We examined mortality rates of adolescents and young adults before and after exiting foster care. We used administrative records to identify individuals born in California 1985 to 2005 who had a foster care episode on or after their 16th birthday. We linked these records to vital death records through 2022 to identify deaths occurring ages 16-23 years. We defined three care statuses: pretransition, transition, and posttransition. Pretransition was age 16 years to the last day of care. Transition was the first 120 days after exiting care. Posttransition was all days after transition. We calculated gender-standardized mortality rates (SMRs) and used a Cox proportional hazards model adjusted for demographics to estimate hazard ratios of total mortality. In total, 1743 deaths occurred among 144 128 individuals. Standardized mortality rates for pretransition, transition, and posttransition per 100 000 person-years were 116, 259, and 209, respectively. Time-varying hazards models detected that these high and disparate rates were driven by higher risk during transition and posttransition for those leaving care before age 20 years. Moreover, the transition period featured particularly heightened risk for those leaving care before age 18 years. Results suggest targeted support during the transition period could help safeguard this population from harm.

BACKGROUND:Post-traumatic stress disorder (PTSD) symptoms can fluctuate substantially over short periods, yet routine screening typically relies on infrequent self-report. Wearable sensors provide continuous behavioural and physiological signals that may help identify periods of elevated risk. AIMS:This study aimed to evaluate whether combining wearable sensor features with daily self-report data could identify short-term PTSD symptom increases among recently discharged veterans. METHOD:Seventy-four veterans wore commercial activity trackers and completed brief daily questionnaires over 87 days. For each participant, we defined an individual baseline by using the first 14 days of PTSD scores. Wearable variables were transformed into baseline-referenced deviation features to capture departures from personal norms. Missing data were addressed with multiple imputation by chained equations. Candidate predictors were prioritised with least absolute shrinkage and selection operator regression, and a set of machine-learning classifiers was evaluated. Primary performance was assessed by using the area under the precision-recall curve (PR AUC). RESULTS:Across feature set sizes (k = 1-25), performance peaked at k = 17. At this iteration, LightGBM achieved the strongest discrimination (PR AUC 0.86 (s.d. 0.07); area under the receiver-operating characteristic curve 0.89 (s.d. 0.04)) with a precision of 0.67 (s.d. 0.08), recall of 0.64 (s.d. 0.08) and F1 of 0.65 (s.d. 0.07). Key predictors reflected a multimodal profile, combining self-reported affect and perceived stress with wearable indicators of sleep continuity, activity variability and autonomic regulation. CONCLUSIONS:Baseline-referenced wearable features combined with daily self-report may help identify near-term PTSD symptom increases among recently discharged veterans with elevated PTSD symptoms and problematic cannabis use. Future work should validate performance in broader PTSD populations, including samples without problematic cannabis use.

INTRODUCTION:A predictive risk model (PRM) was trained to stratify risk among children investigated for alleged maltreatment based on the likelihood of future child protection involvement. In the current brief, we assess the model's ability to differentiate risk of adverse events not used to build the model (i.e., arrest, death) among adolescent populations investigated following reported maltreatment to guide prevention-oriented services. METHODS:Child welfare and vital statistics records were obtained through a data use agreement. Among adolescents born in 2000 and 2001 and investigated for alleged maltreatment between ages 11 and 17 (n = 72,340), risk scores were calculated using a random forest algorithm based on information available at the time of maltreatment report. The records of these adolescents were then linked to arrest and death records. RESULTS:Among adolescents investigated for maltreatment, 5.8% experienced a juvenile arrest or death before age 21. Of those who experienced an arrest or death, 43.9% fell in the highest risk decile. CONCLUSIONS:A PRM trained to predict foster care placement had strong external validity in predicting both future arrests and deaths. The average time from investigation to adverse event indicates a meaningful window for interventions to be delivered focused on supporting and stabilizing adolescents and their families.

