infants less than 31 weeks by metaanalysis, although outcomes from individual studies were not significant.50 Systematic reviews and metaanalyses can also report observational studies, although this is less common owing to differences in study design that limit pooled estimates.51 Observational Study Design in Comparative Effectiveness Research The recent funding availability for CER has resulted in increased attention to observational studies. This raises controversy because of bias limitations inherent in observational study design. One of the most common problems with nonrandomized studies is the uneven distribution of unmeasured confounders. Another major issue is confounding by indication, meaning that the patients believed most likely to benefit from a treatment are the ones most likely to receive it, which exaggerates the actual treatment effect in the analysis. Time frames of the study cohort can present difficulties, with new entrants and attrition. Finally, practice and policy changes that occur during data collection can affect analysis. The data source can affect applicability to other settings, such as the characteristics of that population, local practice patterns, and resource availability.52 Although quantitative methods to minimize the effect of bias are beyond the scope of this paper, the Institute of Medicine has recommended explicit attention to methodologic considerations of observational study design.1,53 Despite these limitations, observational studies provide a mechanism for answering clinical questions for which RCTs are not feasible for a variety of reasons. Observational studies offer potential benefit for clinical questions in which the required sample size would be prohibitive.54 This could include evaluations of adverse events, such as comparing the effect of differing lengths of initial antibiotic therapy on subsequent Neonatal Comparative Effectiveness 837 development of necrotizing enterocolitis.55 Studies have found that confounding by indication is less problematic for evaluating unanticipated harms than for evaluating beneficial effects.47 The large number of patients available for observational studies also facilitates studies of relatively rare diseases, such as the effect of antifungal therapy in extremely low birth weight infants with invasive candidiasis.56 It can be useful in comparing the effects of similar therapies for which the sample size to detect treatment differences may be prohibitive, such as comparing types of antenatal corticosteroids on subsequent hearing and neurodevelopmental impairment.57 Observational studies are also important when randomization is not feasible owing to ethical considerations, or practical issues related to the study question at hand, such as questions of treatment adherence or usage outside of trials, such as during evaluation of total body cooling for hypoxic ischemic encephalopathy,58 or geographic or demographic effects on treatment results.54 Finally, observational studies, particularly those using already existing data, are far less expensive and time consuming than RCTs. For clinical questions where an RCT would be cost or time prohibitive, observational data represent an alternative to expedite advancing the evidence basis for clinical decision making.54 Data Sources for Observational Study Design Observational studies can be conducted as a prospective cohort, as was done in comparing antihypotensive therapies for extremely preterm infants.59 In this example, an observational design was chosen over an RCT owing to lack of physician equipoise in treating hypotension, wide practice variability complicating identification of inclusion and exclusion criteria, and the potential for enrollment or selection bias when enrolling a vulnerable patient population shortly after birth.35,59 More commonly, to obtain the sample size that confers an advantage to observational studies, they can be accomplished via secondary analysis of already collected data. Data sources include disease registries, electronic health records data, and administrative data. Very low birth weight registries such as the National Institute of Child Health and Human Development Neonatal Research Network and Vermont Oxford Network, or disease registries such as the Extracorporeal Life Support Organization database and the Congenital Diaphragmatic Hernia Study Group, are gathered via primary data abstraction of prespecified data elements. Benefits to this approach include trained abstractors, ongoing quality assurance regarding data collection, and discrete coded variables that relate to the disease of interest. Limitations to this approach include lack of granularity of the data fields, differences in definitions between data sources, and differences in the way that detailed individual data (eg, serial laboratory values) are aggregated into discrete variables. For example, the Vermont Oxford Network collects information on an infant’s respiratory support at 36 weeks postmenstrual age. This type of data does then not allow comparison of therapies that require the need to discriminate between the differences in respiratory requirements that may be relevant only before or after this particular data point. Following the US government meaningful use incentives for use of electronic health records systems, chart review for clinical research is rapidly becoming a more viable large-scale option.60 Secondary