Graphical Bayesian Network Research Articles

Econometric Identification of Causal Effects: A Graphical Approach with a Case Study

It is well known that causal inference relies on untestable a-priori causal assumptions. Identification refers to whether a causal relationship can be inferred from observed statistical associations; it requires an understanding of what statistical associations are induced by those causal assumptions. Since the assumptions are untestable, a transparent description of their statistical consequences helps the readers. However, the relation between causal assumptions and their induced statistical associations may not be obvious. In this paper I describe a technique known as Directed Acyclical Graphs or Graphical Bayesian Network or Graphical Causal Models. The technique was developed in the computer science literature in the 1980s (Pearl 2009) although it has antecedents in path analysis developed by Philip and Sewall Wright beginning in the 1920s (Wright 1921). In addition to describing the technique, I illustrate its application to a case study of a research issue in auditing.

Calculating and understanding the value of any type of match evidence when there are potential testing errors

It is well known that Bayes' theorem (with likelihood ratios) can be used to calculate the impact of evidence, such as a `match' of some feature of a person. Typically the feature of interest is the DNA profile, but the method applies in principle to any feature of a person or object, including not just DNA, fingerprints, or footprints, but also more basic features such as skin colour, height, hair colour or even name. Notwithstanding concerns about the extensiveness of databases of such features, a serious challenge to the use of Bayes in such legal contexts is that its standard formulaic representations are not readily understandable to non-statisticians. Attempts to get round this problem usually involve representations based around some variation of an event tree. While this approach works well in explaining the most trivial instance of Bayes' theorem (involving a single hypothesis and a single piece of evidence) it does not scale up to realistic situations. In particular, even with a single piece of match evidence, if we wish to incorporate the possibility that there are potential errors (both false positives and false negatives) introduced at any stage in the investigative process, matters become very complex. As a result we have observed expert witnesses (in different areas of speciality) routinely ignore the possibility of errors when presenting their evidence. To counter this, we produce what we believe is the first full probabilistic solution of the simple case of generic match evidence incorporating both classes of testing errors. Unfortunately, the resultant event tree solution is too complex for intuitive comprehension. And, crucially, the event tree also fails to represent the causal information that underpins the argument. In contrast, we also present a simple-to-construct graphical Bayesian Network (BN) solution that automatically performs the calculations and may also be intuitively simpler to understand. Although there have been multiple previous applications of BNs for analysing forensic evidence--including very detailed models for the DNA matching problem, these models have not widely penetrated the expert witness community. Nor have they addressed the basic generic match problem incorporating the two types of testing error. Hence we believe our basic BN solution provides an important mechanism for convincing experts--and eventually the legal community--that it is possible to rigorously analyse and communicate the full impact of match evidence on a case, in the presence of possible errors.

Bayesian inference of T Tauri star properties using multi-wavelength survey photometry

There are many pertinent open issues in the area of star and planet formation. Large statistical samples of young stars across star-forming regions are needed to trigger a breakthrough in our understanding, but most optical studies are based on a wide variety of spectrographs and analysis methods, which introduces large biases. Here we show how graphical Bayesian networks can be employed to construct a hierarchical probabilistic model which allows pre-main sequence ages, masses, accretion rates, and extinctions to be estimated using two widely available photometric survey databases (IPHAS r/i/Halpha and 2MASS J-band magnitudes.) Because our approach does not rely on spectroscopy, it can easily be applied to homogeneously study the large number of clusters for which Gaia will yield membership lists. We explain how the analysis is carried out using the Markov Chain Monte Carlo (MCMC) method and provide Python source code. We then demonstrate its use on 587 known low-mass members of the star-forming region NGC 2264 (Cone Nebula), arriving at a median age of 3.0 Myr, an accretion fraction of 20+/-2% and a median accretion rate of 10^-8.4 Msol/yr. The Bayesian analysis formulated in this work delivers results which are in agreement with spectroscopic studies already in the literature, but achieves this with great efficiency by depending only on photometry. It is a significant step forward from previous photometric studies, because the probabilistic approach ensures that nuisance parameters, such as extinction and distance, are fully included in the analysis with a clear picture on any degeneracies.

