Change Talk Subtypes as Predictors of Alcohol Use Following Brief Motivational Intervention

Abstract

Objective:

To examine the relative importance of client change language subtypes as predictors of alcohol use following motivational interviewing (MI).

Method:

Participants were 164 heavy drinkers (57.3% female, mean age = 28.5 years, 13.4% Hispanic/Latinx, 82.9% White) recruited during an emergency department visit who received MI for alcohol and HIV/sexual risk in a randomized controlled trial. MI sessions were coded with the Motivational Interviewing Skill Code (MISC) and the Generalized Behavioral Intervention Analysis System (GBIAS). Variable importance analyses used targeted maximum likelihood estimation to rank order change language subtypes defined by these systems as predictors of alcohol use over 9 months of follow-up.

Results:

Among GBIAS change language subtypes, higher sustain talk around change planning was ranked the most important predictor of drinks per week (b = −5.57, 95% CI [−8.11, −3.02]) and heavy drinking days (b = −2.07, 95% CI [−3.17, −0.98]); this talk reflected (a) rejection of alcohol abstinence as a desired change goal, (b) rejection of specific change strategies, or (c) discussion of anticipated challenges in changing drinking. Among MISC change language subtypes, higher sustain talk around taking steps—reflecting recent escalations in drinking described by a small minority of participants—was ranked the most important predictor of drinks per week (b = 22.71, 95% CI [20.29, 25.13]) and heavy drinking days (b = −2.45, 95% CI [1.68, 3.21]).

Conclusions:

Results challenge the assumption that all sustain talk during MI is a negative prognostic indicator and highlight the importance of the context in which change language emerges.

The tenets of Motivational Interviewing (MI) posit that increasing change talk (CT; client statements supporting a change in behavior) and reducing sustain talk (ST; client statements that support maintaining a behavior) during an MI session are key processes through which MI effects behavior change (Miller & Rose, 2009). Research on the association between client change language and subsequent outcomes, however, has been decidedly mixed. Meta-analyses indicate that the amount of CT in an MI session does not predict client outcomes, whereas greater ST predicts worse outcomes, and a greater proportion of CT relative to ST predicts better outcomes (Magill et al., 2018; Pace, Dembe, Soma, Baldwin, Atkins & Imel, 2017. In these meta-analyses, the study samples varied with respect to treatment versus non-treatment seeking status. However, this sample-based characteristic was tested as a moderator by Magill and colleagues (2018), and while the sample was majority non-treatment seeking, this factor did not predict substantive differences in effect magnitude. In individual studies, there is some evidence to suggest that change talk is a more reliable predictor in adult, treatment seeking samples (Moyers et al., 2009, Houck et al., 2018) and sustain talk is a more reliable predictor among young adult, non-treatment seeking samples (Apodaca et al., 2014). Therefore, consideration of context, approaches that use a fine-grained analysis, and examination of client language subtypes may yield better prediction of outcomes and offer more insights into mechanisms of change.

The Motivational Interviewing Skills Code (MISC 2.5; Houck, Moyers, Miller, Glynn, & Hallgren, 2010) provides one approach to differentiating change language subtypes. In addition to coding the valence of client utterances (i.e., unique units of meaning) as either CT, ST, or Follow Neutral, the MISC also codes change language subtypes using the DARN-C system, which includes codes for Desire, Ability, Reasons, Need, or Commitment to change with additional codes for Taking Steps and Other change language. Some research has indicated that having more Commitment language especially during the latter stages of an MI session is particularly predictive of greater reductions in alcohol use following MI (Amrhein, Miller, Yahne, Palmer, & Fulcher, 2003). However, these findings have been inconsistent across studies with Ability and Desire showing effects in some studies (Gaume, Bertholet, Faouzi, Gmel, & Daeppen, 2013; Gaume, Gmel, & Daeppen, 2008) and Reasons and Other change talk identified as important in a recent meta-analysis (Magill et al., 2019).

The Generalized Behavioral Intervention Analysis System (GBIAS; Kahler et al., 2016) provides an alternative approach to coding subtypes of client change language. The GBIAS codes the topics on which utterances focus without specifying the motivational valence of the utterance. In addition to codes for biomedical topics, logistics, psychosocial topics, and socializing, the GBIAS includes seven topic codes for utterances focused on the target behavior (TB) for the intervention, which in this case involved either alcohol use or risky sexual behavior: Patterns and Contexts of the TB, Benefits of the TB, Drawbacks of the TB, Benefits of TB Change, Drawbacks of TB Change, Barriers and Facilitators to TB change, and TB Change Planning (see Method section for details on code definitions). When GBIAS topic codes are combined with MISC coding, each utterance can be understood according to (a) what aspect of a TB the utterance was focused on and (b) whether that utterance was supportive of TB change, supportive of sustaining the TB, or neutral. Although certain topics are considerably more likely to involve CT (e.g., Drawbacks of the TB), any utterance on a TB-related topic can be coded as either CT or ST (Kahler 2016). For example, when discussing Drawbacks of the TB, ST would involve a client stating that a given negative consequence of a TB does not apply to them or is unimportant to them. The combination of MISC and GBIAS coding systems can allow for nuanced examination of whether CT or ST around certain topics is especially important in predicting client behavior change.

Study Aims

The purpose of the present study was to examine the relative importance of specific subtypes of CT and ST, based upon topic codes from the GBIAS, as predictors of alcohol use outcomes following MI. We also sought to add to the literature on process coding in MI by conducting parallel analyses examining subtypes of CT and ST as coded by the MISC. Data for these analyses came from a clinical trial addressing drinking and sexual behavior that increases HIV risk in emergency department patients where a single, 60-minute MI session was delivered, and was found to be associated with reduced alcohol use and sex risk behavior relative to brief advice (Monti, Mastroleo, Barnett, Colby, Kahler, & Operario, 2016). We previously published on the development and validation of the GBIAS within this study population (Kahler et al., 2016). Analyses of MISC-coded data from this trial have found that overall, neither the total amount (i.e., session level average) nor the growth (i.e., session-level slope) in CT or ST during the MI session predicted alcohol use at follow-up (Janssen et al., 2019). Additionally, certain MI-consistent counseling techniques, such as complex reflections and open-ended questions, were found to be more strongly associated with participants moving from ST to CT when compared to techniques like simple reflections or paraphrasing reflections (Laws et al., 2018).

