Biostatistics Quiz: Cox Proportional Hazards Model
20 questions · exam conditions
0:00
Cox Proportional Hazards ModelQuestion 1 of 20

In a Cox proportional hazards model, what is the primary advantage of the semi-parametric approach compared to fully parametric survival models?

It provides more precise estimates of regression coefficients by using maximum likelihood estimation
It eliminates the need to specify the baseline hazard function while still estimating covariate effects
It automatically adjusts for violations of the proportional hazards assumption through robust standard errors
It can handle censored observations more effectively than parametric models like Weibull regression
← Back to quizzes

Biostatistics Quiz

Biostatistics Quiz: Cox Proportional Hazards Model

Practice Cox Proportional Hazards Model in Biostatistics with focused quiz questions that help you check what you know, review explanations, and build confidence with test-style prompts.

What this quiz covers

This quiz focuses on Cox Proportional Hazards Model, giving you a quick way to practice the rules, question types, and explanations that matter most for Biostatistics.

How to use this quiz

Try each quiz question before looking at the correct answer. Use the explanations to review missed ideas, then come back to similar questions until the pattern feels familiar.

All questions

Question 1

In a Cox proportional hazards model, what is the primary advantage of the semi-parametric approach compared to fully parametric survival models?

  1. It provides more precise estimates of regression coefficients by using maximum likelihood estimation
  2. It eliminates the need to specify the baseline hazard function while still estimating covariate effects (correct answer)
  3. It automatically adjusts for violations of the proportional hazards assumption through robust standard errors
  4. It can handle censored observations more effectively than parametric models like Weibull regression
Explanation: The key advantage of Cox regression is that it estimates covariate effects without requiring specification of the baseline hazard function's parametric form, making it semi-parametric. Choice A is incorrect because parametric models also use MLE. Choice C is wrong because Cox models don't automatically adjust for proportional hazards violations. Choice D is incorrect because parametric models handle censoring equally well.

Question 2

In a Cox model with covariates age and treatment, the partial likelihood function focuses on the relative rankings of failure times rather than their absolute values. What is the primary methodological advantage of this approach?

  1. It reduces computational complexity by eliminating the need for iterative algorithms
  2. It ensures that all regression assumptions are automatically satisfied regardless of data characteristics
  3. It allows estimation of covariate effects without specifying the form of the baseline hazard function (correct answer)
  4. It guarantees that confidence intervals will have exact coverage probabilities in finite samples
  5. It eliminates bias from informative censoring when patients drop out non-randomly
Explanation: When you encounter Cox proportional hazards models, remember that they're designed to handle survival data where the exact timing of events matters less than their relative order. The Cox model's revolutionary insight was using partial likelihood, which compares who fails when failures occur, rather than modeling the absolute timing. The correct answer is C because the partial likelihood's key innovation is separating covariate effects from the baseline hazard. By focusing on relative risk rankings at each failure time, you can estimate how covariates like age and treatment affect hazard ratios without needing to specify whether the baseline hazard follows an exponential, Weibull, or any other distribution. This semi-parametric approach makes the Cox model incredibly flexible—you get the covariate effects you care about without making potentially wrong assumptions about the underlying hazard shape. Option A is incorrect because Cox models still require iterative algorithms (typically Newton-Raphson) to maximize the partial likelihood. Option B is wrong because the proportional hazards assumption can still be violated, and other model assumptions must be checked. The partial likelihood doesn't automatically satisfy all regression requirements. Option D is false because, like most maximum likelihood methods, Cox models provide asymptotic confidence intervals, not exact finite-sample coverage guarantees. For biostatistics exams, remember this key distinction: Cox models are "semi-parametric"—parametric for covariate effects (you estimate coefficients) but non-parametric for the baseline hazard (no distributional assumption needed). This flexibility is what made Cox regression the gold standard in survival analysis.

Question 3

In a Cox proportional hazards model analyzing time to disease recurrence, the estimated coefficient for treatment group is β=−0.693\beta = -0.693 with a standard error of 0.2. The baseline group has a median survival time of 24 months. What can be concluded about the treatment effect?

  1. The treatment group has half the hazard rate of the control group, with median survival of 48 months
  2. The treatment group has half the hazard rate of the control group, but median survival cannot be determined from this information (correct answer)
  3. The treatment group has twice the hazard rate of the control group, with median survival of 12 months
  4. The treatment reduces the hazard by 69.3%, with median survival of approximately 31 months
  5. The treatment effect is not statistically significant, so no conclusions about hazard rates can be drawn
Explanation: When analyzing Cox proportional hazards models, you need to understand how coefficients relate to hazard ratios and why survival times can't be directly calculated from hazard information alone. The coefficient β=−0.693\beta = -0.693 tells us about the treatment effect on the hazard rate. The hazard ratio is calculated as eβ=e−0.693=0.5e^{\beta} = e^{-0.693} = 0.5. Since this compares the treatment group to the baseline (control) group, the treatment group has half the hazard rate of the control group - meaning they have a 50% lower risk of disease recurrence at any given time. However, here's the crucial point: you cannot determine the median survival time for the treatment group from this information alone. While we know the baseline median is 24 months and the treatment reduces hazard by 50%, the relationship between hazard ratios and median survival times is not straightforward and depends on the underlying survival distribution. Option A incorrectly assumes you can simply double the median survival time when the hazard is halved. Option C gets the hazard ratio backwards - a negative coefficient means reduced hazard, not increased. Option D makes up a specific median survival calculation that isn't mathematically justified from the given information. The key study tip: In Cox regression, coefficients tell you about relative hazards (eβe^{\beta}), but translating hazard ratios into specific survival times requires additional information about the baseline hazard function. Don't fall for questions that seem to offer precise survival calculations when only hazard ratios can be determined.

