Book summary “How to read a paper”: summary

Introduction

You face a constant flood of medical research, yet clinical decisions require dependable facts rather than intuition, anecdotes, or flawed consensus panels. This information overload leaves you vulnerable to misleading pharmaceutical trials, biased reviews, and poorly designed studies. Trisha Greenhalgh and the book How to Read a Paper show you how to navigate this challenge through evidence-based medicine.

Evidence-based medicine requires you to critically appraise research papers using structured methodological rules rather than relying on intuition or flawed authority.

The book guides you through a logical progression from searching for medical literature to evaluating primary study designs and statistical methods. You learn how to appraise specific research types like drug trials, diagnostic tests, and systematic reviews before exploring guidelines and implementation.

How can you efficiently find the right studies, spot hidden methodological flaws, and apply trustworthy findings to your daily patient care?

1. Searching the Medical Literature

Medline alone contains over 20 million references. You must match your literature search strategy to your clinical purpose while prioritizing synthesized, pre-appraised sources like systematic reviews.

Practicing evidence-based medicine requires a structured five-step process starting from formulating questions and tracking down evidence to critical appraisal, implementation, and evaluation.

Professor Sackett outlined that simply reading papers or doing critical appraisal is insufficient if the initial question is wrong or if the findings are never implemented or evaluated in practice. You can convert your information needs into answerable questions by breaking down the clinical problem. A well-formulated clinical question must explicitly define the patient group or problem, the specific manoeuvre or intervention being considered, the comparison manoeuvre, and the clinical outcome.

Health professionals approach the medical literature for three distinct purposes: browsing for awareness, looking for quick answers for patient care, and thoroughly surveying the literature for research. Synthesised information sources like systematic reviews and point-of-care tools sit at the top of the evidence hierarchy because they translate primary research into actionable guidance and minimize bias. Systematic reviews pool data from multiple clinical trials using rigorous protocols, sometimes combining smaller studies via meta-analysis and forest plots. Point-of-care resources such as Clinical Evidence, DynaMed, and ACP PIER provide regularly updated, evidence-based summaries designed for bedside use. Since 2008, Cochrane systematic reviews have covered diagnostic tests.

Effective literature searching requires matching the search strategy to the clinical purpose, prioritizing higher levels of evidence like systematic reviews, and utilizing specialized tools and citation chaining. You should decide on the purpose of your search: browsing, seeking an answer to a clinical question, or a comprehensive review. Go for the highest level of evidence you can. For keeping abreast of new developments, use synopses such as POEMS. Make yourself familiar with the specialised resources in your own field and use these routinely. When searching the Medline database for primary research, increase efficiency by doing two broad searches and combining them. Citation chain an older paper to identify recent publications on a topic. Use federated search engines such as TRIP or SUMsearch which search multiple resources simultaneously. Consult human sources including expert librarians and specialists in the field. Try an online self-study course to improve searching skill and confidence.

2. Getting Your Bearings and Study Designs

Critical appraisal requires evaluating a paper based on its methodological design rather than the interestingness of its hypothesis or the significance of its results, as the vast majority of published articles contain flaws.

Most published papers follow the standard IMRAD format. You must choose the appropriate primary study design based on your specific clinical question.

• laboratory experiments;
• clinical trials;
• surveys;
• organizational case studies.

In 1979, Dr Stephen Lock wrote about rejecting papers with irremediable flaws. Reviewers note that fundamental methodological flaws frequently occur even in high-quality journals. Only 1 percent of medical research is claimed to be free of methodological flaws. You should inspect the methods section before looking at the results.

Randomised controlled trials are considered the gold standard for evaluating therapeutic interventions because random allocation creates identical groups, isolating the effect of a single variable.

• parallel group comparison;
• crossover;
• placebo-controlled;
• factorial design.

Trials can be compromised by flawed randomisation methods, exclusion criteria, failure to blind assessors, and publication bias. Cohort studies and case-control studies are observational designs primarily used to determine the etiology, prognosis, or causes of diseases rather than to test treatments.

