chapter 5 data analysis -...

142

CHAPTER 5

DATA ANALYSIS

The first and foremost procedure in the data analysis stage was to

verify the quality of collected data for finalizing the tools required for further

analysis.

5.1 ANALYZING THE QUALITY OF DATA

This section involves procedures adopted in verifying and cleaning

of data for further analysis. This included steps such as

Verification of missing values

Identification of Outliers

Analysis of Normality

Analysis of validity and Reliability

5.1.1 Verification of Missing Values

The responses collected from 500 respondents using structured

questionnaire was entered in SPSS 17 under different variable names. To

identify missing variables a frequency test was done. Missing responses were

noticed in 105 cases where respondents fail to mark their responses related to

certain questions which were critical in analysis point of view and hence these

cases were deleted. After deletion of missing responses, 395 usable responses

were obtained.

143

5.1.2 Identification of Outliers

Outliers are created due to various reasons such as data entry

errors, sampling errors as well as biased responses from the respondents.

Some cases of outliers are noticed and were treated in the following manner

in this research. Statisticians have devised several ways to detect univariate

outliers. Grubbs' test is particularly easy to follow. This method is based on Z,

which is calculated as the difference between the outlier and the mean divided

by the SD. If Z is large, the value is far from the others.

Hair et al (1998) suggest that as common rule of thumb, z scores

can range from ± 3 to ± 4 for samples of more than 80.In this research to

determine outlier, an outlier calculator which performs Grubbs' test available

at www.graphpad.com was used. Three outliers located by this procedure

were eliminated.

The multivariate assessment of outliers was conducted using the

DfBeta Influence Statistics method using SPSS17. To estimate effect of

outliers in the study, the following rules were used:

where n = Sample Size i.e. 384(minimum required), and hence all cases

where DfBeta > 0.0914 shall be considered as outlier. The above procedures

detected 7 cases as outliers, which were eliminated.

Outliers represent cases whose scores are substantially different

from all others in a particular set of data. A univariate outlier has an extreme

144

score on a single variable, whereas a multivariate outlier has extreme scores

on two or more variables (Kline, 2005). A common approach to the detection

of multivariate outliers is the computation of the squared Mahalanobis

distance (D2) for each case. This statistic measures the distance in standard

deviation units between a set of scores for one case and the sample means for

all variables (centroids). Typically, an outlying case will have a D2 value that

stands distinctively apart from all the other D2 values. Therefore to

countercheck for multivariate outliers, squared Mahalanobis distance (D2)

was verified from the Amos output. A review of these values showed minimal

evidence of serious multivariate outliers.

5.1.3 Analysis of Normality

Many of the statistical methods require the assumption that the

variables observed are normally distributed. With multivariate statistics, the

assumption is that the combination of variables follows a multivariate normal

distribution. Since there is no direct test for multivariate normality, we

generally test each variable individually and assume that they are multivariate

normal if they are individually normal, though this may not necessarily the

case. In SEM model, estimation and testing are usually based on the validity

of multivariate normality assumption, and lack of normality will adversely

affect goodness-of-fit indices and standard errors (Baumgartner and Homburg

1996; Hulland et al 1996; Kassim 2001).

To assess normality, skewness and kurtosis are commonly used by

the statisticians. Skewness refers to the symmetry of a distribution whereas

kurtosis relates to the peakedness of a distribution. A distribution is said to be

normal when the values of skewness and kurtosis are equal to zero

(Tabachnick and Fidell; 2001). However, there are few clear guidelines about

how much non-normality is problematic.It is suggested that absolute values of

145

univariate skewness indices greater than 3.0 seem to describe extremely

skewed data sets (Chou and Bentler 1995). Regarding kurtosis, there appears

that kurtosis index greater than 10.0 may suggest a problem.

Analysis for univariate normality done using Kolomogorov-

Smirnov test with Lillefors significance correction revealed that none of the

variables are normally distributed. However skewness was between -0.9 to

0.1 range showing most of the data negatively skewed. Non-normality of the

data was anticipated as most of the respondents preferred to agree or strongly

agree to the survey dimensions indicating bulk of the values (including the

median) lie to the right of the mean. In this study, all the variables fall under

the kurtosis value of 3, inferring kurtosis was not problematic in this research.

Amos16.0 provides normality checks for data including skewness ,

kurtosis indexes and Mardia’s coefficient which is a test of multivariate

normality. Critical ratios provided by Amos output as attached to kurtosis

represents Mardia’s normalized estimate of multivariate kurtosis. Bentler

(2005), has suggested that, in practice, values > 5.00 are indicative of data

that are non-normally distributed. To correct for non-normality in the

underlying database, use of Bollen-Stine bootstrap and associated p-value was

considered in this study. For all constructs to moderate the effect of

multivariate non-normality, the maximum likelihood (ML) estimation, which

is relatively robust against departures from multivariate normality even in a

small manner (Anderson and Gerbing 1988; Sweeney 2000; Tabachnick and

Fidell 2001), was applied with Bollen-Stine bootstrap procedure. The boot

strap sample of 1000 was adopted in this study.

5.1.4 Analysis of Validity and Reliability

In undertaking a statistical analysis, unidimensionality should be

always assessed first, prior to examining reliability and validity (Hair et al.

146

1995). This step reduces the possibility of misspecifications (Gerbing and

Anderson 1988), because the analysis of reliability and validity is based on

the assumption of unidimensionality (Nunnally and Bernstein 1994). Validity

determines whether the scale truly measures what it was intended to measure.

