Oral Exam Preparation for Week 1 (BIS 630)

1. Concepts from Introductory Statistics

Main Question:
"What statistical concepts from introductory courses are foundational for survival analysis?"

Follow-up Questions:

 "How is regression used in survival analysis?"

 "Why are residuals important in model diagnostics?"

Answer Key Points:

 Concepts include: random variables, distributions (pdf, cdf), estimation methods (MLE), hypothesis
testing, regression models, and diagnostics.
 Survival analysis extends these ideas with time-to-event data, censoring, and specialized models
like the Cox model.

2. Definition and Characteristics of Survival Time

Main Question:
"What is survival time, and what are its typical characteristics?"

Follow-up Questions:

 "Can you give examples of survival time in different fields?"

 "Is survival time typically normally distributed?"

Answer Key Points:

 Survival time measures duration from an origin point to a specific event (e.g., death, relapse).
 Key characteristics: non-negative, typically right-skewed, often censored.
 Normal distribution is inappropriate; special models are needed.

3. Types of Censoring

Main Question:
"What are the different types of censoring in survival data?"
Answer Key Points:

 Types of censoring:
o Right censoring: Event has not yet occurred by the end of observation (most common).
o Left censoring: Event occurred before observation starts.
o Interval censoring: Event known to occur within a time interval.

 In right-censoring, we only know the event time exceeds a certain threshold.

4. Key Functions in Survival Analysis

Main Question:
"What are the key functions used to describe survival data?"

Follow-up Questions:

 "How are the survival function and hazard function related?"

 "What is the cumulative hazard function?"

Answer Key Points:

5. Independent Censoring Assumption

Main Question:
"What does the independent censoring assumption mean in survival analysis?"

Follow-up Questions:

 "Why is this assumption important for estimation?"

 "Can you give an example where independent censoring might be violated?"

Answer Key Points:

  Independent censoring means the mechanism causing censoring is unrelated to the risk of the
event.
 Violations (e.g., if patients leave a study due to worsening health) can bias survival estimates.
 Independent censoring ensures valid estimation using Kaplan-Meier, Cox, and other standard
methods.

6. Likelihood for Right-Censored Data

Main Question:
"How is the likelihood function constructed for right-censored survival data?"

Follow-up Questions:

 "How does censoring affect the likelihood contribution of an observation?"

 "What assumptions are needed?"

Answer Key Points:

 Likelihood for each individual:

o If event observed: contributes f(ti).
o If censored: contributes S(ti).

 Full likelihood:

 Assumes independence between censoring and event times.

Oral Exam Preparation for Week 2 (BIS 630)

📈 1. Kaplan-Meier Estimator for S(t)

Main Question:
"What is the Kaplan-Meier estimator, and how is it constructed?"

Follow-up Questions:

 "Why is it called a step function?"

 "How does censoring affect the Kaplan-Meier estimate?"

Answer Key Points:

 KM estimator formula:

 Step function: updates only at event times.

 Censoring affects the number at risk ni but not the number of events di.
 At each event time: conditional probability of surviving multiplied together.

🔍 2. Cumulative Hazard Function Estimation

Main Question:
"How can the cumulative hazard function be estimated from survival data?"

Follow-up Questions:

 "What is the Nelson-Aalen estimator?"

 "How is it related to the Kaplan-Meier estimator?"
Answer Key Points:

 Conventionally, Kaplan-Meier is used for survival function, Nelson-Aalen for cumulative hazard.

🧮 3. Estimating Quantiles (tp)（图）

Main Question:
"How do we estimate quantiles, such as the median survival time, in survival analysis?"

Follow-up Questions:

 "What if the survival curve is flat at exactly 1-p?"

 "How does the R survival package handle this?"

⚔️4. Log-Rank Test for Comparing Two Survival Distributions

Main Question:
"What is the purpose of the log-rank test, and how is it performed?"

Follow-up Questions:

 "What are the null and alternative hypotheses?"

 "Why is the log-rank test powerful under proportional hazards?"
性质：

Independent censoring；non- parameters；most powerful ph；kai-square；

🎯 5. Practical Notes on Log-Rank Test

Main Question:
"What are the limitations of the log-rank test?"

