Week 6
Lecture
The three reasons for studying associations (goals for quantitative research)
Three possible aims: causal inference, prediction, description
Distinguish them makes sense: crucial differences in design, statistical methods, interpretation,
evaluation, role for theory/subject knowledge.
What do we do with associations: find causal effects (causal inference, counterfactual prediction),
predict the future (prediction), describe patterns (description).
Causal inference Description
Ruin your analysis by lack of statistical adjustment. Ruin your results by statistical adjustment.
Results are affected by confounding bias. Results have no interpretation anymore.
Causal inference Prediction
Goal is counterfactual prediction: what if. Not only Goal is predicting the future or the prpesent. If
about what is, but also about what could be. you know A, what can you see about B. No
causal interpretations.
Could it have been predicted
Strategy: may be data-driven, regression analysis can be used for predictions on the individual level, no
randomization/stratification/weighting/matching.
Interpretation: results have intrinsic meaning (regression coefficients represent estimates of the
(average) effect, coefficient in OLS, adjusted proportions, RR, RD), CI, P. usually no interest in
interpretation of coefficients for individual predictors.
Evaluation: how good is my prediction model?
- Dichotomous/categorical outcome: how often is your prediction right? SP, SE, VW-, VW+.
- Continuous outcome: how wrong can you expect your prediction to be? Mean absolute error,
root mean squared error.
- After randomization: repeat experiment, try to reproduce results
- After observational study: transparency about assumptions, reproducing results not as
informative (bias will remain), there is always some unmeasured confounding, but how
important is it?
Role of theory is not important. Reverse causality is no problem. No confounding, intermediate, collider,
selection bias. Positivity may improve predictive power. Good consistency required to make prediction
model useful. Exchangeability not relevant.
Lecture
The three reasons for studying associations (goals for quantitative research)
Three possible aims: causal inference, prediction, description
Distinguish them makes sense: crucial differences in design, statistical methods, interpretation,
evaluation, role for theory/subject knowledge.
What do we do with associations: find causal effects (causal inference, counterfactual prediction),
predict the future (prediction), describe patterns (description).
Causal inference Description
Ruin your analysis by lack of statistical adjustment. Ruin your results by statistical adjustment.
Results are affected by confounding bias. Results have no interpretation anymore.
Causal inference Prediction
Goal is counterfactual prediction: what if. Not only Goal is predicting the future or the prpesent. If
about what is, but also about what could be. you know A, what can you see about B. No
causal interpretations.
Could it have been predicted
Strategy: may be data-driven, regression analysis can be used for predictions on the individual level, no
randomization/stratification/weighting/matching.
Interpretation: results have intrinsic meaning (regression coefficients represent estimates of the
(average) effect, coefficient in OLS, adjusted proportions, RR, RD), CI, P. usually no interest in
interpretation of coefficients for individual predictors.
Evaluation: how good is my prediction model?
- Dichotomous/categorical outcome: how often is your prediction right? SP, SE, VW-, VW+.
- Continuous outcome: how wrong can you expect your prediction to be? Mean absolute error,
root mean squared error.
- After randomization: repeat experiment, try to reproduce results
- After observational study: transparency about assumptions, reproducing results not as
informative (bias will remain), there is always some unmeasured confounding, but how
important is it?
Role of theory is not important. Reverse causality is no problem. No confounding, intermediate, collider,
selection bias. Positivity may improve predictive power. Good consistency required to make prediction
model useful. Exchangeability not relevant.