Large Language Models (LLMs) are increasingly involved in high-stakes domains, yet how they reason about socially-sensitive decisions still remain underexplored. We present a large-scale audit of LLMs’ treatment of socioeconomic status (SES) in college admissions decisions using a novel dual-process framework inspired by cognitive science. Leveraging a synthetic dataset of 30,000 applicant profiles grounded in real-world correlations, we prompt 4 open-source LLMs (Qwen 2, Mistral v0.3, Gemma 2, Llama 3.1) under 2 modes: a fast, decision-only setup (System 1) and a slower, explanation-based setup (System 2). Results from 5 million prompts reveals that LLMs consistently favor low-SES applicants—even when controlling for academic performance—and that System 2 amplifies this tendency by explicitly invoking SES as compensatory justification, highlighting both their potential and volatility as decision-makers. We then propose DPAF, a dual-process audit framework to probe LLMs’ reasoning behaviors in sensitive applications.

An increasing number of families in the United States have opted for homeschooling as an alternative to the formal schooling system. Estimates suggest that in the United States, between 5 % and 11 % of the school-age population participates in homeschooling each year-more than 1.5 million students across the country. Recent high-profile cases of severe child abuse and torture among homeschooled children, however, have highlighted the lack of mechanisms for the detection and prevention of child maltreatment. Though these cases have raised awareness about the potential risks of homeschooling, calls for tighter regulations are often met with resistance. Homeschooling advocates argue that increased regulation in the private homes of families criminalizes parents without sufficient evidence of heightened risk. Meanwhile, critics posit that policy intervention is needed to protect the subset of students at increased risk of maltreatment given the deinstitutionalized and isolated nature of homeschooling settings. In this discussion article, we examine the evidence for heightened risk of abuse and neglect in homeschooling environments through both a review of research and an examination of the policies surrounding homeschooling and child maltreatment across U.S. states. Because of the lack of reliable data and the topic's highly political nature, empirical evidence is limited. We found that although some states have attempted to add explicit child protections to homeschooling law, most efforts have been unsuccessful.

This brief report examines academic achievement disparities among California public school students with varying levels of child protection service (CPS) involvement, used as a proxy for maltreatment risk. Using linked birth records, public school performance data, and CPS records, we analyzed fifth grade English Language Arts (ELA) and mathematics achievement for children born 2004–2006 (n = 672,706). We compared outcomes across levels of CPS involvement (no reported maltreatment, maltreatment investigation, open case for substantiated maltreatment, foster care placement) and made descriptive comparisons to the general population and children born with public insurance. Overall, 48.7

BACKGROUND:Approximately 5 % of children and adolescents in foster care are placed in group home settings, with adolescents making up the largest age group in these placements. Group home placement is designed to support individuals with high acuity or specialized needs that cannot be met in a family-based setting. Prior research has highlighted the vulnerabilities of adolescents in group homes, including behavioral issues, mental health concerns, and juvenile justice system involvement. OBJECTIVE:The current study investigated the relationship between placement status and arrest rates among adolescents experiencing a group home placement. PARTICIPANTS:Child protection system records from California were used to identify adolescents (aged 13-17) who were placed in group homes in 2014 (n = 2437). These records were linked to California Department of Justice arrest records for 2014 and 2015. METHODS:Descriptive differences in arrest status were examined by demographics and placement status after initial group home placements with t-tests and logistic regression. The association between placement setting and likelihood of arrest was examined with survival analysis. RESULTS:In our population of adolescents who experienced a group home placement, 16.7 % were arrested during the study window. Adjusted hazard ratios for arrest were higher when adolescents lived in group homes (aHR = 1.57, p = .029) or ran away from care (aHR = 5.62, p < .001), compared to periods in which adolescents had transitioned to reunification or guardianship. During periods where adolescents were in family foster care settings, arrest rates were comparable to those who had exited to reunification or guardianship. CONCLUSION:Periods when an adolescent was in a group home setting were associated with a heightened rate of arrest compared to those reunified with their families. We also found that arrests commonly occurred and reoccurred for young people residing in group homes, underscoring the need to understand whether specific protocols and practices in these facilities contribute to increased arrests.

