Psychological Networks in Clinical Populations: A tutorial on the consequences of Berkson's Bias

Abstract

<p>In clinical research, populations are often selected on the sum-score of diagnostic criteria such as symptoms. Estimating statistical models where a subset of the data is selected based on a function of the analyzed variables introduces Berkson’s bias, which presents a potential threat to the validity of findings in the clinical literature. The aim of the present paper is to investigate the effect of Berkson’s bias on the performance of the two most commonly used psychological network models: the Gaussian Graphical Model (GGM) for continuous and ordinal data, and the Ising Model for binary data. In two simulation studies, we test how well the two models recover a true network structure when estimation is based on a subset of the data typically seen in clinical studies. The network is based on a dataset of 2,807 patients diagnosed with major depression, and nodes in the network are items from the Hamilton Rating Scale for Depression (HRSD). The simulation studies test different scenarios by varying (1) sample size and (2) the cut-off value of the sum-score which governs the selection of participants. The results of both studies indicate that higher cut-off values are associated with worse recovery of the network structure. As expected from the Berkson’s bias literature, selection reduced recovery rates by inducing negative connections between the items. Our findings provide evidence that Berkson’s bias is a considerable and underappreciated problem in the clinical network literature. Furthermore, we discuss potential solutions to circumvent Berkson’s bias and their pitfalls.</p>

Obsessive–compulsive symptoms in eating disorders: A network investigation

“…The nonthreshold sensitivity analyses yielded similar results (see Supplement). Although De Ron et al () have rightly alerted network researchers about the risks of the distorting effects of Berkson's Bias, our networks did not appreciably differ as a function of having versus not having a symptom severity threshold as an inclusion criterion.…”

“…To ensure clinically relevant psychopathology, we excluded 160 individuals (11%) who scored below a clinical cutoff on the eating disorder examination questionnaire (EDE‐Q) version 6.0 (Fairburn & Beglin, ; threshold of two) and 150 (10%) who scored below a clinical cutoff on the Yale–Brown obsessive compulsive scale—self‐report (Y‐BOCS‐SR; Steketee, Frost, & Bogart, ; threshold of 16). Because using a symptom severity threshold as an inclusion criterion can alter network structure via Berkson's Bias (De Ron, Fried, & Epskamp, ; Rohrer, ), we also analyzed the data without applying an EDE‐Q threshold (n = 370; see Supplement) as a sensitivity analysis.…”

“…Because using a symptom severity threshold as an inclusion criterion can alter network structure via Berkson's Bias (De Ron, Fried, & Epskamp, 2019;Rohrer, 2018), we also analyzed the data without applying an EDE-Q threshold (n = 370; see Supplement) as a sensitivity analysis.…”

The replicability and generalizability of internalizing symptom networks across five samples.

“…This study had numerous strengths, including the inclusion of five samples and the examination of replicability within samples and across samples from the same population as well as generalizability to samples from different populations. Recent work suggests Berkson's Bias, which occurs when relationships in a subpopulation (e.g., a clinical sample for which a clinical severity cutoff was an inclusion criterion) differ from those in the general population, is a concern when interpreting networks estimated from clinical samples (de Ron, Fried, & Epskamp, 2019). The multi-sample approach and comparison of clinical and nonclinical samples allowed us to detect the extent to which Berkson's Bias occurred, and this may explain why coefficients of similarity and correlations were stronger between nonclinical networks than between the clinical network and nonclinical networks.…”

Psychometric network models from time-series and panel data

“…At the partial correlation levels, these weak marginal correlations lead to stronger partial correlations of an unexpected sign because the correlation is weaker than can be expected due to the links between positive/negative affect and third variables, such as self-esteem and optimism. It may even be that these variables act as a common effect between positive and negative affect (De Ron, Fried, & Epskamp, 2019;Epskamp, Waldorp, et al, 2018), in which case a strong partial correlation of an unexpected sign may also be expected. To check if the results were influenced by the subset of indicators chosen to assess positive affect and negative affect, I also estimated a panel-lvgvar model for only positive and negative affect using all indicators.…”