2 Trial Design

Sound inference: reliant upon sound design and trial conduct $\to$ poor design cannot be accommodated at the analysis stage.

Trialists objectives $\to$ minimise bias and maximising efficiency: controlling random error. Bias usually caused by structural design defect.

Key principles/issues include:

• •

  bias $\to$ systematic design defect

• •

  replication $\to$ uncertainty/precision

• •

  control

• •

  randomisation

• •

  treatment blinding/masking

• •

  ethics (WMA Declaration of Helsinki, 2001).

Perhaps natural to begin a clinical investigation by trying a new treatment on a group of individuals and observing the outcome (e.g phase I or II trial).

One might consider measuring a treatment effect using the patients’ change from baseline?

When assessing efficacy, however:

• •

  investigators and patients expectations, enthusiasm and feeling of inclusion may in turn may affect their judgement, feeling of well being and outcome measurement

• •

  time-trends may be implicated

• •

  regression to the mean could be confound results

• •

  one needs a yard-stick

A control group provides a yard-stick as to what might have happened in the absence of the experimental treatment. Controls may receive standard or a placebo treatment.

Many early clinical trials which were, so-called, open and uncontrolled have been subsequently found to have yielded spurious results and the treatments were later abandoned, (Miao, 1977, Silverman, 1985).

Definitive assessment should be in relation to the effectiveness of an alternative treatment (controlled).

Controlled trials $\to$ patients allocated to receive the experimental or control treatment.

Controls should be concurrent, not historical.

The randomised double-blind controlled trial represents the gold-standard.

Treatment effect is based upon the difference between experimental and control group outcomes.

Further reading ICH E10 focuses upon issues surrounding choice of control group.

randomisation

does not mean haphazard, but rather, that treatments are assigned according to chance, often equal

random allocation

with a known chance of receiving each treatment, but not predictable!

Aim: to yield treatment groups which are indeed comparable in terms of extraneous factors.

Purpose of randomisation:

• •

  avoid bias due to differences in clinical and demographic characteristics

• •

  support the independence assumption underlying many statistical procedures

• •

  Differences between random samples $\to$ known sampling distributions, control of type I error, $\alpha$.

Justification - equipoise? Uncertainty exists with regards to the superiority of the treatment being compared
Randomisation techniques vary by design.

Further reading

• •

  Altman DG, Bland JM (1999) BMJ 318: 1209

• •

  Altman DG, Bland JM (1999) BMJ 319: 703-704

Analogous to “tossing a coin”: independent random treatment allocation with fixed probability, say, 0.5.

Random number generators $[0,1)$, sealed envelopes and treatment packs prepared in advance to aid blinding.

Multi-centre trials often use a central computer-based randomisation service.

Advantages:

• •

  easy to implement

• •

  analysis via standard statistical methods (t-test, Chi-squared test etc.).

Disadvantages:

• •

  likely to obtain unequal numbers in treatment groups (particularly in small samples)

• •

  does not ensure balance over prognostic factors, for example, age, and disease severity

• •

  lack of adaptability $\to$ can give rise to ethical issues.

Also called “restricted randomisation” the purpose is to balanced groups with regard to numbers of participants.

Example: two treatments A and B, block size 4:

• •

  AABB, ABAB, ABBA, BBAA, BABA, BAAB

• •

  $\left(\genfrac{}{}{0pt}{}{4}{2}\right)=6$ different blocks.

Procedure:

• •

  choose blocks at random from those available (e.g from the six above if two treatments and a block size of 4)

• •

  assign patients accordingly as they present.

Block size

multiple of number of treatments; fixed/randomly varying, keep short to prevent incomplete blocks.

Advantages

balance. Maximum difference is $b/2$ for blocks of size $b$ and two treatments.

Disadvantages

assignment may become known if blocking factor revealed and statistical methods often presume simple randomisation.

There is a problem in small samples (especially), simple or block randomisation might lead to imbalance with regard to an important factor. The purpose of stratified randomisation is to aid between group comparability (balance) over important characteristics.

Procedure

separate block randomisation for each stratum.

Example

stratification by menopausal status in breast cancer trial, by centre in multi centre trial.

Advantages

increases efficiency, reduces potential bias.

Disadvantages

number of strata may limit usefulness, complicates proceedings.

Analysis may be conducted to include stratification variables in the fitted model.

see handout for age group and gender example.

Non-random treatment allocation: allocation probabilities adjusted according to patient imbalance.

The purpose is to balance between treatment groups with regard to important prognostic factors. The procedure uses simple randomisation when groups are balanced: when imbalanced allocate next patient to treatment so that imbalance is minimised (using current value of minimisation score).
blinding can be an issue.

Alternatively consider weighted randomisation: weighting in favour of the treatment which would minimise imbalance.

Analysis: include prognostic factor in the model.

Example: cancer trials (prognostic factors), Senn, 1977 asthma trial.

Hand-out example: balance over gender and steroid usage.

Further reading:

• •

  Treasure T, MacRae KD (1998) BMJ 317: 362–363

• •

  Altman (1991), Section 15.2.3

• •

  Piantadosi (1997), Section 9.4.2.

Sometimes called “masking”. The purpose is to avoid bias due to differences in treatment or outcome assessment (conscious or subconscious). We consider several levels of blinding

single-blind

patient does not know which treatment is being received

double-blind

neither patient nor physician knows the treatment allocation (Gold-Standard)

triple-blind

as double-blind also the monitoring group and data analyst do not know which group receives test and which control treatment

open

all are knowledgeable about the treatment allocation.

There are several methods to achieve this

• •

  placebo control group

• •

  double-dummy method (for comparison of two active treatments with different appearance/formulation)

• •

  ‘sham’ treatment arm.

There is a problem …blinding is not always easy or indeed possible but one can:

• •

  blind assessment / analysis

• •

  conceal the treatment allocation until the patient is entered into trial.

Blinded trials are more complex: an independent monitoring committee is needed to deal with ethical aspects and trial termination. Ethically, blinding can be broken (and should be) for a patient when knowledge of treatment allocation is essential (e.g. Serious Adverse Event, SAE).

Attention: some studies are claimed to be double-blind, but in fact are not! (e.g. revealing of treatment by specific side effects).

Further reading: Day SJ, Altman DG (2000). BMJ 321:504.

Informed consent

can it be obtained?

Randomisation

equipoise assumption.

Choice of control therapy

active or placebo?

Ethics committees

All research on human subjects must be subject to review by a research ethics committee (REC).

• •

  World Medical Association: Declaration of Helsinki

  http://www.wma.net

• •

  Governance Arrangements for NHS Research Ethics Committees

  http://www.corec.org.uk

Further reading: Altman (1991), Section 15.2.9 and references given there, also ICH E10.

A complication in clincal trials is non-adherence to the protocol. For example, patients may switch treatments, fail to complete treatment, miss scheduled visits, withdraw etc.

The choice of analysis set is an important consideration:

• •

  intention-to-treat (ITT): analyse as randomised (effectiveness)

• •

  per protocol: analysys based upon adherers only (true biological effect: efficacy).

Analysis - typically based upon ITT principle: conservative/pragmatic in superiority.

Note in trials of equivalence/non-inferiority anti-conservative and both analysis sets should be considered.