November 20, 2024
FPCI \(\to\) we can’t know whether some \(X\) causes \(Y\) to change for any individual case.
Correlation as a solution is prone to a bias: confounding.
We can use correlation as evidence of causality and plausibly solve confounding, if we make some assumptions, and we can defend those assumptions.
Story above not false, but misleading:
Blatantly false or misleading information about COVID-19 and COVID vaccine circulated widely on social media See here.
Even public officials circulated misinformation.
Misleading information can affect behavior, including reducing willingness to be vaccinated.
Beyond changing policies, does giving a platform to vaccine skeptics make people less likely to be vaccinated?
What if we look at social media users:
This is a common source of confounding:
Sometimes confounding called “selection effects”.
Does knowing transgender people increase support for funding gender-affirming care?
In the 2023 AMS election, UBC students were asked vote on whether to increase student fees by $8 to cover gender-affirming care.
We imagined both potential outcomes in a prior lecture (and answered a survey about it!)
If you had a friend on campus who was trans/non-binary, would you have voted “yes” or “no” on this measure?
If you did not have a friend on campus who was trans/non-binary, would you have voted “yes” or “no” on this measure?
If we just examined the correlation between having trans/non-binary friends and support for increasing AMS fees for gender affirming care…
What could be a source of confounding?
What if we could do this:
What if Meta (the monopolist corporation formally known as Facebook) conducted a test of a new algorithm that classified vaccine misinformation in social media posts and shared links?
What if we could do this:
FPCI: We can never know the causal effect of \(X\) on \(Y\) for a specific case.
Correlation of \(X\) and \(Y\) for different cases may suffer from confounding
Under certain assumptions, we can treat correlation as an inference (or estimate about) the average causal effect of \(X\) on \(Y\).
Experiments give us unbiased (no confounding) average causal relationship between \(X\) and \(Y\), if two key assumptions are met:
\(^*\)Technically, there are other assumptions, but not important for this class
Cases in “treatment” and “control” are the same in terms of potential outcomes, on average:
assumption is that we aren’t adding confounding in the design of the experiment
If in the Vaccine misinformation experiment, “Treatment” group
Multiple differences between “treatment” and “control”
Experiments are the best solution to confounding/FPCI
strong severity says that evidence is convincing to extent assumptions are checked
But experiments have their limits:
BC CDC reports, e.g. hospitalization rates between those who are vaccinated vs. unvaccinated
What could confound this correlation? (to the board)
Clinical trials are experiments. Correlation between treatment and health outcomes are causal (do not have confounding) assuming that…
TO THE BOARD
Why does it matter that clinical trials use placebos and are “double blind”?
Vaccine clinical trials…
These design features of the experiment ensure that only difference between “assigned to treatment” and “assigned to control” is the vaccine.
We can know whether vaccination caused reduction in:
or caused increase in:
Might appear to be the only valid solution:
Experiments, under easy-to-check assumptions, let us find an unbiased causal relationship between \(X\) and \(Y\) using correlation
Internal Validity
A research design (choice of which cases to compare using correlation) has internal validity when the correlation of \(X\) and \(Y\) it finds is the true causal effect of \(X\) on \(Y\) / does not suffer from confounding. (unbiased FOR THOSE CASES)
Does the efficacy of vaccines in clinical trials translate to real world use??
What can we manipulate?
Who/what cases can we study?
External Validity
is the degree to which the causal relationship we find in a study is relevant to the causal relationship in our causal question/claim
Study has external validity if the relationship found is true for the cases we are interested in
Study has external validity if the causal variable in the study maps onto the concept/definition of the cause in the causal claim.
More internal validity (unbiased estimate of causal effect) comes at the cost of external validity (relevance of study sample or cause to the theory)