Testing the reliability of survey data is the pre-requisite for data

analysis and inference. Reliability analysis tests whether a scale consistently

reflects the subset it measures (Churchill 1979; Nunnally and Bernstein

1994). By consistency it is firstly meant that a respondent should score

questionnaire the same way at different times. Secondly, two respondents

with the same attitude towards service quality should identically score the

survey. According to Field, (2005), values between 0.7 and 0.8 of Cronbach’s

items make a difference and in extreme cases they can lead to a negative

Cronbach's alpha (Field 2005). In this study reverse scored items were not

included as it may lead to problems in reliability of the data if the respondents

answer without proper understanding of the question. In this study both

reflective and Formative measures were used. The approaches to test

reliability of these constructs are different. The reliability of reflective

constructs was ascertained using the above criterion.

As formative constructs are composed of different aspects of a

construct, their indicators are not necessary to correlate with each other.

Diamantopoulos and Winklhofer (2001), stated that “it is not clear that

reliability is a concept that applies well to formative constructs”. This

statement was also supported by Diamantopoulos and Siguaw (2006) and

Rossiter (2002) and hence concluded that no reliability test are mandatory for

formative indicators. Reliability evaluation for formative constructs is in

ascertaining the absence of multicollinearity (Diamantopoulos and

147

Siguaw 2006). Muticollinearity can be tested using Variance Inflation Factor

(VIF). The guidelines applied in this regard were as follows:

VIF should be less than 3.3 (Diamantopoulos and Siguaw

2006).

If VIF is less than 10 explains the absence of Collinearity

(Hair et al 1998).

Various validity and reliability criteria adopted in this study were

explained in Table 4.5 above.

5.2 ANALYSIS OF SERVICE QUALITY DIMENSIONS

The next step in the analysis procedure was to explore the service

quality construct and confirm the existence of various dimensions by which it

was assumed to be formed. This was done in two stages

Exploratory factor Analysis using Factor 7.0 developed at the

Rovira I Virgili University, Spain

Confirmatory factor Analysis using Amos 16.0

5.2.1 Exploratory Factor Analysis

The indicator variables related to Service quality construct were

subjected to an exploratory factor analysis to identify the underlying factors

and to test whether the factors extracted are similar to the dimensions

proposed in the study. The analysis was conducted by Factor 7.0,which is a

freeware program developed at the Rovira i Virgili University, Spain by

Urbano Lorenzo-Seva and Pere J. Ferrando( 2005). An important feature of

this program was that it generates goodness of fit of the data simultaneously.

28 scale items were used to measure service quality in the Banking context as

148

explained in the previous chapters. In order to identify the naturally occurring

dimensions of service quality all 28 items were subjected to a factor analysis.

This approach was recommended in the literature as a means of identifying

actual, rather than perceived, factor groupings (Rosen and Surprenant, 1998).

The role of factor analysis is to identify the components or factors that derive

from a set of variables, i.e. to identify the subset of correlated variables that

form a subset which is reasonably uncorrelated with other subsets (Hair et al

1998; Tabachnick and Fidell; 2001).

An Exploratory Maximum Likelihood factor analysis with varimax

rotation was performed as it incorporates common, specific and error variance

and was appropriate when the objective was to identify the minimum number

of factors associated with the maximum explanation of variance (Hair et al

1998). The items that load higher than 0.5 are retained while low loading

items are dropped. In general, higher factor loading is considered better, and

typically loadings below 0.30 are not interpreted. As general rule of thumb,

loadings above 0.71 are excellent, 0.63 very good, 0.55 good, 0.45 fair, and

0.32 poor (Tabachnick and Fidell 2007).

The Exploratory Maximum Likelihood factor analysis identified

five components with an Eigen value greater than 1, which together explained

over 66.36 percent of the variance indicated a good fit and hence it was

assumed that model represents the data. The Kaiser-Meyer-Olkin Measure of

Sampling Adequacy was 0.926 and the Bartlett Test of Sphericity was

significant (p<0.001) with a Chi Square value of 7203.0 with 378 degrees of

freedom which was considered to be very good for further analysis and

provided support for the factorization (Table 5.1). The Goodness of fit

statistics are shown in Table 5.2.

149

Table 5.1 Adequacy of the correlation Matrix

Table 5.2 Goodness of fit statistics after EFA of service quality construct

150

Table 5.3 below provides the details of each factor along with items

contributing it with component loadings for each item. The total number of

items for service quality construct got reduced to 27 as one item could not

load more than 0.45 in factor extraction. Therefore the item “technically

skilled staff” was deleted.

Table 5.3 Factor loadings of service quality construct

151

Explained Variance and reliability of rotated factors as obtained

from the output of Factor 7.0 analysis (Table 5.4) shows adequate reliability

for extracted factors.

Table 5.4 Explained Variance and Reliability of Rotated Factors

The following conclusions were drawn from the exploratory factor

analysis conducted.

There existed five underlying factors which represent the service

quality construct in the banking context in Kerala.

Each item was mainly related to only one factor except for cross

loading shown by certain indicators which can be theoretically

justified as correlations among reflective measures are expected

and possibility of respondents conceive a different factor

perception for certain indicators cannot be ruled out.

Some of the “Image” related indicators showed considerable

cross loadings with “Human” factor. Indicator variables such as

“Helpful to customers” and “punctual in service delivery”

showed higher loading to “Human” dimension which can be

justified on content grounds as these indicators represent

contribution from the employees also. Hence it was decided to

include these indicators along with “Human” dimension for

further analysis.

152

One indicator variable attached to “products and services”

dimension namely “less documentation for products and

services” was showing higher loading to “Convenience” factor

and hence decided to include this indicator along with

“Convenience” factor for further analysis.

Two indictors “modern systems for service” and “wide network

of ATM for easy service” were showing higher loadings to

“products and services” dimension rather than “System” and

hence decided to include them with “products and services” for

further analysis.

On indicator variable “technically skilled staff” showed loading

less than 0.45 and hence excluded from further analysis.

The next step was to conduct a confirmatory factor analysis for the

service quality dimensions identified.

