A Framework for Estimating Posttreatment Moderation of Treatment-by-Dosage Effects in Individual-Patient Meta-Analysis: An Illustration Using Project Harmony

Abstract

Making causal statements regarding dose-response in treatments for posttraumatic stress disorder (PTSD) and alcohol/other drug use disorders (AODs; PTSD+AOD) is difficult because (a) dosage is rarely randomized and (b) self-selected dosage can be affected by treatment assignment. In the present study, we sought to clarify causal inferences regarding treatment-by-dosage interactions in PTSD+AOD treatment using Project Harmony, an individual-patient meta-analytic data set of behavioral, pharmacological, and combination PTSD+AOD treatments (k = 36; N = 4,046). Using propensity score weighting and moderated multilevel “net treatment difference” modeling, trauma-focused (TF) treatments, whether integrated or nonintegrated with AOD treatment, outperformed treatment as usual by greater margins on reductions in PTSD and alcohol use as dosage increased. Furthermore, appropriately treating dosage as a posttreatment covariate and moderator revealed effects for TF treatments on drug use that had not been detected in previous studies. Implications for approaches to increasing TF-treatment attendance and greater use of causal-inference methodologies with dose-response analyses are discussed.

Posttraumatic stress disorder (PTSD) and alcohol/other drug use disorders (AODs) frequently co-occur. Approximately 46% of people with PTSD have a lifetime comorbid AOD, and 36% of people with AOD have current comorbid PTSD (Gielen et al., 2012; Pietrzak et al., 2011). Each disorder can be debilitating on its own, but disability and psychiatric outcomes are exacerbated when they co-occur (e.g., Bowe & Rosenheck, 2015; Peirce et al., 2008). PTSD, AOD, and particularly comorbid PTSD+AOD are all associated with high rates of treatment dropout (e.g., Coffey et al., 2016; Hien et al., 2009; Imel et al., 2013; Kline et al., 2025; Mills et al., 2012; Roberts et al., 2022; Simpson et al., 2021).

Given the links between PTSD and AOD and their mutually exacerbating effects, several treatments have been developed to address this comorbidity. Behavioral interventions that have been tested with patients diagnosed with PTSD, AOD, or comorbid PTSD+AOD can be categorized based on their focus: exclusively targeting PTSD (e.g., cognitive-processing therapy for PTSD; Resick et al., 2024), exclusively targeting AODs (e.g., relapse prevention; Marlatt & Donovan, 2005), and targeting both PTSD and AOD (e.g., concurrent treatment of PTSD and substance use disorders [COPE]; Back et al., 2015). Psychotherapies targeting PTSD also vary in whether they are trauma-focused (TF); some facilitate processing of the trauma and its meaning (e.g., prolonged exposure, Foa et al., 2018; cognitive-processing therapy, Resick et al., 2024), and others do not (e.g., seeking safety; Najavits, 2002). Distinguishing between TF and non-TF psychotherapies and integrated and nonintegrated interventions provides key information about their divergent outcomes and mechanisms of action (Hien et al., 2022). In parallel to these behavioral interventions, pharmacological interventions for PTSD+AOD also vary (a) to the extent to which they target PTSD (e.g., Brady et al., 2005) or AOD (e.g., Petrakis et al., 2016) and (b) whether they have been delivered in combination with some of the behavioral treatment platforms listed above (Foa et al., 2013; Hien et al., 2015; Petrakis et al., 2020).

The Impact of Treatment Dosage on PTSD+AOD Outcomes

For the purpose of this study, we operationalize treatment dosage as the proportion of (a) planned behavioral-intervention sessions that were attended by the patient and/or (b) intended dosage of PTSD- or AOD-targeting medication. A complex relationship exists between treatment dosage (i.e., the proportion of a treatment completed) and PTSD+AOD treatment outcomes. PTSD+AOD behavioral interventions vary not only with respect to their specific content and techniques but also with respect to the number of sessions and/or medication doses that define a theoretical “full dosage” of the intervention. In Project Harmony, a meta-analysis of 36 randomized controlled trials (RCTs; combined N = 4,046) of interventions for PTSD+AOD (Hien et al., 2019, 2023; Saavedra et al., 2021), the total number of planned sessions for PTSD+AOD behavioral interventions ranged from four (cognitive restructuring; Stappenbeck et al., 2015) to 25 (seeking safety; Myers et al., 2015; Najavits et al., 1998, 2005). However, as summarized by Sechrest et al. (1979; see also Sagarin et al., 2014), systematically identifying the impact of dosage on PTSD+AOD outcomes across intervention types is challenging because (a) interventions differ in the theoretical definition of what constitutes a full or meaningful dose (the planned strength of the treatment), (b) interventions differ in their achieved dose (e.g., for behavioral treatments, the number of sessions attended), and (c) dose-response effects cannot be directly interpreted as causal effects in the absence of random assignment to dosage.

