72 T.S. Riall

are used to compare proportions for categorical or ordinal values with two or more independent groups. In the case of more than two groups, the resulting P-value indicates an overall difference between the groups but does not provide pairwise comparisons. When expected cell frequencies are less than five, Fisher exact tests should be used. For matched samples, the McNemar test is used for two variables and the Cochran Q test for three or more.

Outcomes/Health Services Research

Health services or outcomes research often involves second- ary data analysis of large administrative datasets that were not initially collected for research purposes.These datasets include Medicare data, Surveillance, Epidemiology, and End Results (SEER) tumor registry data, hospital discharge data, tumor registries,the National Cancer Data Bank,and the Nationwide Inpatient Sample. The use of such data is complex and poses many analytical difficulties. It is absolutely critical to under- stand if the particular dataset can actually answer your ques- tion.For example,you might want to use SEER tumor registry data to evaluate several published algorithms for predicting additional axillary node positivity in patients with a positive sentinel lymph node biopsy. You need to carefully look through the documentation. SEER collected information on sentinel lymph node biopsy after 2002. However, they give only the final nodal status of the axilla, and you are unable to separate the status of the sentinel nodes from the status of the remainder of the axilla, so you cannot do the study.

In addition, the coding in administrative datasets changes over time. For example, new diagnosis and procedure codes are added and staging schemes are altered, and it is easy to make errors. Therefore, you must thoroughly understand the dataset. I recommend working with someone who has used the dataset before. Download all the relevant coding manuals and information from associated websites and be sure you understand changes in coding schemes over time. Finally, the

manipulation of these datasets requires significant expertise in data management and should not be performed without the help of an experienced data manager/biostatistician.

Administrative data are observational. As such, they are susceptible to significant confounding and selection bias.This often requires advanced statistical techniques to decrease the inherent selection bias, and there may be situations in which we cannot overcome this bias at all. For example, in a 2008 study by Giordano et al.(see selected references),the authors used SEER data to compare outcomes in 43,847 men receiv- ing or not receiving active treatment for lowand intermedi- ate-risk prostate cancer. A previous study using SEER data showed that active treatment was associated with improved survival.This study showed that the observed association was likely the result of selection bias by evaluating cancer mortal- ity and non-cancer mortality in the two groups. Patients who received active treatment also had significantly lower noncancer mortality,suggesting significant selection bias in which patients receive therapy and which do not.

Steps in Outcomes Research

There are generally six steps in developing an outcomes research project. You must identify a researchable question, develop a conceptual model, identify the primary outcome (and secondary outcomes if desired), identify critical inde- pendent variables (predictors and potential confounders), identify appropriate measures of each variable, and develop an analysis plan.

For example, we wanted to use Medicare claims data to evaluate current cholecystectomy rates in Medicare benefi- ciaries presenting with gallstone pancreatitis. We wanted to evaluate factors that predicted cholecystectomy on initial hospitalization. We then wanted to look at gallstone-related readmission rates in patients who did or did not undergo cholecystectomy.We developed the conceptual model shown in Fig. 5.3. We hypothesized that admitting service, hospital