🔬 BCBR • Premium Review

Health research is a systematic process for producing generalisable knowledge about health, disease, care or health systems. It differs from routine clinical care because the primary intent is knowledge generation; it differs from audit because audit usually asks whether practice meets a predefined standard.

Start with the problem, not the method. A common exam trap is choosing a sophisticated design before defining exactly what is unknown and why it matters.

Clinical example: repeated admissions for acute exacerbation of COPD may generate a research problem. “Are our readmissions high?” is audit-like; “Among adults discharged after AECOPD, does a structured inhaler-and-follow-up bundle reduce 30-day readmission compared with usual discharge?” is an analytic research question.

Research cycle: identify problem → review evidence → formulate question/objectives → select design/population/variables → ethics → collect/manage data → analyse → interpret → disseminate.

A good research question specifies the population, exposure/intervention, comparator where relevant, and outcome. FINER—feasible, interesting, novel, ethical and relevant—is useful for judging whether the question deserves a study.

The primary objective should determine the main outcome, sample-size assumptions and principal analysis. Secondary objectives add information but should not compete with an ill-defined primary objective.

Null hypothesis usually states no specified association/difference; the alternative states that one exists. Hypotheses are most useful for analytical questions, not every descriptive study.

Exam distinction: aim is broad; objective is specific and measurable. “To study hypertension” is not an adequate objective; “to estimate prevalence of hypertension among adults ≥18 years in X district during 2026” is.

A literature review maps what is already known, how previous studies measured the problem, where findings disagree and what gap remains. It should justify the new study rather than become a list of article summaries.

Build the search from concepts. Combine synonyms with OR; combine different concepts with AND. Record databases, dates, search terms and selection logic so the search can be reproduced.

Read methods and results, not only abstracts. Extract population, design, sample size, exposure/intervention, outcome definitions, effect estimates, precision, limitations and applicability.

Exam trap: a literature review is not complete merely because many papers were found. Relevance, validity and synthesis determine its usefulness.

A concept paper is a concise bridge between an idea and a full protocol. It should communicate the problem, rationale, question, objectives, proposed methods, expected value, feasibility and broad resource needs.

A strong concept paper allows a reviewer to answer three questions quickly: Is the problem important? Can the proposed design answer it? Can the team realistically execute it?

Keep methods proportional to the question. If the objective is prevalence estimation, a cross-sectional sampling plan may be appropriate; if the objective is causal treatment effect, a trial may be needed.

Before expanding to a protocol, challenge assumptions about recruitment rate, event frequency, measurement availability, cost, expertise and time.

Prevalence = existing cases / population assessed; it describes burden. Cumulative incidence = new cases / population initially at risk over a specified period; incidence rate uses new events divided by person-time at risk.

Prevalence depends on both incidence and duration. A chronic disease may have high prevalence despite modest incidence because patients live with it for years.

Person-time is useful when follow-up differs. Ten persons followed one year and twenty persons followed six months contribute 20 person-years in total, assuming each is at risk for those periods.

Exam trap: do not place existing cases in the denominator for cumulative incidence if they are not at risk of becoming incident cases.

Case reports and case series describe unusual presentations or patterns but lack a comparison group. They are excellent for signal generation and poor for estimating comparative effects.

Cross-sectional studies measure exposure and outcome at one time/short window and are efficient for prevalence. Their major causal limitation is temporality: it may be unclear whether exposure preceded outcome.

Ecological studies compare group-level exposure and outcome. They are useful for population hypotheses but individual-level inference can produce ecological fallacy.

Choose descriptive designs when the purpose is to describe distribution or generate hypotheses—not because they are simply easier.

Case-control studies begin with outcome status and look backward or otherwise ascertain prior exposures. They are efficient for rare diseases and usually estimate an odds ratio. Selecting controls from the population that produced the cases is central.

Cohort studies classify participants by exposure and observe outcome occurrence. They can estimate incidence and risk/rate ratios and establish temporality more naturally, but may be inefficient for very rare outcomes.

Nested case-control and case-cohort approaches can improve efficiency inside cohorts while retaining advantages of a defined source population.

Classic exam clue: rare outcome + multiple past exposures often points toward case-control; rare exposure + multiple subsequent outcomes often favours cohort.

Randomised controlled trials allocate interventions by chance to create comparable groups on average. Random sequence generation prevents predictable assignment; allocation concealment prevents recruiters from knowing the next assignment.

Blinding addresses performance and/or ascertainment bias after allocation; it is not the same as allocation concealment. Placebo is one possible control, not a defining requirement of every RCT.