In the current study, we conducted an exploratory analysis that examined a relatively large number of candidate variables—14 CT/ST: Topic subtypes and 14 CT/ST: DARN-C subtypes—as predictors of drinking outcome using a previously described targeted maximum likelihood estimation procedure (TMLE; Gruber, 2015) that provides a particularly robust method for examining variable importance (Hubbard et al., 2013). This procedure allowed us to rank order each candidate variable according to how strongly it predicted our primary outcomes of interest: drinks per week and number of heavy drinking days over 9 months of follow-up. We then complemented these exploratory analyses with qualitative description of the types of client language that characterized the most important candidate variables.

Methods

Transparency and Openness

We report how we determined our sample size, all data exclusions (if any), all manipulations, and all measures in the study, and we follow Journal Article Reporting Standards for quantitative research in psychology (Applebaum, Cooper, Kline, Mayo-Wilson, Nezu, & Rao, 2018). All data, analysis code, and research materials are available upon request to the corresponding author. Data were analyzed using SPSS v26 (IBM Corp., 2019) and R version 3.6.0 (R Core Team, 2013); see further details provided below. This study’s design and its analysis were not pre-registered.

Parent Trial Procedure

Data for these analyses came from a randomized controlled trial testing the efficacy of MI compared to brief advice for reducing heavy alcohol use and condomless sex among patients in the emergency department (Monti et al., 2016). Participants were included in the trial if they (a) met criteria for hazardous drinking (total score ≥ 8 for males; ≥ 6 for females on the Alcohol Use Disorders Identification Test, AUDIT; Saunders, Asland, Babor, De la Fuente, & Grant, 1993) or endorsed at least one episode of heavy drinking (≥ 5 drinks for males; ≥ 4 drinks for females) in the past three months; and (b) reported engaging in condomless sex, consuming alcohol/other drugs prior to or during sex, or engaging sex with a non-steady partner or being with a steady partner where monogamy is questioned or unknown. For further details on the parent trial procedures and outcomes, see (Monti et al., 2016).

Of the 184 participants who were randomized to receive MI, 169 completed the session, and 164 of the sessions were audio-recorded and transcribed for coding and analysis. Sample size for the present study was not determined through a power analysis; instead, all available data were used. In the parent trial, MI recipients, compared to brief advice recipients, reported fewer follow-up heavy drinking days, drinks per week, and days of condomless sex across 9 months of follow-up. Given that the amount of discussion of alcohol use in MI sessions was far greater than discussion of sex risk (Kahler et al., 2016), we focused the present paper on alcohol use discussions and outcomes rather than on sex risk. University and hospital Institutional Review Boards approved all study procedures.

Motivational Intervention for Alcohol and Sexual Risk Reduction

The dual-target MI session was based on the principles of Miller and Rollnick (2002) and was conducted by six doctoral- and masters-level providers. Providers received 20 hours of training and participated in weekly group supervision meetings, including audio session review. The MI session involved exploration of the pros and cons of alcohol use and sexual risk behavior, personalized feedback on alcohol consumption and consequences, normative comparisons, and protective behaviors to reduce risk. Consistent with an MI framework, the intervention allowed the client to determine which risk behaviors were of greatest concern and what, if any, change goals, such as abstinence from drinking or reducing drinking, were most appealing to them. If clients agreed, a change plan for reducing alcohol and/or sexual risk was completed.

Measurement

Participants completed an in-person baseline assessment and follow-up assessments at 3, 6, and 9 months post intervention (83%, 86%, and 88% retention rates, respectively). Alcohol use outcomes were assessed with the Time-Line Follow-Back for alcohol use and sexual behavior (TLFB-SS; Carey, Carey, Maisto, Gordon & Weinhardt, 2001). The TLFB-SS is a structured, calendar-aided interview adapted from the Alcohol TLFB (Sobell & Sobell, 2003; Sobell, Sobell, Klajner, Pavan, & Basian, 1986). The measure was used to assess behavior over the prior 30 days. In the current analyses alcohol use at baseline, 3, 6, and 9 months was assessed based on (a) number of heavy drinking days (≥ 4/5 drinks for women/men) and (b) log-transformed average number of alcoholic drinks consumed per week. At baseline, we used a contemplation ladder (0 = I never think about drinking less to 10 = My drinking has changed; Hogue, Dauber, & Morgenstern, 2010) to assess readiness to change drinking, which was used as a covariate in all analyses.

Coding Procedures

Coding was conducted of de-identified and transcribed MI sessions in three passes, each conducted by a separate rater. First, a rater segmented transcripts into distinct speech acts, defined as an utterance that performs a specific function in communication such as representing a fact, expressing a feeling, or asking for information (Austin, 1975; Searle, 1969). A second rater assigned GBIAS codes. A third rater assigned MISC codes. Twelve raters were trained in and conducted each of these aspects of the coding process.

Rater Training and Coding Reliability Procedures.

Raters received approximately 60 hours of training on each of the GBIAS and MISC systems including (a) didactic overview, (b) group coding practice with corrective feedback, and (c) individual coding practice with group corrective feedback. Inter-rater reliability and agreement were assessed throughout data collection. After every fifth session was coded, a segment of no more than 10 pages of transcript containing at least 300 utterances (range = 300 to 700) was randomly chosen from the pool of completed interviews and assigned to a second rater for reliability coding. Reliability and agreement statistics were computed, and disagreements were discussed in biweekly consultation meetings to identify decision rules about future coding of similar utterances.

MISC Coding

The Motivational Interviewing Skill Code (MISC 2.5; Houck et al., 2010) was used to code client change language. Client language was classified using the MISC as either CT or ST and as reflecting one of seven subtypes: Desire, Ability, Reasons, Need, Commitment, Taking Steps, and Other. Thus, there were a total of 14 different combinations of MISC codes that could be applied to a given utterance. For ease of presentation, we name the MISC code valence first (CT or ST) and then the specific subcode, e.g., CT: Desire; ST: Desire. Statements were coded as Follow Neutral if they did not focus on a TB or have clear motivational valence.

