Activity space during treatment with medication for opioid use disorder: Relationships with personality, mood, and drug use

Abstract

Introduction:

Activity space in people with substance use disorders (SUDs) has been assessed for theoretical reasons and for detection/prevention of relapse. In this observational study, we relate passively obtained activity space measures to mental states and behaviors relevant to the success of treatment for opioid use disorder. Our long-term goal is to use such data to assess risk in real time and to recognize when SUD patients might benefit from a just-in-time intervention.

Methods:

We used GPS data from 238 urban residents in the first 16 weeks of stabilization on medication for opioid use disorder to test preregistered hypotheses about activity space (distance traveled, number of locations, time spent moving, and psychosocial-hazard levels of neighborhoods where participants spent time) in relation to certain static variables (personality, mood propensities) and time-varying treatment-relevant behaviors such as craving and use of opioids and cocaine.

Results:

The most consistent findings were that 1) mobility decreased over the course of the study; 2) neuroticism was associated with overall lower mobility; 3) trait-like positive mood (averaged from momentary ratings) was associated with higher mobility; 4) participants who used cocaine more frequently had lower mobility; 5) early in treatment, participants spent less time moving (i.e., were more sedentary) on days when they were craving. Some of these findings were in the expected direction (i.e., the ones involving neuroticism and positive mood), and some were opposite to the expected direction (i.e., we expected cocaine use to be associated with higher mobility); others (e.g., changes in mobility over time or in relation to craving) involved nondirectional hypotheses.

Conclusions:

Real-time information that patients actively provide is valuable for assessing their current state, but providing this information can be burdensome. The current results indicate that certain static or passively obtained data (personality variables and GPS-derived mobility information) are relevant to time-varying, treatment-relevant mental states and drug-related behavior, and therefore might be useful when incorporated into algorithms for detecting need for intervention in real time. Further research should assess how population-specific these relationships are, and how these passive measures can best be combined with low temporal-density, actively-provided data to obtain valid, reliable assessments with minimal burden.

1. Introduction

One core concept in behavioral geography is activity space, the set of locations through which people move in their daily lives. The term originated in the late 1960s, in studies of residential migration, as an outgrowth of action space, which was defined as the set of locations that people have at least secondhand knowledge about and would consider residing in (Adams, 1968; Wolpert, 1965). Action space was conceptualized as a subset of all locations on earth. Activity space was the subset of a person’s directly experienced action space, especially the subset encountered in a typical day (Brown & Moore, 1970; Horton & Reynolds, 1970). Current usage of the term activity space usually casts it not as a subset of a larger space, but as an extension of what might otherwise be static measures, such as place of residence, employment, or recreation—the logic being that activity space, on any given time scale, reflects people’s experiences more fully than places of residence or work (Jones & Pebley, 2014; Perchoux et al., 2013).

Activity space can be assessed retrospectively by questionnaires, interviews, or diaries (e.g., Chaix et al., 2012; Inagami et al., 2007; Mason & Korpela, 2009; Shareck et al., 2014; Vallee et al., 2011; Wong & Shaw, 2011) or in real time by tracking with global-positioning-systems (GPS) (e.g., DaViera et al., 2020; Duncan et al., 2021; Heller et al., 2020; Kim et al., 2019; Saeb et al., 2017; Saeb et al., 2015; Wiehe et al., 2008). Once assessed, it can be conceptualized and analyzed via two broad approaches.

In one approach, activity space is expressed in terms of inferred exposures. The regions covered by participants’ tracks are coded for their demographic or other features. The analysis proceeds largely on the basis of types of environments encountered. This approach, which partly abstracts the spatial element out of the data, has shown how momentary exposures can be associated with health and behavioral measures, above and beyond place of residence (e.g., Duncan et al., 2016; Park & Kwan, 2017). This approach has been incorporated into algorithms for detection or prediction of mood or behavior (Ameko et al., 2018; Cai et al., 2018), under the assumption that exposures influence outcomes. For example, in healthy participants, momentary exposure to nature may reduce unwanted rumination (Lopes et al., 2020), while momentary exposure to disordered neighborhoods may increase pain and fatigue (Cornwell & Goldman, 2020). In clinical populations, such as people being treated for substance-use disorders (SUDs), momentary exposure to drug cues might increase risk of craving or lapse, as we have found in momentary-assessment studies that did not include geolocation tracking (Epstein et al., 2009).

