OBJECTIVES:   Outcomes research often employs observational designs (e.g., disease registries, administrative health care datasets, chart reviews).  Researchers using observational data may find various amounts of missing data for confounders when analyzing the association between an exposure (such as treatment use) and an outcome (such as an adverse event). This abstract examines the case when a potentially important confounding variable has a large amount of missing data and compares the analytic methods that may be used in this situation. METHODS:   Strategies for handling missing confounder information include: (1) ignore confounders with lots of missing values; (2) exclude cases that are missing a confounder value; (3) impute a value for the confounder; (4) include missing as a separate category in the analysis.  Data from a disease registry were used as the basis for simulations to compare the odds ratio for risk of death in patients who received a treatment compared to those without treatment.  Both a clinical measurement and a subjective physician assessment are known to confound the relationship between treatment and death. RESULTS:   The most problematic pattern of missing data was informative missing data.  In one simulation, the clinical measurement was a strong predictor of death; however, it was disproportionately missing in patients who had died.  The physician assessment predicted death strongly among patients missing the clinical measurement, but only weakly in patients not missing the clinical data.  Different approaches to the missing confounder data either exacerbated or ameliorated the problem. CONCLUSIONS:  Excluding cases can create misleading results due to selection bias. Combining all missing values into a separate category can create data “noise” (i.e., classification error); however, this may be the most transparent strategy and least likely to bias results. It is important to include all cases and all potential confounders in the analysis of outcomes research studies.