- Email: [email protected]
Empirical bayes estimation of subgroup effects in clinical trials
Abstracts
283
Empirical Bayes Estimation of Subgroup Effects in Clinical Trials C l a r e n c e E. D a v i s a n d D i a n e Leffingwell
University of North Carolina, Chapel Hill, North Carolina (10) At the completion of a clinical trial it is often desirable to compare the treatments within subgroups of patients. The results of the subgroup analysis are usually reported for the subgroups within which sizable treatment differences are found. This practice can lead to an overestimate of the difference between treatments within the subgroups reported. One way of adjusting for this bias is to use empirical Bayes methods that shrink the extreme estimates toward the overall measure of treatment difference. Both point and confidence interval estimates of the subgroup differences can be obtained. The computations required for such an analysis are presented for the case when treatment effect is measured by risk ratios, and an example using subgroup data from the Lipid Research Clinics Coronary Primary Prevention Trial is used to illustrate the method. Analyzing Survival from Initial Diagnosis Avital C n a a n a n d L o u i s e R y a n
Harvard School of Public Health and Dana-Father Cancer Institute, Boston, Massachusetts (11) The usual goal when analyzing clinical trials is to evaluate a treatment effect, fiowever, clinicians are often interested in using the databases collected from such trials to study disease natural history. In particular, survival from the time of initial diagnosis and related prognostic factors are often of interest. Although many such analyses appear in the medical literature, it is incorrect to simply apply standard survival techniques, with each patient's survival measured from the time of diagnosis. The reason is that patients entering a clinical trial constitute a biased sample: those who die relatively soon after diagnosis will be underrepresented in the sample. We describe an approach that correctly analyzes time from initial diagnosis, based on a proportional hazards model. The likelihood is constructed by conditioning on each patient's time of entry to the study, leading to nonstandard risk sets. We illustrate our approach with several data examples and compare it with a standard analysis that ignores the time of entry to the study.
A n a l y s i s of Squared Tables of Ordinal Data with Applications to C l i n i c a l T r i a l s N a z n e e n Shariff University of Edinburgh, Scotland (12) The pharmaceutical industry gives rise to ordinal data from clinical trials of newly manufactured drugs. The proportional odds model is assumed to fit the data. A simulation study compares existing methods with a new method of application (Wilcoxon Van Elteren test) in testing the null hypothesis formulated through the parameters of the model. The methods employed are evaluated on a real data set of square ordinal responses from the clinical trial of a new drug. Theoretical results pertaining to the new Wilcoxon Van Elteren method are derived and verified on the real data set.
Baseline Covariate lntrablock Correlations J o h n Matts, L a V e r n e Christie, a n d t h e P O S C H G r o u p University of Minnesota, Minneapolis Minnesota (13) The permuted block design is perhaps the most frequently used design for randomized clinical trials. The usual practice in analysis is to ignore the blocking. The consequence of ignoring the blocking is reflected in the intrablock correlation coefficient (R). Such an analysis is conservative or anticonservative depending on whether R is positive or negative, respectively. Using data from the Program on the Surgical Control of the Hyperlipidemias (POSCH), we examined the values of R for baseline covariates based on the actual randomization employed. The blocking scheme within each stratum was 2, 4, 6, 6 . . . . ; within each clinic, 18 strata were possible. We also calculated strata-adjusted Rs that reflect the effect of accounting for the ~tratification but ignoring the blocking within strata. Some values were negative. However, in general the Rs were positive and <0.10. Several values of R were larger. In these cases, a data collection change or eligibility criteria change was reflected. Thus, for most variables ignoring the blocking had little effect. However, large effects were possible.
283
Empirical Bayes Estimation of Subgroup Effects in Clinical Trials C l a r e n c e E. D a v i s a n d D i a n e Leffingwell
University of North Carolina, Chapel Hill, North Carolina (10) At the completion of a clinical trial it is often desirable to compare the treatments within subgroups of patients. The results of the subgroup analysis are usually reported for the subgroups within which sizable treatment differences are found. This practice can lead to an overestimate of the difference between treatments within the subgroups reported. One way of adjusting for this bias is to use empirical Bayes methods that shrink the extreme estimates toward the overall measure of treatment difference. Both point and confidence interval estimates of the subgroup differences can be obtained. The computations required for such an analysis are presented for the case when treatment effect is measured by risk ratios, and an example using subgroup data from the Lipid Research Clinics Coronary Primary Prevention Trial is used to illustrate the method. Analyzing Survival from Initial Diagnosis Avital C n a a n a n d L o u i s e R y a n
Harvard School of Public Health and Dana-Father Cancer Institute, Boston, Massachusetts (11) The usual goal when analyzing clinical trials is to evaluate a treatment effect, fiowever, clinicians are often interested in using the databases collected from such trials to study disease natural history. In particular, survival from the time of initial diagnosis and related prognostic factors are often of interest. Although many such analyses appear in the medical literature, it is incorrect to simply apply standard survival techniques, with each patient's survival measured from the time of diagnosis. The reason is that patients entering a clinical trial constitute a biased sample: those who die relatively soon after diagnosis will be underrepresented in the sample. We describe an approach that correctly analyzes time from initial diagnosis, based on a proportional hazards model. The likelihood is constructed by conditioning on each patient's time of entry to the study, leading to nonstandard risk sets. We illustrate our approach with several data examples and compare it with a standard analysis that ignores the time of entry to the study.
A n a l y s i s of Squared Tables of Ordinal Data with Applications to C l i n i c a l T r i a l s N a z n e e n Shariff University of Edinburgh, Scotland (12) The pharmaceutical industry gives rise to ordinal data from clinical trials of newly manufactured drugs. The proportional odds model is assumed to fit the data. A simulation study compares existing methods with a new method of application (Wilcoxon Van Elteren test) in testing the null hypothesis formulated through the parameters of the model. The methods employed are evaluated on a real data set of square ordinal responses from the clinical trial of a new drug. Theoretical results pertaining to the new Wilcoxon Van Elteren method are derived and verified on the real data set.
Baseline Covariate lntrablock Correlations J o h n Matts, L a V e r n e Christie, a n d t h e P O S C H G r o u p University of Minnesota, Minneapolis Minnesota (13) The permuted block design is perhaps the most frequently used design for randomized clinical trials. The usual practice in analysis is to ignore the blocking. The consequence of ignoring the blocking is reflected in the intrablock correlation coefficient (R). Such an analysis is conservative or anticonservative depending on whether R is positive or negative, respectively. Using data from the Program on the Surgical Control of the Hyperlipidemias (POSCH), we examined the values of R for baseline covariates based on the actual randomization employed. The blocking scheme within each stratum was 2, 4, 6, 6 . . . . ; within each clinic, 18 strata were possible. We also calculated strata-adjusted Rs that reflect the effect of accounting for the ~tratification but ignoring the blocking within strata. Some values were negative. However, in general the Rs were positive and <0.10. Several values of R were larger. In these cases, a data collection change or eligibility criteria change was reflected. Thus, for most variables ignoring the blocking had little effect. However, large effects were possible.