A different approach to activity space, to be emphasized here, is to analyze it in terms of its own spatial and temporal configuration irrespective of environmental characteristics. GPS tracks are coded by distance traveled, number of places visited, distance from residence, and other temporal patterns, fluctuations, or trends. This approach, which treats activity space more like a behavioral phenotype akin to what might be measured in nonhuman animals, lends itself to more person-centric questions, such as participants’ propensities (or abilities) to explore new places or vary their routines (e.g., Ai et al., 2019; Alessandretti, Lehmann, et al., 2018; Alessandretti, Sapiezynski, et al., 2018). Like the inferred-exposure approach, this spatiotemporal approach has been used to detect or predict behavior in both healthy and psychiatric populations (Bae et al., 2018; Canzian & Musolesi, 2015; Jaques et al., 2017; Pratap et al., 2019). The mechanistic underpinnings of this approach to risk detection are not always as clear as those for an exposure-based approach, but, as we discuss below, changes in movement could reflect changes in aspects of mentation that confer risk (e.g., of drug use in people with SUDs).

These two broad approaches to activity space—exposure and spatiotemporal—are not mutually exclusive, but, in most publications, attention usually focuses on one or the other. That was the case for our own prior work, which was (to our knowledge) the first to use GPS tracks from urbanites in the stabilization phase of outpatient treatment with medication for opioid use disorder (MOUD) (Epstein et al., 2014). We have used those data for both theoretical/mechanistic inquiries (via inferential statistics) and case-by-case classification and prediction (via machine learning), but always through the lens of inferred exposures (e.g., how encounters with physically disordered or well-maintained environments were associated with momentary changes in mood, drug craving, or lapse to drug use).

We now report on our use of activity space as a behavioral phenotype, an approach that elucidates aspects of geospatial behavior and its person-level correlates distinct from environmental exposure. We used geographical momentary assessment (GMA) – a combination of GPS and ecological momentary assessment (EMA) – to assess mobility and self-reported mental states in adults during their first 16 weeks of MOUD. Our immediate goal was to explore and describe how certain psychological states and traits relate to spatial and temporal activity space. Our long-term goal is to use static and passively-obtained data (e.g., personality traits and GPS-derived activity space measures) to assess risk in real time and to recognize when patients might benefit from a just-in-time intervention (Carpenter et al., 2020) during MOUD. We are interested in static and passive data because providing data is an ongoing burden on the study participant, a burden that could lead to noncompliance with (or discontinuation of) self-reporting.

To structure our analyses, we preregistered four sets of hypotheses and a detailed statistical plan (https://osf.io/vh8xr). The first three sets of hypotheses treat GPS data in purely spatial terms (as indicators of mobility) and are the main foci of this report. The fourth set of hypotheses abstracts GPS data into Neighborhood Hazard Scores (an indicator of inferred exposures); this set is included here because information about the places people visited provides context for our primary mobility measures. Some of our hypotheses were directional, based on our logical expectations concerning relationships between personality (including typical mood state) and size/diversity of activity space. Many of our hypotheses were nondirectional; we tested them because identifying and exploring relationships between certain variables and activity space could be valuable for real-time prediction of stress, drug craving, and drug use, a long-term goal for our laboratory and for addiction research.

Hypothesis Set 1 (non-directional): Activity-space size and diversity (area, distance, time moving, and number of places visited) would vary by time in treatment. Our rationale was that stabilization in treatment is a time of behavioral change, likely to be reflected by changes in travel routines (e.g., avoidance of areas with drug-related cues, and/or seeking of support services in new areas).

Hypothesis Set 2: Larger and more diverse activity space (larger area/distance traveled, more locations visited) would be associated with certain personality traits (Extraversion, and Openness to Experience), certain mood propensities (more positively valenced moods at randomly assessed moments, at the person level), and certain behaviors (cocaine use and more frequent social interactions, at the person level). Smaller and less diverse activity space (area/distance/locations) would be associated with other traits (Neuroticism), mood propensities (trait-like negative mood, and low-arousal moods), and behaviors (opioid use). Personality was assessed with the NEO Personality Inventory (240-item NEO-PI-3) (McCrae et al., 2005).

Our rationale for testing personality and mood as correlates of activity space was that such associations have been found in studies with other populations (Ai et al., 2019; Alessandretti, Lehmann, et al., 2018). Because the findings of those studies were mixed (see Discussion), we formulated our hypothesis on conceptual grounds, taking into account our expectations for people stabilizing on MOUD. Specifically, we previously found that mood, stress, and craving are highly salient with respect to behavioral outcomes during MOUD treatment (Panlilio et al., 2021; Panlilio et al., 2019; Preston et al., 2018), which suggests that these factors should also be considered when studying activity space in this population.