In RCTs of PTSD+AOD, particularly with behavioral interventions, the de facto assumption of “more is better” has not always been supported. Treatment-dosage effects for non-TF interventions seem to be curvilinear in nature such that the dosage for optimal treatment outcomes appears to be about half of available sessions (Hien et al., 2012), whereas the effectiveness of integrated and single-disorder TF interventions typically appears to increase linearly with increasing dosage (Berke et al., 2019; Holmes et al., 2019; Mills et al., 2016; Schacht et al., 2017; Straus et al., 2022; Szafranski et al., 2019; but for an exception, see McLean et al., 2022). However, in nearly every PTSD+AOD psychotherapy (and medication) RCT to date, intervention dosage is not randomly assigned but more often self-selected. Factors that influence how many sessions patients choose to attend and/or what percentage of medication doses they take as prescribed may serve as confounders of the effects of dosage on PTSD+AOD-treatment outcomes. Moreover, selection-bias effects on dose-response outcomes may even, in fact, be built into the treatment process. The patient (and perhaps the therapist) may modulate the amount or length of treatment received based on the patient’s progress on the outcome. According to Barkham et al. (2006), “the level of improvement and treatment duration are mutually regulated so that treatments tend to end when clients, on average, have improved to a degree or level that is good enough” (p. 161). PTSD+AOD treatments may offer flexible ranges of doses based on therapist-patient decisions or indicators of patient progress (e.g., Galovski et al., 2012).

Several individual-level predictors of attendance in PTSD+AOD treatment have been identified (Kline et al., 2025; Mills et al., 2016), many of which are also related to PTSD+AOD-outcome trajectories. In PTSD+AOD treatments, demographic characteristics, such as younger age and lower education level, predict lower attendance (Belleau et al., 2017; Brady et al., 2001; Myers et al., 2015; Pinto et al., 2011). Baseline clinical predictors, such as higher PTSD-symptom severity (Morgan-Lopez et al., 2013, 2014; Myers et al., 2015; Sannibale et al., 2013), alcohol-use severity (Kline et al., 2021; Myers et al., 2015), and drug-use frequency (Sannibale et al., 2013), also predict lower attendance. Many of these variables have also been identified as moderators of PTSD+AOD-treatment outcomes (e.g., Mills et al., 2016; Ruglass et al., 2017). The relationship between dosage and PTSD+AOD outcomes may be a causal effect or a spurious (confounded) relationship because of individual characteristics that predict both dosage and outcomes. Thus, RCTs in which dosage is not randomly assigned may produce findings that do not properly represent the causal effects associated with dose-response.

Advances in Causal Moderation: Implications for Dose-Response in PTSD+AOD Treatment

The quantitative-methods literature has long addressed causal inference in observational studies (Rosenbaum, 1984; Shadish et al., 2002). Variants of propensity-scoring approaches have served as an important method of adjusting for covariate imbalance across treatment groups that are not randomly assigned (Rosenbaum & Rubin, 1983; Saavedra et al., 2023; West et al., 2014). Propensity-score methods have also been developed for moderation analysis to address potential selection biases in RCT and non-RCT data (Bansak, 2021; Dong, 2012). However, the frameworks of Bansak (2021) and Dong (2012) are structured for moderation involving pretreatment moderators. The issue with self-selected dosage is that dosage may not only moderate treatment outcomes but also in and of itself (a) be affected by the treatment and (b) affect PTSD+AOD outcomes; this would also make self-selected dosage a de facto mediator, consistent with the idea that posttreatment confounding and mediation are statistically, although not conceptually, equivalent (MacKinnon et al., 2000; Rosenbaum, 1984, 2020). Recent advances in potential-outcomes mediation analysis (MacKinnon et al., 2020; Saavedra et al., 2023; VanderWeele & Vansteelandt, 2009) allow for the modeling of variables, such as dosage, as variables that may simultaneously moderate treatment effects and are also intermediate in the causal sequence between treatments and outcomes.