Intention-to-treat preserves the randomised comparison by analysing participants according to original allocation. Per-protocol analysis asks a different question and can reintroduce selection differences.

Clinical trial interpretation requires effect size, precision, harms, adherence, missing outcomes and applicability—not p value alone.

Internal validity asks whether the study result is credible for the participants studied. External validity asks whether it applies to the target population/settings.

Selection bias arises from systematic differences in who enters or remains in comparison groups. Information bias arises from systematic measurement/classification error. Confounding is distortion by a third factor associated with exposure and outcome but not simply an intermediate step.

Prevent bias at design stage where possible: appropriate sampling/control selection, standardised measurement, blinding, follow-up procedures, restriction/matching/randomisation. Analysis can adjust some confounding but cannot magically repair poor measurement or severe selection bias.

Effect modification is not a nuisance to eliminate: it may represent a real difference in effect across subgroups and should be interpreted as such.

Qualitative research explores meanings, experiences, processes and context—questions poorly reduced to a numeric frequency alone. Common methods include in-depth interviews, focus groups and observation.

Sampling is often purposive: participants are selected because they can illuminate the question. Adequacy is judged by information richness and analytic sufficiency rather than a conventional power calculation.

Analysis commonly involves coding, category development and themes. Credibility is strengthened by reflexivity, triangulation, transparent analytic procedures and attention to discrepant cases.

Exam distinction: qualitative research is not ‘unscientific’ or simply a small quantitative study; its question, sampling logic, data and analytic framework are different.

Define every variable operationally before data collection: what exactly is measured, by whom, with what instrument, unit, timing and cut-off. Ambiguous variables create irreparable analysis problems.

Nominal variables are unordered categories; ordinal categories have rank; continuous variables can take values across a scale; discrete variables are counts. The measurement scale influences summaries and statistical tests.

Reliability is consistency/reproducibility; validity is whether the instrument measures what it is intended to measure. A measurement can be highly reliable yet systematically wrong.

Diagnostic accuracy: sensitivity = P(test+ | disease); specificity = P(test− | no disease). Predictive values depend strongly on disease prevalence/pre-test probability.

For roughly symmetric continuous data, mean and standard deviation are useful; for skewed data, median and interquartile range usually communicate centre and spread better.

Standard deviation describes variability among observations. Standard error describes uncertainty/precision of an estimated statistic and generally falls as sample size increases.

Always inspect distributions. A single mean can conceal skewness, outliers or multimodality. Histograms, box plots and sensible tabulation are diagnostic tools before inferential testing.

Exam trap: mean ± SD is not automatically appropriate for every numerical variable. First ask what the distribution looks like.

Probability sampling gives population units a defined selection mechanism. Simple random, systematic, stratified, cluster and multistage sampling answer different logistical and precision needs.

Stratification samples within predefined subgroups and can improve representation/precision. Cluster sampling selects natural groups such as villages or schools and usually requires analysis/sample-size methods that account for intracluster correlation.

Convenience, consecutive, quota, purposive and snowball methods are non-probability approaches. They can be appropriate for some questions but limit population inference if treated as representative probability samples.

Separate sampling from allocation: random sampling selects people from a population; randomisation allocates enrolled trial participants to interventions.

Sample size is driven by the primary objective and analysis. For prevalence estimation, expected prevalence and desired absolute precision matter; for comparing groups, expected effect size, variability/event rate, alpha and power matter.

For a single proportion, a common large-sample approximation is n = Z² p(1−p)/d², before design-effect, finite-population or non-response adjustments where applicable.

Power = 1−β: probability of detecting the prespecified effect if it truly exists under the model. Increasing power or demanding a smaller detectable effect generally increases required sample size.

Do not inflate sample size blindly. State assumptions and justify them from prior evidence/pilot data; then account for anticipated non-response/loss to follow-up and complex sampling when relevant.

A confidence interval combines an effect estimate with its statistical precision. Narrow intervals are more precise; wide intervals signal substantial uncertainty.

For a ratio measure such as RR or OR, the null value is 1; for an absolute difference it is 0. A 95% CI crossing the null generally corresponds to non-significance at the matching two-sided 5% test.

A p value is the probability, under the null/model assumptions, of data at least as incompatible with the null as those observed. It is not the probability that the null hypothesis is true.

Type I error is false positive (controlled by α); Type II error is false negative (β). Statistical significance does not equal clinical importance.

Start with the research question and outcome type, then ask: how many groups, independent or paired, distribution/assumptions, and whether adjustment is needed.

Two independent approximately normal means → independent t test; paired approximately normal differences → paired t test. Categorical association → chi-square when assumptions are adequate, with exact alternatives for sparse tables.

