Professional Documents
Culture Documents
Chapter 04
Chapter 04
ANALYSIS
4.1 Introduction
This chapter presents the results of the analysis performed with the use of SPSS and Partial Least
Square (PLS) 3.2.8 path modeling. Besides, initial data screening and preliminary analysis are
discussed that includes missing value analysis and common method bias/variance test.
Subsequently, the results of the current study are presented in three different forms: the
descriptive statistics, results obtained from the measurement model, and the results of structural
model which represents the hypothesized structural paths. Lastly, the result of complementary
PLS-SEM analysis, meant to examine the moderating effects in the structural model, are also
elaborated.
A total of 535 sets of questionnaires were distributed to the users of online food delivery
applications in Karachi. Table 4.1 shows the response rate of the users within the city. The
questionnaires returned were 365, representing a valid response rate of 68.22%. The response
rate was relatively high because of the use of a self-administered questionnaire (Farouk, Abu
essential (Hair, 2007), to avoid any possible violation of the key underlying assumptions of the
application of multivariate techniques. This study proceeded with the first step of preliminary
Because of the way the data were collected personally by the researcher, testing nonresponse
bias could not be carried out because all participants were given two weeks to fill in the
questionnaires after which the researcher personally collected them. In this manner, the
difference between those who responded earlier and later was a non-issue.
Table 4.2 describes the profile of the participants the majority of the participants fell within the
age 25-30 years 30.4%, about 21.4% fell within the age of 31-40 years, 24.7% were less than 25
years, and 15.1% were between 41 and 50 years old and the remaining 8.5% were Above 50
years old. Male participants dominated the response rate (52.1%) as compared to 47.9% Female
counterparts. Concerning marital status, most participants were married. Moreover, the salaried
participants are 25.2% is in the range of 25,001 – 45,000, 31% lies within 45,001 – 60,000,
25.5% and 18.4% were 60,001 – 75,000 and above 75,000 respectively. The growth of E-
commerce is on rise, so the online purchases uses more often and the respondents measuring
30.7% uses 3 times, 27.1% uses 2 times,% and 24.4% and 17.8% use above 4 times and one time
respectively.
Descriptive analysis was run to obtain the descriptive scores where the maximum and minimum
scores, standard deviation, and the mean of all variables were assessed.
As mentioned earlier in chapter three, a five-point Likert scale was used that ranged from “1 =
strongly disagree to 5= strongly agree” Table 4.4 exhibits the mean scores of the variables
ranging from 2.205 to 3.695 and the standard deviation scores are ranging from 0.90381 to
Henseler and Sarstedt (2013) postulated that the goodness of fit index is not an appropriate tool
to validate a research model. Henseler and Sarstedt (2013)) also opposed the goodness of fit
index as an appropriate tool to validate a research model. Based on the recent progress about
PLS path modeling in model validation unsuitability, a two-step process was adopted in this
study as suggested by Henseler et al. (2009), Chin (1998), and Henseler et al. (2009) to evaluate
the PLS-SEM path (see Figure 4.1). The first step was to assess the measurement model for
reliability and validity of the constructs used in this study. In the second step, the structural
model was assessed for the hypothesized structural relationships. Confirmatory factor analysis
(CFA) was performed to validate the measurement model (outer model) by examining the
relationship between items/indicators and their respective underlying constructs, using PLS-SEM
by using software known as Smart PLS 3.2.8 developed by Ringle, Wende, and Will (2005).
According to Hair et al. (2011); Hair, Ringle, and Sarstedt (2013), when determining the internal
consistency reliability, the individual item’s reliability content validity, discriminant validity, and
convergent validity must be ascertained. To ensure the validity and reliability of the model, the
study in hand following the recommendations of (Vinzi et al., 2010), who stressed that outer
loadings of the individual items and the average variance extracted value should be at least.
