- In a study of heart disease risk factors, researchers find that the association between dietary fat intake and heart disease risk becomes non-significant (p=0.15) when physical activity is added to the model, compared to being significant (p=0.02) in the simple regression. Physical activity is negatively correlated with both fat intake and heart disease risk. What does this scenario best illustrate?
Answer: Physical activity confounds the fat intake-heart disease relationship
When you encounter questions about changing statistical significance after adding variables to a model, you're dealing with fundamental epidemiological concepts about third variables and their relationships.
The key insight here is recognizing what happens when physical activity enters the model. Originally, dietary fat intake showed a significant association with heart disease risk (p=0.02). However, when physical activity was added, this association became non-significant (p=0.15). Since physical activity is negatively correlated with both fat intake and heart disease risk, it's acting as a confounding variable - it's associated with both the exposure (fat intake) and outcome (heart disease), creating a spurious association between them.
Option A is incorrect because mediation would show physical activity lies in the causal pathway between fat intake and heart disease, but the scenario describes physical activity as independently related to both variables. Option B is wrong because an interaction effect would involve testing whether the fat-heart disease relationship varies across levels of physical activity - there's no mention of interaction terms or effect modification here. Option D misses the mark because measurement error wouldn't explain why adding a third variable changes the association's significance.
Study tip: Remember the classic confounder pattern: when adding a variable makes a previously significant association disappear, and that variable is correlated with both exposure and outcome, think confounding. This is why randomized controlled trials are so valuable - they minimize confounding through randomization.
- A biostatistician notices that when both systolic and diastolic blood pressure are included in a model predicting cardiovascular events, the confidence intervals for both coefficients become very wide, but removing either variable results in a significant predictor. The correlation between systolic and diastolic pressure is 0.83. What statistical issue is present?
Answer: Multicollinearity between systolic and diastolic blood pressure
When you encounter regression problems where adding variables makes coefficients unstable or non-significant, but individual variables work fine alone, you're likely dealing with multicollinearity. This occurs when predictor variables are highly correlated with each other, making it difficult for the model to determine each variable's unique contribution.
The key clues here point directly to multicollinearity: the high correlation (0.83) between systolic and diastolic blood pressure, wide confidence intervals when both are included together, and significant results when either is used alone. When predictors are highly correlated, the regression model struggles to partition the variance between them, leading to unstable coefficient estimates and inflated standard errors (hence the wide confidence intervals).
Looking at the wrong answers: (A) Confounding involves unmeasured third variables affecting the relationship, but the issue here is clearly between the measured predictors themselves. (C) Heteroscedasticity refers to non-constant error variance across observations, which would affect all coefficients similarly rather than specifically when variables are combined. (D) Non-linear relationships would typically show poor model fit or residual patterns, not the specific pattern of instability when combining correlated predictors.
Study tip: Remember the multicollinearity warning signs: high correlations between predictors (typically >0.7), coefficients that change dramatically when variables are added/removed, and wide confidence intervals despite adequate sample size. This is especially common in biostatistics with physiologically related measures like different blood pressure readings, multiple lab values, or related anthropometric measurements.
- In a study of factors affecting medication adherence, researchers find that when both income and insurance status are included in the model, the standard error for the income coefficient increases from 0.12 to 0.28, while the coefficient estimate changes minimally. The correlation between income and insurance status is 0.74. What statistical phenomenon is occurring?
Answer: Income and insurance status exhibit multicollinearity affecting precision
When you encounter a regression problem where adding a variable dramatically increases standard errors while coefficients remain stable, you're likely dealing with multicollinearity. This occurs when predictor variables are highly correlated with each other, making it difficult for the model to distinguish their separate effects.
The key diagnostic clues here point directly to multicollinearity: the standard error for income more than doubled (0.12 to 0.28) when insurance status was added, while the coefficient estimate stayed roughly the same. The high correlation between income and insurance status (0.74) confirms these variables share substantial overlap. When predictors are highly correlated, the model struggles to partition variance between them, inflating standard errors and reducing statistical precision without necessarily changing the coefficient estimates much.
