· Null hypothesis: There is no difference between the two groups being analyzed
· P-value: The probability of obtaining the observed result—or one more extreme—if the null hypothesis were true. A small P-value would make is reject the null hypothesis.
· Confidence interval (CI): Quantifies the uncertainty in measurement. It is usually reported as a 95% CI, which is the range of values within which we can be 95% confident that the true value for the whole population lies.
· Power: The probability of detecting a statistically significant difference between the control and experimental groups. Power = 1 – β
· Type 1 Statistical error (α): The probability of rejecting the null hypothesis, when it is actually true.
· Type 2 Statistical error (β): The probability of accepting the null hypothesis, when it is actually false.
· Confounder: A third variable, associated with both the exposure and the outcome, does not mediate their relationship, and should be adjusted for when analyzing exposure-outcome relationship (example: smoking for the coffee drinking—cardiovascular risk association). It must be considered in small RCTs. We can never rule out some residual confounding in observational studies (case-control, longitudinal cohort).
· Effect modifier: A third variable, which is NOT associated with either the exposure or the outcome, but changes significantly the exposure-outcome relationship (example: smoking for the contraceptive—thrombosis association). Effect modification is also known as interaction.
· Intention to treat analysis (or “analyze as randomized”): We analyze outcomes based on the initial randomized group assignment, regardless of whether participants received the corresponding intervention. This maintains the random treatment assignment given at the outset of the study, and thus preserves the study from bias.
· Relative Risk: The ratio of risks between an experimental and control group, in cohort and RCT studies.
· Absolute Risk Reduction: Absolute difference between the exposed (experimental group) vs. the unexposed group (control group); used to estimate the Number Needed to Treat.
· Relative Risk Reduction: The proportion of the baseline risk reduced by the intervention (Absolute risk reduction/absolute risk in control population)
· NNT (number needed to treat): 1/ARR. The number of patients needed to treat to avoid one bad outcome (or to prevent one event).
· NNH (number needed to harm): 1/ARI (where ARI is the absolute risk increase) The number of patients needed to expose to cause one additional bad outcome.
· Sensitivity: The probability of a positive test in people with disease
· Specificity: The probability of a negative test in people without disease
· Positive Predictive Value: Proportion of people with a positive test who truly have the disease
· Negative Predictive Value: Proportion of people with a negative test who truly do not have the disease
· Likelihood ratio of a positive test: Probability of an abnormal result in the population with the disease, divided over the probability of that result in the population without the disease. LR (+) of 6 means that this test result is 6 times more likely to occur in a patient with the disease than in a patient without the disease. LR(+) = Sensitivity/(1-Specificity)
· Likelihood ratio of a negative test: The probability of a normal result in the population without the disease, divided over the probability of that result in the population with the disease. LR(-) = (1- Sensitivity)/Specificity
· Censoring: In RCTs, censoring means that follow up length is not the same for all surviving subjects, either because of losses to follow-up, or because the study ended and those randomized last were observed for a shorter time. Survival methods (Kaplan Meier curves, log-rank tests, Cox models, etc.) take censoring into consideration when comparing two groups.
· Kaplan-Meier Curve: (survival curve) It usually plots the cumulative probability of survival in each study arm, starting from 100%, although it may plot cumulative risk.
· Cox proportional Hazards Model: Analyzes time-to-event data, comparing hazard functions between arms. Hazard = risk per unit of time. It applies the principle of censoring, and allows adjustment for potential confounders.
· Case-Control Study: Examines associations between exposure and outcome by sampling subjects with the outcome of interest, matching them against subjects without the outcome, and then comparing the exposure in each. Most important concern is bias, such as recall bias.
· Longitudinal Cohort study: Measures potential exposures at baseline, and then follows up subjects to determine the occurrence of outcomes by exposure group. Non-randomized design may cause bias, including confounding by indication.
Randomized Controlled Study (RCT) Checklist
· Who were the study participants? • Was randomization adequate?
· Was it blinded? • How many were lost to follow-up? • How many crossed over? • Were all patients who entered the trial accounted for at its conclusion? • Was it analyzed on an intention to treat basis?
· What was the number needed to treat? • What were the adverse effects?
· Do the results apply to your patients?
Negative RCT Checklist
· Did they recruit the number of patients they intended to?
· Was the event rate in the control group as high as expected?
· Were losses to follow up and cross-overs as infrequent as expected?
· Was there intervention fidelity?
ŕIf the answer is “NO” to any of the above questions, the possibility of lower than expected power should be considered. But remember: if the treatment were efficacious, there should be at least a trend towards improvement in the intervention arm.
· Placebo-controlled design is not ethical because a treatment is available. The new treatment is expected to have similar efficacy, and have better safety, convenience, tolerability, acceptability, or cost.
· The threshold for non-inferiority significance (null hypothesis line) should be clinically and ethically appropriate. Monitoring and pre-specified stopping rules are also essential.
· The test statistic is the same as for superiority analysis (e.g., Relative Risk), but we examine the one-sided confidence interval (CI) in the direction of possible harm.
· The null hypothesis line represents the amount of harm we are willing to accept included in the one-sided CI. We reject the null hypothesis if the CI does not include that line.
· Losses to follow up, cross-overs, and low intervention fidelity increase the probability of type 1 error (as opposed to superiority RCTs, in which they increase the prob. of type 2 error). Quality of care in control arm should be state-of-the art. As treated analysis must be included.