GBIAS Coding

The GBIAS assigned each client and counselor utterance a topic code as described in depth in Kahler et al. (2016). The GBIAS uses a hierarchical topic coding system including broad categories of biomedical, logistics, socializing, and psychosocial topics with decimal-level codes providing additional specificity. For this paper, those codes were aggregated into “other” non-TB focused discussion.

Whenever a TB was discussed, topic codes within the category of Intervention Targeted Behaviors were used, with a separate code specifying whether the target behavior was alcohol or sex. We limited our analysis to utterances on the TB of alcohol use and the following seven topic codes. The Patterns and Contexts topic code captured conversations about a client’s drinking pattern (how much and how often) including the contexts in which the behavior occurs, positive and negative global judgments of one’s own drinking, and comparison of one’s drinking to others’ drinking. Benefits of the TB and Drawbacks of the TB codes captured respectively discussions of positive and negative consequences one has or could experience from drinking. Benefits of TB Change and Drawbacks of TB Change captured respectively discussion of benefits and negative effects that one might experience if cutting down on or stopping drinking. A topic code for Change Planning captured discussions of possible drinking goals—which could include abstinence, change in frequency or quantity of drinking, or harm reduction strategies—and intermediate change steps towards meeting those goals. A topic code for Barriers and Facilitators of Change captured conversation about personal and social/environmental factors that would impede or facilitate efforts to change drinking. Each utterance receiving a topic code also received a MISC code of either CT, ST, or Follow Neutral. In the present analyses, we focused on utterances receiving both a TB topic code and a MISC code of either CT or ST. Therefore, there were 14 total GBIAS-derived change language subtypes, e.g., CT: Patterns and Contexts; ST: Patterns and Contexts.

Coding Reliability

Thirty-nine tapes were randomly selected for double coding with the GBIAS and the MISC. The mean Cohen’s kappa for coding what target behavior (alcohol, sex, or neither) was under discussion was .90, indicating substantial agreement between raters (Landis & Koch, 1977). For count indicators in the GBIAS and MISC, we calculated one-way intraclass correlation coefficients (ICCs). Interpretation benchmarks for the interrater agreement indicated by these coefficients are as follows: poor <.40, fair .40 to .59, good .60 and .74, and excellent > .74 (Cicchetti, 1994). The segments selected for double coding for the GBIAS and MISC were independent of one another, and therefore we could not calculate ICCs for GBIAS topic codes according to whether they were CT vs. ST. Reliability estimates (ICCSs) and frequencies for each code in the selected segments are provided in Supplemental Table 1. ICCs were above .70 for all but one GBIAS variable, Drawbacks of TB Change (ICC = 18). For the MISC, ICCs were above .70 for eight of fourteen variables. Two subtypes, CT: Ability and ST: Other had ICCs in the .40 to .50 range, indicating fair reliability. Further, ICCs could not be calculated for four of the subtypes (ST: Desire, ST: Need, ST: Commitment, ST: Taking Steps), each of which was very rarely coded within the chosen segment by either the first or second rater (i.e., 75^th percentile = 0 with a maximum of 0 to 3).

Analytic Strategy

Descriptive Analysis

First, we calculated univariable statistics for client sociodemographic information, baseline and follow-up alcohol use, and the frequency of utterances in each of the MISC-derived change language subtypes and each of the GBIAS topic-derived subtypes. We also calculated the tetrachoric correlations (Kirk, 1973) between the dichotomized MISC subtypes and GBIAS subtypes scores using the R package psych.

Variable Importance Analysis

The current study employed an analytic method, which has been previously described in the literature (Hubbard et al., 2013), to rank order subtypes of CT and ST (i.e., the 14 CT/ST: Topic subtypes and the 14 CT/ST: DARN-C subtypes) based on how strongly they predicted the primary alcohol use outcomes of drinks per week and of number of heavy drinking days during follow-up. These analyses used dichotomous MISC and GBIAS measures where numbers of subtype utterances were calculated for each participant, and then values of ‘higher’ or ‘lower’ were assigned based on a median split. We used the counts of utterances on a given subtype, rather than the proportion of utterances on that subtype, because we believed a priori that the content of discussion rather than the proportion of discussion is most relevant clinically for understanding outcomes; for example, whether one discusses making a plan for change could be more important than the proportion of the session dedicated to that topic. Proportion indicators may be ideal when the balance of ambivalence is the conceptual emphasis, whereas count indicators may ideal when specific content and its predictive value is the conceptual emphasis. Dichotomized variables were used because this allowed us to make no assumption about a linear relationship between change subtype frequencies and outcomes; this is especially important in the context of zero-inflated and skewed data. Because some codes were quite rare and sample size was modest, it was not feasible to make multiple categories for each variable beyond a dichotomy. The use of dichotomies allows us to interpret estimates for each variable simply as the predicted difference in the outcome associated with being higher versus lower on a change language subtype. The analytic model was developed with an ensemble machine learning algorithm called SuperLearner (v2.0–26) (van der Laan & Rose, 2011; van der Laan, Polley, & Hubbard, 2007). Targeted maximum likelihood estimation (TMLE; Gruber, 2015) was used to estimate the relationships between each measure and the outcomes of interest.

