Using sensitivity analyses in Bayesian Networks to highlight the impact of data paucity and direct future analyses: a contribution to the debate on measuring and reporting the precision of likelihood ratios

- We demonstrate a resampling method for carrying out sensitivity analyses on observational data used within Bayesian networks
- The results of sensitivity analyses can be used to inform an analyst of where further work will have its greatest impact
- The results of sensitivity analysis may also indicate whether the basis for an opinion is robust
- We describe the interpretation of sensitivity analysis results and the difference to classic frequentist sampling variation

Bayesian networks are being increasingly used to address complex questions of forensic interest. Like all probabilities, those that underlie the nodes within a network rely on structured data and knowledge. Obviously, the more structured data we have, the better. But, in real life, the numbers of experiments that can be carried out are limited. It is thus important to know if/when our knowledge is sufficient and when one needs to perform further experiments to be in a position to report the value of the observations made. To explore the impact of the amount of data that are available for assessing results, we have constructed Bayesian Networks and explored the sensitivity of the likelihood ratios to changes to the data that underlie each node. Bayesian networks are constructed and sensitivity analyses performed using freely available R libraries (gRain and BNlearn). We demonstrate how the analyses can be used to yield information about the robustness provided by the data used to inform the conditional probability table, and also how they can be used to direct further research for maximum effect. By maximum effect, we mean to contribute with the least investment to an increased robustness. In addition, the paper investigates the consequences of the sensitivity analysis to the discussion on how the evidence shall be reported for a given state of knowledge in terms of underpinning data.