8 Técnico de Caja chica Recibe del encargado del FCMF DNCCPC los fondos de caja chica de las pólizas presentadas
FINAL DEL PROCEDIMIENTO
The proposed religious management of regret as one religious pathway to coping with personal death anxiety was only tested on Time 2 survey participants who still rued their identified unresolved major regret (n = 33). Hence, it was tested separately from the other two proposed religious pathways to coping with personal death anxiety, namely, afterlife beliefs and psychosocial maturity.
5.3.1.1. Structural Equation Modelling (SEM)8
Given that both afterlife beliefs and psychosocial maturity as two proposed religious pathways to coping with personal death anxiety had identical initial predictor (intrinsic religiosity) and final dependent variable (personal death anxiety), these two religious pathways were combined into one single model (see Figure 1a, p. 33). Since there were two latent mediators (afterlife beliefs and psychosocial maturity), this combined model was evaluated using SEM. EQS 6.1 for Windows was used to run the SEM.
Assessment of SEM model fit. The overall fit of a hypothesized SEM model to the observed data can be tested by several statistics. Four commonly used statistics are Chi Square, Non-normed Fit Index (NNFI), Comparative Fit Index (CFI) and Root Mean Square Error of Approximation (RMSEA). A non-significant Chi Square test, a value of .90 or greater for NNFI and CFI, and a value of .08 or smaller for RMSEA all indicate a good fit of the model to the observed data.
SEM model selection. SEM model selection aims to ‘select a plausible approximating model that balances model bias and sampling variance’ (Stanley & Burnham, 1999, p.
475). In comparison of two or more models, the model with smaller values of Akaike’s information criterion (AIC) and consistent version of the AIC (CAIC) represents a better fit of the hypothesized model (Hu & Bentler, 1995).
The baseline afterlife beliefs and psychosocial maturity joint mediation SEM structural model would be compared to another model with neuroticism being incorporated into the baseline model (see Figure 2, p. 36). The purpose of comparing these two models is to examine the significance of the confounding mediating effect of neuroticism on the negative relationship between intrinsic religiosity and personal death anxiety, over and above afterlife beliefs and psychosocial maturity.
Parametric SEM assumptions. Many statistical tests assume a normal distribution.
There are three typical indices of non-normality to evaluate the distribution of the data, namely, univariate skewness, univariate kurtosis, and multivariate kurtosis (Finney &
DiStefano, 2006). The general rule of thumb for the acceptable value of skewness and kurtosis that indicate univariate normality is within the +2 to -2 range, but some authors use the more lenient range from +3 to -3 while other use the more stringent range from +1 to -1 (Garson, 2010). In this dissertation, I chose the +2 to -2 range criteria.
In addition to testing univariate normality, parametric SEM requires that the observed data has a multivariate normal distribution (cf. Finney & DiStefano, 2006). The
assumption of multivariate normality implies that each variable is normally distributed with respect to each other in the correctly specified model. Assessments of univariate skewness and kurtosis only serve as an initial test for multivariate normality. If any of the observed scores indicate substantial univariate non-normality, then multivariate normality cannot be assumed (West, Finch & Curran, 1995). However, even if all of the observed variables have univariate normal distribution, the joint distribution may not be multivariately normal (West et al., 1995). Mardia’s statistics is a measure of multivariate kurtosis, which is very important for testing this SEM parametric assumption. A value of Mardia’s statistic less than 3 indicates that the data are multivariate normal (Finney & DiStefano, 2006; Garson, 2010).
Several remedial estimation techniques approximate X2 and estimates of standard errors more closely to their true values to deal with non-normal data9 for a SEM measurement or structural model. For moderately non-normal data, maximum
likelihood based on normal theory estimation is fairly robust to these conditions (Finney & DiStefano, 2006). West and colleagues (1995) recommended the Satorra-Bentler statistics (S-B Scaling) for sample size smaller than 200 and for distribution with moderate departure from normality (e.g., skewness = 2, kurtosis = 7).
It was decided that if the assumptions of normality were supported, SEM with
maximum likelihood estimation and listwise case exclusion (N = 133) was used. If the assumptions of normality were violated, non-parametric SEM with S-B Scaling and listwise case exclusion (N = 133) was used. Listwise is preferred to pairwise deletion with regard to SEM because “SEM requires that…sample size should be consistent across every element of the covariance matrix” (Williams, Gavin, & Hartman, 2004, p.
