-- David Sackett, 1996
Note that this approach gives equal emphasis to:
This glossary provides definitiond and some explanation of commonly used evidence-based medicine terms. We are always keen to remain useful, so please don't hesitate to contact us if there are terms you would like us to add.
The probability that an individual will experience the specified outcome during a specified period. It lies in the range 0 to 1, or is expressed as a percentage. By contrast with common usage, the word "risk" may refer to adverse events (such as myocardial infarction) or desirable events (such as cure).
The absolute difference in risk between the experimental and control groups in a trial. It is used when the risk in the experimental group exceeds the risk in the control group, and is calculated by subtracting the AR in the control group from the AR in the experimental group. This figure does not give any idea of the proportional increase between the two groups: for this, relative risk (RR) is needed.
The absolute difference in risk between the experimental and control groups in a trial. It is used when the risk in the control group exceeds the risk in the experimental group, and is calculated by subtracting the AR in the experimental group from the AR in the control group. This figure does not give any idea of the proportional reduction between the two groups: for this, relative risk (RR) is needed.
A method used to prevent selection bias by concealing the allocation sequence from those assigning participants to intervention groups. Allocation concealment prevents researchers from (unconsciously or otherwise) influencing which intervention group each participant is assigned to.
The application of the results from clinical trials to individual people. A randomised trial only provides direct evidence of causality within that specific trial. It takes an additional logical step to apply this result to a specific individual. Individual characteristics will affect the outcome for this person.
People involved in making decisions on health care must take relevant individual factors into consideration. To aid informed decision-making about applicability, we provide information on the characteristics of people recruited to trials. See also generalisability.
The risk of the event occurring without the active treatment. Estimated by the baseline risk in the control group. The baseline risk is important for assessing the potential beneficial effects of treatment. People with a higher baseline risk can have a greater potential benefit.
Systematic reviews of RCTs or individual RCTs are the best method for revealing the effects of a therapeutic intervention as their study design mitigates against bias either in favour of or against the treatment.
However, sometimes other evidence is sufficient to assign causality and in this case an RCT would not be ethical. In other cases RCTs are not practical. RCTs are unlikely to adequately answer clinical questions in the following cases:
Systematic deviation of study results from the 'true' results/effects on patients, because of the way(s) in which the study is conducted.
A trial is fully blinded if all the people involved are unaware of the treatment group to which trial participants are allocated until after the interpretation of results. This includes trial participants and everyone involved in administering treatment or recording trial results. Ideally, a trial should test whether people are aware of which group they have been allocated to. This is particularly important if, for example, one of the treatments has a distinctive taste or adverse effects. Unfortunately such testing is rare. The terms single and double blind are common in the literature but are not used consistently. Therefore, we attempt to report specifically who is unaware of treatment allocation.
Randomisation by a pattern to produce the required number of people in each group.
A study design that examines a group of people who have experienced an event (usually an adverse event) and a group of people who have not experienced the same event, and looks at how exposure to suspect (usually noxious) agents differed between the two groups. This type of study design is most useful for trying to ascertain the cause of rare events, such as rare cancers.
Case-control studies can only generate odds ratios (OR) and not relative risk (RR). Case-control studies provide weaker evidence than cohort studies but are more reliable than case series.
Analysis of series of people with the disease (there is no comparison group in case series). Case series provide weaker evidence than case-control studies.
A cluster randomised study is one in which a group of participants are randomised to the same intervention together. Examples of cluster randomisation include allocating together people in the same village, hospital, or school. If the results are then analysed by individuals rather than the group as a whole bias can occur.
The unit of randomisation should be the same as the unit of analysis. Often a cluster randomised trial answers a different question from one randomised by individuals. An intervention at the level of the village or primary care practice may well have a different effect from one at the level of an individual patient. Therefore, trying to compensate by allowing for intra-class correlation coefficients or some other method may not be appropriate.
A non-experimental study design that follows a group of people (a cohort), and then looks at how events differ among people within the group. A study that examines a cohort, which differs in respect to exposure to some suspected risk factor (e.g., smoking), is useful for trying to ascertain whether exposure is likely to cause specified events (e.g., lung cancer). Prospective cohort studies (which track participants forward in time) are more reliable than retrospective cohort studies.
