Ghent University Academic Bibliography

Advanced

Causal inference: sense and sensitivity versus prior and prejudice

Els Goetghebeur UGent (2011) Computermodelle in der Wissenschaft : zwischen Analyse, Vorhersage and Suggestion. In Nova Acta Leopoldina. Abhandlungen der Deutschen Akademie der Naturforscher Leopoldina 110, nr. 377. p.47-64
abstract
The age old quest for the golden grail of causal answers has been at the heart of science for centuries, yet fascinating methodological progress continues to be made. The more recent statistical framework of potential outcomes considers for instance a whole set of `counterfactual responses under different exposures’ when only one of them was actually observed. It helps highlight the precise meaning of a causal estimand and links the real with the counterfactual world when the goal is not only to predict outcomes under a similar data structure as the one observed, but also after one starts intervening. The new formalism generated a stream of research involving `structural’ models for potential outcomes and lead to new insights in needed assumptions with corresponding study designs and inference. Causal questions evolve too, such as in epigenetics or plant breeding for instances, where one now wishes to learn about the causal effect of a gene or snip on a trait. The breakthrough insights owe dues to more sophisticated mathematical statistical models, that cope with missing data, stimulated by a growing digital and computing capacity. They also revealed the danger of making highly complex models easy and fast to fit. The epidemiologic literature has for instance been flooded with regression models with results all too quickly interpreted causally. Mistakes are easily made when a model is fitted in a matter of seconds, while thinking through needed assumptions and justified interpretation requires important insight in causal pathways with direct and indirect effects, followed by abstract reasoning at different levels to decide for which variables one should and should not (!) adjust to avoid bias. Computer scientists have been instrumental in developing the tools of causal graphs for this purpose. Much more work is needed to let this penetrate in mainstream applied research. Heckman, McFadden, Newey, Pearl, Rubin, Robins … are among the many who made major contributions to the field. We will discuss the power and limitations of some of these new methods by drawing on recent studies such as those on the prevention of HIV, the assessment of harmful effects of VIOXX and antidepressants, the impact of hospital infections on mortality …
Please use this url to cite or link to this publication:
author
organization
year
type
bookChapter
publication status
published
subject
book title
Computermodelle in der Wissenschaft : zwischen Analyse, Vorhersage and Suggestion
editor
Thomas Lengauer
series title
Nova Acta Leopoldina. Abhandlungen der Deutschen Akademie der Naturforscher Leopoldina
volume
110, nr. 377
pages
47 - 64
publisher
Wissenschaftlichen Verlagsgesellschaft
place of publication
Stuttgart, Germany
ISSN
0369-5034
ISBN
9783804728028
language
English
UGent publication?
yes
classification
B2
copyright statement
I have transferred the copyright for this publication to the publisher
id
2029352
handle
http://hdl.handle.net/1854/LU-2029352
date created
2012-02-14 10:14:01
date last changed
2017-01-02 09:54:41
@incollection{2029352,
  abstract     = {The age old quest for the golden grail of causal answers has been at the heart of science for centuries, yet fascinating methodological progress continues to be made. The more recent statistical framework of potential outcomes considers for instance a whole set of `counterfactual responses under different exposures{\textquoteright} when only one of them was actually observed. It helps highlight the precise meaning of a causal estimand and links the real with the counterfactual world when the goal is not only to predict outcomes under a similar data structure as the one observed, but also after one starts intervening. The new formalism generated a stream of research involving `structural{\textquoteright} models for potential outcomes and lead to new insights in needed assumptions with corresponding study designs and inference. Causal questions evolve too, such as in epigenetics or plant breeding for instances, where one now wishes to learn about the causal effect of a gene or snip on a trait. The breakthrough insights owe dues to more sophisticated mathematical statistical models, that cope with missing data, stimulated by a growing digital and computing capacity. They also revealed the danger of making highly complex models easy and fast to fit. The epidemiologic literature has for instance been flooded with regression models with results all too quickly interpreted causally. 
Mistakes are easily made when a model is fitted in a matter of seconds, while thinking through needed assumptions and justified interpretation requires important insight in causal pathways with direct and indirect effects, followed by abstract reasoning at different levels to decide for which variables one should and should not (!) adjust to avoid bias. Computer scientists have been instrumental in developing the tools of causal graphs for this purpose. Much more work is needed to let this penetrate in mainstream applied research. Heckman, McFadden, Newey, Pearl, Rubin, Robins {\textellipsis} are among the many who made major contributions to the field. We will discuss the power and limitations of some of these new methods by drawing on recent studies such as those on the prevention of HIV, the assessment of harmful effects of VIOXX and antidepressants, the impact of hospital infections on mortality {\textellipsis}},
  author       = {Goetghebeur, Els},
  booktitle    = {Computermodelle in der Wissenschaft : zwischen Analyse, Vorhersage and Suggestion},
  editor       = {Lengauer, Thomas},
  isbn         = {9783804728028},
  issn         = {0369-5034},
  language     = {eng},
  pages        = {47--64},
  publisher    = {Wissenschaftlichen Verlagsgesellschaft},
  series       = {Nova Acta Leopoldina. Abhandlungen der Deutschen Akademie der Naturforscher Leopoldina},
  title        = {Causal inference: sense and sensitivity versus prior and prejudice},
  volume       = {110, nr. 377},
  year         = {2011},
}

Chicago
Goetghebeur, Els. 2011. “Causal Inference: Sense and Sensitivity Versus Prior and Prejudice.” In Computermodelle in Der Wissenschaft : Zwischen Analyse, Vorhersage and Suggestion, ed. Thomas Lengauer, 110, nr. 377:47–64. Stuttgart, Germany: Wissenschaftlichen Verlagsgesellschaft.
APA
Goetghebeur, E. (2011). Causal inference: sense and sensitivity versus prior and prejudice. In T. Lengauer (Ed.), Computermodelle in der Wissenschaft : zwischen Analyse, Vorhersage and Suggestion (Vol. 110, nr. 377, pp. 47–64). Stuttgart, Germany: Wissenschaftlichen Verlagsgesellschaft.
Vancouver
1.
Goetghebeur E. Causal inference: sense and sensitivity versus prior and prejudice. In: Lengauer T, editor. Computermodelle in der Wissenschaft : zwischen Analyse, Vorhersage and Suggestion. Stuttgart, Germany: Wissenschaftlichen Verlagsgesellschaft; 2011. p. 47–64.
MLA
Goetghebeur, Els. “Causal Inference: Sense and Sensitivity Versus Prior and Prejudice.” Computermodelle in Der Wissenschaft : Zwischen Analyse, Vorhersage and Suggestion. Ed. Thomas Lengauer. 110, nr. 377. Stuttgart, Germany: Wissenschaftlichen Verlagsgesellschaft, 2011. 47–64. Print.