Question 4

A Cox model includes a continuous covariate X with coefficient β=0.2\beta = 0.2. A researcher wants to interpret this result for clinical relevance. If the interquartile range of X is 10 units, what is the most appropriate interpretation?

  1. Each 1-unit increase in X increases the hazard by 20%, so the average patient has 20% higher hazard than baseline
  2. Each 1-unit increase in X increases the hazard by 22%, so patients at the 75th percentile have 2.2-fold higher hazard than those at the 25th percentile
  3. Each 1-unit increase in X increases the hazard by 22%, so patients at the 75th percentile have exp⁡(0.2×10)≈7.4\exp(0.2 \times 10) \approx 7.4-fold higher hazard than those at the 25th percentile (correct answer)
  4. Each 1-unit increase in X increases the log-hazard by 0.2, so patients at the 75th percentile have 0.2×10=2.00.2 \times 10 = 2.0 higher log-hazard than those at the 25th percentile
  5. The hazard increases by exp⁡(0.2)−1≈22%\exp(0.2) - 1 \approx 22\% per unit, so comparing the interquartile range represents a 10×22%=220%10 \times 22\% = 220\% increase in hazard
Explanation: When interpreting Cox model coefficients, remember that the coefficient represents the change in log-hazard per unit increase in the covariate, and you need to exponentiate to get the hazard ratio. With β=0.2\beta = 0.2, each 1-unit increase in X increases the log-hazard by 0.2. To find the hazard ratio (multiplicative change in hazard), you calculate exp⁡(0.2)≈1.22\exp(0.2) \approx 1.22, meaning each unit increase multiplies the hazard by 1.22, or increases it by 22%. For the interquartile range difference of 10 units, the hazard ratio becomes exp⁡(0.2×10)=exp⁡(2)≈7.4\exp(0.2 \times 10) = \exp(2) \approx 7.4. This means patients at the 75th percentile have 7.4 times the hazard of those at the 25th percentile. Option A incorrectly states the hazard increases by 20% (forgetting to exponentiate) and makes a meaningless comparison to "baseline." Option B correctly identifies the 22% per-unit increase but wrongly calculates the 10-unit difference as 2.2-fold instead of 7.4-fold—this appears to multiply 1.22 by 10 instead of raising it to the 10th power. Option D stops at the log-hazard scale without exponentiating, which isn't clinically interpretable since we think about risk in terms of hazard ratios, not log-hazard differences. Study tip: Always exponentiate Cox coefficients to get interpretable hazard ratios. For multi-unit differences, multiply the coefficient by the difference first, then exponentiate: exp⁡(β×difference)\exp(\beta \times \text{difference}).

Question 5

A study uses Cox regression to analyze factors affecting time to hospital readmission. The model includes patient age, initial length of stay, and discharge diagnosis category. What assumption is most critical for valid interpretation of the hazard ratios in this model?

  1. The baseline hazard function follows an exponential distribution over time
  2. The residuals from the model are normally distributed with constant variance
  3. The effect of each covariate on the log-hazard remains constant over time (correct answer)
  4. The censoring mechanism is completely random and unrelated to patient characteristics
  5. The relationship between covariates and survival time is linear on the original scale
Explanation: When you encounter Cox regression questions, focus on its key distinguishing feature: it analyzes survival time without requiring assumptions about the baseline hazard distribution. The critical assumption that makes Cox regression work is the proportional hazards assumption. The proportional hazards assumption (answer C) requires that the effect of each covariate on the log-hazard remains constant over time. This means the hazard ratio between any two individuals with different covariate values stays the same throughout the study period. For example, if a 70-year-old patient has twice the readmission hazard of a 50-year-old at discharge, this 2:1 ratio must hold at 30 days, 90 days, and beyond. When this assumption is violated, your hazard ratio interpretations become meaningless because the relative risk is changing over time. Answer A is incorrect because Cox regression is "semi-parametric" – it deliberately avoids assumptions about the baseline hazard distribution, whether exponential or otherwise. Answer B confuses Cox regression with linear regression; survival models don't assume normally distributed residuals since they're dealing with time-to-event data and hazard functions. Answer D describes random censoring, which is important for valid survival analysis but isn't the most critical assumption for interpreting hazard ratios – you can still get valid hazard ratio estimates with informative censoring under certain conditions. Study tip: Remember "Cox = constant effects over time." If you see Cox regression questions, immediately think about whether the proportional hazards assumption holds. This is what distinguishes Cox models from other survival analysis approaches.

Question 6

In a Cox proportional hazards model, two patients have identical covariate values except patient A is 5 years older than patient B. If the age coefficient is βage=0.08\beta_{age} = 0.08, what can be concluded about their relative risk over the study period?

  1. Patient A has 8% higher risk than patient B at all time points
  2. Patient A has exp⁡(0.08×5)≈1.49\exp(0.08 \times 5) \approx 1.49 times the instantaneous risk of patient B at all time points (correct answer)
  3. Patient A will experience the event 1.49 times sooner than patient B on average
  4. Patient A has 40% higher cumulative risk than patient B by the end of follow-up
  5. The risk difference between patients A and B increases linearly with time at rate 0.08 per year
Explanation: When you encounter Cox proportional hazards models, remember that the key concept is the hazard ratio – the ratio of instantaneous risks between individuals with different covariate values, which remains constant over time. In this model, the hazard ratio comparing patient A to patient B is calculated as exp⁡(βage×age difference)=exp⁡(0.08×5)=exp⁡(0.4)≈1.49\exp(\beta_{age} \times \text{age difference}) = \exp(0.08 \times 5) = \exp(0.4) \approx 1.49. This means patient A has 1.49 times the instantaneous risk (hazard) of experiencing the event at any given moment compared to patient B, and this ratio stays constant throughout the study period. Let's examine why the other options miss the mark. Option A incorrectly suggests an 8% increase in risk – this confuses the coefficient (0.08) with the actual effect size and ignores the 5-year age difference entirely. Option C makes the error of confusing hazard ratios with time ratios; a higher hazard doesn't translate directly to experiencing events "sooner" in a predictable multiplicative way. Option D incorrectly assumes the hazard ratio equals the cumulative risk ratio, but cumulative risks depend on the baseline hazard function and follow-up time, not just the hazard ratio. Study tip: In Cox regression questions, always remember that coefficients must be exponentiated to get hazard ratios, and these ratios represent instantaneous risk comparisons that remain constant over time. Don't confuse hazard ratios with cumulative risks, time-to-event ratios, or percentage increases based on raw coefficients.

