Peter Soyster, PhD and Leighann Ashlock, PhD
The factors that motivate or drive behavior (in a functional analytic/behavioral chain analysis- sense) are often different for different people. The exact situational antecedents that might increase the probability of a behavior for one person may decrease the probability for another. While this is not a particularly groundbreaking assertion for mental health clinicians, it remains that the majority of primary research provided to clinicians (and on which evidence-based practice is based) is not conducted in a way that allows for individualized analyses of quantitative data (Fisher et al., 2018; Molenaar, 2004). Put simply, routine psychological science is not analyzing data in a way that would likely be most helpful to clinicians.
Given the degree of heterogeneity in behavioral expression that coincides with almost all current DSM-5 disorder categories, deepening our understanding of intraindividual differences through the behavioral level of analysis is a promising path forward in personalizing the treatment delivery of psychopathology (Fisher et al., 2019). Despite this promise, implementation of personalized interventions has been limited, and remains unwieldy to most clinicians (Rodebaugh, Frumkin, & Piccirillo 2020).
Idiographic and Nomothetic
Successful personalization likely requires data at the level of the individual, as opposed to the larger population. Generally, psychology has approached understanding psychopathology through nomothetic research. That is, deriving an understanding of what is mainly true of a population by investigating interindividual variation at the group level. In clinical practice, this has led to the therapist’s dilemma – where treatment “best practices”
have been established from group-level research, as opposed to clinicians selecting treatment based on their own client (Levine et al., 1992).
The aim of idiographic research is to understand what is normally true of a single person by investigating intraindividual variation. Often this is achieved by collecting data from each person, anchored in time as they go throughout their daily life (e.g., ecological momentary assessment). Idiographic and nomothetic approaches are not antagonistic to each other but instead provide unique information that is differentially useful depending on context. We propose that person-specific models that reliably predict symptoms or behaviors before they occur could resolve the therapist’s dilemma in the context of the clinic, improving intervention and treatment efficacy.
Recently, we published a paper in the Psychology of Addictive Behaviors that applied statistical classification methods to idiographic time series data to identify person-specific predictors of future alcohol drinking-relevant behavior, affect, and cognitions in a college student sample (Soyster et al., 2021).
We collected time series data from college students who had endorsed consuming alcohol at some point in their lifetime. They were sent eight mobile phone surveys per day for 15 days. Each survey assessed the number of drinks consumed since the previous survey, as well as affect, alcohol craving, drinking expectancies, perceived alcohol consumption norms, impulsivity, and social and environmental context.
Each individual’s data were split into training and testing sets, so that trained models could be validated using person-specific out-of-sample data. Elastic net regularization, a machine learning technique, was used to select a set of 40 variables to then be used to predict alcohol consumption, craving, or wanting to drink two hours into the future per person. Averaging across participants, results showed that accurate out-of-sample predictions of future drinking were made 76% of the time. Meaning, these person-specific models were able to classify an average of 76% of time points per participant as either drinking or non-drinking events about two hours before they occurred. For craving alcohol, the mean out-of-sample R² value was 0.27. For wanting to drink, the mean out-of-sample R² value was 0.27.
We also tested the relative accuracy of nomothetic predictions. The pooled elastic net regression model showed accurate out-of-sample predictions of future drinking 78% of the time. For craving alcohol, the mean out-of-sample R² value was 0.31. For wanting to drink, the mean out-of-sample R² value was 0.33. In sum, the pooled data, using a single set of coefficients, were able to predict future outcomes as well as or better than the more idiosyncratic coefficients.
Although we did not observe a significant difference between the pooled and person-specific approaches in the prediction of future drinking, craving alcohol, and wanting alcohol, we argue that using person-level data is the most practical application of the present methodology within a clinical setting. A single clinician could use these methods with any client without needing to rely on data from others in order to achieve commensurate prediction accuracy. Such accuracy could be used to identify personalized optimal periods for the delivery of intervention materials (e.g., just-in-time interventions).
Overall, these models could be applied to ongoing monitoring and early warning systems, bolstering and refining the delivery of intervention materials. Based on each client’s data, these methods could be used to identify when an intervention is needed and which intervention components would likely be most effective. However, the question remains, when should clinicians opt for individualized models over nomothetic ones (or vice-versa)? This is a challenging and open question.
Many currently available ‘personalized’ insights into the factors driving substance use are based on group-level characteristics of the client (e.g., here is what motivates cannabis consumption in sexual and gender minority communities). In our opinion (see: going beyond what we can currently show definitively with data), we find an assumption of mechanistic homogeneity based on socially-defined demographic characteristics to be disconnected from the experience of many clinicians.
One of the first things researchers learn as they start taking statistics training is that the mean value of a variable is the “the best guess we can make about the actual value of a variable at any given instance.” However, many of us seem to have forgotten the prefix of this axiom; “in the absence of additional information…”. Clinicians are experts in getting ‘additional information from their clients. We argue that idiographic approaches to behavioral analysis provide clinicians with the opportunity to reliably analyze this information to inform their clinical practice.