General considerations in assessing risk of bias in non-randomized studies

Reporting of randomized trials is relatively straightforward and, increasingly, guided by the CONSORT statement (Moher 2001). A similar consensus statement, STROBE, for the reporting of observational epidemiological studies has been developed, although much more recently (Vandenbroucke 2007). Therefore, the quality of reporting of information required to assess the risk of bias is likely to be less good for NRS.  This is likely to hinder any assessment of risk of bias.


A protocol is a tool to protect against bias; when registered in advance of a study starting, it proves that aspects of study design and analysis were considered in advance of starting to recruit, and that data definitions and methods for standardizing data collection were defined. Because of the need for research ethics approval, all randomized trials must have a protocol, even if protocols vary in their quality and the items that they specify; many randomized trials, particularly those sponsored by industry, also have detailed study manuals. Historically, researchers have not had to obtain research ethics approval for many NRS, and primary NRS rarely report whether the methods are based on a protocol. Therefore, the protection offered by a protocol often does not exist for NRS. The implications of not having a protocol have not been researched. However, it means, for example, that there is no constraint on the tendency of researchers to ‘cherry-pick’ outcomes, subgroups and analyses to report, which happens to a greater or lesser extent even in randomized trials where protocols exist (Chan 2004).


In common with randomized trials, dimensions of bias to be assessed include selection bias (concerning comparability of groups, confounding and adjustment), performance bias (concerning the fidelity of the interventions, and quality of the information regarding who received what interventions, including blinding of participants and healthcare providers), detection bias (concerning unbiased and correct assessment of outcome, including blinding of assessors), attrition bias (concerning completeness of sample, follow-up and data) and reporting bias (concerning publication biases and selective reporting of results). Assessment of risk of bias in randomized trials has developed by identifying the design features which are used to prevent each of these dimensions, and noting whether each trial fulfils the requirements. Risk of bias assessments for NRS should proceed in the same way, with pre-specification of the features to be assessed in the protocol, recording what happened in the study, and a judgement of whether this was adequate, inadequate or unclear as a method to avoid risk of this particular bias. Determining these features is likely to require expert input from an epidemiologist, and will depend in part on the clinical question. Particular care should be given to the assessment of confounding (see Section


The reason for careful attention to the design features of primary studies (such as how participants were allocated to groups, or which parts of the study were prospective) rather than design labels (such as ‘cohort’ or ‘cross-sectional’) is because it is hypothesized that the risk of bias is influenced by the specific features of a study rather than a broad categorization of the approach taken. Furthermore, terms such as ‘cohort’ and ‘cross-sectional’ are ambiguous and cover a diverse range of specific study designs. No empirically-derived list is available of study design features that are relevant to the risk of bias, although a shortlist can be constructed from evidence and theory about the risk of bias in aetiological studies and randomized trials (see Section 13.2.2 and 13.4.2).


Because of the diversity of NRS, different methods may be needed to assess NRS with different design features. One important distinction is between studies in which allocation to groups is by outcome (e.g. case-control studies) and studies in which allocation to groups is more directly related to interventions. In the former type of study, it is the exposure of interest, rather than the outcome, that is most susceptible to bias; review authors need to ask whether researchers assessing the exposure were masked to whether participants had experienced the outcome or not (i.e. were cases or controls). Case-control studies are well suited to investigating associations between rare outcomes and multiple exposures, so may have an important role in generating evidence about the potential adverse effects and unintended beneficial effects of interventions. They have also been used to evaluate large-scale public health interventions such as accident prevention and screening (MacLehose 2000), which are difficult or expensive to evaluate by randomized trials. However, review authors should familiarize themselves with epidemiological considerations that particularly apply to such studies (Rothman 1986). Note that some analyses of patient registries also have similarities with case-control studies: for example, if the entire database is divided into groups of patients who have or have not experienced a particular outcome and exposures associated with the outcome are investigated. Review authors require a deeper knowledge of epidemiology when assessing the risk of bias in NRS, compared with randomized trials.