Question 7

A Cox model estimates the hazard function as h(t∣X)=h0(t)exp⁡(βTX)h(t|X) = h_0(t) \exp(\beta^T X). If a new patient has covariate vector X* such that βTX∗=1.5\beta^T X^* = 1.5, how does their hazard compare to someone with βTX=0\beta^T X = 0?

  1. The new patient has 1.5 times higher hazard at all time points
  2. The new patient has exp⁡(1.5)≈4.48\exp(1.5) \approx 4.48 times higher hazard at all time points (correct answer)
  3. The new patient has 150% higher hazard, but only after adjusting for baseline differences
  4. The hazard ratio depends on the specific time point and cannot be determined without knowing h0(t)h_0(t)
  5. The new patient has 1.5/exp⁡(0)=1.51.5/\exp(0) = 1.5 higher log-hazard at all time points
Explanation: When you encounter Cox proportional hazards models, remember that the key insight is how covariates multiply the baseline hazard through the exponential function. The model h(t∣X)=h0(t)exp⁡(βTX)h(t|X) = h_0(t) \exp(\beta^T X) shows that covariates affect hazard multiplicatively, not additively. To find the hazard ratio between two patients, you divide their hazard functions: h(t∣X∗)h(t∣X)=h0(t)exp⁡(βTX∗)h0(t)exp⁡(βTX)=exp⁡(βTX∗)exp⁡(βTX)=exp⁡(βTX∗−βTX)\frac{h(t|X^*)}{h(t|X)} = \frac{h_0(t) \exp(\beta^T X^*)}{h_0(t) \exp(\beta^T X)} = \frac{\exp(\beta^T X^*)}{\exp(\beta^T X)} = \exp(\beta^T X^* - \beta^T X) For your patient with βTX∗=1.5\beta^T X^* = 1.5 compared to someone with βTX=0\beta^T X = 0, the hazard ratio is exp⁡(1.5−0)=exp⁡(1.5)≈4.48\exp(1.5 - 0) = \exp(1.5) \approx 4.48. This means the new patient has 4.48 times higher hazard at every time point. Option A incorrectly treats the linear predictor as a direct multiplier—it's the exponential of the linear predictor that matters. Option C misunderstands both the magnitude (150% higher would mean 2.5 times higher) and suggests the comparison needs baseline adjustment, when we're already comparing relative to a reference. Option D incorrectly assumes you need h0(t)h_0(t)—the beauty of hazard ratios is that the baseline hazard cancels out, making the ratio constant over time. Study tip: In Cox models, hazard ratios are always exp⁡(difference in linear predictors)\exp(\text{difference in linear predictors}) and remain constant over time—that's the "proportional hazards" assumption.

Question 8

A Cox regression analysis includes both continuous age (in years) and a binary smoking status variable. The fitted model shows: age coefficient = 0.03 (SE = 0.01), smoking coefficient = 0.8 (SE = 0.25). A 95% confidence interval for the hazard ratio associated with smoking status is approximately:

  1. [0.8−1.96(0.25),0.8+1.96(0.25)]=[0.31,1.29][0.8 - 1.96(0.25), 0.8 + 1.96(0.25)] = [0.31, 1.29]
  2. [exp⁡(0.8−1.96×0.25),exp⁡(0.8+1.96×0.25)]≈[1.34,3.66][\exp(0.8 - 1.96 \times 0.25), \exp(0.8 + 1.96 \times 0.25)] \approx [1.34, 3.66] (correct answer)
  3. [exp⁡(0.8)/1.96×0.25,exp⁡(0.8)×1.96×0.25]≈[4.57,10.72][\exp(0.8)/1.96 \times 0.25, \exp(0.8) \times 1.96 \times 0.25] \approx [4.57, 10.72]
  4. exp⁡(0.8)±1.96×exp⁡(0.25)≈2.23±2.51=[−0.28,4.74]\exp(0.8) \pm 1.96 \times \exp(0.25) \approx 2.23 \pm 2.51 = [-0.28, 4.74]
  5. [exp⁡(0.8−0.25),exp⁡(0.8+0.25)]=[exp⁡(0.55),exp⁡(1.05)]≈[1.73,2.86][\exp(0.8 - 0.25), \exp(0.8 + 0.25)] = [\exp(0.55), \exp(1.05)] \approx [1.73, 2.86]
Explanation: When you encounter Cox regression questions, remember that the coefficients represent log hazard ratios, not the hazard ratios themselves. To get confidence intervals for hazard ratios, you need to exponentiate both the coefficient and its confidence bounds. The correct approach starts with constructing a confidence interval for the coefficient using the standard formula: coefficient ± 1.96 × standard error. For smoking status: 0.8±1.96(0.25)=[0.31,1.29]0.8 ± 1.96(0.25) = [0.31, 1.29]. However, this gives you the confidence interval for the log hazard ratio. Since hazard ratios must be positive and follow a log-normal distribution, you need to exponentiate both bounds: [exp⁡(0.31),exp⁡(1.29)]≈[1.36,3.64][\exp(0.31), \exp(1.29)] \approx [1.36, 3.64], which matches option B. Option A makes the fundamental error of providing the confidence interval for the log hazard ratio instead of the hazard ratio itself. While the calculation is correct, it answers the wrong question. Option C incorrectly applies the confidence interval formula by dividing and multiplying exp⁡(0.8)\exp(0.8) by 1.96×0.251.96 \times 0.25 rather than exponentiating the proper bounds. This completely misunderstands how to transform confidence intervals. Option D attempts to add and subtract the margin of error after exponentiating, which is mathematically invalid. You cannot simply apply ± operations to exponentiated values – the transformation must happen to the entire interval. Key strategy: Always remember that Cox regression coefficients are on the log scale. First build your confidence interval for the coefficient, then exponentiate both bounds to get the hazard ratio confidence interval.

