How to write a retrospective observational study

Retrospective observational studies are non-randomised, non-interventional analyses of existing data relating to patients, care received and outcomes. They have several advantages over prospective randomised studies, but their main disadvantage is the ever present influence of bias and confounding effects on any derived associations (Table 1). As such, most retrospective observational studies are used to generate a hypothesis rather than demonstrate causality. That said, a clinically applicable and well-performed retrospective observational study on the right topic will always be of interest to editors, reviewers, readers and clinicians alike. This article will outline areas that will enable the authors to maximise their chance of publication acceptance. For any study undertaken in a healthcare setting, it is important that the appropriate permissions are granted [1]. The Medical Research Council and NHS Health Research Authority provide a decision tool to help determine whether your work requires research ethics committee approval (http://www.hra-decisiontools.org.uk/ethics). For a retrospective observational study, the key question is whether it can produce generalisable or transferable findings. This is difficult, as whilst the findings may tell us about trends and associations they will not prove causality and might best be used to generate a hypothesis. Therefore, most retrospective observational studies may be considered as audit (where there are clear predetermined standards) or service evaluation (where standards are absent). Service evaluations are designed solely to define or judge current care. There are no accepted standards associated with the intervention or outcomes chosen, and the intervention is one in current use according to guidelines, consensus, professional standards and patient preference. Data are collected routinely and no additional prospective data are sought. There is currently no requirement to prospectively register a study protocol for a retrospective study, which makes it difficult to uncover various poor research practices [2]. That said, the host Trust or organisation may wish to review your protocol before data extraction proceeds. In the UK, it is likely that if the proper processes are followed to ensure anonymity of patient data, the need for consent will be waived. In other countries, full ethical review may be required as it would be for a large, randomised study. Just because formal ethical approval is not required does not negate material ethical issues with a retrospective observational study. Ethical approval can instead be sought by the host Trust or a local university, which increases the probability of having the paper accepted for publication. The most important aspect of the study design is ensuring that you ask an appropriate clinical question on an interesting topic. The advantage of retrospective observational studies is that they take less time to execute, which means that research questions can better elucidate current clinical unknowns. The study might look to estimate the incidence of a rare event or describe the clinical characteristics of a specific population. It could also look for differences between groups with different exposures as a formal case control or cohort study. It would be possible to go even further and use an outcome metric that would be important to patients, such as days alive and at home at 90 postoperative days [3]. The authors should consider and describe the setting, location and dates of recruitment. There is a balance between getting sufficient patients to make any analysis meaningful whilst avoiding historical data that are not representative of current practice; this is a form of temporal bias. Power calculations are not appropriate for these studies and working out how many participants to include can be challenging. Regardless, the method used to determine the sample size must be described in detail. Eligibility criteria should be described clearly as well as the sources and methods of participant selection. Outcomes, exposures, potential confounders and colliders should be thought about and defined clearly. The source of data should be considered, as well as its completeness and accuracy. Sources of bias and the methods used to address these should be listed and justified (Table 2). To include a large sample of patients from a relatively short and clinically relevant time period, large national databases, such as Hospital Episode Statistics, can be utilised. These data are collected to ensure hospitals receive payment for the care they deliver but secondary use of anonymised datasets is permitted. Insights into the safety of day-case paediatric tonsillectomy were gained from analysis of this dataset, which highlighted national variations in practice and suggested that planned overnight stays following tonsillectomy are often unwarranted [4]. Another example of a high-quality retrospective study looked at outcomes in 887,495 patients admitted to hospital between 2010 and 2019 treated for five acute surgical conditions [5]. The study showed that the average number of days alive and out of hospital at 90 days was similar for emergency and non-emergency surgery strategies. Other sources of data that retrospective studies analyse commonly include: National Hip Fracture Database; Danish Anaesthesia Database; National Pain Database; Intensive Care National Audit and Research Centre (ICNARC) database; and Trauma Audit Research Network (TARN) dataset. All statistical methods and software packages used must be described in detail. In any observational work, there will always be measured and unmeasured confounders [6]. Diagrams between variables can help identify decisions about which confounders are important and arrive at a small set of variables to adjust for. This is where sample size is important too, as selecting many confounders for a small dataset will yield unreliable results. When the event is rare, propensity matching may be useful when making comparisons between groups [7]. However, the means used to correct for confounding are wide-ranging and can vary from simple multivariable corrections to more sophisticated methods, such as instrument variable analysis (which is the mirror opposite of propensity scoring). Data-driven confounder selection is common, but this may lead to mislabelling mediators or colliders as confounders as well as exclusion of important but weak confounders. Residual confounding will nearly always be present, despite all the complex statistical corrections and manipulations that can be performed. This can be exacerbated by a failure to consider how variables interact, incorrect modelling and categorising or dichotomising continuous data. Sensitivity analyses can be used to assess how robust an association is to potential unmeasured or uncontrolled confounding [8]. Yet, the most important aspect is the research team remaining reflexive about the limitations of the methods used, and not overinterpreting derived associations. Once the methods have been defined and refined, and following data collection and analysis, writing up your project is the final challenge. The authors' task is to convince the editors and reviewers that their study is worthy of publication in a reputable journal. The best papers are concise and easy to follow. For a retrospective observational study, the strengthening the reporting of observational studies in epidemiology (STROBE) statement and checklist should be followed closely [9]. A participant flow diagram helps to visually display the numbers of patients at each stage of the study. Where individual patients are excluded or their data lost, reasons for this must be given. The discussion section provides an opportunity to contextualise the results and discuss their limitations. Interpretation must be done cautiously, especially where confounding and bias are likely. For example, an overinterpreted wide confidence interval for an association will likely result in rejection, as will a very marginal yet statistically significant result which lacks clinical significance. The easiest way to gauge what is likely to be published is by looking at previous similar studies and considering why they were accepted. When done well, retrospective observational studies can have reasonable clinical impact if they are novel, interesting and educational (Table 3). These are one of the most common types of original article submitted to Anaesthesia and yet only a small fraction are accepted for publication; this is because they rarely influence a change in clinical practice and might be seen as non-experimental, secondary research (Fig. 1). Notwithstanding this, the results can be still useful, as the barriers to completion are relatively surmountable compared with primary prospective studies, and conclusions can be generated in a much shorter timeframe. The chief requirements for retrospective observational studies are that they are: clinically relevant; interesting to readers; simply written; easy to follow; not overanalysed with complicated statistical corrections; and reflexive regarding the inherent difficulties demonstrating causality. Retrospective observational studies have an important place in clinical research and will continue to form a key part of the evidence base in peri-operative medicine. MC is an Editor of Anaesthesia. No other competing interests declared.