Follow-up Questions:

 "What happens if survival curves cross?"

 "Is log-rank test still valid without proportional hazards?"

Answer Key Points:

 If survival curves cross, log-rank test loses power.

 It tests global differences across time, not just at a single point.
 Formal proportional hazards assumption is important (but log-rank doesn't check it explicitly).
Oral Exam Preparation for Week 3 (BIS 630)

Cox Proportional Hazards (PH) Model Basics

Main Question:
"What is the Cox proportional hazards model and what are its key components?"

Follow-up Questions:

 "What is the baseline hazard?"

 "What does the model assume?"

Answer Key Points:

 The Cox PH model:

 Semi-parametric: no assumptions about the shape of h0(t), but assumes proportional hazards.
 The baseline hazard h0(t) is the hazard when Z=0.
 Hazard ratio between two individuals depends only on covariates, not on time.

Hazard Ratio Interpretation

Main Question:
"What is a hazard ratio and how do you interpret it in the Cox model?"

Follow-up Questions:

 "What if you have a continuous covariate?"

  "How would you interpret an estimated HR of 0.3?"

Partial Likelihood (No Ties)

Main Question:
"How is the partial likelihood constructed in the Cox model?"

Follow-up Questions:

 "What is the role of the risk set?"

 "Why is it called 'partial'?"

Handling Tied Event Times

Main Question:
"What methods are used to handle tied survival times in Cox models?"

Follow-up Questions:
  "What are the trade-offs between the methods?"
 "Which method is used by default in R?"

Answer Key Points:

 Three main methods:

1. Exact method: computationally expensive, accurate.
2. Breslow: simplest, but can bias estimates toward null if many ties.
3. Efron: better approximation, used as default in R (coxph).
 Tied times happen when events occur at the same observed time.
 Use ties = "efron" in coxph().
Oral Exam Preparation for Week 4 (BIS 630)

1. Hypothesis Testing in Cox Models

Main Question:
"What are the three main types of hypothesis tests used for Cox models, and how do they differ?"

Follow-up Questions:

 "When would you prefer using a score test instead of a Wald test?"
 "What is the advantage of the likelihood ratio test?"

Answer Key Points:

 Wald test: Based on the asymptotic normality of MLE; easy after fitting the model.
 Score test: Only needs estimation under the null; used when fitting full model is difficult.
 Likelihood Ratio test: Compares the likelihood under null and alternative; most robust among the three.

2. Different Types of Residuals in Cox Models

Main Question:
"Can you describe different types of residuals used in Cox regression and their purposes?"
Follow-up Questions:

 "Which residual is best for detecting outliers?"

 "Which residual is used to check the functional form of covariates?"

Answer Key Points:

 Cox-Snell residuals: Assess overall model fit (should follow Exponential(1)).

 Martingale residuals: Check functional form of covariates; skewed, not good for outlier detection.
 Deviance residuals: Symmetric version of Martingale; used for identifying outliers.
 Schoenfeld residuals: Check time-dependence of covariates, especially PH assumption.

3. Visual Tools for Model Fit Evaluation

Main Question:
"What are the key plots used to assess the fit of a Cox model?"

Follow-up Questions:

 "How would you interpret a Cox-Snell residual plot?"

 "Why are martingale residual plots useful?"

Answer Key Points:

 Cox-Snell plots: Should approximate a straight line with slope 1 through the origin.
 Martingale residual plots: Help determine if a covariate needs transformation (nonlinearity). (functional
form)
 Deviance residuals: Index plots detect outliers.
 Warning: Visual tools are subjective; interpretation can vary, especially with small samples.

4. Examining the Proportional Hazards Assumption

Main Question:
"How can you evaluate the proportional hazards assumption in Cox models?"

Follow-up Questions:

 "What are the graphical methods available?"

 "What formal tests can be conducted?"

Answer Key Points:

  Graphical methods:
o Log(-log(Survival)) plots: Curves should be roughly parallel.
o Scaled Schoenfeld residuals vs time: Horizontal line suggests PH assumption holds.

 Formal test:
o Based on scaled Schoenfeld residuals (Grambsch and Therneau test).
o Test H0: PH assumption holds vs Ha: PH assumption violated.