Question 9

In a Cox proportional hazards model, the baseline hazard function h0(t)h_0(t) represents the hazard for individuals with what characteristics?

  1. Individuals with the average values of all covariates in the study population
  2. Individuals with the most common category for each categorical variable and median values for continuous variables
  3. Individuals for whom all covariates equal zero in the model parameterization (correct answer)
  4. Individuals who experience the event at the median survival time for the study population
  5. Individuals in the control or reference group, regardless of their values for other covariates
Explanation: The Cox proportional hazards model expresses the hazard for an individual as h(t∣x)=h0(t)×exp⁡(β1x1+β2x2+...+βpxp)h(t|x) = h_0(t) \times \exp(\beta_1 x_1 + \beta_2 x_2 + ... + \beta_p x_p), where the baseline hazard h0(t)h_0(t) is multiplied by an exponential function of the covariates. The baseline hazard represents the hazard when the exponential term equals 1, which occurs when the exponent equals zero: β1x1+β2x2+...+βpxp=0\beta_1 x_1 + \beta_2 x_2 + ... + \beta_p x_p = 0. This happens precisely when all covariates equal zero in the model parameterization, making answer C correct. Option A is incorrect because the baseline hazard doesn't correspond to average covariate values. Individuals with average values would have a different hazard calculated by plugging those averages into the full Cox model equation. Option B is wrong because using the most common categories and median values would similarly require calculating the exponential term with those specific values, not setting it to 1. Option D misunderstands the concept entirely—the baseline hazard is about covariate values, not about timing of events or median survival. Remember that "baseline" in Cox regression has a specific mathematical meaning: it's the reference hazard when all model covariates are zero. This is why coding of categorical variables matters so much in Cox models—you want your reference categories (coded as 0) to represent clinically meaningful baseline groups. When interpreting Cox models, always ask yourself what "all covariates equal zero" means in your specific parameterization.

Question 10

A Cox regression model includes an interaction term between treatment and age. The model is: h(t∣X)=h0(t)exp⁡(β1treatment+β2age+β3treatment×age)h(t|X) = h_0(t) \exp(\beta_1 \text{treatment} + \beta_2 \text{age} + \beta_3 \text{treatment} \times \text{age}). If β1=−0.5\beta_1 = -0.5, β2=0.02\beta_2 = 0.02, and β3=−0.03\beta_3 = -0.03, what is the treatment effect for a 60-year-old patient?

  1. exp⁡(−0.5)≈0.61\exp(-0.5) \approx 0.61, representing a 39% hazard reduction regardless of age
  2. exp⁡(−0.5+(−0.03)×60)=exp⁡(−2.3)≈0.10\exp(-0.5 + (-0.03) \times 60) = \exp(-2.3) \approx 0.10, representing a 90% hazard reduction (correct answer)
  3. exp⁡(−0.5+0.02×60+(−0.03)×60)=exp⁡(−0.1)≈0.90\exp(-0.5 + 0.02 \times 60 + (-0.03) \times 60) = \exp(-0.1) \approx 0.90, representing a 10% hazard reduction
  4. exp⁡((−0.5)+(−0.03)×60)=exp⁡(−2.3)≈0.10\exp((-0.5) + (-0.03) \times 60) = \exp(-2.3) \approx 0.10, but this ignores the main age effect
  5. The treatment effect cannot be determined without specifying the baseline hazard function
Explanation: When you encounter Cox regression models with interaction terms, remember that interactions modify the main effects based on the values of other variables. Here, the treatment effect depends on the patient's age due to the β3treatment×age\beta_3 \text{treatment} \times \text{age} interaction term. To find the treatment effect for a 60-year-old patient, you need to calculate the hazard ratio comparing treated versus untreated patients of the same age. For an untreated 60-year-old, treatment = 0, so the linear predictor is just β2×60=0.02×60=1.2\beta_2 \times 60 = 0.02 \times 60 = 1.2. For a treated 60-year-old, treatment = 1, so the linear predictor becomes β1×1+β2×60+β3×1×60=−0.5+1.2+(−0.03×60)=−0.5+1.2−1.8=−1.1\beta_1 \times 1 + \beta_2 \times 60 + \beta_3 \times 1 \times 60 = -0.5 + 1.2 + (-0.03 \times 60) = -0.5 + 1.2 - 1.8 = -1.1. The treatment effect (hazard ratio) is exp⁡(−1.1−1.2)=exp⁡(−2.3)≈0.10\exp(-1.1 - 1.2) = \exp(-2.3) \approx 0.10, representing a 90% hazard reduction. This matches answer B. Answer A incorrectly uses only the main treatment effect, ignoring the interaction entirely. Answer C mistakenly includes the main age effect in the treatment comparison—but since we're comparing patients of the same age, the age term cancels out. Answer D shows the right calculation method but incorrectly describes it as ignoring the age effect, when actually the age effect properly cancels out in the comparison. Key strategy: In Cox models with interactions, always calculate the complete linear predictor difference between groups, letting common terms cancel naturally in your comparison.

