Gynaecology And Infertility

A potential problem that can occur with randomised controlled trials is where the intervention is not targeted at individual patients, but at groups of patients. This can happen where the intervention is an information package for the management of menorrhagia in primary care9 or a clinical guideline for the management of infertility.10 In these studies, randomising patients to receive management using the information package or clinical guidelines would have introduced contamination, since GPs would have been expected to manage both study (information leaflet, clinical guidelines) and control patients. Potentially this could underestimate the true effectiveness of the intervention. Therefore, clusters of patients (e.g. general practices) were randomly allocated to receive intervention (e.g. information leaflets, clinical guidelines) or control. Cluster randomisation should only be carried out when there is a strong justification for doing so.

The primary implication of cluster randomised trials is that the measurements on individuals are not statistically independent of one another; that is measurements from individuals within the same cluster will be correlated to one another. This has implications in the design (e.g. sample size), conduct (e.g. informed consent), analysis and reporting. Cluster randomised trials should adjust for this clustering when determining the number of patients required. The sample size that would be required if patients were to be randomised must be inflated by a factor which takes into account the extent of the clustering and the size of the cluster.11 The extent of the correlation is measured by the intra-cluster correlation coefficient (ICC)12 and researchers are required to have some indication of this, in order that the study can be adequately powered. Studies that fail to adequately inflate the sample size will suffer from a type II error.

Similarly, the correlated responses obtained from each cluster have an implication for the statistical analysis, since standard statistical tests (e.g. t-test) assume that observations are independent of one another. There are a number of approaches to analysing cluster randomised trials and these are detailed elsewhere.13 Failure to account for the correlated responses in the analyses will result in an increased type I error.

Clustering of outcomes can also occur in infertility trials where alternative treatments are being compared. For example, in randomised controlled trials comparing IVF with ICSI the unit of allocation varies between patients,14 oocytes15 and cycles.16 Often, outcomes such as implantation rate and fertilisation rate are considered. These are both expressed as percentages out of the total number of oocytes retrieved. Hence, in trials that randomise patients (couples) or cycles and report implantation or fertilisation rates, there will be clustering of the outcome since oocytes are clustered within patients or cycles.1416 Hence, in these studies adjustment should be made in the analysis to adjust for the correlated outcomes assessed (on each oocyte) within patients (or cycles). In trials that randomise by patients and report fertilisation of implantation rates, some adjustment is required. However, for outcomes such as live birth rate or pregnancy rate no adjustment is required since the percentages are expressed out of the total number of patients randomised. Bhattacharya et al.16 randomised cycles and reported implantation rates per transferred embryo. However, they noted that the difference in implantation rates was likely to be wider than that reported due to failure to adjust for the clustering of embryos/oocytes transferred to each woman. Studies where oocytes have been randomised have no clustering implications since oocytes retrieved from the same women are randomly allocated to receive ICSI or IVF.

When conducting randomised controlled trials in infertility, consideration should be given to the unit of randomisation and the outcome measures to be applied. When there is implicit clustering in the data, the statistical analysis should account for this using the methods described above.13


These are controlled experimental studies where treatment allocation is performed on the basis of patient unit numbers or days of the week when the patients are recruited. Although this design of treatment allocation affords an element of chance, it cannot be considered to be genuine randomisation. This type of design may still appeal to those involved in laboratory trials involving incubation or cryopreservation of human embryos. In these cases, it may be easier and cheaper to use a certain protocol for all embryos on alternate days or alternate weeks rather than change the protocol or a freezer setting for each embryo or each woman. The consequent loss of allocation concealment will lead to serious inclusion bias as some patients may be deliberately excluded. This, is turn, can exaggerate treatment effects.