In the present study, we estimate treatment-by-dosage interactions using the proportion of intervention sessions attended/medication doses taken as the indicator of dosage. We used a combination of propensity score weighting and potential-outcomes mediation analysis to assess whether the influence of dosage on clinical outcomes varies across PTSD+AOD behavioral, medication, and combination interventions, which would (a) provide a new synthesis of evidence regarding treatment-type-by-dosage interactions in PTSD+AOD treatments and (b) clarify whether previous single-study assessments of treatment-dosage interactions were not spurious in nature. We used data from Project Harmony (Hien et al., 2023) to investigate this question. In the current study, we also provide an empirical illustration of conducting a dose-response analysis that mimics what may have been found under joint random assignment of both intervention conditions and dosage (Bansak, 2021; MacKinnon et al., 2020). Given earlier results for non-TF interventions (e.g., Hien et al., 2012) and the findings of the Project Harmony parent study (Hien et al., 2023), we probed the possibility that non-TF interventions may produce a curvilinear relationship between dosage and response wherein higher dosage would lead to weaker effects on treatment outcomes than moderate dosage (Hien et al., 2012). Consistent with existing research for TF interventions (e.g., Schacht et al., 2017), we hypothesized that higher dosage would be related to improved PTSD+AOD outcomes. We also explored the relationship between dosage and response in (a) integrated versus nonintegrated PTSD+AOD interventions and (b) PTSD- and AOD-targeting medications.

Transparency and Openness

This study involved an analysis of existing data rather than primary data collection. Details about the Project Harmony study design were reported in Saavedra et al. (2021).

Preregistration

Project Harmony was preregistered in the International Prospective Register of Systematic Reviews (PROSPERO identifier: CRD42019146678). Primary outcomes analyses mediation and moderation analyses published by Hien et al. (2023) were described in the preregistration, but posttreatment moderation of treatment-by-dosage effects reported in this study were not preregistered.

Data, materials, code, and online resources

Access to the integrated data are subject to approval by Project Harmony investigators and for certain subsets of data, terms of data-use agreements established with original RCT principal investigators (PIs). Item parameters used to score latent PTSD and AOD severity psychometric scale scoring of outcomes are available in the supplemental material to the primary-outcomes report (Hien et al., 2023). Mplus measurement modeling code was made available in supplemental material published by Morgan-López, Hien, et al. (2022); code for the current analysis is available in the Supplemental Material available online.

Reporting

This study involved the analysis of previously harmonized secondary data as opposed to new data collection.

Ethical approval

The study was reviewed and approved as exempt human subjects research by the Institutional Review Boards of two of the Project Harmony PIs’ institutions, and data only from RCT participants who consented to the secondary analysis of their data were included.

Method

Study design, selection, and participants

The analysis was completed in compliance with the Preferred Reporting Items for Systematic Review and Meta-Analysis–Individual Patient Data Statement (Stewart et al., 2015); study selection is shown in Figure 1 and detailed in Hien et al. (2023). The criteria for inclusion of studies in Project Harmony consisted of (a) randomized clinical trials of a psychological, pharmacological, or combination intervention targeting either PTSD symptoms, AOD symptoms, or both; (b) pre- and posttreatment collection of PTSD and AOD outcomes; (c) participant minimum age of 18; and (d) participant current diagnosis at baseline of AOD (i.e., alcohol and/or other drug abuse or dependence as defined in the fourth edition of the Diagnostic and Statistical Manual of Mental Disorders [DSM-IV-TR], American Psychiatric Association [APA], 2000; substance use disorder as defined in the fifth editions of the DSM; DSM-5, APA, 2013; DSM-5-TR, APA, 2022) and either full or subthreshold PTSD (specific to the diagnostic criteria for the specific study, under DSM-IV-TR, APA, 2000; or DSM-5, APA, 2013). A total of 248 studies were identified in our initial screen, and 49 studies met our eligibility criteria. Of those 49 studies, we contacted study PIs and/or first authors inquiring about the availability of raw data for the study and their interest in joining (what would later become) the Consortium on Addiction, Stress, and Trauma (see Acknowledgments). Raw data were acquired from a total of 36 studies, representing 81 total treatment arms across studies (total N = 4,046) for what would become the primary Project Harmony data set; generalizability tests (i.e., differences in published descriptives between the studies acquired and those that were not) show that the studies acquired represented a random sample of the defined universe of PTSD+AOD treatment studies (Hien et al., 2023).

Fig. 1.

Preferred Reporting Items for Systematic Review and Meta-Analysis–Individual Patient Data statement study-selection flow diagram.

Coding of treatment classifications.

Groupings of treatment classes for this analysis included (a) nonmanualized, community-based treatment as usual (TAU); (b) evidence-based, manualized AOD behavioral treatments (e.g., relapse prevention); (c) integrated non-TF behavioral treatments (e.g., seeking safety); (d) nonintegrated TF behavioral treatments (e.g., prolonged exposure, cognitive-processing therapy); (e) integrated TF behavioral treatments (e.g., COPE, integrated cognitive-behavioral therapy, cognitive-behavioral therapy + structured writing); (f) placebo medications; (g) PTSD-targeting medications; (h) AOD-targeting medications; and (i) TF behavioral treatments paired with AOD medications. For information about the studies included under each classification, including details regarding expert ratings of treatment classifications through a survey of PTSD+AOD treatment experts (N = 38) found in Hien et al. (2022), see Table S1 in the Supplemental Material.