Looking at the wrong answers: A) Confounding would involve a third variable affecting both the predictor and outcome, but here we're seeing variance inflation, not bias correction. C) Overfitting typically involves too many variables relative to sample size and would affect model performance metrics, not specifically increase standard errors due to correlation. D) Mediation would require insurance status to be in the causal pathway between income and adherence, and we'd expect to see the income coefficient decrease substantially when the mediator is added.
Remember this pattern: dramatic increases in standard errors when adding correlated predictors signals multicollinearity. Look for high correlations between predictors (generally >0.7) and inflated standard errors as your diagnostic criteria. This is a precision problem, not a bias problem.
- The researchers initially find that higher medication dose is associated with worse glucose outcomes (positive coefficient). However, when patient weight is added to the model, the dose coefficient becomes negative (indicating improvement). Given that heavier patients typically receive higher doses, what does this pattern suggest?
Answer: Patient weight confounds the dose-outcome relationship
When you encounter a situation where adding a third variable to a regression model dramatically changes the coefficient of your main predictor, you're likely dealing with confounding. This is a fundamental concept in observational studies where variables can mask or distort true relationships.
In this scenario, patient weight is a confounder because it's associated with both the exposure (medication dose) and the outcome (glucose change). Heavier patients receive higher doses, but they may also have different baseline characteristics that affect treatment response. The initial positive coefficient for dose was misleading because it didn't account for the fact that higher doses were being given to patients who were inherently harder to treat. Once weight is controlled for in the model, you can see the true effect of dose: higher doses actually improve outcomes when comparing patients of similar weight.
Let's examine why the other options miss the mark. Option A describes mediation, where weight would be on the causal pathway between dose and outcome—but weight determines dose allocation, not the reverse. Option B suggests multicollinearity is problematic, but correlation between predictors only becomes an issue when it prevents meaningful interpretation; here, controlling for the confounder is exactly what's needed. Option D misinterprets the pattern—the medication isn't less effective in heavier patients; rather, heavier patients needed higher doses to achieve the same benefit.
Key takeaway: When a coefficient's sign flips after adding a variable, suspect confounding. The new model likely reveals the true relationship by controlling for an important lurking variable.
- A study examines factors affecting infant birth weight. When socioeconomic status (SES) is added to a model already containing maternal smoking and maternal age, the coefficient for smoking changes from -245g to -180g, while the coefficient for maternal age remains nearly unchanged. What can be concluded?
Answer: SES is a confounder for the smoking-birth weight relationship only
When you encounter questions about coefficient changes after adding variables to a regression model, you're dealing with confounding detection. This is a fundamental concept in epidemiology and biostatistics for identifying variables that distort true relationships.
The key evidence here is that adding SES causes smoking's coefficient to change substantially (from -245g to -180g) while maternal age's coefficient remains nearly unchanged. This pattern indicates that SES is correlated with both smoking and birth weight, creating confounding specifically for the smoking-birth weight relationship. When we control for SES, we get a more accurate estimate of smoking's true effect.
Choice A correctly identifies this scenario. SES confounds only the smoking relationship because only smoking's coefficient changed meaningfully when SES was added.
Choice B is incorrect because multicollinearity would affect both variables if it existed equally. Since only smoking's coefficient changed, SES doesn't have equal relationships with both predictors.
Choice C misinterprets the results. Coefficient changes don't indicate variable importance - they reveal confounding. The magnitude of coefficient change tells us about bias correction, not predictor strength.
Choice D wrongly suggests model misspecification. The fact that adding SES clarifies the smoking effect actually indicates improved model specification, not specification error. Missing variable bias would show different patterns.
Study tip: When coefficients change after adding a variable, think confounding rather than model problems. The variable that shows coefficient change is the one being confounded. Stable coefficients suggest no confounding for those relationships.
