|Home | About | Journals | Submit | Contact Us | Français|
Alcohol dependence has been described as a relapsing condition and it has been proposed that alcohol lapses could potentially be explained by dynamic associations between contextual, interpersonal, and intrapersonal risk factors. Yet, few studies have tested the associations between risk factors in the prediction of lapse dynamics. The current study was a secondary analysis of data from the COMBINE study (n=1,383; COMBINE Study Research Group, 2003). The goal of the current study was to examine static (alcohol dependence severity, treatment history, marital status, psychiatric symptoms) and dynamic (negative affect, craving, stress) predictors of heavy drinking during the course of treatment and up to one year following treatment. Results from dynamic latent difference score models indicated that higher levels of static and dynamic risk and increased dynamic risk over time were significantly associated with greater increases in heavy drinking. Likewise, more frequent heavy drinking and higher static risk predicted higher levels of dynamic risk. In addition, changes in dynamic risk factors significantly mediated the association between changes in heavy drinking and both psychiatric symptoms and treatment history. It is important to note that while the effects of static and dynamic risk factors in the prediction of heavy drinking were statistically significant, the magnitude of the effects were small. The current study provided partial support for a dynamic model of relapse; however future research using intensive longitudinal data collection and more advanced statistical techniques could further elucidate lapse dynamics and potentially improve relapse prevention planning.
Across several studies, it has been shown that the most common outcome following a behavior change attempt is a return to engaging in the undesired behavior (Polivy & Herman, 2002). For example, it has been shown that more than 60% of individuals will have at least one drinking episode (i.e., “lapse”) in the first year following alcohol treatment (Maisto, Pollock, Cornelius, Lynch, & Martin, 2003; Whitford, Widner, Mellick, & Elkins, 2009). Several authors have proposed that high rates of lapsing might be partially explained by the complexity of the addictive behavior change process (Connor, Symons, Feeney, Young, & Wiles, 2007; Donovan, 1996; Niaura, 2000; Skinner, 1989; Warren, Hawkins, & Sprott, 2003) and many models of relapse have been developed that attempt to characterize the antecedents of lapse events (Annis, 1986; Cronkite & Moos, 1980; Litman, 1986; Ludwig & Wikler, 1974; Marlatt & Gordon, 1985; Sanchez-Craig, 1976; see Connors, Maisto, & Donovan, 1996 for a review).
One of the most widely cited models of relapse, the cognitive-behavioral model, was first proposed by Marlatt and Gordon (1985) in their influential text on relapse prevention. The cognitive-behavioral model of relapse, which combines cognitive (e.g., beliefs about one’s ability to abstain), behavioral (e.g., coping responses), and situational/environmental antecedents of substance use lapses, was largely derived from a taxonomy of relapse situations developed by Marlatt and Gordon (1980). The relapse taxonomy, which was based on qualitative interviews with clients who had experienced drinking lapses following treatment, consisted of three hierarchically arranged levels that distinguished between the intrapersonal and interpersonal precipitants (level 1); eight categories of antecedents within the level 1 precipitants (level 2); and specific subdivisions for five of the eight level 2 categories. The eight subdivisions within the two level 1 categories, included coping with negative emotional states, coping with negative physical-psychological states, enhancement of positive-emotional states, testing personal control, and giving in to temptations and urges; and coping with interpersonal conflict, social pressure, and enhancement of positive emotional states. Five of these subdivisions were further divided on level 3 (e.g., Coping with negative emotional states was segregated into Coping with frustration and/or anger and Coping with other negative emotional states).
The relapse taxonomy and cognitive behavioral model of relapse have been very influential in the field of addiction, making significant contributions to clinical practice and stimulating the development of relapse prevention strategies. Due to the widespread popularity of Marlatt’s model, the National Institute on Alcohol Abuse and Alcoholism (NIAAA) funded a large scale study (the Relapse Replication and Extension Project (RREP)) to test the reliability and validity of the taxonomic system for classifying relapse episodes. Investigators in the RREP recruited 563 clients with alcohol dependence from alcohol treatment programs in three geographically distinct areas in the United States and conducted bimonthly prospective assessments of drinking and potential relapse antecedents for one year. Results from the RREP raised significant methodological issues concerning the reliability (Longabaugh, Rubin, Stout, Zywiak, & Lowman, 1996), construct validity (Maisto, Connors & Zywiak, 1996), and predictive validity (Stout, Longabaugh, & Rubin, 1996) of Marlatt’s model. Based on the findings in the RREP, a major re-conceptualization of the relapse taxonomy was suggested (Donovan, 1996).
Following up on this recommendation, Witkiewitz and Marlatt (2004) proposed a re-conceptualization of the cognitive-behavioral model of relapse as a nonlinear dynamic system. The dynamic model of relapse builds upon several previous studies of relapse risk factors (Connors, Maisto, & Zywiak, 1996; Lowman, Allen, & Miller, 1996; Miller, Westerberg, Harris, & Tonigan, 1996; Shiffman, Balabanis, Paty, Engberg, Gwaltney, Liu et al., 2000) by incorporating the characterization of distal and proximal risk factors proposed by Shiffman (1989; see also Donovan, 1996). Distal risks, which are thought to increase the probability of relapse, include background variables (e.g., alcohol dependence) and relatively stable pre-treatment characteristics (e.g., expectancies). Proximal risks actualize, or complete, the distal predispositions and include transient lapse precipitants (e.g., stressful situations) and dynamic individual characteristics (e.g., negative affect). Combinations of precipitating and predisposing risk factors are innumerable for any particular individual and may create a complex system in which the probability of relapse is greatly increased. The system is further characterized using the temporal definitions of tonic and phasic processes, where tonic processes represent stable factors and phasic processes represent transient risk precipitants (Grace, 2000). Risk factors may operate within either tonic or phasic processes. For example, momentary self-efficacy (i.e., phasic) has been shown to predict smoking lapses above and beyond that which is predicted by baseline (i.e., tonic) self-efficacy (Shiffman et al, 2000).
