‘LinkedIn, LinkedIn on the screen, who is the greatest and smartest ever seen?’: A machine learning approach using valid LinkedIn cues to predict narcissism and intelligence

Härtel, Tobias M.; Schuler, Benedikt A.; Back, Mitja D. Article — Published Version ‘LinkedIn, LinkedIn on the screen, who is the greatest and smartest ever seen?’: A machine learning approach using valid LinkedIn cues to predict narcissism and intelligence Journal of Occupational and Organizational Psychology Provided in Cooperation with: John Wiley & Sons Suggested Citation: Härtel, Tobias M.; Schuler, Benedikt A.; Back, Mitja D. (2024) : ‘LinkedIn, LinkedIn on the screen, who is the greatest and smartest ever seen?’: A machine learning approach using valid LinkedIn cues to predict narcissism and intelligence, Journal of Occupational and Organizational Psychology, ISSN 2044-8325, Wiley Periodicals, Inc., Hoboken, NJ, Vol. 97, Iss. 4, pp. 1572-1602, https://doi.org/10.1111/joop.12531 This Version is available at: https://hdl.handle.net/10419/313767 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by-nc/4.0/ 1572 | J Occup Organ Psychol. 2024;97:1572–1602. wileyonlinelibrary.com/journal/joop Received: 27 October 2023 | Accepted: 2 July 2024 DOI: 10.1111/joop.12531 RESEARCH ARTICLE ‘LinkedIn, LinkedIn on the screen, who is the greatest and smartest ever seen?’: A machine learning approach using valid LinkedIn cues to predict narcissism and intelligence Tobias M. Härtel1 | Benedikt A. Schuler2 | Mitja D. Back3 This is an open access article under the terms of the Creative Commons Attribution-NonCommercial License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes. © 2024 The Author(s). Journal of Occupational and Organizational Psycholog y published by John Wiley & Sons Ltd on behalf of The British Psychological Society. Selected results from an earlier version of this manuscript were presented at a conference: Härtel, T. M., Schuler, B. A., & Back, M. D. (2023, August 4–8). Using valid cues to predict narcissism and intelligence from LinkedIn profiles [Conference presentation]. The 83rd Annual Meeting of the Academy of Management (AOM), Boston, USA. https:// journ als. aom. org/ doi/ abs/ 10. 5465/ AMPROC. 2023. 12425 abstract. 1School of Business Administration and Economics, Osnabrück University, Osnabrück, Germany 2Institute of Management and Strategy, University of St.Gallen, St. Gallen, Switzerland 3Department of Psychology, University of Münster, Münster, Germany Correspondence Tobias M. Härtel, School of Business Administration and Economics, Osnabrück University, Rolandstr. 8, 49069 Osnabrück, Germany. Email: [email protected]e Abstract Recruiters routinely use LinkedIn profiles to infer applicants' individual traits like narcissism and intelligence, two key traits in online network and organizational contexts. However, little is known about LinkedIn profiles' predictive potential to accurately infer individual traits. According to Brunswik's lens model, accurate trait inferences depend on (a) the presence of valid cues in LinkedIn profiles containing information about users' individual traits and (b) the sensitive and consistent utilization of valid cues. We assessed narcissism (selfreport) and intelligence (aptitude tests) in a sample of 406 LinkedIn users along with 64 LinkedIn cues (coded by three trained coders) that we derived from trait theory and previous empirical findings. We used a transparent, easy- to- interpret machine learning algorithm leveraging practical application potentials (elastic net) and applied state- of- theart resampling techniques (nested crossvalidation) to ensure robust results. Thereby, we uncover LinkedIn profiles' predictive potential: (a) LinkedIn profiles contain valid information about narcissism (e.g. uploading a background picture) and intelligence (e.g. listing many accomplishments), and (b) the elastic nets sensitively and consistently using these valid cues attain prediction accuracy (r = .3 5/. 41 for narcissism/intelligence). The results have practical implications for improving recruiters' accuracy and foreshadow | 1573 PREDICTING INDIVIDUAL TRAITS FROM LINKEDIN INTRODUCTION With the growth of online networks, practices like ‘cybervetting’ and ‘social media assessments’ (Cubrich et al., 2021; Hartwell & Campion, 2020) inform recruiters' selection decisions. LinkedIn is the most popular online professional network with 950 million users (LinkedIn, 2023). Recruiters routinely use applicants' LinkedIn profiles to infer traits (e.g. Roulin & Levashina, 2019), such as the presence of an About section as a signal of trait selfpresentation (Van de Ven et al., 2017). LinkedIn profiles do not only provide information available in resumés (e.g. educational/professional experiences) or online social networks, such as Facebook (e.g. number of connections), but add information like followed interests and recommendations (Fernandez et al., 2021). Despite calls for empirical insights (Roth et al., 2016; Van Iddekinge et al., 2016), LinkedIn's predictive potential to infer individual traits remains unclear. There are few robust and mixed findings on the LinkedIn information (cues) signalling traits, leading to contradictory conclusions as to whether LinkedIn profiles allow for accurate inferences (Fernandez et al., 2021) or not (Roulin & Stronach, 2022). Also, recruiters were shown to achieve only modest accuracy when inferring traits on LinkedIn (e.g. Roulin & Levashina, 2019; Van de Ven et al., 2017). This may be due to LinkedIn's inherent lacking capacity to signal valid trait information or perceivers' lack of sensitive (weighting valid cues more strongly than nonvalid cues) and/or consistent (weighting cues the same way across targets) use of such information. This study aims to illuminate LinkedIn's predictive potential for trait inferences using a twofold lens model (Brunswik, 1956) approach (Figure 1). First, we identify robust LinkedIn cues conveying valid trait information. Second, we examine machine learning algorithms' accuracy, acting as automated perceivers sensitively and consistently using these valid cues (Bleidorn & Hopwood, 2019; Tay et al., 2020). If valid information is present in LinkedIn profiles, automated perceivers will detect and consistently weight these cues across targets, thereby revealing LinkedIn's predictive potential. We examine LinkedIn profiles (N = 406) coded by three trained coders, alongside users' narcissism and intelligence test scores. Narcissism contributes to understanding workplace outcomes (Judge et al., 2006) and is expressed in online selfpresentation (Gnambs & Appel, 2018). Intelligence is a potentials and limitations of automated LinkedInbased assessments for selection purposes. KEYWORDS Brunswikian lens model, cybervetting, LinkedIn, machine learning, trait assessment Practitioner points • LinkedIn profiles hold predictive potential to infer applicants' individual traits, such as narcissism and intelligence. • Information about which LinkedIn cues provide valid trait signals and which do not is useful for improving recruiters' cybervetting accuracy. • The use of machine learning algorithms to automate LinkedInbased trait assessment may offer a nonintrusive approach to complement traditional trait testing in selection. • Before automated LinkedInbased trait assessments may be applied, more research is needed on the psychometric properties, adverse impacts, and applicant reactions. 