 Alternative approach:
o Add time-dependent covariates to the model and check their significance.

5. Time-dependent Covariates Approach

Main Question:
"How can time-dependent covariates be used to check the PH assumption?"

Follow-up Questions:

 "What is the interpretation if the time-dependent covariate is significant?"

 "Why can't we treat time-dependent covariates the same way as baseline covariates?"

Answer Key Points:

 Introduce an interaction term between a covariate and time (e.g., Z1×tZ1 \times tZ1×t).
 If interaction term is significant, PH assumption is violated.
 Time-dependent covariates change at each event time; complicate partial likelihood calculation.

Oral Exam Preparation for Week 5 (BIS 630)

1. Introduction to Time-Dependent Covariates

Main Question:

"What is a time-dependent covariate, and how does it differ from a time-independent covariate?"

Follow-up Questions:

 "Can you give examples of external and internal time-dependent covariates?"

 "Why is it risky to control for internal time-dependent covariates?"

Answer Key Points:

 Time-dependent covariate: A variable whose value can change over time during the follow-up period.

 External covariate: Changes independently of the subject (e.g., air temperature).

  Internal covariate: Depends on the subject's health status (e.g., blood pressure).

 Risk: Controlling for internal covariates affected by treatment may obscure the true treatment effect.

2. Modeling with Time-Dependent Covariates

Main Question:

"How is the Cox model modified to include time-dependent covariates?"

Follow-up Questions:

 "What is the interpretation of the hazard ratio in a time-dependent Cox model?"

 "Is the hazard ratio still constant over time?"

Add a interaction term of the covariate and the functions of time

3. Stanford Heart Transplant Study

Main Question:

"Why do we need to use a time-dependent Cox model in the Stanford heart transplant study?"

Follow-up Questions:

 "What is immortal time bias?"

 "How do you define the transplant covariate in this study?"

Answer Key Points:

 Patients initially have no transplant, status changes when they receive transplant.

 Immortal time bias: If a subject must survive long enough to receive a treatment, incorrectly modeling it

can falsely exaggerate treatment effects.

 Correct model: Define Z(t)as 1 if transplanted before time t, 0 otherwise.

4. Data Formatting for Time-Dependent Covariates

Main Question:

"How should the data be formatted to fit a time-dependent Cox model?"

Follow-up Questions:

 "What is the (Start, Stop, Status) format (counting process format)?"

 "How is the event indicator coded across time intervals?"

Answer Key Points:

 Data needs to be split into multiple intervals where covariate values are constant.

 Each record includes Start time, Stop time, and Status (event indicator: 1=event, 0=no event).

 Only the last interval (if uncensored) has event = 1; others have event = 0.

5. Estimation in Time-Dependent Cox Model

Main Question:

"How is the partial likelihood modified for time-dependent covariates?"

Follow-up Questions:

 "Why do we need to know covariate values at each death time?"

 "What challenges might arise when estimating with internal time-dependent covariates?"

6. Exercise: Cumulative Exposure Example

Main Question:

"In the smoking and lung cancer example, how would you model cumulative exposure?"

Follow-up Questions:
  "What would the hazard ratio represent?"

 "Why is cumulative exposure considered time-dependent?"

Oral Exam Preparation for Week 6 (BIS 630)

Motivation for Parametric Models

Main Question:

"Why might we consider using parametric survival models instead of Cox models?"

Follow-up Questions:

 "What are the advantages of parametric models?"

 "When is a parametric model particularly useful?"

Answer Key Points:

 PH assumption in Cox may not hold.

 Parametric models can be more efficient if correctly specified.

 Easier to handle complex censoring.

 Direct interpretation on survival time (especially AFT models).

2️⃣ Popular Parametric Distributions

Main Question:

"What are the common distributions used in survival analysis and their characteristics?"

Follow-up Questions:

 "How does the Weibull distribution relate to exponential?"

 "Which distributions allow crossing hazards?"

Answer Key Points:

  Common distributions:

o Exponential: constant hazard (memoryless).

o Weibull: increasing or decreasing hazard depending on shape parameter.

o Log-normal: hazard increases then decreases (non-monotonic).

o Log-logistic: similar to log-normal but different tail behavior.

