Data analysis

Data analysis is presented more in detail in each article. An overview of the anal-ysis is presented in the following section and summarized in Table 1. Articles 1 and 2 are based on data (n=206) that included also the largest IT services compa-nies. It is estimated that the 10 largest IT service companies account for roughly one-third of the revenue for the entire sector (Rönkkö & Peltonen 2012). Because of their importance to the Finnish economy and software industry at large, the two articles examined strategic learning among all sizes of software companies, in-cluding also the largest firms in the sector. To account for any size related effects, Article 2 used firm size as a control variable in the tested models. Due to the na-ture of Article 1, the tested measurement model did not include control variables.

Articles 3, 4, and 5 concentrate on data from 182 SMEs. The role of software SMEs in Finland has become particularly important after the decline of Nokia, as they have, for example, hired former Nokia and subcontractor employees that recently entered the job market (Rönkkö & Peltonen 2012). This focus allowed more detailed conclusions to be drawn for this specific context, specifically, the question of whether resource-scarce SMEs can benefit from strategic planning, EO, and strategic learning to be addressed. This also allowed the influence of possible confounding factors present in large organizations to be minimized. The definitional criteria of SMEs (less than 500 employees) is based on the U.S. gov-ernment’s classification of SMEs and was chosen because of the wide use of this classification in prior EO and learning research (e.g., Kreiser et al. 2010), thus enabling future studies to better compare the results of this study to those previ-ously completed.

3.6.1 Analysis methods

Following Hinkin´s (1995) scale development procedure, Article 1 employed both explorative factor analysis (EFA) and confirmatory factor analysis (CFA) as the main analysis methods to construct and validate measures of strategic learning.

The aim of these analyses was to examine the stability of the factor structure and provide information to facilitate the refinement of the new measure. The explora-tory factor analysis was conducted using SPSS software and principal axis factor-ing. Because the different strategic learning dimensions were expected to corre-late with each other, the Promax rotation method was chosen because its use is recommended when factors are assumed to correlate. The use of the oblique rota-tion method is advised, especially in the social sciences since behavior rarely functions independently, as it theoretically produces more accurate and more re-producible solutions (Thurston 1947). Confirmatory factor analysis was conduct-ed with LISREL 8.80. Fit analysis was conductconduct-ed using the Maximum Likelihood estimation. Considering that the items are non-normally distributed, all the anal-yses were also conducted using the Robust Maximum Likelihood technique de-veloped by Satorra and Bentler (1988). However, following the recommendation of Curran, West, and Finch (1996), the Maximum Likelihood estimation was ap-plied as a default method to determine whether similar results were obtained in the current study.

Article 2 employs a non-parametric approach to structural equation modeling, namely partial least squares (PLS) analysis (Chin, 1998). The PLS technique has enjoyed increasing popularity as a key multivariate analysis method in various research disciplines including strategic management (Hair, Ringle & Sarstedt 2013; Hulland 1999). PLS was chosen instead of the more traditional covariance-based SEM methods for three reasons. First, PLS is advantageous when sample sizes are small, providing more robust estimations and statistical power (Reinartz, Haenlein & Henseler 2009). Second, PLS allows the modeling of latent variables and simultaneous assessment of both measurement and structural models (Chin 1998). Third, it is considered one of the most suitable techniques when hypothesis testing is exploratory in nature rather than confirmatory (Hair, Sarstedt, Ringle &

Mena 2012). The PLS path modeling analyses were performed using SmartPLS 2.0 software (Ringle, Wende & Will 2005), with a path weighting scheme.

Hierarchical ordinary least squares (OLS) regression was chosen for data analysis in Articles 3 and 4. The main reason for choosing OLS regression was the expec-tation of non-linearity in the studied relationships. Unlike regression, that has established methods for handling non-linear relationships, structural equation modeling has no established tool for handling non-linear relationships (Gefen,

Straub & Boudreau 2000). Researchers have increasingly promoted the use of confidence intervals as the preferred inferential statistical method (e.g., Cashen &

Geiger 2004; Nickerson 2000), especially when interpreting interaction and non-linear terms. As noted by Brambor, Clark, and Golder (2006: 74) “it is perfectly possible for the marginal effect of X on Y to be significant for the substantively relevant values of the modifying variable Z even if the coefficient on the interac-tion term is insignificant.” This suggests that interacinterac-tion coefficients and the coef-ficients of the variables constituting the interaction (direct effects) cannot be in-terpreted as such when the interacting variables are continuous and do not include zero in the permissible range as is the case with non-linear and interaction terms applied in Articles 3 and 4. Following Brambor et al.’s (2006) advice, Articles 3 and 4 computed the marginal effect of the explanatory variable at various values of the moderator(s) and plotted them in figures at selected values to illustrate the interactions to provide further evidence on the study findings. In the figures in Articles 3 and 4, the confidence interval area signals 95% certainty for the line and can be used to determine when the marginal effect is significant. Simply put, when the confidence interval area is entirely on one side of the horizontal zero line, we can say with 95% confidence that the true value of the parameter is in our confidence interval and thus that the interaction effect is significant and positive (or negative, depending on the direction of the line). The analyses were performed using Stata 12 software.

Article 5 applies non-hierarchical k-means cluster analysis to identify and com-pare groups of companies with different strategic learning levels. K-means cluster analysis minimizes the variance within each cluster while maximizing between group variance (Punj & Stewart 1983). The non-hierarchical clustering method was chosen as it allows observations to switch cluster membership and is less sensitive to outliers than hierarchical methods (Ketchen & Shook 1996; Punj &

Stewart 1983). To determine whether the identified strategic learning clusters vary in terms of performance, one-way ANOVA was conducted. Tukey´s post hoc analysis was used to test which clusters statistically significantly differ from each other in terms of performance. The analyses were performed using SPSS version 20.

Table 1. Research characteristics and methodologies used in the articles exploitation  strategies     Mediator  =  Strategic  

1From  Orbis  secondary  database;  IV=independent  variable;  DV=dependent  variable