The dynamic model of relapse has generated enthusiasm among researchers and clinicians who have observed these processes in their data and their clients (see Ashton, n.d.; Brandon, Vidrine, & Litvin, 2007; Cohen & Sutker, 2006; Hunter-Reel, McCrady, & Hilderbrandt, 2009; McCarthy, Piasecki, Fiore, & Baker, 2006; Stanton, 2005). Yet, the dynamic model of relapse is a theoretical model and has not yet been subjected to a rigorous empirical test. Empirical analyses of the RREP data by Miller and colleagues (1996) and Connors and colleagues (Connors, Maisto, & Zywiak, 1996) partially inspired the development of the theoretical dynamic model of relapse (Witkiewitz & Marlatt, 2004) and many aspects of the model are supported by these early empirical studies. Miller and colleagues (1996) used data from the Albuquerque New Mexico site of the RREP in the estimation of prospective relapse models that incorporated six domains of relapse risk factors, life events (e.g., interpersonal stress), cognitive variables (e.g., self-efficacy), coping resources, craving, and pre-treatment characteristics (e.g., alcohol dependence), as predictors of drinking outcomes at the six month assessment. Pre-treatment characteristics were assessed at baseline and considered “static antecedents” (p. 156), whereas proximal risk factors (life events, cognitions, coping, and craving) were measured at the four month assessment point and described as “dynamic antecedents” (p. 156). Results indicated that, with the exception of life events, all of the proximal risk factors were significantly associated with six month outcomes and proximal risk factors were stronger predictors of relapse than the distal risk factors.
Using data from the Buffalo NY site of the RREP, Connors and colleagues (Connors, Maisto, & Zywiak, 1996) estimated a path model of drinking outcomes (frequency, intensity, and drinking consequences) in months seven through twelve following treatment predicted by five relapse risk factor domains: background characteristics (e.g., psychiatric symptoms), alcohol involvement (e.g., alcohol dependence), treatment factors (e.g., treatment satisfaction), coping skills, and stressors. The model incorporated direct effects of baseline background characteristics and alcohol involvement, the six month assessments of the treatment factors, coping skills, and stressors (i.e., proximal influences), as well as the indirect effects of distal influences on drinking outcomes via the proximal influences. Background characteristics, alcohol involvement, treatment, and coping skills were all significant predictors of drinking frequency and intensity, and both alcohol involvement and coping skills were significantly associated with drinking consequences. The indirect effects were relatively small in magnitude, although the authors did note the somewhat larger effect of background characteristics impacting outcomes via treatment factors and treatment factors impacting outcomes via coping skills.
Analyses by Miller and colleagues (1996) and Connors and colleagues (Connors, Maisto, & Zywiak, 1996) provided partial support for some of the hypotheses of the dynamic model of relapse. Namely, both studies showed that distal and proximal influences play an important role in predicting post-treatment drinking outcomes and lapse events. There are also several aspects of the dynamic model that were not addressed in either study. Both studies focused on drinking outcomes at either a single time-point (Miller et al., 1996) or averaged across time (Connors, Maisto, & Zywiak, 1996) and neither study examined the potential reciprocal effects of drinking on proximal risk factors; thus ignoring temporal associations between risk factors and drinking. Likewise, both studies focused on post-treatment associations between risk and alcohol lapses, and did not address whether proximal risk factors, which are modifiable, influenced alcohol use during treatment. Both studies also relied on composites of risk domains and used statistical techniques that did not take into account composite measurement error. If there was a high degree of measurement error in the composites, then meaningful risk factor effects could have been obscured (Jaccard & Wan, 1995). In addition, the sample sizes for both studies (n=122; n=142) were rather small for testing complex models.
The goal of the current study was to address the limitations of previous studies (Miller et al., 1996; Connors et al., 1996) using data from the COMBINE study (COMBINE Study Research Group, 2003). The current study was designed to examine changes in proximal risk and drinking across multiple time points in a large sample of individuals (n = 1383) and estimate reciprocal associations between proximal risk and heavy drinking using statistical techniques that take into account measurement error. Specifically, the current study used latent variable modeling techniques to examine the static and dynamic associations between distal risk factors (including alcohol dependence, marital status, pre-treatment psychiatric problems, and pre-treatment self-efficacy), proximal risk factors (including craving, perceived stress, and negative mood), and frequency of heavy drinking during the course of treatment and up to one year following treatment.
The data for this study are from the COMBINE study (“Combined Pharmacotherapies and Behavioral Interventions for Alcohol Dependence;” COMBINE Study Research Group, 2003), a multi-site randomized trial. A total of 1383 subjects across 11 research sites were randomized into 9 treatment groups, described below. Treatment was provided for 16 weeks and participants were followed for one year following treatment.
The sample was recruited from inpatient and outpatient referrals at the study sites and throughout the community. Prior to baseline, 4965 volunteers were screened by telephone to determine whether the individual met eligibility criteria. Participants were excluded if they were dependent on another drug besides alcohol, nicotine, or cannabis, recently used opioids, had a serious mental illness, had any other medical condition that could disrupt study participation, had taken one of the study medications 30 days prior to baseline, or took medication that could raise the potential risks of the study. To be included in the study, subjects needed to have a minimum of 14 drinks (females) or 21 drinks (males) average per week over a successive 30-days in the 90-day period prior to beginning abstinence. Additionally, participants needed to have two or more days of heavy drinking in the 90-day period with the last drink being within 21 days of enrollment. Heavy drinking days was defined as 4 drinks for females and 5 drinks for males. Following meeting eligibility criteria, subjects were required to produce a breath alcohol level of zero before completing consent and baseline assessments.