Our rationale for testing opioid use versus cocaine use as a correlate of activity space follows findings that cocaine is often preferred away from home, while opioids are often preferred at home (Badiani et al., 2019; De Pirro et al., 2018).

Hypothesis Set 3 (non-directional): The size and diversity of each participant’s activity space (area/distance/locations) would be associated with mood, craving, stress, substance use during treatment (polydrug use in particular), exposure to drug cues, and attrition from MOUD treatment. Our rationale was that all of these were aspects of functioning that could change during stabilization into MOUD treatment and that each of them could plausibly be reflected in changes in travel routines.

Hypothesis Set 4 (non-directional): The non-activity-space variables examined in Hypothesis sets 2 and 3 would also be associated with inferred exposure to psychosocial hazards (Glass et al., 2009) in the neighborhood where the participant submitted the EMA report. Our rationale was that we have previously shown that neighborhood exposure is related to drug use during treatment for MOUD (in ways that, as we noted earlier, were sometimes in unexpected directions; Epstein et al., 2014), but we have not previously assessed neighborhood characteristics in the context of mobility or personality.

2. Methods

2.1. Participants

The participants and data-collection procedures, which have been described in detail previously, were part of a natural-history study designed to determine how drug use, craving, and other behavioral/mental events were associated with exposure to stressors among MOUD patients (Panlilio et al., 2021; Panlilio et al., 2019; Preston et al., 2018). To focus on behavioral and mood changes during stabilization onto MOUD treatment, we only analyzed data from participants (N=225) who began a course of MOUD as part of our study (i.e., we excluded participants already receiving MOUD before study enrollment). Seventy-eight percent of participants were male; 64% Black/African-American, 33% White/Euro-American, 2% more than one race, and <1% Asian or of unknown race/ethnicity. Mean age was 46.5 years (SD=9.7, range=[26, 68]). All participants received buprenorphine (56%) or methadone (44%) and lived in or near Baltimore.

2.2. Data collection and measures

Participants used smartphones or GPS recorders to provide geographical data (continuous GPS tracking) and EMA data. Mobility was calculated as distance travelled per day, area covered per day, proportion of time within each day spent moving >0.9 m/s, and number of new locations visited each week. Moving >0.9 m/s could occur if walking, running, or riding in a vehicle. A location was defined as a place where the participant spent at least 15 consecutive minutes within a 25-ft radius; a new location was defined as a place that participants had not visited during the previous calendar week. More details on calculations of mobility metrics are in the online Supplement.

EMA self-report data analyzed here consisted of randomly prompted reports, scheduled to occur three times daily during the participant’s self-reported usual waking hours. These reports included responses to questions about mood, drug use, exposure to drug-related cues, and socializing. The study derived mood ratings (range 1–5) from endorsements of each of 24 mood adjectives (Stull et al., 2022), providing three scores: Positive Mood, Negative Mood, and Low Energy (the latter being a subset of Negative Mood, reflecting endorsement of the items “worn out,” “exhausted,” “sleepy,” and “fatigued”). The study assessed cocaine and opioid use by responses indicating whether either type of drug had been used “since you got to your present location,” or (if they indicated that they were currently walking from one place to another) “within the last 5 minutes.” Exposure to drug cues was a composite score based on items such as recently seeing someone selling or using drugs. Social behavior was assessed with two variables, based respectively on whether the participant indicated they were currently “alone (with no one)” or “socializing.”

The NEO Personality Inventory (240-item NEO-PI-3) (McCrae et al., 2005) was administered to most participants before the end of their fourth week in the study, but was not administered in the earliest wave of the study, so the sample size (173) was smaller for this measure.

In an analysis described earlier (Panlilio et al., 2020), the study divided participants into 6 drug-use clusters based on longitudinal analysis of urinalysis results. The current analyses used these clusters as a categorical and objective measure of drug use during treatment, complementing the self-report measures of drug use obtained through EMA. The clusters represent groups of participants with little or no use of opioids and cocaine (Negative cluster); persistent use of opioids, cocaine, or both (Opioid cluster, Cocaine Cluster, or Both cluster); or intermittent use of opioids, cocaine, or both (Sporadic cluster). We assigned those who left the study early to a Dropout cluster, reflecting treatment attrition.

