Detailed STAT 509 - Lesson 10: Late Phase Studies - Interim Analyses

A complete set of 100 vocabulary flashcards for STAT 509 Lesson 10, covering interim monitoring, group sequential methods (Pocock, Haybittle-Peto, O'Brien-Fleming), alpha-spending functions, futility, adaptive designs, and DSMB/IRB protocols.

Frequentist interim-monitoring focus

Control the overall Type I error rate while repeatedly examining accumulating data.

DSMB framework

Data Safety Monitoring Boards use interim information to assess participant safety, efficacy, trial conduct, and whether a study should continue or change.

Repeated interim testing problem

From a frequentist perspective, repeatedly testing accumulating data increases the Type I error rate unless the testing plan adjusts for repeated looks.

Interim analysis frequency in multi-center trials

Often only once or twice per year; can detect treatment effects nearly as early as continuous monitoring.

Group sequential analysis

A design in which only a few prescheduled statistical analyses are conducted as data accumulate.

R

The number of planned analyses, including interim analyses and the final analysis.

Test statistic at analysis r

The statistic calculated from all accumulated data available at the r-th analysis.

Boundary point/critical value

A prespecified cutoff used to decide whether evidence is sufficiently strong to reject the null hypothesis and stop the trial.

Group sequential stopping principle

At an interim analysis, terminate with rejection of $H_0$ when the test statistic crosses the prespecified rejection boundary.

Boundary-selection objective

Choose boundaries so the overall significance level across all analyses does not exceed the desired alpha.

Equal information/accrual assumption

Assumes $n$ new patients are accrued at each of $R$ analyses, for total sample size $R \times n$.

Pocock approach

Uses the same significance level or critical boundary at every scheduled analysis.

Pocock main advantage

Provides the best chance of early trial termination of the three methods shown.

Pocock main disadvantage

It spends enough alpha early that the final analysis uses a stricter significance level than the usual $0.05$.

Pocock R=3 final-stage example

With three analyses, the p-value cutoff is $0.0221$; a final $p=0.035$ would fail the sequential plan even if significant at $0.05$ without interim looks.

Haybittle-Peto approach

Uses a very stringent boundary for interim looks and approximately the conventional boundary at the final analysis.

O'Brien-Fleming approach

Uses extremely stringent early boundaries that gradually relax, leaving the final analysis close to the conventional alpha level.

Why Haybittle-Peto and O'Brien-Fleming are attractive

They avoid substantially penalizing the final analysis.

Early-stopping drawback of O'Brien-Fleming/Haybittle-Peto

It is difficult to attain statistical significance early unless the treatment effect is very strong.

Popular group sequential choice

O'Brien-Fleming is popular because it preserves close to the desired alpha at the final analysis.

R=2, O'Brien-Fleming Analysis 1 boundary

$$B=2.782, p=0.0054$$

R=2, Haybittle-Peto Analysis 1 boundary

$$B=3.0, p=0.002$$

R=2, Pocock Analysis 1 boundary

$$B=2.178, p=0.0294$$

R=2, O'Brien-Fleming Analysis 2 boundary

$$B=1.967, p=0.0492$$

R=2, Haybittle-Peto Analysis 2 boundary

$$B=1.960, p=0.0500$$

R=3, O'Brien-Fleming Analysis 1 boundary

$$B=3.438, p=0.0006$$

R=3, Haybittle-Peto Analysis 1 boundary

$$B=3.291, p=0.0010$$

R=3, Pocock Analysis 1 boundary

$$B=2.289, p=0.0221$$

R=3, O'Brien-Fleming Analysis 2 boundary

$$B=2.431, p=0.0151$$

R=3, Haybittle-Peto Analysis 3 boundary

$$B=1.960, p=0.0500$$

R=4, O'Brien-Fleming Analysis 1 boundary

$$B=4.084, p=0.00005$$

R=4, Haybittle-Peto Analysis 1 boundary

$$B=3.291, p=0.00100$$

R=4, Pocock Analysis 1 boundary

$$B=2.361, p=0.0182$$

R=5, O'Brien-Fleming Analysis 1 boundary

$$B=4.555, p=0.000005$$

R=5, O'Brien-Fleming Analysis 5 boundary

$$B=2.037, p=0.0417$$

Example 10.1 disease

Non-Hodgkin's lymphoma.

Example 10.1 treatments

Cytoxan-prednisone (CP) versus cytoxan-vincristine-prednisone (CVP).

Example 10.1 primary endpoint

Presence/absence of tumor shrinkage; identified as a surrogate variable.

Example 10.1 sample size

126 patients.

Example 10.1 Pocock alpha cutoff

$0.0158$ at each of the five analyses.

Example 10.1 analysis 5 result

$CP = 23/67$ versus $CVP = 31/59$; result $0.0158 < p < 0.10$, failing to meet the sequential threshold.

Example 10.1 clinical rates

At final analysis, CVP appeared clinically better at $53\%$ success versus $34\%$ for CP.

REMATCH trial

Cited as an example of O'Brien-Fleming use in a clinical trial.

Group sequential drawback: R

The number of scheduled analyses $R$ must be fixed before the trial begins.

Group sequential drawback: Spacing

Traditional group sequential plans require equal spacing between analyses with respect to patient accrual.

Alpha spending function approach

A flexible interim-monitoring framework developed to overcome fixed-number and equal-spacing restrictions.