The final sample included 1,383 participants, 31% were female and 69% were male, 23% of the study population were ethnic minorities (76.3% Non-Hispanic White, 11.6% Hispanic American, 7.8% African American, and 4.1% Other). The subjects’ median age was 44 years, 71% had at least 12 years of education, and 42% were married. Research retention rates did not differ significantly between groups. Within treatment, 94% completed all drinking data, while one year post treatment 82.3% completed the drinking data.
Upon meeting inclusion and exclusion criteria, subjects completed a baseline assessment and were randomly assigned to one of nine treatment groups1. The Medical Management groups (n=607) included: Naltrexone, Acamprosate, Naltrexone + Acamprosate, and Placebo. The Combined Behavioral Intervention (CBI) groups (n=776) consisted of: Naltrexone + CBI, Acamprosate + CBI, Naltrexone + Acamprosate + CBI, Placebo + CBI, and CBI-only (COMBINE Study Research Group, 2003). Subjects received treatment for a total of 16 weeks; participants receiving study medication were offered 9 Medical Management visits and those who received CBI had a maximum of 20 sessions. Participants were followed for 52 weeks post-treatment and seen at the site for assessments at 10, 36, and 52 weeks following treatment.
Percent heavy drinking days was used as the primary outcome variable because it combines both frequency and intensity of drinking. The Form-90 interview (Miller & Del Boca, 1994) was used to calculate Percent Heavy Drinking Days (PHD). Heavy drinking was defined as 4 or more drinks per day for women and 5 or more drinks per day for men. In the COMBINE study, drinking measures were derived in the prior 30 days at baseline and at during and post treatment visits. In the current study we examined heavy drinking across the following assessment periods, which were matched to the proximal risk assessments described below:during treatment at weeks 1, 2, 4, 8, 12, and 16 weeks, 10 weeks following treatment (assessment of the 30 days prior to the 10 week posttreatment assessment), 36 weeks following treatment (assessment of the 30 days prior to the 36 week posttreatment assessment), and 52 weeks (assessment of the 30 days prior to the 52 week posttreatment assessment).
Negative mood, perceived stress, and craving were incorporated as proximal risk factors, as described in more detail below.
During treatment negative affect was estimated using thehe Profile of Mood States – Brief (POMS; McNair, Lorr, & Droppleman, 1971), which was administered at baseline, after the first two weeks of treatment, and every four weeks during the 16 weeks of treatment. Participants were queried as to how they were feeling during the past week using 30 adjectives describing feelings and moods (e.g., Tense, Angry, Annoyed, etc.), with ratings for each adjective ranging from 0 (not at all) to 4 (extremely). Ratings on the 30 items were combined into six mood subscales: Tension, Depression, Anger, Vigor, Fatigue, and Confusion. For the current study, the Tension, Depression, Anger, and Fatigue subscales were included as indicators of a proximal risk latent variable at each time point. The internal consistency reliabilities for each of the subscales exceeded α = 0.70.
Post-treatment negative affect was measured by the 53-item Brief Symptom Inventory [BSI; (Derogatis, 1993), which assesses self-reported psychiatric symptomatology in a variety of domains via ratings on a 5-point scale (0 = Not at all, 4 = Extremely). Scores on the Depression, Anxiety, and Hostility domains were used as indicators of a negative affect latent variable following treatment. The BSI was administered at baseline, weeks 8 and 16 during treatment, and weeks 10, 36, and 52 following treatment.
The Obsessive Compulsive Drinking Scale (OCDS; Anton, Moak & Latham, 1995) is a 14-item self-report instrument assessing drinking-related thoughts, urges to drink, and the ability to resist thoughts and urges to drink, was administered at baseline, after the first two weeks of treatment, every four weeks during the 16 weeks of treatment, and 10 weeks following treatment. Items were rated on a five-point Likert-type scale, with lower ratings indicating less craving. Two of the 14 items, assessing quantity and frequency of drinking in the past week, were excluded from all analyses and a total score was calculated by summing the remaining 12 items at each time point. Reliability of the scale was greater than α = 0.90 at all time points.
The Alcohol Abstinence Self-Efficacy [AASE; (DiClemente, Carbonari, Montgomery, & Hughes, 1994)] scale determined client confidence to abstain from alcohol in high-risk situations via self-report ratings on a 5-point scale (1=Not at all, 5=Extremely) encompassing subscales of negative affect, positive/social, physical concerns, and withdrawal/urges. In the COMBINE trial, the AASE was administered at baseline and during the last week of treatment. The AASE confidence scale score exhibited excellent internal consistency (Cronbach α = 0.97). The AASE was administered at baseline, the last week of treatment, and 10 weeks following treatment.
The four item version of the Perceived Stress Scale (PSS; Cohen, Kamarck, & Mermelstein, 1983) was used to measure perceptions of stressfulness over the past week on a five point Likert-type scale (from never to very often). The PSS was administered during the first two weeks of treatment, then every two weeks during the first 12 weeks of treatment, during the last week of treatment, and 36 weeks following treatment.
Five variables were selected as relatively baseline characteristics that have been previously been associated with alcohol relapse: self-efficacy (Witkiewitz, van der Maas, Hufford, & Marlatt, 2007); alcohol dependence severity (Witkiewitz, 2008), marital status (Dawson, Grant, Stinson, Chou, Boji, & Ruan, 2006), treatment history and psychiatric problems (Hufford, Witkiewitz, Shields, Kodva, & Caruso, 2003). Other risk factors were considered based on their availability in the COMBINE data, including readiness to change, social support, drinking network size, and family history of alcohol problems. These factors were not related to the outcomes in preliminary testing and were thus eliminated from the models in the current study. Self-efficacy was measured by the AASE at baseline. Alcohol dependence severity was based on scores from the Alcohol Dependence Scale (ADS, Skinner & Horn, 1984), a 25 item measure of alcohol withdrawal symptoms, impaired control over drinking, awareness of compulsions to drinking, tolerance, and drink-seeking behavior. Psychiatric problems were assessed using the Global Severity Index of the 53-item Brief Symptom Inventory [BSI; (Derogatis, 1993), which assesses self-reported psychiatric symptomatology in a variety of domains (e.g., depression, anxiety, somatization, psychoticism) via ratings on a 5-point scale (0 = Not at all, 4 = Extremely). Internal consistency of the Global Severity Index in this sample was excellent (Cronbach α = 0.97).