Based on GPS coordinates associated with each randomly prompted EMA report, we used data from the Baltimore Neighborhood Indicators Alliance (bniajfi.org) to identify the neighborhood where a report was completed and to calculate a psychosocial hazard score associated with that neighborhood. The developers of the hazard scores (Glass et al., 2009) intended them to reflect “stable and visible features of neighborhood environments that can give rise to a heightened state of vigilance, alarm, or threat in residents.” We note that—even though residence in higher-hazard neighborhoods is associated with adverse health outcomes—we have previously found that people on MOUD in Baltimore tended to experience less stress and less negative mood while in neighborhoods rated by nonresidents (researchers), via a formal rating system, as being high in signs of social disorder, physical disorder, and drug activity (Epstein et al., 2014). The current analyses are not intended to address reasons for that finding, but can address its reliability across neighborhood-rating methods. We calculated hazard as the sum of Z-transformed scores from each of 10 separate categories, based on rates of: violent crime; vacant or abandoned residential properties; complaints about street conditions; liquor outlet density; 911 calls concerning shootings, assaults, narcotics, automobile accidents, or fire; percentage of population >25 years old with less than high-school diploma or equivalent; median household income (reverse scored); female-headed households with children under 18; unemployment; and households below the poverty line. A higher score indicates greater psychosocial hazard. Hazard data were available for each year of the study (2009–2017), but only for neighborhoods within city limits; analyses involving hazard were based on the available data.

2.3. Statistical analysis

Study staff performed all analyses in R version 4.2.1ba. We performed multilevel modeling using the lme4 package version 1.1–29 (Bates et al., 2015). The criterion for statistical significance was p < 0.05, with p < 0.1 considered marginal. Point estimates and fitted curves are reported with 95% confidence intervals; nonoverlapping intervals provide a conservative indication that estimates differ (conservative in that curves sometimes have overlapping intervals even when differing significantly). To facilitate comparison of effects, the study standardized numeric predictors and outcomes (mean = 0, standard deviation = 1). Raw means and standard deviations are in the Supplement. We plotted regression results with sjplot version 2.8.11 (Lüdecke, 2023).

For Hypothesis 1, we first fit two models (unconditional means and unconditional growth) for each spatial outcome (area, distance, time moving, number of locations, and neighborhood hazard) to determine whether these outcomes varied between participants by average level and changes over time. Fit statistics (BIC and AIC) indicated that the best fits were for models where intercepts and effects of time were allowed to vary between participants. The random intercepts and/or slopes for many individual participants differed from zero, and the effect of time was significant (p<.0001) in these unconditional growth models, except one (proportion of time moving, p=.22). To estimate the effects of predictors independently of personal trends over time, all subsequent multilevel models included varying intercepts and slopes for time (Hypotheses 2, 3 and 4).

Unlike NEO scores, which are considered traits, the EMA-based measures in Hypotheses 2 and 3 varied over time within participants. However, EMA data often exhibit trait-like aspects in addition to momentary variation (Bolger & Laurenceau, 2013). Accordingly, we partitioned EMA-based predictors into two parts: (1) the participant’s personal mean, or proportion for binary self-reported drug-use measures, to quantify differences between participants, and (2) momentary deviations from the participant’s personal mean—i.e., intra-individual differences between the personal mean and the current observation, to quantify the effects of changes within the participant. These between- and within-participant predictor values, which can be thought of as trait and state, respectively, were both allowed to interact with time, to assess whether their effects changed over the course of the study. In addition to the random slope for time (mentioned above), to avoid overfitting or underfitting, all multilevel models with a time-varying predictor were run (1) with a random slope for the predictor, (2) without a random slope for the predictor, and (3) with random slopes for the predictor and the interaction of the predictor with time; only the model with the lowest BIC was used for hypothesis testing. This model-selection procedure provides a good balance between power and risk of Type I error (Matuschek et al., 2017). Detailed results and effects plots for all tests of Hypotheses 2, 3 and 4 are in the Supplement.

Although we modeled activity space as an outcome and other variables (mood, stress, drug use, and socializing) as predictors, the relationships could be bidirectional. NEO personality scores, and some of the personal-mean EMA variables, are presumably traits that existed prior to the study and could potentially have causal effects on the outcomes studied here. However, our primary interest in this study is not causal inference, but to identify NEO personality variables and passively-obtained activity space measures that could be useful for detecting or predicting treatment-relevant states assessed with EMA. For the purpose of assessing need for a just-in-time intervention (i.e., predicting increased risk of lapse) during treatment for opioid use disorder, a variable could be valuable regardless of whether it is a cause, an effect, or simply correlated (e.g., sharing a common cause) with the treatment-relevant state (Bzdok & Ioannidis, 2019). Accordingly, we interpret all results as representing relationships between variables, not as evidence for (or against) causality.