Information fraction $\tau$

The fraction of the trial's total planned statistical information available at an interim analysis.

$\tau$ for fixed-sample mean comparison

$\tau = \frac{n}{N}$, where $n$ is current sample size and $N$ is target sample size.

$\tau$ for time-to-event trial

$\tau = \frac{d}{D}$, where $d$ is events observed and $D$ is target total events.

Alpha spending function $\alpha(\tau)$

An increasing function describing how much of the total Type I error has been spent by information fraction $\tau$.

$\alpha(0)$ property

Alpha spending at trial start ($\tau=0$) is equivalent to $0$.

$\alpha(1)$ property

Alpha spending at trial end ($\tau=1$) is equal to alpha, the desired overall significance level.

Meaning of 'spending alpha'

Each interim analysis uses or spends part of the total allowable Type I error.

Cumulative alpha interpretation

At the r-th analysis, $\alpha(\tau_r)$ is the probability under $H_0$ that any of the first $r$ analyses has rejected $H_0$.

Critical-value computation

Sequential critical values require numerical integration of the relevant joint distribution.

Compromise spending function levels

Example yielding significance level $0.012$ at interim and $0.04$ at final.

Early stopping bias

Treatment-effect estimates are biased when a trial terminates early; bias is larger the earlier the stop occurs.

Futility assessment

A plan to terminate a trial when results indicate that additional enrollment is unlikely to change the ultimate conclusion.

Curtailed sampling

Another term for stopping early because continuation is unlikely to alter the conclusion.

Unconditional power

Probability, calculated at the start, of obtaining significance under a prespecified alpha and alternative effect.

Conditional power

Probability of rejecting $H_0$ at trial end given current data and an assumption about future outcomes.

Low conditional power interpretation

If conditional power is very small, continuing the trial may be futile.

Coin example: Reject $H_0$ definition

Test for a fair coin with $500$ tosses; reject $H_0$ if heads $X \geq 272$ at $0.025$ significance level.

Conditional power when rejection is guaranteed

Conditional power equals $1$; for example, if $X=272$ after $400$ tosses.

Coin example futility conclusion

If only $200$ heads exist after $400$ tosses, the probability of reaching $272$ is extremely small, making continuation futile.

Clinical scenario: Favorable interim trend

If current data favors rejection, high conditional power suggests the trend is unlikely to disappear.

Clinical scenario: Negative interim trend

If data is consistent with $H_0$, evaluate if a reversal is likely; if not, termination may be considered.

Adaptive design

A design that prespecifies how study features may change in response to observed interim results.

Adaptive change example

Increase sample size or terminate the study based on emerging results, if the rule was prespecified.

Confirmatory adaptive-design requirement

Must maintain statistical validity, with Type I error control being critical.

Dose-finding adaptive emphasis

Assign more participants to treatments with favorable responses; Type I error control is less central than identifying effective doses.

Single-center annual IRB report goal

Addresses whether the study remains safe and whether continuation is appropriate.

IRB report topic: Oversight

Compliance with governmental and institutional oversight.

IRB report topic: Eligibility

Review of eligibility, with low frequency of ineligible patients entering the trial.

IRB report topic: Treatment

Review whether most patients are adhering to the treatment regimen.

IRB report topic: Response

Summary of response.

IRB report topic: Survival

Summary of survival.

IRB report topic: Adverse events

Summary and review of adverse events.

IRB report topic: Safety rules

Safety-monitoring rules, possibly including statistical criteria for safety endpoints.

IRB report topic: Quality assurance

Audit and other quality-assurance reviews.

Multi-center trial

A trial conducted at multiple centers, with one or more clinical investigators at each location.

Multi-center advantage: Enrollment

Larger sample size and faster patient accrual, especially for rare diseases.

Multi-center advantage: External validity

Broader interpretation and generalizability across participants and geographic regions.

Multi-center advantage: Scientific merit

Greater merit through collaboration among experienced clinical scientists.

Multi-center disadvantage: Planning

More complex planning and greater expense.

Multi-center disadvantage: DCC

Need for a data coordinating center to store and monitor data and organize investigators.

Multi-center disadvantage: Leadership

Need for strong leadership and keeping investigators involved/motivated.

NIH DSMB requirement

The NIH requires a DSMB for an NIH-sponsored multi-center clinical trial.

FDA DSMB statement

The FDA does not require a DSMB for every multi-center trial, though many companies use them regularly.

DSMB primary advantage

Protect participant interests and safety while maintaining scientific integrity.

DSMB independence

Should be financially and scientifically independent of study investigators for objective decision-making.

Typical DSMB size

Approximately $3$ to $10$ experts.

DSMB Expertise areas

Medicine, statistics, epidemiology, data management, clinical chemistry, and ethics.

Study investigators on DSMB

None of the study investigators should be DSMB members.

DSMB masking

The DSMB should not be masked to treatment assignment when evaluating the trial.

DSMB reporting line

Reports directly to the trial sponsor (NIH or company) rather than to study investigators.

DSMB question: Baseline

Are treatment groups comparable at baseline?

DSMB question: Protocol

Should the protocol be modified?

DSMB question: Quality

Are the data of sufficient quality?

Major DSMB disadvantage

Some trial-specific expertise may be sacrificed to preserve impartiality.