Separability of Effects in the Analysis of Complex Observational Data

Observational studies contribute to the overall body of knowledge, frequently capturing how medicine is practiced in a real-life setting, where patients and therapies are diverse and not limited by research protocol. However, they can present analytic challenges for these reasons, among them non-ignorable treatment assignment resulting in confounding by indication, and the presence of structural zeros in the data. This study addresses these two challenges in the context of an observational study of children receiving sedation outside of the operating room, with the goal of estimating the risk of adverse events associated with the use of propofol. Since each child received at least one sedative, the risk without any drugs was not observable. Logistic regression was used initially, using propensity scores to address the non-ignorable treatment assignments. Propofol was found to be no more or less risky than other drugs used for the same purpose. This analysis has limitations caused by the presence of structural zeros, leading to uncertainty regarding interactions between drugs and resulting inability to provide a risk profile as guidance for the physician. Next, a methodology called Separability of Effects (SE) is developed. It states that SE is present when the stepwise process with respect to a class of variables leads to the same conclusion whether subjects exhibiting those variables are included or not. The sedation unit data was used to illustrate the method and to show how it addresses limitations of the regression analysis. Meaningful differences among drug combinations were found, despite non-significant interaction terms in the regression. A risk profile for propofol was developed using the SE method. Third, a simulation study was conducted where a hypothetical set of children receiving no drug was added to the observed sedation data. The effect of the addition of these data on the main effect of propofol and its interactions is presented. The results support the conclusion that, in the absence of baseline risk data, the SE method provides information on the risks associated with drug combinations not necessarily detected by interaction terms of regression models. The results are consistent over changes in the proportion of hypothetical data.

Name thesis:degree_name

Doctor of Philosophy

Level thesis:degree_level

doctoral

Discipline thesis:degree_discipline

Epidemiology and Biostatistics

Grantor dc:publisher

Case Western Reserve University School of Graduate Studies

Year dc:date

2013

Author dc:creator

• O'Riordan, Mary Ann

Contributors dc:contributor

• Debanne, Sara

• Biostatistics
• Biomedical Research
• Epidemiology
• Medicine
• Pharmacoepidemiology
• Drug safety
• Structural zero

Statement dc:rights

• unrestricted
• This thesis or dissertation is protected by copyright: all rights reserved. It may not be copied or redistributed beyond the terms of applicable copyright laws.

Language dc:language

English

Repository record dc:identifier

http://rave.ohiolink.edu/etdc/view?acc_num=case1365080790

OAI identifier oai:identifier

oai:etd.ohiolink.edu:case1365080790