OBJECTIVE:Cannabis use is common among U.S. military veterans, particularly those experiencing posttraumatic stress disorder (PTSD), poor sleep, and elevated stress. While often used to self-manage these symptoms, the impact of cannabis on day-to-day symptomology remains unclear. This study examined the daily associations among cannabis use, PTSD symptoms, perceived stress, and sleep quality using intensive longitudinal data. METHOD:Seventy-four recently separated U.S. veterans (age = 33.5 years; 80% male; 61% non-Hispanic White) who endorsed past-month cannabis use and elevated PTSD symptoms completed a 3-month daily diary study. Participants provided 4,307 person-days of data via a mobile app. Measures included daily cannabis use (hours high), PTSD symptoms, perceived stress, and sleep quality. Dynamic structural equation modeling (DSEM) was used to estimate within-person lagged and same-day associations, adjusting for relevant covariates. RESULTS:Day-to-day analyses revealed that elevated PTSD symptoms and poor sleep quality each predicted greater perceived stress the following day. Greater number of hours high was associated with less perceived stress the following day. Perceived stress, in turn, predicted both higher PTSD symptoms and poorer sleep quality. In post hoc analysis, stress emerged as a significant mechanism of change in the day-to-day lagged model. In particular, we show greater cannabis use is linked to improved sleep and PTSD symptoms through lower perceived stress. CONCLUSION:Cannabis may offer temporary relief and appears to interrupt the day-to-day cycle linking PTSD, stress, and poor sleep. (PsycInfo Database Record (c) 2026 APA, all rights reserved).

Background and aims. Cannabis use is increasingly prevalent among U.S. veterans, with high rates of both recreational and problematic use. Veterans often use cannabis to manage symptoms associated with mental health problems such as depression and posttraumatic stress disorder (PTSD). Prior work has noted mixed results on the longitudinal associations between cannabis use and depression. Studying these associations at the daily level can lead to improved clarity. Design. The present study examined the daily associations between cannabis use and depression in veterans using dynamic structural equation modeling (DSEM). We also explored these associations for those veterans who screened positive for posttraumatic stress disorder (PTSD) compared to those who did not. Setting. All participants were recruited using advertisements from BuildClinical, an NIH approved recruitment vendor. Participants. The sample consisted of 74 veterans who provided daily data for 87 consecutive days. Measurement. Cannabis was assessed asking how many hours each individual spent high each day, depressed mood was assessed using a sliding scale from not depressed to very depressed each day, and PTSD was assessed using the PTSD checklist. Findings. Among the full veteran sample, results revealed a bidirectional negative association. Specifically, on days when veterans reported greater depression, they reported fewer hours “high” the next day. Conversely, on days when veterans reported a greater number of hours high, they reported less depression the next day. Among veterans screening positive for PTSD, on days when they reported more depression, they reported fewer hours high the next day (no association was noted for cannabis use predicting depression). However, for those who did not screen positive for PTSD, on days when veterans reported greater number of hours high, they reported less depression the next day. Conclusions. These results highlight the need for further research on the effect of individual differences in cannabis use patterns among veterans with PTSD on health outcomes. Clinically, these results highlight the importance of targeting the pros and cons of cannabis use for depression symptom relief. Future research should incorporate daily objective measures of cannabis use to refine treatment strategies for veterans managing PTSD or depressive related distress.

The application of machine learning algorithms to daily diary data represents a valuable tool for improving dynamic prediction of posttraumatic stress disorder (PTSD) symptom escalations. This prospective, intensive longitudinal study aimed to evaluate whether combining baseline (static) and daily diary (dynamic) predictors with machine learning can help forecast clinically significant PTSD symptom increases among veterans. Participants were 74 recently discharged U.S. veterans (Mage = 33.5 years) who completed twice-daily diary surveys for up to 87 days via a mobile app, yielding 4,307 diary days. The outcome was a binary indicator of clinically significant daily PTSD symptom increase (> 1.0 standard deviation above a participant's individual mean over the first 2 study weeks). Random forest models identified top predictors; LASSO regression estimated effect sizes among top predictors. Daily negative affect was the top predictive variable, OR = 1.33, retained in 100% of LASSO iterations. Daily depressed mood, OR = 1.35; anxious mood, OR = 1.15; and perceived stress, OR = 1.13, were also reliably retained. Variables involving alcohol, cannabis use, and baseline impulsivity were less robust but remained prominent predictors of PTSD symptom escalations. Post hoc interaction analyses showed that co-occurring high negative affect and anxiety yielded a > 55% probability of PTSD symptom escalation. The findings show that daily affective states, especially negative mood and stress, strongly predict PTSD symptom increases in veterans. Using machine learning and high-frequency tracking, advances in personalized, real-time PTSD care are possible. Findings support just-in-time interventions for when veterans need help most: in the moment.

