Thread by @realscientists

One of my favorite subjects to talk about: causality and methodologies

Epidemiology is all about measurement and causal inference. Sociology is more theoretical but it does touch on methodologies quite a bit.

Epid is bound by its methods. Sociology is bound by its theories. So to conduct epid studies, math or stats makes the most sense. Sociology can use any method so long as the method makes sense to the study

I taught two semesters of sociological research methods. So I've become really interested in the social critiques of epid methods and vice versa. Let's get it

Identifying Associations between variables:

1. Need to compare at least two groups

2. Comparison groups are defined as classes (categories, levels, locations) of one variable

3. Comparison between groups is made with respect to their values of a second variable

4. One of the variables should be identifiable as exposure and the other as an outcome

Describing Associations:

1. Identify exposure and outcome variables

2. Use attributes
a. direction/shape
b. “increase” / “decrease” / “change” may imply causation

c. Magnitude (strength)
d. Absolute vs relative difference
e. Precision (statistical significance)

An association between two variables can be because of:

1. Lack of exchangeability (the groups being compared are too different)

2. Bias (e.g. another factor is impacting the results)

3. Random chance

4. Causality

What is causal inference?

The goal of determining whether an exposure or treatment (A) causes an outcome (Y)

Causality can only be “confirmed” through longitudinal (cohort) studies

Measuring Causality: The Counterfactual

Counterfactual: the outcome that would have been observed under an exposure value that the person did not actually experience (“counter to the fact”).

What do we need to know to determine whether marijuana smoking (exposure, A) causes Alzheimer's (outcome, Y)?

1. How many of these people would have developed Alzheimer's had they not smoked Marijuana?

Counterfactual theory assumptions:

Well‐defined interventions: Only one version of exposure (treatment) as opposed to loosely defined

No interference: a subject’s counterfactual outcome does not depend on other subject’s treatment (may not hold for infections)

Deterministic: ignores quantum mechanics; assumes action (A) always results in the same value of outcome (Y) with probability=1

Real Life:

Individual causal effects cannot be determined: causal inference is a missing data problem

We need another definition of a causal effect that requires weaker assumptions

Average causal effects:

When the exposure is randomly allocated, the two groups are exchangeable. One represents the counterfactual exposure experience of the other, on average.

Average causal effects can be estimated from studies under:

No assumptions: ideal randomized studies (clinical trials)

Strong assumptions: observational studies

Measuring Average Causal Effects:

How many of these people would have developed Alzheimer's had they not smoked Marijuana?

Compare marginal (unconditional) probabilities:

Probability Alzheimer’s would have been observed if the person had not smoked marijuana

vs

Probability Alzheimer’s would have been observed if the person had smoked marijuana

So now let's get into the social limitations

What is the Exposure and Outcome?

"What is the effect of race on educational attainment?"

Compare marginal (unconditional) probabilities:

Probability educational status would have been observed if the person was white

vs

Probability educational status would have been observed if the person was non-white

Essentially we are asking:

“What would a non-white person’s education be like if they were white?”

How is this helpful to policy?

Non-Manipulable Exposures:

1. Race is not something we can intervene on

2. The associated counterfactuals tend to be meaningless

3. We cannot create an intervention that changes the person’s race

4. Is research useful if we cannot create an intervention/policy/treatment?

Most of the time we are looking for the effects of racism (not race)

Race is often used as a proxy for other effects such as:

Neighborhood income
School quality
Labor market stratification etc.

Effects may vary by how we define race:

Skin color
Parental skin color and its perception
Cultural context
Genetic background