Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Jul 8.
Published before final editing as: J Stud Alcohol Drugs. 2026 Jun 4:10.15288/jsad.25-00478. doi: 10.15288/jsad.25-00478

Predictors of Follow-Up Engagement in an Adolescent Alcohol Longitudinal Cohort

Tommy Gunawan a, Ty Brumback b, Michael L Thomas c, Laika Aguinaldo d, Daniel Lopez e, Luis F S Castro-de-Araujo f, Herry Patel d, Emily Schulze d, Sonja C Eberson d, Mary J Meloy d, Duncan B Clark g, David B Goldston h, Kate B Nooner i, Sandra A Brown d, Susan F Tapert d, Vijay A Ramchandani j, Alejandro D Meruelo d
PMCID: PMC13339738  NIHMSID: NIHMS2187523  PMID: 42240465

Abstract

Objective:

Attrition can bias longitudinal findings, including those on alcohol use. We examined predictors of follow-up engagement in a multi-site neuroimaging cohort of youth.

Method:

Data were drawn from the first nine annual waves of the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA), which recruited 831 participants aged 12–21 years in 2013–2014 (52% female). Follow-up engagement was modeled as the total number of completed follow-up visits (0–9, excluding baseline) using Poisson regression with multiple imputation and family-cluster–robust standard errors. Secondary analyses examined categorical engagement levels. Primary predictors included exceeding age- and sex-specific baseline alcohol use thresholds, externalizing symptoms, race, and demographic and family factors.

Results:

Overall engagement was high, with participants completing an average of approximately 9 of 10 possible visits. In Poisson models, exceeding baseline alcohol use thresholds was associated with a modest reduction in follow-up visits, although this did not reach statistical significance (IRR = 0.940, 95% CI: 0.881–1.004, p = .064). Externalizing symptoms were not significantly associated with engagement (IRR = 1.001, 95% CI: 0.998–1.003, p = .647). Sibling co-enrollment was associated with greater engagement (IRR = 1.058, 95% CI: 1.021–1.097, p = .002), while other factors showed small or non-significant associations.

Conclusions:

Follow-up engagement was generally high across adolescence into young adulthood. Exceeding baseline alcohol use thresholds showed a modest association with reduced engagement, although not statistically significant in the count model. Externalizing symptoms were not associated with engagement. Categorical analyses suggested a stronger association between exceeding baseline alcohol use thresholds and lower engagement groups. These findings may help identify participants who could benefit from additional engagement efforts.

Keywords: Engagement, Longitudinal cohort, Alcohol use, Retention strategies, Sibling enrollment, Youth behavior

Introduction

Longitudinal studies are critical for understanding developmental trajectories, health outcomes, and behavioral change over time, enabling examination of covariation, temporal precedence, and prediction. Sustaining high participant levels across multiple waves of data collection is challenging (Bitsko, 2022; Feldstein Ewing et al., 2018; Hanna et al., 2014; Smith et al., 2017; Teague et al., 2018). Attrition can include study withdrawal as well as intermittent nonresponse (i.e., completing some but not all follow-up visits). Both forms threaten the validity and generalizability of findings, reduce statistical power (Deke et al., 2015), and if non-random, can introduce bias in findings (Bamer et al., 2020; Peterson et al., 2012; Weuve et al., 2012). Importantly, attrition is often conceptualized dichotomously (retained vs. dropped), which may obscure meaningful heterogeneity in participation patterns. Intermittent nonresponse can differ from complete disengagement in both causes and consequences for bias, statistical power, and analytic strategies (e.g., imputation assumptions, weighting approaches). Differentiating high, intermittent, and minimal engagement may therefore yield more precise insight into retention dynamics and help identify subgroup-specific risk profiles that would not be evident in a binary framework.

Several factors have been suggested to influence attrition, including age, sex, socioeconomic status (SES), and substance use. Developmental stage may shape participation patterns: younger participants often depend on parental involvement and transportation (Smith et al., 2017), whereas older adolescents and young adults may face competing schedule demands, life transitions, disinterest, or geographic mobility that threaten longitudinal study participation (Costello et al., 1999; Odierna & Schmidt, 2009). Sex differences in retention have also been observed in adolescent cohorts, with males more likely to miss assessments than females. For example, the National Longitudinal Study of Adolescent to Adult Health (Add Health) reported higher response rates among female participants across waves (Harris et al., 2019), and a meta-analysis found that studies with greater proportions of male participants tended to have lower overall retention rates (Teague et al., 2018).

Ethnicity may also shape patterns of longitudinal engagement through cultural norms surrounding family involvement, trust in research institutions, and differential exposure to structural barriers such as transportation access or mobility (George et al., 2014; Yancey et al., 2006). Hispanic ethnicity, in particular, has been examined in U.S.-based retention studies due to documented disparities in research participation and representation. Evaluating ethnicity in conjunction with SES and family factors may help clarify whether observed engagement patterns reflect structural context rather than individual behavioral risk.

The influence of SES on attrition is complex and context dependent. Levine (Levine, 2017) posits that SES operates through distinct psychosocial orientations: autonomy (greater perceived control, expectations for independent decision-making, prioritization of one’s own schedule) and interdependence (heightened obligations to coordinate decisions with family or work networks). These orientations, together with associated differences in time and resource flexibility, institutional trust, and perceived opportunity costs, shape engagement and retention behaviors (e.g., responsiveness to outreach, capacity to reschedule, transportation planning). Some studies observed that higher-income families faced greater scheduling constraints or perceived less benefit from continued participation (Thornton et al., 2016; Wallace & Gibb, 2013). These inconsistencies underscore the importance of evaluating SES in conjunction with other factors in relation to study engagement.

