The paper considers meta-analysis of diagnostic studies that use a continuous score for classification of study participants into healthy or diseased groups. Classification is often done on the basis of a threshold or cut-off value, which might vary between studies. Consequently, conventional meta-analysis methodology focusing solely on separate analysis of sensitivity and specificity might be confounded by a potentially unknown variation of the cut-off value. To cope with this phenomena it is suggested to use, instead, an overall estimate of the misclassification error previously suggested and used as Youden's index and; furthermore, it is argued that this index is less prone to between-study variation of cut-off values. A simple Mantel—Haenszel estimator as a summary measure of the overall misclassification error is suggested, which adjusts for a potential study effect. The measure of the misclassification error based on Youden's index is advantageous in that it easily allows an extension to a likelihood approach, which is then able to cope with unobserved heterogeneity via a nonparametric mixture model. All methods are illustrated at hand of an example on a diagnostic meta-analysis on duplex doppler ultrasound, with angiography as the standard for stroke prevention.

Measurement error affecting the independent variables in regression models is a common problem in many scientific areas. It is well known that the implications of ignoring measurement errors in inferential procedures may be substantial, often turning out in unreliable results. Many different measurement error correction techniques have been suggested in literature since the 80's. Most of them require many assumptions on the involved variables to be satisfied. However, it may be usually very hard to check whether these assumptions are satisfied, mainly because of the lack of information about the unobservable and mismeasured phenomenon. Thus, alternatives based on weaker assumptions on the variables may be preferable, in that they offer a gain in robustness of results. In this paper, we provide a review of robust techniques to correct for measurement errors affecting the covariates. Attention is paid to methods which share properties of robustness against misspecifications of relationships between variables. Techniques are grouped according to the kind of the underlying modeling assumptions and the inferential methods. Details about the techniques are given and their applicability is discussed. The basic framework is the epidemiological setting, where literature about the measurement error phenomenon is very substantial.

In this paper, a new allocation rule for treatment assignments in sequential clinical trials is proposed. The stratified and randomized play-the-winner rule (SRPWR) is an extension of the randomized play-the-winner rule to more than two treatments. It is applicable to cases where the probabilities of success of a treatment depend on both treatments and known confounders (e.g., patient's age, gender and disease status). On average, demonstrate that the SRPWR assigns more patients to the better treatment, while eliminating the selection bias and allowing delayed responses to treatments.

Simulated data sets are used to evaluate conditional and unconditional maximum likelihood estimation in an individual case-control design with continuous covariates when there are different rates of excluded cases and different levels of other design parameters. The effectiveness of the estimation procedures is measured by method bias, variance of the estimators, root mean square error (RMSE) for logistic regression and the percentage of explained variation. Conditional estimation leads to higher RMSE than unconditional estimation in the presence of missing observations, especially for 1:1 matching. The RMSE is higher for the smaller stratum size, especially for the 1:1 matching. The percentage of explained variation appears to be insensitive to missing data, but is generally higher for the conditional estimation than for the unconditional estimation. It is particularly good for the 1:2 matching design. For minimizing RMSE, a high matching ratio is recommended; in this case, conditional and unconditional logistic regression models yield comparable levels of effectiveness. For maximizing the percentage of explained variation, the 1:2 matching design with the conditional logistic regression model is recommended.

The traditional concept of a P-value comparing an observed binomial probability to a prescribed value is extended to the ordered trinomial case in which target proportions have been specified for `excellent', `acceptable' and `unacceptable' quality. The resulting trinomial probabilities are summarised by calculating two aggregate probabilities, relating to outcomes unequivocally better than, and unequivocally worse than, that actually observed, based on these assumed target proportions. Accumulations of exactly calculated tail probabilities on a mid-P basis are recommended.

Logistic regression is frequently used in many areas of applied statistics. The maximum likelihood estimates (MLE) of the logistic regression parameters are usually computed using the iterative Newton—Raphson method. It is well known that these estimates are biased. Several methods are proposed to correct the bias of these estimates. Among them Firth (1993) and Cordeiro and McCullagh (1991) proposed two promising methods. The conditional exact method (CMLE) is popular for small-sample estimates, and is available in many software packages. In this article we compare these methods in terms of their bias. In general, our extensive simulations show that the methods proposed by Cordeiro and McCullagh and by Firth work well, though Cordeiro and McCullagh is slightly better in our simulations. In case of separation, Firth or CMLE can be used; however, a judicious approach is required when there is a wide variation in results. Two real data analyses are given exhibiting these properties. The data analysis also includes bootstrap results.

A major interest in gene expression microarray studies is to develop an accurate classifier which can be adopted in clinical practice. The usage of large numbers of genes with small data samples may lead to overfitting in classification, and generate promising, but often nonreproducible results. Therefore, assessing the reproducibility of a classifier is necessary. Appropriate methods for validating a developed classifier and estimating its predicting accuracy are discussed. In addition, some mistakes that can arise in the cross validation process are reviewed using published articles in prominent medical journals, to prevent the indefinite results of a classifier development from leading to inappropriate treatment.

Measuring the benefit of screening mammography is difficult due to lead-time bias, length bias and over-detection. We evaluated the benefit of screening mammography in reducing breast cancer mortality using observational data from the SEER-Medicare linked database. The conceptual model divided the disease duration into two phases: preclinical (T₀) and symptomatic (T₁) breast cancer. Censored information for the bivariate response vector ( T₀, T₁) was observed and used to generate a likelihood function. However, the contribution to the likelihood function for some observations could not be calculated analytically, thus, censoring boundaries for these observations were modified. Inferences about the impact of screening mammography on breast cancer mortality were made based on maximum likelihood estimates derived from this likelihood function. Hazard ratios (95% confidence intervals) of 0.54 (0.48—0.61) and 0.33 (0.26— 0.42) for single and regular users (vs. non-users), respectively, demonstrated a protective effect of screening mammography among women 69 years and older. This method reduced the impact of lead-time bias, length bias and over-detection, which biased the estimated hazard ratios derived from standard survival models in favour of screening.