A potential problem in some randomised trials arises when patients or their clinicians refuse to be randomised on grounds of strong treatment preferences. Exclusion of these patients may affect the generalisability of the results as participants may not be representative. Yet recruitment of these patients may introduce substantial bias especially when it is impossible to blind them. In addition, compliance and satisfaction may be higher with the preferred intervention.17

This is particularly so when the 'new' treatment is only available within the context of the trial or when, as in trials in unexplained infertility, one of the arms comprises a 'no treatment' or 'expectant management' group. Dissatisfaction with the allocation may lead to differential compliance and follow-up resulting in groups which cannot then be assumed to be similar. The outcome measures could also be affected by how satisfied patients are with their allocated treatment. The effect of patient preference on outcome would depend to a great extent on the specific outcomes being assessed. If the principal outcome is death or live birth, then the effect of patient preferences is likely to be small. If the principal outcome is satisfaction with care, then the effect of patient preference is large.18 Under such circumstances the conventional randomised trial will underestimate the relationship between the intervention and the outcome, i.e. show the minimum effect size. Conversely a comparison between two groups of patients who have chosen their treatment and thus optimised their treatment choice will be considered to represent the maximum effect size. An intervention in question will have an effect size between the minimum and maximum as derived from the randomised and the preference part of a partially randomised patient preference trial.18

To deal with patient preferences within a trial, the use of a partially randomised patient preference (PRPP) trial has been suggested.19 Patients with strong preferences are allowed their desired treatment. Those without such views are subjected to randomisation. For example in a trial of medical and surgical termination of pregnancy we end up with four groups-randomised to medical, randomised to surgical, prefer medical and prefer surgical.

Potential disadvantages with PRPP trials include effects of the trial size. The size of a total PRPP cohort will need to be much larger than for a conventional randomised controlled trial. As already mentioned, the size of the randomised cohort needs to be the same as in a conventional trial. In addition, the number of patients in the non-randomised preference arms needs to be of equivalent size. The numbers quickly add up to generate a total sample size double that for a conventional trial. This has the predictable effect of adding to the cost and duration of the trial. Entry of a disproportionate number of patients into either the randomised or the preference arms is also a problem, as the trial will not be completed unless the appropriate number have been recruited into the two components of the trial. The situation may be further complicated by patients favouring one treatment over another, making comparison of the two groups in the preference arm more difficult.

A further problem with this approach lies in the analysis. Any comparison using the non-randomised groups is unreliable because of unknown and uncontrolled confounders. Patient preference designs have been used in trials of termination of pregnancy20'21 and menorrhagia.22 The evidence to support the use of PRPP trials compared with conventional randomised trials is limited. A randomised comparison of the two strategies by Cooper et at.22 suggested that the extra cost and complexity were not justified in the context of medical versus surgical treatment of menorrhagia.

A conventional randomised trial could address the effect that patient preference has on outcome by recording this information before allocation.23 This would allow resources to be concentrated on recruiting as many patients as possible into the randomised comparison group but would allow stratification of the results by initial preference.


Often in reproductive health care the aim is to show that one treatment is as effective (equivalence), or no less effective (non-inferior), as another. The methodology for equivalence trials differs to that of superiority trials in design, analysis and interpretation. In designing equivalence trials, attention must be given to defining an equivalence margin. This is the difference in effect that would be deemed to be 'clinically insignificant'.24 In comparison with superiority trials, larger sample sizes are needed to demonstrate equivalence. In the analysis of equivalence trials, conventional statistical testing has little relevance and interpretation of results should be conducted though use of confidence intervals in relation to the predefined equivalence margin.25 Statistical significance is demonstrated if the upper and lower limits of the 95% confidence interval do not cross the equivalence margin.25 In superiority trials, the most conservative analysis is by intention to treat (ITT). In an equivalence trial, however, a 'per protocol' (PP) analysis is usually considered statistically more conservative. This is because an ITT analysis may blur the comparison between groups and lead to an increased chance of declaring the two treatments as equivalent when they are not. The decision about which should be the primary form of analysis (ITT or PP) in an equivalence study is not straightforward.26 It depends on the particular characteristics of the study, including the definitions adopted for the ITT and PP analyses and the risk of bias.27

100 Pregnancy Tips

100 Pregnancy Tips

Prior to planning pregnancy, you should learn more about the things involved in getting pregnant. It involves carrying a baby inside you for nine months, caring for a child for a number of years, and many more. Consider these things, so that you can properly assess if you are ready for pregnancy. Get all these very important tips about pregnancy that you need to know.

Get My Free Ebook

Post a comment