Talks to the LGC
1. January, 14th., 2026, Peter Froggatt Centre , QUB.
Dr. Lucas Kook, University of Vienna, Austria, gave a talk to the Local Group entitled: Probing Causal Models: Criticism, Falsification, and Sensitivity Analysis
Lucas opened the talk by discussing elements of the Smoking and Lung Cancer debate, early in which the 'causal' role of cigarette smoking was challenged by eminent statisticians like Burch, and Fisher who argued for a common underlying agent which caused both. While such a mechanism was possible no evidence for its existence was found and by the late 1960s the debate was settled in favour of smoking as the evidence from observational studies accumulated. In 2023, Philip Dawid called this the “biggest success story of causal inference”. Today, we know that smoking is only one cause of lung cancer, but it is by far the most important. A major lesson to be learned here is that: 'Causal models are credible only to the extent that they withstand criticism, falsification, and sensitivity analysis.
He next outlined a general causal model as a set of distributions P with Pobs ∈ P, such that the observed data X follows Pobs. It extends statistical models to describe distribution of X under different interventions..Of course, one had to be able to intervene on X, but in some circumstances this was not possible.
In relation to model criticism we should. if the causal model is true, consider whether is it consistent with our domain knowledge? What are its testable implications and can the resulting conditional independences be tested? Falsification tests were one way to operationalise model criticism. When the casual model was wrong we should examine how sensitive are our conclusions in relation to: (a) untestable model assumptions, and (b) robustness against violations of “no unobserved confounding”.
Lucas went on to introduce structural casual models (SCM) more formally noting that 'Observing' and 'Doing' were different operations. An SCM consists of a collection of d noise variables with distribution PN and d structural assignments Xj := fj(pa(Xj), Nj) for j ∈ {1, . . . , d}, where pa(Xj) ⊆ {X1, . . . , Xd} \ {Xj} are called parents of Xj. The SCM induces a graph, the noise variables are assumed to be jointly independent, Interventional distributions are obtained by replacing structural assignments and it is assumed that interventions on all variables are possible with changing anything else. With these arrangements, the Local Markov Property holds for the governing variables X1, . . . , Xd, namely, for all j ∈ {1, . . . , d}, Xj ⊥⊥ nd(Xj) | pa(Xj) and provides a framework in which testable implications may be explored in order to potentially falsify a causal model. (Here, nd(Xj) are the non-descendents of Xj).
Next he moved on quickly to illustrate the method on an interesting case-study involving a casual model of protein interaction network showing how the falsification testing could be used to refine the structure of the causal model, referring, en passant, to the COMETS package (Kook & Lundborg, 2024) available on CRAN for this purpose.
Finally, he turned to causal sensitivity analysis (details omitted) and illustrated this idea by returning to the smoking and lung cancer debate and to Cornfield's Inequality which addresses the question: How strong would a hidden confounder’s (U) influence on exposure (X) and outcome (Y) need to be to explain away the observed association. If the observed association is RRXY = 10, then it transpires that RRUX ≥ 10 and RRUY ≥ 10 are needed to explain away the association. For instance, if the “smoking gene” occurs with prevalence 2% among non-smokers, then at least 20% of smokers would have to carry it.
This talk, a breath of fresh on the subject of causal analysis, was received with acclaim by the audience which appreciated the tight logical development and the questioning, sometimes, sceptical, nature of the exposition. The speaker was thanked in the usual way.
Lucas handled several questions from the audience mainly about relying on methodology which involved untestable statistical assumptions. It was, for example, well known that David Cox advised making minimal assumptions and testing carefully the validity of those that were made. The rather naïve assumption of no unmeasured confounding was thought unlikely to apply in many healthcare settings. Such shortcomings made the methods proposed in the talk essential. When the discussion abated Lucas was thanked again for a very stimulating talk.
Key Reference:
L. Kook and A. R. Lundborg. Algorithm-agnostic significance testing in supervised learning with multimodal data. Briefings in Bioinformatics, 25(6), 2024. doi: 10.1093/bib/bbae475
2. February 11th, QUB.
Michelle Cowley-Cunningham, formerly of Dublin City University. gave a talk to the Local Group entitled: The innocent v the fickle few. How psycholinguistics helps explain jurors’ understanding of DNA evidence as a Random-Match-Probability (RMP) following judges’ directions
Dr Cowley-Cunningham began by pointing out the relevance of language, context and Judges' directions to the interpretation of evidence and hence to outcome in criminal trials. She pointed in particular to the work of the Innocence Project (1992) founded in the Benjamin N. Cardozo Law School, New York. Its published studies highlighted examples of the importance of Experimental Psychology – Conditions, (b) Single v Multiple Exemplar Cueing (e.g., use of ‘nonetheless’ - meaning "in spite of what has just been said"), (c) Refutations + Single v Multiple Exemplar Cueing and (d) Judges’ Directions. In all, the project had helped free more than 250 innocent people from prison.
Next, Michelle considered the example of Madeleine McCann and DNA in the Media. Madeleine went missing on 03 May 2007 and in September 2007, UK tabloids published headlines claiming that DNA evidence found in the boot of a hire car was a "100% match" to her DNA. This was a very strong claim and Michelle went on to explain that the interpretation of DNA evidence was not context free. In the Madeleine case this had had the effect of wasting time and misdirecting police resources.