Family-level characteristics may also influence longitudinal engagement. Research on retention in pediatric and adolescent cohorts suggests that household-level stability, caregiver involvement, and family organization are associated with sustained participation across waves (Booker et al., 2011; Robinson et al., 2007; Teague et al., 2018). Sibling co-enrollment may reinforce participation through shared transportation, coordinated scheduling, and normative reinforcement of study involvement within the household. In contrast, family history of alcohol problems may index broader familial stress, instability, or competing demands that could shape engagement behaviors. Despite these plausible mechanisms, family-level predictors are rarely modeled alongside individual demographic and behavioral factors in adolescent longitudinal research.

Substance use, especially alcohol drinking, has been linked to higher rates of study attrition in youth-focused longitudinal research, and family history of alcohol problems may represent an additional contextual risk factor shaping engagement. An analysis of 30 adolescent and young-adult cohorts found that heavier baseline alcohol use was consistently associated with higher subsequent loss-to-follow-up (Keyes et al., 2020). Studies from national samples such as the National Longitudinal Survey of Youth (Bureau of Labor Statistics, U.S. Department of Labor, 2023) suggest that early alcohol use may contribute to irregular participation, avoidance of structured follow-up, or greater overall disengagement. Large-scale adolescent studies such as AddHealth (Harris et al., 2019) suggest that youth engaging in risky behaviors may be more likely to drop out of long-term studies.

The present study extends this literature by moving beyond binary retained/dropped frameworks by modeling engagement both continuously and categorically, allowing differentiation of patterns such as high, intermittent, and minimal participation. It also evaluates alcohol use alongside baseline demographics (age, sex, ethnicity, SES) and family context (family history of alcohol problems, sibling co-enrollment)—factors rarely examined together—thereby clarifying whether associations between substance use and follow-up engagement persist after accounting for demographic and familial influences. Identifying substance use–related patterns independent of sociodemographic factors has practical implications for retention planning, as it can inform targeted outreach, tailored scheduling approaches, or adaptive engagement strategies for higher-risk youth.

In this study, patterns of engagement were examined over ten years in a multisite longitudinal study of youth enrolled at ages 12–21 years and followed annually. We examined predictors of follow-up engagement, modeled both as a continuous count of completed visits and as categorical patterns of participation (high, intermittent, minimal). By distinguishing these engagement patterns, we aimed to better characterize heterogeneity in longitudinal participation and inform retention strategies tailored to developmental and behavioral risk profiles. Predictors were baseline demographic factors (age, sex, ethnicity, SES), exceeding baseline alcohol use thresholds, and family history of alcohol use problems, and sibling enrollment in the same study. Based on prior research, we hypothesized that participants exceeding baseline alcohol use thresholds, older age at baseline, and male sex would complete fewer follow-up visits.

Importantly, externalizing symptoms represent a major developmental correlate of both substance use and engagement-related behaviors (Cox et al., 2021; Farmer et al., 2016). This raises the question of whether associations between alcohol use and follow-up reflect substance-specific effects or broader behavioral dysregulation. Accordingly, externalizing symptoms were included to evaluate whether alcohol-related differences in engagement persist beyond general externalizing liability.

Method

Analyses used baseline through 9-year follow-up data from the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA) (Brown et al., 2015), a longitudinal study investigating the impact of alcohol use on brain development in youth recruited at ages 12 to 21 (N = 831; mean age = 16.1 years, SD = 2.5) and followed annually. Participants were recruited from the regions surrounding five research sites (UC San Diego, Duke University, University of Pittsburgh, SRI International, and Oregon Health & Science University). One-third of participants had a sibling co-enrolled in the study. Participants who formally withdrew from the study or died during follow-up (n = 91) were excluded from the primary engagement models because they were no longer eligible for ongoing annual assessments. Thus, the primary analyses examine follow-up engagement among participants eligible for continued participation.

The NCANDA study protocol was reviewed and approved by the institutional review board at each site. Written informed consent was obtained from all participants; for those under 18 years of age, youth assent and parent/guardian consent were obtained. All analyses adhered to ethical guidelines for research involving human subjects, including the Declaration of Helsinki.

Retention Procedures

NCANDA employed a structured, multi-pronged retention framework across sites, including ongoing participant contact (e.g., mailed communications, electronic outreach, reminder calls), logistical supports (e.g., flexible scheduling, transportation assistance), and age-adaptive engagement strategies. Assessments were primarily conducted in person, including neuroimaging visits, with limited remote accommodations when necessary (e.g., COVID-19 disruptions).

Measures

Baseline assessments included comprehensive substance use, mental health, demographic, neurocognitive, and neuroimaging measures.

Alcohol use at baseline was defined based on whether participants exceeded age- and sex-specific drinking thresholds at enrollment, consistent with National Institute on Alcohol Abuse and Alcoholism (NIAAA) guidelines (National Institute on Alcohol Abuse and Alcoholism, 2011). This binary variable distinguished youth with minimal versus higher exposure and does not represent heavy or problematic drinking.

Family history of alcohol problems was assessed as a binary variable based on combined youth and parent report (Rice et al., 1995). Socioeconomic status (SES) was assessed using a modified MacArthur Sociodemographic Questionnaire (Giatti et al., 2012). Baseline externalizing symptoms were measured using the Youth Self Report (YSR) externalizing T-score and included as a covariate to account for potential confounding between alcohol use and broader behavioral dysregulation.

Race was included as a primary predictor using collapsed categories (White [reference], Black, Asian/Other) to ensure stable estimation.

Outcome Definition

The primary outcome was the total number of completed follow-up visits across nine annual assessments (range 0–9, excluding baseline). A visit was considered complete if the participant contributed data to at least one core study domain.

For comparability with prior literature, a secondary categorical outcome was defined based on the proportion of completed follow-up visits: High Follow-Up (≥80%; 8–9 visits), Intermittent Follow-Up (3–7 visits), and Minimal Follow-Up (<30%; 1–2 visits). These thresholds were selected to align with prior retention research (Robinson et al., 2007).