Sir Austen Bradford Hill, Sir Richard Doll, and Sir Richard Peto followed up 40,000 male British doctors divided into four smoking cohorts. The landmark study achieved a 94 percent follow-up rate, demonstrating a dose-response relationship between smoking and lung cancer mortality.

While traditional hierarchies place systematic reviews at the top and case reports at the bottom, the hierarchy of evidence should be applied flexibly as a rule of thumb rather than mechanically. The traditional hierarchy of evidence scale spans 7 points. Case reports and case series describe individual patient histories and adverse reactions, offering immediate real-world insights that can be published rapidly.

Dr W.G. McBride published a case report alerting colleagues worldwide to the link between thalidomide taken in early pregnancy and absent limbs in newborns. Evaluating scientific contribution requires examining risk of bias, consistency, directness, and precision rather than relying on a rigid pyramid.

3. Assessing Methodological Quality

A research paper stands or falls on the strength of its methods section, requiring you to check recruitment, blinding, and sample size.

Trisha Greenhalgh states that a study sinks or swims on methodological quality. You must evaluate five essential questions:

• Was the study original?
• Whom is it about?
• Was it well designed?
• Was systematic bias avoided?
• Was it large enough and continued for long enough to make the results credible?

Clinical trials often suffer from recruitment bias. Pharmacokinetic studies of new drugs tested exclusively in 23-year-old healthy male volunteers are clearly not applicable to the average elderly female. You must follow these selection rules:

• Check how participants were recruited for recruitment bias.
• Examine who was included in the study.
• Review who was excluded from the study.
• Assess whether participants were studied in real-life circumstances.

Failure to keep outcome assessors blind to patient group allocations allows expectation bias to distort trial results.

Majeed and colleagues performed an RCT demonstrating that recovery time after laparoscopic gallbladder removal was no quicker than open surgery, a discrepancy from previous studies likely due to meticulous blinding where patients left surgery with identical blood-stained dressings. You evaluate masking with specific checks:

• Check if outcome assessors were kept blind or masked to group allocation.
• Look for evidence of performance bias caused by unblinded assessors.

Underpowered studies lead to Type II errors, which are erroneous conclusions that an intervention has no effect. Common statistical power stipulated for clinical trials is 80% to 90%, alongside a 1 in 20 odds threshold for a statistically significant p-value less than 0.05. You apply these quantitative checks:

• Extract the sample size from the paper.
• Extract the duration of follow-up.
• Extract the completeness of follow-up.
• Perform a sample size calculation if the paper lacks one and claims no difference.

Studies with low follow-up rates below the 70% threshold are generally considered untrustworthy. You defeat dropout bias with strict adherence to final verification:

• Check if results were analysed on an intent-to-treat basis.
• Confirm the figures yourself to ensure withdrawals were properly accounted for.

4. Statistics for the Non-Statistician

You must look past p-values and relative risks by examining confidence intervals and absolute risk reductions.

Selecting appropriate statistical tests requires understanding whether data are measured as actual numbers or arbitrary ranks and whether they follow a normal distribution. Statistical tests are classified as parametric, which assume data follow a specific distribution like the normal curve, or non-parametric, which focus on rank orders. Using parametric tests on skewed or non-normally distributed data produces spurious and misleading estimates.

• Check if the data are normally distributed before applying parametric tests.
• Use non-parametric rank order tests if data cannot be transformed into a normal distribution.

Correlation coefficients do not prove causation, and correlation must be carefully distinguished from mathematical regression equations. The Pearson correlation coefficient measures association strength between normally distributed variables but does not indicate direction or cause. Regression provides an equation to predict a target variable from independent variables, but establishing true causation requires meeting specific epidemiological criteria.

• Ensure every r-value is accompanied by a p-value or confidence interval.
• Check that variables in a correlation are structurally independent.
• Evaluate Bradford Hill criteria before assuming causality from an association.

Confidence intervals provide a continuum of the strength of evidence, overcoming the arbitrary dichotomous limits imposed by p-values. While a p-value indicates the probability that an outcome arose by chance based on an arbitrary cutoff, confidence intervals estimate the range within which the true population difference lies across repeated trials. Larger trials produce narrower confidence intervals, making their findings more definitive.

