Causality is the relationship between cause and effect.
- The principle of causality is that all events have a cause.
- Indicates a logical relationship between 2 events (a cause and an effect) and an order between them (the cause precedes the effect)
- In medicine, establishing causality:
- Helps identify the cause of a disease
- Enables the best possible management of the patient
- Allows researchers to develop the best diagnostic tests
Causality versus correlation
“Correlation is not causation.”
- Causality means that 1 event was caused by another vent
- Correlation (or association) means that 2 things are connected, but it does not imply causality.
Below is a graph depicting the relationship between drowning deaths and eating ice cream. As ice cream consumption goes up, so do drowning deaths. However, this study is only showing a correlation rather than causation. Eating ice cream does not cause drowning deaths. Rather, on hot days, people are more likely to eat ice cream, and they are more likely to go to the beach and drown. Thus, temperature is a confounding factor, leading to an observed relationship when in reality there is no causality.
Bradford Hill criteria
- Also known as Hill’s criteria for causation
- A group of 9 principles useful in establishing epidemiologic evidence of a causal relationship between a presumed cause and an observed effect
- If a majority of the principles are satisfied, the relationship between variables is likely causal.
- Widely used in public health research
The 9 principles:
- Strength: What is the effect size or how big is the relationship? How large is the relative risk or odds ratio?
- Consistency: Do I see the relationship frequently? Is it reproducible?
- Specificity: Does this exposure exclusively cause this outcome?
- Temporality: Does the exposure come before the outcome?
- Biological gradient/dose-response relationship:
- Does more of the exposure cause more of the outcome?
- Does removal of the exposure decrease risk of the outcome?
- Plausibility: Does the relationship make sense biologically?
- Coherence: Does the observed relationship fit into the general knowledge of science and medicine?
- Experiment: Can a randomized controlled experiment be conducted (in humans or animals)?
- Can we devise an analogous relationship between the outcome and another exposure?
- Have I explored other theoretical possibilities for the observed relationship?
Example: Applying the Bradford Hill criteria to establish causality
As seen in the table below, a majority of the principles are satisfied, so you can be reasonably sure that smoking actually causes lung cancer.
|Strength||Yes||There is a strong relative risk (RR) association between smoking and lung cancer.|
|Consistency||Yes||This ↑↑ RR has been reproduced across many cohort studies|
|Specificity||No||Smoking can lead to many different outcomes, and other exposures can also lead to lung cancer.|
|Temporality||Yes||Smoking precedes the development of lung cancer in the vast majority of cases.|
|Biological gradient||Yes||The more you smoke, the higher your RR of lung cancer.|
|Plausibility||Yes||In the lab, it has been shown that lung tissue exposed to the carcinogens found in cigarette smoke shows an increase in genetic mutations.|
|Coherence||Yes||Certain chemicals within cigarette smoke are carcinogens and thus ↑ risk for lung cancer: this idea fits in with our larger understanding of medicine and science|
|Experiment||Yes||We have exposed laboratory animals to smoke, and they have developed cancer.|
|Analogy||Not really||Other options have been explored, and there may be other potential possibilities.|
A causal relation between 2 events exists if the occurrence of the 1st event causes the 2nd event.
- The 1st event is then referred to as a cause and the 2nd event the effect.
- A correlation between 2 events does not imply causation.
- However, if there is a causal relationship between 2 events, they will be correlated.
- A causal pathway (the pathway from the cause to the effect) can be:
- Direct: The factor causes the disease without intermediary steps.
- Indirect: The factor causes the disease but only through 1 or more intermediary steps.
Types of causal relationships
There are 4 types of causal relationships or factors based on whether or not the exposure was necessary to develop the outcome, and whether exposure is sufficient on its own to cause the outcome. These 4 types are:
- Necessary and sufficient
- Necessary but not sufficient
- Sufficient but not necessary
- Neither necessary nor sufficient
Example #1: Necessary and sufficient
- Necessary: It is impossible to have the outcome without the exposure.
- Sufficient: It is all that is needed to produce the outcome.
- Example: Infection with coronavirus is both necessary and sufficient to cause the disease SARS.
Example #2: Necessary and not sufficient
- Necessary: The exposure is required to develop the outcome.
- Not sufficient:
- The exposure needs to be aided by some other factor in order to produce the outcome.
- Individual factors cannot produce the outcome by themselves.