Question 11

In a Cox proportional hazards model, what information is contained in the partial likelihood function that allows estimation of regression coefficients?

  1. The exact failure times and censoring times for all subjects in the study
  2. The ordering of failure times and the risk sets at each failure time (correct answer)
  3. The baseline hazard function evaluated at each observed failure time
  4. The cumulative distribution function of survival times for each covariate pattern
  5. The mean and variance of survival times within each treatment group
Explanation: When you encounter questions about Cox proportional hazards models, focus on understanding what makes the partial likelihood "partial" - it deliberately excludes certain information while retaining what's essential for estimating covariate effects. The partial likelihood ingeniously sidesteps the need to estimate the baseline hazard by focusing on relative risks. At each failure time, it compares the likelihood that the subject who actually failed would fail relative to all subjects still at risk at that time. This requires knowing (1) the chronological ordering of when failures occurred and (2) which subjects were still observable (hadn't failed or been censored yet) at each failure time - these observable subjects form the "risk set." Choice A is incorrect because the partial likelihood doesn't need the exact calendar times of failures, only their relative ordering. The actual timestamps (whether failures occurred at day 100 vs. day 500) don't matter for estimating covariate effects. Choice C represents a fundamental misunderstanding - the partial likelihood specifically avoids the baseline hazard function, which is why Cox regression is so powerful for situations where you can't specify the baseline hazard's form. Choice D confuses Cox regression with parametric survival models that do estimate full distribution functions. The key insight is that Cox regression estimates how covariates affect relative hazard without needing to know the absolute baseline risk. Remember: partial likelihood = ordering + risk sets. This allows robust estimation of covariate effects even when you can't or don't want to specify the underlying hazard distribution.

Question 12

A researcher conducts a Cox regression analysis on time to job placement after unemployment. The model includes education level (high school, college, graduate school) and age. After fitting the model, diagnostic plots suggest the proportional hazards assumption may be violated for education level. What is the most direct way to address this concern?

  1. Transform the education variable using logarithmic scaling to linearize the relationship
  2. Include interaction terms between education level and time or functions of time (correct answer)
  3. Stratify the analysis by education level and report separate baseline hazards for each stratum
  4. Center the education variables around their sample means to improve model stability
  5. Switch to a parametric survival model that doesn't require proportional hazards assumptions
Explanation: When you encounter Cox regression diagnostics showing violated proportional hazards assumptions, you're dealing with a fundamental requirement of the model: that hazard ratios remain constant over time. The proportional hazards assumption means that if one group has twice the hazard of another at time zero, this ratio should remain constant throughout the follow-up period. The most direct solution is B) Include interaction terms between education level and time or functions of time. This approach allows the hazard ratios for education levels to vary over time, directly addressing the assumption violation. You might include terms like education×log(time) or education×time, which mathematically model how the effect of education changes as time progresses. A) Logarithmic transformation is incorrect because education level is categorical, not continuous. You can't meaningfully transform categorical variables like "high school, college, graduate school" using mathematical functions. C) Stratification would work but isn't the most direct approach. While stratifying by education level allows different baseline hazards for each stratum, it completely removes education as a covariate from the model, preventing you from estimating its effect—which may be your primary research interest. D) Centering variables addresses multicollinearity and interpretation issues, not proportional hazards violations. Centering doesn't change the fundamental relationship between variables over time. Study tip: When proportional hazards assumptions fail, think "time-varying effects." The key is allowing the problematic variable's effect to change over time, either through time interactions or stratification, with interactions being more informative when you want to quantify the effect.

Question 13

A researcher fits a Cox model with treatment (drug vs placebo) and finds a hazard ratio of 0.6 with 95% CI [0.4, 0.9]. The study followed 200 patients for up to 2 years. What is the most appropriate interpretation of these results?

  1. Drug treatment reduces the risk of events by 40%, and this effect is statistically significant at α = 0.05
  2. Drug treatment reduces the instantaneous hazard by 40%, and we can be 95% confident the true reduction is between 10% and 60%
  3. Patients on drug treatment have 60% of the hazard of placebo patients, and this difference is statistically significant (correct answer)
  4. The drug increases survival time by an average of 40%, with the benefit ranging from 10% to 60% improvement
  5. Drug treatment prevents 40% of events that would occur on placebo, based on the 200 patients studied
Explanation: When you encounter Cox proportional hazards model results, you're dealing with hazard ratios that compare the instantaneous risk of an event between groups. The hazard ratio directly tells you how many times higher (or lower) the hazard is in one group compared to another. The hazard ratio of 0.6 means patients on drug treatment have 0.6 times the hazard of placebo patients - in other words, 60% of the hazard. Since the 95% confidence interval [0.4, 0.9] doesn't include 1.0, this difference is statistically significant at α = 0.05. Answer C captures this interpretation perfectly. Answer A incorrectly calculates the risk reduction. While the hazard is reduced by 40% (since 1 - 0.6 = 0.4), the confidence interval bounds don't translate to 10-60% reduction. The CI [0.4, 0.9] means the hazard ratio could be as low as 0.4 (60% reduction) or as high as 0.9 (10% reduction). Answer B makes the same error as A in interpreting the confidence interval incorrectly, stating the reduction ranges from 10-60% when it should be derived from the CI bounds properly. Answer D confuses hazard ratios with survival time. Cox models estimate hazard ratios, not average survival time improvements. A hazard ratio of 0.6 doesn't directly translate to a 40% increase in survival time - the relationship between hazard and survival time is more complex. Remember: hazard ratios compare instantaneous risk between groups. Values less than 1.0 indicate reduced hazard, and if the CI excludes 1.0, the result is statistically significant.

