Medizinische Universität Graz Austria/Österreich - Forschungsportal - Medical University of Graz
Gewählte Publikation:
Riedl, R.
On the use of balancing scores in cohort and matched case-control studies
Doktoratsstudium der Medizinischen Wissenschaft; Humanmedizin; [ Dissertation ] Medical University of Graz; 2014. pp. 160
[OPEN ACCESS]
FullText
- Autor*innen der Med Uni Graz:
- Riedl Regina
- Betreuer*innen:
- Berghold Andrea
- Altmetrics:
- Abstract:
- In observational studies the presence of many confounders, defined as variables associated with both, exposures and outcome may induce a bias in the estimates of association if not properly accounted for. However, this is challenging when the number of confounders is large and the outcome is rare. In such a case balancing scores like the propensity score (PS) or the disease risk score (DRS) have been applied to address confounding in design and analysis of observational studies. However, the performance of estimates of conditional exposure effects based on score adjusted regression models from cohort and case-control data has only been studied in limited simulations, and there is lack of theoretical results. This thesis has two parts. First, we give an overview of the use of balancing scores in the design and analysis of observational studies. We then derive the asymptotic bias in estimates of conditional exposure effects when scores, including the PS or the DRS are used as adjustment variables in multivariable regression models for the analysis of cohort data, and as matching variables in case-control studies. The magnitude of the bias in estimates of conditional exposure effects is then computed numerically and in simulations for several settings. In cohort analysis, adjusting for the PS or the DRS instead of the confounders itself yields unbiased estimates for linear exposure effects. For non-linear exposure effects in both study designs, unbiased estimates are given for the PS only under the null hypothesis of no association. The estimates obtained from matching cases and controls on the DRS and analyzing the data using conditional logistic regression are unbiased. All other settings result in bias. The magnitude of this bias as a function of various study parameters is assessed in detail for Poisson and logistic regression for cohorts, and for matched case-control studies when conditional or unconditional logistic regression adjusted for the matching scores are used to analyze the data. We additionally investigate the bias in simulations with more complex confounder correlations based on a real dataset and observed similar results as for the analytical calculations. In the second part of the thesis we study the impact of having received a blood transfusion on cancer risk in the U.S. elderly, based on a sample from the SEER (Surveillance, Epidemiology and End Results)-Medicare database. The risk is investigated for cancer overall and for specific sites and subtypes. Various summary score methods, including matching and adjustments are applied to account for possible confounders. We observed an increased cancer risk shortly after the receipt of a blood transfusion for cancer overall and for cancers of the stomach, colon, liver, kidney, renal pelvis and ureter, myeloma, leukemia and for Hodgkin lymphoma and non-Hodgkin lymphoma. Risk was not elevated for transfusions received long before cancer diagnosis with the exception of liver cancer. However, after careful adjustment for confounding, no long term effect of transfusion on risk was observed. Our findings suggest that transfusions in most patients that were later diagnosed with cancer may be prompted by an undiagnosed cancer or a precursor to cancer. In summary, this thesis contributes to the understanding of results from score adjusted regression models for cohort and from case-control data where scores were used to match cases and controls. Our work also contributes to the understanding of the relationship between blood transfusions and cancer risk.