly. Answer C is incorrect: 'always preferred' is overconfident. Adding irrelevant variables can hurt model quality (overfitting, multicollinearity). The right number of variables matches the underlying cost structure.
Question 5
Hint
The most influential added point usually combines an unusual X value with a large vertical departure from the current pattern.
Answer B. The observation (27, 20) is far outside the existing X range and also far below the established pattern. A point with both high leverage and a large residual has the greatest ability to pull the least-squares line. That combination makes this observation the most influential on the fitted line.
Why not A: This choice focuses only on the extreme X value. Even though it has high leverage, it sits on the existing trend, so it extends the pattern rather than materially changing the fitted line.
Why not C: This choice notices the point is off the pattern, but its X value is still within the observed range. Without extreme leverage, it is less influential than a similarly unusual point far to the right.
Why not D: This choice overweights the residual and ignores leverage. Although it is below the pattern, its X value is inside the existing range, so it has less ability to move the overall line.
Question 6
Hint
When the model looks strong overall but related predictors look weak one by one, think about shared explanatory power.
Answer A. This pattern is classic for multicollinearity. The predictors together explain much of the variation in sales, but because they overlap heavily, each coefficient's standard error is inflated. That can leave the overall model significant while individual t-statistics are not.
Why not B: This mistake treats a monthly data set as automatically pointing to time-series error correlation. Serial correlation affects inference, but it is not the usual explanation for a strong overall model paired with individually insignificant correlated predictors.
Why not C: This choice assumes the linear form is the main problem. Nonlinearity can weaken fit or bias coefficients, but it does not create the specific pattern where the model works well overall while overlapping predictors each appear weak alone.
Why not D: This choice comes from linking insignificant t-statistics to any standard error problem. Heteroskedasticity can distort standard errors, but it does not specifically explain a high R-squared with a significant overall F-test and weak individual predictors.
Common questions
Can adjusted R-squared decrease when R-squared increases?
Yes. Here, R² rises from 0.700 to 0.710, but adjusted R² falls from approximately 0.673 to 0.669 because the improvement does not offset the added-predictor penalty.
Does k include the intercept in adjusted R-squared?
No. In the formula used here, k counts predictors only. Labor hours, machine hours, and setups make k equal 3.
Why did I get 0.684 instead of 0.669?
You used the new R² but kept the old predictor count. With 25 observations and 3 predictors, n - k - 1 equals 21, not 22.
Practice BAR like the real exam
The free ChatCPA simulator: real exam layout, timed testlets, starting with a question on this topic. No account needed to start.
Open the free simulator →More on Data analytics, regression analysis, and quantitative methods
- Calculate Predicted Cost Using Simple Linear Regression
- Correlation vs. Causation: Confounding and Selection Bias
- Time-Series Forecasting: Match the Method to the Pattern
- Sensitivity Analysis vs Scenario Analysis: One Input or Many
All Data analytics, regression analysis, and quantitative methods practice →
Questions from the ChatCPA bank of 17,658 CPA exam questions, each with a written reason for every wrong answer. ChatCPA is built by Nicholas Miller, CPA (Oregon #14907). How these pages are made. Spot an error? Tell us.