Associations between risk factors and drinking outcomes were examined using a dynamic longitudinal latent variable modeling framework. Specifically, the multivariate longitudinal associations were examined using latent difference score models (McArdle, 2001; McArdle & Hamagami, 2001). Latent difference score models are an extension and combination of both latent growth models and autoregressive models, in which the goal is to evaluate longitudinal growth across time while estimating the dynamic inter-relationships between multiple change processes (McArdle & Grimm, 2010; McArdle & Nesselroade, 1994). The latent difference score separates true score variance and measurement error in calculating change between observed variables across repeated measures (McArdle & Hamagami, 2001) using a latent variable model, rather than an observed change score as reliability and validity of both observed and residualized change scores have been questioned (Cronbach & Furby, 1970). Each observed score (e.g., percent heavy drinking days (PHD)) is modeled as an indicator of a latent variable that represents the true score (e.g., expected PHD, if measured without error) and a latent variable that represents the degree of measurement error (i.e., residual variance). The change in true scores across time can then be modeled using three separate longitudinal parameterizations in a single model (Figure 1): autoregressive paths between adjacent true scores (i.e., PHD in week 26 regressed on PHD in week 16); latent difference scores (labeled d1-d3, p1-p3 in Figure 1) that represent the proportional change (α parameters in Figure 1) from one time-point to the next, and a latent slope that represents a constant change over time (β parameters in Figure 1).
The latent slope is considered to be a measure of the degree of “change in changes” over time, after controlling for the autoregressive and proportional change components, as well as the influence of bivariate change. The latent slope can be thought of as a nonstationarity parameter, whereby if the latent slope equals zero then there is essentially a stationary process that does not change over time and a non-zero slope indicates that the degree of change in the latent variable from one time-point to the next changes over time. A positive latent slope would characterize a process that tends to change by a greater degree across time. For instance, in the COMBINE data, average PHD remains stable from week 8 to week 12 to week 16 at 17%, but then increases to 22% at week 26, thus the degree of change in PHD is increased from week 16 to week 26. A negative latent slope characterizes a process that changes less across time. Bivariate change is modeled by the cross-lagged paths from the true score at time t-1 to the latent difference score at time t (γ parameters in Figure 1). For more information on the parameterization of the model, interested readers are referred to detailed descriptions of the latent difference score model and its extensions (see McArdle, 2001; McArdle & Grimm, 2010).
Bivariate latent difference score models require several constraints for model identification. In the current study a series of models with various parameter constraints were estimated and compared using χ2 difference tests. The results from alternative models are available from the author and only the final bivariate latent difference score models will be described in the current study. In the final models residual variances of the measurement errors were constrained to be equivalent for each measure across time, the β parameters were constrained at 1.0 to reflect a constant rate of change2, and for the during-treatment models the α-parameters (representing proportional change) were constrained to equality. Constraints on the γ parameters, which measure the degree of bivariate change, are presented in the results section.
In the current study, models were conducted in several steps for two time periods: during treatment and following treatment. First a set of models were estimated to examine changes in percent heavy drinking days and proximal risk factors during the course of the 16-week treatment. Weekly percent of heavy drinking days reported at the time of each proximal risk assessment point (weeks 1, 2, 4, 8, 12, and 16 during treatment) were used as drinking outcomes3 and all available measures of proximal risk were incorporated as indicators of a proximal risk latent variable, described below. Second, models of heavy drinking and proximal risk over the 52 weeks following treatment were estimated. For the post-treatment models, percent of heavy drinking days in the 30 days prior to each COMBINE assessment point (10 weeks, 36 weeks, and 52 weeks following treatment) were used as drinking outcomes and all available measures of proximal risk (described above) were incorporated as indicators of a proximal risk latent variable, described below. For each time period, univariate difference score models of each outcome (proximal risk and percent heavy drinking days) were estimated separately and then combined into bivariate difference score models. After estimating unconditional bivariate difference score models (without additional covariates), the baseline covariates (i.e., distal risks) were incorporated as predictors of the constant change in heavy drinking and proximal risk in the bivariate difference score models. Finally, mediation analyses were conducted to determine whether the proximal risk factors mediated the associations between distal risk factors and heavy drinking. Mediation analyses were conducted using the product of coefficients method (MacKinnon, Lockwood, Hoffman, West, & Sheets, 2002).
All models were estimated using Mplus version 6 (Muthén & Muthén, 2010). Considering the complex sampling design in the COMBINE study (participants recruited from 11 academic sites), all parameters were estimated using a weighted maximum likelihood function and all standard errors were computed using a sandwich estimator (the MLR estimator in Mplus). MLR provides the estimated variance-covariance matrix for the available data and therefore all available data were included in the models. Maximum likelihood is a preferred method for estimation when some data are missing, assuming the data are missing at random (Schafer & Graham, 2002). The fit of all models was evaluated by χ2 values, the Root Mean Square Error of Approximation [RMSEA; (Browne & Cudeck, 1993)], and the Comparative Fit Index [CFI; (Bentler, 1990)]. Models with RMSEA < 0.05 and CFI >0.95 were considered a good fit to the observed data (Hu & Bentler, 1999) and models with RMSEA < 0.08 and CFI >0.90 were considered a reasonable fit.
