• No results found

7.1 Theoretical Model

7.1.1 Measurement Method

The theorized model was analyzed by employing partial least squares (PLS) path analysis with the statistical software application SmartPLS 2.0 (Hair, Hult, Ringle and Sarstedt, 2014; Ringle, Wende and Will, 2005) and SPSS analyses. The PLS variance-based approach emerged as more suitable as a traditional multiple regression procedure as this part of research was exploratory in nature and the research objective was to predict structural relationships of the theoretical model that regarded a formative measurement model (see Hair, Ringle, and Sarstedt, 2011for an overview of PLS application).

167

Table 42. Means, Standard Deviations and Aggregated Team Level Intercorrelations.

Variable M SD 1 2 3 4 5 6 7 8 9 10 11 12 13 Performance T1a 61.47 7.98 Performance T2b 66.37 8.78 .20 Messages Total T1 34.32 20.54 -.09 -.09 SAE T1 13.15 10.35 -.13 .13 .69** SAT T1 9.58 6.24 -.00 .13 .41** .12 TMS T1 14.84 8.27 -.15 .28* .76** .69** .54** Reflexivity T1 0.97 1.72 .08 .10 .46** .49** .48** .64** Messages Total T2 89.92 34.55 .02 -.03 .80** .52** .35** .55** .33** SAE T2 21.68 15.20 .02 .19 .57** .64** .12 .54** .45** .63** SAT T2 24.13 13.10 -.02 .20 .22 .20 .34 .15 .24 .46** .38** TMS T2 41.55 19.82 -.03 .26* .43** .47** .18 .37** .34** .61** .71** .69** Reflexivity T2 1.97 2.07 -.17 .23 .20 .4** -.04 .22 .20 -.21 .68** .45** .66** Language Diversity 0.29 0.46 -.16 .05 .03 .02 .02 .11 -.13 -.13 -.01 -.09 -0.4 .08 Ethnical Diversity 0.82 0.38 -.14 .02 .06 .04 .30* .21 .20 -.18 -.21 -.13 -2.0 -.25 .11 Note. N = 62 teams.

a Correlations are based on z-score transformed percentages. * p < .05.

** p < .01. aT1 = Task1. bT2 = Task2.

168

PLS has been applied in various fields of research such as in the area of spectral analysis in the chemical industry (Haaland & Thomas, 1988), feedback control (Piovoso & Kosanovich, 1994; Piovoso, Kosanovich & Pearson, 1992), discriminant analysis (Barker & Rayens, 2003), marketing and product quality research (Fornell and Bookstein, 1982; Mejdell & Skogestad, 1991), organizational and management research (e.g., Hulland, 1999, Konradt et al., 2015; Sosik, Kahai and Piovoso, 2009, Yoo & Kanawattanachai, 2001) and other applications. PLS allows for flexibility in modeling. For example, PLS allows to model conditions in experiments and use indicators or scales as measures of latent variables (Sosik et al., 2009). PLS has also been recommended as a robust tool for early stage research where the theoretical background is still developing (Falk & Miller, 1992), as is the case here.

From a measurement standpoint, PLS enables to conduct combined regression within the same statistical procedure in addition to applying reliability and validity statistics to test the underlying theoretical model (Wold, 1982). Further, PLS estimates relationships among latent variables by including measurement errors in the observable indicators (Fornell & Bookstein, 1982), deals with unreliability and heteroscedasticity issues efficiently (Martens & Naes, 1989), and ultimately does not make general assumptions about data distribution, observation independence or variable metric (e.g., Henseler, Ringle, & Sinkovics, 2009). PLS gets around the data normality assumption for estimating model parameters by using

subspaces (Martens & Naes, 1989; Sosik et al., 2009). As such, the data are converted into pseudo-variables (i.e., scores) to capture the variability in the data related to the predictor by presenting a linear combination of variables (e.g., Sosik et al., 2009, Wold, 1985). The modeling statistic facilitates modeling of a relatively large number of indicator variables that are either formatively or reflective. This research employs the formative measurement model, as the formative indicators cannot be substituted for each other but instead combine to give rise to the latent variables (Fornell & Bookstein, 1982). With the formative measurement model, a multiple regression model is estimated with the latent construct as the dependent variable and the assigned indicators as independent variables (Hair, Ringle & Sarstedt, 2011). Thus, the indicators are represented with regression weights leading to their latent variable (Hair, Hult, Ringle, and Sarstedt, 2014).

PLS is capable to estimate models with relatively small sample sizes (Chin and Newsted, 1999), which characterizes group research and this study (n = 62). For the PLS analyses, the minimum sample size should equal to either 1) 10 times the largest number of formative indicators used to measure one latent variable or 2) 10 times the largest number of

169

a structural path leading to a latent variable (Chin, 1997, 1998). Based on the rule of 10 (10 x 4 formative indicators and 10 x 6 structural paths leading to a latent variable) a sample size of 60 was required. This study exceeded the required sample size (n = 62) which was considered adequate for generating stable parameter estimates. One dummy variable was created to represent the experimental condition of reflexivity intervention (0 = no, 1= yes), which allowed for assessing mean differences between both conditions. Similar to Konradt and colleagues (2015), while testing different hypotheses, these dummy variables were applied to control for additional effects of the experimental conditions (see also Preacher & Hayes, 2008). Prior to analysis, all indicator variables needed to be standardized as variables were measured on different scales. Given that the PLS draws on standardized latent variable scores, data is automatically standardized (i.e., z-standardized, where each individual indicators has a Mean of 0 and Variance of 1) through the PLS-SEM algorithm (Hair et al., 2014). As previously stated, 62 observations were used to build the formative measurement model.

The theoretical model of volume of messages of cognitive behavior indicators and reflexivity were assessed. Although, using communication volume for examining a theoretical model regarding reflexivity and cognitive behaviors in PLS is a rather new approach, evidence exists of including communication volume in path model analyses to explain cognitive states (e.g., Yoo and Kanawattanachai, 2001).Variable distributions were inspected and statistics calculated to test normality using the Shapiro-Wilk test. Statistics indicated that the normality assumptions were not met for all indicators and a positively skewed distribution was present, confirming PLS to be the best approach in testing the underlying theoretical model. PLS comprises testing of two models 1) a measurement model (i.e., outer model), specifying the relationships between latent variables and their associated manifest (i.e., observed variables), and 2) a structural model (i.e., inner model), relating latent variables to other latent variables (Chin, 1998). Hence, after assessing the measurement model, the structural model of the theoretical-based model is considered.