Hence, the items that had lower loading were deleted to improve the data quality (Hair et al.,
2011, 2013).
first, as pre-requisite for assessing the outer model. In line with the criteria suggested by the
Joseph F Hair, Rolph E Anderson, Barry J Babin, and William C Black (2010b); Hair, Sarstedt,
Hopkins, and G. Kuppelwieser (2014), convergent validity is attained by meeting the criteria that
factor loading of each item is above 0.6 and no single loading of an item from other construct is
higher than the construct being measured. Table 4.4 shows that these items had loading between
Internal consistency reliability is the extent to which all items on a particular subscale measure
the same concept (McCrae, Kurtz, Yamagata, & Terracciano, 2011). The acceptable value for
composite reliability defined in the literature (Hair et al., 2011) should not be lower than the
threshold value of 0.7, and the average variance extracted (AVE) acceptable value should be at
least 0.5. The table 4.5 shows that all the variables were highly reliable, and the AVE value of
each variable was above than the cutoff point of 0.50, which shows that the measurement model
was reliable for further analyses. The Cronbach’s alpha (α) was also calculated to validate the
internal consistency of the constructs. As per the rule of thumb given by George and Mallery
(2003), the value of greater than 0.9, 0.8 and 0.7 were classified as excellent, good and
acceptable respectively. Table 4.5 below shows the AVE, Cronbach’s alpha, and composite
Table 4.4 Construct Reliability, Cronbach’s Alpha, Composite Reliability, and AVE of Latent
Variables
Construct Items Loading Cronbach’s CR AVE
Alpha
Usefulness 0.877 0.915 0.730
USE1 0.839
USE2 0.846
USE3 0.860
USE4 0.871
Ease of Use 0.868 0.872 0.717
EOU1 0.804
EOU2 0.868
EOU3 0.857
EOU4 0.855
Money Availability 0.716 0.841 0.638
MA1 0.824
MA2 0.777
MA3 0.794
latent variable is different from other latent variables. In this study, discriminant validity was
predicted by the AVE values, the criteria proposed by (Fornell & Larcker, 1981). Discriminant
validity was obtained by evaluating the correlation between the latent variables along with the
square root of AVE (Fornell & Larcker, 1981). Fornell and Larcker (1981) recommended the use
of AVE of 0.50 or greater to assess discriminant validity. They also recommended that the
square root of AVE should be above the value of the latent variables. To examine discriminant
validity, this study examined the model’s external consistency and compared the value of AVE
of all latent variables. The table 4.5 shows that the ‘square root’ of AVE was greater than the
correlation among the latent variables, indicating adequate discriminant validity (Fornell &
Larcker, 1981). After performing CFA, none of the variables were dropped even the deletion of
some items. However, Hair et al. (2013), argued that a variable with two items should not be
subject to removal.
Additionally, the heterotrait-monotrait ratio of correlations (HTMT), which is the estimation tool
to estimate the factors correlation, was used (Henseler, Ringle, & Sarstedt, 2015; Pittino et al.,
2018). The HTMT is a newly developed method for the PLS-SEM to evaluate discriminant
validity. The HTMT criterion approach significantly outclasses the old approaches to determine
the discriminant validity for instance, the Fornell and Larcker (1981) criterion and also (partial)
cross-loadings, that are generally incompetent to notice the lack of discriminant validity. The
discrimination among the latent constructs and the HTMT values were less than one as shown in
Table 4.6. Therefore, following the HTMT criteria, all variables achieved discriminant validity.
In addition to HTMT ratios’ evaluation, the HTMT values were assessed via the PLS algorithm
procedure. The results obtained were significantly lower than 1, signifying that all the latent
constructs attained discriminant validity (Henseler et al., 2015). The results suggested that
discriminant validity was established at HTMT 0.90, which means that the values for inter-
construct ratio were below 0.90 and that the confidence intervals did not contain the value of 1.0
According to Hair et al. (2013), the paths with non-significant or signs of opposite direction to
the hypothesized relationships do not accept the hypotheses, while significant paths empirically
provide statistical evidence of causality of the relationships in the model. Before the mediating
effect was tested, bootstrapping with a resample of 500 was run to get the t-value to assess if
measurement model was established for reliability and validity. According to Hair, Black, Babin,
Anderson, and Tatham (2006), a structural model addresses the hypothesized structural
relationships in a research model. PLS - structural equation modeling estimates the inner model
for the direct hypothesized structural relationships among the constructs by exhibiting the t-
values of each structural path as coefficients. The path coefficients are same like beta values in
regression line analysis, where the value of the beta represents the coefficients of regressions and
t-values decide the significance level of the relationship path (Henseler et al., 2009). The t-value
of 1.645 for one-tailed and 1.967 or greater for 2-tailed are considered as significant (J. J. Hair et
al., 2014).
