Eutanasia, ensañamiento terapéutico y muerte digna

CFA is to be performed using the second sub-samples, to cross validate factors derived from EFA. CFA can assess the factor structure’s quality through testing the overall model and item loadings significance and goodness-of-fit of alternative models (Joreskog & Sorbom, 1989). The multiple iterations of CFA, with the maximum likelihood estimation (MLE) method further purify items within each construct.

Unfitted items would be deleted from the measurement model. This further strengthens both internal and external consistency of the scale items (Sethi & King, 1994). Often, CFA needs to be done through statistical software such as AMOS and LISREL.

The initially hypothesised model requires modification where applicable (Hair et al., 2010). The modification can be performed based on indicators such as modification indices (MI), standardised residuals, path estimates, squared multiple correlations (SMC) and qualitative review. The MI provides a mean to improve an initially specified model that does not fit the data satisfactorily. While there is no threshold indicated, Raykov and Marcoulides (2006) suggested that researchers may consider making changes to parameters associated with the highest MI. It is advisable to change the parameters one at a time, starting with the largest MI. Refer to Table 3.13 for the summary of other requirements for model diagnostics. The model diagnostics are essential to suggest model changes through an empirical trial-and-error approach (Hair et al., 2010).

140 Table 3.13: Model Diagnostics Requirements

Despite no common guidelines or consensus on which goodness-of-fit indices to use, there are at least 30 of the indices that are now available (MacKenzie et al., 1991).

However, Marsh et al. (2010) and Sweeney and McFarland (1993) suggested that item loadings, adjusted goodness-of-fit indices, Chi-square significance, degrees of freedom, Root Mean Square Error of Approximation (RMSEA), Tucker-Lewis Index (TLI), Comparative Fit Index (CFI) and results from competing models should be presented.

Hair et al. (2010) recommended at least one of the goodness-of-fit (GoF) indices for the following measures should be reported:

(i) Absolute measure - χ²/df, p-value, CMIN/DF, GFI, RMSEA, PCLOSE (ii) Incremental measure – CFI, TLI, NFI

(iii) Parsimony – PNFI, PCFI, PRATIO

Absolute measure fit indicates how well a structural equation model explains the relationships derived from the sample data (Worthington & Whittaker, 2006). Chi-square, commonly used as a test statistic where non-significant p-value is desired, it has been argued by many experts that it should be used as a measurement fit instead (Joreskog & Sorbom, 1993). A smaller Chi-square indicates a better model fit. Even though a closer Chi-square value to the degrees of freedom is deemed to indicate a good model fit (Thacker, Fields, & Tetrick, 1989). Despite Taylor and Todd (1995) suggesting that Chi-square value of five times larger than the degrees of freedom while

Model Diagnostic Requirement

Standardised Residuals < 2.5 no problem

> 4.0 possible problem Path Estimates (Constructs to Items) min. 0.5, ideally ≥ 0.7 ; and be significant Squared Multiple Correlations (SMC) or Reliability ≥ 0.4

141 Carmines and McIver (1981) suggested two to three times larger is acceptable, Ullman (1996) believed that maximum two times higher should be the rule of thumb. The Normed Chi-square (CMIN/DF) of less than three indicates a good absolute fit (Hair et al., 2010).

Browne and Cudeck (1993), Hair et al. (2010), Marsh et al. (2004) indicated that the maximum value of RMSEA that can be accepted is 0.08 whereas anything below 0.05 indicates a close fit. A PCLOSE of above 0.05 is desirable. Even though Mulaik et al. (1989) suggested the use of fit Index (GFI) and Adjusted Goodness-of-fit Index (AGFI), both indices have been widely argued to be much affected by sample size, number of indicators and not sensitive to detecting miss-specified models (Sharma et al., 2005). Several studies discourage the use of these two measures in determining the model fit due to its adverse effects over several factors (e.g. Hu & Bentler, 1998, 1999; Sharma et al., 2005).

Incremental measures compare a specific structural equation model to a baseline structural equation model in order to improve a model’s fit to the data (Worthington &

Whittaker, 2006). The typical baseline comparison model is the null model in which all the variables are independent of each other or uncorrelated (Bentler & Bonnett, 1980).

A higher fit indices value indicates a better model (Meyer et al., 1993). Hinkin (1995, p.

976) reported that most of the articles that he analysed used 0.85 threshold as an acceptable value for incremental fit indices. Bentler and Bonett (1980), Hair et al.

(2010) and Marsh et al. (2004) suggested 0.9 for the indices of TLI and CFI. However, for a larger model (more than 24 indicators) and smaller sample size situation (around 200), Sharma et al. (2005) proposed a threshold of 0.8 as acceptable.

Parsimony measure serves as a criterion for choosing between alternative models. Mulaik et al. (1989) suggested the Parsimony Goodness-of-Fit Index (PGFI) and the Parsimonious Normed Fit Index (PNFI). The PGFI is based upon the GFI by

142 adjusting for loss of degrees of freedom while the PNFI also adjusts for degrees of freedom however it is based on the NFI. Mulaik et al. (1989) indicated a complex model will probably lower the value of these indices substantially as compared to other GoF indices. As there is no specific threshold for these two indices, Mulaik et al. (1989) suggested that it is possible to obtain parsimony fit indices within the 0.50 region.