TMLE is a two-stage estimation procedure. In stage 1, SuperLearner (van der Laan et al., 2007) was used to develop a weighted combination result model for the outcome, E(Y | X), where Y is the outcome and X is the set of all change language subtypes and the following covariates: gender, the baseline value of the outcome, baseline value of the alcohol contemplation ladder, total number of client utterances coded in the session (given that greater amount of client talk may increase the likelihood of being classified as high on any given subtype and represent a confound) and time (two dummy-coded variables for month 6 and month 9 with month 3 as the reference). The weighted combination arises from weights assigned by the SuperLearner algorithm to minimize the cross-validated mean squared error of the predictions. This leads to an adaptive solution based on a combination of the most effective learning algorithms. In our analysis, the model library included six algorithms: 1) generalized linear models (GLM) using main terms only; 2) GLM using forward and backward selection to select from all possible main and second order terms; 3) lasso penalized regression (Friedman, Hasti, & Tibshirani, 2010; Friedman et al., 2021); 4) Bayesian additive regression trees (Chipman, McCulloch, & Dorie, 2014); and 5–6) two neural networks with different architectures (Ripley & Venables, 2021). We specified a 20-fold cross validation process.

In Stage 2, TMLE targets the Stage 1 estimated variable importance of each change language subtype (A) based on our model, defined as the mean level of the outcome when A = 1 minus the mean level for the outcome when A = 0. These models corrected for potentially informative missingness in the outcome by modeling P(missing | X) and for potential confounding using a propensity score, P(A = 1 | W), where W is all covariates, including the time dummy codes, and the other change language subtypes. In this stage, the initial outcome regression model was targeted towards optimizing estimation of the corrected marginal relationship between a given change language subtype predictor and the outcome. The model evaluated the marginal average treatment effect (ATE) corresponding to a one-unit change in value in each predictor, i.e., going from ‘lower’ to ‘higher’ (coded 0 and 1, respectively) on a given change language subtype based on a sample median split. The ATEs were calculated assuming that predictors affect outcome equally over time.

Analyses were carried out in R v 3.6.0 (R Core Team, 2013) using SuperLearner (v2.0–26) and code adapted from the tmle package.(Polley, LeDell, Kennedy, Lendle, & van der Laan, 2019; Gruber, van der Laan, & Kennedy, 2021). Source code and data used for these analyses are available at https://osf.io/k3qvg/, (Kahler, 2022). After obtaining the ATE estimates, predictors were ranked based on estimate magnitude.

The variable importance analysis was exploratory, and thus triangulation of three key considerations guided how the results were interpreted. Specifically, we focused our discussion on the three to four variables in each set of competing variables (i.e., approximately the top quartile of variables examined) with the largest ATE estimates. These estimates reflect the largest expected difference in alcohol outcomes for sessions with versus without a higher level of a given client language variable. We also considered the precision of estimates as measured by the width of their confidence intervals, since estimates with relatively narrow confidence intervals are those less likely to be influenced by chance (Poole, 2001). Finally, because we had two different but related measures of alcohol use as outcomes (i.e., number of drinks, number of heavy drinking days), we also considered whether variables had a similar rank ordering of importance across both outcomes, since variables that have a large influence on both outcomes would be considered more robust as predictors.

Results

Descriptive Statistics

Table 1 details the sociodemographics, alcohol use, and MISC and GBIAS code summaries for study participants. Overall, sessions contained an average of 820.2 (SD = 359.1; range = 211 to 2186) client utterances, which was very highly correlated with the total length of the session in minutes, r = .78. Of the client utterances, an average of 465.0 (SD = 211.2; range = 148 to 1400) were focused on the topic of alcohol. Of these alcohol utterances, an average of 122.8 (SD = 60.9; range = 8 to 357) were coded as CT, and 50.6 (SD = 30.8; range = 6 to 55) were coded as ST. An average of 108.4 (SD = 47.1; range = 30 to 301) alcohol-focused utterances were coded as Follow Neutral, with the remainder (M =183.2, SD = 138.8; range = 23 to 1057) coded as continuations of the prior utterance; these utterances were not considered in analyses.

Table 1.

Descriptive Statistics for Primary Study Variables (N=164)

Baseline		n(%) or M	SD	Min	Max
	Female gender	94 (57.3)
	Age (Years)	28.5	9.2	18	60
	Hispanic/Latinx	22 (13.4)
	Race
	American Indian	4 (2.4)
	Black/African Am.	18 (11.0)
	White	136 (82.9)
	Multiple races	6 (3.7)
	Contemplation Ladder1	4.9	3.9	0	10
Alcohol Use		Baseline (n = 164)		M3 (n = 136)		M6 (n = 141)		M9 (n = 144)
		Median	25,75 th	Median	25,75 th	Median	25,75 th	Median	25,75 th
	Heavy Drinking Days per Month	3	1, 7	2	0, 4	2	0, 5	1	0, 4
	Drinks Per Week	7.7	2.4, 15.4	4.8	1.4, 11.4	4.4	1.3, 12.6	3.5	0.2, 10.2
Topics		Number of CT Utterances by Topic				Number of ST Utterances by Topic
		Median	25,75 th	Min	Max	Median	25,75 th	Min	Max
	Patterns & Contexts	13.5	8, 21.8	0	62	8.5	5, 16	0	44
	Barriers & Facilitators	4	0, 9	0	29	2	0, 5	0	23
	Benefits of the TB	3	1, 6	0	29	16	10, 25.8	2	73
	Drawbacks of the TB	53	36, 71.5	3	186	7	4, 15	0	74
	Benefits of TB change	6	2.3, 13	0	60	0	0, 0	0	5
	Drawbacks of TB change	0	0, 0	0	6	0	0, 0.8	0	13
	Change Planning	25	11, 39	0	116	2	0.3, 5	0	27
DARN-C		Number of CT Utterances by MISC Subtype				Number of ST Utterances by MISC Subtype
		Median	25,75 th	Min	Max	Median	25,75 th	Min	Max
	Desire	1	0, 3	0	27	0	0, 0	0	7
	Ability	6	3, 12.8	0	45	2	0, 6	0	32
	Reason	74	46.3, 106.8	6	281	44	29, 67.8	7	204
	Need	1	0, 2	0	11	0	0, 0	0	6
	Commitment	13.5	5, 22	0	58	1	0, 2	0	10
	Taking Steps	0	0, 1	0	18	0	0, 0	0	5
	Other	54	36.5, 77.8	7	168	10	6, 16	0	7

Note. CT = Change Talk. ST = Sustain Talk. TB = Target Behavior. DARN-C = Desire, Ability, Reason, Need, Commitment. M3 = month 3. M6 = month 6. M9 = month 9. 25,75^th = 25^th and 75^th percentile for the variable.

