Chapter 5 Data Analysis and Results
5.3 Constructs and Their Items
5.3.4 Injunctive norm construct
Injunctive norm refers to one’s perceived beliefs of what an individual thinks others expect him/her to do. The construct was assessed from two different aspects: the IN_Perception construct, which had four items, was designed to measure the perceived expectations from others (e.g. family, friends, etc.); IN_Comply, which had four items, was designed to identify which people were perceived to influence the respondent’s decision making (e.g. family, friends, etc.). Figure 5.4 illustrates the causalities between the two constructs and their indicators. After confirming the capabilities of the indicators measuring the respective sub-constructs (i.e. IN_Perception and IN_Comply) of the injunctive norm construct, a product term was created based on the valid indicators as an index of injunctive norm.
The dimensionality and the internal consistency reliability tests of the four items from IN_Perception showed that the items were unidimensional, which suggested they were measuring the same factor. The Cronbach’s alpha of IN_Perception was .915, (a result >.70 indicates internal consistency reliability). The factor analysis also indicated that the four items grouped under the IN_Comply were unidimensional; their Cronbach’s alpha was .899, which was also greater than .70. In the convergent validity assessments, the congeneric measurement models of IN_Perception and IN_Comply indicated that the four items for IN_Perception were capable of measuring IN_Perception individually, but not the items for IN_Comply.
According to the assessment of the IN_Perception’s four-item congeneric model (as illustrated in Panel A, Figure 5.4), all four indicators’ factor loadings were significant for the IN_Perception construct. The fit indices of the IN_Perception’s congeneric model also indicated the model’s good fit for the data, showing that the items could measure the construct individually. The chi-square of the IN_Perception
congeneric model was 4.211 and the p-value was .122. The p-value of the bootstrapping was evaluated because the absolute value of the multivariate kurtosis was greater than four; the p-value was .917. These results suggested that the proposed relationship between the four items and the IN_Perception construct fit the sample collected from survey respondents. Other fit indices also led to the same conclusion. The SRMR of the model was .0121 and the values of its CFI and TLI were .989 and .966, respectively. The IN_Perception’s RMSEA and PCLOSE values were .060 and .320, which suggested that the discrepancy between the proposed congeneric model and the sample population was minor and not statistically significant. Therefore, these four indicators were used in the later CFA before creating the injunctive norm’s product term.
Figure 5.4 IN congeneric measurement models
Panel A
Panel B
For the IN_Comply four-item congeneric model (as illustrated in Panel A, Figure 5.4), all four items’ factor loadings towards the IN_Comply construct were significant. However, the fit indices of IN_Comply suggested that the proposed
relationship to the variables did not fit the sample population. The chi-square and the p- value of the IN_Comply congeneric model were 14.732 (df was 2) and .001, respectively, and the p-value of the model’s bootstrapping was .021. These results indicated that the difference between the congeneric model and the sample population could not be ignored and was statistically significant. The values of the model’s RMSEA and PCLOSE also supported this conclusion. The IN_Comply congeneric model’s RMSEA and PCLOSE were .143 and .009, respectively. However, the same conclusion was not reached for all the fit indices. The model’s SRMR, CFI and TLI values were .0240, .989 and .966, respectively, which suggested the model could fit the sample population (Table 5.2 shows the desired results for fit indices). An examination of the factor loadings and the variances of the indicators showed that INFriendInflu (‘My friends have much influence on my decision to purchase (or not purchase) energy- saving light bulbs’) had the highest factor loading (more than 1.00) and negative variance (-.046, which was inadmissible, as variance cannot be negative). Therefore, the indicator was removed from the scale in measuring the IN_Comply concept (as illustrated in Panel B, Figure 5.4). The revised model was a good fit for the data, as indicated by its fit indices (x2(1)=1.380, p=.240, bootstrapping p =1.000, CFI=.999,
TLI=.997, RMSEA=.035, PCLOSE=.405 and SRMR=.0117). Therefore, these three indicators were used for the later CFA before creating the injunctive norm’s product term.
5.3.5 Perceived behavioural control constructs
As discussed in Section 3.3.6, two constructs measured two aspects of perceived behavioural control: PBC_Perception, which had four reflective indicators, measured the respondents’ perceived ability to purchase energy-saving light bulbs; PBC_Situational, which had five reflective indicators, assessed the information regarding the respondents’ perceived situational influence. Figure 5.5 illustrates the causalities between these two constructs and their corresponding indicators.
