Detailed STAT 509 Lesson 6 - Clinical Trial Objectives, Endpoints, and Estimation

A comprehensive vocabulary study set for STAT 509 Lesson 6, covering clinical trial endpoints, censoring, dose-finding, risk calculations, survival analysis (Kaplan-Meier, Logrank), and estimation methods like bootstrapping.

Clinical trial objective

The specific goals of a trial that must be stated in specific terms; success should depend on obtaining a valid result rather than a predetermined numerical outcome.

Endpoint (outcome)

A quantitative measurement determined for each study participant that is required by the scientific objectives.

Primary objective / primary endpoint

The principal objective or outcome of a clinical trial, which forms the basis for the trial's sample-size calculation.

Secondary objective / secondary endpoint

Additional objectives or outcomes beyond the primary one, with statistical power constrained by the sample size selected for the primary objective.

Hard endpoint

An endpoint well-defined in the protocol, definitive with respect to the disease process, and requiring no subjectivity.

Soft endpoint

An endpoint that does not relate strongly to the disease process or requires subjective assessment by investigators and/or patients.

Continuous measurement

A numerical measurement on a continuous scale, such as blood pressure, weight, or blood chemistry variables.

Event time

An endpoint measuring the time until an event occurs, such as survival time or time to cancer recurrence.

Count

The number or frequency of occurrences, such as migraine-headache frequency or use of rescue medication.

Binary endpoint

An outcome with only two categories, such as recurrence/no recurrence or success/failure.

Ordered categories

Categories with a meaningful order, such as pain levels (absent, mild, moderate, severe) or NYHA status.

Unordered categories

Categories without a natural ranking, such as types of adverse experiences (GI vs. cardiac).

Repeated measurements

Endpoints that are assessed multiple times during the course of a study.

Censoring

Occurs when some follow-up information is available for a participant, but the exact event time is not fully observed.

Right censoring

Censoring where the event has not been observed by the last known follow-up time due to study termination, loss to follow-up, or death from an unrelated cause.

Left censoring

Occurs when the initiation time for the subject, such as the exact time of diagnosis, is unknown.

Interval censoring

Occurs when a subject is not observed for a period and it is unknown exactly when the event occurred within that specific interval.

Type I right censoring

A scenario where all subjects begin at the same time and the study ends at a fixed, predetermined time.

Type II right censoring

A scenario where all subjects begin at the same time and the study terminates once a predetermined proportion of subjects have experienced the event.

Type III right censoring

Random censoring commonly found in clinical trials due to staggered entry and unequal follow-up times.

Independent censoring assumption

The assumption in event-time methods that the censoring mechanism and the event occurrence are not related.

Informative censoring

A situation where patients tend to be censored before an event occurs, which can lead to an underestimation of the event rate.

Disease-specific death problem

The risk of bias because causes of death may not be independent and determining the exact cause of death can be difficult.

Surrogate endpoint

An outcome measured in place of a biologically definitive or clinically meaningful endpoint; it typically tracks disease progress or extent.

Central validity question for surrogates

Whether the surrogate endpoint is strongly associated with the definitive outcome.

Surrogate causal pathway

The theoretical sequence where the disease affects the surrogate endpoint, which in turn affects the definitive endpoint.

CD4 count

A common surrogate endpoint example used in AIDS clinical trials.

Tumor-size reduction

A common surrogate endpoint example used in cancer clinical trials.

Intraocular pressure

A common surrogate endpoint example used in glaucoma clinical trials.

Phase I study (dose-finding)

An early trial aimed at determining the optimal biological dose ($OBD$) of a drug.

Dose-escalation

A design component specifying how doses are increased for subsequent groups of subjects.

Therapeutic index criterion

Defining an optimum dose as the dose with the maximal separation between risk and benefit.

Minimum effective dose (MED)

For example, the dose that eliminates mild-to-moderate pain in $80\%$ of participants.

Maximum nontoxic dose (MND)

The highest dose associated with serious side effects in no more than $1$ of $20$ patients.

Maximum tolerated dose (MTD)

For a cytotoxic cancer drug, the dose yielding serious but reversible toxicity in no more than $30\%$ of patients.

Dose-limiting toxicity (DLT)

A specific definition of toxicity used to govern dose-increase or dose-decrease decisions in a trial.

Continual reassessment method (CRM)

A dose-finding algorithm that fits a mathematical model to accumulating data and updates the estimate of the optimal dose for the next cohort.

Fibonacci sequence in dosing

A sequence ($1, 1, 2, 3, 5, 8, ...$) where dose increments are based on the sum of the two previous levels.

Risk difference (RD) formula

$RD = p_T - p_P$, where $p_T$ is the treatment event probability and $p_P$ is the placebo event probability.

Relative risk (RR) formula

$$RR = \frac{p_T}{p_P}$$

Odds formula

$\text{odds} = \frac{p}{1-p}$ for an event probability $p$.

Odds ratio (OR) formula

$OR = \frac{\frac{p_T}{1-p_T}}{\frac{p_P}{1-p_P}}$, comparing the odds of an event between two groups.

Null value for odds ratio

$OR = 1.0$, indicating that the odds are equal in both groups.

When OR approximates RR

Occurs when the event probabilities in both groups are relatively small (rare events).

Estimated relative risk from 2 x 2 table

$\hat{RR} = \frac{\frac{a}{a+b}}{\frac{c}{c+d}}$ where $a$ and $c$ are successes in Groups 1 and 2.

Estimated odds ratio from 2 x 2 table

$\hat{OR} = \frac{ad}{bc}$ where $a, b$ are group 1 success/failure and $c, d$ are group 2 success/failure.