Analysis of data from only those participants who remained at the end of the study. Compare with intention-to-treat analysis, which uses data from all participants who enrolled.
The confidence interval measures the probability that an actual value will fall within a defined range of values. This can be calculated at any level but most common is 95%. The 95% confidence interval (or 95% confidence limits) would include 95% of results from studies of the same size and design in the same population. This is close but not identical to saying that the true size of the effect (never exactly known) has a 95% chance of falling within the confidence interval. If the 95% confidence interval for a relative risk (RR) or an odds ratio (OR) crosses 1, then this is taken as no evidence of an effect. The practical advantages of a confidence interval (rather than a P value) is that they present the range of likely effects.
A trial in which participants are assigned to two or more different treatment groups. The term is often used to refer to controlled trials in which treatment is assigned by a method other than random allocation. When the method of allocation is by random selection, the study is referred to as a randomised controlled trial (RCT). Non-randomised controlled trials are more likely to suffer from bias than RCTs.
In a randomised controlled trial (RCT), controls refer to the participants in its comparison group. They are allocated either to placebo, no treatment, or a standard treatment. It is important that the nature of the control is specified as there may be an interaction between the intervention and control groups that affects the results, for example, if the 'control' has possible therapeutic qualities, such as, saline for eye infections.
A measure of association that indicates the degree to which two variables change together in a linear relationship. It is represented by r, and varies between 1 and +1. When r is +1, there is a perfect positive relationship (when one variable increases, so does the other, and the proportionate difference remains constant). A correlation coefficient of 0 indicates no linear correlation. When r is 1-, there is a perfect negative relationship (when one variable increases the other decreases, or vice versa, and the proportionate difference remains constant). This does not rule out a relationship, it just excludes a linear relationship.
A trial in which participants receive one treatment and have outcomes measured, and then receive an alternative treatment and have outcomes measured again. The order of treatments is randomly assigned. Sometimes a period of no treatment is used before the trial starts and in between the treatments (washout periods) to minimise interference between the treatments (carry over effects). Interpretation of the results from crossover randomised controlled trials (RCTs) can be complex.
A study design that involves surveying a population about an exposure or condition, or both, at one point in time. It can be used for assessing prevalence of a condition in the population.
Cross sectional studies should never be used for assessing causality of a treatment.
Crude summation of the raw data with no weighting (as distinguished from meta-analysis).
A method for measuring disease burden, which aims to quantify in a single figure both the quantity and quality of life lost or gained by a disease, risk factor, or treatment. The DALYs lost or gained are a function of the expected number of years spent in a particular state of health, multiplied by a coefficient determined by the disability experienced in that state (ranging from 0 [optimal health] to 1 [deaths]). Later years are discounted at a rate of 3% per year, and childhood and old age are weighted to count for less.
Refers to the ability to trace a statement from its most condensed form through to the original evidence that supports it. This requires not only the data but also all the methods used in the generation of the condensed form to be explicit and reproducible. We see it as an important component of the quality of evidence-based publications.
In the medical literature, effect size is used to refer to a variety of measures of treatment effect.
The term effect size is often used to refer to a standardised mean difference: a statistic for combining continuous variables (such as pain scores or height), from different scales, by dividing the difference between two means by an estimate of the within-group standard deviation.
Standardised mean differences are very difficult to interpret because combining heterogenous scales provides statistical accuracy at the expense of clinical intelligibility.
The occurrence of a dichotomous outcome that is being sought in the study (such as myocardial infarction, death, or a 4-point improvement in pain score).
In determining the power of a trial, the event rate is as important as the number of participants. If the occurrence of an event is very low, even in a very large trial, there may not be sufficient events to detect differences in outcomes between treatment and control groups.
A study in which the investigator studies the effect of intentionally altering one or more factors under controlled conditions.
The validity of the results of a trial beyond that trial.