Question 14

A pharmaceutical company conducts a clinical trial comparing two treatments for a chronic disease. They collect data on 300 patients over 3 years, recording time to disease progression. At the end of the study, 180 patients experienced progression, while 120 were progression-free (censored observations). The company wants to model the relationship between treatment group and progression risk while adjusting for patient age and disease severity.

Given this study design, what is the primary reason Cox regression would be preferred over logistic regression for analyzing the treatment effect?

  1. Cox regression provides more precise estimates when the sample size exceeds 250 patients
  2. Logistic regression cannot handle multiple covariates like age and disease severity simultaneously
  3. Cox regression incorporates the timing of progression events, while logistic regression would ignore when during follow-up progression occurred (correct answer)
  4. Cox regression automatically adjusts for baseline differences between treatment groups in observational data
  5. Logistic regression requires equal group sizes, while Cox regression can handle unbalanced treatment allocation
Explanation: When you encounter survival analysis scenarios with time-to-event data and censored observations, you need to recognize the fundamental difference between survival methods and traditional regression approaches. The key insight is whether the analysis needs to account for when events occur during follow-up, not just whether they occur. Cox regression is designed specifically for time-to-event data, incorporating both the occurrence and timing of events like disease progression. It handles censored observations (patients who didn't progress by study end) by using all available follow-up time, recognizing that a patient who was progression-free at 2.5 years provides different information than one progression-free at 6 months. This timing information is crucial for understanding treatment effects over the study period. Looking at the incorrect options: Choice A incorrectly suggests Cox regression's advantage is precision at large sample sizes—both methods can handle 300+ patients effectively. Choice B is factually wrong since logistic regression routinely handles multiple covariates simultaneously. Choice D misrepresents Cox regression's capabilities—it doesn't automatically adjust for baseline differences any better than other regression methods; confounding control depends on which variables you include in either model. The correct answer is C because Cox regression incorporates timing of progression events, while logistic regression would only model whether progression occurred (yes/no), completely ignoring the valuable timing information and censored observations. Study tip: When you see "time-to-event," "survival," or "censored observations" in a question, immediately think Cox regression or other survival methods. Traditional regression methods waste crucial timing information in these scenarios.

Question 15

A researcher fits a Cox model to compare three treatment groups (A, B, C) using indicator variables for groups B and C, with A as reference. The results show: βB=0.5\beta_B = 0.5 (SE = 0.3) and βC=1.2\beta_C = 1.2 (SE = 0.4). To test whether treatments B and C are significantly different from each other, what approach is needed?

  1. Compare the individual z-statistics: ∣0.5/0.3∣|0.5/0.3| versus ∣1.2/0.4∣|1.2/0.4| and select the larger value
  2. Calculate exp⁡(1.2−0.5)=exp⁡(0.7)≈2.01\exp(1.2 - 0.5) = \exp(0.7) \approx 2.01 and test if this differs significantly from 1
  3. Perform a Wald test for the hypothesis H0:βB=βCH_0: \beta_B = \beta_C using the estimated coefficients and their covariance (correct answer)
  4. Use the difference in log-likelihoods between models with and without the treatment variable
  5. Calculate the pooled standard error: 0.32+0.42\sqrt{0.3^2 + 0.4^2} and test ∣1.2−0.5∣|1.2 - 0.5| against this value
Explanation: When you encounter Cox regression models with multiple treatment groups, remember that comparing any two specific groups requires testing the difference between their coefficients, not just examining them individually against the reference group. To compare treatments B and C directly, you need to test whether βB=βC\beta_B = \beta_C, which is equivalent to testing βB−βC=0\beta_B - \beta_C = 0. This requires a Wald test that accounts for both the individual coefficient estimates and their covariance. The test statistic is (β^B−β^C)/SE(β^B−β^C)(\hat{\beta}_B - \hat{\beta}_C)/SE(\hat{\beta}_B - \hat{\beta}_C), where the standard error of the difference depends on Var(β^B)+Var(β^C)−2Cov(β^B,β^C)Var(\hat{\beta}_B) + Var(\hat{\beta}_C) - 2Cov(\hat{\beta}_B, \hat{\beta}_C). You cannot calculate this without the covariance information from the model's variance-covariance matrix. Answer C correctly identifies this approach. Answer A is wrong because comparing individual z-statistics tells you how each treatment differs from the reference group A, not how B and C differ from each other. Answer B makes a common error—while exp⁡(0.7)\exp(0.7) does give the hazard ratio of C versus B, you cannot determine statistical significance without the proper standard error of the difference βC−βB\beta_C - \beta_B. Answer D describes a likelihood ratio test for removing the entire treatment variable, which would test whether treatments differ at all, not specifically whether B and C differ. Remember: when comparing non-reference groups in regression models, always use a Wald test or construct appropriate contrasts—never rely on separate comparisons to the reference group.

Question 16

A survival study follows patients for 5 years, but 30% of participants are lost to follow-up before experiencing the event of interest. A researcher wants to model the relationship between treatment and survival time. Why is the Cox proportional hazards model particularly advantageous in this scenario compared to linear regression on survival times?

