Analytic Study Design

These are structured around comparison. Descriptive studies are about what exists. Here you are asking why these differences exist and whether they are compatible with causal interpretation. Analytical inference is way more rigorous for this reason.

They do not automagically establish causation though! You can do this only if the underlying assumptions are defensible (“is temporality established” “are we doing an effect measure rather than a frequency” “how did you do your cohort”?) Define → Measure → Justify

Cohort Studies

See Ch.7 of the Rothman book.

Important in understanding casuality. You follow people over time (this is how they are different than cross-sectional studies). People are classified based on their exposure. So,

Select on Exposure -> Wait for Outcome

What you’ve done here is enforced temporality. You always do this in cohort studies.

Note that hypotheses for analytical studies can come from descriptive studies lol. You could use these to esablish the exposure. So what about the outcome? This should align with the scientific question that we have in mind (the hypothesis).

These have the most information about time.

Retro and Prospective

What you think they are. But the logic remains the same! The most important thing is the preservation of temporality.

You generally get much larger studies from retrospective studies (see chapter on how cohort studies are generally pretty expensive).

Loss to follow-up is a big problem in prospective studies. Think about long onsets and rare diseases.

Note: There’s also Ambidirectional Cohort Studies! They’re rare and also called “Callback Studies”. You’re basically ‘welding’ two timelines together to tell a better story (esp with temporality). Data inconsistencies/management are challenging. You get all the bummers of retro and prospective studies.

Bias in Studies

Think of the contingency table.

	O+	O-
E+	a	b
E-	c	d

You are talking about misclassifying when a → b or c → d or vertically.

• There are non-differential misclassifications when the misclassification is the same in exposed and underexposed groups OR in cases and non-cases (outcomes). This distortion is uneven! Here, the misclassification biases towards the null. This is slightly better?
• There is differential misclassification where there is some preferential ‘direction’ of misclassification; it differs between groups. It’s what you think. You mess up the causal inference in any directions.

Exposure ----(Systematic/Structural Distortion = Bias)----> Outcome

Think about the contingency table. With bias, two things are affected (a) who gets in the table (selection bias)? (b) how and where are they placed in the table (information bias)?

Think of how attrition (people leaving study) can be related to exposure or outcome risk. This is what we’re talking about (if you lose someone they do not fulfill the study design requirements).

The key biases are: selection (when we get people into a study: Loss-to-follow-up is a subtype, retrospective more prone), information (allocate people into the cells in the contingency table: retrospective studies more prone), recall (retrospective studies more prone). Just think through how you’re recruiting people, the time horizon, disease rarity, etc when you decide which type of study is more prone.

TODO

• Induction period versus follow-up period?
• Non-differential misclassification