Statistical Analyses

Primary Analysis: Count Model of Follow-Up Engagement

The primary analysis modeled follow-up engagement as a count outcome using Poisson regression with robust standard errors. Follow-up engagement was modeled at the participant level as the number of completed annual follow-up visits from 0 to 9, excluding baseline. This approach preserves the full distribution of participation and avoids reliance on arbitrary categorical thresholds. Models were estimated across multiply imputed datasets and pooled using Rubin’s rules (Rubin, 1987), with standard errors clustered at the family level to account for non-independence among co-enrolled siblings.

Predictors included baseline age, sex, Hispanic ethnicity, SES, sibling co-enrollment, exceeding age- and sex-specific baseline alcohol use thresholds, family history of alcohol problems, race, and externalizing symptoms. Analyses including externalizing symptoms were restricted to participants with available baseline externalizing data (n=734).

Secondary Analyses: Categorical Models of Follow-Up Classification

To facilitate comparison with prior studies, follow-up engagement was also modeled categorically using multinomial logistic regression with High Follow-Up as the reference group. Pairwise logistic regression comparing Minimal and Intermittent Follow-Up groups was also conducted (Supplementary Table A1). Multinomial models were estimated using the nnet::multinom function (Ripley & Venables, 2025), with family-clustered robust standard errors (Cameron & Miller, 2015).

Interaction Analysis

A sensitivity analysis examined an interaction between baseline age and alcohol use to evaluate whether associations with follow-up engagement varied across developmental stages. This interaction was tested in complete-case multinomial models and further characterized using model-derived predicted probabilities across representative ages. Due to instability in multiply imputed models, the interaction term was not included in the primary pooled analyses.

Missing Data

Baseline covariates exhibited minimal item-level missingness (<1%). Little’s test (Li, 2013) indicated data were not missing completely at random (χ2(244) = 765, p < .001), and analyses proceeded under a missing at random assumption. Multiple imputation by chained equations (mice package version 3.18.0) (Buuren & Groothuis-Oudshoorn, 2011, p. 20) was applied to baseline predictors and auxiliary variables (m = 5 datasets), with estimates pooled using Rubin’s rules. The outcome variable was fully observed and not imputed.

Additional Sensitivity Analyses

Robustness of findings was evaluated through complete-case models, alternative thresholds for follow-up classification (Supplementary Table B, Supplementary Figure A), and testing non-linear effects of SES using a quadratic term. These analyses did not materially alter the primary findings.

Multicollinearity

Variance inflation factors (Craney & Surles, 2002) indicated no problematic multicollinearity (all VIFs < 4, max = 3.78 for age). Due to collinearity between site and SES, site indicators were excluded.

Results

Sample Characteristics

The dataset included 545 High Follow-Up Completers (74%), 162 Intermittent Follow-Up Completers (22%), and 33 Minimal Follow-Up Completers (4%) (Table 1). Across the full analytical sample (n = 740), participants completed an average of 9.1 out of 10 possible visits, indicating an overall follow-up rate of approximately 91%. Baseline demographic and behavioral characteristics were generally similar across follow-up groups, with the exception of exceeding baseline alcohol use thresholds and sibling co-enrollment, which differed significantly by classification.

Table 1. Summary Statistics by Follow-Up Classification.

Number of visits reflects all visits including baseline (descriptive only). P-values reflect comparisons across groups using ANOVA or chi-square tests. Race was excluded from the table for simplicity; no significant differences in race distribution were observed across follow-up groups (χ2 = 9.25, df = 6, p = 0.160).

Variable High Follow-Up Completers (n=545) Intermittent Follow-Up (n=162) Minimal Follow-Up (n=33) p-value Test Type
Age (Mean ± SD) 16.09 ± 2.55 16.14 ± 2.37 16.64 ± 2.58 0.475 ANOVA
# Visits Completed (Mean ± SD) 9.79 ± 0.40 6.58 ± 1.40 2.18 ± 0.77 – –
% Male 47.9% 52.5% 39.4% 0.333 χ2
% Hispanic (Yes) 12.3% 10.5% 12.1% 0.824 χ2
% Exceeds Alcohol Drinking Threshold (Yes) 13.2% 25.9% 15.2% <0.001 χ2
SES (Mean ± SD) 8.00 ± 2.23 8.02 ± 2.20 7.94 ± 2.29 0.982 ANOVA
% Siblings Enrolled (Yes) 37.4% 35.8% 6.1% 0.001 χ2
% Family History Alcohol Problem (Yes) 9.4% 11.1% 18.2% 0.237 χ2
Open in a new tab

Primary Analysis: Count Model of Follow-Up Engagement (Table 2)

Table 2. Poisson Regression Predicting Number of Completed Follow-Up Visits (0–9).

Poisson regression results predicting the number of completed follow-up visits (range 0–9, excluding baseline) in the NCANDA cohort. Estimates are pooled across multiply imputed datasets using Rubin’s rules, with standard errors adjusted for clustering at the family level. Results are presented as incidence rate ratios (IRRs) with 95% confidence intervals, reflecting the relative change in the expected number of completed visits associated with each baseline predictor.

Predictor Estimate SE IRR 95% CI p-value
Intercept 2.057 0.019 7.822 (7.530, 8.126) <.001
Age (centered) −0.002 0.004 0.998 (0.989, 1.006) 0.606
Male (vs Female) −0.004 0.020 0.996 (0.958, 1.035) 0.825
Hispanic Ethnicity (Yes vs No) −0.001 0.031 0.999 (0.940, 1.061) 0.965
Race: Black (vs White) −0.065 0.035 0.937 (0.874, 1.004) 0.063
Race: Asian/Other (vs White) −0.019 0.029 0.981 (0.927, 1.039) 0.521
Exceeds Alcohol Drinking Threshold (Yes vs No) −0.062 0.033 0.940 (0.881, 1.004) 0.064
Externalizing Symptoms (centered) 0.001 0.001 1.001 (0.998, 1.003) 0.647
Socioeconomic Status (centered) 0.002 0.005 1.002 (0.993, 1.012) 0.655
Sibling Co-enrolled (Yes vs No) 0.056 0.018 1.058 (1.021, 1.097) 0.002
Family History of Alcohol Problems (Yes vs No) −0.047 0.038 0.954 (0.885, 1.028) 0.219
Open in a new tab

Exceeding age- and sex-adjusted alcohol use thresholds at baseline was associated with fewer completed follow-up visits (IRR = 0.940, 95% CI: 0.881–1.004, p = .064), corresponding to approximately a 6% reduction in expected follow-up participation, although this did not reach statistical significance. Externalizing symptoms were not significantly associated with follow-up engagement (IRR = 1.001, 95% CI: 0.998–1.003, p = .647), indicating no meaningful relationship.