At baseline the average percentage of heavy drinking days was 65.52% (SD = 28.57%) and during the 30 days prior to baseline only 0.9% of the sample (n = 12) reported no heavy drinking days. By the end of treatment the average percentage of heavy drinking days was 16.78% (SD = 28.61%) and 49.7% of the sample (n = 687) reported no heavy drinking days. An inspection of the descriptive statistics for all primary measures in Table 1 demonstrates the marked reductions in percent heavy drinking days from baseline to up to one year following treatment. There were also reductions in negative mood, stress, and craving during treatment, and improvements in self-efficacy and psychological health immediately following treatment.
Preliminary modeling of each proximal risk factor (craving, stress, and negative mood) indicated that each of the proximal risk factors were reliably associated with percent heavy drinking days at each time-point and across time. Likewise, the associations between risk factors ranged from r = 0.44 to r = 0.75 (p < 0.005) across time, thus there was considerable overlap in the proximal risk factors and consistency in their prediction of heavy drinking. Rather than presenting all subsequent models separately for each proximal risk factor (which would require the presentation of three models for each of the bivariate difference score models presented below) longitudinal confirmatory latent factor models that combined the joint variance of the proximal risk factors were estimated separately for during treatment (see top portion of Figure 2) and post-treatment models (top portion of Figure 1), in order to simplify the presentation of results. Both models provided a reasonable fit to the observed data (during treatment: RMSEA = 0.06 (90% Confidence Interval 0.058–0.062), CFI = 0.90; post-treatment: RMSEA = 0.06 (90% Confidence Interval 0.058–0.067), CFI = 0.95) and standardized factor loadings exceeded 0.61 for all indicators, with the majority of factor loadings exceeding 0.88.
The first goal of the bivariate difference score modeling was to determine the dynamic bivariate associations between percent heavy drinking days (PHD) and proximal risk across time during treatment.4 Several alternative models were estimated with varying constraints on the γ parameters, which measure the degree of dynamic bivariate change. The first model (baseline model) estimated PHD and proximal risk as independent of one another (i.e., γ parameters were fixed to zero). This fully constrained model provided a reasonable fit to the data (RMSEA = 0.047 (90% Confidence Interval 0.046–0.049), CFI = 0.93) and served as the baseline model. The second model (model 2) estimated PHD predicting the change in proximal risk, with no bidirectional association of proximal risk predicting change in PHD (i.e., the γp parameters were fixed to zero), such that the only cross-lag association was from the PHD factors predicting the change in proximal risk. This model fit worse than the baseline model (χ2 difference test (1) = −2.96, p = 1.00). In the third model (model 3) estimated proximal risk predicting the change in PHD, with no bidirectional association of PHD predicting change in proximal risk (i.e., the γd parameters were fixed to zero), such that only the cross-lag association from proximal risk predicting the change in PHD was estimated. This model fit significantly better than the baseline model based on a χ2 difference test (χ2 (1) = 56.79, p < 0.0001). The final model (bidirectional model) estimated the bidirectional association between changes in PHD and changes in proximal risk (i.e., the γd and γp parameters were jointly estimated), such that PHD predicted changes in proximal risk and proximal risk predicted changes in PHD. The bidirectional model fit better than the baseline model (χ2 (2) = 60.54, p < 0.0001), but was not a significant improvement over model 3 (χ2 (1) = 3.75, p = 0.053) in which only the cross-lag association from proximal risk to changes in PHD was estimated. Thus, model 3 (shown in Figure 2) with RMSEA = 0.04 (90% Confidence Interval 0.045–0.049) and CFI = 0.93 was retained as the optimal model.
Table 2 provides a summary of unstandardized coefficients from the final bivariate latent difference score model (model 3, Figure 2). Three parameters are of particular interest: the constant change parameters (which represent the how much the behavior fluctuates over time), the proportional change (which represents the effect of each variable on itself), and the bivariate associations (which represent the associations between drinking and proximal risk). The estimates for the constant change parameters (βd and βp) indicate that changes in percent heavy drinking days tended to increase over time (i.e., heavy drinking was less stable over time) and changes in proximal risk decreased over time (i.e., became more stable) during treatment (PHD slope = 0.33 (SE = 0.04), p < 0.001; proximal slope = −0.43 (SE = 0.06), p < 0.001), with significant individual variation around the slope. The negative proportional change coefficient for PHD (αd = −1.50 (SE = 0.04), p < 0.001) and proximal risk (αp= −0.49 (SE = 0.05), p < 0.001) indicated more frequent heavy drinking and higher levels of proximal risk were associated with a decrease in heavy drinking frequency and proximal risk at a subsequent time-point, respectively, which could be an indication of regression to the mean for both PHD and proximal risk across each adjacent time-point.
Bivariate associations between proximal risk and PHD were estimated by the covariance of the growth factors and by the cross-lag regressions of PHD difference scores on the latent proximal factors. First, the covariance between the growth factors indicated that individuals who had higher PHD initially also had greater increases in proximal risk over time (B (SE) = 0.13 (0.02), p< 0.001). There was also a strong positive association between the constant change in proximal risk and constant change in PHD over time (B (SE) = 0.20 (0.04), p < 0.001), thus individuals who reported increased proximal risk over time had greater increases in PHD over time. The cross-lag coefficients are interpreted within the context of the constant change coefficients. Given that there is a general trend for an increase in PHD over time, the significant and positive cross-lag coefficient (γp = 0.01 (SE = 0.003), p < 0.001) indicates that individuals with higher levels of proximal risk at a prior time point are likely to have more of an increase in PHD at a subsequent time point. In other words, higher levels of proximal risk predicted greater increases in heavy drinking over time.