¹Contemplation Ladder for alcohol use ranges from 0 = I never think about drinking less to 10 = My drinking has changed.

Supplemental Tables 2a–2c show correlations among the dichotomized GBIAS change language subtypes, among dichotomized DARN-C change language subtypes, and between dichotomized GBIAS subtypes and dichotomized DARN-C subtypes, respectively. Correlations among the 14 GBIAS change language subtypes ranged from −.14 to .74, with large correlations (r = .50 or greater) seen in 10 of the 91 pairwise correlations. Correlations among the 14 DARN-C subtypes ranged from −.41 to .61, with large correlations seen in 8 of the 91 pairwise correlations. Finally, the 196 pairwise correlations between GBIAS subtypes and the DARN-C subtypes ranged from −.32 to .83, with 14 correlations of r = .50 or greater.

Variable Importance Results

GBIAS Change Language Subtypes

Tables 2a and 2b summarize the relative importance of GBIAS-derived change language subtypes, along with their confidence intervals, in predicting the two alcohol outcomes at follow-up. For drinks per week, ST regarding the topic of Change Planning was the most important predictor of the outcome. Counterintuitively, the coefficient was negative indicating that being above the median on this type of ST versus below the median was associated with less drinking controlling for baseline drinking, baseline readiness to change, and the other subtypes of CT and ST in the model. The second and third most important variables were ST: Drawbacks of TB Change and ST: Benefits of the TB. As would be expected, higher ST about drawbacks of changing drinking (e.g., “I don’t want to be that weird guy in the corner who can’t have a drink.”) and about the benefits of the TB (e.g., “It [drinking] really helps to shut some of the noise off”) predicted more drinking. Although of a similar magnitude as the effect of ST: Benefits of the TB, the effect of ST: Benefits of TB Change had a relatively wide confidence interval.

Table 2a.

Relative Importance of MISC: GBIAS Topic Subtypes in Predicting Weekly Average Drinking (N=164), Most to Least Important

Variable	Rank	Estimate	SD	p	95% CI, Lower	95% CI, Upper	CI Width
ST: Change Planning	1	−5.57	1.30	<.001	−8.11	−3.02	5.09
ST: Drawbacks of TB Change	2	3.75	1.36	.006	1.08	6.43	5.34
ST: Benefits of the TB	3	2.97	1.05	.005	0.91	5.02	4.11
ST; Benefits of TB change	4	−2.96	2.14	.166	−7.16	1.23	8.39
CT: Change Planning	5	2.79	0.98	.005	0.86	4.72	3.86
CT: Drawbacks of TB Change	6	−2.78	0.80	.001	−4.34	−1.21	3.13
ST: Patterns and Contexts	7	−2.03	1.96	.302	−5.88	1.82	7.70
CT, Barriers and Facilitators	8	1.89	0.82	.020	0.29	3.49	3.20
ST: Barriers and Facilitators	9	1.49	0.90	.098	−0.28	3.26	3.54
CT: Benefits of the TB	10	−1.47	1.23	.232	−3.88	0.94	4.82
CT: Benefits of TB Change	11	−1.46	1.05	.165	−3.51	0.60	4.11
ST: Drawbacks of the TB	12	1.11	1.18	.347	−1.20	3.43	4.63
CT: Patterns and Contexts	13	0.33	1.04	.753	−1.72	2.37	4.09
CT: Drawbacks of the TB	14	−0.17	1.12	.878	−2.38	2.03	4.41

Note. MISC = Motivational Interviewing Skill Code. GBIAS = Generalized Behavioral Intervention Analysis System. CI = confidence interval. SD = standard deviation. Estimate reflects difference in the outcome based on higher versus lower score on the topic using a median split. CT = Change Talk. ST = Sustain Talk. TB = Target Behavior.

Table 2b.

Relative Importance of MISC: GBIAS Topic Subtypes in Predicting Heavy Drinking Days (N=164), Most to Least Important

Variable	Rank	Estimate	SD	p	95% CI, Lower	95% CI, Upper	CI Width
ST: Change Planning	1	−2.07	0.56	<.001	−3.17	−0.98	2.19
CT: Change Planning	2	1.95	0.53	<.001	0.91	2.99	2.08
ST: Patterns and Contexts	3	−1.59	1.29	.218	−4.11	0.94	5.05
CT: Drawbacks of TB Change	4	1.27	0.91	.164	−0.52	3.06	3.58
CT: Benefits of TB Change	5	−1.09	0.46	.019	−1.99	−0.18	1.81
ST: Drawbacks of TB Change	6	0.79	0.82	.333	−0.81	2.40	3.21
ST: Barriers and Facilitators	7	0.62	0.45	.166	−0.26	1.50	1.76
ST: Drawbacks of the TB	8	0.57	0.60	.345	−0.61	1.75	2.36
ST: Benefits of the TB	9	0.52	0.52	.320	−0.50	1.53	2.04
ST: Benefits of TB Change	10	−0.50	1.01	.619	−2.47	1.47	3.94
CT: Drawbacks of the TB	11	0.42	0.52	.418	−0.60	1.44	2.04
CT: Benefits of the TB	12	−0.17	0.57	.765	−1.29	0.95	2.24
CT: Barriers and Facilitators	13	0.16	0.49	.750	−0.80	1.11	1.91
CT: Patterns and Contexts	14	0.01	0.48	.982	−0.94	0.96	1.90

Note. MISC = Motivational Interviewing Skill Code. GBIAS = Generalized Behavioral Intervention Analysis System. CI = confidence interval. SD = standard deviation. Estimate reflects difference in the outcome based on higher versus lower score on the topic using a median split. CT = Change Talk. ST = Sustain Talk. TB = Target Behavior.

