Grant Statistical Methods for Partially Controlled Studies ; RO1 EY 014314-01 For

2

PROGRESS REPORT

01/03/2003

Grant “Statistical Methods for Partially Controlled Studies”; RO1 EY 014314-01 for 05/2002-02/2003

PI: Constantine Frangakis, PhD (on leave for 01/2003-04/2004*).

Co-PI: Donald B. Rubin, PhD

*PI during the leave of Dr. Frangakis: Ronald Brookmeyer, PhD

We are making good progress on the specific aims of the grant as we originally proposed. There have been no major personnel changes to the grant in the past year. Dr. Brookmeyer is temporarily acting as the principal investigator during Dr. Frangakis’ leave that began January 2003 as previously arranged.

The PI and Co-PI have published or have had in press the work given in the reference list below, with the help of the writing of the proposal and of the grant. Two of the articles are followed by discussion in the journals in which they will be published. The contribution of the work is briefly described below.

The article by Barnard, Frangakis, Hill and Rubin (2003, BFHR) addresses types of randomized studies where, first, some subjects do not adhere to the randomized (controlled) treatment, and second, some subjects do not provide the outcomes on which the treatment effects are to be evaluated. In such studies, the article develops methods for evaluating (a) the effect of the randomized treatment, and (b) the effect of the treatment received by the subjects. The methods are used to evaluate the effect of school choice vouchers on student performance in NY City, and prepares the work needed to address also more complex studies such as the Baltimore Needle Exchange Program, and surrogate outcomes, proposed in the grant. The article is progressing along the line of work in Frangakis and Rubin (Biometrika, 1999, 365--379, and Barnard et al., 2002, Case Studies in Bayesian Statistics), and addresses a number of important practical problems that are found in such studies: complex designs of the controlled treatment (such as using propensity scores, Rosenbaum and Rubin, Biometrika, 1983, 41--55); missing covariate data; multivariate outcomes; and semi-continuous outcomes, with many subjects having outcomes piling-up at a single value. The article will be followed by discussion by researchers including Bengt Muthen (UCLA, Education) and Alan Krueger (Economics, Princeton) and rejoinder in JASA.

The article by Frangakis and Rubin (2002, FR02) addresses types of studies of similar but more general structure and purpose than those in BFHR. Specifically, Frangakis and Rubin (2002) set a framework for studies where there is a treatment, a post-treatment variable (surrogate), and an outcome, and where we wish to evaluate the effect of the treatment on outcome that occurs with (and without) the effect of treatment on the surrogate. For defining and estimating such effects, FR02 reveals limitations of existing frameworks, and advantages of ``principal stratification'', the new framework.

The article by Frangakis, Rubin and and Zhou (2002, henceforth FRZ) applies ``principal stratification'' in a study of Advance Directives, of importance to elders and their next of kins. The discussions of Mealli and Rubin (2003a, 2003b) explore in further depth and provide additional insight on the type of assumptions that are useful in studies with surrogate variables, with application in Psychology and Econometrics. FR02, whose published version in Biostatistics benefited substantially from the writing and revision of the grant proposal, is followed by discussion by researchers including Sir David Cox (Oxford, Statistics) and Els Goetghebeur (Ghent University, Mathematics and Informatics) and rejoinder.

Additional work that was made possible partly by the grant includes that developed in Frangakis and Varadhan (2002; FV02), and in Varadhan and Frangakis (2003; VF03). FV02 develop methods to estimate relative risk of seasonal factors, which involves addressing the problem that null values of that risk are on the boundary of the parameter space. The method is applied to validate that across both hemispheres, the risk of suicide within a country is larger in months with more sunshine. This supports, together with other observations, that low melatonin levels can be a trigger of short-term bounces in depression among depressed subjects. VF03 address studies where the case-crossover design is applied to estimate short-term effects. VF03 show that the standard analyses of a typical such design have length bias and low precision. VF03 also proposes a method to address these problems.

REFERENCES

Barnard, J., Frangakis, C. E., Hill, L., and Rubin, D. B. (2003). Principal stratification approach in broken randomized experiments: A case study of School Choice vouchers in NY City. (Frangakis and Hill are corresponding authors). To appear in the Journal of the American Statistical Association, (with discussion).

Frangakis, C. E., and Rubin, D. B. (2002). Principal stratification in causal inference. Biometrics, 58, 20--29.

Frangakis, C. E., Rubin, D. B., and Zhou, X. H. (2002). Clustered encouragement design with individual noncompliance: Bayesian inference and application to Advance Directive Forms. Biostatistics (with discussion) 3, 147-- 164.

Frangakis, C. E., and Varadhan, R (2002). On Confidence Intervals for Seasonal Risk of Suicides, with Null Values on the Boundary. Epidemiology 13, 734--737.

Mealli, F., and Rubin, D.B. (2003a). Assumptions allowing the estimation of direct causal effects: discussion of `Healthy, Wealthy, and Wise ? Tests for direct causal paths between Health and Socioeconomic status' by Adams et al '' Journal of Econometrics, 112, 79--87.

Mealli, F., and Rubin, D.B. (2003b). ``Discussion of `Estimation of intervention effects with noncompliance: alternative model specifications,' by Booil Jo,'' Joural of Educational and Behavioral Statistics, in press.

Varadhan, R and Frangakis CE (2003). Revealing and addressing length bias and low precision in the case-crossover design (manuscript in preparation).

CONFERENCE TALKS

``Revealing and addressing length bias and heterogeneous effect periods in case-crossover designs,’’ Joint Statistical Meetings, New York, NY, August 2002.

``Evaluating the impact of the Baltimore Needle Exchange Program'', (invited), International Biometrics Society (ENAR), Annual Meeting, Tampa FL, (for 2003).

``Evaluating the impact of the Baltimore Needle Exchange Program'', (invited), American Mathematic Society, Joint Summer Research Conference, Mount Holyoke College, MA, 2002.