Examining the data is an essential step before starting the analysis since biases, outliers, or missing values can lead to violation of the multivariate analysis assumptions and consequently might lead to insignificant results or biases (Hair, et al., 2010).
4.1.1 Data Screening and Checking and Replacing Missing Values
First, data screening means that the research should ensure that the data are clean and ready to be used before conducting further analysis. Any data that might lead to biases, for example, responses answering all (1) strongly disagree or all (5) strongly agree, mainly when the questionnaire contains a negative question; this might demonstrate that the respondents did not read the items thoroughly before answering. Furthermore, the inconsistency in answers for the items within one dimension was also screened; these two screening steps were followed as a proactive approach to eliminate the possibility that the
134
respondents answered the question randomly. The following Table (14) indicates the number of questionnaires that contain answers with all 5, 4, 3, 2, and 1 to this criterion and that were removed from any further analysis.
Table 14:Data screening, detecting valid answers Type of answer Number Respondents
All 5 21 8,44,65,72,73,91,92,113,144,145,213,221,
226,267,307,310,320,356,369,371,377
All 4 8 120,122,219,252,316,319,329,430
All 3 1 182
All 2 0 -
All 1 0 -
Total 30
Second, missing values are defined as the unavailable valid value of one or more items and can lead to reducing the sample size and affect the data generalizability; some cases might lead to biased results (Hair, et al., 2010). In this research, no missing data were found since the questionnaire was distributed using an online platform with an option to make it mandatory for the respondent to answer all the questions; otherwise, the questionnaire will not be submitted. Nonetheless, to confirm this, descriptive statistics and frequency tests were run on the SPSS in order to confirm the absence of missing data. The results are provided in Table (15).
Table 15: Data screening, Missing Data analysis item
N Minimu
m
Maximu
m item
N Minimu
m
Maximu Vali m
d
Missin g
Vali d
Missin g CSA1
401 0 1 5 EPHY
1
401 0 1 5
CSA2
401 0 1 5 EPHY
2
401 0 1 5
CSA3
401 0 1 5 EPHY
3
401 0 1 5
CSA4
401 0 1 5 EPHY
4
401 0 1 5
DPO1
401 0 1 5 EPHY
5
401 0 1 5
DPO2 401 0 1 5 TP1 401 0 1 5
DPO3 401 0 1 5 TP2 401 0 1 5
135
DPO4 401 0 1 5 TP3 401 0 1 5
DPO5 401 0 1 5 TP4 401 0 1 5
DPO6 401 0 1 5 TP5 401 0 1 5
LIP1
401 0 1 5 CtextP
1
401 0 1 5
LIP2
401 0 1 5 CtextP
2
401 0 1 5
LIP3
401 0 1 5 CtextP
3
401 0 1 5
LIP4
401 0 1 5 CtextP
4
401 0 1 5
LIP5
401 0 1 5 CtextP
5
401 0 1 5
LIP6
401 0 1 5 CtextP
6
401 0 1 5
ECOG 1
401 0 1 5 CtextP
7
401 0 1 5
ECOG 2
401 0 1 5 CtextP
8
401 0 1 5
ECOG 3
401 0 1 5 Count
P1
401 0 1 5
ECOG 4
401 0 1 5 Count
P2
401 0 1 5
EEMO 1
401 0 1 5 Count
P3
401 0 1 5
EEMO 2
401 0 1 5 Count
P4
401 0 1 5
EEMO 3
401 0 1 5 Count
P5
401 0 1 5
EEMO 4
401 0 1 5
No missing data were found
As can be seen from the previous Tables (14 and 15), the analysis start with 431 respondent it was reduced to 401 after screening the responses. Accordingly, all the following tests have used the 401 valid responses.
4.1.2 Negatively Worded Items
All answers to the reversed questions were studied and reversed to fit with the normal questionnaire items. In this research, four negatively worded questions were found in the Employee Engagement questionnaire, as shown in the following Table (16):
Table 16: Negatively worded Items
Item Wording
ECOG2 I often think about other things when performing my job. (r)
136
EEMO3 I often feel emotionally detached from my job. (r) EPHY3 I avoid working overtime whenever possible. (r) EPHY5 I avoid working too hard. (r)
Since this is a (5) point Likert scale questionnaire, the answers were studied, and the responses were reversed. Answers with (5) were reversed to (1, 4 to 2) and vice versa;
nonetheless, (3) remained the same.
4.1.3 Test of Outliers
Outliers are responses with unique combinations of answers that may be distinguished from other responses by their high or low observations for one or more items (Hair, et al., 2010).
It is observations that do not adhere to a pattern similar to most of the data (Rousseeuw &
Van Zomeren, 1990). According to Hair et al. (2010), outliers might obviously affect statistical analysis; however, they should be assessed within the context.
The outlier detection can be in three forms, univariate, bivariate or multivariate. In univariate outlier detection, each item of the questionnaire should be examined, either by using graphical methods such as box and plot or by standardising the answers and comparing them against cut points (Hair, et al., 2010). Usually, the cut point is (-+1.95), which is the Z value for (95%) confidence intervals. In bivariate outlier detection, the variables are paired, and a scatter plot is detected. Finally, in multivariate outlier detection, several variables in the research are tested to find their outliers in a combined manner. In another way, the outliers are calculated concerning the research variables combined. It should be noted that responses classified as outliers in one test are not necessarily an outlier in another test.
