CHAPTER 2: DEVELOPMENT OF MEASURES TO BE USED IN UNDERTAKING
2.4.1 Factor structure of the ERI-E
The changes made to the ERI-E have substantially improved its reliability. The additional items have meant a scale of 36 items has been constructed that is a moderate to good fit for the data, and the internal consistency is adequate for the subscales as shown in Table 9, particularly for the general community sample. There was little difference in fit between the two models, with both a single factor solution, and a multifactor solution having almost identical fit indices. This was in spite of fit indices being chosen which favour more parsimonious models. When there is no difference between the suitability of models choosing the more parsimonious model is often preferred, but in this case the fit indices already favour these models. Given the lack of a difference between the fit indices for these models a decision about which model is preferred was based on three factors; what previous research has shown in terms of everyday risk taking being a single domain or multiple domain construct; analysis of other psychometric properties, such as internal consistency, item loadings, and domain correlations to indicate which model is most suitable; and the face validity of the items in each domain.
In previous research a majority of the literature suggests that everyday risk-taking is a multifactor construct (e.g. Tulloch and Lupton, 2003; Weber & Blais, 2002). Although the domains of everyday RTB may require more development, the relatively consistent, but only moderate, correlations between the domains of the ERI-E (see Table 11) indicate they are measuring distinctly different aspects of everyday risk taking. The noticeably higher correlation of the risks involving personal danger domain with the overall scale indicates this construct plays a particularly important role in everyday risk taking.
In terms of the second factor (psychometric properties), the factor loadings for the six-factor model are higher than those for the single factor model, indicating that there is some degree of domain specificity in everyday risk taking. Item loadings for the single factor model average 0.52, and for the multiple domain model average 0.62 (for the community sample). In regards to the correlations between the factors (see Table 11), and with overall everyday risk taking, the relationships are as would be expected for a multi domain model, moderate correlations between individual domains, and in almost all cases the highest correlation being with the overall scale, rather than another domain. The highest correlation between subscales is 0.69, and the lowest is 0.39, so the shared variance between domains is between 15% and 50%. This shows a degree of convergent and discriminatory validity, that there is some shared variance indicates convergent validity, they share some common factor. That all subscale correlations are below 0.7 indicates discriminant validity, in all cases they share less than half of the explained variance with any
other subscale. Further evidence of discriminant and convergent validity is from analysing which subscales have the highest and lowest correlations. The lowest subscale correlation is between risks to others and social risks, which is expected as one is concerned about the effect on the individual, and the other the effect on other people. The highest correlation is between risks involving personal danger, and health risks. These are similar domains in that both involves physical damage to an individual, one from injury, the other from sickness, and therefore were expected to be more highly correlated than other domains. In general the level of internal consistency within domains is also adequate (particularly for the community sample), especially given the relatively low number of items in each domain.
In terms of the third factor analysis of the items indicates the items are appropriate for the domains, a measure of face validity. In most instances the items are clearly measuring an aspect of the domain they are expected to measure, for example all the items involving risks to
belongings clearly belong in this category.
Significant improvements were also made to the loadings of each item onto the relevant subscale from the previous research with the ERI-E, particularly for the community sample. Previously these were between 0.30 and 0.67 (Hunt, 2008), and have now improved to between 0.47 and 0.70 for the community sample, and are 0.33 or above for the other samples. These improvements have been made without significant increases being made in the average inter- correlation of individual items. The ERI-E therefore does not suffer from the problem of bloated specifics, where high internal consistency is due to high correlations between similar items (Cattell, 1973; in Kline, 2000). Individual items also have good psychometric properties, with every item having a full range of responses, and no items having extreme values for skewness or kurtosis. The means for a majority of items is close to the middle of the range, approximately 3. There were significant differences between the samples. The sample which was the best fit for the data was the community sample. This sample had the widest range of ages represented, and was closest to a random sample. The average age for this sample was also significantly higher than for other samples, and the higher internal consistencies may reflect a more consistent view of risk in the different domains that develops with age. These results are in contrast to research on the DOSPERT which found it more consistent for student samples than for community samples, (Blais & Weber, 2006a). This suggests that if the aim of research is to extrapolate research results into conclusions for the general population, the ERI-E is a better questionnaire to use than the DOSPERT, but the DOSPERT may be better for measuring risk taking relevant in tertiary student populations.
