Common Method Bias or Variance

CFA helps in understanding the extent of the common method bias. Common method bias or common method variance is a type of bias that could occur in certain situations when collecting data. In this particular case, since the data was collected using the same instrument, there is a possibility that there could be common method bias. Common method bias remains a threat to validity in Information systems research when using one method and despite the

majority of IS research using a single data collection method, only few studies investigated and mentioned it (Straub et al., 2004). Therefore, in this section, the author will investigate whether there are common method bias or common method variance effects.

The test that the researcher will use is the ―common method factor‖ technique for studies that do not measured a common factor explicitly (MacKenzie & Podsakoff, 2012) such as the case in this study.

To investigate for common method bias, the researcher introduced a common latent factor (CLF) to the CFA measurement model (see Appendix 8:

Common Method Bias Adjusted Model). However, introducing the CLF introduced an issue where the model cannot run and AMOS outputs the iteration limit reached message. Upon investigating the regression weights, the researcher noted that regression weights for RD items were very high (above one), what is known as a Heywood case (Hoyle, 2012; Kenny, 2014; Kline, 2012). One approach to fixing such an issue is to define the parameter estimate for the items to be the same, 1. However, once that is done, the model ran but a negative error appeared for item SI_1. Therefore, the researcher defined the error term for that item to be .4 (close to the error estimate of the second item for the same construct since they both measure the same thing), so that it doesn‘t become negative. Some authors discussed possible causes of such illogical values, including the possibility of model specification error or that it is caused by an issue in the sample (e.g. wide differences). In this study, because it is important to ensure that this is not a specification error, we assess discrepancies in section 6.4.2. Goodness of fit indices can also indicate if there are critical issues in the model as they would

indicate bad fit. Significant problems or issues in the model can be identified through the examination of these two areas.

The next step, then, was to compare standardised regression weights with and without the CLF. Comparison of both (see Appendix 7) shows some big differences between the regressions weights for both models. Therefore, this indicates there is a common method bias and that a large portion of the variance is being explained by the CLF.

As a result of this finding, the researcher was presented with two options at this stage, drop the affected items and continue without adding the CLF, or, to continue with the CLF added to the hybrid model. Simply dropping the items and continuing would mean that the information captured for these dropped items would be lost. Therefore, the researcher decided to take a more robust approach to deciding which of the two options to take.

First, the researcher would re-check the validity and reliability of the model that has a CLF (I.e. The common method bias adjusted model). If any

reliability or validity issues appear, the researcher would try to remedy them. Then, if an acceptable model is reached, it would be compared to the other model (the one without any affected items and with no CLF). Finally, the decision would be made based on the validity and reliability of each model first, and then, based on the GOF parameters and whether there were any issues (.e.g. influences) caused by the CLF.

Reliability and validity tests for the models are presented in Appendix 9 (Model without CLF) and Appendix 8 (Model with CLF).

Reliability and validity testing of the common method bias adjusted model (i.e. the model with CLF) shows a number of reliability and validity concerns (see Appendix 8 for full analysis). These concerns were related to: TandFC, ReInv, SRE, and RD constructs. The researcher attempted to remedy some of these concerns by dropping low loading items for these aforementioned constructs. Unfortunately, this led to further issues. Therefore, the researcher stopped here with regard to this option.

Next, the author dropped the affected items and made some minor

adjustments to the model. This resulted in the model without CLF presented in Appendix 9. Then, the reliability and validity of this model was tested and no reliability and validity issues were found.

Based on the above, although common method bias existed and was

influencing a number of items, those items were dropped. Common method bias might have occurred and affected these items as a result of some questions affecting how the respondent should respond next (Straub et al., 2004).

As explained before, another approach would have been to keep these items and continue with the analysis while having the CLF. However, the author has tested the model with CLF and it was clear that it had a significant negative effect on the model as well, causing reliability and validity issues. Therefore, for pragmatic reasons, the following affected constructs and items were dropped: FC_3 (Item), SI_INF (Construct), and RD (Construct).

In subsequent sections, the study will continue using the final modified model without CLF presented in Appendix 9 and also shown here for reference:

5.7.5 Outliers

Outliers are cases where some values are substantially different from others in the data set.

Using AMOS, the researcher investigated multivariate outliers using the Mahalonobis d-squared values which indicate observations farthest from the centroid (Byrne, 2010). Outliers can affect the model‘s GOF. However, in this study, the achieved model is very good. Still, the researcher decided to

investigate possible outliers, to see if there were any observations considered very far away from the rest of the observations.

A single observation was removed as it had a Mahalanobis d-squared value of over 100.000, over 15.000 in difference to the next observation which had a Mahalanobis d-squared value of 85.917. Therefore, reducing the number of cases (n:496).