The factor analysis performed in SPSS suggested that there was only one main factor extracted from the four items of PBC_Perception. These results indicated that these four items were unidimensional. The Cronbach’s alpha for PBC_Perception was .706, which was greater than .70, indicating internal consistency reliability. There
was also only one main factor extracted from the five items for PBC_Situational. In other words, these five items were also unidimensional. The Cronbach’s alpha for the PBC_Situational was .776, which is a desirable result. However, the evaluations of the two grouped items’ convergent validity suggested that the items could not be used individually to represent their related constructs in the CFA stage.
According to PBC_Perception’s four-item congeneric measurement model (as illustrated in Panel A, Figure 5.5), all four items’ factor loadings were significant for the PBC_Perception construct. However, the results from the PBC_Perception’s congeneric measurement model suggested that the proposed relationship did not fit the sample perfectly. The chi-square and the p-values of the congeneric models were 7.502 and .023, respectively, which indicated that the discrepancy between the proposed congeneric model and the sample could not be ignored. The bootstrapping p-value was not reviewed because PBC_Perception’s data distribution was normal at the multivariate level. Other fit indices also suggested that the discrepancy between the model and the sample could not be ignored. The SRMR of the PBC_Perception was .0357, which indicated a good fit. Although the CFI of the PBC_Perception congeneric model was .980, its TLI was .940, which suggested that the model did not fit the sample perfectly. The RMSEA of the PBC_Perception congeneric model was .094 (which indicated mediocre fit), and the PCLOSE value was .133 (which indicated that the discrepancy between the model and the sample may not be statistically significant).
While examining the modification indices suggested by the AMOS, a constraint was suggested to add to covary the indicators EasyFind (‘It is easy for me to find energy-saving light bulbs whenever I need to buy them’) and TimePressure (‘High time demands on me made it much harder to purchase energy-saving light bulbs on my last visit to the store’). Covarying these two indicators also made sense from theoretical perspective, as time pressure could also affect the impression of whether or not it was easy to find a product. Therefore, a constraint was added as suggested (as illustrated in Panel B, Figure 5.5). According to the fit indices, this revised PBC_Perception model was a good fit for the data (x2(1)=2.806, p=.094, bootstrapping p =1.000, CFI=.993,
TLI=.960, RMSEA=.076, PCLOSE=.219 and SRMR=.0200). Therefore, the PBC_Perception model illustrated in Panel B, Figure 5.5 was used in later data analyses.
For the congeneric measurement model of PBC_Situational (as illustrated in Panel A, Figure 5.5), the results indicated that the proposed relationship (i.e. the five
items and their PBC_Situational construct) did not fit the sample population. The chi- square of the PBC_Situational was 79.306 and the p-value was less than .001. The p- value of the bootstrapping was inspected because PBC_Situational had a non-normal distribution. The p-value of the PBC_Situational’s bootstrapping was less than .001. These results indicated that the discrepancy between the PBC_Situational congeneric model and the sample population was major and could not be ignored. The other fit indices also suggested the same conclusion. The PBC_Situational’s SRMR was .0819 (>.08 indicates a poor model fit). The CFI and TLI of the PBC_Situational were .838 and .676, respectively, which suggested the model was a poor fit for the sample. The RMSEA and PCLOSE were .218 and less than .001, which also led to the conclusion of poor model fit (Table 5.2 shows the desired results for fit indices).
Figure 5.5 PBC congeneric measurement models
Panel A
Panel B
The decision to remove EnvtConcern (‘Concerns for the environment’) and Affordability indicators was made. The reasons for removing EnvtConcern were based
on the suggestion from the modification indices, and also theoretically the indicator should have pre-existed, rather than being prompted in the situation. The reasons for removing Affordability were based on the factor loadings and the suggested modification indices. From theoretical perspective, dropping the Affordability indicator could better reflect reality in New Zealand, as energy-saving light bulbs are not very expensive and consumers may not even consider whether or not they are able to purchase the products, in terms of their budget. The revised PBC_Situational construct was as illustrated in Panel B, Figure 5.5; the model’s fit indices suggested that the model was a good fit for the data (x2(1)=.469, p=.493, bootstrapping p =1.000,
CFI=1.000, TLI=1.008, RMSEA<.001, PCLOSE=.637 and SRMR=.0087). Therefore, the PBC_Situational model illustrated in Panel B, Figure 5.5 was used in later data analyses.