Lecture 11: Reporting and communicating Bayes factors Communication of numerical results where the context is typically non-numeric • The end-users of statistical inference are vey often less convenient with using numbers (quite typical for lawyers) • Understanding of probabilistic reasoning should not be taken for granted • We need tools to demonstrate what a Bayes factor means • Moreover, Bayes factors are often difficult to assign (estimate) with high precision • Reporting the magnitude of a Bayes factor on an ordinal scale is often to be preferred in opposite to providing its numerical value The forensic evidence evaluation is made in a context represented by two competing propositions/hypotheses Hm : The incriminating proposition (often referred to as the main proposition) • can state a specific source of a recovered material • can state a specific activity that gave rise to the recovered material • can state that the recovered material belongs to a certain category (forged pass-ports, ignitable liquids, toxic materials,…) Ha : The alternative proposition. Comprises the relevant opposite to Hm The scale of conclusions used at the Swedish National Forensic Centre (NFC) Level Magnitude of the Bayes factor/likelihood ratio Verbal equivalent +4 at least one million extremely strong support for Hm +3 between 6000 and one million +2 between 100 and 6000 +1 between 6 and 100 support to some extent for Hm 0 between 1/6 and 6 neither support for Hm nor for Ha –1 between 1/100 and 1/6 –2 between 1/6000 and 1/100 –3 between 1/(one million) and 1/6000 –4 at most 1/(one million) strong support for Hm support for Hm support to some extent for Ha support for Ha strong support for Ha extremely strong support for Ha Symmetric scale: The intention is to answer – as far as possible – the question put by the commissioner. This question is what is usually reformulated as the main proposition Hm. Why the limits 6, 100, 6000 and 1 million? Should the number of levels have been 7 on both sides of zero, then a logarithmic scale would have been a proper choice. The levels here were obtained by requiring that • level +2 should give a posterior probability P(Hm | E ) of at least 0.99 when the prior probability P(Hm) is at least 0.5 • the limit for level +4 should be one million – corresponds with a reasonable limit for DNA evidence with respect to a Swedish population …and by then forcing a logarithmic increase between levels Pros and cons with an ordinal scale of conclusions compared to reporting the explicit Bayes factor/likelihood ratio Pros: • The use of a scale enables the use of elicited Bayes factors/likelihood ratios with low precision – it is sufficient to ensure that a certain limit is crossed • The end-users get used to a limited number of evidentiary “strengths” and do not have to consider numerical differences from case to case – What is the significant difference between a Bayes factor of 350 and a Bayes factor of 300 ? • As was previously mentioned, a majority of forensic casework types still lacks explicit data to assign Bayes factors with a high accuracy Cons: • The interval lengths of the scale are by necessity strongly increasing with level Large Bayes factors (e.g. around 100000) become underestimated by just ensuring the lower level Who is the end-user? In a substantial part of the forensic literature the enduser is referred to as the court, the jury or the judge In Sweden, it is not the judge (or the court) that requests the forensic analysis – it is the leader of the preliminary investigation, very often a prosecutor. Hence, who should at first understand the conclusions from a forensic analysis is the prosecutor – in order to deem on its value for the prosecution. When educating and communicating with an end-user… …it is difficult to centre on the meaning of the Bayes factor/ likelihood ratio! In general, when a request of a forensic analysis is made, the commissioner would expect a categorical answer – Was it this gun that fired the bullet? – The forensic findings support that the gun fired the bullet (level +2). – Do you mean that you are not 100 % sure that it was this gun? – I do not address the proposition, only the forensic findings under each proposition. They are less probable if there was another gun that fired the bullet. – Then, since you have up to +4 in your scale, shall I take it that your +2 means you are 50 % certain that the bullet was fired from this gun? Instead of always addressing the likelihood ratio in the education, if we try to address Bayes theorem… – The reported conclusion +2 means that the forensic findings are at least 100 times more probable if this gun fired the bullet compared to if another gun fired it. Hence you should multiply your prior odds with 100. – What is