1574 | HÄRTEL et al. key predictor of job performance (Sackett et al., 2022), and recruiters believe they can assess it from LinkedIn (Hartwell & Campion, 2020). We derived 64 LinkedIn cues based on theoretical/empirical evidence to signal these traits. This ensured relevance of cues, addressing content validity concerns in automated approaches. State- of- theart resampling (nested crossvalidation) provided robust results and an easy- to- interpret machine learning algorithm (elastic net) maximized practical applicability. We enrich the literature on trait expression in selection contexts by (1) adding robust LinkedIn cues signalling narcissism (e.g. uploading background pictures) and intelligence (e.g. listing schools with many followers) to the information bases providing valid trait signals (e.g. resumés, job interviews; Härtel et al., 2024; DeGroot & Gooty, 2009). Also, we enrich the literature on automated trait assessments for selection purposes by (2) showing that automated perceivers can accurately infer traits not only from online social networks like Facebook (Azucar et al., 2018) but also from professional networks like LinkedIn (prediction accuracy r = .35/.41 for narcissism/intelligence). Combining these contributions (3) clarifies mixed findings on LinkedIn's potential for accurate trait inferences: LinkedIn offers the possibility to make accurate trait inferences when sensitively and consistently incorporating valid cues (as identified in this study) into trait inferences (like automated perceivers do). Applying the lens model to (automated) trait inferences based on LinkedIn profiles This study examines the accuracy of inferring individual traits of an unknown target person from their observable LinkedIn profile information. Following the lens model (Brunswik, 1956; see Back & Nestler, 2016; Hammond, 1996; Karelaia & Hogarth, 2008; Nestler & Back, 2013), latent traits are inferred indirectly in this situation by drawing on observable cues (Figure 1). The necessary prerequisite for accuracy is the presence of valid cues – LinkedIn information associated with users' traits (left side of the lens model). LinkedIn may contain such valid information (Fernandez et al., 2021), but prior findings were instable (Roulin & Levashina, 2019; Roulin & Stronach, 2022). This raises concerns of whether LinkedIn profiles actually hold the capacity for accurate trait inferences. FIGURE 1 Brunswikian lens model in the context of automated cybervetting based on LinkedIn profiles. | 1575 PREDICTING INDIVIDUAL TRAITS FROM LINKEDIN The sufficient prerequisite for accurate LinkedInbased trait inferences is that valid cues are used sensitively and consistently when forming judgements (right side of the lens model). Sensitive cue utilization means that perceivers prioritize valid cues over nonvalid cues – being sensitive for what relevant and irrelevant cues are. Consistent cue utilization means weighting the cue information in the same way across targets – applying consistent judgement rules to each target (Back & Nestler, 2016). Although LinkedIn profiles likely contain some valid trait information, recruiters' inferences have been lacking accuracy (Roulin & Levashina, 2019; Roulin & Stronach, 2022; Van de Ven et al., 2017). This may stem from the lack of sensitivity and/or consistency in using valid cues. For instance, recruiters use the presence of a profile picture as a cue for agreeableness (Van de Ven et al., 2017), despite no actual association (Roulin & Levashina, 2019; Van de Ven et al., 2017), indicating a lack of sensitivity. Also, human perceivers are typically inconsistent in applying judgement rules across targets, indicating a lack of consistency (Karelaia & Hogarth, 2008). The lens model's value lies in its scope covering human and automated approaches (Cannata et al., 2022; Tay et al., 2020) – the human recruiter is replaced by a machine learning algorithm in the automated approach. Algorithms can be construed as automated perceivers constructed to sensitively and consistently use valid cues to make inferences. They autonomously learn which cues are (not) valid based on training data (sensitivity) and apply consistent weights across targets when inferring traits in new data (consistency). Focusing on the automated perceiver thus allows us to quantify LinkedIn's predictive potential purified from the imperfections of human judgement: If LinkedIn profiles contain a valid cue base signalling narcissism/intelligence, the automated perceiver should achieve prediction accuracy because it fulfils the sufficient prerequisite for accuracy. Previous research on valid cues signalling traits on LinkedIn Initial research examined cue validities using small explorative LinkedIn cue sets (9 ≤ NCues ≤ 22) in relatively small samples (97 ≤ N ≤ 154; Roulin & Levashina, 2019; Roulin & Stronach, 2022; Van de Ven et al., 2017). This exploratory approach offers initial insights; for example, users high on selfpresentation often included an About section, providing opportunity for selfpromotion (Van de Ven et al., 2017). Yet, some findings were puzzling – selfpresenters showing higher education (Van de Ven et al., 2017) – and instable – none of the valid cues in Roulin and Levashina (2019) was confirmed by Roulin and Stronach (2022) and vice versa. While these findings suggest limited predictive potential of LinkedIn, the exploratory nature warrants caution. In contrast, using a deductive approach recently enabled to demonstrate the existence of a valid LinkedIn cue base for Big Five traits. Fernandez et al. (2021) addressed prior limitations by reporting cue validities of (a) a wider range of LinkedIn cues (NCues = 33) (b) based on theoretical underpinnings to signal Big Five traits (c) in a large sample of 607 users. They suggest that LinkedIn could offer a sufficient cue base for accurate trait inferences, at least for selected Big Five traits. We address mixed prior findings by expanding on valid LinkedIn cues signalling individual traits in terms of narcissism/intelligence. We therefore (a) examine an extensive cue set (NCues = 64) (b) derived from trait theory and empirical findings ensuring relevance and interpretability (c) in a sample of 406 users. We prioritize a high ratio of potential valid cues per trait to avoid underestimating predictive potentials. Also, while single traitcue links typically show small effect sizes, combining subtle cues can reveal meaningful information. Further, we identify valid cues not only from correlations but based on feature importance in nested crossvalidated models. This addresses concerns of capitalizing on chance and overfitting by prioritizing robust cues based on their ability to predict novel data. Automated approach to trait inferences based on LinkedIn Automated approaches applied to online social networks like Facebook based on various cue sets (e.g. posts/pictures/likes) achieved remarkable prediction accuracy for individual traits (Azucar et al., 2018; 1576 | HÄRTEL et al. Settanni et al., 2018), including narcissism (e.g. Garcia & Sikström, 2014; Sumner et al., 2012) and intelligence (e.g. Kosinski et al., 2013, 2014). Unlike social networks like Facebook, LinkedIn focuses on professional identity (Hartwell & Campion, 2020), making it advantageous for selection contexts due to fewer issues with adverse impact, acceptance, and legal matters (Stoughton et al., 2015). An initial attempt to transfer the automated approach to LinkedIn used automated languagebased Big Five assessments, but found no prediction accuracy (Roulin & Stronach, 2022). This raised calls for less reliance on limited textual information, showing weak trait associations (Chen et al., 2020; Holtzman et al., 2019). Generally, research using automated approaches typically applies datadriven, explorative approaches without preselecting cues based on theory/empirical evidence (Settanni et al., 2018). While useful for uncovering novel associations and theory building, this lacks content validity and has limitations: (a) related constructs might be measured rather than trait content, (b) cues' interpretability may suffer reducing application potentials (‘black box’), and (c) generalizability across online networks is constrained (Alexander et al., 2020; Bleidorn & Hopwood, 2019; Tay et al., 2020). We quantify LinkedIn's predictive potential by applying the automated approach. We go beyond textual cues by targeting an extensive set of cues that depict qualitatively different types of information reflecting LinkedIn's broad information spectrum. This aligns with the ‘good information’ principle (Back & Nestler, 2016), positing that interpersonal judgements' accuracy depends on the quantity and quality of accessible information. Incremental valid information enhances accuracy (Borkenau et al., 2004; see Azucar