• Check the upper 95% confidence interval of a negative trial to see if a much larger trial could show a clinically significant benefit.

Expressing intervention effects through absolute risk reduction and the number needed to treat provides practical clinical meaning for patients. Relative risk reductions can exaggerate the appearance of a treatment benefit, whereas absolute risk reduction and the number needed to treat translate population statistics into individual patient outcomes and expected treatment burdens.

Fahey and colleagues surveyed 182 health authority board members using four different rehabilitation descriptions for heart attack victims, where 137 respondents failed to recognize that a 20% relative risk reduction, a 3% absolute risk reduction, a survival change from 84% to 87%, and 31 patients needed to treat all described the exact same underlying trial results.

• Calculate the absolute risk of death for the control group and the intervention group.
• Calculate the relative risk reduction as 1 minus the ratio of intervention risk to control risk.
• Calculate the absolute risk reduction as the difference between control risk and intervention risk.
• Calculate the number needed to treat as the reciprocal of the absolute risk reduction.

5. Papers That Report Drug Trials

You must guard against industry presentation tactics and surrogate endpoints when evaluating trials of drug treatments.

The pharmaceutical industry employs sophisticated marketing tactics and direct-to-consumer advertising to influence clinicians and patients. Drug companies use drug reps and detailer professionals alongside direct-to-consumer advertising to shape prescribing habits. Patient requests driven by advertising can significantly alter your prescribing behavior for drugs like antidepressants.

Clotrimazole pessaries for vaginal thrush transitioned from prescription to over-the-counter availability and were subsequently advertised on prime-time television and the internet.

Industry presentation tactics cast products in the most favorable light. You must watch out for techniques designed to mask true clinical performance:

• Think up a plausible physiological mechanism why the drug works and find a surrogate endpoint heavily influenced by the drug.
• Select a patient population, clinical features, and trial length that reflect the maximum possible response.
• Compare the product only with placebos or with a competitor at a sub-therapeutic dose.
• Include pilot study results in definitive studies to inflate perceived randomisation numbers.
• Omit mention of trials with fatalities or serious adverse drug reactions.
• Maximise visual impact through graphics departments without labeling axes or showing confidence intervals.
• Become master of the hanging comparative.
• Invert the standard hierarchy of evidence so anecdote takes precedence over randomised trials.
• Name at least three local opinion leaders using the drug and offer starter packs.
• Present a cost-effectiveness analysis showing the expensive product works out cheaper.

Surrogate endpoints predict distant clinical outcomes but are not direct measures of harm or benefit. While they reduce trial cost and duration, they carry significant risks of misleading conclusions.

The use of albuminuria to evaluate dual renin-angiotensin blockade in hypertension showed a positive surrogate marker effect, but the combination simultaneously produced a doubly effective rate of the fatal side effect hypokalaemia.

Before starting a patient on drug therapy, you must follow a structured three-step process to avoid therapeutic chaos:

• Identify for the patient the ultimate objective of treatment.
• Select the most appropriate treatment using all available evidence.
• Specify the treatment target and criteria for stopping or changing treatment.

In treating high blood pressure, the ultimate objective is preventing target organ damage. The specific treatment choice involves antihypertensive drugs or salt restriction, while the treatment target is a Phase V diastolic blood pressure of less than 90 mmHg.

The Consolidated Standards of Reporting Trials statement provides a standardized reporting checklist for randomized controlled trials. Produced in 1996 by an international working group and updated in 2010, the checklist ensures methodological transparency and quality.

When a drug rep visits you, follow practical rules to maintain clinical objectivity:

• See representatives only by appointment and focus on a single product of interest.
• Take charge of the interview and ask directly for information instead of listening to a sales routine.
• Request independent published evidence from reputable peer-reviewed journals.
• Do not look at promotional brochures containing unpublished material and misleading graphs.
• Ignore anecdotal evidence such as medical celebrity endorsements.
• Ask for evidence using the STEP acronym covering safety, tolerability, efficacy, and price.
• Evaluate evidence stringently using the CONSORT checklist and check sample size.
• Do not accept the newness of a product as an argument for changing to it.
• Decline to try the product via starter packs or small-scale uncontrolled studies.
• Record the content of the interview in writing.