- In a regression model predicting cholesterol levels, the correlation between age and years of education is -0.15, between age and income is 0.25, and between education and income is 0.72. If all three variables are included as predictors, which two variables are most likely to exhibit problematic multicollinearity?
Answer: Education and income, because they have the strongest correlation
When analyzing multicollinearity in regression models, you need to identify which predictor variables are highly correlated with each other, as this can create statistical problems like unstable coefficient estimates and inflated standard errors.
The key is examining the correlation coefficients between predictor variables. Generally, correlations above 0.7 (or below -0.7) indicate potentially problematic multicollinearity, while correlations between 0.3-0.7 suggest moderate relationships that may warrant attention.
Looking at the given correlations: age-education (-0.15), age-income (0.25), and education-income (0.72), the education-income correlation of 0.72 is by far the strongest and exceeds the typical threshold for concern. This makes biological sense—higher education levels generally lead to higher-paying careers, creating a strong positive relationship between these variables.
Option A is incorrect because the strength of correlation, not its direction, determines multicollinearity risk. A correlation of -0.15 is quite weak. Option B misses the mark because while the reasoning about age and career advancement has some merit, the actual correlation between age and income (0.25) is only moderate. Option D is wrong because multicollinearity depends on the specific correlations between variable pairs, not their general categorization as demographic factors.
Study tip: When evaluating multicollinearity, focus on the magnitude of correlations, not their sign or conceptual relationships. Look for correlations above 0.7 in absolute value as your primary red flag, and always examine the actual numbers rather than making assumptions based on variable types.
- A biostatistician examines a model predicting cognitive decline that includes age, education years, and a derived variable 'cognitive reserve score' calculated as (education years × 2) + 15. The model fails to converge properly. Which issue is most likely causing this problem?
Answer: Perfect multicollinearity between education and cognitive reserve score
When regression models fail to converge, the most common culprit is a problem with the predictor variables themselves, particularly how they relate to each other mathematically.
Perfect multicollinearity occurs when one predictor variable can be expressed as an exact linear combination of another predictor. In this case, the cognitive reserve score is calculated as (education years × 2) + 15. This means cognitive reserve score = 2 × education + 15, creating a perfect linear relationship between education years and cognitive reserve score. When the regression algorithm tries to estimate separate coefficients for both variables, it encounters an impossible mathematical situation—there are infinite combinations of coefficients that would produce the same model fit. This causes the convergence failure, making option B correct.
Option A is incorrect because having three predictors (age, education, cognitive reserve) typically requires a much smaller sample size than would cause convergence issues—usually you need at least 10-15 observations per predictor. Option C is wrong because non-normal outcomes don't prevent model convergence; they might affect the validity of statistical tests but won't stop the algorithm from finding coefficient estimates. Option D is incorrect because heteroscedasticity (unequal variance in residuals) doesn't prevent convergence either—it affects the reliability of standard errors but doesn't create computational impossibilities.
Remember this pattern: when you see a regression model that won't converge, immediately check whether any predictors are mathematical transformations of others. Perfect multicollinearity is the classic cause of convergence failure in regression analysis.
- In a model predicting surgical complications, a researcher finds that when both surgeon experience (years) and surgeon volume (procedures per year) are included, the confidence interval for experience becomes very wide (1.2 to 8.7) compared to when it's included alone (2.1 to 4.8). The correlation between experience and volume is 0.69. What should the researcher conclude?
Answer: Multicollinearity is reducing precision of the experience coefficient
When you encounter regression problems where coefficient confidence intervals dramatically widen after adding variables, you're likely dealing with multicollinearity. This occurs when predictor variables are highly correlated with each other, making it difficult for the model to separate their individual effects.
Here, the evidence clearly points to multicollinearity: the experience coefficient's confidence interval nearly doubles in width (from 2.7 units wide to 7.5 units wide) when volume is added, and the correlation of 0.69 between predictors is quite high. This happens because experienced surgeons typically perform more procedures, so the model struggles to distinguish whether better outcomes come from experience or volume. The increased uncertainty inflates the confidence interval, but this doesn't mean the variable isn't important—just that its effect is harder to estimate precisely. This confirms answer C is correct.