Older baseline age was not significantly associated with completed visits (IRR = 0.998, p = .606), and male sex was also not associated with follow-up engagement relative to females (IRR = 0.996, p = .825). Hispanic ethnicity was not associated with follow-up participation (IRR = 0.999, p = .965). Race effects were small and did not reach statistical significance, with Black participants completing slightly fewer visits than White participants (IRR = 0.937, p = .063), and participants categorized as Asian/Other showing no meaningful difference (IRR = 0.981, p = .521).

Family-level predictors showed a more consistent association with engagement. Sibling co-enrollment was associated with a modest increase in completed follow-up visits (IRR = 1.058, p = .002), while a positive family history of alcohol problems was not significantly associated with completed visits (IRR = 0.954, p = .219). Socioeconomic status was not significantly associated with follow-up engagement (IRR = 1.002, p = .655).

Secondary Analyses: Categorical Follow-Up Classification

To facilitate comparison with prior literature, follow-up engagement was also modeled categorically (High, Intermittent, Minimal) using multinomial logistic regression (Table 3). Consistent with the primary count model, exceeding baseline alcohol use thresholds was associated with increased odds of Intermittent versus High follow-up (OR = 2.76, 95% CI: 1.68–4.54, p < .001), while sibling co-enrollment was associated with reduced odds of Minimal follow-up (OR = 0.11, 95% CI: 0.03–0.48, p = .003) (Figure 1). Other predictors, including age, sex, socioeconomic status, Hispanic ethnicity, and family history of alcohol problems, were not significantly associated with follow-up classification in the primary multinomial model.

Table 3. Secondary Multinomial Logistic Regression Results Predicting Follow-Up Classification.

Multinomial logistic regression results predicting follow-up classification using High Follow-Up Completers as the reference. Odds ratios represent the likelihood of being in the Intermittent or Minimal Follow-Up groups relative to the High Follow-Up group. Coefficients and standard errors are pooled across multiply imputed datasets, with standard errors cluster-robust at the family level.

Predictor Classification Coefficient Estimate Standard Error (SE) Odds Ratio 95% Confidence Interval p-value
Age (Centered) Intermittent Follow-Up −0.062 0.040 0.94 (0.87, 1.02) 0.120
Minimal Follow-Up 0.070 0.082 1.07 (0.91, 1.26) 0.392
Male Intermittent Follow-Up 0.206 0.182 1.23 (0.86, 1.76) 0.257
Minimal Follow-Up −0.335 0.378 0.72 (0.34, 1.50) 0.376
Hispanic Intermittent Follow-Up −0.253 0.312 0.78 (0.42, 1.43) 0.418
Minimal Follow-Up −0.096 0.544 0.91 (0.31, 2.64) 0.859
Alcohol Use Intermittent Follow-Up 1.020 0.253 2.76 (1.68, 4.54) <0.001
Minimal Follow-Up −0.216 0.571 0.81 (0.26, 2.47) 0.706
SES Intermittent Follow-Up −0.008 0.041 0.99 (0.92, 1.08) 0.848
Minimal Follow-Up 0.017 0.080 1.02 (0.87, 1.19) 0.832
Sibling Enrolled Intermittent Follow-Up −0.015 0.209 0.99 (0.65, 1.49) 0.943
Minimal Follow-Up −2.170 0.737 0.11 (0.03, 0.48) 0.003
FH Alcohol+ Intermittent Follow-Up 0.070 0.302 1.07 (0.59, 1.94) 0.816
Minimal Follow-Up 0.645 0.458 1.91 (0.78, 4.68) 0.160
Open in a new tab

Figure 1: Forest Plot of Odds Ratios (ORs) and 95% confidence intervals (CIs) for Predictors of Follow-Up Classification in Multinomial Logistic Regression.

Figure 1:

Open in a new tab

The model compares the likelihood of classification as Intermittent Follow-Up (n=162) or Minimal Follow-Up (n=33) relative to the High Follow-Up (n=545) reference group. Statistically significant predictors are shown in blue. The vertical dashed red line indicates the null value (OR=1). Exceeding baseline alcohol use thresholds at enrollment was significantly associated with increased odds of being in the Intermittent Follow-Up group, while sibling co-enrollment was strongly protective against classification into the Minimal Follow-Up group; all other predictors were not statistically significant.

In a supplementary pairwise logistic regression directly comparing Minimal and Intermittent Follow-Up groups (Supplementary Table A1), exceeding baseline alcohol use thresholds did not distinguish these groups, whereas sibling co-enrollment remained strongly associated with reduced odds of Minimal relative to Intermittent follow-up. Older age and male sex were also associated with reduced odds of Minimal versus Intermittent engagement in this comparison.

Interaction Between Age and Alcohol Use

We conducted a sensitivity analysis to examine whether the association between baseline alcohol use and follow-up engagement varied across developmental stages by including an age × alcohol use interaction. In complete-case multinomial models, inclusion of the interaction term significantly improved model fit (χ2 = 7.37, df = 2, p = .025).