A randomised controlled trial (RCT) only provides direct evidence of causality within that trial. It takes an additional logical step to apply this result more generally. However, practically it is necessary to assume that results are generalisable unless there is evidence to the contrary. If evidence is consistent across different settings and in different populations (e.g., across ages and countries) then there is evidence in favour of external validity. If there is only evidence from atypical settings (e.g., teaching hospital when most cases are seen in primary care), then one should be more sceptical about generalising the results. Generalisability is not just a consequence of the entry requirements for the trial, but also depends on the population from which the trial population was drawn (see also applicability).
A factorial design attempts to evaluate more than one intervention compared with control in a single trial, by means of multiple randomisations.
A person with the target condition (defined by the gold standard) who has a negative diagnostic test result.
A person without the target condition (defined by the gold standard) who has a positive diagnostic test result.
The "fixed effects" model of meta-analysis assumes, often unreasonably, that the variability between the studies is exclusively because of a random sampling variation around a fixed effect (see also random effects).
Evidence-based health care resources often have great difficulty in providing good-quality evidence on harms (adverse effects of treatment). Most RCTs are not designed to assess adverse effects adequately: the sample size is too small, the trial too short, and often information on adverse effects is not systematically collected. Often a lot of the harms data are in the form of uncontrolled case reports. Comparing data from these series is fraught with difficulties because of different numbers receiving the intervention, different baseline risks, and differential reporting. We aim to search systematically for evidence on what are considered the most important harms of an intervention. The best evidence is from a systematic review of harms data that attempts to integrate data from different sources. However, because of these difficulties and following the maxim "first one must not do harm" we accept weaker evidence of harms. This can include information on whether the intervention has been either banned or withdrawn by regulatory authorities because of the risk of harms.
Broadly equivalent to relative risk (RR); useful when the risk is not constant with respect to time. It uses information collected at different times. The term is typically used in the context of survival over time. If the HR is 0.5 then the relative risk of dying in one group is half the risk of dying in the other group.
HRs take account of more data than RRs and are particularly useful when assessing chronic conditions or long-term sequelae.
In the context of meta-analysis, heterogeneity means dissimilarity between studies. It can be because of the use of different statistical methods (statistical heterogeneity), or evaluation of people with different characteristics, treatments, or outcomes (clinical heterogeneity). Heterogeneity may render pooling of data in meta-analysis unreliable or inappropriate. Finding no significant evidence of heterogeneity is not the same as finding evidence of no heterogeneity. If there is a small number of studies, heterogeneity may affect results but not be statistically significant.
Similarity (see heterogeneity).
The number of new cases of a condition occurring in a population over a specified period of time.
Analysis of data for all participants based on the group to which they were randomised and not based on the actual treatment they received.
Where possible we report ITT results. However, different methods go under the name ITT. Therefore, it is important to state how withdrawals were handled and any potential biases, e.g., the implication of carrying last result recorded forward will depend on the natural history of the condition.
A method for controlling for withdrawal of participants from a trial and allowing data regarding their progress up until withdrawal to be included in the analysis of results. As an example, assume that there are weekly assessments over 8 weeks after the baseline observation of the patient. If a patient withdraws from the study after the third week, then this value is "carried forward" and assumed to be his or her score for the 5 missing data points up until the end of the trial. The assumption is that a patient will improve gradually from the start of the study until the end, so that carrying forward an intermediate value is a conservative estimate of how well the person would have done had he or she remained in the study. There are clear limitations to this approach as the progress of patients on treatment or placebo is often not linear, but it does allow for more data to contribute to the analysis, which increases the power of the trial to detect differences between groups.
The ratio of the probability that an individual with the target condition has a specified test result to the probability that an individual without the target condition has the same specified test result.
A statistical technique that summarises the results of several studies in a single weighted estimate, in which more weight is given to results of studies with more events and sometimes to studies of higher quality.
We use meta-analysis to refer to the quantitative methods (usually involving weighting) used to integrate data from trials. This is logically distinct from a systematic review, which is defined by an explicitly systematic search and appraisal of the literature with or without an accompanying meta-analysis of the data. Meta-analysis is also distinct from data pooling, which is based purely on the raw data. It should be noted that the statistical package RevMan assumes that all outcomes are adverse and therefore if RevMan states that the results for a beneficial outcome favour control this means the beneficial outcome is more likely with the experimental intervention.