Answer A is wrong because wider confidence intervals don't indicate lack of significance—they indicate reduced precision. Answer B misses the mark because interaction terms wouldn't solve the fundamental correlation problem between the main effects. Answer D overstates the relationship—while the variables are correlated, a 0.69 correlation doesn't constitute complete confounding, which would require near-perfect correlation.
Study tip: Watch for the classic multicollinearity pattern: high correlations between predictors (typically >0.7) plus widened confidence intervals when variables are included together. This is different from confounding, which involves relationships between predictors and outcomes, not just between predictors themselves.
- A researcher studies factors affecting infant birth weight. In the crude analysis, maternal smoking shows a coefficient of -180g (p=0.03). After adding maternal age, education, and income to the model, the smoking coefficient becomes -220g (p=0.01). What does this pattern most likely indicate?
Answer: The socioeconomic variables are confounders that were masking the true harmful effect of smoking
The smoking coefficient became more negative (stronger harmful effect) and more significant after adjusting for socioeconomic variables, indicating these variables were positive confounders masking smoking's true effect. Smokers may have lower socioeconomic status, which independently affects birth weight. B) is wrong - multicollinearity would cause instability and loss of significance, not stronger effects. C) is incorrect - if these were mediators, adjusting would reduce the smoking effect. D) is backwards - we're examining smoking as the exposure, not as a confounder.
- In a logistic regression model with predictors X1, X2, and X3, the correlation matrix shows all pairwise correlations between 0.15-0.35. However, when X3 is regressed on X1 and X2, the R² is 0.78. The condition number of the design matrix is 12.4. What should the analyst conclude?
Answer: Severe multicollinearity exists despite low pairwise correlations, as X3 is highly predictable from X1 and X2
This demonstrates that pairwise correlations can miss multicollinearity. The R²=0.78 when regressing X3 on X1 and X2 shows that 78% of X3's variance is explained by the other predictors, indicating severe multicollinearity despite low pairwise correlations. A) is wrong - pairwise correlations don't capture all multicollinearity patterns. B) misses the severity indicated by the high R². D) confuses overfitting with multicollinearity - the high R² in auxiliary regression directly indicates multicollinearity.
- In a multiple regression model predicting systolic blood pressure, a researcher finds that the variance inflation factor (VIF) for age is 8.2, for weight is 2.1, and for height is 7.9. Which statement about this model is most accurate?
Answer: Age and height likely have problematic multicollinearity that should be addressed
Variance inflation factors (VIF) help you detect multicollinearity in regression models by measuring how much the variance of a coefficient increases due to correlation with other predictors. Understanding VIF thresholds is crucial for building reliable models.
A VIF of 1 indicates no multicollinearity, while values above 5-10 suggest problematic correlation between predictors. Looking at these results, age (VIF = 8.2) and height (VIF = 7.9) both exceed the common threshold of 5, indicating they're highly correlated with other variables in the model. This makes sense clinically—age and height often correlate with weight and each other, creating redundancy that inflates standard errors and makes coefficient estimates unstable.
Option A is incorrect because while all VIFs are technically below 10, using 10 as a strict cutoff is overly lenient. Most statisticians prefer the more conservative threshold of 5, which age and height exceed. Option C misinterprets VIF—a lower VIF indicates less multicollinearity, not greater importance. Weight's VIF of 2.1 simply means it's less correlated with other predictors, not that it's the strongest predictor of blood pressure. Option D is wrong because you don't automatically reject models based on average VIF exceeding 5. Instead, you address the problematic variables through removal, combination, or other techniques.
When you encounter VIF questions, remember that values above 5 warrant attention, and the solution isn't rejecting the entire model but rather addressing the specific problematic predictors through variable selection or transformation techniques.