BACKGROUND:Child protective services (CPS) rely on referrals of child maltreatment to address abuse and neglect. CPS referral rates vary geographically in the type of maltreatment reported and community characteristics. OBJECTIVE:This study examined the relationship between CPS referral rates and community characteristics, based on reporter type and geography. We investigated (a) the relationship between CPS referral rates and community-level characteristics, including how these relationships vary by reporter type; and (b) the relationships between community-level characteristics and CPS referral rates across geographic areas. PARTICIPANTS AND SETTING:The study used statewide data from California on CPS referrals for children aged 5 years or older from 2018 to 2022. METHODS:Referrals were geocoded to census tracts based on the family's address. Regression models were run for each reporter type, using referral rate as the outcome and community-level social determinants of health indicators from the Healthy Places Index (economic, education, health insurance, clean environment, housing, neighborhood conditions, social, transportation, racial composition, and urbanicity). Geographically weighted regression was used to examine spatial heterogeneity. RESULTS:Referral rates had different relationships with community-level characteristics by reporter type. Spatial heterogeneity was identified by community characteristics and reporter type. Variations occurred in magnitude, direction, and statistical significance. CONCLUSIONS:CPS referral rates varied by community characteristics depending on the referral source and geographic area. These findings suggest that prevention and intervention programs may not be universally effective across communities and regions, highlighting the importance of considering spatial variability when developing and providing such services.

United States military veterans face heightened vulnerability to sleep disturbances due to factors such as irregular sleep schedules, combat-related stress, and co-occurring mental health disorders. These sleep disturbances are often exacerbated by substance use, including alcohol and cannabis, as veterans may rely on these substances to self-medicate for stress and sleep issues. However, the interplay between sleep quality, substance use, and perceived stress remains poorly understood, particularly on a day-to-day basis. This study aimed to explore the dynamic associations between these factors using daily diary data collected over three months from 74 veterans with elevated PTSD symptoms and problematic cannabis use. Data from this study are secondary analysis. Dynamic structural equation modeling (DSEM) was employed to examine both within-day and day-to-day lagged associations between sleep quality, perceived stress, and substance use (alcohol and cannabis). Results showed that worse sleep quality was associated with higher perceived stress the next day, which in turn predicted greater alcohol consumption. Additionally, stress mediated the relationship between poor sleep quality and increased alcohol use. For cannabis, while no day-to-day lagged effects were observed, within-day analyses revealed that higher cannabis use was associated with lower stress and better sleep quality that same night. These results highlight the complex and bidirectional relationships between sleep, stress, and substance use among veterans, underscoring the need for interventions that address these dynamics holistically. Future research should further explore these interactions using real-time data to inform tailored interventions for improving sleep and mental health outcomes in this population.