 Weibull generalizes exponential (Weibull with shape = 1 is exponential).

 Log-normal and log-logistic allow crossing hazards.

Accelerated Failure Time (AFT) Model Basics

Main Question:

"What is the structure of the accelerated failure time model?"

Follow-up Questions:

 "How is the acceleration factor interpreted?"

 "How is the survival function shifted under AFT?"

AFT assumption: covariates affect survival time by accelerating or decelerating the entire survival distr by a constant

factor

4️⃣ Interpretation of Coefficients in AFT Models

Main Question:
"How do you interpret a positive or negative coefficient in an AFT model?"

Follow-up Questions:

 "What happens if time ratio: exp(β) > 1?"

 "What if exp(β) < 1?"

Answer Key Points:

 exp(β) > 1: Time is stretched → slower event → longer survival.

 exp(β) < 1: Time is compressed → faster event → shorter survival.

Example:

 exp(β) = 1.25 → survival time increases by 25%.

 exp(β) = 0.8 → survival time decreases by 20%.

5️⃣ Weibull Model: Both PH and AFT

Main Question:

"Why is the Weibull distribution special in survival analysis?"

Follow-up Questions:

 "What happens to the coefficients under PH vs AFT for Weibull?"

Answer Key Points:

 Weibull satisfies both Proportional Hazards (PH) and AFT properties.

 In AFT: log(T) is linear in covariates.

 In PH: log(h(t)) is linear in covariates.

 Transformation between models:

o From AFT to PH coefficients: −γβ or −β/σ.

6️⃣ Proportional Odds Model

Main Question:

"What is the proportional odds (PO) model in survival analysis?"

Follow-up Questions:

 "How is it different from proportional hazards?"

  "Which distribution satisfies PO?"

7️⃣ Model Checking: Graphical Methods

Main Question:

"How can we graphically assess the fit of a parametric survival model?"

Follow-up Questions:

 "What plots are used for exponential, Weibull, log-logistic, log-normal?"

Percentile - percentile plot: exam the aft assumption

If the aft is valid, the plot should be a line through origin and the slope is the acceleration factor.
Handle non-PH:

1. Time dependent

2. Parametric model

3. Stratified model

Oral Exam Preparation for Week 10 (BIS 630)

Stratified Cox Models

Main Question:

"When and why would you use a stratified Cox model?"

Follow-up Questions:

 "How is the hazard function defined in a stratified Cox model?"

 "What are the limitations of using stratification?"

can we develop a different summary to quantify the covariate’s effect

Restricted mean survival time (RMST)

Testing for Heterogeneous Effects

Main Question:

"How can you test if the covariate effect differs across strata in a stratified Cox model?"

Follow-up Questions:

 "What does the likelihood ratio test compare?"

 "What is the null hypothesis?"

Answer Key Points:

 Test whether β is constant across strata.

 Compare:

o Stratified model: shared β across strata

o Separate models: different β per stratum

 Null: No interaction between covariates and strata

Restricted Mean Survival Time (RMST)

Main Question:

"What is restricted mean survival time, and how is it estimated?"

Follow-up Questions:

 "How does RMST compare to median survival or hazard ratio?"

 "What is the interpretation of RMST?

Interpretation: average time survived over a t0 period.

We can see that

- Only observed restricted survival times are included

- Inversely weight each individual by the probability of observing the survival time

Comparing Groups Using RMST

Main Question:

"How can RMST be used to compare treatment groups?"

Follow-up Questions:

 "What are the pros and cons of using RMST for treatment effect?"

 "What’s the difference between RMST difference and HR?"

Regression Using RMST: Pseudo-Values and IPCW

Main Question:

"How do we use regression to model RMST?"

Follow-up Questions:

 "What are pseudo-values?"

 "What is the IPCW approach?"

We can see that

- Only observed restricted survival times are included
- Inversely weight each individual by the probability of observing the survival time

Oral Exam Preparation for Week 11 (BIS 630)

Definition and Impact of Left Truncation

Main Question:

"What is left truncation, and how is it different from censoring?"

Follow-up Questions:

 "What are some real-life examples of left truncation?"

 "What is the consequence of not adjusting for left truncation?"