Multilevel models can handle missing data, but if the missingness is not random, estimates can be biased. This bias can be reduced by multiple imputation of the missing values. To examine whether missing data were biasing our results, we conducted sensitivity analyses using multiple imputation with the mice package version 3.14.0 (van Buuren & Groothuis-Oudshoorn, 2011). For each set of outcomes (three sets based on time scale: daily distance/area/moving, weekly locations, and thrice-daily neighborhood hazard at the time of random prompts), all missing values of the outcome and all predictors were imputed simultaneously based on all nonmissing values of the outcome and predictors. For all models, the results did not differ substantially from the results with non-imputed data, suggesting that missing data did not bias the results.

3. Results

3.1. Change over time (Hypothesis Set 1)

Results are shown as coefficients for time in Tables 1, 2, and 3. Distance per day, area per day, and new locations visited per week decreased significantly over time in all models that included time and EMA variables as predictors (Table 1), consistent with the unconditional growth models described above (with time as the only predictor). However, in models with predictors that did not vary over time (personality variables in Table 2 and drug-use cluster in Table 3), mobility levels only changed significantly (decreased) over time for the locations outcome. (See Discussion regarding non-straightforward findings like this one: we mostly interpret them as parts of an overall portrait of complexity, not especially consequential in isolation.)

Table 1.

Multilevel modeling results for time-varying predictors

Abbreviations:

Hyp = Hypothesis number (3 and 4 were nondirectional); Exp. = Expected; Obs = Observed; + = Positive relationship between predictor and outcome; − = Negative relationship between predictor and outcome; b = Between participants; w = Within participants; × = Interaction; *** = p < .001; ** = p < .01; * = p < .05; (*) = p < .1

Table 2.

Multilevel modeling results for neuroticism, extraversion and openness (NEO)

Abbreviations:

Hyp = Hypothesis number (3 and 4 were nondirectional); Exp. = Expected; Obs = Observed; + = Positive relationship between predictor and outcome; − = Negative relationship between predictor and outcome; × = Interaction; *** = p < .001; ** = p < .01; * = p < .05; (*) = p < .1

Table 3.

Multiple regression results with drug use cluster as a predictor

Abbreviations:

Hyp = Hypothesis number (3 and 4 were nondirectional); Exp. = Expected; Obs = Observed; + = Positive relationship between predictor and outcome; − = Negative relationship between predictor and outcome; × = Interaction; Opi = Opioids; Coc = Cocaine; Spor = Sporadic, Drop = Dropout; *** = p < .001; ** = p < .01; * = p < .05; (*) = p < .1

3.2. Variables expected to be associated with higher mobility (Hypothesis Set 2)

Greater positive mood, as predicted, was associated with greater distance, area, and locations (Table 1). This association was between participants (trait-like). However, self-reported cocaine use was associated with mobility in the direction opposite to what we predicted: more self-reported cocaine use was associated with less distance, area, and time spent moving >0.9 m/s (Table 1). This was also between participants (trait-like). This unexpected finding was further supported by analysis of the drug-use clusters: participants whose urinalysis indicated persistent use of cocaine throughout the study had lower levels of area and distance traveled.

The model of daily area traveled as a function of being alone showed a marginal interaction effect: participants who were alone more often did not show the decrease in this measure of mobility over time that typified other participants. Socializing and two NEO traits (Extraversion and Openness to Experience) that were expected to be associated with higher mobility were not (see Figure 1).

Figure 1.

Summary plots of marginal effects from models of each activity-space variable (area, distance travelled, number of locations visited, time spent moving, and psychosocial hazard) as a function of Neuroticism, Extraversion and Openness (NEO) scores, corresponding to models in Table 2. All variables were standardized (mean = 0, standard deviation = 1). Red lines show activity-space measures at a high value of the NEO variable, and blue lines show activity-space measures at a low value of the NEO variable. Lighter-colored bands indicate 95% confidence intervals. Only the interaction plots are shown here. Main-effects plots and low/high NEO values used for plotting are in the Supplement. Text and asterisks indicate significant regression coefficients: ***=p<.001; *=p<.05; (*)=p<.1.

3.3. Variables expected to be associated with lower mobility (Hypothesis Set 2)

Higher neuroticism was consistently associated with lower values of all four measures of mobility (see Figure 1). A marginally significant interaction occurred between neuroticism and area traveled per day (though not the other mobility measures): participants with higher neuroticism covered more area, early in the study.

