WV) have a high linear dependence and multicollinearity can be considered a
problem^18.
Having employed regression techniques with summated scale variables to identify
collinearity (albeit at an apparently moderate level) between the Employee
Empowerment and Work Values constructs, the next step is to seek a suitable
remedy for this situation.
A number of remedies for multicollinearity are discussed by Cohen et al. (2003:
425 - 430 ) including: (i) the combination of independent variables if they are
measuring the same or very similar concepts (p. 426); (ii) the collection of
additional data to increase sample size and improve the precision of the structural
coefficient estimates (p. 427); and (iii) removal of one or more independent
variables (p. 430).
Remedy (i) is not appropriate in this instance because the discriminant validity
estimates from the measurement model (CFA 5:8) confirm that the EM, WV and
ML constructs do indeed measure discrete concepts. Remedy (ii) is recommended
for future research to pursue this analysis using a larger sample. For the current
research, remedy (iii) is followed and the Work Values construct is removed
(because this construct has the non-significant structural coefficient). Remedy
(iii) (removal of one or more independent variable) is also supported, specifically
in an SEM context, by Kline (2005: 57).
The decision to remove the Work Values construct is further supported following
the outcome of a second approach to investigating the issue of the non-significant
WV→EPA path. This second course of action is described as follows.
Following the parsimonious principle (see e.g. Kline 2005: 145-147 and Raykov
and Marcolides 2006: 41-43), which guides researchers to find a parsimonious
model that maintains a satisfactory fit to the data, it is common in SEM analyses
to remove non-significant parameters (see also Byrne 2010: 185).
Removing parameters in this way requires researchers to compare the alternative
‘nested’ models using the chi square difference test (Δχ^2 ) with the objective of
finding “a parsimonious model which fits the data reasonably well” (Kline 2005:
18
NB – Garson (citing Belsley et al. 1980) notes that it is possible to find one
multicollinearity diagnostic (in this case the condition index) indicating a problem while
another (the VIF) does not.