- A researcher examining predictors of lung function includes smoking history (pack-years), current smoking status (yes/no), and cigarettes per day in the same model. The model produces a condition number of 45 and several non-significant coefficients despite strong individual correlations. What is the most likely explanation and solution?
Answer: Multicollinearity among smoking variables; consolidate or remove variables
When you encounter regression models with high condition numbers and unexpectedly non-significant coefficients despite strong individual correlations, you're likely dealing with multicollinearity. This occurs when predictor variables are highly correlated with each other, making it difficult for the model to distinguish their individual effects.
The smoking variables in this model—pack-years, current smoking status, and cigarettes per day—are inherently related. Pack-years incorporates both duration and intensity of smoking, while cigarettes per day measures current intensity, and current smoking status indicates whether someone is actively smoking. The condition number of 45 (values >30 suggest serious multicollinearity) confirms this problem. When predictors overlap conceptually like this, the model struggles to assign unique effects to each variable, inflating standard errors and making coefficients non-significant even when the variables individually predict the outcome well.
Option A is incorrect because sample size issues typically produce different symptoms—you'd see wide confidence intervals across all variables, not this specific pattern of individual significance disappearing in the joint model. Option C misses the point; adding more variables would likely worsen multicollinearity rather than solve it. Option D incorrectly assumes the relationships are non-linear when the real issue is redundancy between predictors.
The solution (B) involves consolidating these overlapping smoking measures—perhaps using just pack-years (which captures cumulative exposure) or creating a composite smoking index, rather than including all three correlated measures.
Study tip: High condition numbers + individually significant predictors becoming non-significant together = multicollinearity. Look for conceptually overlapping variables in the model.
- A study examines whether exercise reduces depression scores. Researchers are concerned that baseline depression severity might confound results. In the crude analysis, exercise shows β=-3.2 (SE=1.1). After adjusting for baseline severity, exercise shows β=-2.8 (SE=1.0). Baseline severity has β=0.65 (SE=0.08, p<0.001). How should this be interpreted?
Answer: Baseline severity is a precision variable that improves model fit but doesn't substantially confound the exercise effect
The small change in exercise coefficient (3.2→2.8, ~12% change) with reduced standard error indicates baseline severity acts more as a precision variable than a strong confounder. It improves model precision without substantially changing the exposure effect. A) is wrong - statistical significance doesn't determine confounding status. B) misses that some confounding is present, just not substantial. D) is incorrect - baseline measures typically predate exposure and aren't on the causal pathway.
- The correlation between weight and BSA in their dataset is r=0.94. When both are included in the regression model predicting clearance, weight coefficient = 0.12 (SE=0.18, p=0.51) and BSA coefficient = 15.2 (SE=8.3, p=0.07). What should the researchers conclude?
Answer: Severe multicollinearity exists due to the mathematical relationship between the variables, requiring selection of one predictor
Since BSA is mathematically derived from weight (and height), including both creates severe multicollinearity (r=0.94 confirms this). The large standard errors and loss of significance for both predictors are classic multicollinearity symptoms. One predictor should be chosen based on clinical relevance. A) misses that neither is truly significant due to multicollinearity. B) ignores that they're mathematically related, not independent measures. D) is wrong - the pattern of large SEs with high correlation indicates multicollinearity, not sample size issues.
- In a study of predictors for ICU length of stay, researchers fit a model including APACHE score, number of organ failures, and mechanical ventilation days. The eigenvalues of the correlation matrix are: 2.41, 0.87, 0.52, 0.20. The condition index (√largest eigenvalue/√smallest eigenvalue) is 3.47. Two predictors show variance inflation factors of 1.8 and 2.1. What should be concluded about multicollinearity?
Answer: Moderate multicollinearity exists but is not problematic since condition index and VIFs are within acceptable ranges
When evaluating multicollinearity in regression models, you need to examine multiple diagnostic measures together rather than relying on a single criterion. Multicollinearity occurs when predictor variables are highly correlated, which can inflate standard errors and make coefficient estimates unstable.
