Grimes, K.F. Schulz: An overview of clinical research: the lay of the land, The Lancet, Vol. 359 (9300), pp. 57-61
Hernán MA, Robins JM (2020). Causal Inference: What If. Boca Raton: Chapman & Hall/CRC
“Exchangeability means people who are treated and untreated have the same probability of the outcome happening due to all mechanisms other than treatment” (E. Murray)
RCTs is the only study design for which exchangeability (no confoudning assumption) is expected to hold (unlike observational research)
Exchangeable: 🐈🐈🐈🐈🐈🐈🐈🐈🐈🐈 vs 🐈🐈🐈🐈🐈🐈🐈🐈🐈🐈
Not exchangeable: 🐈🐈🐈🐈🐈🐈🐈🐈🐈🐈 vs 🐿️🐿️🐿️🐿️🐿️🐿️🐿️🐿️🐿️🐿️
vs
vs
In non-interventional setting epidemiologic studies are especially susceptible to bias, including confounding
Non-exchangeability → study compares apples with non-apples and makes erroneous conclusions attributing the found association to the effect of exposure (washing with water alone vs with water + soap) on the outcome (time to getting spoiled), while it is the confounding that can explain the observed association
vs
In non-interventional setting epidemiologic studies are especially susceptible to bias, including confounding
Non-exchangeability → study compares apples with non-apples and makes erroneous conclusions attributing the found association to the effect of exposure (washing with water alone vs with water + soap) on the outcome (time to getting spoiled), while it is the confounding that can explain the observed association
vs
Random allocation to intervention groups
Both groups (active vs control) will be treated identically in all respects except for the intervention
Allocation concealment: personnel performing randomization are prevented from knowing the patients' treatment allocation
Patients and investigators remain unaware of what treatment was given until the study is completed
Patients are often analysed according to the group to which they were allocated, irrespective of whether they actually received the intended intervention (intention to treat analysis, ITT)
When evaluating effect of Z, we evaluate the intention-to-treat effect
The path A → Y is the effect of treatment actually received on the outcome
Total population, N | Random treatment assignment (Z), n | 28-day mortality (Y=1), n | |
---|---|---|---|
1000 | Z=1 (treated) n=482 | 262 | |
Z=0 (comparator/placebo) n=518 | 290 |
Percent Relative Effect: (1-0.97)=0.03=3% decrease in relative risk
Risk difference
Relative risk
The treatment allocation is no longer only dependent on the randomization but also can be dependent on patients' characteristics (L)
Patients outcome is not independent of patient's characteristics (L)
Patients characteristics (L) is the confounder
When computing per-protocol effect, need to adjust for L
Source of selection bias in RCTs
Washing apples with water and soap keeps apples fresh longer vs washing with water alone
Randomization secures exchangeability between the treatment arms
Intention-to-treat (ITT) effect is an effect of the randomization on the outcome
Per protocol analysis aims to investigate the effect of the actual treatment on the outcome
If adherence to the treatment is not perfect, the Intention-to-treat effect ≠ Per-protocol effect
Loss-to-follow-up introduces selection bias to RCT results and needs to be addressed in the analysis
Random error
Systematic error
Internal validity
External validity (= generalizability)
Internal validity → external validity
Not applicable for interpretation of the results of observational studies
H0 - null hypothesis (hypothesis of no association)
Risk difference is zero or the risk ratio is 1
Is a hypothesis of no association between two variables in a superpopulation
The groups we compare were sampled in a random fashion from a superpopulation
Is not about the observed study groups
Rothman, K., Greenland, S., & Lash, TL. (2008). Modern Epidemiology, 3rd Edition. Philadelphia, PA: Lippincott Williams & Wilkins. https://www.nature.com/articles/d41586-019-00857-9
Is the probability that a test statistic (computed using the data) would be greater than or equal to its observed value, assuming that the test hypothesis is correct and all assumptions hold
Results are claimed to be “significant” or “not significant” according to whether the p-value is less than or greater than an arbitrary cutoff value, usually 0.05, which is called the alpha level of the test
The observed difference can be statistically significant
When the model used to compute it is wrong
Due to chance
Dichotomization of study results based on p-values is harmful
A small p-value → the data are unusual if all the assumptions used to compute the test statistics (including the null hypothesis) were correct
When a study is large, very minor effects or small assumption violations can lead to "statistically significant" results of the null hypothesis tests
Rothman, K., Greenland, S., & Lash, TL. (2008). Modern Epidemiology, 3rd Edition. Philadelphia, PA: Lippincott Williams & Wilkins. https://www.nature.com/articles/d41586-019-00857-9
Table1: balance is desirable but may not be perfect
Accuracy = Validity + Precision
Precision = no of random error
Validity = no systematic error
Selection bias + information bias
Internal validity → external validity
Interpret the RCT results concentrating on the validity and precision, but not p-values
“Harms should always be viewed as important whether they are labelled primary or secondary” (CONSORT statement)
Existence and nature of adverse effects
Withdrawal of participants due to an adverse event → loss-to-follow-up (selection bias)
Often unexpected and unpredictable
Harms-related stopping of an RCT
CONSORT (Consolidated Standards of Reporting Trials)
Separately reporting anticipated and unexpected adverse events
For each study arm
The absolute risk of each adverse event, including recurrent
Number of participants withdrawn due to harms
Interpretation
Benefit-risk balance
Not all reported adverse events are necessarily caused by the intervention
Completely unpredictable adverse effects → hard to study in RCT
Post-authorization safety study (PASS) (non-interventional)
Carried out after a medicine has been authorized to obtain data on safety
Intended effect in RCTs are beneficial effects of the interventions/treatments
Intended effect → efficacy
Unintended effects in RCTs are undesired effects, or harms
Adverse effects can lead to loss-to-follow-up (selection bias) in RCT
Importance of reporting
Unpredictable adverse effects → Post-authorization safety study (PASS) (non-interventional)
Grimes, K.F. Schulz: An overview of clinical research: the lay of the land, The Lancet, Vol. 359 (9300), pp. 57-61
Keyboard shortcuts
↑, ←, Pg Up, k | Go to previous slide |
↓, →, Pg Dn, Space, j | Go to next slide |
Home | Go to first slide |
End | Go to last slide |
Number + Return | Go to specific slide |
b / m / f | Toggle blackout / mirrored / fullscreen mode |
c | Clone slideshow |
p | Toggle presenter mode |
t | Restart the presentation timer |
?, h | Toggle this help |
Esc | Back to slideshow |