Pretreatment participant demographic and clinical characteristics.

The following individual-level variables were treated as covariates: gender, age, race/ethnicity, education level, population type (civilian, veteran, incarcerated), diagnosis of major depressive disorder at baseline, and nonstudy concomitant psychotropic pharmacotherapy use at baseline. The average age in the sample was 39.0 (SD = 11.2 years). The sample consisted of 53% males and 47% females and was 65% White, 25% African American, 7% Hispanic, and 3% other. Almost half (46%) of the sample were veterans, 50.7% had a high school education or less, 54.8% reported major depression at baseline, and 60.0% reported use of nonstudy-related psychiatric medications. Within each trial, randomization was expected to balance these variables across treatment arms. However, grouping of treatment classes across studies (see below) potentially compromises the original within-study randomization structure (Brincks et al., 2018; Morgan-López, McDaniel et al., 2022; Saavedra et al., 2021). This grouping turns a series of RCTs into a large de facto quasi-experiment and requires propensity score weighting for rebalancing of treatment classes across covariates. Before propensity score weighting, as reported in the supplemental material in Hien et al. (2023), covariate imbalance across conditions was observed that was statistically significant and/or had an effect size > |0.10| for age (d = 0.11), gender (d = 0.13), race/ethnicity (d = 0.15), marital status (d = 0.14), percentage baseline depression (d = 0.13), and percentage nonstudy medications (d = 0.19). Propensity weighting rendered imbalance on all of these variables to nonsignificant (ps > .9, ds < |0.06|; see Table 1).

Table 1.

Covariate Differences Across Treatment Classes

	Unweighted covariate differences				Propensity-score-weighted differences
	F(13, 3895)	p value	r 2	d	F(13, 3895)	p value	r 2	d
Age	0.79	.67	.003	0.11	0.35	.98	.001	0.06
Gender	1.08	.37	.004	0.13	0.35	.98	.001	0.06
Race/ethnicity	1.75	.04	.006	0.15	0.41	.96	.001	0.06
Education level	1.31	.19	.004	0.13	0.42	.96	.001	0.06
Marital status	1.54	.09	.005	0.14	0.44	.95	.001	0.06
Population type	0.55	.89	.002	0.09	0.09	.99	0	0
% Major depression at baseline	1.1	.35	.004	0.13	0.34	.98	.001	0.06
% Taking nonstudy psychotropic medication at baseline	2.8	< .001	.009	0.19	0.46	.94	.001	0.06

Treatment dosage as a posttreatment moderator.

As noted earlier, treatment dosage is defined as the proportion of (a) planned behavioral-intervention sessions that were attended by the patient and/or (b) intended dosage of PTSD- or AOD-targeting medication. The mean dosage across all conditions was 57.8%, ranging from a low of 48.1% for TF/nonintegrated treatment to a high of 59.5% for TF/integrated treatments. In the parent-study article (Hien et al., 2023), an effect size of d = 0.34 was observed for the unweighted mean of the unadjusted effects of treatment on dosage across treatment classes. Two challenges regarding dosage arise in the present analyses: (a) Dosage is affected by treatment assignment (and, in turn, affects treatment outcomes), given the literature cited earlier documenting reductions in treatment adherence among TF treatments, and (b) dosage is being treated as a moderator, requiring specialized methods for estimation of causal effects involving nonrandomized moderators that also treat dosage as a mediator (MacKinnon et al., 2020) but without formally testing mediation effects of dosage (Rosenbaum, 1984; Saavedra et al., 2023) because dosage is not thought of as a mechanism of action per se.

Outcomes of interest

Latent PTSD severity.

Latent PTSD-severity scores were estimated under the moderated nonlinear factor-analysis (MNLFA) framework (Bauer, 2017, 2023; Bauer & Hussong, 2009). A unidimensional latent variable for PTSD (Hien et al., 2023; Morgan-López, Hien et al., 2022) was estimated using 42 indicators of PTSD (21 symptoms from a clinical interview, 21 self-report symptoms) that were harmonized across the studies that had PTSD item-level data. The 42 symptoms include the 16 PTSD symptoms that are common to both the DSM-IV-TR (APA, 2000) and DSM-5 (APA, 2013) and DSM-5-TR (APA, 2022) diagnostic systems, the one symptom (sense of foreshortened future) that is unique to DSM-IV-TR (APA, 2000), and the four symptoms that are unique to DSM-5 (APA, 2013) and DSM-5-TR (APA, 2022). Details on MNLFA scale score estimation and item parameters are included in the supplemental materials of Hien et al. (2023).