The condition index of 3.47 indicates moderate multicollinearity. Values below 10-15 are generally considered acceptable, though some sources use stricter cutoffs around 5. The eigenvalue pattern shows one small eigenvalue (0.20), which suggests some linear dependence among predictors, but it's not approaching zero (which would indicate severe problems). The VIFs of 1.8 and 2.1 are well below the common threshold of 5-10, indicating that multicollinearity isn't severely inflating the variance of coefficient estimates.
Answer A is incorrect because a single small eigenvalue doesn't automatically indicate "severe" multicollinearity requiring immediate correction—you must consider all measures together. Answer B incorrectly concludes no significant multicollinearity exists; while the levels aren't severe, the condition index of 3.47 does suggest moderate multicollinearity is present. Answer C misinterprets the eigenvalue pattern—perfect prediction would result in an eigenvalue of exactly zero, not 0.20.
Answer D correctly synthesizes all the evidence: moderate multicollinearity exists (condition index above ideal levels, one smaller eigenvalue), but it's within acceptable ranges that won't seriously compromise the analysis.
Study tip: Always evaluate multicollinearity using multiple diagnostics together. Memorize the key thresholds: condition index <15, VIF <5-10, and eigenvalues not approaching zero.
- In a Cox regression model, a researcher includes both systolic blood pressure (continuous) and hypertension status (binary: ≥140 mmHg). The model shows: systolic BP hazard ratio=1.02 per mmHg (95% CI: 0.98-1.06, p=0.31), hypertension status HR=2.1 (95% CI: 1.4-3.2, p=0.001). What is the most likely explanation?
Answer: Both variables capture the same underlying construct, creating multicollinearity and unstable estimates
Including both continuous systolic BP and a binary version based on the same variable creates perfect multicollinearity in the threshold region. The hypertension variable is completely determined by systolic BP values, leading to unstable estimates where the continuous effect becomes non-significant. A) misses the multicollinearity issue. C) is wrong - if there were a true threshold effect, we'd expect both to be significant or use splines. D) incorrectly frames this as confounding rather than multicollinearity.
- A researcher studying emergency department wait times includes both 'time of day' (coded 1-24 hours) and 'shift' (day/evening/night) as predictors. Day shift is defined as hours 7-15, evening shift as hours 15-23, and night shift as hours 23-7. The model shows unstable coefficient estimates. What type of statistical issue is present?
Answer: Multicollinearity due to overlapping measurement of time
When you encounter regression models with unstable coefficient estimates, you should immediately think about multicollinearity—when predictor variables are highly correlated with each other, making it difficult for the model to separate their individual effects.
In this case, the researcher has created a classic multicollinearity problem by including both 'time of day' (1-24 hours) and 'shift' categories that are directly derived from those same hours. The shift variable is essentially a categorical transformation of the continuous time variable. When hour = 10, you automatically know it's day shift; when hour = 20, it's evening shift. This perfect mathematical relationship between predictors causes the regression coefficients to become unstable because the model cannot determine which variable is actually driving any observed effects.
Looking at the wrong answers: (A) Temporal autocorrelation refers to correlation between observations over time, not between predictor variables themselves. (C) Confounding involves external variables affecting both predictors and outcomes, but here the issue is the direct relationship between the predictors. (D) Non-linear relationships would show up as poor model fit or residual patterns, not unstable coefficients.
The correct answer is (B) because multicollinearity from overlapping measurement is exactly what's happening—two variables measuring essentially the same underlying construct (time) in different formats.
Study tip: Watch for multicollinearity whenever you see predictors that could be mathematical transformations of each other. Always check correlation matrices before building regression models, and avoid including both continuous variables and categorical versions of the same measure.
- In a study of factors affecting vaccination rates, researchers include both neighborhood income and percentage of residents with college degrees as predictors. These variables have a correlation of 0.81. The coefficient for neighborhood income becomes non-significant when education percentage is added, but education percentage remains significant. What conclusion is most appropriate?
