Design Issues: Policy Trials Professor David Torgerson Director, York Trials Unit

Policy Trials These are similar to ordinary RCTs. We need to undertaken trials that are unbiased and cost effective. »Avoiding selection bias at randomisation; »Cluster trial design; »Efficient design.

Selection Bias Selection bias can occur in non- randomised studies when group selection is related to a known or unknown prognostic variable. If the variable is either unknown or imperfectly measured then it is not possible to control for this confound and the observed effect may be biased.

Effects of selection bias Observational data on hormone replacement therapy consistently shows that this reduces cardiovascular disease, stroke, dementia. Trial evidence shows the opposite. BECAUSE women taking HRT tended to be different and at lower risk from these problems than women not taking HRT.

Randomisation Randomisation (or a similar technique, such as minimisation) removes selection bias across a ‘population’ of RCTs by ensuring all variables that may affect outcome are balanced across treatment groups at baseline. Other techniques may allow the introduction of selection bias.

Subversion Subversion of the allocation mechanism introduces selection bias. This occurs when the next allocation can be predicted and participants are then selected to match a desired allocation rather than having the allocation assigned at random.

Subversion - evidence Schulz [1] has described incidents of researchers subverting allocation by looking at sealed envelopes through x-ray lights. Researchers have confessed to breaking open filing cabinets to obtain the randomisation code. In a survey [2] of 25 researchers 4 admitted to keeping ‘a log’ of previous allocations to try and predict future allocations. Case study of a subverted trial. Schulz JAMA 1995;274:1456. Brown et al. Stats in Medicine, 2005,24:3715.

Mean ages of groups

Example of Subversion

Recent Blocked Trial “This was a block randomised study (four patients to each block) with separate randomisation at each of the three centres. Blocks of four cards were produced, each containing two cards marked with "nurse" and two marked with "house officer." Each card was placed into an opaque envelope and the envelope sealed. The block was shuffled and, after shuffling, was placed in a box.” Kinley et al., BMJ 325:1323.

What is wrong here? SouthamptonSheffieldDoncaster DoctorNurseDoctorNurseDoctorNurse Kinley et al., BMJ 325:1323.

Problem? If block randomisation of 4 were used then each centre should not be different by more than 2 patients in terms of group sizes. Two centres had a numerical disparity of 11. Either blocks of 4 were not used or the sequence was not followed.

Evidence from a systematic review. In a systematic review of the use of calcium supplements to enhance weight loss Trowman et al found a significant relationship between calcium use and reductions in body weight. HOWEVER, examination of baseline characteristics found that people with lower body weights tended to be allocated to the calcium group. In no single trial was this difference significant but in a meta-analysis of baseline weights the difference was highly significant (p = 0.006).

Meta-analysis of baseline body weight. Trowman et al. The impact of baseline imbalances should be considered in systematic reviews: a methodological case study. Journal of Clinical Epidemiology 2007;60:

Why subversion? “To provide best patient care…” “He fancied her! She was pretty!” “Individual was putting younger fitter individuals into the intervention, they were trying to improve the results” “Prefer to do certain procedures” “Researcher over rode the random allocation, thought there should be same numbers in each group” Hewitt et al. J Clin Epidemiol 2009;62:261-69

Concealment: Recommendations Allocation sequence must be independently generated and kept secret from the people who are enrolling participants. A secure method of giving allocation to the recruiters must be developed, coin tossing, or opaque envelopes are inadequate.

Cluster trials In most drug RCTs people are randomised as individuals to treatment. However, many policy trials need to randomise intact groups (e.g., schools, prisons, hospitals; periods of time). These trials can have problems because often difficult to conceal allocation of cluster from person recruiting participants.

Recruitment Bias A key issue is individual participant recruitment into cluster trials. There are a number of ways where biased participant recruitment can occur, which can lead to baseline imbalances in important prognostic factors. In an individually randomised trial we avoid recruitment bias by concealing the random allocation from the potential participant and researcher until AFTER they have consented to be in the trial and have been recruited. In cluster trials sometimes this is not possible.