SE of log odds ratio

$SE[\ln(\hat{OR})] = \sqrt{\frac{1}{a} + \frac{1}{b} + \frac{1}{c} + \frac{1}{d}}$, used to construct confidence intervals.

Fisher's exact test

A test preferred for equal binary proportions when the expected-cell-count conditions for normal approximations are not met.

Mantel-Haenszel procedure

A method to calculate an adjusted odds ratio by combining individual estimates from different strata.

Mantel-Haenszel assumption

The assumption that the odds ratio is constant across all strata, even if event rates differ between them.

Dose-response / trend analysis

Tests whether increasing dose is associated with a monotone increase or decrease in response.

Jonckheere-Terpstra (JT) trend test

A nonparametric test based on the sum of Mann-Whitney-Wilcoxon comparisons across ordered pairs of dose groups.

JT supports increasing trend

Indicated by a large positive JT statistic.

Cochran-Armitage (CA) trend test

A trend test specifically for binary data that can utilize actual numerical dose levels.

Kaplan-Meier (KM) estimator

A nonparametric method for estimating survival probability $S(t)$ over time in the presence of censoring.

Kaplan-Meier formula

$\hat{S}(t) = \prod_{t_k \le t} \left( 1 - \frac{d_k}{n_k} \right)$ where $d_k$ is failures and $n_k$ is the risk set.

S(t) definition

The probability of an individual surviving beyond time $t$.

KM risk set (n_k)

The number of patients at risk and still under observation immediately before failure time $t_k$.

Why KM is a step function

Because the survival estimate only changes at times when a failure or event actually occurs.

Change-from-baseline score

$D_i = Y_i - X_i$, where $X_i$ is the baseline measurement and $Y_i$ is the final measurement.

Variance of a difference formula

$$Var(Y-X) = Var(Y) + Var(X) - 2Cov(X,Y)$$

Precision of change scores

The variance of the change can be smaller than the variance of a single measurement if the within-person correlation $\rho$ is high and positive.

ANCOVA model for treatment effect

$E(Y_i) = \beta_0 + \beta_1 T_i + \beta_2 X_i$, where $\beta_1$ represents the adjusted difference between treatments.

Logrank test

A Mantel-Haenszel-type nonparametric test used to compare the survival curves of different groups.

Logrank Z statistic

$Z = \frac{O-E}{\sqrt{\sum V_k}}$, which follows a standard normal distribution under the null hypothesis.

Generalized Wilcoxon test

A nonparametric survival-curve comparison that extends the Wilcoxon rank-sum test but does not assume proportional hazards.

P-value limitation in clinical trials

They do not convey the magnitude of clinical effect and are heavily influenced by sample size.

Bootstrap

A resampling method used to estimate the variability or variance of an estimator by sampling with replacement.

Bootstrap variance formula

$$Var_{boot}(\hat{\theta}) = \frac{1}{B-1} \sum_{b=1}^{B} (\hat{\theta}_b - \bar{\theta}_{boot})^2$$

Exploratory analysis

An analysis beyond the primary protocol objectives, often used to generate hypotheses rather than confirm them.

Multiple-testing problem

The inflation of Type I error rate that occurs when many hypothesis tests are performed on the same data set.

Subset analysis

Comparing treatment groups within specific subgroups, such as age or gender, often as an exploratory exercise.

Wait-loss primary endpoint example

Mean absolute change from baseline weight at $1$ year.

NYHA status

An example of an endpoint measured in ordered categories.

Type I right censoring occurrence

Common in laboratory animal experiments but rare in human trials.

Life-table analysis

A method of survival analysis mentioned in the lesson notes.

Two required measurements for event-time analysis

Follow-up time and an indicator for whether that time represents an event or a censoring point.

Effectively identifying OBD

The goal of sequentially rising doses in dose-finding trials.

Fibonacci dose-ranging example sequence

$D$, then $2D$, then $3D$, then $5D$, eventually moving to $8D$, $13D$, etc.

Modified Fibonacci design

A dose-finding design that may include rules for de-escalation as well as escalation.

Calculated risk difference for 0.25 vs 0.05

$$0.20$$

Calculated Relative Risk for 0.25 vs 0.05

$$5.00$$

Calculated Odds Ratio for 0.25 vs 0.05

$$6.33$$

Calculated risk difference for 0.30 vs 0.10

$$0.20$$

Calculated Relative Risk for 0.30 vs 0.10

$$3.00$$

Calculated Odds Ratio for 0.30 vs 0.10

$$3.86$$

Normal-approximation z test

A test used to compare two binary proportions when sample sizes are large.

SAS PROC FREQ /cmh

The SAS syntax option used to perform a Mantel-Haenszel analysis.

Two-sided trend alternative

Allows for the possibility of either a monotone increasing or monotone decreasing response ordering.

JT tests population medians

The formal statistical parameter compared by the Jonckheere-Terpstra test.

Homogeneous-variance assumption

The assumption that all groups in a parametric trend test share a common population variance.

CONTRAST statement in PROC GLM

The SAS method used to implement a parametric trend test for continuous data.

Kaplan-Meier robustness

The property that the estimator is not adversely affected by unusual data points compared to parametric models.

Meaning of KM step down

Visual indication that an event or failure has occurred at that specific point in time.

Geometric mean and CV

Descriptive statistics preferred when measurement data are lognormally distributed.

Patient as own control

The strategy of using baseline subtraction to improve precision in clinical measurements.

Proportional hazards assumption

An assumption required by the logrank test where hazard functions are constant ratios over time.

Mesothelioma Example 6.4 logrank p-value

$$p = 0.48$$

Mesothelioma Example 6.4 Wilcoxon p-value

$$p = 0.63$$

Sampling with replacement probability

In the bootstrap, each original observation has a probability of $1/N$ for selection on each draw.