The basic purpose of the study under focus was to examine the direct relationships between the
dependent variable (Online Impulse Buying Behavior) and mediating variable (Urge to Buy
Impulsively), and secondly, to assess the hypothesized relationships among the constructs
through a structural model. A total of 07 direct relationships with Online Impulse Buying
Behavior (dependent variable) were tested in this study. Six (06) of the total hypothesis were
supported. Further, a total number of five (05) direct relationships with Urge to Buy Impulsively
(mediating variable) were tested in which all hypotheses supported (see Table 4.8 and Table 4.9).
Figure 4.3 displays the path coefficients, t-values, p-values along with the standard deviation
values. Based on these results, decision was made to support or reject a hypothesis. The t-values
were obtained from bootstrapped procedure (with 500 sampling iterations for 365 cases
observations). Hair et al. (2013) argued that bootstrapping serves as a proxy of parameters for
standard error. As Hair et al. (2013) explained, the paths that are non-significant or showing
signs the opposite direction to the hypothesized do not support prior hypotheses while significant
paths empirically support the proposed causal relationship. Before the mediating effect was
tested, bootstrapping with a resample of 500 was run to get the t-value to assess if the direct
The result from the output of the PLS algorithm and bootstrapping showed a positive significant
association between Cash on Delivery and OIBB (β = 0.178, t = 4.151). Therefore, Hypothesis 1
supported.
The result from the output of the PLS algorithm and bootstrapping showed a positive significant
supported.
The result from the output of the PLS algorithm and bootstrapping showed a positive significant
association between Ease of Use and OIBB (β = 0.131, t = 2.542). Therefore, Hypothesis 3
supported.
The result from the output of the PLS algorithm and bootstrapping showed a positive significant
association between Money Availability and OIBB (β = 0.017, t = 0.440). Therefore, Hypothesis
The result from the output of the PLS algorithm and bootstrapping showed a positive significant
association between Use of Credit Card and OIBB (β = 0.181, t = 2.973). Therefore, Hypothesis
5 supported.
The result from the output of the PLS algorithm and bootstrapping showed a positive significant
supported.
The result from the output of the PLS algorithm and bootstrapping showed a positive significant
supported.
The result from the output of the PLS algorithm and bootstrapping showed a positive significant
association between Cash on Delivery and UTBI (β = 0.228, t = 3.800). Therefore, Hypothesis 8
supported.
The result from the output of the PLS algorithm and bootstrapping showed a positive significant
association between Cash on Delivery and UTBI (β = 0.262, t = 4.641). Therefore, Hypothesis 9
supported.
The result from the output of the PLS algorithm and bootstrapping showed a positive significant
association between Ease of Use and UTBI (β = 0.177, t = 3.378). Therefore, Hypothesis 10
supported.
The result from the output of the PLS algorithm and bootstrapping showed a positive significant
association between Ease of Use and UTBI (β = 0.138, t = 3.378). Therefore, Hypothesis 11
supported.
Hypothesis 12: Usefulness is positively related to UTBI.
The result from the output of the PLS algorithm and bootstrapping showed a positive significant
association between Ease of Use and UTBI (β = 0.218, t = 4.880). Therefore, Hypothesis 12
supported.
As Hair et al. (2014) stated that a mediation test is done to know whatever a mediating variable
enhance the impact of an independent variable on a dependent variable. Several techniques can
be used for mediation test such as Baron and Kenny (Baron & Kenny, 1986), Sobel test (Sobel,
1982), and bootstrapping (Preacher & Hayes, 2004, 2008). The re-sampling mediation technique
(bootstrapping) was used in this study to test the indirect effect of each potential variable because
this is one of the most rigorous and powerful procedures for testing the mediation effect (Hayes,
2009; Zhao et al., 2010). According to Hair et al. (2014), bootstrapping for mediation analysis is
best suited for PLS-SEM because it can be applied to small sample size.