5.2.2 Confirmatory Factor Analysis-Service Quality Dimensions

The primary objective of conducting CFA was to determine the

ability of a predefined factor model to fit an observed set of data. It provides

estimates for each parameter of the measurement model. The various

parameters used for evaluation of the model are shown in Table 5.5.

Table 5.5 Various Parameters to be considered for model evaluation

Sl.NO Parameter 1 Factor loadings,

2 Factor Variances 3 Covariance 4 Indicator Error Variances 5 Error Covariances

CFA is useful in

153

Testing the significance of a specific factor loading.

Testing the relationship between two or more factor loadings.

Testing whether a set of factors are correlated or uncorrelated.

Assessing the convergent and discriminant validity of a set of

measures.

CFA has strong links to structural equation modeling and hence the

procedures involved are as explained under heading 4.4.Prior to validating the

full structural model with all latent variables, it was required to validate each

of the measurement models as a preliminary step. The measurement model is

the part of an SEM model that deals with the latent variables and their

indicators. The measurement model was evaluated for validity like any other

SEM model, using goodness of fit measures. The major data considerations to

be addressed before conducting CFA are

Table 5.6 Various Data Considerations

Sl.No. Data Considerations

1 Absence of missing data

2 Absence of outliers

3 Adequacy of sample size

4 Existence univariate and multivariate normality

The data were found free from missing values and outliers as

explained in headings 5.1. Unfortunately, there is no easy way to determine

the sample size needed for CFA. There are some very rough guidelines for

sample sizes: less than 100 is considered “small” and may only be appropriate

for very simple models; 100 to 200 is “medium” and may be an acceptable

minimum sample size if the model is not too complex; and greater than 200 is

154

“large”, which is probably acceptable for most models (Kline, 2005).

Analysis of normality was done in the univariate level and multivariate level

as explained in the heading 5.1.3.

Maximum likelihood (ML) estimation method was used in all

analysis using Amos.16. Maximum likelihood “aims to find the parameter

values that make the observed data most likely (or conversely maximize the

likelihood of the parameters given the data)” (Brown, 2006). It has several

desirable statistical properties:

it provides standard errors (SEs) for each parameter estimate,

which are used to calculate p -values (levels of significance)

and

it provides confidence intervals, and its fitting function is used

to calculate many goodness-of-fit indices

5.2.2.1 Measurement Model for “Image” Dimension

The seven indicator variable model of “Image” dimension was

suggesting poor fitting model in the first estimate. The normed alpha,

RMSEA and CFI were above the permissible level. On verification of

modification indices two indicator variables “ img2” and “img8” were

showing cross loadings to many other variables and was found to be a major

cause for poor fit and hence were removed. The resulting model was found to

be good fitting model with recommended indices as illustrated in Figure 5.1.

All the paths shown in the model are significant as critical ratios were

above 1.96.

155

img

.39

img1 i1.40

img3 i3.76

img5 i4.55

img6 i5

.74

.39

img9 i7

.62

.87.63.63

CMIN/df-2.96,CFI-0.98,SRMR-0.027,RMSEA-0.07,PClose-0.166HoelterNat0.05-287, Composite reliability-0.706,AVE-0.698

Figure 5.1 Measurement Model for "Image" dimension

5.2.2.2 Measurement Model for “Human” Dimension

The seven indicator variable model of “Human” dimension was

suggesting poor fitting model in the first estimate. The normed alpha,

RMSEA, and NFI were above the permissible level. On verification of

modification indicators indicator variables “img7” was showing cross

loadings to many other variables and was found to be a major cause for poor

fit and hence were removed. The resulting model was found to be much better

but still needed modification except for low squared multiple correlation of

indicator variable “Human5” which is considered for removal in the second

stage. The resulting model was found to be good fitting model with

recommended indices as illustrated in Figure 5.2. All the paths shown in the

model are significant as critical ratio were above 1.96.

156

Human

.44

img4 h1.65

human1 h3.84

human2 h4.92

human3 h5.64

human4 h6

.80

.81

.66

.96.91

CMIN/df-2.74,CFI-0.99,SRMR-0.019,RMSEA-0.07,PClose-0.12HoelterNat0.05-234, Composite reliability-0.83,AVE-0..829

Figure 5.2 Measurement Model for "Human" dimension

5.2.2.3 Measurement Model for “Convenience” Dimension

The initial five indicator variable model reported a poor level of fit

as the RMSEA (0.134) was outside the recommended tolerances. To modify

the model, the indicator variable “convei1” was removed due to poor squared

multiple correlation. The resulting model was found to be good fitting model

with recommended indices as illustrated in Figure 5.3. All the paths shown in

the model are significant as critical ratio were above 1.96.

Convenience

.53 convei2 h3

.44 convei3 h4

.65 convei4 h5

.51 prd1 h6

.71

.73

.81.66

CMIN/df-3.34,CFI-0.995,SRMR-0.015,RMSEA-0.079,PClose-0.2

HoelterNat0.05-434, Composite reliability-0.835,AVE-0.727

Figure 5.3 Measurement Model for "Convenience" dimension

157

5.2.2.4 Measurement Model for “Products and Services” Dimension

The four indicator variable model of “products and services”

dimension was suggesting poor fitting model in the first estimate. The normed

alpha and RMSEA were above the permissible level. An indicator variable

“prd2” was removed from further analysis due to poor loading to get a well fit

model with all indices considered above the desired level and with significant

paths as illustrated in Figure 5.4.

prd&services

.47prd3 p2

.66sys1 p3

.66sys2 p4

CMIN/df-1.13,CFI-0.99,SRMR-0.013,RMSEA-0.017,PClose-0.491 HoelterNat0.05-234, Composite reliability-0.705,AVE-0.765