Answer: Multicollinearity makes both coefficients unreliable for interpretation
When you encounter predictors with high correlation (like the 0.81 seen here), you're dealing with multicollinearity. This occurs when independent variables are so closely related that they essentially measure similar underlying constructs. The key warning sign is when coefficients change dramatically or lose significance when correlated variables are included together.
With correlation of 0.81 between neighborhood income and education percentage, these variables share about 66% of their variance (0.81²). This creates unstable coefficient estimates because the regression model struggles to separate their individual effects. When predictors are highly correlated, small changes in the data can cause large swings in coefficients, making individual coefficient interpretation unreliable.
Option C is correct because multicollinearity above 0.8 typically renders individual coefficients untrustworthy for interpretation, even if one remains "significant." The standard errors become inflated and the coefficients become unstable.
Option A incorrectly assumes you can rank variable importance when multicollinearity exists. The apparent "significance" of education doesn't mean it's more important—it could simply be due to which variable the model arbitrarily assigns the shared variance to.
Option B suggests mediation, but this requires specific analytical approaches beyond simple multiple regression. You can't conclude mediation just from observing coefficient changes in the presence of correlation.
Option D is backwards—when multicollinearity exists, you typically need theoretical justification for which variable to remove, not automatic removal based on statistical significance alone.
Study tip: When correlation between predictors exceeds 0.8, focus on the multicollinearity problem first before interpreting any individual coefficients.
- A researcher studying depression scores includes both 'years of education' and 'college degree (yes/no)' in the same regression model. The model shows unstable coefficients and large standard errors for both education variables. What is the most appropriate next step?
Answer: Remove one of the education variables to address multicollinearity
When you encounter regression models with unstable coefficients and large standard errors for related variables, you're likely dealing with multicollinearity - a situation where predictor variables are highly correlated with each other. This is exactly what happens when you include both "years of education" and "college degree (yes/no)" in the same model, since these variables measure overlapping aspects of educational attainment.
Multicollinearity doesn't bias your coefficients, but it makes them unstable and inflates standard errors, making it difficult to determine the individual effect of each variable. The solution is to remove one of the correlated variables, keeping the one that's most theoretically relevant or has better measurement properties.
Option A is too extreme - education may be an important predictor, so removing both variables would eliminate potentially valuable information from your model. Option B misses the point entirely; even though education is measured differently in each variable, they're still highly correlated and create multicollinearity problems. The different measurement approaches don't prevent the statistical issue. Option D suggests logarithmic transformation, but this addresses skewness or non-linear relationships, not multicollinearity between variables.
Option C correctly identifies that removing one education variable will resolve the multicollinearity while preserving the educational information in your model.
Study tip: When you see unstable coefficients and large standard errors for conceptually related variables, immediately think multicollinearity. The fix is usually variable selection, not transformation. Always check correlation matrices between your predictors before building complex models.
- A researcher is analyzing factors associated with blood pressure using multiple linear regression. The correlation matrix shows: age-BMI (r=0.85), age-exercise (r=-0.73), BMI-exercise (r=-0.79). When fitting separate models, age has coefficient 2.1 (SE=0.3, p<0.001), but in the full model with all three predictors, age has coefficient 0.4 (SE=0.8, p=0.62). What is the most likely explanation?
Answer: Multicollinearity among predictors is causing instability in coefficient estimates and inflated standard errors
The high correlations between predictors (0.73-0.85) indicate severe multicollinearity. The dramatic change in age coefficient (2.1→0.4) and inflated standard error (0.3→0.8) with loss of significance are classic signs of multicollinearity causing unstable estimates. A) is incorrect - this describes confounding symptoms but age showing different effects suggests multicollinearity. C) is wrong - sample size issues wouldn't cause such dramatic coefficient changes. D) is incorrect - effect modification would require testing interactions, not explaining coefficient instability.