et al., 2018; Settanni et al., 2018). Including a diverse array of cues is thus important to depict predictive potentials: If valid cues exist but are not coded, predictive potentials are underestimated. We base our cue set on theory and empirical evidence, establishing conceptual cue–trait links. We prioritize elastic nets over more complex algorithms, emphasizing transparency of how inferences are made over gaining some predictive power. Elastic nets are regularized regressions that avoid insample overfitting and optimize out- of- sample prediction accuracy by shrinking the regression coefficients of cues that contain less predictive information towards zero (see Hastie et al., 2009; James et al., 2013, for an introduction). This makes them parsimonious models, similar in interpretation to multiple linear regressions, but better at predicting new data (see Analytical Approach). Transparent selection decisions are essential for ensuring fair and legal procedures (Goretzko & Israel, 2022). Relevance of narcissism and intelligence in LinkedIn profile assessment We heed calls to prioritize alternative traits in cybervetting (Mönke & Schäpers, 2022; Settanni et al., 2018). Grandiose narcissism is a form of entitled selfimportance with agentic and antagonistic components (Back, 2018) evoking organizational consequences (Campbell et al., 2011) beyond the Big Five (Judge et al., 2006). Due to assertive charm, individuals high on narcissism often propel into managerial roles (Wille et al., 2019). While some job roles benefit from moderate narcissism (Satornino et al., 2023), high levels pose risks due to overconfidence and disregard for others (O'Boyle et al., 2012). Research (Blair et al., 2008; De Fruyt et al., 2009; Judge & LePine, 2007) and the practitioner literature (Rotolo & Bracken, 2022; Schwarzinger, 2022) advocate assessing narcissism in hiring, despite challenges using traditional selection methods (e.g. selfreports, job interviews) due to socially desirable responding (Bensch et al., 2019; Kowalski et al., 2018) and impression management (Paulhus et al., 2013). Online networks serve individuals high on narcissism as platform to exercise their need for external validation by providing selfenhancement opportunities (Gnambs & Appel, 2018; McCain & Campbell, 2018). Thus, LinkedIn may offer an alternative to assess narcissism. Intelligence – the ability to process complex information facilitating higher order thinking like reasoning/problemsolving (Gottfredson, 1997) – is the most potent trait to predict job performance across occupations (Sackett et al., 2022), particularly in complex jobs (Salgado et al., 2003) and highlevel leadership positions (Ones & Dilchert, 2009). Whereas little is known about expressing intelligence online (Schroeder et al., 2020), recruiters use LinkedIn to infer intelligence (Roulin & Stronach, 2022; see Brown & Campion, 1994; Kluemper & Rosen, 2009) and consider it an effective source to do so | 1577 PREDICTING INDIVIDUAL TRAITS FROM LINKEDIN (Hartwell & Campion, 2020). LinkedIn may offer a noninvasive alternative to established but less commonly applied intelligence tests (König et al., 2010; Zibarras & Woods, 2010). METHOD Sample/procedure The sample comprised 406 (243 female) Germanspeaking LinkedIn users with substantial subscribers (M = 297.15, SD = 468.32) recruited through online postings, participant recruitment platforms, and university lecture announcements. The average age was 29.47 (SD = 9.69), 283 (69.70%) participants held a bachelor's degree or higher, 258 (63.55%) were studying (37.59% business/economics), and 215 (52.96%) were working ≥ 20 weekly hours in various sectors/professions. Participants completed an online survey assessing demographics and individual traits. At survey start, participants connected their LinkedIn profiles with a LinkedIn showcase page. We saved profiles in various formats, including pdf exports to assess profile length, profile pictures as separate images, and screenshots of other relevant information. For details on the survey and psychometric tests, see the Codebook, https:// osf. io/ 4ruqj/ . Measures Individual traits Table 1 presents descriptive statistics and Cronbach's alphas for traits. The 18- item Narcissistic Admiration and Rivalry Questionnaire (NARQ; Back et al., 2013) measured narcissism with 6- point scales ranging from 1 (do not agree at all) to 6 (agree completely). The NARQ captures the subdimensions of narcissistic admiration (agentic selfpresentation) and narcissistic rivalry (antagonistic selfdefence), together constituting a theoretically informed assessment of grandiose narcissism. To ensure a comprehensive intelligence measure, we included fluid intelligence (reasoning/problemsolving) and crystallized intelligence (accumulated knowledge) as general intelligence subdimensions (Cattell, 1963; Nisbett et al., 2012). To manage survey duration, we used short scales offered by a wellestablished openaccess repository for measurement instruments in German (GESIS – Leibniz Institute for the Social Sciences, n.d.). The short version of the Hagen Matrices Test (HMT- S; Heydasch et al., 2012, 2013) measured fluid intelligence (M = 4.31, SD = 1.45). The test consists of six items where testtakers select the figure that logically completes a 3 × 3 matrix from eight alternatives. The short version of the Berlin Test for the Assessment of Fluid and Crystallized Intelligence (BEFKI GC- K; Schipolowski et al., 2013; Wilhelm et al., 2014) measured crystallized intelligence (M = 8.78, SD = 2.04). The test comprises 12 items presenting declarative knowledge questions from various fields (e.g. ‘In a wellknown painting by Dalí, “deliquescent clocks” are depicted. To which style can this painting be assigned?’), with testtakers selecting the correct answer from four options. Scores for both tests are based on the number of correctly completed items. We aggregated both measures' zstandardized scores (r = .26, p < .001) to obtain a comprehensive, accurate intelligence indicator by reducing measurement error and enhancing content validity (Breit et al., 2024). This also simplifies interpretation in personnel selection. The modest correlation of r = .26 between fluid and crystallized intelligence reflects their conceptual differences as related but distinct constructs (Ackerman et al., 2001; Ziegler et al., 2012). LinkedIn cues Cue derivation was based on trait theory and empirical findings ensuring content validity and interpretability. The cue set was developed before coding, without knowing the results, though not preregistered. 1578 | HÄRTEL et al. The fourstep approach yielded 64 LinkedIn cues, categorized into two types. Objective cues (n = 55) were straightforward to code by a single coder (e.g. counting skills). Subjective cues (n = 9) required judgement (e.g. rating physical attractiveness) and were rated by two coders on scales from 1 (not at all) to 6 (completely). The three coders were thesis candidates or research interns and received extensive training. Interrater agreement was good to excellent (Koo & Li, 2016; ICC3,k ≥ .84; see Table 1). We touch upon the fourstep cue derivation process for illustrative purposes; details are provided in the Supplemental Results Section A in Appendix S1, https:// osf. io/ 4ruqj/ . (1) We derived cues transferring theoretical implications/empirical findings on narcissism/intelligence from broader nononline contexts to LinkedIn. For example, individuals high on narcissism are motivated to attain leadership (Benson et al., 2016), serving as platform to earn admiration (Campbell & Campbell, 2009). Their assertiveness (Härtel et al., 2021) helps narcissistic individuals attain group leadership (Grijalva et al., 2015), managerial ranks (Wille et al., 2019), and prestigious leadership roles (Watts et al., 2013). On LinkedIn, this may manifest through leadership positions/ skills in the Experience/Skills sections. The same may hold true for intelligence as a key leadership attribute and predictor of leadership emergence (Judge et al., 2004). (2) We transferred theoretical implications/empirical findings on associations