Various approaches to statistical analysis have been applied in the context of judgement analysis research to explore judgement policies. Smith and Gilhooly (2006) used judgement analysis to explore general practitioners (GP) decision policies about whether to prescribe medication to treat patients with symptoms of depression, and compared the results achieved using two distinct analysis techniques. Regression analysis was used to determine the relative importance of different cues available to the GP and those that significantly influenced decision making. Smith and Gilhooly (2006) compared this to a ‘fast and frugal’ model (Gigerenzer & Todd, 1999), which suggests that people use minimal cues to make decisions and therefore it is not a case of how different information is weighted but rather which cues are ‘critical’ to decision making. Model fit was tested by examining how well the regression model and the ‘fast and frugal’ model could predict GP prescription decisions, with the regression model being found to be slightly more effective. The fact that the regression model was only minimally more predictive despite the additional information was used to support the value of the fast and frugal approach.
Smith and Gilhooly (2006) noted that the fast and frugal model works best when measuring dichotomous judgements (such as ‘Yes’ versus ‘No’), as these provide a definitive indication of the impact of each addition of a piece of cue information. This means the number of cues needed to result in a judgement outcome can be clearly identified. This approach would not apply well to the financial elder abuse judgement analysis given the continuous scale dependent variables. In addition, the cues in the
financial abuse case scenarios included a mixture of ordinal and dummy variables that could not be meaningfully translated to test a fast and frugal model.
Multiple regression analysis is the predominant approach taken in judgement analysis research, modelling participant’s judgement policy using their weighting of the cue information (Cooksey, 1996). The different cues can then be directly compared, identifying those with a significant influence on judgement. Results are often reported in research at a group overall level. Approaches to summarising regression analysis findings have included identifying the number of participants who gave each cue the strongest weighting (MacCormick & Parry, 2006) as well as reporting the average weighting of each cue across the group in graphical form (Harries & Gilhooly, 2003).
Conducting multiple regression analysis in the context of the financial elder abuse judgements involved a slightly different process and interpretation, due to the nature of the financial abuse cues; a number of which had to be recoded into dummy variables (see the example in Table 6.6). When conducting regression analysis with dummy variables, the unstandardized beta coefficients should be used in the regression equation rather than the standardised beta coefficient (Cohen, 1983). This is because the unstandardized beta coefficient represents the impact of a unit change in the cue (i.e. presence or absence of the dummy variable category), which is more meaningful than the standardized beta coefficient that represents the impact of a change of one standard deviation unit (Field, 2009).
In a regression equation with one dummy variable, unstandardized coefficients are interpreted as showing the change in dependent variable versus the reference category (Hardy, 1993). For instance, using the example in Table 6.6, the change in mean certainty of abuse where the identifier of the abuse was a family member versus the individual (‘You’). With multiple dummy variables, interpretation alters slightly, to show the impact of each category on the dependent variable, controlling for change in the dependent variable caused by all the other independent variables (Hardy, 1993). Although the interpretation of unstandardized coefficients is useful to establish the effects of each cue-category, it does mean that the regression beta weights cannot be compared to assess cue weighting.
In the financial abuse case scenarios, the impact of being in one particular cue- category versus another was examined using t-tests (Hardy, 1993). T-tests were
used to establish if there was a significant difference in certainty of abuse where a case involves one type of financial problem compared to another. The Bonferroni correction was applied to determine an adjusted significance level to account for the fact that multiple t-test comparisons were needed. The aim of this was to reduce the likelihood of a Type I error resulting in incorrect identification of a significant effect (Field, 2009). This approach was selected in preference to conducting a One-way Anova to compare the cue categories, because this would have required treating the cues as single independent variables rather than groups of dummy variables. Despite the resulting effect of having to adjust the significance level, it is hoped this resulted in a more consistent treatment of the cues throughout the analysis.
Although the cue categories had to be represented as separate dummy variables, it was also important to consider how the overall impact of each cue could be determined. For instance, looking at the overall effect of the financial problem suspected rather than comparing each type of financial problem. This issue was addressed by running multiple regression analyses excluding each cue group in turn. The R2 for the model without each cue group could then be subtracted from the R2 for
the model overall, to establish a ‘usefulness coefficient’ or squared semi-partial correlations (Cooksey, 1996). The significance of the R2 for each cue group
(usefulness coefficient) could then be tested by conducting an incremental F test (Hardy, 1993). This compares the R2 for the reduced versus the full regression model
to establish if there is a significant change in judgements as a result of the cue, having controlled for the other cues. Working examples of this are provided in the reporting of the results.
Multiple regression analysis was conducted to explore how the financial elder abuse cues influenced professionals’ judgements of certainty of abuse, as well as likelihood of action being taken. This approach was not used to explore the relationships between the financial elder abuse cues and the action choices. One option for such analysis could have been to conduct logistic regression analysis to predict each action choice (e.g. ‘Monitor situation’ either 'Yes' or 'No') based on the levels of the financial elder abuse cues. It was decided not to conduct this analysis because participants could select any combinations of the action choices, therefore, analysis of each individual action choice in isolation would not have reflected the nature of how professionals were able to select from the full range of actions.
Professionals’ choice of what actions they would take in response to financial elder abuse was instead analysed by considering the ‘highest’ level of action selected in response to each scenario. For instance, for the social care and health professionals, ‘Monitor situation’ was at the lower end of the action choices, whilst ‘Implement safeguarding procedures’ was the highest or strongest action that could be selected (See Table 6.3 for full details of the action options). The most common action choice was also examined, and in addition, correlation analysis was conducted to explore the relationship between both certainty of abuse and likelihood of action scores, and the number of actions and strength of actions selected.