6. Papers That Report Complex Interventions

Complex interventions involve multiple interacting components operating at various levels, requiring a theoretical core combined with flexible non-core adaptations.

Unlike simple interventions with an on-off design, complex interventions present severe implementation challenges. They require a development phase with qualitative interviews or pilot trials to optimize components before full testing. The Medical Research Council produced updated guidance for evaluating complex interventions in 2008.

A PhD student studied yoga classes for diabetes control. Despite a development phase, the trial showed no impact on diabetes control because participants failed to practice exercises at home and many found them too difficult. The age subgroup most keen to try yoga in the diabetes trial was 60 to 69.

In a review of 19 studies reviewed on school-based feeding programmes, six different mechanisms were identified through which nutrition or school performance could theoretically improve. Only 27% of implementation studies included an explicit theory of change.

Appraising a paper on a complex intervention requires answering ten specific questions regarding the problem, developmental phase, core components, theoretical mechanism, and process evaluations.

You must evaluate these studies using specific appraisal steps:

• Ask what specific problem the complex intervention aims to solve.
• Examine what was done in the developmental phase to inform design.
• Identify core components that must be standardized versus non-core components adapted to context.
• Determine the explicit theoretical mechanism of action of the intervention.
• Evaluate whether outcome measures used were sensible and comprehensive.
• Analyze the findings across various outcome measures.
• Check what qualitative process evaluation was carried out alongside the trial.
• Assess whether negative findings can be explained by implementation failure or inadequate optimisation.
• Examine how variation across different subgroups is explained by refining the theory of change.
• Evaluate whether authors' recommendations for further research are justified.

7. Diagnostic Tests and Clinical Prediction Rules

You evaluate diagnostic test performance by mapping outcomes in a two-by-two table to calculate sensitivity, specificity, and likelihood ratios.

A jury of ten people judging ten men for murder, where three are actual murderers and seven are innocent, is used to illustrate how a two-by-two table works. The horizontal axis represents the truth based on the gold standard, while the vertical axis represents the verdict or test result. By categorizing outcomes into true positives, false positives, false negatives, and true negatives, you calculate core performance metrics.

Five key features define a diagnostic test's performance when compared to a gold standard:

• sensitivity measures the ability to detect true positives;
• specificity measures the ability to exclude true negatives;
• positive predictive value tells you what a positive test result means;
• negative predictive value tells you what a negative test result means;
• accuracy reflects the overall proportion of correct test results.

Andersson and colleagues validated urine glucose testing against the gold standard glucose tolerance test using a population denominator of 1000 participants. The study found a 22.2 percent sensitivity, a 99.3 percent specificity, a 46.2 percent positive predictive value, a 97.9 percent negative predictive value, and a 97.2 percent overall accuracy.

Using 100 healthy male medical students to validate a cholesterol test would create a severely biased sample because it omits women, children, and older people. Evaluating a paper about a diagnostic test requires ten specific questions:

• Question One: Is this test potentially relevant to my practice?
• Question Two: Has the test been compared with a true gold standard?
• Question Three: Did this validation study include an appropriate spectrum of participants?
• Question Four: Has work-up (verification) bias been avoided?
• Question Five: Has expectation bias been avoided?
• Question Six: Was the test shown to be reproducible both within and between observers?
• Question Seven: What are the features of the test as derived from this validation study?
• Question Eight: Were confidence intervals given for sensitivity, specificity and other features of the test?
• Question Nine: Has a sensible normal range been derived from these results?
• Question Ten: Has this test been placed in the context of other potential tests in the diagnostic sequence for the condition?

Likelihood ratios provide a powerful way to express how much a test result shifts the odds of a patient having a target disorder.