.69

.81.81

Figure 5.4 Measurement Model for "products & services" dimension

5.2.2.5 Measurement Model for “System” Dimension

The five indicator variable model related to “system” dimension

was suggesting poor fitting model in the first estimate. The normed alpha,

RMSEA and CFI were above the permissible level. As per modification

indices, an error correlation was added between indicator variables “sys6”

and “sys7” considering the theoretical grounds, as to correlate error terms

there needs to be a strong theoretical justification behind such a move

(Joreskog and Long 1993) to develop a well-fit and significant model as

illustrated in Figure 5.5. These variables represent responses related to user-

friendly website and up to date web site and hence theoretically there is a

158

chance for their error variables to have correlation. All the paths shown in the

model are significant as critical ratio were above 1.96.

system

.73sys3 s1

.84sys4 s2

.82sys5 s3

.66sys6 s4

.81

.85

.90 .92

CMIN/df-3.74,CFI-0.99,SRMR-0.011,RMSEA-0.07,PClose-0.08HoelterNat0.05-235, Composite reliability-0.839,AVE-0.858

.64sys7 s5

.80

.36

Figure 5.5 Measurement Model for "system" dimension

5.2.2.6 Structural Model for Service Quality Construct

Structural equation models with latent variables (SEM) are more

often used to analyse relationships among variables. The relationships among

latent variables were tested only after obtaining a statistically significant well-

fitting model which represents the data.

The statistical significance of relationships among Service quality

and its extracted dimensions such as Image, Human, Convenience, products

and Services and system were of interest to this study. The well-fit

measurement models of service quality dimensions such as Image, Human,

Convenience, products and Services and system are taken together to arrive at

a fitting structural model for service quality. The model developed is

illustrated in Figure 5.6.Two important considerations are used to test the

statistical significance using Amos output.

159

The critical ratio (C.R.), which represents the parameter estimate

divided by its standard error; as such, it operates as a z-statistic in testing that

the estimate is statistically different from zero. Based on a probability level

of .05, the critical ratios are to be > ±1.96 for statistical significance.

Non-significant parameters, with the exception of error variances, can be

considered unimportant to the model; in the interest of scientific parsimony

they should be deleted from the model (Barbara.M.Byrne 2010).The standard

residual co-variance should be less than 2.58 to conclude statistically

significant co-variance between two variables (Barbara.M.Byrne 2010).

Hence such observations can also be considered for exclusion in further

analysis.

The first model developed needed re-specification as the standard

residual co-variance between some of the variables was above 2.58. The

model re-specification on the basis of modification indices was adopted to

finalize a good-fitting model explaining the service quality construct.

However a scientific theory based reasoning is essential in adopting

suggestion offered by modification indices in an urge to find better fit for the

structural model. The indicator variable “image 4” attached to “Human”

dimension was selected for removal at re-specification stage due to two

reasons

The squared multiple correlation<0.5

This item was showing cross loadings with many other

variables and hence problematic.

160

The re-specified model is illustrated in Figure 5.7.

Img

.43img1 x1.40img3 x2.72img5 x3.55img6 x4.39img9 x5

.74

.63

.63.65.85

Human

.66human1 x7.83human2 x8.90human3 x9.66human4 x10

.95.81

.81.91

Convenience

.52convei2 x11.42convei3 x12.62convei4 x13.56

prd1 x14

.75.65.72

.79

Prd&services

.43prd3 x15.72sys1 x16.62sys2 x17

.85.65.79

System

.73sys3 x18.85sys4 x19.81sys5 x20.67sys6 x21.64sys7 x22

.82

.80

.92.85.90

.33.49

.60

.78

.59

.57

.47

.51

.45

.34

.47img4.68 x6

.35

Fig 5-6 Confirmatory model for Service quality construct-1CMIN/df-2.11,CFI-0.96,SRMR-0.043,RMSEA-0.054,PClose-0.186

HoelterNat0.05-213,Re-specification needed as some of std.redidual covariences >2.58

Figure 5.6 Confirmatory model for Service quality construct-1

161

img


.75.63

.63

.65

.85

Human


.96

.81

.81.91

Convenience


prd1 x14

.75

.65.72

.79

Prd&services

.43prd3 x15.72sys1 x16.62sys2 x17

.85.65

.79

System


.82

.80

.92.85

.90

.33 .48

.60

.77

.59

.57

.47

.51

.45

.34

.35

Fig 5-7 Confirmatory model for Service quality construct-CMIN/df-1.7,CFI-0.976,SRMR-0.037,RMSEA-0.0434,PClose-0.929

HoelterNat0.05-267,

Figure 5.7 Confirmatory model for Service quality construct

The stages in development of a confirmed model are summarized

as follows.

Stage-1: from the initial list of 28 indicator variables, 1 variable

related to “technically skilled staff” was removed for

poor loading at exploratory factor analysis.

162

Stage-2: Five variables were removed at the stage of development

of measurement models for dimensions identified after

exploratory factor analysis.

Stage-3: One variable was removed in the re-specification stage

of evaluating the confirmatory model for service quality.

The details of dimensions with their indicators in perceived service

quality scale developed for banking context are given in Table 5.7. The

overall reliability of the scale was 0.922.The model fit summary and

estimates are provided in Appendix 2.

Table 5.7 Variables after Confirmatory Factor Analysis

163

5.2.2.7 Validation of the Perceived Service Quality Scale

To demonstrate the soundness of measurement scale developed,

first of all, it was necessary to address the issue of Common methods variance

(CMV). Common methods variance can be a major source of measurement

error in data collection when variables are latent and measured using the same

survey at one point of time. CMV may inflate the true correlations among

latent constructs and threaten the validity of conclusions. Harman's single-

factor test is most widely known approach for assessing CMV in a single-

method research design (Podsakoff and Organ 1986). In single-factor test, all

of the items in the study are subjected to exploratory factor analysis (EFA).

