Core research concepts

Complete cases or multiple imputation?

Choose a missing-data strategy from the mechanism and estimand, not a percentage rule.

Prediction leakage, overfitting and optimism

Keep information boundaries intact when developing and evaluating a prediction model.

From research question to estimand

Turn a broad clinical idea into a defined target quantity before selecting the study design and analysis.

Time zero, index date and follow-up

Align eligibility, exposure assignment and the start of outcome follow-up so the comparison answers the intended question.

Confounding

When groups differ in causes of the outcome, the comparison can mislead.

Effect modification and interaction

When the size of an effect differs across levels of another variable.

Directed acyclic graphs

A diagram of assumed causal relationships that guides what to adjust for.

Missing data

Why values are missing matters more than how many are missing.

Sample size and statistical power

How much information the study needs, given what it is trying to estimate.

Simulation-based power and sample-size planning

How to design, run and report a reproducible power simulation when a standard formula does not fit.

Why events per variable is a floor, not a sample-size target

A model-capacity check can expose an under-sized analysis but cannot establish its power or precision.

Multiple testing and multiplicity

Testing many questions raises the chance of a false positive unless the strategy accounts for it.

Confidence intervals and p-values

What each one actually says, and what neither of them says.

The estimand

The precise statement of what effect you are trying to estimate, settled before the analysis rather than implied by it.

The Table 2 fallacy

In a table of adjusted coefficients, only the exposure row estimates the effect you designed the study to measure.

Selection bias and colliders

Choosing or keeping people on the basis of something both the exposure and the outcome affect creates an association that is not there.

Immortal time bias

A stretch of follow-up during which the outcome could not have happened, quietly credited to the treated group.

Clustering and the ICC

People in the same ward or clinic resemble each other, so they carry less independent information than their number suggests.

Calibration and discrimination

Two different things a risk model can be good or bad at, and only one of them is usually reported.

Generalizability and transportability

Whether a result found in one group of people applies to a different group, which is a question about effect modifiers rather than about representativeness.

Measurement error and misclassification

Variables are measured imperfectly, and the direction the resulting bias runs is not the one most people assume.

Regression to the mean

Extreme measurements tend to be less extreme next time, for reasons that have nothing to do with any intervention.

Temporality and reverse causation

The cause has to come first, and in observational data it often turns out not to have.

The ecological fallacy

An association between group averages need not hold, or even point the same way, for individuals.

Non-collapsibility

An odds ratio changes when covariates are added even with no confounding at all, which is a property of the measure and not a sign of bias.

Information bias

Bias arising from how data were gathered rather than from who was studied.

Absolute and relative effects

The same result expressed as a ratio and as a difference can sound entirely different, and only one of them tells a patient what to expect.

Number needed to treat

How many patients must be treated for one to benefit, which is an absolute effect turned upside down.

Prevalence, incidence and person-time

How often disease is present, how often it appears, and why the denominator decides which question you answered.

Surrogate outcomes

A measurement standing in for the outcome that matters, which is faster to observe and frequently misleading.

Composite outcomes

Several events counted as one, which buys statistical power and can hide what actually happened.

Patient-reported outcomes

Outcomes reported by the patient without a clinician's interpretation, which need the same rigour as any other measurement.

The minimal clinically important difference

The smallest change a patient would notice and care about, which decides whether a statistically significant result matters.

Pre-test probability and likelihood ratios

What a test result means depends on how likely the condition was before the test was ordered.

Randomization, concealment and blinding

Three separate protections that are routinely confused, and only one of them is what makes a trial randomized.

Loss to follow-up and attrition

People leave studies for reasons connected to their outcome, which is what makes their absence a bias rather than a loss of numbers.

Prespecification and the analysis plan

Deciding what will be analysed before seeing the data, which is what makes a p-value mean what it claims.

Registration and reporting bias

The published literature is a biased sample of the research that was done, and registration exists to make that visible.

Reproducibility and analytic transparency

Whether someone else, given the same data, would get the same answer, and whether anyone could check.

Ethics review and consent

What makes research on people permissible, of which informed consent is one requirement among several.

Privacy and de-identification

Removing names does not make data anonymous, and the risk depends on what remains rather than on what was deleted.

Equity and representation

Who gets studied determines who the evidence serves, and both exclusion and extraction are failures.

Choosing a study design from the research question

Match the design family to what the question asks, how the exposure arrived and what the data can support.

Descriptive studies and precision-based planning

Plan a descriptive estimate around how precise it needs to be, rather than around a test nobody is running.

Risk of bias and certainty of evidence

Judge whether each result could be systematically wrong, then how much confidence the whole body of evidence supports.

Planning when the sample size is already fixed

When the data already exist, judge what the fixed number can support before analysis instead of computing power afterwards.