Identification Problems For example, in a cluster trial of back pain treatments equal number of patients with same severity of back pain will be present in both clusters. The problem lies in how to identify such patients to include them in the interventions. Unless one is very careful different numbers and types of patient can be selected.

UK BEAM Trial The UKBEAM pilot study used a cluster design. Eligible patients were identified by GPs for trial inclusion. GP practices were randomised to usual care or extra training. The ‘primary care team’ were trained to deliver ‘active’ management of backpain.

UK BEAM participant recruitment P = 0.06 P = 0.01

UKBEAM pilot study.

Another musculoskeletal trial In 2002 I joined a steering group for a trial of training GPs to identify and treat a common musculoskeletal condition. GPs were to recruit the participants. With the BEAM experience we KNOW what WILL happen. GPs WILL recruit more patients if they are trained. Did they?

Why would you do that? “You learn nothing by being kicked by the same mule twice”.

Consent Bias This occurs when consent to take part in the trial occurs AFTER randomisation. Another danger in Cluster trials. For example, Graham et al, randomised schools to a teaching package for emergency contraception. More children took part in the intervention than the control. Graham et al. BMJ 2002;324:1179.

Consent bias? Knowledge of emergency contraception at baseline

Consent Bias? Because more children consented in the intervention group we would expect their knowledge to be less (as we include children less likely to know). Conversely we get a volunteer or consent effect with the intervention group only those most knowledgeable agreeing to take part.

Trial Consent Problems Even when it is possible to identify all eligible members of a cluster some may not consent to take part in the trial. If there is differential consent, in particular, this can lead to selection bias again. To prevent this we must use the same approach as we do for individually randomised trials: recruit participants on the basis that the can get either intervention and then randomise.

Hip Protector Trial Kannus. N Eng J Med 2000;343:1506. At this point trial is balanced for all co-variates

Hip Protector Trial Selection Bias

Dilution effects In a cluster trial of accident prevention among young children 25% of parents in the experimental arm did not receive the intervention. Clearly this will reduce the power of that trial AND dilute any likely ‘treatment’ effect. Kendrick et al. BMJ 1999;318:980.

Review of Cluster Trials Because of the ‘BEAM’ problem we decided to undertake a methodological review of cluster trials. We identified all cluster trials published in the BMJ, Lancet, NEJM since Puffer et al. BMJ 2003;327:785.

Results We identified 36 relevant trials. ONLY 13 had identified participants prior to randomisation. Of the 23 not identifying participants a priori 7 showed evidence of differential recruitment or consent. Other biases included differential of inclusion criteria or attrition. In total 14 (39%) showed evidence of bias.

Underestimate of problem Only in 5 papers did authors alert reader to possible problem. Subsequently one of the trials that ‘looked’ OK was published elsewhere where recruitment bias was admitted to have occurred.

Baseline Characteristics InterventionControl Live in Flat40%23%P < Married67%59%P = 0.07 Access to help 80%70%P = 0.04 P values adjusted for clustering. Jordhoy Palliative Medicine :43-49.

Cluster Trials: Should I do one? Yes, BUT do them properly. Is it possible to avoid doing them and do an individually randomised trial?

Contamination An important justification for their use is SUPPOSED ‘contamination’ between participants allocated to the intervention with people allocated to the control.

Spurious Contamination? Trial proposal to cluster randomise practices for a breast feeding study – new mothers might talk to each other! Trial for reducing cardiac risk factors patients again might talk to each other. Trial for removing allergens from homes of asthmatic children.

Patient level contamination In a trial of counselling adults to reduce their risk of cardiovascular disease general practices were randomised to avoid contamination of control participants by intervention patients. Steptoe. BMJ 1999;319:943.