Instead of relying solely on a binary normal or abnormal cut-off, you can combine likelihood ratios with pre-test probability using a Fagan nomogram to calculate the post-test probability of a disease. A serum ferritin level between 18 and 45 micrograms per litre has a likelihood ratio of 3 for diagnosing iron-deficiency anaemia. Based on a prevalence of one in twenty, you start with a 5 percent pre-test probability in the UK population and reach a 15 percent post-test probability.

Combining multiple diagnostic tests, signs, and symptoms into clinical prediction rules helps you sharpen diagnostic focus for complex medical challenges. Women presenting with a breast lump are typically offered a triple combination of tests: fine needle aspiration, X-ray mammogram, and ultrasound. You develop and evaluate these prediction rules through distinct stages:

• Stage one: develop the rule by establishing the independent and combined effect of explanatory variables such as symptoms, signs, or diagnostic tests on the diagnosis.
• Stage two: assess these explanatory variables in different populations.
• Stage three: conduct an impact analysis measuring the impact of applying the rule in a clinical setting in terms of patient outcome, clinician behaviour, and resource use.

8. Systematic Reviews and Meta-Analysis

Systematic reviews use explicit, reproducible methods to pool data into forest plots while accounting for statistical heterogeneity.

Journalistic reviews allow authors to selectively pick studies supporting a pet theory while ignoring contradictory data. Systematic reviews rely on predefined objectives, rigorous search strategies, and quality standards independent of study results.

Nobel laureate Linus Pauling published a review claiming vitamin C cures the common cold based on selected supportive studies. A true objective analysis of all available data showed no effect.

You must ensure that a systematic review covers several procedural requirements:

• contain a statement of objectives, sources and methods
• conducted in a way that is explicit, transparent and reproducible

Appraising a review requires working through five core questions:

• What is the important clinical question that the review addressed?
• Was a thorough search carried out of the appropriate database(s) and were other potentially important sources explored?
• Was methodological quality assessed and the trials weighted accordingly?
• How sensitive are the results to the way the review has been performed?
• Have the numerical results been interpreted with common sense and due regard to the broader aspects of the problem?

Electronic databases alone are insufficient for a comprehensive search. Knipschild and colleagues found that electronic databases yielded only 22 out of 61 total trials on vitamin C and cold prevention, requiring manual Index Medicus searches and extensive reference tracking to uncover the rest.

Meta-analysis results are commonly visualized using forest plots. Individual trial point estimates and 95 percent confidence intervals appear along a horizontal line against a vertical line of no effect at relative risk 1.0.

Cuijpers et al. analyzed eight trials comparing cognitive behavior therapy to discontinued pharmacotherapy for depression. Individual trials lacked sufficient sample size for significance, but the pooled diamond showed a significant reduction in relapse with a 2.61 pooled relative risk and a narrower confidence interval from 1.58 to 4.31.

When trials in a meta-analysis are statistically heterogeneous, mathematical averaging becomes misleading. Thompson analyzed ten trials of cholesterol-lowering strategies showing an initial Chi-square statistic of 127, indicating high heterogeneity. Correcting for the age of participants reduced the Chi-square to 45, proving that cholesterol reduction is more effective in younger individuals.

9. Guidelines, Economics, and Quality Improvement

You must appraise clinical guidelines, economic evaluations, and quality improvement case studies through structured domain frameworks.

The AGREE instrument structures the evaluation of clinical guidelines through six distinct quality domains:

• Scope and purpose
• Stakeholder involvement
• Rigour of development
• Clarity and presentation
• Applicability
• Editorial independence

Economic evaluations are categorized into four main types depending on how costs and outcomes are measured: cost-minimisation analysis, cost-effectiveness analysis, cost-utility analysis, and cost-benefit analysis. When comparing treatments, you must use incremental costs rather than average costs per case. Drug X costs £100 and cures 10 out of 20 patients. Drug Y costs £120 and cures 11 patients, making the incremental cost to cure that one extra patient £400.