340).
5.3.1.2. Bivariate Correlations
To be consistent with SEM analysis, bivariate correlations with listwise case exclusion (N = 133) were used to examine the relationships among variables specified in the two hypothesized mediation models (Figure 1a, p. 33; Figure 2, p. 36). Similarly, bivariate correlation with listwise case exclusion (N = 137) was used to assess the test-retest reliability of the three death-related constructs (personal death anxiety, personal dying anxiety and perceived attraction to death).
Because the number of people reporting an unresolved major regret and who still rued it was low (n = 33), bivariate correlations with pairwise case exclusion were used to examine the proposed religious management of regret as the third religious pathway to coping with personal death anxiety.
Parametric assumptions. Skewness and kurtosis were used as measures of normality to determine the choice of correlational analysis. If the assumptions of normality were supported, then Pearson’s correlation was used. If the assumptions of normality were violated, then Spearman’s rho was used.
5.3.1.3. Hierarchical Linear Regression
Two hierarchical linear regressions with listwise case exclusion were run to identify any additional predictors of the two-factor personal death anxiety that were not specified in the final mediation model. Listwise deletion was chosen because it is “very robust to the violations of MCAR [missing completely at random] (or even MAR [missing at random])” for predictor variables for “virtually any kind of regression” (Allison, 2009, p.
75). In addition, listwise deletion usually produces accurate estimates of the true standard errors (Allison, 2009, p. 76).
It should be noted that because the hypothesized protective role of religious management of regret on lowering personal death anxiety only applied to Time 2 survey participants who still rued their unresolved major regret, the variance of personal death anxiety that were explained by religious management of regret would not be partialled out in the additional hierarchical linear regression.
First, bivariate correlations with listwise case exclusion (N = 124) were run to identify any significant correlates of personal death anxiety. Then, those significant correlates of personal death anxiety were entered into a hierarchical linear regression to select significant predictors that could explain an additional amount of variance of personal death anxiety, over and above what was accounted for by intrinsic religiosity, afterlife beliefs, psychosocial maturity and possibly neuroticism.
Assumptions of regression. For a regression model to generalise, four additional underlying assumptions must be met. The first assumption is the assumption of no multicollinearity, which means the assumption of no perfect linear relationship between two or more predictors. To meet the assumption of no multicollinearity, the average values of variance inflation factor (VIF) should not be substantially greater than 10 (Bowerman & O’Connell, 1990) and the value of tolerance must be above 0.2
(Menard, 1995).
The second assumption is the assumption of homoscedasticity, which means
predictors are assumed to have the same variances. Homoscedasticity is signified by a plot of regression standardised residual against regression standardised predicted residual with randomly and evenly dispersed points.
The third assumption is the assumption of independent residuals, which means the residual terms for any two observations are assumed to be uncorrelated. The Durbin-Watson test statistics value of 2 indicates that residuals are uncorrelated while a value less than 1 or greater than 3 signals correlations among residuals (Durbin & Watson, 1951; Field, 2009).
The fourth assumption is the assumption of normally distributed errors, which means residuals are assumed to be random. Normality of residuals is indicated by a histogram
depicting a bell-shaped curve of the distribution of the residuals and by a normal probability plot with observed residuals falling on a straight line.
If any of the additional regression assumptions were violated, then specific correction formulas were applied to address the corresponding violated assumption.
5.3.1.4. Independent t-Test or Mann-Whitney Test
Additional analyses were run to test for any difference in the levels of personal death anxiety between people who still rued a major unresolved regret and those who did not have a major lingering regret.
Parametric assumptions. In addition to normal sampling distribution, an independent t-test requires homogeneity of variance. Skewness and kurtosis were used as measures of normality. Levene’s test for equality of variances was used to test the assumption of homogeneity of variances. If the assumptions of normality and of homogeneity of variances were supported, then an independent t-test was used. If the assumptions of normality and/or of homogeneity of variances were violated, then a Mann-Whitney test was used.
5.3.1.5. Content Analysis for Qualitative Data
The content analysis involved reading through the written texts several times to discover emerging themes. Each theme represents a repetitive topic mentioned frequently by the participants and was labelled with a code. Next, the codes were sorted into broader domains based on their similarities.