The role of digital data (and especially big data) has had an impact on social science research methods as a result of the 'computational and algorithmic turn’. In order to fully appreciate such a turn, we can contrast the difference between the two cultures of modelling’ (Breiman, 2001; Gentle et al., 2012).

One difference with conventional quantitative analysis is geared towards the estimation of main effects and less interaction ones. Conventional modelling seeks to estimate the simultaneous contributions of several covariates in predicting an outcome variable (Y). And yet, as brilliantly argued by Ragin (2008), most social science problems require a different problem logic from the one above – to do with understanding combinations of factors associated with different outcomes rather than isolating the role of individual predictors.

Ragin contrasted the independent variables approach with the ‘configuration of factors’ one. Traditional multivariate methods can include interactions by adding interaction terms (Jaccard and Turrisi, 2003), but this is not as common as one might expect. There are many reasons for this, but one, in particular, is very practical and becomes extremely pressing in the context of large and high- resolution datasets, which are more and more common in digital social research: it is, namely, that large numbers of potential interactions can be reached very quickly by only considering two-way ones.

Computational approaches have developed techniques to deal with this problem of achieving a fast test of thousands of possible interactions among predictors in order to determine which ones are important to predict a given dependent variable and which have no impact. Computational methods like classification and regression trees (CART) and decision tree chi-square automatic interaction detection (CHAID) identified meaningful interactions and used them to predict a dependent variable in line with the critique and recommendations made by Ragin about the importance of the ‘configurations of conditions’ (as opposed to simply independent variables). 

We will discuss one example represented by the development of model-based recursive partitioning. This approach is an improvement on the use of CART, which is also a method coming from the ‘algorithmic culture’ of modelling, and that has useful applications in social science.

download

Giuseppe Alessandro Veltri - bio holds a BSc in Psychology of Communication from the University of Siena, an MSc in Social Research Methods (Statistics) from the Methodology Institute of the London School of Economics (LSE) and a PhD in Social Psychology from the LSE.

He is Full Professor of Research Methodology and Cognitive Sociology at the Department of Sociology and Social Research of the University of Trento.

He is coordinator of the Data Science postgraduate degree at University of Trento and coordinator of the doctoral programme in Sociology. He was Senior Lecturer at University of Leicester.

He has been Lecturer at University of East Anglia and a scientific fellow at the European Commission JRC Institute for Prospective Technological Studies (IPTS). Before joining the IPTS, he has been a research associate at the Institut Jean Nicod (Ecole Normale Supérieure) in Paris.

He has published in scientific journals such as Nature, PLOS One, Computers in Human Behavior, Public Understanding of Science, Big Data & Society, Behavioural Public Policy and others.