Section XII: More Observational Studies

What is a case-crossover study?

You compare a person to themselves at another time to see if a short-term exposure triggered their event.

Why compare a person to themselves?

So that stable characteristics (sex, SES, genetics) can’t confound the results.

What is a case-crossover study used for?

To test whether a brief, transient exposure triggers an acute event.

It’s perfect when you believe the exposure acts like a momentary trigger, not a long-term cause

What conceptual question does a case-crossover study answer?

“Did something that happened shortly before the event cause it to happen at that moment?”

Why does case-crossover remove all time-invariant confounders?

Because each person is compared to themselves.

Anything about them that doesn’t change (sex, SES, baseline health, personality, genetics) cancels out.

This is why it’s used when individual-level confounding is hard to measure.

What is the key conceptual risk in a case-crossover study?

Time-varying confounding.

Exposure might follow time patterns (weekday vs weekend, seasons), which can create confounding.

Why use case-control instead of a cohort?

Because the outcome is rare, the follow-up period would be too long, or because exposure data must be reconstructed from historical sources.

This is a study of chronic, not acute, causation.

Why can’t standard regression answer multilevel questions?

Standard regression assumes independence.

But people within the same environment are not independent, leading to false precision and incorrect inferences.

Why is temporality the core conceptual limitation of cross-sectional studies?

Because measurement occurs at a single time point, cross-sectional studies answer prevalence questions but cannot answer “which came first?”

Thus they cannot test causal direction.

What is the “hazard window” in case-crossover studies?

The short time immediately before the acute event where exposure is measured.

Why does case-crossover automatically control for sex, race, SES, genetics, and personality?

These characteristics do not change across time windows; self-matching removes them as confounders.

Why must the control window be chosen carefully in a case-crossover design?

Time-varying confounders (season, weekday patterns, stress cycles) can bias results if windows are not comparable.

What exposures are appropriate for case-crossover designs?

Exposures that fluctuate over time and could trigger acute events quickly (pollution spikes, temperature extremes, alcohol binges).

Why are chronic exposures a poor fit for case-crossover studies?

Chronic exposures do not vary meaningfully across windows, so there is no exposure contrast within the same person.

Key limitation of case-crossover vs case-control?

Case-crossover cannot estimate prevalence; it only tests short-term triggers of acute events.

In MI–air pollution studies, what is the exposure and what is the outcome?

Exposure = short-term pollution level; Outcome = myocardial infarction.

In case-crossover, is MI the exposure or the outcome?

MI is the outcome; exposures are measured before it occurs.

What measure of association is typically used in case-crossover?

A matched (conditional) odds ratio.

What is the ecologic fallacy?

Incorrectly inferring individual-level associations from group-level data.

Example of ecologic fallacy from lecture?

States with high alcohol use also have high depression rates, but this does not mean individuals who drink are depressed.

Why can’t ecologic data infer individual-level causal effects?

Group-level associations may be confounded by unmeasured contextual variables.

What problem do multilevel models solve?

They separate individual-level from group-level effects in nested data structures (students in schools, patients in clinics).

When do we need multilevel modeling?

When observations are clustered and not independent, causing biased estimates and incorrect standard errors.

What bias occurs if nested data are analyzed as independent?

Standard errors become too small, leading to false statistical significance.

What is the purpose of a DAG?

To visualize causal assumptions and identify which variables should be adjusted for (confounders) and which should not (mediators, colliders).

How does a confounder appear on a DAG?

Arrows point from the confounder to both the exposure and the outcome.

How does a mediator appear on a DAG?

Exposure → mediator → outcome.

How does a collider appear on a DAG?

Exposure → collider ← another variable.

Which must be prioritized: internal or external validity?

Internal validity, because results must be accurate before being generalized.

Why can cross-sectional studies not establish temporality?

Exposure and outcome are measured at the same time → impossible to know which came first.

Why does case-crossover always establish temporality?

The exposure window is explicitly defined before the acute event occurs.

How does matching in case-control differ from case-crossover matching?

Case-control studies match different individuals, while case-crossover studies match a person to themselves.

Does matching eliminate confounding in case-control studies?

It controls confounding for the matched variable, but not for unmeasured confounders.

Why is matching unnecessary in case-crossover studies?

Self-matching already controls all time-invariant confounders by design.

When is case-crossover inappropriate?

When exposures do not vary, or when outcomes do not have clear, short induction periods.

What happens if exposure has weekly or seasonal cycles?

Time-related confounding biases the association if control windows do not match hazard windows.

What is a “referent window” in case-crossover?

A control time period where the event did not occur; exposure during this period is compared to the hazard window.

Why is referent-window selection crucial?

Poor window selection introduces time-varying confounding and distorts the odds ratio.

What bias is eliminated by self-matching?

Confounding by stable characteristics such as sex, SES, and genetics.

What biases are NOT eliminated by self-matching?

Measurement bias, exposure misclassification, and confounding by time-varying factors.

Why is case-crossover good for suicide attempts or overdose triggers?

These outcomes have abrupt onset and exposures (e.g., stress, intoxication) fluctuate rapidly.

When are ecologic studies the only feasible option?

When exposures operate only at the population level (e.g., state laws, taxation policies).

What key limitation does multilevel modeling acknowledge?

That individuals are influenced by their environments; observations within clusters are not independent.

How does Lecture 12 connect to confounding from Lecture 9?

Case-crossover removes many confounders by design, but DAGs identify remaining time-varying confounders requiring adjustment.