Food science problems are complex. Scientists may be able to capture more of the complexity of an investigated theme if they were able to integrate related studies. Unfortunately, individual studies are usually not designed to allow such integration, and the common statistical methods cannot be used for analyzing integrated data. The modeling technique of Bayesian networks has gained popularity in many fields of application due to its ability to deal with complexity, but has emerged only recently in food science. This thesis used data from experiments on sensory satiation as case studies. The objective was to explore the use of Bayesian networks to combine raw data of independently performed but related experiments to build a quantitative model of sensory satiation. Methods This thesis started with introducing the theoretical background of Bayesian networks to food science. The available data from various independent experiments on sensory satiation were then examined for their potential to be combined. Finally, the outcomes obtained using Bayesian networks on a single dataset were compared with the published outcomes of the respective study, in which classical statistical procedures were used to analyze the data. Results Two hurdles were identified when combining the data of related studies that were performed independently and without the intention of combining their data. The first hurdle was a lack of essential information for reliable estimations of parameters of the combined model network. This information could be obtained by deriving it from existing information in the individual studies or by performing extra experiments; these practices are, however, not always possible. The second hurdle was a possible conflict in causal relationships underlying the individual experimental designs, which can cause misleading analyses of the combined dataset. This was the case for some experiments that involved the control of secondary explanatory variables. As such, an approach termed as Global Experimental Design was proposed in this thesis as a solution to overcome these hurdles. This approach emphasizes the building of an overall network prior to designing individual studies. In comparison to using the classical statistical procedures, more information can be extracted using Bayesian networks. This technique could make use of the domain knowledge in a transparent manner as well as empirical data with missing values. Conclusions It is possible to combine raw data from related studies for a meaningful analysis if effort is made in the phase of experimental design. The approach of Global Experimental Design outlines this phase with the building of an overall network. By using Bayesian networks as a tool for exploratory analysis, scientists are able to gain more insights into a research domain.
There are no comments yet. You can post the first one!
Post a comment
Please log in to use this service. Login as Wageningen University & Research user or guest user in upper right hand corner of this page.