between narcissism/intelligence and cues in related contexts like online social networks and resumés to LinkedIn. For instance, individuals high on narcissism often boost their grandiosity through online posting (McCain & Campbell, 2018). This selfpromotion is effective with large audiences, providing attention and popularity (Marshall et al., 2020). Indeed, narcissism is associated with large online networks (Gnambs & Appel, 2018). On LinkedIn, this may manifest through frequent posting/liking/commenting and large network sizes. As another example, intelligence was shown to be expressed in listing scholastic awards on resumés (Cole et al., 2003), as it enhances learning and academic/occupational success (Kuncel et al., 2004; Ng et al., 2005). Accordingly, intelligence might be expressed through honours/awards in LinkedIn's Accomplishments. (3) We added cues linked to narcissism/intelligencerelated traits in prior LinkedIn research. For instance, whereas selfpresenters were found to exhibit less smiling on LinkedIn (Van de Ven et al., 2017), individuals high on narcissism smile in social situations (Back, Stopfer, et al., 2010). To clarify this association on LinkedIn, we examined genuine, wide smiles. Regarding intelligence, for example, we investigated whether it manifests in detailed descriptions of professional experiences, addressing conflicting results (Roulin & Levashina, 2019; Roulin & Stronach, 2022). (4) Research suggests that LinkedIn cues like the number of skills/interests on CEO profiles can indicate narcissism (Aabo & Eriksen, 2018; Cragun et al., 2020). We added such cues as potential narcissism indicators, though some cues' validity (e.g. selfreferential language use) is questionable (Carey et al., 2015). Analytical approach To get an impression of cue validities, we computed bivariate correlations of narcissism/intelligence with the respective LinkedIn cues. We winsorized cues (Wilcox, 2011) by setting extreme values |z| > 4.47 to ± 4.47. Winsorizing reduces extreme values' disproportionate influences while retaining their predictive value, reducing overfitting to the current sample, and ensuring generalizability to new samples. The threshold of 4.47 is based on Chebyshev's inequality (Saw et al., 1984), showing that at least 95% of the data fall within ± 4.47 standard deviations of the mean, regardless of the distribution. This limits winsorizing to extreme values representing no more than 5% of the data, usually much less. We used a machine learning approach (see Hastie et al., 2009; James et al., 2013; Lantz, 2019, for an introduction) to illuminate LinkedIn's potential for trait inferences. The discipline of machine learning seeks to ‘construct computer systems that automatically improve through experience’ (Jordan & Mitchell, 2015, p. 255): Machine learning algorithms are programs consisting of rule sets enabling them to automatically improve their performance on a task by learning generalizable patterns from training | 1579 PREDICTING INDIVIDUAL TRAITS FROM LINKEDIN data. The learned patterns can then be used to perform the task on new data. For example, these algorithms can learn associations between LinkedIn profile cues and traits from training data and then predict traits for new users based on their profiles. Multiple linear regression models perform poorly on prediction tasks as they typically overfit the training data (Zou & Hastie, 2005). Overfitting means that models learn spurious patterns from the training data that are not generalizable to new data. Multiple linear regressions optimize explained variation by minimizing the sum of the squared deviations of the predicted from the observed values of a target feature in the training data (i.e. insample). By construction, they adapt to the training data (and thus to its specificities and biases) to the point where the model predicts new data less well. This makes multiple linear regression unsuitable for prediction tasks. To address overfitting, we fitted elastic nets (Zou & Hastie, 2005) using the respective derived LinkedIn cues as independent variables (features)1 to predict the dependent variables narcissism/intelligence (target features). Elastic nets are regularized regressions that optimize prediction performance on new data (i.e. out- of- sample) by automatically selecting a penalty λ that shrinks those overfitted regression coefficients that hurt out- of- sample prediction accuracy towards zero (James et al., 2013). This regularization prevents overfitting and enhances generalizability. Automatic selection of λ to maximize out- of- sample prediction accuracy characterizes elastic nets as machine learning algorithms. The elastic net algorithm typically selects the optimal penalty λ using kfold crossvalidation (Hastie et al., 2009; James et al., 2013; Kohavi, 1995). The data are split into k (typically 10) equalsized parts, with nine parts constituting the insample training data to select the optimal penalty λ and the remaining part serving as the out- of- sample test data to estimate the prediction performance of the elastic net given the penalty λ. Specifically, elastic nets fit regressions with a fixed number of penalties λ (typically 100) to the training data. Then, each regression is used to predict the target feature based on the features using the test data. Having computed the chosen prediction performance metric (e.g. mean squared error) for each regression with its specific penalty, the elastic net selects the penalty λ for which the crossvalidation error (i.e. the chosen prediction performance metric on the test set) is smallest. This procedure is repeated k times so that each part of the data has served as test data once. This yields k elastic nets each with a different λ and out- of- sample prediction performance. Finally, the elastic net with the penalty λ that resulted in the smallest out- of- sample prediction error is fitted to all available data. This model can then be used to predict the target feature based on the features with which the elastic net was trained. Through crossvalidation, elastic nets learn the penalty λ that minimizes out- of- sample prediction error, enabling them to learn more generalizable association patterns than multiple linear regression models. By shrinking coefficients of less predictive cues towards zero, they sensitively use cues. Furthermore, they use cues consistently by fitting a generalizable model that is consistently applied to all LinkedIn users. This makes elastic nets highly sensitive and consistent, and thus, a useful algorithm for unravelling LinkedIn's predictive potential based on the lens model principles. Importantly, elastic nets maintain interpretability as the coefficients can be obtained from the final model and interpreted similar to multiple linear regression (Alexander et al., 2020). Combining prediction performance and interpretability made elastic nets the ideal algorithm for this study. Previous research has used kfold crossvalidation (Kohavi, 1995) to evaluate machine learning algorithms' prediction accuracy. However, this method leads to optimistically biased performance estimates as it combines hyperparameter tuning (e.g. optimize the penalty λ) and model selection based on the estimated prediction performance (Cawley & Talbot, 2010; Varma & Simon, 2006). By using the same data for hyperparameter tuning and estimating prediction accuracy, information from hyperparameter tuning leaks into the model selection (e.g. by selecting the penalty λ with the smallest out- of- sample prediction error as the final model). kfold crossvalidation thus allows algorithms to learn what model fits well to the test data, thereby inflating estimated prediction accuracy. 1We included gender (0/1 = female/male) and age as controls, as traits show consistent differences for gender/age (e.g. Nisbett et al., 2012; Weidmann et al., 2023), which may also affect LinkedIn cues. For example, men often list leadership skills, while women smile more (Fernandez et al., 2021). 