Model-derived predicted probabilities (Supplementary Table C; Supplementary Figure B) indicated a developmentally contingent pattern. Among adolescents who did not exceed baseline alcohol use thresholds, the probability of High Follow-Up remained consistently high across ages, with modest increases over time. In contrast, among adolescents who exceeded age- and sex-specific baseline alcohol use thresholds, the probability of High Follow-Up was substantially lower at younger ages but increased with age, while Intermittent Follow-Up was elevated in early adolescence and declined over time. Minimal Follow-Up remained relatively low across age groups.

Attempts to incorporate the interaction into multiply imputed models resulted in unstable estimates, and therefore the interaction term was not included in the primary pooled models. These findings should be interpreted as exploratory but suggest that alcohol-related differences in engagement may be most pronounced earlier in adolescence.

Sensitivity Analyses

A series of sensitivity analyses supported the robustness of the primary findings. Alternative thresholds for defining follow-up classification (≥70%, ≥80%, ≥85%; Supplementary Table B) yielded similar patterns of association, with exceeding baseline alcohol use thresholds consistently predicting reduced engagement and sibling co-enrollment associated with greater retention.

Across model specifications, exceeding baseline alcohol use thresholds remained the most consistent predictor of reduced follow-up engagement, while externalizing symptoms showed minimal practical impact on participation.

Discussion

This study examined follow-up engagement in a longitudinal adolescent cohort, focusing on sociodemographic characteristics, alcohol use, and family factors, among participants who remained active in the study. Using multiply imputed data to address missingness, we identified baseline characteristics associated with follow-up engagement and emphasized the value of proactive retention strategies in adolescent cohort studies. In primary analyses modeling engagement as a continuous count of completed follow-up visits, exceeding age- and sex-specific baseline alcohol use thresholds was associated with fewer completed visits over the follow-up period, although this association did not reach statistical significance in the participant-level count model. Externalizing symptoms were not significantly associated with engagement, suggesting that the observed relationship between alcohol use and follow-up participation is less likely to be attributable to broader externalizing liability. Secondary multinomial models yielded a consistent pattern, with exceeding baseline alcohol use thresholds associated with increased odds of intermittent relative to high follow-up, and sibling enrollment associated with reduced odds of minimal follow-up. Hispanic ethnicity, age, SES, and family history of alcohol problems showed limited or inconsistent associations across modeling approaches.

Engagement Patterns and Alcohol Use Risk

Exceeding baseline alcohol use thresholds showed a modest and directionally consistent association with reduced follow-up engagement, reflected in a small reduction in completed visits across the full range of participation. In categorical models, this pattern manifested as greater odds of intermittent follow-up relative to high engagement, suggesting that exceeding baseline alcohol use thresholds is more strongly associated with irregular participation than with complete disengagement. This pattern is consistent with prior work linking alcohol involvement to reduced study compliance (Duan et al., 2009; Keyes et al., 2020; McCabe et al., 2006). Although externalizing symptoms were not significantly associated with engagement, the observed pattern suggests that alcohol-related differences in engagement reflect substance-specific or context-specific processes rather than general behavioral dysregulation. The interaction with age further indicates that this association may be developmentally contingent: early youth who exceeded baseline alcohol use thresholds showed especially low probabilities of high follow-up, while in older youth engagement improved with age among those who exceeded baseline alcohol use thresholds. These findings highlight the analytic challenge of selective lower engagement in substance-focused cohorts and the importance of tailoring retention strategies to developmental stage and risk profile. Although the magnitude of alcohol-related attrition was modest, as minimal follow-up occurred in only a small proportion of participants, these results suggest the value of adaptive approaches that anticipate reduced engagement among higher-risk youth.

Demographic and Family-Level Predictors

Sex, SES, and Hispanic ethnicity were not strongly associated with follow-up engagement in a consistent or substantively meaningful way across models, although no consistent or statistically significant effects were observed in the continuous count model. We interpret these findings cautiously, as the SES variable may reflect both individual-level and site-level variation, potentially obscuring more nuanced effects of financial strain or resource access on retention. Race showed small, non-significant differences in the primary count model; however, these findings should be interpreted cautiously given the use of collapsed categories and the fact that the study was not designed to examine mechanisms underlying racial differences in retention. These findings are broadly consistent with prior work from the NCANDA consortium demonstrating high retention across demographic groups (Nooner et al., 2021), and may reflect the effectiveness of structured retention strategies in mitigating structural barriers to participation (Bamer et al., 2020; Nooner et al., 2021). Sibling enrollment was consistently associated with greater engagement, including both reduced odds of minimal follow-up in categorical models and a modest increase in completed visits in the count model. This pattern supports the hypothesis that shared family participation may reinforce ongoing engagement through mechanisms such as coordinated scheduling, shared transportation, or parental reinforcement. Although sibling co-enrollment is not typical in all cohort designs, it may serve as a proxy for broader household-level engagement processes relevant to longitudinal research. Household-level engagement may be a useful consideration in retention planning. Family history of alcohol problems showed a small, non-significant association with reduced engagement in the count model but was not a consistent predictor across categorical analyses.

Lessons from NCANDA Retention Strategies

NCANDA’s high follow-up rates—reflected by an average of over 9 completed visits out of 10 possible assessments—may reflect its multi-pronged retention framework. This included ongoing participant outreach, flexible scheduling, logistical support, and age-adaptive engagement strategies. Such efforts may have contributed to sustained participation despite developmental transitions and competing demands, as evidenced by the generally small effects of demographic variables on engagement. Treating retention as a continuous outcome further highlights how these strategies may support not only overall retention but also reduce variability in participation across individuals. NCANDA’s approach aligns with best practices from the literature (Cottler et al., 1996; Smith et al., 2017; Twitchell et al., 1992) and may serve as a model for other adolescent cohort studies. However, the intensity of NCANDA’s design—including annual in-person assessments and neuroimaging—may limit generalizability to studies with less frequent or fully remote protocols.