There were other potential concerns, jurors could have difficulty with the concept of random match probabilities (RMPs). For example in a city with a population of 1 million (106) people and a RMP of 1 in 100,000 (105) ten potential suspects are expected, raising the idea of a reference set of suspects. The interpretation of RMPs could also lead to confusion as the idea of coincidence being in play (innocence) was less likely with higher RMPs. Moreover, psychological experiments in other fields had shown that the units in which RMPs are expressed can to lead to confusion. For example, patients with a mental illness were judged more likely to pose a risk to the public when released if the risk was presented as a frequency statement rather than a probability: i.e., 1 in 20 rather than 0.05. (Slovic, Monahan & MacGregorm, 2000). In addition, there was published evidence that narrative was important and that cueing and judges' directions could influence juror's decisions.
Michelle went on to describe her pilot study, which focussed on examining the role of exemplar cueing in which DNA random-match-probabilities are presented in the presence or absence of exemplar prompting linguistic cues. She randomly allocated 64 members of the general public (37 women and 27 men) to one of 4 experimental categories concerning a jewellery store robbery vignette involving a male suspect. These were: (1) Single Exemplar ‘You learn that the chance that the suspect would match the blood drop if he were not the source is 0.0001%’, (2): Multiple Exemplar ‘You learn that the chance that the suspect would nonetheless match the blood drop if he were not the source is 0.0001%’, (3) Single Exemplar + ‘You learn that there is a record of the suspect's debit card being used at a petrol station on the other side of town, and the card was used at the same time as the attempted robbery’, (4) Multiple Exemplar + ‘You learn that there is a record of the suspect's debit card being used at a petrol station on the other side of town, and the card was used at the same time as the attempted robbery’. The outcome measure was Guilty, Undecided and Not Guilty, and the main result is shown in the Figure, where a clear shift to Not Guilty conclusions and indecision is apparent with increased cueing and refutational evidence. Michelle presented other tables on judges' cueing with broadly similar results (see recording for details).
This was an excellent talk, measured, and very topical. The audience thanked Michelle in the usual way. Michelle then dealt with questions and comments, The concept of the reference set was commented on. One wondered how the suspect before the court had been selected? The controversial Letby case was raised in this context and the behaviour of the some Metropolitan Police officers was raised. The British Justice System was assumed to be fair and self-righting, but, increasingly, these assumptions were being questioned. The absence of a database logging details of judicial errors over time (Hillsborough and the Post Office Scandal being prime examples) was seen as an barrier to improvement.
When the discussion terminated Michelle was thanked, enthusiastically, again.
Key Reference
Cowley M. The Innocent v The Fickle Few': How Jurors Understand Random-Match-Probabilities and Judges' Directions when Reasoning about DNA and Refuting Evidence. J Forensic Sci & Criminal Inves. 2017; 3(1): 555601. DOI: 10.19080/JFSCI.2017.03.555601
3. April 29th, QUB.
Jessica Doohan, Department of Mathematics, University of Limerick, gave a talk to the Local Group entitled: Comparing and combining of generalised additive models and neural networks in practice.
Jessica opened her talk by reviewing the structure of Linear Regression (LR) models and Generalised Additive regression Models (GAMs) in which the fns are non-linear possibly non-parametric functions (e.g. splines). She compared these with the structure of a classical NN involving input and hidden layers.
GAMs could capture non-linearities and had the advantage of retaining interpretability, but were limited by the additive structure and the need to pre-specify interactions. NNs on the other hand were highly flexible and could capture interactions well, but were difficult to interpret and prone to over-fitting. Accordingly the key research questions are (a) which model performs best and (b) what are the factors influencing their performance?
To answer these questions Jessica spent some time describing in depth a PRISMA-based systematic Review of this area using the Scopus citation database (details omitted). It resulted in identifying 445 papers of which 143 met the comprehensive study eligibility criteria leading to the inclusion of 430 data sets for analysis. As the primary comparison was between GAMs and NNs various performance metrics were used in the studies. It was found that RMSE, AUC and R2 covered 90% of the datasets and Jessica focussed on the log ratio statistics defined, as: loge(RMSENN/RMSEGAM), loge[(1-AUCNN)/(1-AUCGAM)] , and loge[(1-R2NN)/(1-R2GAM)], respectively. These have the property that loge(ratio) > 0 ==> NN and loge(ratio) < 0 ==> GAM. Mixed-effects modelling was adopted to account for dependent datasets arising in the same paper. Thus, Paper ID was the random effect and the fixed effects were: Year, Domain area, Sample size, Number of predictors and NN complexity. Overall, the results were inconclusive but the trend in time among studies with smaller sample sizes was towards better performance using NNs although GAMs remained a strong and interpretable option.
Next Jessica introduced a major technical innovation by considering a mixture of regressions with 2 components which could follow any two of an LM or GAM or NN models viz:
where the μik follow:
She went on to describe methods of estimation using the E-M algorithm, highlighting a clever updating approach using Weighted Least Squares and finally discussing the complex patterns of data which the Mixtures of Regression model could handle comfortably (extensive details in recording & paper).
This was a tour de force of statistical modelling and the audience showed their appreciation in the usual way. Jessica dealt with a number of questions about the underlying modelling assumptions and possible extensions and concluded by saying that this research was work in progress and that new models were still under development. The audience congratulated her again on an excellent presentation and talk.
The Chair thanked everyone for attending and said that the next talk would be in October after the Summer break. He reminded the audience about the forthcoming CASI conference in Portlaoise ( May 11th - May 13th).
Key Reference
Doohan, J., Kook, L., & Burke, K. (2026). Comparison of generalised additive models
and neural networks in applications: A systematic review. Expert Systems with
Applications, 131082.
Event report by Professor Gilbert MacKenzie
October, 2026.