Latent alcohol-use severity.

Latent alcohol-use-severity scores were estimated under MNLFA using two indicators: number of days of alcohol use in the past 30 days and any alcohol use to intoxication in the past 30 days.

Latent drug-use severity.

Latent drug-use-severity scores were also estimated under MNLFA. Binary indicators of any use in the past 30 days of the following substances were employed to support a six-indicator latent drug-use variable: heroin, other opioids (excluding heroin), sedatives, cocaine, other stimulants (excluding cocaine), and hallucinogens.

Data-analytic strategy

First, to address the imbalance on covariates that occurs with “mixing and matching” treatment arms across studies (e.g., Hien et al., 2023; Morgan-López, Hien, et al., 2022; Saavedra et al., 2021), propensity scores were estimated using a multinomial logit model incorporating all pretreatment covariates (i.e., excluding dosage; see below) and study-level fixed effects using SAS Proc GLIMMIX. Covariate balance across treatment classes was achieved for all covariates after propensity score weighting; all postweighting balance checks were below d < |0.10| (see supplemental materials in Hien et al., 2023).

Outcome model.

We used a propensity-score-weighted, three-level multilevel latent growth modeling structure estimated in Mplus (Version 8; Muthén & Muthén, 2017) that accounted for two forms of clustering: (a) repeated observations within participants and (b) participants within studies (i.e., “one-stage” meta-analysis of individual-patient data; Burke et al., 2017). At the between-persons level, random intercepts and slopes were both specified; at the study level, only a random intercept was specified to account for (residual) variation in outcomes across studies, consistent with the optimal variance component structure from previous analyses of Project Harmony data (Hien et al., 2023; Hill et al., 2024). Primary predictors were intervention dummy-variable indicators based on the treatment classification coding described in Hien et al. (2022, 2023) for each treatment class across studies: (a) TF (yes/no), (b) integrated (yes/no), (c) PTSD medication (yes/no), (d) AOD behavioral (yes/no), (e) AOD medication (yes/no), and (f) placebo medication (yes/no); treatment-class interaction effects were included as appropriate (e.g., TF × Integrated Treatments). Variables representing treatment-by-dosage interactions were created across all treatment-class combinations that occurred in the study by creating Number of Sessions × Dummy-Product Variables interactions (West et al., 1996); curvilinear treatment by dosage (i.e., Treatment × Dose² terms) were explored across all models but were removed because none were significant. Dosage was treated as a “mediator,” whereby the above-described treatment classes were modeled as predictors of dosage, which, in turn, predicted outcomes. For the outcome variables (PTSD severity, alcohol-use severity, drug-use severity), dosage, the treatment-class indicators, and the treatment-by-dosage interaction terms predicted within-treatments changes (i.e., from pretreatment to posttreatment) and posttreatment changes (i.e., from posttreatment to 12-month follow-up) in outcome slopes.

The incorporation of treatment-by-dosage interactions, in the context of dosage also being intermediate in the causal sequence between treatment classes and outcomes, parallels the structure of the potential-outcome mediation model proposed by VanderWeele and Vansteelandt (2009). However, whereas the typical interest in this model is formal testing of indirect/mediated-effect differences between treatment and control conditions (MacKinnon et al., 2020), our interest is in the differences in the dosage effect on outcomes between experimental treatments and TAU conditions on the direct-effect estimates net of (a) treatment-condition differences in dosage and (b) dosage effects on outcomes (i.e., Rosenbaum’s [1984] “net treatment difference” approach). Missing data were addressed via multiple imputation for clustered data using the R package mice (van Buuren & Groothuis-Oudshoorn, 2011) with 20 imputed data sets. Results were combined across the 20 data sets using Rubin’s (1976) rules for combining results in multiple imputation.

Results

We report key findings from tests of causal moderation of treatment dosage on PTSD+AOD outcomes. All other results not reported here replicated findings from Project Harmony primary-outcomes analyses as reported by Hien et al. (2023).