CMV is assumed to exist if

a single factor emerges from unrotated factor solutions, or

a first factor explains more than 50% the variance in the

variables (Podsakoff and Organ 1986)

The EFA conducted with all variables in the study yielded five

distinct factors with an eigenvalue above 1. The first factor accounts for

22.5% of the variance at unrotated stage and all factors together account for

66.5% of the total variance. When the initial solution was rotated using a

varimax rotation in principal component analysis the same factor accounts for

less than 12% of the total variance and hence confirmed that CMV was not a

major concern in this study.

Convergent validity was established when the relationship between

measurement items and the factor were significantly different from zero.

Based on this criterion, critical ratios were used to evaluate the statistical

significance. Parameters which have a critical ratio greater than 1.96 were

considered significant based on the level of p=0.05 (Anderson and Gerbing

1988). In this study, all of the measurement items represented their factors

164

significantly, as the critical ratio of every item exceeded the 1.96 value;

hence, all of the measurement items satisfied the convergent validity test

(Table 5.8). Also, the standardized regression weights should be significantly

linked to the latent construct and have at least loading estimate of 0.5 and

ideally exceed 0.7 (Hair et al 2006).In this study the factor loading ranged

from 0.629 to 0.920 and no loading was less than recommended 0.5.

Table 5.8 Estimates and squared multiple correlation of all indicators

The convergent validity assessment also included the measure of

construct reliability and average variance extracted. According to Fornell and

Larcker (1981), variance extracted refers “the amount of variance that is

captured by the construct in relation to the amount of variance due to

165

measurement error”. Further, Fornell and Larcker (1981), suggested that

variance extracted to be a more conservative measure than construct

reliability. The other criteria used to assess convergent validity are:

As a rule of thumb good reliability is suggested if Cronbach’s

alpha estimate is t higher than 0.7.

Variance extracted (VE) for a construct should be larger than

0.5 indicate reliable factors (Hair et al 1995,Holmes-Smith

2001)

As a rule of thumb composite reliability is considered high if

squared multiple correlation R2 (“smc”) greater than 0.5,

moderate if between 0.3 and 0.5 and poor if less than 0.3

(Holmes-Smith 2001),suggesting construct reliability

Online CONSTRUCT VALIDITY Calculator version 2.0 available

at http://www.hishammb.net/cvc2 is used for calculating construct reliability

and variance extracted by each dimensions used for service quality

(Table 5.9).

Table 5.9 Composite Reliability and Variance Extracted by each constructs

Discriminant validity was confirmed by examining correlations

among the constructs. As a rule of thumb, a 0.85 correlation or higher

indicates poor discriminant validity in structural equation modeling

166

(David 1998). None of the correlations among variables were above 0.85

(Table 5.10). The results suggested adequate discriminant validity of the

measurement

Table 5.10 Correlations among constructs

Further, to confirm discriminant validity the squared inter construct

correlation (SIC) were calculated and compared with average variance

extracted. All variance extracted (AVE) estimates in the Table 5.9 were larger

than the squared inter construct correlation estimates (SIC) provided in Table

5.10. Therefore it was confirmed that the indicators have more in common

with the construct they were associated with than they do with other

constructs.

Nomological validity was tested by examining whether the

covariances between the constructs in the measurement model make sense.

The construct covariances are used to assess this. All the covariances were

positive and significant as seen in Table 5.11 confirming nomological

validity.

From the above observations, it was confirmed that the scale

developed was having adequate psychometric soundness for measuring

perceived service quality of banking services in Kerala, India

167

Table 5.11Covariances among constructs

Multi-group Comparison. A multi-group CFA was conducted to identify potential developmental differences in factor structure. In a multi-group CFA, measurement scale was tested to check whether the items comprising a particular measuring instrument operate equivalently across different populations (e.g., gender, age, bank type etc). There are three primary steps in a multi-group CFA:

Determining the factor structure of the measure across each group freely estimating the factor loadings (unconstrained model);

Determining the factor structure of the measure across each group constraining the factor loadings to be equal (constrained model);

Comparing the goodness-of-fit indices between the constrained and unconstrained models.

Differences between groups were assessed by comparing the goodness- of-fit indices of the model with factor loadings constrained to be equal to the unconstrained base model. If significant differences are observed between the constrained and unconstrained model goodness-of-fit indices, this indicates that factor structure is not same. If no significant differences are

168

observed between the constrained and unconstrained model goodness-of-fit indices, this indicates factor structure is considered to be the same.

A multi-group comparison was conducted to determine whether the scale has the same theoretical structure for each bank type. No significant differences between the constrained and unconstrained models were identified. The comparative indices are listed in Table 5.12 below.

Table 5.12 Comparative goodness of fit for nested models

All indices are showing similar values and hence it can be assumed that model fits to all type of population. Thus the construct validity of the measurement model for service quality is fully established. The objective of the study to understand the various dimensions and indicators that can form a valid scale to measure perceived service quality in the kerala context was thus achieved.

5.3 STRUCTURE OF PERCEIVED SERVICE QUALITY CONSTRUCT

An important issue to be addressed in this study was whether Perceived service quality needs to be defined as a formative or a reflective construct. A reflective construct implies that the separate dimensions of PSQ ,such as image, human, convenience, products and services and system are actually different manifestation of the PSQ construct and as such ‘‘reflect’’

169

the content of PSQ, whereas a formative construct suggests that PSQ is defined as the outcome formed of its dimensions. For example, increases in any one of the dimensions say “Human”, if results in an increase in all the other dimensions of PSQ, then PSQ should be conceptualized as reflective. On the other hand, when an increase in any one of the dimensions increases the overall magnitude of PSQ, without necessarily affecting the rest of the dimensions, PSQ should be defined as formative. The researcher conceptualized PSQ as first-order reflective and second-order formative construct on theoretical grounds and content validity need to be established for this assumption.

For statistical validity of both reflective and formative models of

perceived service quality and to understand which model represents the data

in a better manner two models as shown in Figures 5.8 and 5.9 are developed

and tested for goodness of fit analysis. Identification of formative indicator

constructs in Amos 16 required following procedures (Jarvis et al 2003;

MacCallum and Browne 1993).

