Day 1 lecture “clean-up” issues

Post on 11-Jan-2016

24 views 0 download

Tags:

description

Day 1 lecture “clean-up” issues. What’s wrong with social epi?. Kaplan Poor theory Individual focus Risk-factor thinking Interdisciplinarity Berkman Too few experiments Poor experimental results. What Causes Disease?. God. Germs. Individual Choice & Behavior. - PowerPoint PPT Presentation

transcript

Day 1 lecture “clean-up” issues

What’s wrong with social epi?

Kaplan• Poor theory• Individual focus• Risk-factor thinking• Interdisciplinarity

Berkman• Too few experiments• Poor experimental results

What Causes Disease?

Germs

God

Individual Choice & Behavior

Social ForcesMarket (failures)

Social NormsDiscrimination

PolicyGovernment Failures

What Causes Disease?

PhysiologyMolec. Bio.GeneticsGerms

BehaviorLife-Style

Choice

Social ForcesMarketsNorms

Racism; Sexism

Fundamental/Upstream

Cause of Disease

Immediate/Proximal

D

Oakes, JM. 2005. "An Analysis of AJE Citations with Special Reference to Statistics and Social Science." AJE 161:494-500.

b) Methodological challenges?

i) Outcome measure?

ii) Proximity = Exposure?

iii) Exposure = Disease?

iv) Social stratification

v) Multilevel phenomena

Group: time 1 Group: time 2

Individual: time 1 Individual: time 2

Source: Oakes JM, JS Kaufman. 2006. "Introduction: Advancing methods in social epidemiology." Pp. 1-18 in Methods in Social Epidemiology, edited by Oakes and Kaufman. San Francisco: Jossey-Bass / Wiley.

v) Multilevel phenomena

3) Counterfactuals

or

Potential Outcomes Model

or

Rubin’s Model

• Idea comes from philosopher David Hume (1711-1776)

“…but for…”

• Advanced by philosopher David Lewis in 1973

“If kangaroos had no tails, they would topple over…”

• Statistician Don Rubin advanced ideas in statistics and epi

Potential outcomes model

• Recent work takes “closet possible world” assumption seriously

Just what are limits to comparative inference?

Ideal is to compare same unit under two scenarios; difference in outcomes is causal effect attributable to scenario.

Let’s compare Ivan (and only Ivan) in environment with and without McDonalds, with all else exactly the same. Difference in Ivan’s BMI is effect of McDonald’s on Ivan.

Trouble is, we cannot observe Ivan under both scenarios. Ivan either lives in an environment with McDonalds or he doesn’t.

The scenario which Ivan does not actually live in is counter to fact, or the counterfactual. Rubin calls same the “potential outcome for Ivan.”

In order to have actual data, we must find a substitute for Ivan’s unobservable counterfactual state. Finding a credible counterfactual substitute is the crux of all sound causal inference.

In fact, randomization to condition is nothing more than technique to generate credible counterfactual substitutes.

BMI = 20

Since we cannot observe BLACK under both scenarios, we

substitute BLUE for the unexposed scenario.

?

If BLUE is a perfect substitute for BLACK but for the exposure (ie, unexposed BLACK), then causal inference is credible.

To extent BLUE is NOT a perfect substitute for BLACK but for the exposure, we have bias or confounding.

Background differences complicates inference.

BLACKExposed Observed

BLACKUnexposed

Counterfactual

BLUEUnexposed

CounterfactualSubstitute

So causal inference is all about finding the best counterfactual substitute for the unobservable counterfactual scenario.

The best ones are said to be exchangeable.

Randomization is a good mechanism to produce a group of subjects that are exchangeable.

Simply, it’s all about the comparison group!

Computer simulation to advance theory/concepts

See Auchinloss & Diez Rouz. In Press. American J Epidemiology

• Methodological individualist

• Nonlinear

• SUTVA not an issue

• Different kind of explanation

• But not empirical…

Agent-based Models