PTSD-severity outcomes

Higher treatment dosage was associated with steeper decreases in PTSD severity from pretreatment to posttreatment: For TAU, b = −0.38, SE = 0.16, z = −2.49, p < .001. Compared with TAU (at the average levels of treatment dosage), statistically significant decreases in PTSD severity from pretreatment to posttreatment were detected for TF psychotherapies (averaged across integrated and nonintegrated treatments), b = −0.30 (SD = 0.13), z = −2.25, p = .02; PTSD medications, b = −0.31 (SD = 0.10), z = −3.14, p = .002; and AOD medications, b = −0.41 ( SD = 0.09), z = −4.67, p < .001. No pretreatment to posttreatment interaction effects with treatment dosage were observed (all ps > .25) for any treatment classes (compared with TAU) except for TF psychotherapies. For TF interventions, the benefits were marginally larger (although not statistically significant) at higher dosage, b = −0.54 (SD = 0.33), z = −1.60, p = .10. This interaction was probed for TF/TAU differences at high, moderate, and low dosages (i.e., simple interactions; Aiken & West, 1991; Bauer & Curran, 2005), but instead of statistical significance at each level, we divided the predicted values at end-of-treatment and 12-month follow-up by the baseline standard deviation, the longitudinal analog of Cohen’s (1988) d effect size (Feingold, 2019). Figure 2 shows that at the end of treatment, the TAU/TF differences were d = 0.38 at high dosage (i.e., 75%), d = 0.24 at the average dosage, and no practical difference (d = 0.10) at low dosage (i.e., 25%). By 12-month follow-up, the TAU/TF differences were d = 0.55 at high dosage, d = 0.46 at the average dosage, and d = 0.25 for low dosage.

Fig. 2.

The effect of dosage on latent PTSD-severity outcomes. Compared with mean TAU dosage, high TAU dosage and TF psychotherapies (at both mean and high dosages) led to greater pretreatment to posttreatment reductions in PTSD severity. The y-axis reflects the standardized score estimated under moderated nonlinear factor analysis (see Hien et al., 2023). PTSD = posttraumatic stress disorder; TAU = treatment as usual; TF = trauma-focused.

Compared with TAU, decreases in PTSD severity from posttreatment through 12-month follow-up were observed for integrated treatments (averaged across TF and non-TF psychotherapies), b = −0.08 (SD = 0.04), t = −2.09, p = .04, and AOD medications, b = −0.23 (SD = 0.05), t = −4.87, p < .001. No other main effects or interaction effects with treatment dosage were detected for any treatment class (compared with TAU) from posttreatment to 12-month follow-up (all ps > .35).

Alcohol-use-severity outcomes

Higher treatment dosage was not related to changes in alcohol-use severity from pretreatment to posttreatment, b = 0.11 (SD = 0.11), t = 0.99, p = .32. Compared with TAU (at average levels of treatment dosage), significant pretreatment to posttreatment decreases in alcohol-use severity were observed for AOD medications, b = −0.66 (SD = 0.17), t = −3.82, p < .001, and placebo medications, b = −0.39 (SD = 0.16), t = −2.56, p < .01. For TF psychotherapies (averaged across integrated and nonintegrated treatments), the comparative decreases in alcohol-use severity from pretreatment to posttreatment relative to TAU were nonsignificant (p = .09, on average). This marginal result was modified by a significant TF-by-dosage interaction effect (see Fig. 3), which showed that steeper reductions in alcohol-use severity were observed at higher dosage of TF treatment, b = −0.53 (SD = 0.15), t = −3.51, p < .001. Probing of these interaction effects at high, moderate, and low dosages showed that at the end of treatment, the TAU/TF differences were d = 0.52 at high dosage (i.e., 75%), d = 0.37 at the average dosage, and d = 0.22 at low dosage (i.e., 25%). By 12-month follow-up, the TAU/TF differences were d = 0.32 at high dosage, d = 0.32 at the average dosage, but no practical difference for low dosage (d = .03).

Fig. 3.

The effect of dosage on latent alcohol-use-severity outcomes. Compared with mean TAU dosage, higher dosage of TF psychotherapies yielded larger effects on alcohol-use severity from posttreatment to follow-up. The y-axis reflects the standardized score estimated under moderated nonlinear factor analysis (see Hien et al., 2023). TAU = treatment as usual; TF = trauma-focused.

Drug-use-severity outcomes

Higher treatment dosage was not related to changes in drug-use severity, b = –0.12 (SD = 0.13), t = –0.91, p = .35. Compared with TAU (at the average levels of treatment dosage), only TF psychotherapies (across integrated and nonintegrated treatments) showed significant decreases in drug-use severity from pretreatment to posttreatment, b = −0.08 (SD = 0.03), t = −2.01, p = .04. This effect may be more pronounced for TF/nonintegrated psychotherapies than TF/integrated psychotherapies because the TF × Integrated interaction term approached significance, b = 0.07 (SD = 0.04), t = 1.72, p = .08. No other main effects or interactions with dosage were observed between either pretreatment to posttreatment or posttreatment to 12-month follow-up.

