Postsurgical infections represent an important cause of morbidity after abdominal surgery. The microbiological diagnosis is not achieved in at least 30% of culture with consequent worsening of patient outcome. In this study, procalcitonin measurement, during the first 3 days after abdominal surgery, has been evaluated for the early diagnosis of postsurgical infection.Ninety consecutive patients subjected to major abdominal surgery at the University Campus Bio-Medico of Rome, have been included. PCT concentrations were measured by time-resolved amplified cryptate emission (TRACE) assay at admission and at the first, second, and third day after surgery. PCT levels were compared using the Mann–Whitney test and by ANOVA test for variance analysis. Receiver operating characteristic (ROC) analysis was performed to define the diagnostic ability of PCT in case of postsurgical infections.PCT values resulted significantly different between patients developing or not developing postsurgical infections. PCT >1.0 ng/mL at first or second day after surgery and >0.5 ng/mL at third day resulted diagnostic for infectious complication, whereas a value <0.5 ng/mL at the fifth day after surgery was useful for early and safety discharge of patients.In conclusion, PCT daily measurement could represent a useful diagnostic tool improving health care in the postsurgical period following major abdominal surgery and should be recommended.

Data can be collected in scientific studies via a controlled experiment or passive observation. Big data is often collected in a passive way, e.g. from social media. In studies of causation great efforts are made to guard against bias and hidden confounders or feedback which can destroy the identification of causation by corrupting or omitting counterfactuals (controls). Various solutions of these problems are discussed, including randomization.

We review recent literature that proposes to adapt ideas from classical model based optimal design of experiments to problems of data selection of large datasets. Special attention is given to bias reduction and to protection against confounders. Some new results are presented. Theoretical and computational comparisons are made. K E Y W O R D Sconfounders, large datasets, model bias, optimal experimental design INTRODUCTIONFor the analysis of big datasets, statistical methods have been developed, which use the full available dataset. For example, new methodologies developed in the context of Big Data and focussed on a 'divide-and-recombine' approach are summarised in Wang et al. 19 Other major methods address the scalability of Big Data through Bayesian inference based on a Consensus Monte Carlo algorithm 13 and sparsity assumptions. 16 In contrast, other authors argue on the advantages of inference statements based on a well-chosen subset of the large dataset. Big datasets are characterised by few key factors. While usually data can be collected in scientific studies via active or passive observation, Big Data is often collected in passive way. Rarely their collection is the result of a designed process. This generates sources of bias, which either we do not know at all or are too costly to control. Nevertheless they will affect the overall distribution of the observed variables. 3,11 Many authors in Ref. 15 argues that analysis of big dataset is effected by issues of bias and confounding, selection bias and other sampling problems (see, for example, Sharpes 14 for electronic health records). Often the causal effect of interest can only be measured on the average and great care has to be taken about the background population, for example, it is possible to consider and analyse every message on Twitter and use it to draw conclusions about the public opinion, but it is known that Twitter users are not representative of the whole population. The analysis of the full dataset might be prohibitive because of computational and time constraints. Indeed, in some cases, the analysis of the full dataset might also be not advisable. 4,6 To recall just one example, where the sample proportion of a self-reported big dataset of size 2300,000 unit has the same mean squared error as the sample proportion from a suitable simple random sample (SRS) of size 400 and a Law of Large Population has been defined in order to qualify this (see Meng 9 ).Recently, some researchers argued on the usefulness of utilising methods and ideas from design of experiment (DoE) for the analysis of big datasets, more specifically from model-based optimal experimental design. They argue that specialThis is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.

Nowadays, in many different fields, massive data are available and for several reasons, it might be convenient to analyze just a subset of the data. The application of the D-optimality criterion can be helpful to optimally select a subsample of observations. However, it is well known that D-optimal support points lie on the boundary of the design space and if they go hand in hand with extreme response values, they can have a severe influence on the estimated linear model (leverage points with high influence). To overcome this problem, firstly, we propose a non-informative “exchange” procedure that enables us to select a “nearly” D-optimal subset of observations without high leverage values. Then, we provide an informative version of this exchange procedure, where besides high leverage points also the outliers in the responses (that are not necessarily associated to high leverage points) are avoided. This is possible because, unlike other design situations, in subsampling from big datasets the response values may be available. Finally, both the non-informative and informative selection procedures are adapted to I-optimality, with the goal of getting accurate predictions.

We review recent literature that proposes to adapt ideas from classical model based optimal design of experiments to problems of data selection of large datasets. Special attention is given to bias reduction and to protection against confounders. Some new results are presented. Theoretical and computational comparisons are made.