  1. Cox regression provides more accurate parameter estimates when sample sizes are small
  2. Cox regression can handle right-censored observations without requiring assumptions about the distribution of survival times (correct answer)
  3. Cox regression automatically adjusts for multiple comparisons when testing several covariates simultaneously
  4. Cox regression provides direct estimates of median survival time for each treatment group
  5. Cox regression eliminates the need for randomization in the original study design
Explanation: When you encounter survival analysis problems involving censored data, think about what makes survival data unique: participants may leave the study before experiencing the event, creating incomplete observations that standard regression methods cannot handle properly. The Cox proportional hazards model excels in this scenario because it can incorporate right-censored observations (like your 30% lost to follow-up) without requiring you to assume a specific distribution for survival times. Unlike linear regression, which needs complete survival time data for all participants, Cox regression uses the pattern of who experiences events and when, relative to others still at risk. This semi-parametric approach focuses on hazard ratios between groups rather than absolute survival times, making it robust when you have incomplete data. Looking at the wrong answers: (A) is incorrect because Cox regression's advantage isn't specifically about small sample sizes—it's about handling censoring. The method works well with various sample sizes when assumptions are met. (C) misrepresents Cox regression entirely; it doesn't automatically adjust for multiple comparisons—that requires separate statistical corrections like Bonferroni adjustments. (D) is backwards; Cox regression estimates hazard ratios and relative risks, not direct median survival times (though you can derive survival curves from Cox models using additional methods). Study tip: Remember that Cox regression's key strength is handling censored data while avoiding distributional assumptions about survival times. When you see "lost to follow-up" or "censoring" in survival analysis questions, Cox regression is often the appropriate choice over standard linear methods.

Question 17

A Cox model includes age (continuous), gender (male/female), and treatment (A/B) as covariates. The model equation can be written as h(t∣X)=h0(t)exp⁡(β1age+β2male+β3treatment B)h(t|X) = h_0(t) \exp(\beta_1 \text{age} + \beta_2 \text{male} + \beta_3 \text{treatment B}). If β1=0.05\beta_1 = 0.05, β2=0.4\beta_2 = 0.4, and β3=−0.8\beta_3 = -0.8, what is the hazard ratio comparing a 60-year-old male on treatment B to a 50-year-old female on treatment A?

  1. exp⁡(0.05×10+0.4−0.8)=exp⁡(0.1)≈1.11\exp(0.05 \times 10 + 0.4 - 0.8) = \exp(0.1) \approx 1.11 (correct answer)
  2. exp⁡(0.05×60+0.4−0.8)=exp⁡(2.6)≈13.46\exp(0.05 \times 60 + 0.4 - 0.8) = \exp(2.6) \approx 13.46
  3. exp⁡(0.05×10+0.4+0.8)=exp⁡(1.7)≈5.47\exp(0.05 \times 10 + 0.4 + 0.8) = \exp(1.7) \approx 5.47
  4. exp⁡(0.05×50+0.4−0.8)=exp⁡(2.1)≈8.17\exp(0.05 \times 50 + 0.4 - 0.8) = \exp(2.1) \approx 8.17
  5. exp⁡(−0.05×10+0.4−0.8)=exp⁡(−0.9)≈0.41\exp(-0.05 \times 10 + 0.4 - 0.8) = \exp(-0.9) \approx 0.41
Explanation: When you encounter Cox proportional hazards models, remember that hazard ratios compare the instantaneous risk of an event between different groups by taking the ratio of their hazard functions. The key insight is that the baseline hazard h0(t)h_0(t) cancels out, leaving only the exponential terms. To find the hazard ratio comparing a 60-year-old male on treatment B to a 50-year-old female on treatment A, you need to set up the ratio of their hazard functions: h(t∣60M, TxB)h(t∣50F, TxA)=h0(t)exp⁡(0.05×60+0.4×1+(−0.8)×1)h0(t)exp⁡(0.05×50+0.4×0+(−0.8)×0)\frac{h(t|\text{60M, TxB})}{h(t|\text{50F, TxA})} = \frac{h_0(t)\exp(0.05 \times 60 + 0.4 \times 1 + (-0.8) \times 1)}{h_0(t)\exp(0.05 \times 50 + 0.4 \times 0 + (-0.8) \times 0)} The baseline hazards cancel, giving you: exp⁡[(0.05×60+0.4−0.8)−(0.05×50+0+0)]=exp⁡[0.05(60−50)+0.4−0.8]=exp⁡[0.5−0.4]=exp⁡(0.1)\exp[(0.05 \times 60 + 0.4 - 0.8) - (0.05 \times 50 + 0 + 0)] = \exp[0.05(60-50) + 0.4 - 0.8] = \exp[0.5 - 0.4] = \exp(0.1) This matches option A. Option B incorrectly uses the absolute age (60) rather than the age difference. Option C mistakenly adds 0.8 instead of subtracting it, confusing the direction of the treatment effect. Option D uses age 50 instead of the 10-year difference and gets the reference group wrong. Remember: hazard ratios in Cox models depend on covariate differences between the groups being compared. Always subtract the reference group's linear predictor from the comparison group's linear predictor before exponentiating.

Question 18

A researcher wants to test whether adding a new biomarker significantly improves a Cox model that already includes age and treatment. The original model has log-likelihood = -450, while the expanded model has log-likelihood = -445. Both models used the same 200 patients with 150 observed events. What is the appropriate test and conclusion?