(1) The scale of measurement for the latent construct was

established by constraining a path from one of the construct’s

indicators to be equal to 1 or by constraining the residual

error variance for the construct to be equal to 1 and

(2) To resolve the indeterminacy associated with the construct

level error term, a formative Construct should emits paths to

at least two unrelated latent constructs with reflective

indicators

at least two theoretically appropriate reflective

indicators, or

one reflective indicator and one latent construct with

reflective indicators

170

Hence it was assumed that formative construct of PSQ emit paths

to five indicators used to measure satisfaction development. This move can be

theoretically justified as PSQ was assumed to cause satisfaction development

among customers. The same indicators were used in reflective PSQ model

also.

Img


.75.63

.63.65.85

Human


.96

.81

.81.91

Convenience


prd1 x14

.75.65.72

.79

Prd&services

.43prd3 x15.72sys1 x16.62sys2 x17

.85.65.79

System


.82

.80

.92.85.90

.33 .48

.60

.77

.59

.57

.47

.51

.45

.34

.34

Fig 5-8 Confirmatory model for Formative Service quality constructCMIN/df-1.79,CFI-0.964,SRMR-0.043,RMSEA-0.045,PClose-0.881

HoelterNat0.05-245,

.33

PSQ

.06

.32

.07

-.06

.32

x.69

satdevelop1e2 .83.70satdevelop2e3 .84.49satdevelop3e4

.70.49

satdevelop4e5.70

.19satdevelop5e6

.44

Figure 5.8 Confirmatory model for Formative Service quality construct

171

.59

Img


.74

.63

.63.65.85

.66

Human


.96.80

.81.91

.51

Convenience


prd1 x14

.74.64.73

.79

.41

Prd&services

.43prd3 x15.72sys1 x16.62sys2 x17

.85.66.78

.30

System


.82.80

.92.85.90

.35

Fig 5-9 Confirmatory model for Reflective Service quality constructCMIN/df-3.95,CFI-0.858,SRMR-0.089,RMSEA-0.088,PClose-0.000

HoelterNat0.05-111,

.00

PSQx

.77

.81

.72

.64

.55

e1

e3

e4

e5

e2

.21satdevelop1e6

.46.22satdevelop2e7 .47.17satdevelop3e8

.41.19

satdevelop4e9.44

.09satdevelop5e10

.29

Figure 5.9 Confirmatory model for Reflective Service quality construct

The model no 1 demonstrated in Figure 5.8 emerged superior to other model with regard to goodness of fit. The model 1 explained the relation among first order factors to second order PSQ construct in the formative manner. This finding was in tune with the initial conceptualization of perceived service quality construct as a multidimensional second order formative construct with five first order reflective constructs. Hence the

172

content validity was further established. The findings confirmed that the structure of perceived service quality in the Kerala construct is hierarchical one formed with five first order dimensions all measured in the reflective manner. Thus objective in this regard is satisfied.

5.4 ANALYSIS OF BANK PERFORMANCE BASED ON SERVICE QUALITY

The previous sections have explained the steps in developing and confirming measurement scale for perceived service quality with regard to banking context. The next step was to analysis of bank’s performance based on perceived service quality for each type of banks like private sector, public sector and new generation in the Kerala context. This study adopted perception alone method suggested by Cronin and Taylor (1992), in SERVPERF analysis for evaluating the performance of each type of banks considered for the study. The SERVPERF score is the cumulative score obtained for each bank on all attributes (items) included in the validated scale of perceived service quality. Table 5.13 gives the SERVPERF score for each type of banks.

Table 5.13 SERVPERF details of each type of banks

173

On verification of the results, it was observed that service quality in

the new generation banks are ahead compared to other types of banks

followed by private sector banks and public sector banks. The retained

indicators after confirmatory factor analysis were used to calculate

SERVPERF scores for each type of banks. In image and human dimensions

private sector banks dominated perceived service quality supremacy whereas

in other three dimensions new generation banks are more accepted in the

Kerala banking context. The public sector banks even though enjoy second

position in system dimension was found last in overall service quality

perceptions.

A test of Homogeneity was performed to check whether the

assumption of homogeneity is violated to make valid inferences. Levene’s

test for homogeneity was not significant (p>0.05) as shown in Table 5.14 and

hence, it can be concluded that population variance of each group are

approximately equal. In order to find out the significant difference in the

perception towards various dimensions of perceived service quality among

the customers of three groups of banks, one-way analysis of variance was

administered. The resulted ‘F’ statistics are illustrated in Table 5.15 which

suggests that except products and services and system dimensions there exists

no significant difference in perceptions of customers in perceived service

quality at 0.05 levels.

Table 5.14Test of Homogeneity of Variances

174

Table 5.15 F-tests for significance among service quality dimensions

** Significant at 0.05 level

5.5 MEASUREMENT MODEL FOR “DESIRED EXPECTATION”

CONSTRUCT

In this study apart from perceived service quality dimensions,

reflective model was conceptualized for the construct “Desired Expectation”.

Hence validation of this construct was also done using Amos 16.0.

The Nine indicator variable model for “Desired Expectation”

dimension was suggesting poor fitting model in the first estimate. Two

indicator variables namely “helpful employees” and “Punctual” were showing

very poor values for squared multiple correlation and hence were removed for

further analysis. The resulting model was showing a better fit but further

improvement was required. As per modification indices, two error

correlations were added between indicator variables “easy transaction” and

175

“system consistency” as well as “speedy action” and grievance redressal”.

