access_time published 03.08.2017

Confused by clinical research methods? Get a clue!

Thomas Perneger


Confused by clinical research methods? Get a clue!


Clinical research used to be simple, a relaxing diversion after a long day at the surgery. A doctor reviewed the charts of all three patients in his or her practice who suffered from an exotic ailment, computed a proportion or two (“a third died”), took photographs of a Petri dish and of a face rash, added a black bar to cover the eyes of Exhibit 2 – done.

That was then. Today researchers are like wedding planners – they obsess about inclusion criteria, cover up adverse events before anyone notices, lose sleep over budget overruns, and miss out on all the fun. Readers of medical journals are not spared. Articles are full of jargon, and it’s not medical Latin like in the old days. Especially maddening are statistical concepts, disguised as single letters (p, r, kappa), cute metaphors (neural network, bootstrap), or outright lies (normality, significance, error). It’s no wonder doctors only read GetWithIt™ updates anymore! To remedy this situation, Swiss Medical Weekly proposes here an interactive learning tool that will turn you into an instant expert in current research methodology.

To proceed, consider each clue, close your eyes, and let free associations flow. Soon a statistical or epidemiological phrase will occur to you. Do not over-think it. For instance, if the clue is “Final ranking at a beauty pageant?” a valid answer is “Miss classification”. Get it? Misclassification! Don’t worry about sloppy spelling or awkward grammar. Click on “show answer” to verify that you got it right. If you are unfamiliar with the scientific concept, read the comment (if you don’t get the joke, you are on your own).

Let’s play!

Miss classification

(misclassification) Error in the measurement of a risk factor or a disease, which typically weakens the observed association (misclassification bias)

Case-control design

Type of study in which patients who have a disease are compared to controls who don't

Intention to treat

Analysis of a clinical trial that includes all participants as initially randomised (regardless of actual treatment received)

Relative risk

Ratio of risks in the exposed and the unexposed

Log transformation

Application of the logarithm function


Statistical synthesis of the results of several studies that address the same research question

Normal function

Bell-shaped probability density distribution; synonym of Gaussian distribution

Mean square error

In analysis of variance, the sum of squared differences between observed and predicted values divided by the degrees of freedom

Odds of exposure

Odds are a function of a proportion, namely the proportion p divided by its complement (1-p). Exposure is synonym of risk factor.

Psycho metric

(psychometric) Related to classic methods for the measurement of unobserved attributes (often psychological variables)

Forward selection

Method for selecting covariates for a multivariable model, which starts with the best univariable predictor, then adds the second best, etc.

Cost utility

Type of economic analysis where costs incurred are related to health utility gains

Two by two table

Common tool in epidemiology; e.g. tabulation of exposure by outcome

Uniform distribution

Distribution that assigns equal probability (or density) to all values on a given interval

Split-half validation

Validation method where a model is built on one half of the data and then applied to the other half


Probabilities of the observed result or more extreme results under the null hypothesis

Censored events

In time-to-event analysis, observations for which the event of interest has not occurred at the time of analysis

Composite endpoint

Occurrence of any of several events that are considered equally relevant and therefore combined in a clinical trial

Parsimonious model

Statistical model that includes as few covariates as possible while explaining as much as possible of the outcome

Single blind

A clinical trial in which the patients do not know which treatment they receive

Bar chart

Graph that represents by the height of a bar a characteristic of interest (mean, percentage, etc.) for several subgroups

Standard error

Level of uncertainty of an estimator (i.e., a value obtained from the data) due to the limited sample size

Number needed to treat

Measure of treatment effectiveness; the number of persons that must be treated in order to avoid one clinically relevant event

Logistic regression

Modelling method used for binary outcome variables

Monte Carlo simulation

Computer-based method that explores a phenomenon through the simulated behaviour of repeated samples

High outlier

Unusually high value of a variable

Within-group variance

In analysis of variance, variability that is not attributable to group membership


Statistical test used in analysis of variance

Sensitivity and specificity

Ability of a test to correctly identify the diseased (sensitivity) and the non-diseased (specificity)

Confidence limit

Upper or lower bound of a confidence interval, which is the range of values of a parameter that are compatible with the observed data

Discrete event modelling

Simulation technique of a complex process that can lead to one or several distinct events