The association between distal risk, proximal risk and heavy drinking was estimated by incorporating five distal risk factors (marital status, baseline self-efficacy, alcohol dependence severity, prior treatment history, and baseline psychiatric symptoms) as predictors of the PHD and proximal risk growth factors. As reported in Table 2, self-efficacy and psychiatric severity were significantly associated with the level of proximal risk, such that individuals with higher levels of self-efficacy and fewer psychiatric symptoms had lower levels of proximal risk. Treatment history and psychiatric severity also predicted the constant change in proximal risk. Individuals who had prior treatment episodes and those with more psychiatric symptoms at baseline also reported a greater increase in proximal risk over time. Marital status and self-efficacy were also associated with the level of PHD, such that being married and having higher self-efficacy at baseline were associated with fewer heavy drinking days. The distal risk factors were not significantly associated with changes in PHD over time.
Overall, the model with proximal and distal risks as predictors of dynamic changes in PHD over time explained 64% to 75% of the variance in PHD at any given time-point, yet only 3% of that variance was uniquely explained by the combination of distal and proximal risks. Squared partial correlation coefficients for the linear association between the proximal/distal risk factors and PHD indicated that proximal risk explained less than 7% of the variance in PHD at each time point and distal risk explained less than 1% of the variance in PHD at each time point.
To evaluate whether changes in proximal risk mediated the association between each of the distal risks and PHD during treatment the indirect effect of each of the distal risk factors in the prediction of PHD slope were estimated based on the significant covariate effects described above. Specifically, the indirect effects of the two significant predictors of proximal risk slope, treatment history and psychiatric severity, in the prediction of PHD slope were estimated. Mediation analyses, using the product of coefficients approach, indicated that the change in proximal risk significantly mediated the association between treatment history and PHD slope (B (SE) = 0.004 (0.002), p = 0.02; 95% CI: 0.001–0.007) and also mediated the association between psychiatric symptoms and PHD slope (B (SE) = 0.001 (0.0001), p = 0.01; 95% CI: 0.0001–0.001). Estimates of effect size for multiple mediator models are not fully developed (MacKinnon, 2008), yet the proximity to 0.0 in the 95% confidence intervals suggest that these are very small mediated effects.
Following the same modeling steps described above for the during treatment models, results of the post-treatment models indicated that the bidirectional model fit significantly better than the baseline model (χ2 (2) = 45.51, p < 0.0001) and was also a significant improvement over models 2 and 3 in which only one set of cross-lag associations were estimated (bidirectional vs. model 2 χ2 (1) = 34.29, p < 0.0001; bidirectional vs. model 3 χ2 (2) = 24.69, p < 0.0001). Thus, the bidirectional latent difference score model with RMSEA = 0.067 (90% Confidence Interval 0.065–0.071) and CFI = 0.93 was retained as the optimal model.
Table 3 provides a summary of unstandardized coefficients from the final bidirectional latent difference score model. The constant change parameters (βd and βp) indicated that both percent heavy drinking days and proximal risk remained relatively stable over time following treatment (PHD slope = −0.24 (SE = 0.33), p = 0.46; proximal slope = 1.86 (SE = 1.24), p = 0.13), with significant individual variation around each slope. Preliminary univariate model testing indicated that constraining the proportional change coefficients following treatment to equality across time resulted in a significant decrement in model fit, thus αd and αp were allowed to vary across time. As seen in Table 3, the proportional change coefficient for PHD was positive at each time point (αd1 = 0.42 (SE = 0.10), p < 0.001; αd2 = 0.25 (SE = 0.09), p < 0.001; αd3 = 0.09 (SE = 0.07), p = 0.20) indicating that frequency of heavy drinking at the end of treatment and 10 weeks after treatment predicted an increase in heavy drinking frequency 10 and 26 weeks after treatment, respectively. Consistent with the during treatment models, the negative proportional change coefficients for proximal risk (αp1= −0.52 (SE = 0.09), p < 0.001; αp2= −0.62 (SE = 0.09), p < 0.001; αp3= −0.51 (SE = 0.09), p < 0.001) indicated individuals with higher levels of proximal risk at one time point endorsed less proximal risk at subsequent time-points.
Again, the bivariate associations between proximal risk and PHD were estimated by the covariance of the growth factors and by the cross-lag regressions of the PHD and proximal risk factors. First, as shown in Figure 3, the covariance between the growth factors indicated that individuals who had higher PHD initially also had greater increases in proximal risk over time (B (SE) = 9.96 (1.14), p< 0.001) and individuals who reported increased proximal risk over time had greater increases in PHD over time (B (SE) = 2.67 (0.92), p = 0.004). Given that the constant change slope of PHD was negative in the post-treatment models, the negative cross-lag coefficient for the effect of proximal risk on the difference in PHD at a subsequent time point (γp = −0.10 (SE = 0.02), p < 0.001) has a similar interpretation as the results from the during treatment models: individuals with higher levels of proximal risk at a prior time point reported less of a reduction in PHD at a subsequent time point. Likewise, the cross-lag coefficient for proximal risk regressed on PHD at a prior time point (γd = −0.44 (SE = 0.12), p < 0.001) indicated that greater drinking frequency at a prior time point is associated with less of a reduction in proximal risk at a subsequent time point.
The association between distal risk, proximal risk and heavy drinking was estimated by incorporating the distal risk factors as predictors of the PHD and proximal risk growth factors following treatment. As reported in Table 3, treatment history, marital status, and psychiatric severity were significantly associated with the level of proximal risk, such that individuals without a treatment history, who were married, and those with fewer psychiatric symptoms reported lower levels of proximal risk. All five covariates were significantly associated with the change in proximal risk over time: the lack of a treatment history, being married, fewer psychiatric symptoms, lower alcohol dependence, and higher self-efficacy predicted a greater decrease in proximal risk over time. Treatment history, marital status and self-efficacy were also associated with the level of PHD, such that those without a treatment history, being married and higher self-efficacy at baseline had fewer heavy drinking days. Being married and fewer psychiatric symptoms also predicted a greater decrease in PHD over time. Overall, the model with proximal and distal risks as predictors of dynamic changes in PHD over time explained 79% to 81% of the variance in PHD at any given time-point, yet only 6% of that variance was uniquely explained by the combination of distal and proximal risks. Based on the squared partial correlation coefficients for the linear association between proximal risk and PHD, proximal risk explained less than 5% of the variance in PHD at each time point and distal risk explained less than 1% of the variance in PHD at each time point.