Non-parametric tests are useful when their rank/distributional framework matches the data; they are not simply ‘tests for small samples’.

A test name without an effect estimate and confidence interval is incomplete reporting. Analysis should quantify the effect, not merely declare significance.

Correlation quantifies association, commonly linear association for Pearson r. Correlation does not establish causation, and r≈0 does not exclude a nonlinear relationship.

Linear regression models a continuous outcome; logistic regression commonly models a binary outcome. Regression can estimate adjusted associations by including measured covariates.

Adjustment is not automatically good: variables should be selected using subject-matter/causal reasoning, not only univariable p values. Overadjustment and collider bias can create misleading results.

Interpret coefficients in their correct scale and report uncertainty. For logistic regression, exponentiated coefficients are odds ratios.

Define target population, source population and study population. Inclusion/exclusion criteria should follow the scientific question and safety needs rather than convenience alone.

Recruitment methods influence representativeness. In case-control studies, controls should represent the exposure distribution in the population that generated the cases.

Track screened, eligible, enrolled and analysed participants. Differences between these groups help readers judge selection and attrition.

Avoid unnecessary exclusions: they can improve homogeneity but reduce generalisability and may unfairly exclude groups from research benefits.

Translate the protocol into work packages: approvals, staff roles, procurement, training, pilot, recruitment, data collection, monitoring, cleaning, analysis and dissemination.

A Gantt chart makes timelines and dependencies visible. Assign responsibility, deliverables and escalation pathways rather than writing ‘team will do’.

Maintain version control for protocol, consent forms, CRFs and SOPs. Deviations should be documented, assessed and handled transparently.

Operational feasibility is a scientific issue: an under-recruited or poorly implemented study may not answer the question regardless of elegant statistics.

Begin with an objective-to-variable matrix. Every item should serve a planned variable or essential operational purpose; unnecessary questions increase burden and error.

Use clear, neutral, single-concept questions. Avoid double-barrelled wording, leading questions, undefined time frames and overlapping response categories.

Choose closed versus open questions deliberately. Define skip patterns, coding, units, permitted ranges and missing-value conventions before field use.

Pilot the tool in people/settings resembling the study population; revise wording, order, timing and workflow before formal data collection.

Standardisation reduces measurement variation: train observers, use SOPs, calibrate instruments and supervise field procedures.

Quality assurance is prospective; quality control detects problems during/after collection. Build both into the workflow rather than cleaning everything at the end.

Protect privacy during interviews/examinations and collect only data necessary for the approved objectives. Sensitive variables require proportionate safeguards.

Monitor completeness, consistency, protocol adherence and recruitment in real time so correctable errors are detected early.

Create a data-management plan before recruitment: identifiers, coding, database structure, access roles, entry, validation, query resolution, backups, audit trail, retention and secure destruction where applicable.

Use a data dictionary/codebook with variable name, label, type, unit, allowed values, missing codes and derivation rules. Never let 0, blank and ‘unknown’ mean the same thing accidentally.

Separate direct identifiers from analysis data when feasible. Pseudonymisation/coding reduces exposure but remains reversible if a linkage key exists; access to the key should be restricted.

Cleaning includes range, logic, duplication and consistency checks. Corrections should be traceable rather than silently overwriting questionable values.

Write the analysis plan from the objectives before looking for interesting p values. Define populations, outcomes, exposures, covariates, summaries, effect measures and handling of missing data.

Begin with participant flow and descriptive analysis. Then estimate the primary effect/association with confidence intervals, followed by prespecified adjusted/sensitivity analyses where appropriate.

Missing data need description and justification of handling. Complete-case analysis can be biased when missingness relates to variables in the analysis.

Interpretation should integrate effect magnitude, precision, bias, confounding, multiplicity, missing data, clinical relevance and external validity.

Core principles include respect for persons/autonomy, beneficence, non-maleficence and justice. Ethical research also requires scientific validity: exposing participants to a study incapable of answering its question is itself problematic.

Risk should be minimised and reasonable in relation to anticipated benefit/knowledge. Participant selection should be fair, and vulnerable groups require additional safeguards against coercion, undue influence or exploitation.

Privacy concerns the circumstances of access to a person; confidentiality concerns protection of information already obtained. Both require planning.

Independent ethics review does not transfer investigator responsibility. Investigators remain accountable for approved conduct, safety and reporting.

Consent is a process, not a signature. Participants need understandable information about purpose, procedures, risks, benefits, alternatives where relevant, confidentiality, voluntariness and withdrawal.

Capacity and voluntariness must be considered. Language, literacy, dependency relationships, emergencies and vulnerability can require adapted procedures and additional safeguards.