1586 | HÄRTEL et al. Cue Measurement α/ICC [CI] naMSD rN [CI] rI [CI] Interests Interests Numerically counted (i.e. interests related to groups, influencers, companies, and schools) 406 40.92 65.94 −.01 [−.11, .09] −.04 [−.14, .05] Median of interests' followers Median of number of followers of interests related to groups, influencers, companies, and schools (assessing numbers of followers was limited to the first 50 interests in each category) 406 56,074 126,762 .16 [.06, .25] −.12 [−.21, −.02] Influencers Number of interests related to influencers divided by total number of interests 406 .03 .05 .07 [−.03, .16] −.02 [−.12, .08] Median of influencers' followers Median of number of followers of interests related to influencers (assessing numbers of followers was limited to the first 50 influencer interests) 406 1,311,217 4,007,637 .05 [−.05, .14] −.01 [−.11, .09] Other Sports activities 0 = not present; 1 = present (i.e. activities involving physical exertion (e.g. soccer captain, marathon)) 406 .03 .16 .09 [−.01, .19] .02 [−.08, .11] Cues potentially signalling intelligence (aggregate of fluid and crystallized intelligence) .60 [.54, .66]/.54 [.47, .60] 406 .00 .79 Profile picture Picture above neckline 0 = not present; 1 = present 381 .65 .48 −.01 [−.11, .09] .17 [.07, .27] Experience Extensive description Total number of words to describe professional positions divided by total number of professional positions 406 8.59 14.29 .02 [−.08, .12] .06 [−.04, .16] TABLE 1 (Continued) | 1587 PREDICTING INDIVIDUAL TRAITS FROM LINKEDIN Cue Measurement α/ICC [CI] naMSD rN [CI] rI [CI] Education Marks 0 = not present; 1 = present (e.g. marks as numerical number, percentile ranks, graduation with honours) 406 .27 .45 .01 [−.08, .11] .06 [−.03, .16] Averaged marks Mean of marks presented as numerical number in German grading system (1.0–4.0) 106 1.63 .45 −.14 [−.32, .06] −.17 [−.35, .03] Educational stations Numerically counted 406 2.64 1.48 .07 [−.03, .16] .14 [.05, .24] Extensive description Total number of words to describe educational stations divided by total number of educational stations 406 5.60 10.61 .00 [−.09, .10] .11 [.01, .20] Average duration Total duration of educational stations in years divided by total number of educational positions 406 3.58 1.61 −.05 [−.15, .05] −.03 [−.13, .06] Skills Industry knowledge Number of skills categorized as ‘industry knowledge’ skills divided by total number of skills 406 .17 .20 .09 [−.01, .18] .00 [−.09, .10] Tools and technologies Number of skills categorized as ‘tools and technologies’ skills divided by total number of skills 406 .13 .17 .07 [−.03, .16] .03 [−.07, .12] Accomplishments Honours/awards Number of accomplishments categorized as ‘honours and awards’ divided by total number of accomplishments 406 .06 .15 .05 [−.05, .14] .18 [.08, .27] Publications Number of accomplishments categorized as ‘publications’ divided by total number of accomplishments 406 .03 .10 .03 [−.07, .12] .15 [.05, .24] Projects Number of accomplishments categorized as ‘projects’ divided by total number of accomplishments 406 .02 .07 .05 [−.05, .14] .10 [.01, .20] Test scores Number of accomplishments categorized as ‘test scores’ divided by total number of accomplishments 406 .01 .03 −.02 [−.12, .07] .11[.02, .21] TABLE 1 (Continued) (Continues) 1588 | HÄRTEL et al. Cue Measurement α/ICC [CI] naMSD rN [CI] rI [CI] Interests Schools Number of interests related to schools divided by total number of interests 406 .11 .09 .00 [−.10, .09] .06 [−.03, .16] Median of schools' followers Median of number of followers of interests related to schools (assessing numbers of followers was limited to the first 50 school interests) 406 50,800 64,121 .12 [.02, .21] .19 [.09, .28] Groups Number of interests related to groups divided by total number of interests 406 .07 .10 −.03 [−.13, .07] .13 [.03, .22] Note: CI = 95% confidence interval. rN = bivariate correlation coefficient between grandiose narcissism as measured by the NARQ (Back et al., 2013) and the winsorized raw scores of the LinkedIn cues. rI = bivariate correlation coefficient between general intelligence, computed as the aggregation of zscores of fluid intelligence measured with the HMT- S (Heydasch et al., 2013) and crystallized intelligence measured with the BEFKI GC- K (Schipolowski et al., 2013), and the winsorized raw scores of the LinkedIn cues. aSome cues have missing values because certain information was not available on the LinkedIn profile, especially due to missing profile pictures and missing or incomplete sections and entries. Correlations in bold are significant at the p ≤ .05 level (twosided). TABLE 1 (Continued) | 1589 PREDICTING INDIVIDUAL TRAITS FROM LINKEDIN Machine learning approach Table 2 shows the performance estimates of the nested crossvalidated elastic nets predicting narcissism/intelligence based on the respective derived LinkedIn cues. The elastic nets outperformed multiple linear regressions and interceptonly models in absolute prediction performance, with lower MSE, RMSE, and MAE. This indicates that elastic nets made smaller errors in predicting narcissism/ intelligence. Interestingly, interceptonly models sometimes outperformed multiple linear regressions, suggesting the latter may overfit the training data. The elastic nets also showed better relative prediction performance, with correlations between predicted and observed target feature values of r = .35/.41 for narcissism/intelligence. These correlations were higher than those of the multiple linear regressions (r = .27/.34 for narcissism/intelligence). Overall, the results indicate that narcissism/intelligence can be predicted from LinkedIn profiles with accuracy comparable to metaanalytical findings on automated trait inferences from online social network information (.29 ≤ r ≤ .40; Azucar et al., 2018; Settanni et al., 2018). We found that the LinkedIn cues driving prediction in the elastic nets (Table 3) were generally similar to those identified in the correlation approach, indicating relatively stable cue validities when controlling for other cues and using resampling. All 10 cues correlated with narcissism were important elastic net predictors (βM ≥ |.017|, C V FI ≥ 9). The five most important elastic net predictors were all significantly correlated with narcissism and indicate that individuals high on narcissism display (a) interests with more followers, (b) less smiling on profile pictures, (c) more accomplishments related to organizations, (d) background pictures, and (e) public speaking skills. Eleven of the 15 LinkedIn cues correlated with intelligence were important elastic net predictors (βM ≥ |.017|, CVFI ≥ 7). Four cues (extensive description of educational stations, number of volunteer experiences, profile length, and English profile) correlated with intelligence but were not important elastic net predictors (βM ≤ |.0 03|, C V FI ≤ 3). The five most important elastic net predictors were all significantly correlated with intelligence and indicate that individuals high on intelligence display (a) schools with more followers, (b) profile pictures above neckline, (c) more accomplishments, (d) less dressedup, trimmed appearances on profile pictures, and (e) more accomplishments related to honours/awards. TABLE 2 Performance and hyperparameter estimates of nested crossvalidated elastic nets, multiple linear regressions, and interceptonly models predicting narcissism and intelligence. Narcissism Intelligence EN MLR IOM EN MLR IOM MSD MSD MSD MSD MSD MSD MSE .905 .081 1.032 .145 .975 .001 .892 .092 .973 .179 .975 .001 RMSE .950 .042 1.014 .070 .988 .000 .943 .049 .982 .091 .988 .000 MAE .760 .054 .818 .062 .793 .041 .744 .057 .780 .082 .782 .035 r.35 .18 .27 .16 .41 .18 .34 .19 R2.07 .01 .28 .07 .10 .03 .25 .06 R2 Adj. .02 .01 .15 .08 .05 .03 .14 .07 λMin .100 .016 .087 .017 Note: All performance measures are computed based on the observed and predicted values on the corresponding target features (i.e. narcissism and intelligence). λMin = penalty selected by the elastic net. The means and standard deviations of the performance measures and λMin are estimated based on the best performing models from the inner folds that are tested on the test data from the outer folds of the tenfold nested crossvalidation approach. Abbreviations: EN, elastic net; IOM, interceptonly model; MAE, mean absolute error; MLR, multiple linear regression; MSE, mean squared error; r, bivariate correlation coefficient; R2, explained variation; R2 Adj., adjusted explained variation; RMSE, root mean squared error. 