Discussion

PTSD+AOD is a complex comorbidity associated with significant treatment need. Although effective forms of PTSD+AOD care exist, not everyone offered PTSD+AOD treatment receives the full intended-treatment course (Coffey et al., 2016; Hien et al., 2009; Imel et al., 2013; Lappan et al., 2020; Mills et al., 2012; Simpson et al., 2021). Questions about treatment dosage (i.e., sessions attended/medication taken) are therefore critical. In the present study, we attempted to address this research need by estimating the causal effects of dosage on PTSD+AOD outcomes for different intervention classes, (a) accounting for covariate imbalance when combining treatment conditions across studies and (b) treating dosage as predictor and moderator of treatment outcome that is also affected by the treatment.

Analyses tested the causal moderating effect of observed dosage separately on PTSD, alcohol-use-severity, and drug-use-severity outcomes. We investigated the effects of multiple types of behavioral, pharmacological, and combined interventions for PTSD+AOD. The TF versus TAU comparison was the only type of PTSD+AOD treatment comparison to interact with proportion of intervention sessions attended on clinical outcomes, that is, a significant dosage moderating effect was observed in alcohol-use-severity-outcome models; a similar moderation effect trended toward significance in PTSD-severity-outcome models and was not significant in drug-use-severity-outcome models. Analyses failed to detect interaction effects involving proportion of sessions attended in models of symptom changes from immediately after treatment to 12-month follow-up, suggesting that maintenance of TF treatment gains (compared with TAU) was not dependent on the proportion of sessions attended during treatment.

Existing research (Holmes et al., 2019; Mills et al., 2016; Schacht et al., 2017; Straus et al., 2022) and theory (e.g., Foa et al., 2018; Resick et al., 2024) suggest that maladaptive posttraumatic cognitions and avoidance of trauma-related cues maintain PTSD. Accordingly, TF interventions prioritize strategies that facilitate approaching trauma-related cognitions and feelings and emotional processing of the trauma. They address self-blame and negative thoughts about oneself, other people, and the world and confront safe yet distressing trauma cues. Hien et al. (2023) previously reported findings from Project Harmony showing that at average levels of dosage across all PTSD+AOD treatment types, TF psychotherapy paired with AOD medications (but not TF psychotherapy alone) showed the largest effect sizes for PTSD and alcohol-use outcomes compared with TAU. By treating the proportion of sessions attended as a factor that influences treatment outcome but is also influenced by treatment type, the current study (a) extended prior work that treated dosage as a simple covariate and (b) highlighted the importance of an explicit trauma focus during PTSD+AOD treatment.

A novel feature of this study is that our analytic approach allowed us to interpret the comparative effectiveness of PTSD+AOD treatment classes relative to TAU. Analyses of dosage effects assume that dosage level is not confounded with individual-patient characteristics. Such lack of confounding is optimally achieved by randomly assigning patients to both treatment dosage and treatment type. In practice, dosage is often not random because patients may miss treatment sessions or prematurely withdraw from treatment altogether before completing the intended number of sessions for systematic reasons. Furthermore, the treatment condition itself may lead to differential dosage given that TF treatments delivered to patients with PTSD+AOD have been associated with relatively high dropout rates in previous reviews (e.g., Simpson et al., 2021). Kehle-Forbes et al. (2025) reported evidence that some patients find TF treatments difficult to tolerate (e.g., Kehle-Forbes et al., 2016) but observed optimal outcomes with close to full TF-treatment compliance. The use of propensity score weighting (Rosenbaum & Rubin, 1983) and net-treatment-difference modeling (Rosenbaum, 1984, 2020) allows researchers to control for multiple key individual-patient characteristics, mitigating potential confounding of predictors affected by treatment assignment (i.e., dosage) by individual-patient characteristics.

The collective findings from Project Harmony are that TF treatments yield greater reductions in alcohol use and PTSD severity compared with TAU at higher levels of intervention session attendance (Hien et al., 2023). Therefore, one clinical implication is that providers should inform patients about a potential dose-response effect during initial collaborative treatment-planning discussions. Clinicians providing TF treatment to patients with PTSD+AODs might also directly encourage and reinforce treatment engagement and retention early in the intervention (e.g., incorporating motivational-interviewing methods; Murphy et al., 2009). Likewise, a critical future direction based on our findings is to study ways to optimize effective TF PTSD+AOD treatments. For example, treatment retention could be incentivized via contingency management (e.g., offering rewards for attending treatment sessions; Peck et al., 2023; Pfund et al., 2022; Schacht et al., 2017) or by focusing on helping patients access more substance-free reinforcers rather than focusing on substance consumption reduction alone (see McKay, 2017). Future work should additionally strive to make trauma processing more potent and efficient, such as by testing massed, condensed, and other intensive delivery formats to provide patients with as much TF content as possible before logistical or emotional barriers can interfere with treatment receipt (e.g., Dell et al., 2023; Weinstein et al., 2023). Additional work on TF treatments addresses the extent to which adjunctive interventions that address many patients’ difficulties with exposures (e.g., pretreatment and posttreatment processing of exposures with care managers; Petrova et al., 2025) have the potential to increase engagement with TF treatments and optimize treatment outcomes.