Opioids had a marginal within-participant relationship (in the predicted direction) on time spent moving (but not the other mobility measures): participants moved less on days when they reported opioid use.

We expected EMA-derived propensities toward negative mood and low energy to be associated with lower mobility but were not. That Neuroticism but not negative-mood propensity (an important component of neuroticism) was associated with low mobility seems counterintuitive; we return to this in the Discussion.

3.4. Variables expected to be associated with mobility (nondirectional, Hypothesis Set 3)

Participants with higher personal-mean levels of drug craving or exposure to drug-related cues generally had lower mobility. Specifically, those with higher personal-mean levels of craving had significantly lower distance travelled and fewer locations visited, and those with more drug-cue exposure were significantly lower on all four mobility measures. The study also found between-person findings involving within-participant interactions with time for some outcomes: later in the study, participants encountered more drug cues in weeks when they visited more locations; early in the study, participants spent less time moving when they were craving.

3.5. Variables expected to be associated with neighborhood-level psychosocial hazard (nondirectional, Hypothesis Set 4)

Participants with higher personal-mean (trait-like) Positive Mood had significantly higher neighborhood-level psychosocial hazard (i.e., they were usually in neighborhoods with more hazards), but momentary (state) Positive Mood was not significantly related to hazard. Participants with higher average levels of being alone had significantly lower levels of neighborhood hazard. Cocaine use was significantly associated with higher momentary hazard, especially later in the study, but opioid use was not. Momentary (state) exposure to drug-related cues was associated with higher hazard, but personal mean (trait) cue exposure was not.

4. Discussion

In our sample of outpatients stabilizing into treatment with MOUD, we found associations between GPS-derived measures of activity space (distance traveled per day, number of locations visited, and time spent moving) and important indicators of changes in life and treatment-related functioning. We preregistered hypotheses, for transparency about the extent to which the results would coincide with our expectations. The results did not support most of the hypotheses: as shown in the right-hand columns of tables 1, 2, and 3, only a minority of specific tests provided at least partial support of our predictions, and some tests of directional hypothesis were opposite to what we had hypothesized. To our knowledge, this is the first report to examine GPS-derived measures of activity space among outpatients in MOUD. Accordingly, our goal was not to test a strongly held set of prior expectations. Rather, we prespecified hypotheses to structure the analyses and characterize the data in a systematic way. Thus, the results are no less informative for being discrepant with our hypotheses. We now discuss each set of results in turn.

4.1. Hypothesis of overall change over time in mobility

Although we did not have a control group of people with MOUD who were not receiving treatment, and we therefore cannot directly attribute effects to treatment, all our participants started treatment at the beginning of the study, which allowed us to assess whether consistent changes occurred in activity space as a function of time since starting treatment. We predicted that the first weeks of MOUD treatment would be a time of change in activity space, and this prediction was supported: distance per day, area per day, and number of new locations per week decreased significantly over time in all models that included time and EMA variables as predictors (though time was not consistently significant when used as a covariate in models focusing on personality variables or drug-use cluster).

4.2. Directional hypotheses for mobility’s associations with mood, personality traits, and cocaine (versus opioid) use

4.2.1. Mood

As we had hypothesized, positive mood was associated with greater distance, area, and number of new locations. These associations were trait-like rather than momentary or situational: participants with high average levels of positive mood in EMA were more mobile than average. Measures of negative mood and low arousal, however, were not reliably associated with activity space at either the trait-like or momentary level, counter to our hypotheses. This absence of an association seems discordant with our finding that mobility was lower in participants with more trait Neuroticism. This discrepancy should be considered in the context of prior findings that people can generate summary statements about their mental or behavioral tendencies that differ from slices of mentation or behavior obtained on single occasions (Epstein, 1979, 1980). Each of those two basic approaches to measurement, representing a “remembering self” and an “experiencing self” (Campara et al., 2021), may be associated with outcomes of interest. As EMA researchers, we often expect that summary measures of momentary assessments will be more sensitive than standard questionnaires (Moore et al., 2016), but this finding reminds us that exceptions exist.

4.2.2. Personality traits

Our hypotheses for the Big Five personality traits, as assessed by the NEO, were only supported for one of these traits: Neuroticism was associated with lower mobility by all four measures (area, distance, moving, and locations). We did not find the positive associations we had hypothesized between mobility and Extraversion or Openness to Experience.

