My current research is focused on the development of statistical methods for analyzing studies with informative missing or censored data. It is well known that in studies in which subjects selectively miss scheduled visits or prematurely withdraw, standard statistical methods (e.g., proportional hazards models, logrank tests, generalized estimating equations, mixed effects models, multiple imputation etc.) can yield biased results. These methods rely on strong assumptions, which cannot be empirically validated and may be scientifically unreasonable. As a result, the way that standard analyses are reported may grossly misrepresent the true level of uncertainty. With this mind, my colleagues and I have been advocating a two stage approach. The first step is to fit flexible models which allow (1) incorporation of all (primary and auxilliary) information that is prognostic for missing/censoring and (2) are valid under the untestable assumption that missingness/censoring is stochastically explained by observed factors. The second step is to perform a sensitivity analysis to assess how inference about the parameters of interest change when one allows the missingness/censoring to depend stochastically on unobserved outcomes. The degree of dependence on unobserved outcome is specified through interpretable ``selection bias'' parameters (such as odds or hazard ratios), which can be constrained by scientific experts. If the inference is insensitive over what experts consider plausible dependencies, then this lends great credence to the analysis in the first step; otherwise, it suggests that one should treat the results of the primary analysis with greater skepticism. This work is published in a series of papers including Rotnitzky, Robins, and Scharfstein (1998), Scharfstein, Rotnitzky, and Robins (1999), Robins, Rotnitzky, and Scharfstein (2000), Rotnitzky, Scharfstein, Su, and Robins (2001), Scharfstein, Robins, Eddings, and Rotnitzky (2001), Scharfstein and Robins (2002), and Scharfstein and Irizarry (2003). These papers involve applications in treatment studies of HIV and schizophrenia.

The sensitivity analysis approach relies heavily on expert opinions about the plausible range for ``selection bias'' parameters. While the methodology is useful in assessing the sensitivity of treatment comparisons to standard assumptions, it may be dissatisfying to some researchers/decision makers becuase a single answer is not provided. In contrast, a Bayesian analysis allows the investigator to draw a ``single'' inference by formally incorporating prior expert beliefs about selection bias into the analysis. In a recent article by Scharfstein, Daniels, and Robins (2003), my colleagues and I have taken a first step toward developing a flexible Bayesian approach, which utilizes informative priors

(elicited from experts) for the non-identifiable degree of selection bias and weak priors for model parameters which are well identified from the data. The Bayesian methodology is applied to setting of missing CD4 counts in a treatment study for HIV.

Some researchers may feel uncomfortable with the sensitivity analysis and Bayesian approaches because of their reliance on quantification of prior beliefs. To address this concern, we have been developing a methodology for constructing informative bounds on treatment effects that rely on soft, yet plausible, belief statements such as ``Sicker subjects at baseline are more likely to drop-out'' or ``Subjects whose health worsens are more likely to drop-out.'' In Scharfstein, Manski, and Anthony (2003), we show how to construct bounds on treatment effects in the realtively simplistic setting in which a baseline and follow-up ordinal outcome are to be measured on subjects, the latter of which is missing on some subjects. The methodology is motivated by and applied to the Good Behavior Game, which was a randomized, prevention trial designed to reduce aggressive behavior in children.

Biostatistics, Causal inference; Longitudinal data analysis; Survival analysis; Missing data; Semiparametric models

George Snedecor Award for best publication in biometry, 1997-98

Ho-Ching Yang Memorial Faculty Award

Advising, Mentoring, and Teaching Award

CV

Scharfstein DO and Irizarry RA: Generalized Additive Selection Models for the Analysis of Non-ignorable Missing Data. Under Revision, Biometrics.

Scharfstein DO, Manski CF, and Anthony JC: On the Construction of Bounds in Prsopective Studies with Missing Ordinal Outcomes: Application to the Good Behavior Game. Under Review, Biometrics.

Scharfstein DO, Daniels M, and Robins JM: Incorporating Prior Beliefs About Selection Bias into the Analysis of Randomized Trials with Missing Data. To Appear, Biostatistics.

Minkovitz C, Strobino D, Hughart N, Miller T, Scharfstein DO, Bishai D, and Guyer B: Introduction of Developmental Specialists in Pediatric Practices: Provider Perspectives from Healthy Steps. Under Revision, Ambulatory Pediatrics.

Scharfstein DO and Robins JM: Estimation of the Failure Time Distribution in the Presence of Informative Right Censoring. Biometrika 89: 617-635, 2002.

Scharfstein DO, Liang KY, Eaton W, Chen LS: The quadratic cumulative odds regression model for scored ordinal outcomes: application to alcohol dependence. Biostatistics 2 (4) 473-483, 2001.

Scharfstein DO, Robins JM, Eddings W, Rotnitzky A: Inference in randomized studies with informative censoring and discrete time-to-event endpoints. Biometrics 57(2):404-413, 2001.

Minkovitz C, Strobino D, Hughart N, Scharfstein D, Guyer B, Healthy Steps Evaluation Team: Early effects of the healthy steps for young children program. Archives of Pediatrics and Adolescent Medicine 155(4):470-9, 2001.

Rotnitzky A, Scharfstein DO, Su TL, Robins JM: Methods for conducting sensitivity analysis of trials with potentially nonignorable competing causes of censoring. Biometrics 57(1):103-113, 2001.

Guyer B, Hughart N, Strobino D, Jones A, Scharfstein D: Assessing the impact of pediatric-based developmental services on infants, families, and clinicians: challenges to evaluating the Healthy Steps Program. Pediatrics 105(3):E33, 2000.

Robins JM, Rotnitzky A, Scharfstein DO: Sensitivity analysis for selection bias and unmeasured confounding in missing data and causal inference models. In: Statistical Models for Epidemiology, the Environment, and Clinical Trials. E Halloran and D Berry, editors. pp 1-95, 2000.

Scharfstein DO, Rotnitzky A, Robins JM: Adjusting for nonignorable drop-out using semiparametric nonresponse models (with discussion). Journal of the American Statistical Association 94(448):1096-1120, 1999.