Besides the use of absolute, incremental, and parsimony measures in evaluating model fit, Hoelter (1983) suggested the use of Hoelter's critical N to measure the sample size adequacy. A Hoelter's N value above 200 is desirable (Hoelter, 1983) while 75 is the minimum value to ensure an acceptable model fit (Kenny, 2014). Table 3.14 displays the guidelines to some of the useful model fit indices. It is important to note that these cutoff values are just rough guidelines and not sufficiently supported by empirical evidence. In reality, large models with at least five factors and 50 items that could not meet the minimal acceptable guidelines of fit are a norm (Marsh, 2007; Marsh et al., 2005).

Table 3.14: Model Fit Indices

Measures Fit Indexes Acceptable Level

Absolute Chi-square (χ²) < 2 times of df

Normed Chi-square (CMIN/DF) < 3

Root Mean Square Error of Estimation

(RMSEA) ≤ 0.08

p of Close Fit (PCLOSE) > 0.05

Incremental Tucker-Lewis Index (TLI) ≥ 0.9

Comparative Fit Index (CFI) ≥ 0.9

Parsimony PNFI > 0.5

Sample Size

Adequacy HOELTER .05 ≥ 75

HOELTER .01 ≥ 75

143 3.9.9 Structural Model Assessment and Hypotheses Testing

Upon assessment of the measurement model as discussed above, the analysis proceeds with structural model assessment and hypothesis testing. To Anderson and Gerbing (1988), this procedure is known as Two-Step Modelling Approach. While the measurement model assessment is to evaluate convergent and discriminant validities (Campbell & Fiske, 1959); the structural model assessment is to conduct the nomological validity (criterion-related validity) (Campbell, 1960; Cronbach & Meehl, 1955; Schumacker & Lomax, 2010). In general, this validity is to measure the relationships between the constructs of the study (Anderson & Gerbing, 1988). A strong relationship indicates the newly developed scales are a good measurement tool in predicting future performance of the study subject (Spector, 1992). A combination of measurement model and structural model assessments provides a comprehensive and confirmatory evaluation of the construct validity (Bentler, 1978).

Anderson and Gerbing (1988) suggested that the researcher may first evaluate if any structural model possesses acceptable goodness of fit indices. A small chi-square value is desirable (Bentler & Bonett, 1980) as explained in Section 3.9.4.1.5 above. In order to obtain a good cross-validation of covariance structures (Cudeck & Browne, 1983), researchers are advised to split the original samples into two distinct sets of samples. The first set of the samples is to develop the measurement model while the latter set of samples is to validate the solution derived from the first set of samples (Anderson & Gerbing, 1988; Ashill & Jobber, 2010).

3.10 Chapter Summary

This chapter displays an array of discussion on the chosen research methodology approach, methods, design, data collection and analysis techniques employed in this research study. The issues addressed are in response to the research questions and

144 objectives identified in Chapter One and literature reviews in Chapter Two. This research adopts a two-phase sequential method. Phase one employed qualitative approach through individual in-depth interviews in data collection. This method is recommended for new scale development. This study follows the scale development procedures as suggested by Ashill and Jobber (2010), Churchill (1979), DeVellis (2003), Hinkin (1995), Malhotra (2007), and Nunnally and Bernstein (1994). The procedures cover four stages:

Stage 1: Specify domain of the construct (Phase 1) Stage 2: Items Generation(Phase 1)

Stage 3: Scale Development (Phase 2)

Stage 4: Scale Evaluation / Validation (Phase 2)

The findings in phase one and the literature reviews act as the basis to form the questionnaire in the second phase of quantitative studies. Non-probability sampling approach was used with purposive sampling through snowballing method in the first phase of the study. In the second phase, the researcher continued employing non-probability sampling methods through snowballing, purposive intercept sampling, and convenience sampling. Collected data were analysed by various statistical analysis techniques, as discussed. NVIVO was used to carry out a content analysis on the in-depth transcription of the study’s first phase. In the second phase, the key statistical techniques employed were EPA, CFA, and SEM in an attempt to ensure the reliability and validity of the constructs, test the research framework and the proposed relationships, as well as confirm the model fit. The findings are presented in the following Chapter 4, Chapter 5, and Chapter 6. Discussions of the implications and suggestions on directions for future research are the subject of Chapter 7 before making the final conclusions of the research.

145 CHAPTER 4: DATA ANALYSIS - QUALITATIVE FINDINGS

4.1 Introduction

As mentioned in Chapter One, this study aims to evaluate the relationship among the theoretical constructs of push motivations (PUSH-M), pull motivations (PULL-M), transnational behaviours (TB), overall satisfaction (OVS), and post-satisfaction intentions (PSI). In doing so, based on the inter-relationship of the constructs, the researcher will propose the ‘Second Home Retirement’ (SHR) Model, as an ultimate research goal. This chapter is the first data analysis part stipulated in stage two of the scale development process. Section 4.2 first displays the qualitative findings from the in-depth interviews. Next, the items generation of push motivations (PUSH-M), pull motivations (PULL-(PUSH-M), and transnational behaviours (TB) is presented.