Implications for Longitudinal Research

Our findings support explicitly modeling alcohol use in retention planning. Exceeding baseline alcohol use thresholds was a directionally consistent predictor of reduced engagement across the full distribution of follow-up participation, and its interaction with age suggests that risk profiles evolve across adolescence. Studies may benefit from tailoring retention strategies based on developmental stage and substance use risk (e.g., flexible scheduling, mobile-based assessments, or targeted outreach for higher-risk youth). Additionally, family-based strategies—such as enrolling siblings or maintaining parental contact—may help reduce variability in engagement, particularly during transitions to greater independence. Given that alcohol use is both a variable of interest and a predictor of engagement, future longitudinal analyses should account for potential selection bias due to differential retention. Approaches such as inverse probability weighting, multiple imputation, or sensitivity analyses will be important for strengthening inference about developmental trajectories. Additional pandemic-specific adaptations that helped maintain high retention during COVID-19 disruptions are described in Nooner et al., 2021.

Limitations

The generalizability of these findings may be influenced by NCANDA’s intensive follow-up structure and hybrid in-person design. Although the count-based modeling approach provided a more granular representation of engagement, some predictors—particularly for minimal follow-up—remain difficult to estimate precisely due to small subgroup sizes. Externalizing symptoms were included as a baseline covariate; however, time-varying behavioral factors or other dimensions of psychopathology may further clarify dynamic relationships between risk behavior and engagement. In the continuous model, race was associated with only small effects, and these findings warrant cautious interpretation, as the study was not designed to evaluate mechanisms underlying racial differences in retention. Additionally, the baseline indicator of exceeding alcohol use thresholds does not capture longitudinal changes in substance use, and future studies should examine time-varying or joint models to better characterize reciprocal relationships between substance use and engagement. Because formal withdrawals and deaths were excluded from the primary count outcome, findings should be interpreted as predictors of engagement among participants eligible for continued follow-up, not as a complete model of all forms of attrition. Other limitations, including measurement constraints for SES, incomplete capture of developmental transitions, and potential non-random missingness, are consistent with prior longitudinal cohort research (Luo et al., 2024; Si et al., 2024). Finally, while the present analyses were hypothesis-driven and focused on interpretable associations, future work may benefit from data-driven approaches, including machine learning methods, to identify higher-order interactions and more complex combinations of predictors associated with longitudinal engagement.

Conclusion

This study identified a modest association between exceeding baseline alcohol use thresholds and reduced follow-up engagement in a longitudinal adolescent cohort, with adolescents who exceeded baseline alcohol use thresholds completing fewer follow-up visits over time. Externalizing symptoms were not significantly associated with engagement, suggesting that the association between baseline alcohol use thresholds and follow-up engagement was not explained by baseline externalizing symptoms. Sibling enrollment was associated with greater engagement, while sociodemographic variables showed limited or inconsistent effects. These findings highlight the importance of modeling engagement as a continuous process and suggest that targeted, developmentally informed, and family-inclusive retention strategies may help sustain participation and reduce bias in longitudinal substance use research.

Supplementary Material

Supplementary Material

Public Health Significance Statement.

This study highlights the importance of keeping young people engaged in long-term research on alcohol use and brain development. We found that adolescents who exceeded age- and sex-specific baseline alcohol use thresholds at study enrollment tended to complete slightly fewer follow-up visits over time, while externalizing behaviors were not clearly related to continued participation. In contrast, having a sibling enrolled in the study was associated with higher participation. These findings may help researchers identify groups who could benefit from additional support to stay engaged in long-term developmental studies, which is essential for understanding how alcohol affects brain health and developing effective prevention and intervention programs.

Acknowledgements:

This work was made possible by support from the National Institute on Alcohol Abuse and Alcoholism (NIAAA), including a career development award to Alejandro Meruelo, MD, PhD (K23 AA026869), and grants supporting the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA: AA021697, AA021695, AA021692, AA021696, AA021681, AA021690, and AA021691). Tommy Gunawan and Vijay Ramchandani were supported by the NIAAA Division of Intramural Clinical and Biological Research (Z1A AA000466). We are deeply grateful to the NCANDA participants and their families for their continued dedication to this longitudinal research. We also thank Kilian M. Pohl and Kevin M. Cummins for their helpful feedback on this manuscript.

Footnotes

Statements and Declarations

All authors declare that they have no conflicts of interest.

Data Availability Statement:

Access to NCANDA data is available through the NIAAA Data Archive (collection 4513, https://nda.nih.gov/edit_collection.html?id=4513).