Quality improvement work differs fundamentally from traditional clinical research. Research seeks transferable truths using standard IMRAD formats. Quality improvement is real-time work by delivery teams focused on changing human performance within a specific context, often reported using SQUIRE guidelines. May et al. evaluated academic detailing in middle America, documenting a population of 260,512 and a median household income of $39,813.

10. Getting Evidence Into Practice and Patient Care

Closing the evidence-to-practice gap requires addressing human factors, organizational absorptive capacity, and shared decision-making with patients.

Health professionals are often slow to adopt proven medical interventions due to a complex set of barriers rather than mere ignorance or stubbornness. Increasing the delivery rate and fidelity of existing proven medical treatments often saves more lives than discovering entirely new drugs.

Artificial surfactant for premature babies was adopted rapidly by pediatricians by 1993 because its effect was immediate and supported by large trials, whereas prenatal dexamethasone for pregnant women took decades to spread because its effects were seen as long-term statistics and the drug was out of patent.

Simply telling professionals about evidence-based medicine through didactic talks and lectures fails to produce sustained changes in clinical behavior. Organizations assimilate innovations and best practices more successfully when they develop an absorptive capacity, receptive context, and specific supportive roles. You must apply these implementation strategies:

• Create an evidence-based culture where decision-making is expected to be based on the best available evidence
• Provide high-quality, up-to-date information sources in every office with protected staff time
• Designate knowledge managers to ensure links to knowledge bases and training
• Deploy knowledge workers to help staff find and apply knowledge
• Designate champions at every level to back and promote evidence-based innovations
• Identify and utilize boundary spanners who maintain strong social ties inside and outside the organization

Evidence-based medicine relies on population averages, but clinical care must integrate individual patient narratives, preferences, and patient-reported outcome measures. A patient declined warfarin because it required stopping grapefruit consumption, a food she had enjoyed for over sixty years, illustrating how personal values diverge from population risk-reduction averages.

Shared decision-making utilizes structured competencies, decision aids, and option grids to help patients deliberate over treatment options without being dictated to by clinicians. You must guide consultations using systematic steps:

• Define the problem clearly
• Portray equipoise by showing no clear professional preference
• Portray treatment options and the option of no treatment
• Provide information in preferred formats and check understanding
• Explore ideas, concerns, and expectations
• Check role preference and involve the patient to the desired extent
• Defer the decision if necessary to allow consideration with family or friends
• Review arrangements at a specified time period

Individualized trial designs such as n of 1 trials and guided trial and error help identify individual patient responders despite the high resource intensity of personalized protocols. March and colleagues conducted an n of 1 trial comparing non-steroidal anti-inflammatory drugs with paracetamol in osteoarthritis patients to uncover individual response patterns. You should manage treatment hurdles through structured protocols:

• Expect analgesic failure because population numbers needed to treat mean individuals are statistically more likely not to benefit
• Pursue analgesic success through a guided process of trial and error, systematically trying one intervention after another

10 Key Ideas

1. Evidence-based medicine requires you to critically appraise research papers using structured methodological rules rather than relying on intuition or flawed authority.
2. Effective literature searching requires matching the search strategy to the clinical purpose, prioritizing higher levels of evidence like systematic reviews, and utilizing specialized tools and citation chaining.
3. Critical appraisal requires evaluating a paper based on its methodological design rather than the interestingness of its hypothesis or the significance of its results, as the vast majority of published articles contain flaws.
4. You must look past p-values and relative risks by examining confidence intervals and absolute risk reductions.
5. You must guard against industry presentation tactics and surrogate endpoints when evaluating trials of drug treatments.
6. Complex interventions involve multiple interacting components operating at various levels, requiring a theoretical core combined with flexible non-core adaptations.
7. You evaluate diagnostic test performance by mapping outcomes in a two-by-two table to calculate sensitivity, specificity, and likelihood ratios.
8. Systematic reviews use explicit, reproducible methods to pool data into forest plots while accounting for statistical heterogeneity.
9. You must appraise clinical guidelines, economic evaluations, and quality improvement case studies through structured domain frameworks.
10. Closing the evidence-to-practice gap requires addressing human factors, organizational absorptive capacity, and shared decision-making with patients.