Waiver or alteration of consent is not an investigator convenience; it requires ethical justification and approval under applicable guidance.

Protocol amendments, continuing review/reporting and safety/deviation reporting should follow the ethics committee and applicable regulatory requirements.

Trial conduct begins with a scientifically justified protocol, appropriate ethics/regulatory approvals and registration where required. Roles and responsibilities must be clear before recruitment.

Randomisation protects allocation comparability; concealment protects enrolment; blinding protects subsequent behaviour/assessment. These solve different bias problems.

Participant safety requires active adverse-event detection, clinical management, documentation and reporting according to protocol and applicable requirements.

Preserve data integrity with source documentation, auditability, protocol adherence and transparent handling of deviations/missing outcomes.

Write the problem in one paragraph, the evidence gap in the next, then state the research question and objectives. A reviewer should understand why the study matters before reaching the methods.

Give a defensible design, target population, key variables/outcomes, approximate scale, feasibility and expected impact. Avoid pretending a concept paper is a compressed thesis.

A concept paper is also a stress test: if the question cannot be expressed clearly in one page, the project may still be conceptually unfocused.

Exam focus: recognise which elements belong in rationale, objective and methods rather than mixing them.

A protocol is the operational blueprint: title, investigators, background/rationale, question/hypothesis, objectives, design, setting, population, eligibility, sampling, sample size, variables, procedures, data management, analysis, ethics, timeline, budget and references as appropriate.

Every section must connect. Primary objective ↔ primary outcome ↔ sample-size calculation ↔ data-collection variable ↔ statistical analysis. Broken links are a common protocol weakness.

Operational definitions prevent ambiguity. Specify how exposure/outcome will be measured, timing, instrument, thresholds and quality procedures.

Prespecification improves transparency, but justified amendments can occur; they should be version-controlled, approved where required and documented.

Research integrity requires honest generation, analysis and reporting of data. Fabrication invents data; falsification manipulates research materials/processes/data; plagiarism appropriates others’ words or ideas without proper attribution.

Authorship should reflect substantive scholarly contribution and accountability, not hierarchy, payment or routine administrative support alone. Other contributions can be acknowledged.

Undisclosed duplicate publication and selective outcome reporting can distort the evidence base. Conflicts of interest should be disclosed and managed.

Image/data manipulation, citation practices, peer-review confidentiality and correction/retraction responsibilities are also part of publication ethics.

A scientific manuscript tells one coherent research story. IMRaD—Introduction, Methods, Results and Discussion—is a useful core structure for original research.

Introduction: what is known, what is unknown, why this study, objective. Methods: enough detail to judge validity/reproduce. Results: report what was found without interpretive inflation. Discussion: interpret, compare, explain, acknowledge limitations and implications.

Tables and figures should be self-explanatory, labelled with denominators/units and not redundantly duplicate the text. Report effect estimates with confidence intervals where appropriate.

Write the abstract last and make it consistent with the manuscript. Avoid causal language when the design supports only association.

A grant proposal must persuade reviewers that the problem is important, the question is answerable, the methods are rigorous, the team can deliver and the requested resources are justified.

Align specific objectives with work packages, milestones, timeline and budget. A budget should follow the methods—not invent activities to consume a budget ceiling.

Anticipate risks: recruitment failure, equipment delays, missing expertise, data-security issues and contingency plans. Include dissemination and potential impact.

Reviewer-friendly writing is precise: define the gap, novelty/value, methods and deliverables without exaggerated claims.

Appraisal starts with validity: was the design appropriate, were groups selected/comparable, were exposures/outcomes measured well, was follow-up adequate and were analyses appropriate?

Then examine results: effect size, confidence interval, absolute effects where useful, harms and consistency. Finally ask applicability to your patient/population and setting.

Do not use journal prestige or p<0.05 as a shortcut for validity. A precise biased estimate remains biased.

For observational studies, distinguish association from causation and consider residual confounding and selection/information bias.

Translate the stem into a domain before reading options: design? measure of frequency? sampling? variable type? bias? statistical inference? ethics? protocol? This prevents distractor-driven reasoning.

Memorise null values: difference=0; ratio=1. Remember sampling ≠ randomisation, allocation concealment ≠ blinding, prevalence ≠ incidence, SD ≠ SE, statistical significance ≠ clinical importance.

For calculation questions, write the formula and identify numerator/denominator before inserting numbers. For design questions, identify how participants entered the study and when exposure/outcome were measured.

Use the 100-MCQ bank after each phase, then revisit incorrect concepts. The goal is not answer recall but rapid recognition of the principle being tested.