References

  1. Bamer AM, McMullen K, Gibran N, Holavanahalli R, Schneider JC, Carrougher GJ, et al. (2020). Factors Associated with Attrition of Adult Participants in a Longitudinal Database: A National Institute on Disability, Independent Living, and Rehabilitation Research Burn Model System Study. Journal of Burn Care & Research; Official Publication of the American Burn Association, 41(2), 270–279. doi: 10.1093/jbcr/irz186. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Bitsko RH (2022). Mental Health Surveillance Among Children—United States, 2013–2019. MMWR Supplements, 71. doi: 10.15585/mmwr.su7102a1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Booker CL, Harding S, & Benzeval M (2011). A systematic review of the effect of retention methods in population-based cohort studies. BMC Public Health, 11(1), 249. doi: 10.1186/1471-2458-11-249. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Brown SA, Brumback T, Tomlinson K, Cummins K, Thompson WK, Nagel BJ, et al. (2015). The National Consortium on Alcohol and NeuroDevelopment in Adolescence (NCANDA): A Multisite Study of Adolescent Development and Substance Use. Journal of Studies on Alcohol and Drugs, 76(6), 895–908. doi: 10.15288/jsad.2015.76.895. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Brown VA (2021). An Introduction to Linear Mixed-Effects Modeling in R. Advances in Methods and Practices in Psychological Science, 4(1), 2515245920960351. doi: 10.1177/2515245920960351. [DOI] [Google Scholar]
  6. Bureau of Labor Statistics, U.S. Department of Labor. (2023). National Longitudinal Survey of Youth 1979 cohort, 1979–2020 (rounds 1–29) [Longitudinal Survey]. Center for Human Resource Research (CHRR), The Ohio State University. [Google Scholar]
  7. van Buuren S, & Groothuis-Oudshoorn K (2011). mice: Multivariate Imputation by Chained Equations in R. Journal of Statistical Software, 45(3), 1–67. doi: 10.18637/jss.v045.i03. [DOI] [Google Scholar]
  8. Cameron AC, & Miller DL (2015). A Practitioner’s Guide to Cluster-Robust Inference. The Journal of Human Resources, 50(2), 317–372. doi: 10.3368/jhr.50.2.317. [DOI] [Google Scholar]
  9. Costello EJ, Erkanli A, Federman E, & Angold A (1999). Development of psychiatric comorbidity with substance abuse in adolescents: effects of timing and sex. Journal of Clinical Child Psychology, 28(3), 298–311. doi: 10.1207/S15374424jccp280302. [DOI] [PubMed] [Google Scholar]
  10. Cottler LB, Compton WM, Ben-Abdallah A, Horne M, & Claverie D (1996). Achieving a 96.6 percent follow-up rate in a longitudinal study of drug abusers. Drug and Alcohol Dependence, 41(3), 209–217. doi: 10.1016/0376-8716(96)01254-9. [DOI] [PubMed] [Google Scholar]
  11. Cox SML, Castellanos-Ryan N, Parent S, Benkelfat C, Vitaro F, Pihl RO, et al. (2021). Externalizing Risk Pathways for Adolescent Substance Use and Its Developmental Onset: A Canadian Birth Cohort Study: Trajectoires de comportements extériorisés et le risque pour l’initiation et l’usage de substances des adolescents : Une étude de cohorte de naissance canadienne. Canadian Journal of Psychiatry. Revue Canadienne de Psychiatrie, 66(10), 887–896. doi: 10.1177/0706743720982429. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Craney TA, & Surles JG (2002). Model-Dependent Variance Inflation Factor Cutoff Values. Quality Engineering, 14(3), 391–403. doi: 10.1081/QEN-120001878. [DOI] [Google Scholar]
  13. Deke J, Sama-Miller E, & Hershey A (2015). Addressing Attrition Bias in Randomized Controlled Trials: Considerations for Systematic Evidence Reviews (Nos. 2015-72; Office of Planning, Research and Evaluation Administration for Children and Families). U.S. Department of Health and Human Services. [Google Scholar]
  14. Duan L, Chou C-P, Andreeva VA, & Pentz MA (2009). Trajectories of peer social influences as long-term predictors of drug use from early through late adolescence. Journal of Youth and Adolescence, 38(3), 454–465. doi: 10.1007/s10964-008-9310-y. [DOI] [PubMed] [Google Scholar]
  15. Farmer RF, Gau JM, Seeley JR, Kosty DB, Sher KJ, & Lewinsohn PM (2016). Internalizing and externalizing disorders as predictors of alcohol use disorder onset during three developmental periods. Drug and Alcohol Dependence, 164, 38–46. doi: 10.1016/j.drugalcdep.2016.04.021. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Feldstein Ewing SW, Chang L, Cottler LB, Tapert SF, Dowling GJ, & Brown SA (2018). Approaching Retention within the ABCD Study. Developmental Cognitive Neuroscience, 32, 130–137. doi: 10.1016/j.dcn.2017.11.004. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. George S, Duran N, & Norris K (2014). A systematic review of barriers and facilitators to minority research participation among African Americans, Latinos, Asian Americans, and Pacific Islanders. American Journal of Public Health, 104(2), e16–e31. doi: 10.2105/AJPH.2013.301706. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Giatti L, Camelo LV, Rodrigues JF, & Barreto SM (2012). Reliability of the MacArthur scale of subjective social status - Brazilian Longitudinal Study of Adult Health (ELSA-Brasil). BMC Public Health, 12(1), 1096. doi: 10.1186/1471-2458-12-1096. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Hanna KM, Scott LL, & Schmidt KK (2014). Retention strategies in longitudinal studies with emerging adults. Clinical Nurse Specialist CNS, 28(1), 41–45. doi: 10.1097/NUR.0000000000000020. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Harris KM, Halpern CT, Whitsel EA, Hussey JM, Killeya-Jones LA, Tabor J, & Dean SC (2019). Cohort Profile: The National Longitudinal Study of Adolescent to Adult Health (Add Health). International Journal of Epidemiology, 48(5), 1415–1415k. doi: 10.1093/ije/dyz115. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Keyes KM, Jager J, Platt J, Rutherford C, Patrick ME, Kloska DD, & Schulenberg J (2020). When does attrition lead to biased estimates of alcohol consumption? Bias analysis for loss to follow-up in 30 longitudinal cohorts. International Journal of Methods in Psychiatric Research, 29(4), 1–9. doi: 10.1002/mpr.1842. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Kim H-Y (2015). Statistical notes for clinical researchers: post-hoc multiple comparisons. Restorative Dentistry & Endodontics, 40(2), 172–176. doi: 10.5395/rde.2015.40.2.172. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Lenth RV, Banfai B, Bolker B, Buerkner P, Giné-Vázquez I, Herve M, et al. (2025). emmeans: Estimated Marginal Means, aka Least-Squares Means (Version 1.11.1) [Computer software]. https://cloud.r-project.org/web/packages/emmeans/index.html