These steps can be theoretically justified as the chance of responses in

correlated manner to these questions was possible due to complimentary

nature of these questions. All the paths shown in the model are significant as

critical ratio were above 1.96 and the model is illustrated in Figure 5.10.In the

further analysis desired expectation was considered as a reflective construct

with seven indicator variables.

desiex

.43Knowledgeable employees e2

.65.64

Safety e3.80 .78Privacy e4.88

.53Easy transaction e6

.73

.68System consistency e7

.82

.57Speedy action e8

.75

.41Grievance redressal e9

.64.34

.26

CMIN/df-2.71,CFI-0.987,SRMR-0.026,RMSEA-0.067,PClose-0.139HoelterNat0.05-248, Composite reliability-0.701, AVE-0.754

Figure 5.10 Measurement Model for "Desired expectation"

5.6 HYPHOTHESIS TESTING – RESEARCH MODEL ANALYSIS

For the analysis of the thesis model, instead of covariance based

structural equation modeling, a variance based or component based Partial

least square (PLS) approach was adopted in this study. PLS-based SEM has

several key advantages over covariance-based SEM, including the following:

176

it always yield a solution, even in complex models

it does not require variables to meet parametric analysis criteria,

such as multivariate normality and large sample sizes

it enables the estimation of parameters in models with formative

LVs as well as reflective and doesn’t give rise to identification

problems as the case in Amos 16.0.

Most relationships between variables describing natural and

behavioral phenomena seem to be nonlinear, with U-curve and S-curve

relationships being particularly common (NedKock 2009). WarpPLS1.0

introduced in 2009 identifies nonlinear (or “warped”, hence the name of the

software) relationships among LVs and corrects the values of path

coefficients accordingly. Hence in this study Warp PLS 2.0 (current version)

was used for analysis of relationships among latent variables. The main

features of Warp PLS 2.0 are

It estimates P values for path coefficients automatically and

hence significance can be easily established.

It estimates several model fit indices for checking whether data

is well represented by the model.

It enables evaluation of measurement model as well as

structural model simultaneously

The software allows users to view scatter plots of each of the

relationships among LVs together with the regression curves

that best approximate those relationships.

177

It calculates variance inflation factor (VIF) coefficients for LV

predictors associated with each LV criterion.

It pre-process the data before SEM analysis and hence make it

easy to correct problems with the data, such as identical column

names, columns with zero variance, and missing values

In this study two constructs namely perceived service quality and

satisfaction were conceptualized as second order constructs. For analysis of

second order constructs using WarpPLS2.0, it is required to calculate the LV

scores at first by creating models with latent variables and indicators without

linking. These LV scores are used to define the second order construct in the

final model. The Path coefficients and associated p-values are obtained by

running WarpPLS 2.0 with a bootstrapping procedure. Boot strapping method

of re-sampling was adopted due to the reason it tends to generate more stable

path coefficients with samples sizes more than 100 (Nevitt and Hancock

2001). Various analysis algorithms used by Warp PLS are Warp3 PLS

Regression, Warp2 PLS Regression, PLS Regression, and Robust Path

Analysis. In this study Warp3 PLS Regression algorithm was used for

analysis.

The estimated model with path co-efficients and corresponding p

values are provided in Figure 5.11. The detailed results of analysis are

provided in Appendice-3. A pre condition for accepting the estimated model

for further interpretation was that the model should fit with the data.

Similarly the various validity and reliability criterion should be met. A model

possessing required reliability and validity conclude that the levels of

measurement errors in the data are relatively less and the results of analysis

credibly tests the hypotheses proposed in the study.

178

Figure 5.11 Estimated Research Model

Latent variable coefficients of the variables in the model are shown

in Table 5.16.

Table 5.16 Latent Variable Coefficients of the variables in the model

179

5.6.1 Model Validation-Verifying the Model Fit

To assess the model fit with the data, it was recommended that the

p-values for both the average path coefficient (APC) and the average

r-squared (ARS) be both lower than .05. In addition, it was recommended

that the average variance inflation factor (AVIF) be lower than 5 (Ned Kock

2009). Table 5.17 below provides the model fit indices with p values of the

estimated model. It was found that, all the three fit criteria were met and can

reasonably assume that the model have acceptable predictive and explanatory

quality as the data is well represented by the model.

Table 5.17 Model fit indices and P values of the Research Model

5.6.2 Validity of Reflective Constructs in the Model

The following rules were adopted to check the validity of reflective

constructs. The reflective constructs used in this model were “Desired

expectation” and the first order dimensions of “Perceived service quality”

construct.

5.6.2.1 Validation of “Desired Expectation Construct”

The construct of “Desired expectation” was found reliable since the

indicators like composite reliability co-efficient(0.927), Cronbach

alpha(0.907) and the average variance extracted (AVE=0.646) obtained after

the estimation of the model were above the threshold limits.

180

To check the Convergent validity loadings of each indicator of the

construct and their p values were considered. All the loadings were above 0. 5

and were significant at p <0.05and thus established convergent validity.

(Table 5.18).

Table 5.18 Factor loadings and p values for “Desired expectation” construct

The latent variable correlations in the model are considered for

ascertaining the discriminant validity of both reflective and formative

constructs. If the square root of the average variance extracted to be higher

than any of the correlations involving that latent variable (the values on the

diagonal latent variable correlation table of Warp PLS output should be

higher than any of the values above or below them, in the same

column).(Table 5.19).

Table 5.19 latent variable correlations of constructs

181

5.6.2.2 Validation of Various Dimensions of PSQ

The tables below (Tables 5.20, 5.21, and 5.22) establishes the

reliability, convergent validity and discriminant validity of the five

dimensions of the perceived service quality construct as per guidelines

mentioned in previous section. The results re-confirmed the findings from

confirmatory factor analysis.

Table 5.20 Reliability analysis of PSQ dimensions

Table 5.21 Factor loadings and p values for PSQ dimensions

182

Table 5.22 Latent variable correlations among PSQ dimensions

(All the correlations significant at p<0.001)

5.6.3 Validity of Formative Constructs

5.6.3.1 Validation of “Customer Satisfaction” Construct

The following conditions were verified to establish validity of

formative constructs

Absence of Multi collinearity was verified by checking the

Variance inflation factors (VIF) and found that they were less

than recommended value of 3.3 (Table 5.23).