Neuroticism includes aspects that could seem likely to limit activity space, such as proneness to anxiety, and sensitivity to social rejection (McCrae et al., 2005). Nonetheless, Ai et al. (2019) found no association between Neuroticism and mobility (in college students in China), and Alessandretti et al. (2018) unexpectedly found that Neuroticism was associated with greater mobility (more changes in travel routines over long time scales). The latter finding occurred in a US sample of over 1,000 participants that included but was not limited to college students; Alessandretti et al. interpreted this finding as a manifestation of emotional and behavioral instability. In the same sample, Alessandretti et al. found that Openness to Experience was similarly associated with more changes in travel routines over long time scales, possibly as a manifestation of curiosity rather than instability. Among these three sets of NEO/mobility findings—by Ai et al. (2019), Alessandretti et al. (2018), and by us—each set included unexpected elements, and each set was partly discrepant with the other sets. We conclude that activity-space metrics do not provide a straightforward readout of personality traits and that relationships between personality and activity space might vary across culture, populations, and specific life demands unique to each person’s daily roles, obligations, and opportunities.

4.2.3. Cocaine (versus opioid) use

On days when participants used opioids (assessed via EMA entries), they were marginally less mobile, in agreement with our drug-specific hypothesis. However, the findings did not support our hypothesis that cocaine use would be associated with greater mobility. Instead, we found a reliable person-level association in the opposite direction; participants who frequently used cocaine had lower mobility.

We based this prediction on a body of evidence that people prefer to use opioids while at home and cocaine while elsewhere (Badiani et al., 2019; De Pirro et al., 2018). In hindsight, we suspect that those prior findings do not logically lead to the prediction we made regarding greater mobility. As some published personal accounts attest, an instance of cocaine use, even outside the home, may lead to continuation of use until the supply is exhausted (Smith, 2022), which could decrease mobility. EMA data (without GPS) suggest that days with cocaine use are typically more sedentary than other days (Epstein & Preston, 2010). Thus, our cocaine finding probably indicates a mismatch between the behaviors we expected (Badiani et al., 2019; De Pirro et al., 2018) and the way in which we examined the GPS data.

4.3. Nondirectional hypotheses for mobility’s associations with drug craving, polydrug use, cue exposure, and MOUD attrition

We did not detect associations between mobility and the likelihood of polydrug use or treatment dropout, events that investigators and clinicians typically hope to predict and prevent (Burgess-Hull et al., 2022). We did find that drug-cue exposure and craving were associated with lower mobility when comparing participants to each other, but not by moments within participants. These trait-like associations are broadly consistent with prior findings of lower mobility as a concomitant of worse psychological functioning (e.g., in people with histories of depression) (Auerbach et al., 2022), but contrast with our own finding that mobility decreases during stabilization into MOUD. Perhaps, like findings we cited earlier—in which more mobility was associated not just with the relatively adaptive traits Extraversion and Openness to Experience, but also with the less adaptive trait Neuroticism (Alessandretti, Lehmann, et al., 2018)—this finding is a warning that patterns of mobility that seem identical (at least, under a particular system of data coding) may reflect very different processes for different people. That is, any single measure, such as distance travelled or number of places visited, cannot be fully understood without also considering contextual information, such as activities engaged in and the person’s perception of the places.

We also noted interaction effects indicating changes over time in relationships between certain predictors (cue exposure, craving) and activity-space measures (new locations visited, time spent moving). These findings may be a warning as well—a reminder that between-person differences can have very different meanings from within-person changes or fluctuations. In an observational study such as this, we can only speculate as to whether, for example, craving led to increased mobility because it increased drug-seeking, or whether treatment promoted abstinence, which increased craving, stress, and mobility. These and other relationships could occur in different people or in the same person at different times. Definitively assessing causality in relationships such as these will require further research, ideally based on experimental manipulations.

4.4. Nondirectional hypotheses for environmental exposures (Neighborhood Psychosocial Hazard)

Higher personal-mean levels of positive mood were associated with more time in higher-hazard neighborhoods, and higher personal-mean frequencies of being alone were associated with more time in lower-hazard neighborhoods. The association between trait-like positive mood and time in higher-hazard neighborhoods may seem counterintuitive, but it broadly supports the reliability of our earlier finding (Epstein et al., 2014), in which we used a different metric of neighborhood disorder (the Neighborhood Inventory for Environmental Typology, based on blockface-level measures of physical and social disorder (Furr-Holden et al., 2010); there, we found that disordered neighborhoods were associated at the momentary level with less stress, less negative mood, and less drug craving. Subtle but important differences could exist between the present finding (that positive mood was higher in participants who spent more time in high-hazard neighborhoods) and the earlier finding (that negative mood was lower in participants while they were in a physically disordered neighborhood).