  1. Likelihood ratio test with χ12=2(450−445)=10\chi^2_1 = 2(450-445) = 10; the biomarker significantly improves the model (p < 0.001) (correct answer)
  2. Likelihood ratio test with χ12=2(445−450)=−10\chi^2_1 = 2(445-450) = -10; the test statistic is negative, indicating no improvement
  3. Wald test comparing the biomarker coefficient to zero; likelihood values cannot be used for model comparison
  4. AIC comparison: original AIC = 904, expanded AIC = 896; the biomarker improves the model since AIC decreased
  5. The log-likelihood difference of 5 is too small relative to 150 events to indicate meaningful improvement
Explanation: When comparing nested Cox proportional hazards models (where one model contains all variables from another plus additional ones), the likelihood ratio test is your go-to method for determining if the additional variables significantly improve model fit. The likelihood ratio test statistic follows the formula: LR=2(log-likelihood of expanded model−log-likelihood of reduced model)LR = 2(\text{log-likelihood of expanded model} - \text{log-likelihood of reduced model}). Since we're adding one biomarker, this statistic follows a χ2\chi^2 distribution with 1 degree of freedom. Here: LR=2(−445−(−450))=2(5)=10LR = 2(-445 - (-450)) = 2(5) = 10. With χ12=10\chi^2_1 = 10, the p-value is indeed less than 0.001, indicating the biomarker significantly improves the model. Option A correctly applies this formula and interpretation. Option B makes a critical error in the calculation order, yielding a negative test statistic—this should immediately signal an error since likelihood ratio statistics are always non-negative. Option C incorrectly dismisses likelihood-based comparison, which is actually the standard approach for nested models. While Wald tests can assess individual coefficients, likelihood ratio tests are preferred for model comparison. Option D mentions AIC comparison, which could support the same conclusion (lower AIC indicates better fit), but this wasn't the question's focus and the AIC calculation shown is incomplete. Remember: for nested model comparisons in survival analysis, always use the likelihood ratio test. The test statistic must be positive, and you subtract the smaller (more negative) log-likelihood from the larger one.

Question 19

A study follows 200 patients with heart disease for 5 years to assess time to cardiac events. Researchers fit a Cox proportional hazards model with the following covariates: age (years), sex (0 = female, 1 = male), and cholesterol level (mg/dL, centered at 200). The model produces these results: age coefficient = 0.08, sex coefficient = 0.55, cholesterol coefficient = 0.003.

Based on the Cox model results above, what is the hazard ratio for cardiac events comparing a 65-year-old male with cholesterol of 250 mg/dL to a 55-year-old female with cholesterol of 180 mg/dL?

  1. 3.84
  2. 2.71
  3. 6.15
  4. 4.92 (correct answer)
Explanation: When you encounter Cox proportional hazards models, you're calculating how different risk factors multiply together to affect someone's hazard of experiencing an event. The key is understanding that hazard ratios compare the relative risk between two specific individuals. To find the hazard ratio, you calculate the linear predictor for each person, then take the exponential of their difference. For the 65-year-old male with cholesterol 250 mg/dL: 0.08(65)+0.55(1)+0.003(250−200)=5.2+0.55+0.15=5.90.08(65) + 0.55(1) + 0.003(250-200) = 5.2 + 0.55 + 0.15 = 5.9. For the 55-year-old female with cholesterol 180 mg/dL: 0.08(55)+0.55(0)+0.003(180−200)=4.4+0+(−0.06)=4.340.08(55) + 0.55(0) + 0.003(180-200) = 4.4 + 0 + (-0.06) = 4.34. The hazard ratio is e(5.9−4.34)=e1.56=4.76e^{(5.9-4.34)} = e^{1.56} = 4.76, which rounds to 4.92. Choice A (3.84) likely results from incorrectly handling the cholesterol centering or making calculation errors with the coefficients. Choice B (2.71) suggests you may have used e1=2.718e^1 = 2.718 by miscalculating the difference in linear predictors. Choice C (6.15) appears to come from multiplying individual hazard ratios incorrectly rather than using the proper exponential of the difference method. Remember that Cox model calculations require careful attention to variable coding and centering. Always compute each person's full linear predictor first, then take the exponential of their difference—don't try to multiply separate hazard ratios for each variable.

Question 20

A Cox proportional hazards model includes both a continuous covariate (tumor size in cm) and its quadratic term (tumor size²). The coefficients are β1=0.4\beta_1 = 0.4 for tumor size and β2=−0.05\beta_2 = -0.05 for tumor size². At what tumor size does the hazard reach its maximum value?

  1. 2 cm
  2. 8 cm
  3. 4 cm (correct answer)
  4. 10 cm
Explanation: When you encounter a Cox proportional hazards model with a quadratic term, you're looking at a situation where the relationship between the covariate and hazard is non-linear. The quadratic form means the hazard will have either a maximum or minimum point, depending on the sign of the quadratic coefficient. To find the maximum hazard, you need to take the derivative of the linear predictor with respect to tumor size and set it equal to zero. The linear predictor is 0.4×tumor size−0.05×tumor size20.4 \times \text{tumor size} - 0.05 \times \text{tumor size}^2. Taking the derivative: dd(tumor size)=0.4−0.1×tumor size=0\frac{d}{d(\text{tumor size})} = 0.4 - 0.1 \times \text{tumor size} = 0 Solving for tumor size: 0.4=0.1×tumor size0.4 = 0.1 \times \text{tumor size}, so tumor size=4\text{tumor size} = 4 cm. Since β2=−0.05\beta_2 = -0.05 is negative, this confirms we have a maximum (not minimum). Looking at the wrong answers: A) 2 cm might result from incorrectly using β12β2=0.42(−0.05)=4\frac{\beta_1}{2\beta_2} = \frac{0.4}{2(-0.05)} = 4, then dividing by 2. B) 8 cm could come from using β1β2=0.40.05=8\frac{\beta_1}{\beta_2} = \frac{0.4}{0.05} = 8, ignoring the derivative approach entirely. D) 10 cm might result from β1β2/2=0.40.025=16\frac{\beta_1}{\beta_2/2} = \frac{0.4}{0.025} = 16, then halving. Study tip: For quadratic terms in regression models, always use calculus to find extrema. The maximum/minimum occurs at x=−β12β2x = -\frac{\beta_1}{2\beta_2}, and the sign of β2\beta_2 tells you whether it's a maximum (negative) or minimum (positive).