After a 3 week period, 391 responses were received,representing a 41.6 % response rate. Removal of responseswith more than 20 % missing data resulted in 69 responsesthat were deemed unusable, leaving 322 usable responsesfor analysis. This sample size far exceeds the minimumrequired to provide sufficient analytical power to conduct aleast squares based partial least squares model fit (Cohen1988). In terms of gender, the total sample had 148 (46 %)females and 174 (54 %) males with a median age range of21–25 years. A comparison of early and late respondents,using an independent samplet-test revealed no significantdifference (Armstrong and Overton1977).Fig. 1Conceptual framework16K.G. Boakye et al.
4 Analyses and resultsTo discover the underlying structure of the data, we conductedan exploratory factor analysis (EFA) (Fabrigar et al.1999).Principal component analysis (PCA) with Varimax rotation(Nunnally1978) was conducted to explore the underlyingfactors. Using the criteria of eigenvalues greater than‘1’,along with examination of the scree plot and a cumulativepercentage of variance explained at 79.75 %, the existenceoffourfactorswassupported(seeTable1).UsingHarman’s one-factor test (Harman1976), we did not findany issue with common method bias in our dataset asnone of the four exogenous variables accounted for morethan half of the variance in the endogenous variable(Podsakoff et al.2003).4.1 Outer (measurement) model assessmentThe data was analyzed using structural equation modeling(SEM). Because our sample size was not sufficient to satisfythe power requirements for a covariance-based SEM approachand did not meet the multivariate normality assumptionsassociated with those methods, we used partial least squares(PLS) to analyze the data and test the overall significance ofthe model. PLS is a multivariate statistical technique that iswidely accepted in the operations management (Peng and Lai2012), information systems (Wetzels et al.2009; Keil et al.2000), and marketing journals (Barclay1991; Hair et al.2012). PLS’s latent variables are weighted composite scoresof the indicator variables, leading directly to explicit factorscores (Chin and Todd1995). In addition, it is less restrictiveon sample size and residual distribution restrictions (multivar-iate normality assumptions) than are found in other analysismodels (Chin et al.2003).In fact, the PLS approach is a robust method, providingboth measurement and structural information in terms ofindicator loadings and path coefficients. PLS parameters areestimated using a resampling approach (i.e., bootstrap orjackknife) since it lacks the classical parametric inferentialframework (Peng and Lai2012; Wold1982). Largely, weused PLS because (1) it has the ability to test the explanatorypower and predictive validity of our model, and (2) it handlesformative constructs without identification problems.