1590 | HÄRTEL et al. TABLE 3 Regression coefficients of nested crossvalidated elastic nets predicting narcissism and intelligence. Cue CVFI βMβSD βFull Narcissism Interests: Median of interests' followers 10 .098 .016 .106 Profile picture: Smiling 10 −.096 .014 −.105 Gender (0/1 = female/male) 10 .089 .016 .094 Accomplishments: Organizations 10 .077 .014 .085 Additional pictures: Background picture 10 .069 .014 .074 Skills: Public speaking 9 .067 .027 .074 Volunteering: Altruistic volunteering 10 .045 .018 .051 Profile picture 10 −.042 .021 −.051 Additional pictures: Additional pictures/videos 10 −.037 .020 −.056 Profile picture: Dressedup, trimmed appearance 9 .033 .016 .039 Other: Sport activities 10 .031 .013 .040 Recommendations: Received recommendations 9.030 .023 .038 Experience: Leadership positions 10 .021 .014 .029 Other: Profile in English 9 .019 .014 .026 Skills 9.017 .014 .019 Profile picture: Professional shot 6 .009 .013 .011 Education: Business studies 5.008 .011 .003 Accomplishments: Courses 9.007 .009 .013 Interests 3−.007 .014 −.015 Profile picture: Stylish/flashy/fashionable appearance 2.003 .008 .007 About: Extensive About section 2 −.002 .006 .000 Skills: Endorsements 1−.001 .004 .000 Profile card: Name with title 1 .000 .001 .000 Intercept 10 .000 .000 .000 Intelligence Interests: Median of schools' followers 10 .105 .012 .103 Profile picture: Picture above neckline 10 .091 .020 .093 Gender (0/1 = female/male) 10 .090 .024 .090 Accomplishments 10 .080 .020 .084 Profile picture: Dressedup, trimmed appearance 10 −.075 .015 −.077 Accomplishments: Honours/awards 10 .058 .015 .059 Interests: Groups 10 .056 .019 .057 Accomplishments: Publications 10 .054 .022 .054 Accomplishments: Test scores 10 .045 .010 .045 Education: Averaged marks 9−.043 .024 −.040 Volunteering: Average duration 10 .040 .013 .042 Skills: Leadership 9−.033 .020 −.032 Education: Educational stations 7.019 .021 .019 Profile picture: Charming facial expression 5 −.019 .026 −.009 Accomplishments: Projects 8.017 .015 .019 Licences and certifications 6−.013 .015 −.004 Age 5.011 .015 .000 | 1591 PREDICTING INDIVIDUAL TRAITS FROM LINKEDIN DISCUSSION This study uncovers LinkedIn's predictive potential to infer traits. Building on the lens model (Brunswik, 1956), we show that LinkedIn profiles contain a valid cue base signalling users' narcissism (e.g. background pictures, public speaking skills) and intelligence (e.g. more accomplishments, schools with more followers), meeting the necessary condition for accurate inferences. Sensitively and consistently utilizing these valid cues, the automated perceiver attained prediction accuracy (r = .35/.41 for narcissism/intelligence). This shows that when the sufficient condition for accurate trait inferences is ensured to be met, LinkedInbased inferences can reach accuracy levels comparable to automated trait inferences from online social networks (Azucar et al., 2018; Settanni et al., 2018) and at the upper bound of recruiters' LinkedInbased trait inferences (e.g. Roulin & Levashina, 2019; Van de Ven et al., 2017). Applying the lens model to (automated) trait inferences based on LinkedIn profiles This study underlines the utility of extending the lens model to automated perceivers to systematically examine trait inference processes (Cannata et al., 2022; Tay et al., 2020): Machine learning algorithms are trained to use valid cues sensitively and consistently, and thus turned out as a useful vehicle to demonstrate that LinkedIn profiles possess predictive potential. Machine learning thus offers a promising alternative methodological approach to lens model studies, avoiding overfitting and prioritizing prediction over explanation, which suits the nature of trait inferences (Yarkoni & Westfall, 2017). Resampling ensures robust results, reducing contradictory conclusions on cue validities. Algorithms like elastic nets effectively handle many intercorrelated cues common in most information bases (James et al., 2013) and feature importance provides practical insights for cue weighting. The automated approach also helps understand the modest accuracy of human perceivers' LinkedInbased trait inferences (Roulin & Levashina, 2019; Roulin & Stronach, 2022; Van de Ven et al., 2017). According to the lens model, this may result from (a) a lack of a valid cue base or (b) a lack of a sensitive and consistent use of valid cues. While Roulin and Stronach (2022) challenged the existence of a valid Cue CVFI βMβSD βFull Profile picture: Professional shot 4 .010 .021 .000 Profile picture 3.009 .016 .000 Interests: Schools 3.007 .011 .000 Education: Marks 3.004 .009 .000 Volunteering: Volunteer experiences 1.003 .011 .000 Skills: Industry knowledge 1−.003 .009 .000 Experience: Leadership positions 3.003 .005 .000 Other: Profile length 3 .003 .005 .000 Education: Extensive description 1.002 .006 .000 Education: Business studies 2−.002 .003 .000 Education: Average duration 1−.001 .002 .000 Recommendations: Received recommendations 1.001 .002 .000 Other: Profile in English 1 .000 .000 .000 Intercept 10 .000 .000 .000 Note: CVFI = crossvalidation fold incidence, that is, the number of outer folds the regression coefficient of a feature was ≠ 0; βM = regression coefficients averaged across outer folds; βSD = standard deviation of regression coefficients across outer folds; βFull = regression coefficients of elastic net trained on full data. Only cues are shown for that CVFI > 0. Cues sorted by |βM|. All values on zscale. TABLE 3 (Continued) 1592 | HÄRTEL et al. LinkedIn cue base based on a small exploratory cue set, Fernandez et al. (2021) found support for it based on deductively derived cues. Deriving an extensive cue set from theory and empirical findings, we identified a rich set of valid LinkedIn cues for narcissism/intelligence as robust predictors in nested crossvalidated models. As such, we found that the automated perceiver, designed to sensitively and consistently use these valid cues, achieved notable prediction accuracy. This suggests that recruiters' lack of accuracy may stem from insensitive and inconsistent use of valid cues, rather than LinkedIn's limited predictive capability. Indeed, comparing the cues used by recruiters in prior research with those used by our automated perceiver model reveals human perceivers' lack of sensitivity. For example, recruiters were found to use the number of subscribers to infer trait selfpresentation (Van de Ven et al., 2017), but this did not signal narcissism in our study. Thus, the sensitive automated perceiver did not use this information. Vice versa, valid cues like ‘profile in English’ used by the automated perceiver were not used accordingly by humans (Van de Ven et al., 2017). Human perceivers are also inconsistent in applying judgement rules, whereas statistical models remove these inconsistencies, typically leading to more accurate inferences (Karelaia & Hogarth, 2008). Extending previous research on valid cues signalling traits on LinkedIn This study advances the field of trait expression in online professional networks by identifying 10/11 easy- to- interpret LinkedIn cues correlated with narcissism/intelligence tested to be robust in nested crossvalidated models. This addresses calls to examine online network information signalling important applicant characteristics (Schroeder et al., 2020). We also enrich the broader literature on trait expression in selection contexts; adding LinkedIn cues to the information bases recruiters can use to infer traits, like resumés (Burns et al., 2014; Härtel et al., 2024), application photos (Fernandez et al., 2017), online social networks (Back, Stopfer, et al., 2010; Stopfer et al., 2014), job interviews (DeGroot & Gooty, 2009), or even handshakes (Stewart et al., 2008). LinkedInspecific cues, like listing interests with many followers for