As with drinks per week, the strongest predictor of heavy drinking days was ST: Change Planning, with more ST on this topic related to less frequent heavy drinking during follow-up. Also counterintuitively, being above the median on CT: Change Planning was associated with more heavy drinking days. All other variables showed a relatively small ATE for heavy drinking days or had substantially wider confidences intervals. Because ST and CT around Change Planning were among the most highly collinear variables in the analyses (tetrachoric r = .50), their relationship with outcome may have been altered when analyzed jointly, including potentially changing the direction of the association, a statistical phenomenon termed “suppression” by Cohen et al. (2003). Therefore, we re-ran TMLE analyses once without ST: Change Planning and once without CT: Change Planning. Removal of CT: Change Planning did not impact the rank ordering of ST: Change Planning, whereas CT: Change Planning moved into the bottom half of predictor importance when ST: Change Planning was removed.

Post Hoc, Mixed Method Examination of Cases with High ST: Change Planning Utterances.

The most consistent, yet surprising, finding to emerge from the variable importance analysis of GBIAS subtypes was that high ST around change planning was predictive of better drinking outcomes. The median number of ST: Change Planning utterances was only 2 (25^th percentile = 0; 75^th percentile = 5, and 95^th percentile = 11). Thus, participants were classified as higher on this variable even when the total amount of conversation involving ST: Change Planning was quite modest. As shown in Table 4, the types of utterances that can be classified as ST: Change Planning are heterogeneous and can include rejection of specific change goals or strategies (e.g., “It’s just not something I plan on doing.”), as well as conversation around challenges to implementing change.

Table 4.

Examples of ST: Change Planning by participants who reported >50% reduction in drinks per week.

Category	Participant	Quotation
Self-efficacy	20 YO F	I want to stop (CT), but it’s kind of like a habit now (ST). P: I’m trying to cut back (CT), and then it will come in front of me, and I can’t say no (ST).
	28 YO F	Just to see, I could stop drinking for the first week, and then Friday picks back up again. (ST) … I’m contradicting myself because I like doing it (ST), but I’m like, ‘it sucks’ (CT).
	23 YO F	I don’t have confidence in myself … Even if I know I can do it, I still in my head think that I can’t.
Obstacles/Contingencies	22YO F	I could almost say I probably won’t stop drinking until I get (psychiatric treatment) (ST).
	28 YO F	I do need to go see somebody [for help with drinking] but I don’t wanna have that piece of paper floating over my head where the court might find out (ST).* If I leave today and I don’t get this job [I applied for] … then I go back into the slump .. . then I will pick up the bottle (ST).
Non-abstinence	26YO M	I think on major holidays I’ll probably drink (ST). It’s all right once in a while to go crazy, have fun (ST)
	20 YO F	So after the baby’s born, I mean like my 21st birthday, I’m going to be bad (ST).
	22 YO M	I have no idea how I could [eliminate hangovers] unless I could just stop drinking. There’s no way to do that (ST).
	23YO F	Everyone’s like, “Oh, just quit.” … I don’t want to just quit, I’m still 23.
Resists harm reduction	20 YO M	I don’t think I should [drive drunk], but if … I was to start drinking alcohol right now, I’d go and drive (ST). I don’t know what the change I could make though is; I mean, I’d have to walk places (ST).
Rejects intermediate steps	35 YO M	I’d probably have to fall off again before I [go back to AA] (ST). I don’t wanna go to AA by myself, I don’t want to get there and all that shit … (ST).
	28 YO F	There’s lots of people that ask me to do stuff … in the evening, take a walk … I’m never gonna do that, I’m too lazy (ST). I’d rather have a drink (ST).
	23YO F	I’m not ready for that step [avoiding Long Islands] yet.

^*She is suing for custody of her children. CT = Change Talk. ST = Sustain Talk

To better understand the relationship between ST utterances around Change Planning, and subsequent reported changes in drinking behavior, we reviewed transcripts from the 19 participants who provided follow-up data and were in the 90^th or greater percentile of ST: Change Planning utterances, which ranged from 8 to 27. Nine of these participants reported a 50% or greater reduction in average drinks per week from baseline (M = 38.5; SD = 29.4) to follow-up (M = 5.2; SD = 8.0). One additional participant reported a substantial 42.4% reduction in drinking post-intervention: baseline drinking of 68.8 drinks per week versus an average of 39.7 drinks per week during follow-up. Of these ten participants who made large changes based on self-reported data, six indicated early in the MI session that they were interested in reducing their alcohol consumption. Two of these, in fact, stated that they had already stopped shortly before enrolling in the study. Four other participants indicated a desire to reduce their drinking later in the MI session. For all of these participants, ST: Change Planning did not consist of resistance to reducing drinking. Rather, the ST largely reflected a substantive exploration of the difficulties they would face in reducing drinking or rejection of potential intermediate steps proposed by the counselor, such as attending self-help meetings. In two cases, ST included resistance to harm reduction suggestions, specifically not driving while intoxicated. Several participants also stated that total abstinence was not a goal even though they were planning to cut down on their drinking. (See Table 4 for examples.)

The nine participants with high ST: Change Planning, who did not report substantial reduction in drinking post-intervention, reported many fewer drinks per week at baseline (M = 8.4; SD = 11.6) than those who did change, and their drinking remained stable, on average, over follow-up (M = 11.4; SD = 9.9). Although each had a meaningful number of CT: Change Planning utterances (range = 27 to 68), seven of these nine discussed only minor changes in drinking (e.g., going from 12 to 10 drinks per week) while rejecting abstinence as a goal and making no firm plans for change. One indicated a desire to continue heavy drinking, and one indicated a need to quit drinking, discussed recent treatment episodes and obstacles to quitting, but never committed to a plan for getting into treatment.

MISC Change Language Subtypes

Tables 3a and 3b summarize the relative importance of DARN-C variables. For drinks per week (Table 3a), the most important variable was ST: Taking Steps. This variable reflects client utterances describing recent increases in drinking since enrolling in the study (e.g., “Me and my girlfriend have been drinking a lot just more recently;” “Before, I wasn’t drinking as much as I am recently;” “But, yes, lately, this past month, I’ve been drinking more than before”). Such utterances were very rare, with only 9.1% (n = 15) of participants having at least one such utterance during the session and thereby classified as higher on ST: Taking Steps. ST: Desire (e.g., “it’s [quitting drinking] not something I feel like I want to do right now”) was the next most important variable with higher values associated with more drinking, as would be expected. Also consistent with expectation, CT: Need (e.g., “I need to stop.”) was associated with less drinking. However, the fourth most important variable, CT: Reason (e.g., “Alcohol, yeah. I need to stop, I know.”), was unexpectedly associated with greater drinking.

