Multiple Imputation by Fully Conditional Specification for Dealing with Missing Data in a Large Epidemiologic Study
Pages 287-295
Yang Liu and Anindya De
DOI: http://dx.doi.org/10.6000/1929-6029.2015.04.03.7
Published: 19 August 2015

Abstract: Missing data commonly occur in large epidemiologic studies. Ignoring incompleteness or handling the data inappropriately may bias study results, reduce power and efficiency, and alter important risk/benefit relationships. Standard ways of dealing with missing values, such as complete case analysis (CCA), are generally inappropriate due to the loss of precision and risk of bias. Multiple imputation by fully conditional specification (FCS MI) is a powerful and statistically valid method for creating imputations in large data sets which include both categorical and continuous variables. It specifies the multivariate imputation model on a variable-by-variable basis and offers a principled yet flexible method of addressing missing data, which is particularly useful for large data sets with complex data structures. However, FCS MI is still rarely used in epidemiology, and few practical resources exist to guide researchers in the implementation of this technique. We demonstrate the application of FCS MI in support of a large epidemiologic study evaluating national blood utilization patterns in a sub-Saharan African country. A number of practical tips and guidelines for implementing FCS MI based on this experience are described.

Application of Generalized Additive Models to the Evaluation of Continuous Markers for Classification Purposes
Pages 296-305
Mónica López-Ratón, Mar Rodríguez-Girondo, María Xosé Rodríguez-Álvarez, Carmen Cadarso-Suárez and Francisco Gude
DOI: http://dx.doi.org/10.6000/1929-6029.2015.04.03.8
Published: 19 August 2015

Abstract: Background: Receiver operating characteristic (ROC) curve and derived measures as the Area Under the Curve (AUC) are often used for evaluating the discriminatory capability of a continuous biomarker in distinguishing between alternative states of health. However, if the marker shows an irregular distribution, with a dominance of diseased subjects in noncontiguous regions, classification using a single cutpoint is not appropriate, and it would lead to erroneous conclusions. This study sought to describe a procedure for improving the discriminatory capacity of a continuous biomarker, by using generalized additive models (GAMs) for binary data.

Methods: A new classification rule is obtained by using logistic GAM regression models to transform the original biomarker, with the predicted probabilities being the new transformed continuous biomarker. We propose using this transformed biomarker to establish optimal cut-offs or intervals on which to base the classification. This methodology is applied to different controlled scenarios, and to real data from a prospective study of patients undergoing surgery at a University Teaching Hospital, for examining plasma glucose as postoperative infection biomarker.

Results: Both, theoretical scenarios and real data results show that when the risk marker-disease relationship is not monotone, using the new transformed biomarker entails an improvement in discriminatory capacity. Moreover, in these situations, an optimal interval seems more reasonable than a single cutpoint to define lower and higher disease-risk categories.

Conclusions: Using statistical tools which allow for greater flexibility (e.g., GAMs) can optimize the classificatory capacity of a potential marker using ROC analysis. So, it is important to question linearity in marker-outcome relationships, in order to avoid erroneous conclusions.