Targeting Item–level Nuances Leads to Small but Robust Improvements in Personality Prediction from Digital Footprints

Abstract

In the past decade, researchers have demonstrated that personality can be accurately predicted from digital footprint data, including Facebook likes, tweets, blog posts, pictures, and transaction records. Such computer‐based predictions from digital footprints can complement—and in some circumstances even replace—traditional self‐report measures, which suffer from well‐known response biases and are difficult to scale. However, these previous studies have focused on the prediction of aggregate trait scores (i.e. a person's extroversion score), which may obscure prediction‐relevant information at theoretical levels of the personality hierarchy beneath the Big 5 traits. Specifically, new research has demonstrated that personality may be better represented by so‐called personality nuances—item‐level representations of personality—and that utilizing these nuances can improve predictive performance. The present work examines the hypothesis that personality predictions from digital footprint data can be improved by first predicting personality nuances and subsequently aggregating to scores, rather than predicting trait scores outright. To examine this hypothesis, we employed least absolute shrinkage and selection operator regression and random forest models to predict both items and traits using out‐of‐sample cross‐validation. In nine out of 10 cases across the two modelling approaches, nuance‐based models improved the prediction of personality over the trait‐based approaches to a small, but meaningful degree (4.25% or 1.69% on average, depending on method). Implications for personality prediction and personality nuances are discussed. © 2020 European Association of Personality Psychology

How Are Personality States Associated with Smartphone Data?

“…A further advantage of XGBoost compared with neural networks is that it can handle missing values appropriately without the need for imputation. The second and third model are ridge regression (Hoerl & Kennard, 1970; James et al, 2013) and lasso regression (James et al, 2013; Tibshirani, 1996), which are two variations of regularized regression algorithms (Xing, Jordan, & Karp, 2001) and which have been widely used, including in personality research (Hall & Matz, 2020; Seeboth & Mõttus, 2018; Stachl et al, 2017; Stachl et al, 2019). Finally, we also explored ordinary least squares regression.…”

Descriptive, predictive and explanatory personality research: Different goals, different approaches, but a shared need to move beyond the Big Few traits

“…For example, technological progress has provided new sources of information (Rauthmann, 2020). Several recent articles describe how personality and its associations with other variables can be assessed through objectively measured behaviour or digital traces of behaviour (e.g., Cooper et al, 2020;Wiernik et al, 2020;Hall & Matz, 2020;Stachl, Au et al, 2020). These approaches offer great potential for noninvasively collecting personality-related information about large numbers of people and possibly over extended periods of time, hence allowing measurement of short-and even longer-term changes in personality.…”

Descriptive, Predictive and Explanatory Personality Research: Different Goals, Different Approaches, but a Shared Need to Move beyond the Big Few Traits

“…Several recent articles describe how personality and its associations with other variables can be assessed through objectively measured behaviour or digital traces of behaviour (e.g. Cooper et al, 2020;Hall & Matz, 2020;Stachl et al, 2020a;Wiernik et al, 2020). These approaches offer great potential for non-invasively collecting personality-related information about large numbers of people and possibly over extended periods of time, hence allowing measurement of short-term and even longer-term changes in personality.…”

“…using machine learning techniques; Wiernik et al, 2020) and then using these digital records-based self-report-approximations for descriptive or predictive purposes. The standard approach so far has been to predict the Big Few first and then use these predictions for whatever is their intended purpose, but recent evidence suggests that predicting narrower traits such as nuances first and using these predictions in subsequent analyses may be preferable (Hall & Matz, 2020). Again, more research is needed before we could recommend generally preferable practices and, therefore, it may be useful to systematically compare different approaches in their performance.…”

“…But often these assessment methods have to be given personality‐relevant interpretation in relation to subjectively rated personality traits before they become useful. For example, on its own, mobile phone sensor data do not have psychological meaning; they do once we know how they track with self‐reported traits (Hall & Matz, 2020; Stachl et al, 2020b; Wiernik et al, 2020). As a result, these methods often approximate subjectively rated traits rather than provide entirely new information, and any issues with self‐reports can spill over to their digital approximations (Tay, Woo, Hickman, & Saef, 2020).…”