The worst fit was for the internet sample, but the worst internal consistencies were for the student sample. The internet sample size was low for a CFA, which may explain why the fit indices were poor (Kline, 2011). One problem with the fit indices used is that larger numbers of items and factors lead to smaller goodness of fit values, due to the omission of correlated error variance, and small theoretically insignificant factor loadings (Cheung & Rensvold, 2002). In summary, the ERI-E is a reliable and valid tool for use with general population samples, but there are problems with internal consistency for student samples, and to a lesser extent the internet sample, which may indicate difficulties for use with international participants. This is similar to the results of other measures of everyday risk taking where internal consistencies vary between groups, including between countries (Blais & Weber, 2006b). It is possible that
everyday risk taking is not perceived in the same way in different countries (even similar ones). Different measures may be required for different groups, in terms of both age and nationality. There were some differences in the internal consistency of the domain scales, but in general Cronbach’s alpha values were similar across scales, less so across samples. The most internally consistent scales varied between samples, with risks involving personal danger generally the most consistent, but for the internet sample risks to others had the highest Cronbach’s alpha value. As expected the internal consistency was highest for the single factor solution, given the larger number of items composing this scale (Huysamen, 2006).
The new scales had on average slightly lower correlations with the full scale measure than for previous research with the ERI-E (Hunt, 2008), but all values exceeded 0.5, and there were no correlations between subscales of less than 0.3. Social risk taking has the lowest correlation with the ERI-E for the student sample, but is of a similar magnitude to other domains. Correlations between domains were higher in the ERI-E than for the DOSPERT overall, but results from the DOSPERT were inconsistent. In one sample using the DOSPERT the social scale had non-significant correlations with all other subscales, and a correlation with the total score that was mostly explained by the items they had in common. In another sample the correlations between subscales were higher, but with most subscale correlations under 0.5 (Weber et al., 2002). The higher subscale correlations shows that the ERI-E domains measure a shared factor of everyday risk taking, something that is less certain with the DOSPERT. In terms of subscale correlation with overall everyday risks, the highest value was for the risks involving personal danger subscale, and the next highest was health risks (Table 11), which was similar to results from the DOSPERT (Weber et al., 2002). The lowest correlation was with risks to others, which also had low correlations with other subscales, particularly social risks and unknown risks. It may be therefore that risks to others may be seen differently to risks that
involve the individuals themselves. The correlation between risks to others and everyday risks was still above 0.6, which was not significantly lower than other domains correlation with the total ERI-E score, so it does seem to be measuring some aspect of everyday RTB.
The multi-factor ERI-E is a suitable tool for use in research. It has good psychometric
properties, with adequate reliability and validity. It is best suited to use with general population samples, with its suitability for use with student populations less certain, primarily due to the poor internal consistency of some subscales. More research is necessary to assess its suitability for use in different countries, but the evidence from this study is that it does have validity in countries other than New Zealand. Although the internet sample fit indices were lower the smaller sample size is a possible reason for this result. The domains used have some validity, but further research is required to provide further evidence of their construct validity,
particularly evidence of convergent and divergent validity. Studies with social phobia and OCD samples would aid this research, as these groups would be expected to have higher correlations with some domains than other, people with social phobics with social risks, and people with OCD with risks to others, and risks to health (depending on the nature of their OCD symptoms).
2.4.2 Factor Structure of the DASS-21
The results of the factor analysis in the present research are as good, or better, than that for previous samples using the DASS-21 (Henry & Crawford, 2005), so it is suitable for use with a New Zealand population. It includes reliable factors for both anxiety and depression, as well as being a reliable measure for the degree of general psychological distress. Although the
quadripartite model fits the data slightly better than the original model, there are some
psychometric difficulties with this that make it not so suitable. Almost half the items do not load adequately onto their specific factor once the loading onto the general psychological distress factor is taken into account. In particular the stress scale items have very low loadings, some of a low negative value. This is consistent with previous research, and the original reason for the development of the quadripartite model (Henry & Crawford, 2005). What this indicates is that items on the stress scale are measuring general psychological distress rather than a separate specific construct of stress. The DASS-21 may be more accurately seen as measuring three factors, subscales for anxiety and depression, and an overall factor of general psychological distress, measured by all 21 items. Stress is not a specific factor, and it is questionable whether this should be used as such in research.
The original model CFA has high factor loadings for all items, and a high correlation between factors. The fit for this model is good, once correlated errors are allowed. These are justified theoretically, as a number of items are quite similar. The items loadings values for the EFA are in general good, but there are some items which load highly onto both the anxiety and stress
scales. The two items which had low loadings onto their respective factors (Items 5 and 9) had similar loadings in past research (Henry & Crawford, 2005). This fits into the CFA results showing that the stress scale is more of a measure of general psychological distress than a scale measuring a separate construct. Results are similar to previous research, and the DASS-21 measures these constructs similarly in New Zealand as it does in other countries.
2.5. Conclusion
The results of this chapter show that is spite of some limitations the ERI-E and the DASS-21 are suitable measures to be used in the present research. The ERI-E is a reliable measure of
everyday RTB across multiple domains for community samples (in particular). The DASS-21 has been shown in previous research to be a good measure of people’s levels of anxiety, depression, and general psychological distress, and these results were replicated with a New Zealand general community sample. Therefore the ERI-E and DASS-21 will be used in the next two chapters to measure the relationships between anxiety, depression, and different domains of everyday RTB.