Table 3a.

Relative Importance of DARN-C Subtypes in Predicting Weekly Average Drinking (N=164), Most to Least Important

Variable	Rank	Estimate	SD	p	95% CI, Lower	95% CI, Upper	CI Width
ST: Taking steps	1	22.71	1.24	<.001	20.29	25.13	4.84
ST: Desire	2	3.64	1.17	.002	1.35	5.92	4.57
CT: Need	3	−3.20	1.03	.002	−5.21	−1.19	4.02
CT: Reason	4	3.04	1.03	.003	1.01	5.06	4.05
CT: Commitment	5	−2.95	1.09	.007	−5.08	−0.82	4.26
CT: Ability	6	−2.88	1.03	.005	−4.90	−0.86	4.04
ST: Need	7	2.57	1.52	.090	−0.40	5.54	5.94
CT: Desire	8	−2.54	1.19	.033	−4.87	−0.21	4.66
ST: Reason	9	−2.34	1.15	.041	−4.59	−0.09	4.50
ST: Commitment	10	−1.83	1.27	.149	−4.32	0.66	4.98
ST: Ability	11	1.57	0.87	.071	−0.13	3.26	3.40
CT: Other	12	−1.18	0.95	.213	−3.03	0.68	3.71
CT: Taking steps	13	0.82	1.09	.449	−1.31	2.96	4.26
ST: Other	14	−0.80	1.28	.532	−3.31	1.71	5.02

Note. DARN-C = Desire, Ability, Reason, Need, Commitment. Estimate reflects difference in the outcome based on higher versus lower score on the Motivational Interviewing Skill Code subtype using a median split. CI = confidence interval. SD = standard deviation. CT = Change Talk. ST = Sustain Talk.

Table 3b.

Relative Importance of DARN-C Subtypes in Predicting Heavy Drinking Days (N=164), Most to Least Important

Variable	Rank	Estimate	SD	p	95% CI, Lower	95% CI, Upper	CI Width
ST: Taking steps	1	2.45	0.39	<.001	1.68	3.21	1.53
ST: Need	2	2.00	0.69	.004	0.64	3.36	2.72
CT: Desire	3	−1.64	0.59	.005	−2.79	−0.49	2.30
CT: Reason	4	1.50	0.49	.002	0.54	2.47	1.94
CT: Need	5	−1.18	0.63	.059	−2.41	0.04	2.45
ST: Reason	6	−1.07	0.69	.120	−2.41	0.28	2.69
CT: Ability	7	−0.99	0.46	.033	−1.89	−0.08	1.81
ST: Desire	8	0.98	0.54	.070	−0.08	2.04	2.12
ST: Other	9	−0.78	0.61	.199	−1.97	0.41	2.38
ST: Ability	10	0.77	0.42	.070	−0.06	1.60	1.66
CT: Other	11	−0.73	0.43	.091	−1.58	0.12	1.69
CT: Commitment	12	−0.67	0.55	.221	−1.74	0.40	2.14
ST: Commitment	13	−0.46	0.62	.457	−1.67	0.75	2.42
CT: Taking steps	14	0.21	0.57	.715	−0.91	1.32	2.23

Note. DARN-C = Desire, Ability, Reason, Need, Commitment. Estimate reflects difference in the outcome based on higher versus lower score on the Motivational Interviewing Skill Code subtype using a median split. CI = confidence interval. SD = standard deviation. CT = Change Talk. ST = Sustain Talk.

In the analysis predicting heavy drinking days (Table 3b), ST: Taking Steps was again the most important variable, although its coefficient was substantially smaller than it was for drinks per week. CT: Reason was again associated unexpectedly with greater drinking. However, whereas CT: Need and ST: Desire had been among the most important variables in predicting drinks per week, it was ST: Need and CT: Desire that were among the most important in predicting heavy drinking. The confidence interval width for ST: Need was wider compared to the other three variables with the largest ATE, suggesting less precision around the estimate of its ATE for heavy drinking.

Discussion

In this study, TMLE provided a rigorous means of examining the relative importance of subtypes of change language—as defined by the commonly used MISC system and the more recently developed GBIAS—as predictors of alcohol use outcomes following an MI intervention. Importantly, these analyses did not rely on p-values to determine variable importance, as is often seen in the literature. Instead, our approach examined point estimates and the precision around the point estimates and used an innovative approach, TMLE, to combine a variety of analytic methods to achieve estimates of relative ranking of the importance of variables while facilitating correct model specification such as the presence of higher order product terms. Results of these exploratory analyses replicated some prior findings on change language subtypes, but also provided a handful of novel insights that can inform future confirmatory investigations and that may have implications for the delivery of MI within a brief intervention.

The variable importance analysis of subtypes of change language defined by the GBIAS topic categories identified one particularly surprising and strong association, namely that higher levels of ST on the topic of Change Planning predicted reduced number of drinks consumed per week and reduced frequency of heavy drinking when simultaneously taking into account levels of the other subtypes of change language. Our post hoc examination of instances of ST around Change Planning, which were relatively rare even in those classified as being higher on this variable, helped clarify the nature of the association found. For the majority of participants higher in ST: Change Planning, the ST utterances emerged from a conversation specifying the type of change goal (e.g., abstinence, harm reduction) and the steps they would take to reach that goal. This finding highlights the importance of conversational context when deciding whether to “soften sustain talk.” If a client indicates an openness to discussing changing their drinking, rejection of certain ideas—coded as ST: Change Planning—may simply reflect the collaborative dialogue supported within an MI framework to identify viable options for behavior change; counselors often solicit discussion of anticipated difficulties in the change planning process. In the context of an MI session with non-treatment seeking clients, counselors only may have engaged in such discussions if the client indicated a willingness to change. Therefore, ST: Change Planning may be serving as a potential marker for those sessions in which a client progressed to deeper changing planning. Our results highlight the limitation in inferring that all ST utterances are negative prognostic indicators and may explain, in part, the relatively weak associations seen in the literature between aggregated ST utterances and later behavior change (e.g., Magill et al., 2018). Although these results are exploratory and require confirmation, they do suggest that counselors can elicit clients’ perceptions of difficulties and challenges, including encouraging clients to explore and reject possible intermediate steps or goals, without being concerned that eliciting such ST will undermine behavior change.