  24. Levine CS (2017). Psychological buffers against poor health: the role of the socioeconomic environment. Current Opinion in Psychology, 18, 137–140. doi: 10.1016/j.copsyc.2017.08.028. [DOI] [PubMed] [Google Scholar]
  25. Li C (2013). Little’s Test of Missing Completely at Random. The Stata Journal, 13(4), 795–809. doi: 10.1177/1536867X1301300407. [DOI] [Google Scholar]
  26. Luo J, van Grieken A, Kruizinga I, & Raat H (2024). Longitudinal associations between socioeconomic status and psychosocial problems in preschool children. European Child & Adolescent Psychiatry, 33(4), 1029–1038. doi: 10.1007/s00787-023-02217-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. McCabe SE, Cranford JA, Morales M, & Young A (2006). Simultaneous and concurrent polydrug use of alcohol and prescription drugs: prevalence, correlates, and consequences. Journal of Studies on Alcohol, 67(4), 529–537. doi: 10.15288/jsa.2006.67.529. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. National Institute on Alcohol Abuse and Alcoholism. (2011). Alcohol Screening and Brief Intervention for Youth: A Practitioner’s Guide. Department of Health and Human Services, NIH Publication No. 11–7805. https://www.niaaa.nih.gov/alcohols-effects-health/professional-education-materials/alcohol-screening-and-brief-intervention-youth-practitioners-guide [Google Scholar]
  29. Nooner KB, Chung T, Feldstein Ewing SW, Brumback T, Arwood Z, Tapert SF, et al. (2021). Retaining Adolescent and Young Adult Participants in Research During a Pandemic: Best Practices From Two Large-Scale Developmental Neuroimaging Studies (NCANDA and ABCD). Frontiers in Behavioral Neuroscience, 14, 597902. doi: 10.3389/fnbeh.2020.597902. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Odierna DH, & Schmidt LA (2009). The effects of failing to include hard-to-reach respondents in longitudinal surveys. American Journal of Public Health, 99(8), 1515–1521. doi: 10.2105/AJPH.2007.111138. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Peterson JC, Pirraglia PA, Wells MT, & Charlson ME (2012). Attrition in longitudinal randomized controlled trials: home visits make a difference. BMC Medical Research Methodology, 12(1), 178. doi: 10.1186/1471-2288-12-178. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. R Core Team. (2021). R: A language and environment for statistical computing [Computer software]. R Foundation for Statistical Computing. https://www.R-project.org/ [Google Scholar]
  33. Rice JP, Reich T, Bucholz KK, Neuman RJ, Fishman R, Rochberg N, et al. (1995). Comparison of direct interview and family history diagnoses of alcohol dependence. Alcoholism, Clinical and Experimental Research, 19(4), 1018–1023. doi: 10.1111/j.1530-0277.1995.tb00983.x. [DOI] [PubMed] [Google Scholar]
  34. Ripley B, & Venables W (2025). nnet: Feed-Forward Neural Networks and Multinomial Log-Linear Models (Version 7.3–20) [Computer software]. https://cran.r-project.org/web/packages/nnet/index.html
  35. Robinson KA, Dennison CR, Wayman DM, Pronovost PJ, & Needham DM (2007). Systematic review identifies number of strategies important for retaining study participants. Journal of Clinical Epidemiology, 60(8), 757–765. doi: 10.1016/j.jclinepi.2006.11.023. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Rubin DB (1987). Procedures with Nonignorable Nonresponse. In Multiple Imputation for Nonresponse in Surveys (pp. 202–243). John Wiley & Sons, Ltd. doi: 10.1002/9780470316696.ch6 [DOI] [Google Scholar]
  37. Si Y, Little RJA, Mo Y, & Sedransk N (2024). Nonresponse Bias Analysis in Longitudinal Studies: A Comparative Review with an Application to the Early Childhood Longitudinal Study. International Statistical Review = Revue Internationale de Statistique, 92(3), 383–405. doi: 10.1111/insr.12566. [DOI] [Google Scholar]
  38. Smith LJ, McNamara PJ, & King AC (2017). Optimizing follow-up and study retention in the 21st century: Advances from the front line in alcohol and tobacco research. Drug and Alcohol Dependence, 175, 171–178. doi: 10.1016/j.drugalcdep.2017.01.045. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Teague S, Youssef GJ, Macdonald JA, Sciberras E, Shatte A, Fuller-Tyszkiewicz M, et al. , & the SEED Lifecourse Sciences Theme. (2018). Retention strategies in longitudinal cohort studies: a systematic review and meta-analysis. BMC Medical Research Methodology, 18(1), 151. doi: 10.1186/s12874-018-0586-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Thornton M, Williams J, McCrory C, Murray A, & Quail A (2016). Growing Up in Ireland National Longitudinal Study of Children: Design, Instrumentation and Procedures for the Child Cohort at Wave Two (13 Years). [Google Scholar]
  41. Twitchell GR, Hertzog CA, Klein JL, & Schuckit MA (1992). The anatomy of a follow-up. British Journal of Addiction, 87(9), 1327–1333. doi: 10.1111/j.1360-0443.1992.tb02741.x. [DOI] [PubMed] [Google Scholar]
  42. Wallace E, & Gibb J (2013). Millennium Cohort Study Participant Engagement Research. [Google Scholar]
  43. Weuve J, Tchetgen Tchetgen EJ, Glymour MM, Beck TL, Aggarwal NT, Wilson RS, et al. (2012). Accounting for bias due to selective attrition: the example of smoking and cognitive decline. Epidemiology (Cambridge, Mass.), 23(1), 119–128. doi: 10.1097/EDE.0b013e318230e861. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Williams R (2016). Understanding and interpreting generalized ordered logit models. The Journal of Mathematical Sociology, 40(1), 7–20. doi: 10.1080/0022250X.2015.1112384. [DOI] [Google Scholar]
  45. Yancey AK, Ortega AN, & Kumanyika SK (2006). Effective recruitment and retention of minority research participants. Annual Review of Public Health, 27(1), 1–28. doi: 10.1146/annurev.publhealth.27.021405.102113. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Material
NIHMS2187523-supplement-Supplementary_Material.docx (275.1KB, docx)

Data Availability Statement

Access to NCANDA data is available through the NIAAA Data Archive (collection 4513, https://nda.nih.gov/edit_collection.html?id=4513).

RESOURCES