TY - JOUR
T1 - Squeezing observational data for better causal inference
T2 - Methods and examples for prevention research
AU - Garcia-Huidobro, Diego
AU - Michael Oakes, J.
N1 - Publisher Copyright:
© 2016 International Union of Psychological Science
PY - 2017/4/1
Y1 - 2017/4/1
N2 - Randomised controlled trials (RCTs) are typically viewed as the gold standard for causal inference. This is because effects of interest can be identified with the fewest assumptions, especially imbalance in background characteristics. Yet because conducting RCTs are expensive, time consuming and sometimes unethical, observational studies are frequently used to study causal associations. In these studies, imbalance, or confounding, is usually controlled with multiple regression, which entails strong assumptions. The purpose of this manuscript is to describe strengths and weaknesses of several methods to control for confounding in observational studies, and to demonstrate their use in cross-sectional dataset that use patient registration data from the Juan Pablo II Primary Care Clinic in La Pintana-Chile. The dataset contains responses from 5855 families who provided complete information on family socio-demographics, family functioning and health problems among their family members. We employ regression adjustment, stratification, restriction, matching, propensity score matching, standardisation and inverse probability weighting to illustrate the approaches to better causal inference in non-experimental data and compare results. By applying study design and data analysis techniques that control for confounding in different ways than regression adjustment, researchers may strengthen the scientific relevance of observational studies.
AB - Randomised controlled trials (RCTs) are typically viewed as the gold standard for causal inference. This is because effects of interest can be identified with the fewest assumptions, especially imbalance in background characteristics. Yet because conducting RCTs are expensive, time consuming and sometimes unethical, observational studies are frequently used to study causal associations. In these studies, imbalance, or confounding, is usually controlled with multiple regression, which entails strong assumptions. The purpose of this manuscript is to describe strengths and weaknesses of several methods to control for confounding in observational studies, and to demonstrate their use in cross-sectional dataset that use patient registration data from the Juan Pablo II Primary Care Clinic in La Pintana-Chile. The dataset contains responses from 5855 families who provided complete information on family socio-demographics, family functioning and health problems among their family members. We employ regression adjustment, stratification, restriction, matching, propensity score matching, standardisation and inverse probability weighting to illustrate the approaches to better causal inference in non-experimental data and compare results. By applying study design and data analysis techniques that control for confounding in different ways than regression adjustment, researchers may strengthen the scientific relevance of observational studies.
KW - Causal inference
KW - Methodology
KW - Methods
KW - Observational studies
KW - Randomised trials
UR - http://www.scopus.com/inward/record.url?scp=85016642979&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85016642979&partnerID=8YFLogxK
U2 - 10.1002/ijop.12275
DO - 10.1002/ijop.12275
M3 - Article
C2 - 27094382
AN - SCOPUS:85016642979
SN - 0020-7594
VL - 52
SP - 96
EP - 105
JO - International Journal of Psychology
JF - International Journal of Psychology
IS - 2
ER -