Rate of illness but not death.
Rate of death.
The ratio of the probability that an individual with the target condition has a negative test result to the probability that an individual without the target condition has a negative test result. This is the same as the ratio (1sensitivity/specificity).
The chance of not having a disease given a negative diagnostic test result (not to be confused with specificity, which is the other way round).
A review or meta-analysis that either did not perform a comprehensive search of the literature and contains only a selection of studies on a clinical question, or did not state its methods for searching and appraising the studies it contains.
Not significant in Clinical Evidence means not statistically significant. In statistical terms this means that the observed difference, or a larger difference, could have arisen by chance with a probability of more than 1/20 (i.e., 5%), assuming that there is no underlying difference. This is not the same as saying there is no effect, just that this experiment does not provide convincing evidence of an effect. This could be because the trial was not powered to detect an effect that does exist, because there was no effect, or because of the play of chance.
One measure of treatment harm. It is the average number of people from a defined population you would need to treat with a specific intervention for a given period of time to cause one additional adverse outcome. NNH can be calculated as 1/ARI. In Clinical Evidence, these are usually rounded downwards.
One measure of treatment effectiveness. It is the average number of people who need to be treated with a specific intervention for a given period of time to prevent one additional adverse outcome or achieve one additional beneficial outcome. NNT can be calculated as 1/ARR . In Clinical Evidence, NNTs are usually rounded upwards.
Absolute measures are useful at describing the effort required to obtain a benefit, but are limited because they are influenced by both the treatment and also by the baseline risk of the individual. If a meta-analysis includes individuals with a range of baseline risks, then no single NNT will be applicable to the people in that meta-analysis, but a single relative measure (odds ratio or relative risk) may be applicable if there is no heterogeneity. In Clinical Evidence, an NNT is provided for meta-analysis, based on a combination of the summary odds ratio (OR) and the mean baseline risk observed in average of the control groups.
Observational studies are the most appropriate form of evidence for the Prognosis, Aetiology, and Incidence/Prevalence sections. The minimum data set and methods requirements for observational studies have not been finalised. However, we always indicate what kind of observational study, whether case series, case-control, prospective or retrospective cohort study.
The odds of an event happening is defined as the probability that an event will occur, expressed as a proportion of the probability that the event will not occur.
One measure of treatment effectiveness. It is the odds of an event happening in the experimental group expressed as a proportion of the odds of an event happening in the control group. The closer the OR is to 1, the smaller the difference in effect between the experimental intervention and the control intervention. If the OR is greater (or less) than 1, then the effects of the treatment are more (or less) than those of the control treatment. Note that the effects being measured may be adverse (e.g., death or disability) or desirable (e.g., survival). When events are rare the OR is analagous to the relative risk (RR), but as event rates increase the OR and RR diverge.
The ratio of events to non-events in the intervention group over the ratio of events to non-events in the control group.
The complement of odds ratio (1OR), similar to the relative risk reduction (RRR) when events are rare.
A trial in which both participant and assessor are aware of the intervention allocated.
Measures of treatment effect or failure such as mortality, morbidity, quality of life, ability to work, pain, etc. Laboratory outcomes are best avoided. Even if there is a strong relationship between a laboratory outcome marker and a clinical outcome it is not automatic that it will hold practice. Outcomes that are markers for clinically important patient-centred outcomes are often called surrogate outcomes (e.g., ALT concentrations are a proxy for liver damage following paracetamol overdose).
Generally used to refer to unpublished data obtained directly from authors of published articles as clarification of their published data.
Population, intervention, comparison, and outcome, all with a time element (PICOt). The current reporting requirements of systematic reviews are: how many RCTs, how many participants in each, comparing what with what, in what type of people, with what results. Each variable needs a temporal element, (how old are the participants, how long is the treatment given for, when is the outcome measured).
A substance given in the control group of a clinical trial, which is ideally identical in appearance and taste or feel to the experimental treatment and believed to lack any disease-specific effects. In the context of non-pharmacological interventions, placebo is usually referred to as sham treatment.