We have not attempted to speculate on the reasons for these possibly counterintuitive associations. In unpublished work, we conducted geonarrative interviews (Kwan & Ding, 2008) with a subset of our participants, asking for their explicit recountings of how specific locations affected their moods. Our main finding was that participants reported few or no reliable associations between place and mood; one interpretation of this is that places sometimes influence mood (and behavior) in ways that are inaccessible to consciousness. The extent to which this happens is an open question: psychology has seen decades of debate on whether people can discern their own motivations and other mental processes, with some authors arguing that we usually can (Baumeister et al., 2011; Dang et al., 2020; Gignac G & Zajenkowski, 2020; Newell & Shanks, 2014) and others arguing that we usually cannot (Atir et al., 2015; Nisbett & Wilson, 1977; Wilson & Schooler, 1991). To the extent that environments can influence behavior without conscious awareness, clinicians may find it all the more useful to detect such influences via GPS or other sensor data; the open question (De La Fabián et al., 2023) is whether detection can be reliable enough to drive momentary interventions or whether it will primarily produce group-level associations such as the ones we have found.

In our current analyses, we found some momentary associations with psychosocial-hazard exposures that we did not find for the mobility measures that were our primary focus. Specifically, independent of their trait-like levels, participants were more likely to be in a higher-hazard neighborhood while they were using cocaine (especially later in the study) or in the presence of drug-related cues. These state-like effects accord with previous findings that the level of neighborhood disorder in the current environment (but not the level in the person’s home neighborhood) is associated with momentary spikes in pain and fatigue (Cornwell & Goldman, 2020), and that people living in high-crime neighborhoods typically find safe places within the neighborhood in which to spend time (DaViera et al., 2020).

At the person level, our mobility data (hypothesis set 3) indicated that higher trait mobility—by all four measures—was related to low overall exposure to drug cues. The positive momentary relationship between psychosocial-hazard and drug cues (hypothesis set 4) could reflect participants’ acquisition or use of drugs in higher-hazard neighborhoods. For the number of new locations per week, a within-participant interaction effect also occurred in which, later in the study, participants visited more locations in weeks when they had higher cue exposure. This could reflect an increased likelihood of visiting a high-hazard neighborhood to seek drugs later in the study.

4.5. Limitations

This study focused intensively on a specific population, people in Baltimore seeking treatment for opioid use disorder involving intravenous heroin use (predating widespread substitution of fentanyl for heroin). Further work will be necessary to determine whether the findings generalize to populations in other areas, with different geographic and demographic situations.

A different frequency of EMA sampling might produce more precise or accurate results. However, a tradeoff exists between sampling frequency and user compliance. The optimal parameters might differ depending on whether the goal is theoretical/mechanistic inquiries into psychological processes vs. practical applications involving automated detection/prediction of risk for individual patients (Yarkoni & Westfall, 2017). Here, we used a moderate number of prompts per day, but our eventual goal is to be able to obtain sufficient information with less frequent prompting, supplemented by passive data.

Clinic-attendance requirements for our patients imposed a requirement for travel (to our clinic, several times per week), reducing the possible extent of between-person differences in activity space. The between-person differences we analyzed were those that occurred over and above all participants’ clinic visits.

4.6. Conclusions

Among the variables we studied, we found that higher mobility was associated with trait-like variables including positive mood and Neuroticism, and that higher neighborhood psychosocial hazard is associated with trait-like positive mood but also with state-like levels of drug craving and contact with drug-related cues. Importantly, though, we found that many of the relationships we assessed were complex and not necessarily in accord with expectations based on available theory or empirical resources. Some of our individual findings—mood measures that were associated with only one of several mobility measures, or associations that changed in unexpected ways across time since start of treatment—may or may not replicate in other samples of people with SUDs. This complexity suggests that, perhaps not surprisingly, for passive measures to be useful in detecting and predicting the need for intervention, the algorithms will not only have to be highly specific for the population, but tailored to specific subsets of the population.

Highlights.

• We assessed mobility as a function of personality, mood, craving and drug use

• Mobility decreased over the first 16 weeks of buprenorphine or methadone treatment

• Neuroticism and persistent cocaine use were each associated with low mobility

• Trait-like positive mood was associated with high mobility

• Mobility was complexly related with other treatment-relevant psychosocial measures

• Mobility was generally related more to between-person traits than momentary states