These analyses did not detect any statistically significant interactions between dosage and any treatment type on drug-use-severity outcomes, although a previously undetected effect of TF treatments on reductions in drug-use severity (net of dosage as a posttreatment covariate) was detected. A previous report from Project Harmony, which treated observed dosage as a pretreatment covariate (Hien et al., 2023), showed that no active treatment was significantly superior to TAU on drug-use outcomes from either pretreatment to posttreatment or posttreatment to 12-month follow-up (TAU pretreatment to posttreatment effect size was d = −0.63). Although one explanation is that many interventions, regardless of dosage, are equally beneficial for reducing drug-use severity among people with PTSD, it is also possible that “true effects” were obfuscated because of one particular methodological issue. It has been noted in the quasi-experimental literature that covariates—or in this case, moderators—that are affected by treatment assignment should be treated in an equivalent manner to mediators and that failing to do so can bias estimates of interest (Loh & Ren, 2025; MacKinnon et al., 2000; Rosenbaum, 1984). It may be that this differential treatment of dosage (i.e., pretreatment vs. posttreatment variable) obscured effects of TF treatment, particularly when explicitly equated against dosage for TAU and other treatment classes. Researchers in the PTSD+AOD treatment space are encouraged to perform similar analyses and, in some cases, reanalysis of the effects of TF treatments on AOD outcomes because these effects may have been dampened by not incorporating posttreatment effects of dosage in outcome models.

The strengths of this study include a large and diverse sample (N = 4,046) resulting from combining data from 36 RCTs, harmonized latent severity indicators derived from disparate outcome-assessment instruments used in the original RCTs, and a novel analytic framework for the examination of moderated comparative effectiveness that critically adjusted for key covariates. At the same time, our findings should be contextualized within study limitations. For example, although this study compared multiple PTSD+AOD treatment approaches to TAU, these categories do not represent an exhaustive list of PTSD+AOD-treatment approaches. Likewise, active treatment classes were compared against TAU, not each other. In addition, we operationalized treatment dosage as the proportion of planned sessions and/or maximum medication dosage (per the protocol) that participants completed; we therefore cannot identify whether PTSD+AOD-severity outcomes would have differed if treatment dosage had been defined by level of engagement during treatment sessions or degree or quality of between-sessions practice (i.e., “homework”) completion; nor can we identify whether participants reached maximum dosage for pharmacotherapies before dropping out of treatment. An additional limitation and a potential opportunity for further study involves the possibility that complex interactions exist between treatment conditions, dosage, and pretreatment factors (e.g., demographics, comorbid depression, traumatic brain injury) such that pretreatment factors may affect treatment tolerance in addition to interacting with self-selected dosage, leading to differential impact on PTSD and/or AOD measurement (Saavedra et al., 2022) and treatment outcomes (Saavedra et al., 2024). Illumination of these complex interactions may benefit from principled and careful use of machine-learning tools. Finally, analyses of long-term durability of PTSD+AOD-treatment gains were limited to 12-month follow-up, precluding us from drawing conclusions about PTSD+AOD-severity trajectories beyond 1 year after treatment.

Conclusions

With the current study, we built on previous work to investigate the differential effect of treatment dosage on PTSD, alcohol-use, and drug-use severity among individuals with PTSD+AOD receiving different types of treatment. Using a novel methodological framework permitting the examination of differential dosage effects, we found that treatment dosage is especially important for patients with PTSD+AOD receiving TF psychotherapy (whether integrated with AOD treatment or not; Back et al., 2015). Our findings also underscore the complex nature of AOD use among individuals with PTSD, including how the relationship between PTSD, alcohol-use, and drug-use severity may vary over time because of factors other than treatment type. Future studies examining a broader set of treatment categories and operational definitions of treatment dosage will be critical for advancing the field’s knowledge of and approaches to the treatment of individuals with PTSD+AOD.

Supplementary Material

Additional supporting information can be found at http://journals.sagepub.com/doi/suppl10.1177/21677026251351276