- Example: A disease is caused by a gene that becomes activated by a particular environmental trigger. Both the gene and environmental trigger are necessary for the disease, but neither is sufficient alone to cause the disease.
Example #3: Sufficient and not necessary
- Sufficient: The exposure alone can produce the outcome.
- Not necessary: The exposure is not the only one that can produce the outcome.
- Example: Both radiation poisoning alone and benzene poisoning alone are sufficient to cause leukemia. Therefore, it is not necessary to have radiation exposure in order to develop leukemia (because you could have been exposed to benzene instead). Thus, radiation poisoning and benzene poisoning are both sufficient but not necessary for the development of leukemia.
Example #4: Neither necessary nor sufficient
- Not necessary: Several factors or exposures have complex interactions that produce the outcome.
- Not sufficient: The individual exposures are not enough to produce the outcome alone.
- Example: prostate cancer. There are multiple risk factors that individually are neither necessary nor sufficient to cause prostate cancer alone. There are multiple combinations of exposures possible, making them all neither necessary nor sufficient.
- This is arguably the most common type of causal relationship encountered in clinical practice.
Reliability and Validity
Reliability refers to the reproducibility of a test or research finding: Is the test or finding repeatable?
- Reliability = reproducibility/consistency/precision
- ↑ Reliability leads to ↓ random error rates
- Reliability improves with ↓ standard deviation and ↑ power
Validity refers to how accurate a test or research finding is: Are the results representative of the real world?
- Validity = accuracy
- Results reflect reality and can be believed.
- Validity depends on the elimination of biases.
- Sensitivity and specificity are measures of validity.
- Internal validity:
- The causal relationships are meaningful within the context of the study.
- Requirements for internal validity:
- External validity or generalizability: Results can be applied to other patients or settings.
Note: An invalid study can still be reliable, but an unreliable study cannot be valid. In other words, a relationship that does not represent the real world (invalid) can be seen multiple times in a study (reliable), but a study that cannot be reproduced (unreliable) cannot represent the real world (validity).
- Reliability is measured quantitatively with a coefficient, typically written as r.
- R is valued between 0 and 1:
- r = 1: perfectly reliable test
- r = 0: complete absence of reliability
- ↑ r = ↑ reliability = ↓ errors
- Investigators typically want r at least 0.9, which means 90% of the data are accurate while 10% is caused by errors.
Threats to reliability and validity
Threats to reliability:
- Poor sampling strategies
- Example: You want to measure the average age in a community. You might go to a retirement center and find an average age of 72 or go to a high school and find an average age of 16. These samples are not representative of the population you are trying to study, so your data are not reliable.
- Instability in the thing being measured
- Example: You want to measure blood pressure, but blood pressure changes throughout the day, based on factors such as activity level and how horizontal you are (lying down versus standing). If these confounding variables are not accounted for, your data will not be reliable.
- Divergences between observers, especially in cases where data collection requires a qualitative assessment by the observer
- Example: If you’re asking observers to assess mood, different observers may have different opinions about how to score different moods. This would decrease reliability.
Threats to internal validity:
- Confounding factors: a variable that creates an artificial relationship or that masks a real relationship between study variables (see the example of ice cream and drowning deaths above)
- Selection bias: the error introduced when the study population does not represent the target population due to some selection preference (see Types of Biases for details)
Threats to external validity:
- Too many exclusion criteria (overly specific study characteristics that do not represent other populations)
- Hawthorne effect (observer effect): People in studies change their behavior because they are being watched.
- Rosenthal effect: The investigator’s expectations about the outcome of a given study affect the actual study outcome.
- Celentano, D, Szklo, M. (2019). From association to causation: Deriving inferences from epidemiologic studies.
- Everitt, BS, Skrondal, A. (2010). The Cambridge Dictionary of Statistics, Cambridge University Press.
- Redmond, CK, Colton, T. (2001) Biostatistics in Clinical Trial, 2001: p. 522.
- Greenberg, RS. (2015). Medical epidemiology. In Population Health and Effective Health Care (5th ed.).
- Kanchanaraksa S. (2008). Evaluation of diagnostic and screening Tests: Validity and reliability. In The Johns Hopkins University Bloomberg School of Public Health.
- Höfler, M. (2005). The Bradford Hill considerations on causality: A counterfactual perspective? Emerging Themes in Epidemiology, 2 (1): 11.