Shrinkage parameter

In empirical Bayes estimation, the relative weight given to the overall mean when it is combined with the local estimate

p <0.001

Low p-value that typically leads to the rejection of the null hypothesis

Pyramid of evidence

Hierarchy of scientific evidence based on the trustworthiness of the methods used (meta-analysis of randomised trials at the top and expert opinion at the bottom)

Degrees of freedom

Number of independent bits of information that are used to compute a statistic or estimate a model

Safety monitoring

Procedures for detecting threats to patient safety in a clinical trial

Patient autonomy

Ethical principle that gives patients ultimate say in decisions that affect them, including participation in research

Recall bias

Systematic error that can affect case-control studies, because the cases may remember their exposure history differently from the controls

Representative sample

Sample drawn through a process that guarantees fair representation of the underlying population in the long run (e.g., at random)

Interaction test

Statistical procedure that examines the effect of a variable as a function of another; e.g. a test to compare the effectiveness of a treatment in men versus women

Adaptive design

Design of a clinical trial that incorporates future decisions based on accumulated data

Tripel blind

(triple blind) Triple blinding ensures that the patient, the researcher, and the data analyst are unaware of treatment allocation

Test of significance

Procedure for choosing between two hypotheses (typically “the treatment doesn’t work” vs “the treatment works”) based on observed data

Translational trials

PStudies that test whether basic science advances are applicable to patient care

Normal deviate

Normally distributed statistic or observation expressed in standard deviation units

Marginal conditioning

Treating marginal distributions as fixed when the association between two variables is examined; e.g. the row and column totals are considered fixed when a kappa statistic is computed

Non-informative prior

In Bayesian statistics, the a priori distribution of a parameter that reflects ignorance of its true value

Multiplicity problem

Proliferation of “significant” results due to type 1 errors when a large number of tests is performed

Law of large numbers

Principle that describes the average behaviour of a procedure when it is repeated many times

Forest plot

Graph used in meta-analysis representing the results of several studies, with their confidence intervals, plus the pooled effect

Direct standardisation

Method for removing confounding, where rates of disease observed in strata of the confounder (e.g., in age groups) are applied to a reference population

Research protocol

Document that describes in detail all procedures to be used in a study

Alpha and beta errors

Type 1 and type 2 errors. The former is the probability of rejecting the null hypothesis when it is true, the latter the probability of accepting the null hypothesis when the alternate is true

Null hypothesis

Hypothesis of no association or no effect that the researcher hopes to reject.

Non-inferiority test

Test of the hypothesis that a treatment is less effective than another by a pre-specified amount

Link function

Mathematical function applied to the expected value of the outcome variable in generalised linear models; e.g. logarithm function applied to the number of events in log-linear models

CONSORT statement

Guideline for the reporting of randomised controlled trials

N-of-1 trial

Trial where two treatments are alternated at random within a single patient

Goodness of fit

Measure of the adequacy of a model for a given dataset

Pairwise deletion

Method for dealing with missing data when correlations are obtained for >2 variables, where each correlation is computed on all available pairs

Disclosure of interests

Authors’ disclosures of financial or other links with private entities that may bias scientific conclusions

Two-sided test

Statistical test that leads to rejection of the null hypothesis if the data deviate from the null value in either direction

Second order uncertainty

In modelling, uncertainty about the values of parameters (whereas first order uncertainty concerns values of individual observations)


Ability of a diagnostic or predictive model to correctly categorise patients with and without a characteristic (e.g., a specific disease)

Imputation of missing data

Replacement of missing data by plausible values

Cross validation

(cross-validation) Method for verifying the properties of a predictive model

Citation count

The number of times an article is cited in the scientific literature

Peer review

Evaluation of research articles or protocols by experts in the field to aid the decision to publish the paper or to fund the study

Good clinical practice

(Good Clinical Practice) Regulations for clinical trials established by the International Conference on Harmonization

Residual plot

In regression modelling, differences between observed and predicted values, plotted against predicted values

Range of values

Lowest and highest values of a variable


Thank you for playing along!




Disclosure statement

No financial support and no other potential conflict of interest relevant to this article was reported.

Picture credit

Header image: Aprescindere /

Thomas Perneger

Thomas Perneger, MD, PhD, Division of clinical epidemiology, Geneva University Hospitals, 6 rue Gabrielle-Perret-Gentil, CH-1211 Geneva


Become a blogger

Would you like to submit an op-ed article? Please contact us!

e-mail contact

Verpassen Sie keinen Artikel!