To perform bootstrapping, the path coefficients were determined first followed by the t-values to
assess the direct relationships between the independent variables and the dependent variable
Once the direct relationships had been tested, the test of the mediation effect was performed.
According to (Hayes, 2009), there are several steps in assessing the relationship. First, a
researcher needs to fit a model through SEM to estimate the relationship between the predictor
and the mediator variables where path “a” is the relationship between the mediator and the
criterion variable and path “b” is the relationship of the three variables. Then, the t-values were
identified. Third, the standard errors (SE) of all indirect effects were determined. From the
structural model assessment of this study, it was found that three variables were significantly
related to Urge to Buy Impulsively and Online Impulse Buying Behavior (see Table 4.10). The
Table 4.10 Mediation Effect Results (Confidence Interval, t-Value, p-Value, and Standard
Deviation)
Hypothesis Relationship Beta SE t value P Values Decision
Hypothesis 13 COD -> UTBI -> OIBB_ 0.057 0.018 3.206 0.001 Supported
Hypothesis 14 ENT -> UTBI -> OIBB_ 0.065 0.021 3.076 0.002 Supported
Hypothesis 15 EOU -> UTBI -> OIBB_ 0.044 0.016 2.694 0.007 Supported
Hypothesis 16 MA -> UTBI -> OIBB_ 0.034 0.014 2.359 0.019 Supported
Hypothesis 17 USE -> UTBI -> OIBB_ 0.054 0.017 3.092 0.002 Supported
4.7.1.3 Structural Model with Moderator
As shown in Table 4.11, Use of Credit Card (moderating variable) showed a t-value of 4.75,
which was more than the cutoff value of 1.967, indicating that the result was statistically
significant. Thus, the result showed credible evidence of the moderating effect of customer
orientation on the relationship between UTBI and OIBB (β = 0.130, t = 4.751, p = 0.000). Hence,
Once the goodness of the path model had been established, the next step was to examine the
hypotheses. By running PLS-SEM (PLS algorithm and bootstrapping), structural model was
assessed (Chin, 2010). First, the predictive power of the structural model was evaluated by the
coefficient of determination (R2 values) of the endogenous construct (Chin, 2010; Henseler et al.,
2009) and the significance level of the path coefficients was determined (Henseler et al., 2014).
The R2 value is the representation of the proportionate variation that can be explained by one or
more predictor variables (Elliott & Woodward, 2007; Hair et al., 2010b). Falk and Miller (1992),
and Hair et al. (2010b) recommended a minimum acceptable threshold of an R2 value of 0.10.
Chin (1998) suggested the R2 values assessment criteria 0.19 as weak 0.33 as moderate and 0.67
as substantial respectively. Table 4.7 illustrates the R 2 of each endogenous latent variable where
To evaluate the criterion of predictive accuracy, the Stone-Geisser’s Q 2 value was utilized
(Geisser, 1974) cited in (Henseler et al., 2014). The blindfolding procedure was performed to
obtain the value of Q2. Generally, there are two different approaches to calculate Q2. They are the
approach develops the path model estimate of both the structural model (scores of the predictor
constructs) and the measurement model (the endogenous construct) of data prediction.
Alternatively, the cross-validated communality approach presents only the construct scores
estimated for the target endogenous construct (excluding the information about the structural
Table 4.13 Summary of the Predictive Relevance of the Endogenous Latent Constructs (Q2)
Constructs Q² Predictive Relevance
Online Impulse Buying Behavior 0.518 Yes
Urge to Buy Impulsively 0.391 Yes
4.10 Summary
This chapter reported the findings of the study. SPSS was used to describe the participants’
profile. However, PLS-SEM analysis was employed to test the reliability and validity of the
measures. The technique of bootstrapping in PLS-SEM analysis was used to test the research
hypotheses. In general, the result indicated that the measurement model was deemed acceptable
based on sufficient evidence of reliability, convergent validity, and discriminant validity. After
the measurement model was assessed, and the structural model was tested. The following chapter
discusses the findings, accompanied by implications to practice, future research, and limitations
of the study.