The indirect effects of all five distal risk factors in the prediction of PHD slope via the proximal risk slope were estimated. The mediation results were entirely consistent with the during treatment models: change in proximal risk significantly mediated the association between treatment history and PHD slope (B (SE) = 1.78 (0.90), p = 0.048; 95% CI: 0.015–3.54) and also mediated the association between psychiatric symptoms and PHD slope (B (SE) = 0.43 (0.20), p = 0.03; 95% CI: 0.04–0.81). No other mediation effects were significant.
The current study evaluated the dynamic associations between proximal and distal risk factors in the prediction of heavy drinking in a sample of individuals who received 16 weeks of treatment for alcohol dependence. Importantly the results of the analyses were largely consistent across time, suggesting a common system that is operating during treatment and following treatment. Likewise, model parameters were also consistent when analyses were conducted separately for those who never quit heavy drinking (and thus did not “lapse”) and by the timing of the first heavy drinking lapse. The present study utilized dynamic latent difference score models, which provided the opportunity to examine bivariate change in proximal risk and heavy drinking across time. The main findings from the analyses can be derived by reviewing the coefficients in Tables 2 and and3.3. Results of the growth model portion of the during treatment and post-treatment latent difference score models indicated that, on average, the frequency of heavy drinking increased over time, whereas proximal risks tended to decrease over time during treatment. Both frequency of heavy drinking changes and proximal risk changes remained relatively stable following treatment. Although it is important to note there was significant individual variability in the changes in heavy drinking and proximal risk both during and following treatment, suggesting that the average patterns were not representative of the entire sample. The significant positive correlation between the random slopes of heavy drinking and proximal risk (Figure 3) indicated that most individuals who experienced an increase in heavy drinking reported an increase in proximal risk.
Bivariate associations indicated that levels of heavy drinking during and following treatment predicted subsequent changes in proximal risk over time, such that individuals who engaged in more frequent heavy drinking tended to report a significant increase in proximal risk over time. Tests of the cross-lagged associations between proximal risk and heavy drinking provided support for the hypothesis that changes in heavy drinking frequency during and following treatment can be predicted from antecedent levels of proximal risk. Likewise, heavy drinking frequency at any given time point following treatment was associated with subsequent changes in proximal risk.
The distal risk factors were less strongly associated with heavy drinking than the proximal risk factors. Baseline psychiatric symptoms predicted the constant change in heavy drinking during treatment and following treatment, marital status predicted the initial level of heavy drinking and the constant change in heavy drinking following treatment, self-efficacy predicted initial level of heavy drinking during and following treatment, and treatment history was associated with initial level of heavy drinking following treatment. It is important to note that all of the selected distal risk factors were significantly associated with proximal risk at some point during treatment and/or following treatment, which is consistent with the idea that distal risks do impact the overall system. Formal mediation analyses provided initial evidence that both treatment history and psychiatric symptoms were associated with the constant change in percent heavy drinking days via changes in proximal risk. One interpretation of these findings is that individuals with a prior treatment history and greater psychiatric symptoms are at greater risk for more frequent heavy drinking if they experience increased proximal risk over time. Thus, clients who initiate treatment with higher levels of psychiatric symptoms and prior treatment episodes might be especially vulnerable to heavy drinking following an increase in proximal risk. Psychoeducation about their higher level of risk could better prepare them to seek additional help at times of increased stress, craving, or negative mood.
Yet, despite the number of statistically significant bivariate associations it is important to note that the magnitude of the effects were quite small, with less than 7% of the variance in drinking outcomes uniquely explained by the combination of distal and proximal risks tested in the current study. The latent dynamic difference score model explained a considerable amount of variance in drinking outcomes, but nearly all of the variance was explained by the drinking indicators and the proximal and distal risk factors added very little to the overall fit of the model. Thus, the clinical significance of the proximal and distal risk factors included in the current study is questionable.
The dynamic model of relapse builds upon previous relapse models by highlighting temporal dynamics and complex interactions between distal and proximal risk factors as part of a multi-dimensional system. Clinically, the dynamic model provides a broader framework for understanding the relapse process and encourages flexibility in relapse prevention planning by tailoring relapse prevention plans to the specific needs of each client. The results from the current study provide support for the hypothesis that alcohol lapses are dynamic and that there are significant bidirectional associations between proximal risk factors and alcohol use over time, such that individuals with higher levels of proximal risk tended to have an increase in heavy drinking and increases in heavy drinking were associated with increased proximal risk at subsequent assessments. Yet, the current study did not provide a full test of the dynamic model of relapse and several crucial proximal risk factors, including coping and momentary self-efficacy, were not considered due to them not being measured by the COMBINE study. Moreover, it is interesting that the small effects of proximal risk observed in the current study were based on more internal factors (e.g., craving, mood, stress), whereas the dynamic model of relapse advocates contextual factors (e.g., setting, effective coping) are crucial to predicting a lapse in the moment.
The COMBINE study was extremely well conducted and provided a wealth of information for the current analyses, including daily drinking measures, multiple measurements of risk factors during and following treatment, and an array of candidate distal risk factors at baseline, yet the data were still not ideal for testing the dynamic model of relapse for many reasons. First, the dynamic model of relapse proposes that proximal risk factors are phasic processes – meaning that they exhibit momentary changes and rapidly adapt to environmental stimuli. In the COMBINE study the proximal risk factors were assessed at weekly and monthly intervals. Although this is an improvement over prior studies that have measured proximal risks at 3- or 6-month intervals it still does not provide the moment-to-moment data necessary to test craving, mood and stress as phasic predictors of heavy drinking. Likewise, although daily drinking data were collected in COMBINE using the timeline follow-back methodology, it is important to note that these data were still reported retrospectively over weekly and monthly intervals. A more sensitive test of the dynamic model of relapse would come from drinking data collected on a daily basis.
