Cnossen MC, van Essen TA, Ceyisakar IE, Polinder S, Andriessen TM, van der Naalt J, Haitsma I, Horn J, Franschman G, Vos PE, Peul WC, Menon DK, Maas AI, Steyerberg EW, Lingsma HF. Clin. Epidemiol. 2018; 10: 841-852.
Affiliation
Center for Medical Decision Making, Department of Public Health, Erasmus Medical Center Rotterdam, Rotterdam, the Netherlands, m.c.cnossen@erasmusmc.nl.
INTRODUCTION: Observational studies of interventions are at risk for confounding by indication. The objective of the current study was to define the circumstances for the validity of methods to adjust for confounding by indication in observational studies.
PATIENTS AND METHODS: We performed post hoc analyses of data prospectively collected from three European and North American traumatic brain injury studies including 1,725 patients. The effects of three interventions (intracranial pressure [ICP] monitoring, intracranial operation and primary referral) were estimated in a proportional odds regression model with the Glasgow Outcome Scale as ordinal outcome variable. Three analytical methods were compared: classical covariate adjustment, propensity score matching and instrumental variable (IV) analysis in which the percentage exposed to an intervention in each hospital was added as an independent variable, together with a random intercept for each hospital. In addition, a simulation study was performed in which the effect of a hypothetical beneficial intervention (OR 1.65) was simulated for scenarios with and without unmeasured confounders.
RESULTS: For all three interventions, covariate adjustment and propensity score matching resulted in negative estimates of the treatment effect (OR ranging from 0.80 to 0.92), whereas the IV approach indicated that both ICP monitoring and intracranial operation might be beneficial (OR per 10% change 1.17, 95% CI 1.01-1.42 and 1.42, 95% CI 0.95-1.97). In our simulation study, we found that covariate adjustment and propensity score matching resulted in an invalid estimate of the treatment effect in case of unmeasured confounders (OR ranging from 0.90 to 1.03). The IV approach provided an estimate in the similar direction as the simulated effect (OR per 10% change 1.04-1.05) but was statistically inefficient.
CONCLUSION: The effect estimation of interventions in observational studies strongly depends on the analytical method used. When unobserved confounding and practice variation are expected in observational multicenter studies, IV analysis should be considered.