that, prior odds??? – Your prior odds are how much you believe in that it was this gun compared to how much you believe it was another gun based on other information in your investigation. – But what I believe shouldn’t affect your conclusions, should it? My standpoint is that the conclusion from a forensic scientist should be objective and selfstanding. – It is objective, yes. But it can never be self-standing when it comes to the probability that this gun actually fired the bullet. It is always relative to the propositions. – I do not follow this at all? Must I be a mathematician to understand your conclusions? However scientifically sound Bayes theorem is when it come to reasoning about evidentiary strength we must never forget that… • Jurists very often – there are good exceptions though – are clever people that used to dislike mathematics in school • It is not yet easy (and not opportune) to quantify the prior odds of a proposition • Taking the reasoning up to court level would meet great cultural objections against the vision that guilt in a case should be verified using mathematical calculations • Today – and for quite a long while – the forensic laboratory cannot come up with accurate likelihood ratios in several forensic case types to “feed” the mathematical exercise These points, however, do not prevent us from… • developing methods for assigning Bayes factors/ likelihood ratios based on rigorous mathematics • showing theoretically how different choices of prior odds would affect the final conclusions about the propositions (e.g. via Bayesian networks) • trying to alleviate the mathematical exercise by illustrating it on simpler form It could be good to start with the ultimate goal of the request for a forensic analysis: There is a question: Was it this gun that fired the bullet? The question is (usually) put because the investigation indicates that this could be the case. Re-formulate the question as a proposition: Hm : This gun fired the bullet! The ultimate goal is to find the veracity (the posterior probability) of this proposition… …without disregarding the indication coming from the investigation itself!! Many people would agree on that when trying to prove that something holds, other possibilities must (successively) be excluded. Other possibilities are represented by an alternative proposition. That proposition may be: Ha : Another gun fired the bullet! …but it is important that the chosen Ha must comprise all relevant alternatives to Hm and… …that the commissioner and the laboratory works with the same alternative proposition! Graphical illustration of Bayes theorem P H h P H a P E H h P E H a prior odds P H h E P H a E posterior odds high high even even Bayes factor, B low low If B is higher than 1 Upward arrow The Findings support the main proposition Hm If B is lower than 1 Downward arrow The Findings support the alternative proposition Ha If B is equal to 1 Horizontal arrow The propositions are equally supported With one specific pair of main and alternative hypothesis the findings always give the same Bayes factor, i.e. the same angle of the arrow. The degree of veracity, however depends on where the arrow starts, i.e. what the prior odds are. prior odds posterior odds very high medium high medium high even medium low low Depicting the scale of conclusions with the arrow model Even prior odds Level Degree of veracity P(Hh | E ) +4 : +3: +2: +1: > 0.999999 > 0.9998 > 0.99 > 0.86 0: between 0.14 and 0.86 Lower limit of the Bayes factor 106 B 6000 B 100 B 6B (1/6 < B < 6) When the evidentiary strength is reported in terms of a level of this scale it may suffice for the commissioner to consider whether the prior odds are “lower than” , “about equal to” or “higher than” 1. The magnitude of the posterior probability may still be appreciated. Nordgaard A., Ansell R., Drotz W., Jaeger L. (2012) Scale of conclusions for the value of evidence. Law, Probability and Risk 11(1): 1-24. And when the alternative proposition is changed?... There may be subsequent arguments for a more narrow proposition: Ha2 : Another gun of same calibre as the seized gun fired the bullet! Since the propositions change, the likelihood ratio must change – it generally decreases with a narrower alternative. May be difficult to comprehend for someone who has taken the view that the forensic conclusion is “categorical”. But if you accept the principle of exclusion, you must admit that there is a difference between trying to exclude Ha and trying to exclude Ha2 The eventual degree of veracity may however be less influenced (sometimes not at all) which can be easily communicated with the arrow model: B with Ha B with Ha2 Not only the likelihood ratio changes with the alternative proposition, so does the prior odds.