Counselling Trial Steptoe et al, wanted to detect a 9% reduction in smoking prevalence with a health promotion intervention. They needed 2000 participants (rather than 1282) because of clustering. If they had randomised 2000 individuals this would have been able to detect a 7% reduction allowing for a 20% CONTAMINATION. Steptoe. BMJ 1999;319:943.

Accepting Contamination We should accept some contamination and deal with it through individual randomisation and by boosting the sample size rather than going for cluster randomisation Torgerson BMJ 2001;322:355.

What about dilution bias? If, in the presence of contamination, we use individual allocation we might observe a difference that is statistically significant but is not clinically or economically significant. Dilution has biased the estimate towards the mean. If we can measure contamination we can deal with this using ‘instrumental’ or CACE analysis. Hewitt et al. Canadian Medical Association Journal 2006;175:347-48

Cluster Trials Can cluster trials give different results? All things being equal this shouldn’t happen (except for a more imprecise estimate). BUT because of the greater potential for selection bias cluster trials MAY give the ‘wrong’ answer.

An example. There are 14 RCTs of hip protectors for the prevention of hip fracture. Nine RCTs are individually randomised trials, whilst 5 are cluster trials (e.g., hospital ward, nursing home). Cluster trials, without exception show a benefit of hip protectors.

Hip Protector Trials Individual RCTSCluster RCTs 1.19 (0.8 to 1.7) (0.5 to 1.7)0.53* 0.93 (0.5 to 1.7) (0.4 to 3.0) (0.1 to 1.4) (0.0 to 1.6)All Cluster trials, bar *, significant, No individual trial was significant 1.49 (0.3 to 7.1) 3.03 (0.6 to 14.8)

Hip Protector Trials: Cluster vs Individually Randomised.

Age differences between ‘good’ cluster and poor cluster trials. Data from Puffer et al.

Cluster trials Are essential to evaluate robustly some policy interventions. Need care in their design especially if there is sequential recruitment to the study. Good evidence shows that a significant proportion of health care cluster trials are poorly undertaken.

Cluster Trials- What Should We Do? Identify ALL eligible people if possible BEFORE randomisation ALWAYS use Intention To Treat analysis Blind the person applying inclusion/exclusion criteria. Blind follow-up/data collection. INCREASE sample size not only for cluster effects but also because of treatment refusal

Questions?

Unequal allocation Most trials randomly allocate into groups that have equal sizes. Given a FIXED total sample size then this approach USUALLY gives the most statistical power. BUT sample sizes may not be fixed but research costs are. So we can put more into one group than another. For example, in a trial of improving female representation in local councils in India, 1/3 rd of councils were randomly allocated to be reserved for women, whilst the remaining 2/3rds could be either men or women (virtually all men). Econometrica 72, 1409

Allocation ratios Often sample sizes are constrained through cost reasons. It is nearly always true that statistical power is maximised if equal numbers are allocated to treatment groups but when the sample size isn’t restricted but the budget is, this is not so. For example, Brannan and John in an RCT of telephoning or canvassing in the UK 2005 general election were constrained by their budget. However, they could have got more power for their study if they had allocated more to the cheaper groups. Brannan and John IPEG report.

Unequal allocation can IMPROVE statistical power. Panagopoulos and Green undertook a field experiment of radio avertising on electoral competition. Resource restrictions meant they could only allocate to 14 cheaper sites out of 281 eligible municipalities. Consequently 151 ‘cheap’ localities were identified and a matched sample of 28 were identified. Could it have been done differently and gain more power? Panagopoulos and Green American Journal of Political Science, 2008,

Unequal allocation Assuming a mean difference of 10% (50 in control and 60 in intervention) a sample of 28 would have 75% power (assuming no covariates or matching). But unequal allocation could have produced more power, if out of the 151 cheap localities 14 had been randomised to intervention and 137 to control there would have been 95% power to detect the same difference.

Policy Trials Have many characteristics that are similar to non-policy trials. RCTs are the most robust method, if undertaken properly, to evaluate policy innovations. We should do more of these – but care needs to be taken over their design.