Using Object Oriented Bayesian Networks to Model Linkage, Linkage Disequilibrium and Mutations between STR Markers

In a number of applications there is a need to determine the most likely pedigree for a group of persons based on genetic markers. Adequate models are needed to reach this goal. The markers used to perform the statistical calculations can be linked and there may also be linkage disequilibrium (LD) in the population. The purpose of this paper is to present a graphical Bayesian Network framework to deal with such data. Potential LD is normally ignored and it is important to verify that the resulting calculations are not biased. Even if linkage does not influence results for regular paternity cases, it may have substantial impact on likelihood ratios involving other, more extended pedigrees. Models for LD influence likelihoods for all pedigrees to some degree and an initial estimate of the impact of ignoring LD and/or linkage is desirable, going beyond mere rules of thumb based on marker distance. Furthermore, we show how one can readily include a mutation model in the Bayesian Network; extending other programs or formulas to include such models may require considerable amounts of work and will in many case not be practical. As an example, we consider the two STR markers vWa and D12S391. We estimate probabilities for population haplotypes to account for LD using a method based on data from trios, while an estimate for the degree of linkage is taken from the literature. The results show that accounting for haplotype frequencies is unnecessary in most cases for this specific pair of markers. When doing calculations on regular paternity cases, the markers can be considered statistically independent. In more complex cases of disputed relatedness, for instance cases involving siblings or so-called deficient cases, or when small differences in the LR matter, independence should not be assumed. (The networks are freely available at http://arken.umb.no/~dakl/BayesianNetworks.)

An ontology-based approach for constructing Bayesian networks

Bayesian networks are commonly used for determining the probability of events that are influenced by various variables. Bayesian probabilities encode degrees of belief about certain events, and a dynamic knowledge body is used to strengthen, update, or weaken these assumptions. The creation of Bayesian networks requires at least three challenging tasks: (i) the determination of relevant variables (nodes), (ii) the determination of relationships between the identified variables (links), and (iii) the calculation of the conditional probability tables (CPTs) for each node in the Bayesian network. Based on existing domain ontologies, we propose a method for the ontology-based construction of Bayesian networks. The method supports (i) the construction of the graphical Bayesian network structure (nodes and links), (ii) the construction of CPTs that preserve semantic constraints of the ontology, and (iii) the incorporation of already existing knowledge facts (findings). The developed method enables the efficient construction and modification of Bayesian networks based on existing ontologies.

A Bayesian network approach for modeling local failure in lung cancer

Locally advanced non-small cell lung cancer (NSCLC) patients suffer from a high local failure rate following radiotherapy. Despite many efforts to develop new dose-volume models for early detection of tumor local failure, there was no reported significant improvement in their application prospectively. Based on recent studies of biomarker proteins' role in hypoxia and inflammation in predicting tumor response to radiotherapy, we hypothesize that combining physical and biological factors with a suitable framework could improve the overall prediction. To test this hypothesis, we propose a graphical Bayesian network framework for predicting local failure in lung cancer. The proposed approach was tested using two different datasets of locally advanced NSCLC patients treated with radiotherapy. The first dataset was collected retrospectively, which comprises clinical and dosimetric variables only. The second dataset was collected prospectively in which in addition to clinical and dosimetric information, blood was drawn from the patients at various time points to extract candidate biomarkers as well. Our preliminary results show that the proposed method can be used as an efficient method to develop predictive models of local failure in these patients and to interpret relationships among the different variables in the models. We also demonstrate the potential use of heterogeneous physical and biological variables to improve the model prediction. With the first dataset, we achieved better performance compared with competing Bayesian-based classifiers. With the second dataset, the combined model had a slightly higher performance compared to individual physical and biological models, with the biological variables making the largest contribution. Our preliminary results highlight the potential of the proposed integrated approach for predicting post-radiotherapy local failure in NSCLC patients.