ECON 440 Lecture 3: Study Design and the Quality of Evidence
Why do we care?
○
How can we estimate them?
○
Causal effects
•
Strengths and weaknesses
○
3 categories of study design
•
Reevaluating your studies
•
Outline
Correlation vs. Causation
What are the impacts of interventions, programs, and policies?
•
What are the causal effects of social exposures (e.g., income, education?)
○
What are the causal effects of physical and behavioral risk factors (e.g. diet, smoking?
○
Where should we intervene to improve health?
•
Does Health Economics & Policy Need Causal Effects?
Same individual(s)
○
What would the outcome among the treated have been if they were instead not treated?
•
Can't measure it - we must estimate it
•
The Challenge of the Unobservable Counterfactual
In expectation
1.
The same as the treated group (in all ways that matter)
2.
Except for the treatment/exposure of interest
3.
To estimate an unbiased effect, the control group must be:
•
With randomization?
▪
In model-based studies?
▪
In simple observational studies?
▪
In quasi-experimental designs?
▪
The same as the treated group, except for the exposure of interest?
○
Is the control group a good counterfactual?
•
Ideally, aided by evidence and arguments presented by the authors
▪
Whether or not this is a reasonable assumption is something you need to assess yourself
○
All these designs make an assumption about the absence of unobserved confounding (yes, even RCTs)
•
The Comparison/Control Group
Why - what does randomization get us?
▪
Why not - what's the question?
▪
RCT - acts as the "Gold Standard"
○
Often exclude important segments of the patient population
▪
Randomized Clinical Trials
○
Field experiments
○
Law of Large Numbers
○
Exposure shouldn't be correlated with observed or unobserved "confounders"
○
Exposure is randomized
•
Mathematical models of infectious disease
○
Agent-based models
○
Dynamic stochastic models
○
Outputs are only as good as the inputs (assumption, parameters, etc.)
▪
The model generates alternative states of the world (counterfactuals)
○
Model-based studies
•
Treated vs. controls
▪
Cross-sectional comparisons
○
Observational studies
•
3 Broad Categories of Study Design
Lecture 3 - Study Design and the Quality of Evidence
Tuesday, January 23, 2018
6:57 PM
ECON 440 Page 1
Document Summary
Lecture 3 - study design and the quality of evidence. What are the causal effects of social exposures (e. g. , income, education?) Can"t measure it - we must estimate it. To estimate an unbiased effect, the control group must be: The same as the treated group (in all ways that matter) All these designs make an assumption about the absence of unobserved confounding (yes, even rcts) Whether or not this is a reasonable assumption is something you need to assess yourself. Ideally, aided by evidence and arguments presented by the authors. Often exclude important segments of the patient population. Exposure shouldn"t be correlated with observed or unobserved confounders The model generates alternative states of the world (counterfactuals) Outputs are only as good as the inputs (assumption, parameters, etc. ) Difference-in-differences, interrupted time series, regression discontinuity, instrumental variables, etc. This is what allows us to infer causality - and is what economists care about.