The delights
of Bayesian probability reasoning are sufficient to draw from me another case
comment!
Our Court of
Appeal, in Manoharan v R [2015] NZCA 237 (11
June 2015), has said that a likelihood ratio of 20 (meaning 20 to 1, or in context, 20 times more likely under the prosecutor's hypothesis than under the defendant's) is “not strong” [52].
This was an
LCN DNA analysis and obviously compared to the usual DNA results of likelihood
ratios in the many millions, 20 is not high. But in the context of a case it
can be, as can be seen from the results of applying Bayes’ Theorem (this was a single-issue case involving independent items of evidence).
I should emphasise that the single-issue nature of this case ("who did it?", an actus reus issue) simplifies the use of Bayes' theorem. Where several elements of the offence charged are in issue, the theorem must be applied to each separately, and only evidence relevant to that issue is used in each application. This prevents, for example, a huge scientific LR on the issue of identity from swamping the issue of intention. For each issue the "probability of guilt" means the probability that that issue is proved to the standard required for it to be established.
The likelihood ratio used by the scientists in declaring a match is not necessarily the same as the likelihood ratio for the activity evidence used by the fact-finder in the trial. This is because the defence may have an innocent explanation for the match which brings the denominator of the likelihood ratio (the probability of getting the match on the assumption that the defendant is innocent) close to the value of the numerator. Then, the probative value of the test result for the prosecutor would be very reduced or even extinguished.
We can compare what the probability of guilt would have been if the scientific LR had been in the millions and the trial LR numerator approximately 1, with what the probability of guilt would be under the revised scientific LR and corresponding new trial LR, for priors-in-combination (by which I mean, the ratio of probability of guilt to probability of innocence, based on all the other evidence in the case and the starting assumption about the probability of guilt and the probability of innocence, for this single-issue case) of various levels. (For multi-issue cases the priors will be assessed separately for each issue.)
I should emphasise that the single-issue nature of this case ("who did it?", an actus reus issue) simplifies the use of Bayes' theorem. Where several elements of the offence charged are in issue, the theorem must be applied to each separately, and only evidence relevant to that issue is used in each application. This prevents, for example, a huge scientific LR on the issue of identity from swamping the issue of intention. For each issue the "probability of guilt" means the probability that that issue is proved to the standard required for it to be established.
The likelihood ratio used by the scientists in declaring a match is not necessarily the same as the likelihood ratio for the activity evidence used by the fact-finder in the trial. This is because the defence may have an innocent explanation for the match which brings the denominator of the likelihood ratio (the probability of getting the match on the assumption that the defendant is innocent) close to the value of the numerator. Then, the probative value of the test result for the prosecutor would be very reduced or even extinguished.
We can compare what the probability of guilt would have been if the scientific LR had been in the millions and the trial LR numerator approximately 1, with what the probability of guilt would be under the revised scientific LR and corresponding new trial LR, for priors-in-combination (by which I mean, the ratio of probability of guilt to probability of innocence, based on all the other evidence in the case and the starting assumption about the probability of guilt and the probability of innocence, for this single-issue case) of various levels. (For multi-issue cases the priors will be assessed separately for each issue.)
A perspective on the size of a LR of 20 is obtained from considering a case where the evidence is not given, compared to one where it is.
- If, without the evidence, the probability of guilt is 0.90, then with the evidence that probability increases to 0.99.
- If, without the evidence, the probability of guilt is 0.60, then with the evidence that probability increases to 0.96.
- And, if, without the evidence, the probability of guilt is 0.30, then with the evidence that probability increases to 0.89.
Update: on 28 October 2015 the Supreme Court refused leave to appeal: Manoharan v R [2015] NZSC 156. The Court could see no basis to indicate that the Court of Appeal had been wrong when it had held that there had been no miscarriage of justice.
Another update: Bernard Robertson has commented on this case in"Likelihood ratios in evidence" [2016] New Zealand Law Journal 22. He points out that the Court at [39] correctly describes the witness's evidence as stating a likelihood ratio, but earlier in the judgment, at [17], it had transposed the conditional when referring to the same evidence. Mr Robertson also points out that the "20 times more likely" assertion should have been clarified because juries will not know what to do with a likelihood ratio, and that in this case source level propositions (where the sample at the scene came from) were mixed with activity level propositions (how it got there).