Table 5.23 Indicator weights and VIFs of “Satisfaction” Construct

183

All the indicator weights were with P values lower than .05 and

hence merit validity in formative latent variable measurement

(Table 5.23).

All square root of AVE shown in the diagonal of the latent

correlation matrix are higher than the correlation involving that

latent variable, establishing Discriminant validity (Table5.24).

Table 5.24 Latent variable correlation of the “Satisfaction” Construct

(All correlations are significant at p<0.001)

5.6.3.2 Validity of other Formative Constructs

Formative indicators were used in this study for measurement of

all constructs except “Desired Expectation” and first order

dimensions of “Perceived Service Quality”. All this measures were found valid as explained below

Absence of Multi collinearity and indicator weights with P

values lower than .05 confirm validity of other formative

constructs (Table 5.25)

The Discriminant validity was established as under heading

5.5.2.1

In this model all the Average variance extracted(AVE) which is

measure of the amount of variance captured by a latent

construct in relation to the variance due to random measurement

error were above 0.5establishing discriminant validity of the

model (Table 5.16).

184

Table 5.25 Indicator weights of formative constructs

5.7 ANALYSIS OF RELATIONSHIP BETWEEN SERVICE

QUALITY DIMENSIONS TO CUSTOMER SATISFACTION

AND BEHAVIORAL INTENTIONS

To find out the service quality dimensions which have significant

relationship to customer satisfaction a model was developed as illustrated in

Figure 5.12 and was estimated using WarpPLS2.0. The model developed was

valid model with regard to fit indices. The significant dimension to have

direct impact on Customer satisfaction at p<0.01 was the “System” dimension

and the “Human” dimension was found significant at p<0.05. With regard to

influence of service quality dimensions to behavioral intentions of the

customer, it was observed that none of the dimensions are significant at0.01

level but two dimensions namely “Human” and “Products and Services”

influence positive behavioral intentions of the customer significantly at 0.05

185

level (Figure 5.13) which can be considered as valid information for drawing

conclusions on customer behavior in the banking context.

Figure 5.12 Model for PSQ dimensions to Satisfaction

Figure 5.13 Model for PSQ dimensions to Behavioral Intentions

186

5.8 SERVICE QUALITY AND CUSTOMER SATISFACTION –

DEMOGRAPHIC FACTORS

The Hypothesis to examine the association between demographics,

and service quality perceived and customer satisfaction was done by

developing another model as illustrated in Figure 5.14 and estimating the

model using Warp3 PLs algorithm with boot strapping procedure. On

verification of model fit indices with p values of the estimated model, it was

found that, all the three fit criteria are met and can reasonably assume that the

model have good fit with the data.

Figure 5.14 Model for demographic factors to PSQ and Satisfaction

On verification of p values for the path co-efficients from each

parameter to latent constructs PSQ and Satisfaction, it was found that, age of

the respondents have significant relation with perceived service quality

(p<0.05) at 0.05 level whereas none of the other demographic factors had

187

significant relation with perceived service quality or Customer satisfaction.

Thus Hypothesis H8 was only partly supported.

The p values for path co-efficients from bank type to PSQ and

satisfaction were found to be significant at 0.01 level (p<0.01) whereas length

of association with the bank was found to have significant relationship with

PSQ and satisfaction at 0.05 level (p<0.05) establishing the support for

hypothesis H9.

5.9 DESCRIPTIVE STATISTICS OF RESPONDENTS

The summary of demographic profile of the respondents was listed

below.

75.3% of the respondents are male

21.6% of the respondents are customers of private banks,37.1%

public sector banks and 41.3% new generation banks

19.2% of the respondents in the age group less than 20,36.1% in

the age group 20-35,32.2% between 35-50 and 12.5% above

50yrs

1.8% of the respondents were matriculates,35.3%

graduates,31.4% post-graduates and 31.4 professionals

10.9%ofthe respondents having annual income less than 2

lakhs,31.4% between 2lakh and 4 lakh,24.9% between 4 lakh

and 6 lakh and 32.7% above 6lakh

2.9% of the respondents were having length of association with

their bank for less than 1yr, 11.4% between 1yr and 3 yr, 15.6%

between 3yr and 5 yrs and 70.1% above 5 yrs.

188

5.10 ANALYSIS OF PATHS-TESTING OF HYPOTHESES

The next step of data analysis involved examining the structural

models in order to test various hypotheses proposed for the study. All the

paths in the model except two were found significant at 0.01level.The

hypotheses found insignificant were, H2a- desired expectation to service

quality and H 6b-Service quality to negative BI. The abstract of hypothesis

tested are provided in Table 5.26. The third objective of the study to identify

the linkages among various constructs used in the study was thus satisfied.

Except two, all other hypotheses proposed were found significant. The

insignificant relation between PSQ and desired expectation can be justified

from the fact that on determining perceived service quality the factors which

contribute are more critical rather than what are the desired expectations

regarding quality parameters. Also the significant relation desired

expectations bear with customer satisfaction underlines its importance in

developing satisfaction. Similarly from the quality perceptions of the sample

under study negative behavioral intentions were not expected exclusively due

to perceived service quality alone and found customer satisfaction has a vital

role in the development of both positive and negative behavioral intentions.

All the paths from indicators to corresponding constructs were

found significant in this study. The measurement variables to first order

service quality latent constructs such as ‘Image’, ‘Human’, ‘Convenience’,

‘Products services’ and ‘system’ were found significant in the confirmatory

factor analysis(Tables 5.8 and 5.20 above).All the indicators irrespective of

formative or reflective were found significant and thus confirmed content

validity of the theory developed.

189

Table 5.26 Results of hypothesis testing

chapter 5 data analysis -...

Documents