There are several risk factors hypothesized by the dynamic model that were either excluded from the COMBINE assessment battery or were not assessed at multiple time periods. Most notably, coping behavior, interpersonal stressors, and contextual risk, which have been proposed as proximal risk factors that immediately precede a lapse (Hunter-Reel et al., 2009; Miller et al., 1996; Witkiewitz & Marlatt, 2004) were not included in the COMBINE assessments. Interpersonal influences, outcome expectancies, and motivation, were included in the COMBINE battery, but were not assessed repeatedly. Given that all of these variables have been hypothesized as potentially phasic processes, it is important for them to be included in the model at multiple time points. All of these limitations could be addressed with an intensive longitudinal research design that measured multiple factors repeatedly over multiple days. Momentary assessment (Stone & Shiffman, 1994), either via an electronic device or interactive voice response methodology, could provide the data necessary to fully test the dynamic model of relapse. Ideally, assessments of coping, interpersonal stress, self-efficacy, craving, mood, and other proximal factors could be collected multiple times per day over the course of several months, and combined with a thorough pre-treatment assessment battery of distal risk factors. Future research with a data set that includes multiple measures of risk factors over multiple days could also take advantage of innovative modeling tools that were designed for estimating nonlinear time-varying dynamics (Chow, Ferrer, & Nesselroade, 2007; Walls & Schafer, 2006).
Finally, the variances of the constant change parameters were significant across all models suggesting that there is significant individual heterogeneity in the change in proximal risk and percent heavy drinking days over time. This finding is consistent with prior research (see Witkiewitz et al., 2007) that has found significant individual variability in drinking trajectories following treatment. This prior research utilized a type of longitudinal mixture model, growth mixture modeling, to explicitly estimate the degree of heterogeneity as separate latent classes of growth trajectories. Future research should explore the utility of dynamic mixture models, an extension of latent difference score modeling that allows for estimating multiple classes of bivariate change (McArdle & Grimm, 2010), in testing the bivariate association between proximal risk and heavy drinking over time. For example, it could be the case that the dynamic prediction of heavy drinking from proximal risk factors is not consistent across all individuals, such that heavy drinking by certain individuals is not influenced by proximal risk factors whereas for other individuals heavy drinking is closely linked to proximal risk. One could then attempt to characterize those who are more or less vulnerable to proximal risk by baseline risk factors, which would provide very useful clinical information about who is most at risk to relapse in response to proximal risk.
Despite the limitations, the current study provided initial empirical evidence for a portion of the dynamic model of relapse, including the hypotheses that risk factors are inter-related and that changes in heavy drinking are reliably influenced by both proximal and distal risk. Yet, the magnitude of the effects of the proximal and distal risk factors included in the current study were rather small, suggesting that other factors might be critically important to include in a dynamic model of relapse. Replication of the current findings in a new dataset and re-estimating the current models with intensive longitudinal data would provide further support for the dynamic model and could potentially lead to the development of dynamic relapse prevention plans.
This research was supported by National Institute on Alcohol Abuse and Alcoholism grant R21AA017137. The authors thank members of the COMBINE Study Research Group for their assistance in facilitating access to COMBINE study data.
Publisher's Disclaimer: The following manuscript is the final accepted manuscript. It has not been subjected to the final copyediting, fact-checking, and proofreading required for formal publication. It is not the definitive, publisher-authenticated version. The American Psychological Association and its Council of Editors disclaim any responsibility or liabilities for errors or omissions of this manuscript version, any version derived from this manuscript by NIH, or other third parties. The published version is available at www.apa.org/pubs/journals/adb
1The parameters derived from the models described below were examined separately by treatment condition and few discernible differences in parameter estimates existed, thus all models are described for the full sample (n =1,383).
2For the during treatment models we first estimated models in which the constant change parameters were freely estimated to reflect the unequal intervals of measurement. This model provided a slightly better fit to the observed data, but would not converge when distal risks were included as covariates. Substantive results of the unconditional model were identical to the model in which the constant change parameter was constrained to 1.0, thus the constrained model was retained in subsequent analyses.
3Percent of heavy drinking days (PHD) reported at the time of each proximal risk assessment point during treatment was used as the drinking outcome for several reasons. First, it is important to note that weekly drinking in the seven days prior to the proximal assessment was computed from the timeline follow-back, which was administered at the same time as the proximal risk assessment. The PHD for the week during and after a proximal risk assessment was computed from the timeline follow-back administered at the time of the subsequent proximal risk assessment (up to 30 days later). The of accuracy of the timeline follow-back over longer intervals of time might be reduced (Hoeppner, Stout, Jackson, & Barnett, 2010), thus the assessment of drinking at the time of the proximal risk assessment could be more precise in reflecting current drinking than the assessment of drinking 30-days following a proximal risk assessment. Likewise, retrospective recall has been shown to be based on experiences in the present (Shiffman, Hufford, Hickcox, Paty, Gnys, & Kassel, 1997), thus it was expected that drinking rates just prior to the proximal risk assessment would be the best approximation of the concurrent drinking and proximal risk.
4The during treatment models were initially estimated separately for two groups defined by when they first had a heavy drinking lapse, excluding those who never abstained (defined as those who had a first heavy drinking episode during the 1st week of treatment) and excluding those who never had a heavy drinking episode during treatment. The first group was defined by those individuals who had a heavy drinking lapse during the first half of treatment and the second group had their first heavy drinking lapse during the second half of treatment. Model parameters were consistent across groups and remained consistent when the excluded participants (those who never stopped heavy drinking and those who never drank heavily during treatment) were included. Thus, the remaining models are based on the full sample (n = 1383).