Ridderprint

Chapter 7 140 Angrave et al., 2016; Bassi, 2011; Boudreau & Ramstad, 2007; Mondare, Douhitt, & Carson, 2011; Roberts, 2009; Schramm, 2006; Zielinski, 2014). This limited statistical savviness would have hampered the spread of people analytics in practice (Angrave et al., 2016; Cascio & Boudreau, 2010; Paauwe & Farndale, 2017). Similarly, the scientific community can be argued to trail behind because of their reliance on associative measures and explanatory modeling. Like other social science fields, the majority of HRM research is conducted via ordinary least squares regression analysis on structured, theory-driven datasets collected via cross-sectional questionnaires. Such analyses are in line with the traditional purpose of HRM research: the generation of explanatory management theories. Nevertheless, even the quality of these explanatory theories can improve once scholars become more open to alternative analytical strategies and techniques (see section 7.3.4 on Machine Learning; Shmueli, 2010; Yarkoni & Westfall, 2017). The purpose of people analytics – to gain actionable insights, potentially using novel, complex forms of HRM data – often requires alternative methodological approaches. Hence, people analytics projects often take a more inductive approach to explore which relationships exist in the data, rather than using the data to confirm scientific theories (Chapter 1; Chapter 3). People analytics researchers are often specifically concerned with howwell a model performs on “new”, unseen data (e.g., predictive accuracy) whereas how well models fit – or explain variance – in the current sample (e.g., information criteria, R 2 ) can be of secondary importance. Optimizing this predictive accuracy of a model requires a different workflow than the one common in conventional HRM research. For instance, people analytics researchers may use cross-validation and bootstrapping in order to train, test, and validate models on partitions of the data (Friedman, Hastie, & Tibshirani, 2001; James et al., 2013; Shmeuli, 2010; Yarkoni & Westfall, 2017). Additionally, this different purpose opens the floor for unconventional statistical models and data mining algorithms that trade off explanatory value and interpretability for predictive accuracy and exploratory capability. Examples include shrinkage (e.g., ridge regression), ensemble (e.g., bagging, boosting), deep learning (e.g., neural networks), Bayesian, and graphical methods (Friedman et al., 2001; James et al., 2013; Jebb, Parrigon, & Woo, 2016; Shmeuli, 2010; Woo et al., 2017; Yarkoni & Westfall, 2017). Such non-traditional methods and workflow can help to uncover new practical and scientific insights and improve the quality of HRM decisions. In this dissertation specifically, I applied network analysis to explore communication channels (Chapter 2) and survival analysis techniques to uncover turnover patterns and the effect of HRM practices thereon (Chapter 6). Both are infrequently applied in conventional HRM research, despite their applicability to a wide range of HRM topics (e.g., onboarding, succession planning). Furthermore, Chapter 3 illustrates how latent class analysis – or bathtub modeling (Croon et al., 2014) – helps to overcome multi-level issues in the HRM-performance relationship and how optimal