Usually, for multivariate tests, multivariate outliers should be detected. Three tests can be used to detect the multivariate outlier: Cook's, Leverage or Mahalanobis distance; the latter are the most commonly used method (Garson, 2012).
137
The Mahalanobis distance can be calculated in SPSS and AMOS. In SPSS, the variables are aggregated and tested in a regressing test where the Mahalanobis distance is calculated. The value is then converted to Chi2- probability distribution, and any value less than (0.001) will be removed (Hair, et al., 2010). While in AMOS, all the variables with their observable items are introduced in AMOS, and the Mahalanobis distance is produced by the programme, the same cut point can be used as SPSS. However, the outliers should be studied within the context and how they might affect the mean of the items to decide on omitting them. The following Table (17) presents the responses IDs (Top 20) and the P-value for the Mahalanobis test.
Table 17:Outlier Detection
ID Mahalanobis p-value
387 1.17296 0.0010756
46 1.36918 0.0019941
124 1.42942 0.0023627
374 1.45942 0.0025632
96 1.47435 0.0026674
337 1.55335 0.0032687
227 1.66875 0.0043094
49 1.70252 0.0046529
370 1.78306 0.0055477
348 1.79239 0.0056586
79 1.80985 0.00587
155 1.91307 0.0072312
189 1.92445 0.0073933
256 2.02207 0.0088862
11 2.0484 0.0093211
355 2.22406 0.0125913
281 2.24274 0.0129783
135 2.24829 0.0130947
35 2.32535 0.0147824
101 2.33252 0.0149463
As can be seen from Table (17), no outliers were detected; all the P-values for the Mahalanobis distance were higher than the cut point (0.001). As a result, (401) valid responses will be used.
4.1.4 Common Method Variance
138
This research follows a self-reported method to collect the data from the designated sample.
Accordingly, this may lead to the risk of common method Variance (CMV). Collecting data from a single respondent for more than one item might create correlations between the items (Podsakoff, et al., 1990). Corrective measures such as statistical and post Hoc remedies can be used to avoid the common method variance issue.
Harman's single factor test is one of the most used approaches to detect CMV (Fuller, et al., 2016). In the test, it is assumed that if CMV exist, a single factor (component) will be responsible for more than (50%) of the variance. This can be detected using an exploratory factor analysis test that includes all the questionnaires in none rotated factor analysis with a principal component analysis in order to produce a variance table (Fuller, et al., 2016). Table (18) provides the results of the CMV test.
Table 18: Common Method Variance Test
Component Initial Eigenvalues Extraction Sums of Squared Loadings Total % Of
Variance
Cumulative % Total % Of Variance
Cumulative %
1 15.954 33.946 33.946 15.954 33.946 33.946
2 4.129 8.786 42.732 4.129 8.786 42.732
3 2.988 6.358 49.090 2.988 6.358 49.090
4 2.150 4.575 53.664 2.150 4.575 53.664
5 1.817 3.866 57.531 1.817 3.866 57.531
6 1.510 3.214 60.745 1.510 3.214 60.745
7 1.363 2.900 63.645 1.363 2.900 63.645
8 1.198 2.550 66.194 1.198 2.550 66.194
9 1.068 2.273 68.468 1.068 2.273 68.468
10 1.005 2.138 70.606 1.005 2.138 70.606
11 .846 1.801 72.407
12 .758 1.614 74.021
13 .734 1.561 75.582
14 .690 1.468 77.049
15 .616 1.310 78.359
16 .607 1.291 79.650
17 .581 1.236 80.886
18 .547 1.163 82.049
19 .530 1.128 83.176
20 .507 1.079 84.255
21 .460 .980 85.235
22 .441 .938 86.173
23 .428 .910 87.083
24 .408 .867 87.950
25 .394 .838 88.788
26 .384 .817 89.605
27 .372 .791 90.396
139
28 .352 .749 91.145
29 .349 .743 91.889
30 .328 .697 92.585
31 .314 .668 93.253
32 .294 .625 93.878
33 .280 .595 94.473
34 .266 .566 95.039
35 .253 .538 95.576
36 .240 .510 96.086
37 .231 .492 96.578
38 .219 .467 97.045
39 .196 .417 97.462
40 .187 .397 97.859
41 .178 .379 98.238
42 .168 .357 98.595
43 .157 .334 98.929
44 .154 .327 99.256
45 .126 .269 99.525
46 .118 .250 99.775
47 .106 .225 100.000
Extraction Method: Principal Component Analysis.
As can be seen in Table (18), introducing all the questionnaire items in an unrotated factor analysis resulted in (47) components, (10) components having Eigenvalues higher than (1), which have a Cumulative variance of (70.606%). Nevertheless, one component is responsible for (33.946%) of the variance, less than (50%); hence, according to Harman's single factor test, the CMV issue was not detected.