narcissism or schools with many followers for intelligence, offer unique insights unavailable in other sources. Our findings also reveal that some cues related to trait expression in other contexts lack predictive power on LinkedIn. For example, narcissism often manifests in physical appearance cues (e.g. professional/formal attire; Back, Schmukle, et al., 2010; Vazire et al., 2008) and indicators of online socializing behaviour (e.g. more contacts/frequent posting; Gnambs & Appel, 2018; McCain & Campbell, 2018), but less so on LinkedIn. LinkedIn's professional norms – maintaining a formal appearance, building a professional network, and posting business content – may render these cues less discriminative (Hartwell & Campion, 2020). Similarly, whereas narcissism often involves smiling in everyday situations (Back, Schmukle, et al., 2010), consistent with Van de Ven et al. (2017), we observed less smiling to be a narcissism predictor. LinkedIn's norms may prompt most users to smile, masking narcissists' agentic smiling tendencies. This could create space for the expression of antagonistic narcissism, resulting in arrogant expressions with fewer genuine smiles. Indeed, less smiling was a strong predictor of narcissistic rivalry (βM = −.106), not admiration (βM = −.011; Supplemental Results Section E in Appendix S1, ht t ps:// osf. io/ 4ruqj/ ). Overall, this highlights the role of situational strength in trait expression (Meyer et al., 2010). Automated approach to trait inferences based on LinkedIn This study advances the literature on automated trait assessments based on online networks by applying the automated approach to LinkedIn using a broad set of theoretically and empirically grounded cues (Bleidorn & Hopwood, 2019; Tay et al., 2020). This way, we obtained high predictivity and the capability to provide explanations for our findings (Yarkoni & Westfall, 2017), both critical for practical applications. The machine learning models' accuracies predicting narcissism/ | 1593 PREDICTING INDIVIDUAL TRAITS FROM LINKEDIN intelligence were r = .35/.41, translating to hit rates of 67.5%/70.5%, indicated by the binomial effect size display (Rosenthal, 2014; Rosenthal & Rubin, 1982). This means individuals testing above average narcissistic/intelligent are correctly identified about 67.5%/70.5% of instances. In absolute terms, these effect sizes are medium by conventional standards (Cohen, 1988) and (very) large by modern standards (Funder & Ozer, 2019; Gignac & Szodorai, 2016). On a relative basis, these accuracy levels exceed work colleagues' accuracy of trait inferences (r = .27; Connelly & Ones, 2010; Youyou et al., 2015) and are comparable to the agreement between self- and supervisor ratings of job performance (r = .35; Harris & Schaubroeck, 1988). However, we emphasize not to overstate the validity of automated LinkedInbased inferences. LinkedIn profiles provide limited insight into a snippet of individuals' traits – they reflect an external view of the publicly revealed professional identity and cannot replace wellestablished psychometric measures. Figure 3 compares the prediction accuracies from our study with those of human and automated perceivers drawing from similar sources like resumés, social media profiles, and LinkedIn profiles. Our accuracies (rNarcissism = .35/rIntelligence = .41) are comparable to metaanalytical results of automated approaches predicting similar traits from social media information like Facebook (r = .40/.29; Azucar et al., 2018; Settanni et al., 2018). These metaanalyses describe the relationship between digital traces and traits as ‘quite strong’ (Azucar et al., 2018, p. 157) and the accuracy level as ‘remarkable’ (Settanni et al., 2018, p. 217). Given LinkedIn's contextually constrained nature focusing on professional identity, the relative accuracy of automated LinkedInbased trait inferences is notable. Comparing our results with human perceivers' accuracy using similar information, the automated perceiver appears to perform better.3 Also note that reported accuracy levels for human perceivers often represent averaged inferences from groups of perceivers (aggregated perceiver approach; Back & Nestler, 2016), with 10 perceivers in Van de Ven et al. (2017), two- to- four in Roulin and Levashina (2019), and six or more in Roulin and Stronach (2022). Single perceivers yield substantially lower accuracy (Back & Nestler, 2016; Van Iddekinge et al., 2016) due to poorer reliability (De Vet et al., 2017). Therefore, automated LinkedInbased trait inferences appear to outperform single recruiters' cybervetting accuracy as commonly practiced in many organizations. Practical implications Recruiters believe they can accurately infer applicants' traits through online profiles (Van Iddekinge et al., 2016). Our study shows that there is some substance to this assumption – LinkedIn profiles possess predictive potential if the contained valid cues are used sensitively and consistently. Yet, although recruiters routinely infer traits based on LinkedIn, they lack accuracy, which can lead to erroneous decisions, like falsely rejecting applicants due to their ‘not suitable’ personality. Training recruiters to use valid LinkedIn cues sensitively and consistently could improve accuracy (Cole et al., 2005; Karelaia & Hogarth, 2008; Powell & Bourdage, 2016). The machine learning approach identifies which LinkedIn information robustly signal traits, providing a foundation for such training. Going further, a hybrid scoring approach combining automated/human perceivers could offer additional value (Cannata et al., 2022). Automated perceivers could continuously scan LinkedIn users to identify promising candidates not actively job searching and screen large applicant pools quickly, allowing recruiters to focus on candidates with desired traits. Human recruiters may then consider subtle nuances and cue interactions (Cole et al., 2007; Knouse, 1994) missed by automated systems. A hybrid system may be more acceptable to applicants, often reacting negatively to decisions made solely by algorithms (Gonzalez et al., 2022). 3Comparing the automated perceiver's accuracy with previous findings on human perceivers might be seen as favouring the former, as our automated perceiver was exposed only to preselected, potentially valid cues. Note that providing the automated perceiver with potentially valid cues was a pragmatic decision to give it access to meaningful LinkedIn information that was timeconsuming to code manually. In contrast, human perceivers with their rich perception have access to much more LinkedIn information. Even if all this additional information was invalid, it would not substantially harm automated perceivers' accuracy due to their sensitivity. Accordingly, preselection of cues likely underestimates rather than overestimates automated perceivers' accuracy. 1594 | HÄRTEL et al. Besides, some may be inclined to interpret our findings suggesting that LinkedInbased trait inferences should be delegated to automated perceivers due to their accuracy and resource savings. Indeed, automated trait inferences could offer a less intrusive, costeffective, and scalable supplement to traditional psychological assessments (Morgeson et al., 2007). Yet, this is too early to call. Although our study provides robust results by applying resampling, it remains a single study with a mediumsized sample of 406 LinkedIn users. Also, automated LinkedInbased assessments pose practical and fairness challenges, so further research is needed before endorsing them. As such, LinkedIn use is dynamic, and cues may evolve over time, for instance, as users become acquainted with the platform, they may upload background pictures more frequently. Also, applicants already tailor their resumés for favourable algorithmic scoring (Caprino, 2019; Weed, 2021) and will likely adjust their LinkedIn profiles once automated assessments become common. These dynamic processes demand continuous crossvalidation and algorithmic recalibration, necessitating efficient, automated cue collection processes. Yet, we coded LinkedIn cues manually, which was timeconsuming and