Answer Key Points:

 Left truncation: Subjects are only observed if their survival time is longer than a truncation point

(delayed entry).

 Censoring: Event is not observed during the study period, but subject is included.

 Left truncation causes sampling bias by excluding early events, biasing survival upwards.

 Example: Atomic bomb survivor study – no data for those who died before 1950 (5 years after

exposure).

Handling Left Truncation in Estimation

Main Question:

"How can survival estimation methods be adjusted for left truncation?"

Follow-up Questions:

 "How is the risk set modified in Cox regression with left truncation?"

 "Can Kaplan-Meier be used under left truncation?"

 Risk sets are not nested, and may increase over time (unlike right-censoring only setting).
5️⃣ Practical Considerations and Assumptions

Main Question:

"What assumptions are required for valid inference under left truncation?"

Follow-up Questions:

 "What happens if event time and entry time are dependent?"

 "How can baseline age be misused?"

Answer Key Points:

 Assumes independence between event time and truncation time, given covariates.

 Violation causes informative truncation, leading to bias.

 Baseline age should be used as truncation time if follow-up starts after a fixed age, not as a

covariate.
Oral Exam Preparation for Week 12 (BIS 630)

🧮 1. Sample Size Calculation in Survival Analysis

Main Question:

"How do we calculate the required sample size for a survival analysis study?"

Follow-up Questions:

 "Why do we need to consider the number of events, not just total sample size?"

 "What assumptions are commonly made when calculating sample size for survival analysis?"

 "What’s the role of the hazard ratio in this process?"

Answer Key Points:

 Sample size is based on number of events (deaths) rather than total subjects.

 Steps:

1. Define effect size: hazard ratio or change in median survival.

2. Calculate number of events using:

3. Adjust for the probability of observing an event to get total required sample size:

 Assumptions often include exponential distribution of survival times.

 Proportional hazards (PH) assumption is typically made.

 K-M estimates or parametric approximations (e.g., median survival time) may be used.

🔁 2. Dependent vs Independent Censoring

Main Question:

"What is dependent censoring, and how does it affect survival analysis?"

Follow-up Questions:
  "How can we detect dependent censoring?"

 "Why is it problematic?"

Answer Key Points:

 Independent censoring means censoring time is unrelated to event time.

 Dependent censoring violates this and can bias survival estimates.

 Detection: Plot risk scores (from survival model) vs. censoring scores (from censoring model).

o If correlated, dependence exists.

 Consequences:

o Positive correlation: overestimates survival

o Negative correlation: underestimates survival

3. Sensitivity Analysis for Dependent Censoring

Main Question:

"How can we check whether dependent censoring influences the result?"

Follow-up Questions:

 "What are the two extreme assumptions in sensitivity analysis?"

 "What would you conclude if the results stay consistent?"

Answer Key Points:

 Sensitivity analysis approaches:

1. Assume censored subjects have shorter survival → treat censored time as death time.

2. Assume censored subjects have longer survival → replace censored time with longest

observed time.

 If all results are similar, outcome is likely robust to dependent censoring.

⚖️4. Inverse Probability of Censoring Weighting (IPCW)

Main Question:

"What is IPCW and how is it used to address dependent censoring?"

Follow-up Questions:

 "How are the weights computed?"

 "How do we incorporate IPCW in a Cox model?"

Answer Key Points:

 IPCW adjusts for dependent censoring by re-weighting observed data.

 Steps:

1. Model the censoring process (censoring is the event).

2. Compute weights as inverse of estimated censoring survival probability G(t)G(t)G(t).

3. Fit Cox model with these weights:

coxph(Surv(start, stop, event) ~ covariates, weights = ..., data = ...)

 May use robust SE with cluster(id).

📄 5. Data Structure for IPCW or Time-dependent Covariates

Main Question:

"How should data be formatted to apply IPCW or fit time-dependent models?"

Follow-up Questions:

 "What is the (start, stop, status) format?"

 "Why is it needed?"
Answer Key Points:

 Use counting process format:

o Each subject’s follow-up divided into intervals where covariates and weights are constant.

o Format: (start, stop, event) — similar to time-dependent Cox.

 IPCW requires time-varying weights, so we need this format to apply the method correctly.