Another contextual consideration derived from our mixed methods exploratory analyses is that participants with higher ST around Change Planning who substantially reduced their drinking tended to be particularly heavy drinkers. Therefore, their expressions of concerns about their ability to change drinking or quit drinking likely reflected their relatively higher severity of use. By contrast, those with higher ST around Change Planning who did not make substantial changes to their drinking were relatively lighter drinkers, even when compared to the sample as a whole. Any change intentions they endorsed tended to be for very modest reductions. For a clinician, these results suggest caution when moving towards change planning, especially in instances where a client’s drinking barely exceeds low risk drinking limits. In sum, results suggest that ST during change planning can only be understood in the context of an individual’s needs, readiness, and capacity for change and highlights the importance of personalizing each brief MI session to maximize potential for behavior change.

A number of other subtypes of change language defined by the GBIAS topic codes showed confidence intervals that did not include zero, suggesting their association with outcome was greater than chance. However, even for these change language subtypes, the relative rankings of their importance were inconsistent across the outcomes of drinks per week versus heavy drinking days or they had relatively wide confidence interval widths. CT around Change Planning showed an association with greater drinking, but follow-up analyses in which ST: Change Planning—with which it was highly collinear—was removed from the model greatly attenuated that association, suggesting that CT: Change Planning only predicted drinking when the effect of ST: Change Planning was accounted for. Therefore, drawing generalizations from exploratory analyses about the relative importance of these change language subtypes needs to be done with great caution, and we do not speculate further about which additional GBIAS-defined subtypes are most worthy of further attention.

In the variable importance analyses of MISC-defined change talk subtypes, higher ST around Taking Steps was the most important predictor of greater drinking during follow-up. This subtype of ST was particularly rare—occurring in only about 10% of participants—and reflected the fact that a small number of participants had recently increased their drinking prior to joining the study. The presence of any such recent escalation in drinking immediately prior to completing an MI session may be an important indicator of poor response to MI to which counselors should attend. Specific attention to the circumstances that led to that escalation can help clarify its motivational significance, including whether a client views this change as temporary. In the present study, which was conducted in participants who were not seeking treatment, those changes were typically not temporary. Studies that include formal booster sessions may be able to track whether participants with recent escalations in drinking become more ready to change after those patterns have persisted long enough to result in negative consequences.

The two next most important predictors of outcome from the MISC subtypes were around Desire and Need. However, although the direction of the prediction was consistent with expectations, whether it was ST or CT around these variables that was more important differed based on the alcohol outcome examined. ST around Desire has been found to be a relatively strong predictor in some but not all studies (Magill et al. 2019); the current study provides further evidence that expressing a lack of desire to change or a desire to persist in current drinking patterns may be prognostic of poor outcomes. CT: Desire was a relatively important predictor of reduced heavy drinking consistent with theory. That ST around the Need to change drinking predicted more drinking while CT: Need predicted less drinking also is consistent with MI theory. However, given that most prior studies have not shown meaningful associations between any CT subtypes and outcome (Magill et al., 2019), further research and replication is needed to determine the relative importance of CT: Need and CT: Desire. Furthermore, these results must be viewed with some caution given that CT around Reasons was associated with more drinking, a direction inconsistent with MI theory. It is important to note that all of these change language subtypes were considered simultaneously in the analyses.

Limitations

This study should be viewed for the exploratory analysis that it represents, suggesting further areas for examining change talk subtypes. All predictor variables were dichotomized, which likely minimizes model misspecification but reduces the amount of variability in each predictor and thus may result in some loss of information about the relationship between a predictor and outcome. The current study conducted reliability analyses on randomly selected session segments of 300 to 700 utterances rather than on entire sessions. Therefore, the frequency of GBIAS and MISC subtypes counts were necessarily reduced, which made it impossible to calculate reliability estimates for very rare codes. The study sample was relatively small given the number of predictors examined and comprised a large proportion of non-Hispanic White participants, limiting generalization to other racial and ethnic groups. Also, all participants were recruited in an emergency department, reported engaging in some risky sexual behavior, and were not seeking alcohol treatment. The impact of that context on discussions of drinking and its risks may differ from other contexts where medical consequences are less salient or when people are specifically seeking alcohol treatment.

Conclusions

Use of TMLE with two systems for coding subtypes of change language led to novel findings with potential clinical implications if they can be replicated and extended. Results suggest, first and foremost, that as long as a client is endorsing some change goals and committing to some steps to support change in drinking, critical discussion of goals and strategies may reflect active engagement in change planning rather than resistance to change per se and may therefore be a valued outcome of MI sessions. Second, recent escalation in drinking prior to an MI appears to be a robust negative prognostic factor. Assuming a causal relationship exists, and modifications to counseling may be needed to interrupt these patterns when present. Finally, the variability of findings across studies about which subtypes of change talk are predictive of outcome, including some relationships being in the opposite direction anticipated, suggests that more nuanced models of change language are needed that can consider the nature of given change language subtypes within the context of all other change language in that session. To inform clinical practice more fully, it also will be important to understand whether specific study settings or sample characteristics modify associations between client talk and outcomes.

Public Health Significance.

This study found that in brief alcohol counseling, clients’ expressions of concern about taking certain steps or setting certain goals around their drinking—which could be considered arguments against behavior change—actually predicted better drinking outcomes. Thus, as long as a client has an intention to change their drinking, frank discussions about what goals and steps they may or may not find acceptable may facilitate behavior change.