Placebo is not the same as giving no treatment and can induce real physiological changes. Whether it is appropriate to compare the experimental with placebo or no treatment depends on the question being asked. Where possible we report on the specific intervention given as a placebo. We include, if available, information on whether participants or clinicians could distinguish between placebo and the intervention.
The ratio of the probability that an individual with the target condition has a positive test result to the probability that an individual without the target condition has a positive test result. This is the same as the ratio (sensitivity/1specificity).
The chance of having a disease given a positive diagnostic test result (not to be confused with sensitivity, which is the other way round).
A study has adequate power if it can reliably detect a clinically important difference (i.e., between two treatments) if one actually exists. The power of a study is increased when it includes more events or when its measurement of outcomes is more precise.
An RCT designed to provide results that are directly applicable to normal practice (compared with explanatory trials that are intended to clarify efficacy under ideal conditions). Pragmatic RCTs recruit a population that is representative of those who are normally treated, allow normal compliance with instructions (by avoiding incentives and by using oral instructions with advice to follow manufacturers instructions), and analyse results by "intention to treat" rather than by "on treatment" methods.
The proportion of people with a finding or disease in a given population at a given time.
A term generally used to describe the public record of the proposed approach for a systematic review. It outlines the clinical question that the review is designed to address, details the criteria against which studies will be assessed for inclusion in the review, and describes how the authors will manage critical appraisal and data synthesis. The aim of the protocol is to prospectively set the research aims and protect against post hoc adjustment of the clinical question in line with what is found in the literature, ensuring that reviews remain question-driven.
See surrogate outcomes.
Occurs when the likelihood of a study being published varies with the results it finds. Usually, this occurs when studies that find a significant effect are more likely to be published than studies that do not find a significant effect, so making it appear from surveys of the published literature that treatments are more effective than is truly the case.
Can occur through both preference for significant (positive) results by journals and selective releasing of results by interested parties. A systematic review can try and detect publication bias by a forest plot of size of trial against results. This assumes that larger trials are more likely to be published irrespective of the result. If a systematic review finds evidence of publication bias this should be reported. Often publication bias takes the form of slower or less prominent publication of trials with less interesting results.
The probability that an observed or greater difference occurred by chance, if it is assumed that there is in fact no real difference between the effects of the interventions. If this probability is less than 1/20 (which is when the P value is less than 0.05), then the result is conventionally regarded as being "statistically significant".
A method for comparing health outcomes, which assigns to each year of life a weight from 1 (perfect health) to 0 (state judged equivalent to death) dependent on the individual's health-related quality of life during that year. A total score of years multiplied by weight can then be compared across different interventions. There is disagreement about the best methods for measuring health-related quality of life.
See best evidence.
A trial using a method of allocating participants to different forms of care that is not truly random; for example, allocation by date of birth, day of the week, medical record number, month of the year, or the order in which participants are included in the study (e.g., alternation).
The "random effects" model of meta-analysis assumes a different underlying effect for each study and takes this into consideration as an additional source of variation, which leads to somewhat wider confidence intervals than the fixed effects model. Effects are assumed to be randomly distributed, and the central point of this distribution is the focus of the combined effect estimate.
The random effects model is more effective than the fixed effects model at allowing for trial heterogeneity but does not remove the heterogeneity caused by elements such as differing trial methods and populations.
A system of allocating patients to treatment groups similar to tossing a coin (the experimental treatment is assigned if the coin lands "heads" and a conventional, "control" or "placebo" treatment is given if the coin lands "tails"). Randomisation is often computer generated as this helps to avoid bias.
A trial in which participants are randomly assigned to two or more groups: at least one (the experimental group) receiving an intervention that is being tested and another (the comparison or control group) receiving an alternative treatment or placebo. This design allows assessment of the relative effects of interventions.
Given data on a dependent variable and one or more independent variables, regression analysis involves finding the "best" mathematical model to describe or predict the dependent variable as a function of the independent variable(s). There are several regression models that suit different needs. Common forms are linear, logistic, and proportional hazards.