labourintensive. The imminent shift towards automated cue extraction – initially demonstrated by studies leveraging web scraping for textual data (Landers et al., 2016; Youyou et al., 2015) and deep learning for image data (Wei & Stillwell, 2017) – is gaining momentum in rapid developments of large language and computer vision models. This will likely diminish manual coding. Generative pretrained transformers can parse textual information to evaluate selfpromotional content in job descriptions or categorize skills, and convolutional neural networks can scrutinize profile images to discern facial expressions, poses, or attire choices. Another concern involves subgroups (dis- )favoured by the algorithm. The algorithm is programmed to minimize error when predicting test scores, which may show subgroup differences for attributes like gender (Kheloui et al., 2023) and ethnicity (Roth et al., 2001). This carries the risk that the algorithm learns the ‘predictive value’ of subgroup differences. In our study, for example, men FIGURE 3 Prediction accuracy in the present study compared to human and automated perceivers' accuracy reported in prior research. Note. Prediction accuracy is defined as the correlation (r) between observed trait test scores and inferred trait scores based on the respective information base. aWhenever available, we consulted results directly referring to narcissism and otherwise drew on related constructs, namely need for popularity, trait selfpresentation, honestyhumility, and extraversion. bAveraged correlation for I- O psychology students and hiring professionals as raters. cThe correlation between cognitive ability test scores and aggregated LinkedInbased cognitive ability ratings by the four perceivers across the two time points was much lower (r = −.02, see Table S2). | 1595 PREDICTING INDIVIDUAL TRAITS FROM LINKEDIN scored higher than women on intelligence (ΔM = .27, t(345.32) = 3.36, p < .001), leading the algorithm to use gender to predict intelligence. Using the elastic net, an interpretable algorithm revealing the cues it uses, discrimination can be addressed by removing biased cues from the model or adjusting their regression weights. Alternatively, subgroup differences can be mitigated in the training data through fair test construction or statistical corrections (Hough et al., 2001). Implementing such measures may then even hold potentials to reduce discrimination, as automated systems are principally immune to (unconscious) human biases (Mönke et al., 2024; Zhang et al., 2020). It is also crucial to test whether the algorithm exhibits varying accuracy across subgroups, constituting another form of discrimination. Limitations and directions for future research Generalizing our findings to the LinkedIn population depends on the representativeness of our sample, which was not restricted to subpopulations like students/recent graduates (cf. Fernandez et al., 2021; Roulin & Levashina, 2019), but meant to reflect LinkedIn's mixed professional population. Indeed, a comparison of age distributions between the LinkedIn population and our sample indicates reasonable alignment.4 Yet, our sample is not representative in several respects. For example, we examined Germanspeaking LinkedIn users. Cultural differences shape motivations and selfpresentation tactics related to using online networks (Jackson & Wang, 2013; Rui & Stefanone, 2013) and may thus affect trait expression and LinkedIn's predictive potential. For instance, whereas uploading background pictures may be a valid narcissism indicator in Germany, it may be a more commonly used feature in other countries, providing less insight into narcissism levels. Also, in creative industries like web design or advertising, it may merely serve as a portfolio. Future research should confirm prediction accuracies and cue validities in larger samples mirroring the LinkedIn population across dimensions like age, gender, education, job roles, and industries. We optimized the automated perceiver to predict trait test scores. The ‘goldstandard’ would be aggregating various trait indicators as prediction criteria, like selfreports, informant reports, and behavioural assessments (Cannata et al., 2022). Further steps preceding practical applications are then examining the extent automated trait inferences predict work outcomes, particularly job performance (criterionrelated validity; Cubrich et al., 2021; Kluemper et al., 2012), and explain incremental variance beyond (i) recruiters' LinkedIn inferred traits and (ii) established tests. More research is needed on adverse impact (Hunkenschroer & Luetge, 2022) and applicant reactions (Oostrom et al., 2023). Until then, organizations should avoid strongly relying on automated LinkedInbased trait inferences. Our findings suggest that automated perceivers infer traits on LinkedIn more accurately than single recruiters, who appear to be less sensitive and/or consistent in using valid cues. Yet, evidence on recruiters' cue utilizations is limited (Roulin & Stronach, 2022; Van de Ven et al., 2017), and our study lacks a direct comparison, warranting caution in drawing final conclusions. Initial evidence suggests that recruiters may indeed use nonvalid cues and overlook valid ones, indicating a lack of sensitivity. Also, human perceivers are typically inconsistent in applying judgement rules. Future research should compare automated and human perceivers directly, disentangling whether recruiters can distinguish valid from nonvalid LinkedIn cues and apply consistent judgement rules. Such research could inform training programs educating about rarely used valid cues and how to use them consistently. It would be fascinating to see if the accuracy of human inferences could be improved through such measures, and whether they could even surpass automated perceivers by capturing subtle idiosyncratic information. 4Percentage of age groups in the LinkedIn population (Dixon, 2023) versus our sample: 18–24 years, 22% versus 34%; 25–34 years, 60% versus 48%; 35–54 years, both 15%; ≥ 55 years, both 3%. Although our sample skews slightly younger on average (Mage = 29), it seems to align more closely with the LinkedIn population compared to prior studies (e.g. Mage = 21/41 in Roulin & Levashina, 2019/Roulin & Stronach, 2022). 1602 | HÄRTEL et al. Youyou, W., Kosinski, M., & Stillwell, D. (2015). Computerbased personality judgments are more accurate than those made by humans. Proceedings of the National Academy of Sciences of the United States of America, 112(4), 1036–1040. https:// doi. org/ 10. 1073/ pnas. 14186 80112 Zhang, L., Van Iddekinge, C. H., Arnold, J. D., Roth, P. L., Lievens, F., Lanivich, S. E., & Jordan, S. L. (2020). What's on job seekers' social media sites? A content analysis and effects of structure on recruiter judgments and predictive validity. Journal of Applied Psycholog y, 105(12), 1530–1546. https:// doi. org/ 10. 1037/ apl00 00490 Zibarras, L. D., & Woods, S. A. (2010). A survey of UK selection practices across different organization sizes and industry sectors. Journal of Occupational and Organizational Psycholog y, 83(2), 499–511. https:// doi. org/ 10. 1348/ 09631 7909X 425203 Ziegler, M., Danay, E., Heene, M., Asendorpf, J., & Bühner, M. (2012). Openness, fluid intelligence, and crystallized intelligence: Toward an integrative model. Journal of Research in Personality, 46(2), 173–183. https:// doi. org/ 10. 1016/j. jrp. 2012. 01. 002 Zou, H., & Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society: Series B (Statistical Methodolog y), 67(2), 301–320. https:// doi. org/ 10. 1111/j. 1467- 9868. 2005. 00503. x SUPPORTING INFORMATION Additional supporting information can be found online in the Supporting Information section at the end of this article. Appendix S1 How to cite this article: Härtel, T. M., Schuler, B. A., & Back, M. D. (2024). ‘LinkedIn, LinkedIn on the screen, who is the greatest and smartest ever seen?’: A machine learning approach using valid LinkedIn cues to predict narcissism and intelligence. Journal of Occupational and Organizational Psychology, 97, 1572–1602. https://doi.org/10.1111/joop.12531