The number of times more likely (RR >1) or less likely (RR <1) an event is to happen in one group compared with another. It is the ratio of the absolute risk (AR) for each group. It is analogous to the odds ratio (OR) when events are rare.
Relative risk is calculated as the absolute risk (AR) in the intervention group divided by the AR in the control group. It is to be distinguished from odds ratio (OR), which is the ratio of events over non-events in the intervention group over the ratio of events over non-events in the control group. In the USA, odds ratios are sometimes known as rate ratios or relative risks.
The proportional increase in risk between experimental and control participants in a trial.
The proportional reduction in risk between experimental and control participants in a trial. It is the complement of the relative risk (1RR).
The chance of having a positive diagnostic test result given that you have a disease (not to be confused with positive predictive value [PPV]), which is the other way around.
Analysis to test if results from meta-analysis are sensitive to restrictions on the data included. Common examples are large trials only, higher quality trials only, and more recent trials only. If results are consistent this provides stronger evidence of an effect and of generalisability.
An intervention given in the control group of a clinical trial, which is ideally identical in appearance and feel to the experimental treatment and believed to lack any disease specific effects (e.g., detuned ultrasound or random biofeedback). Placebo is used for pills, whereas sham treatment is used for devices, psychological, and physical treatments.
By convention, taken to mean statistically significant at the 5% level. This is the same as a 95% confidence interval not including the value corresponding to no effect.
The chance of having a negative diagnostic test result given that you do not have a disease (not to be confused with negative predictive value [NPV], which is the other way around.
A measure of effect size used when outcomes are continuous (such as height, weight, or symptom scores) rather than dichotomous (such as death or myocardial infarction). The mean differences in outcome between the groups being studied are standardised to account for differences in scoring methods (such as pain scores). The measure is a ratio; therefore, it has no units.
SMDs are very difficult to interpret and combining heterogenous scales provides statistical accuracy at the expense of clinical intelligibility.
Means that the findings of a study are unlikely to have arisen because of chance. Significance at the commonly cited 5% level (P <0.05) means that the observed difference or greater difference would occur by chance in only 1/20 similar cases. Where the word "significant" or "significance" is used without qualification in Clinical Evidence, it is being used in this statistical sense.
Analysis of a part of the trial/meta-analysis population in which it is thought the effect may differ from the mean effect. Subgroup analysis should always be prespecified. Otherwise, they provide weak evidence as they are underpowered so are more suited for hypothesis generation. If many tests are done on the same data this increases the chance of finding spurious effects and some kind of correction is needed (e.g., Bonferroni). Given independent data, and no underlying effect, 1 time in 20 a significant result would be expected by chance.
Outcomes not directly of importance to patients and their carers but predictive of patient-centred outcomes.
A review in which specified and appropriate methods have been used to identify, appraise, and summarise studies addressing a defined question. It can, but need not, involve meta-analysis). In Clinical Evidence, the term systematic review refers to a systematic review of RCTs unless specified otherwise.
In Clinical Evidence, we aim to avoid saying there was a non-significant trend. Alternatives include, "fewer people died after taking treatment x but the difference was not significant" or "The difference was not significant but the confidence intervals covered the possibility of a large beneficial effect" or even, "The difference did not quite reach significance".
A person without the target condition (defined by a gold standard) who has a negative test result.
A person with the target condition (defined by a gold standard) who also has a positive test result.
The soundness or rigour of a study. A study is internally valid if the way it is designed and carried out means that the results are unbiased and it gives you an accurate estimate of the effect that is being measured. A study is externally valid if its results are applicable to people encountered in regular clinical practice.
A measure of effect size used when outcomes are continuous (such as symptom scores or height) rather than dichotomous (such as death or myocardial infarction). The mean differences in outcome between the groups being studied are weighted to account for different sample sizes and differing precision between studies. The WMD is an absolute figure and so takes the units of the original outcome measure.
A continuous outcome measure, similar to standardised mean differences but based on one scale so in the real units of that scale. Ideally should be replaced by a discrete outcome and a relative risk; however, we use WMD if this is not possible.