Objective: To examine prospective, bidirectional associations between homelessness and substance use frequency among young adults receiving substance use treatment in the United States. We also investigated potential differences across demographic subgroups. Methods: Young adults (N = 3717, Mage = 20.1, 28% female, 7.3% sexual/gender minority, and 37% non-Hispanic White) receiving substance use treatment in the U.S. completed assessments at intake, 3 months, 6 months, and 12 months post-intake. Latent growth curve models with structured residuals (LGC-SR) were used to examine cross-lagged associations between homeless days and frequency of substance use and associated problems. Models were stratified by sex, race/ethnicity, and sexual and/or gender minority status. Results: Overall, days spent homeless (mu(slope )= -0.19, p = 0.046) and substance use frequency (mu(slope1 )= -6.19, p < 0.001) significantly decreased during treatment, with no significant cross-lagged associations between homeless days and substance use frequency. However, results differed by race and ethnicity. For non-Hispanic White young adults, greater substance use at treatment entry was associated with steeper declines in homeless days between-persons (phi(standardized) = -0.14, p = 0.04). For African Americans, homeless days at treatment entry were associated with greater increases in substance use between-persons (phi(standardized) = 0.29, p = 0.04). No significant differences were found by sex or sexual/gender minority status. Conclusions: Despite overall declines in homelessness and substance use during treatment, these outcomes may unfold differently for non-Hispanic White and African American young adults. More support may be needed for African American young adults reporting homelessness at treatment entry.

To promote the early identification and support of children at risk for developmental issues and other adversities, seven of Orange County, California’s largest birthing hospitals have adopted a “hospital screening” process comprised of an initial clinical data scan, followed by the completion of a more robust bedside assessment. This approach to hospital screening and referral, termed the “Bridges Program,” is used to connect parents of high-risk newborns to home visiting services aimed at providing support and resources. Orange County also has four birthing hospitals that do not participate in the Bridges Program, which allowed for the comparison of births occurring between 2011 and 2012 at Bridges hospitals (n = 53,302) to births at non-Bridges hospitals (n = 25,146) on various child protection outcomes by age five, accounting for program engagement and demographic differences. Bridges assessments successfully identified births at higher risk for child protection service involvement, and high-risk Bridges births not forwarded for agency outreach were significantly more likely to be reported and substantiated for maltreatment compared to low risk Bridges births (RR = 1.18, CI = [1.08, 1.30]; RR = 1.30, CI = [1.10, 1.54]). These results highlight a subset of births that would benefit from alternative strategies for engaging with home visiting services.

The purpose of this study is to estimate the rate of emotional disturbance (ED) among children in foster care and assess the validity of the national foster care census data (AFCARS) measure of ED. This study used linked child protection and Medicaid records from 2014 and 2015, for the states of California and Wisconsin, as well as data from AFCARS, a federal population census of children in foster care which states are mandated to contribute to. ED is defined by AFCARS and includes an array of mental and behavioral health diagnoses. According to AFCARS, 13% of CA children in foster care and 15% of WI children in foster care had an ED, whereas Medicaid claims produce rates of 45% and 48%, respectively. Rates of ED among children in congregate care were underestimated by 43–46 percentage points, with substantial proportions having diagnoses of disruptive behavioral disorders. Despite the AFCARS ED measure being cited in congressional testimonies and its wide use in research, results from this study suggest that the AFCARS ED estimates are an unreliable metric for use in research, policy, or practice.

Linking administrative records across programs can yield person-centered information, including client characteristics, public service trajectories, and outcomes and help to answer policy-related questions. Several solutions are available for undertaking record linkage, producing linkage keys for merging data sources for positively matched pairs of records. In this session, we will demonstrate a new application of the Python RecordLinkage package to family-based record linkages with machine learning algorithms for probability scoring, which we call probabilistic record linkage for families (PRLF). First, we will demonstrate the utility of PRLF with a simulation of administrative records and assess linkage accuracy with variations in match rates and data degradation. Second, we will compare generalized linear model estimates across three record linkage solutions (PRLF, ChoiceMaker, and Link Plus). Findings from the simulation study indicate linkage accuracy is largely influenced by degradation (e.g., missing data fields, erroneous or incomplete values) compared to the proportion of simulated matches between datasets. Results from the methods comparison using real world data indicate that all three solutions, when optimized, provide similar results for researchers. We discuss the strengths of our process, such as the use of ensemble methods, to improve match accuracy. We then will identify caveats of record linkage in the context of administrative data. The tool was developed in Python to allow for researchers to work with open-source software and adjust the basic workflow to fit their linkage needs. We will identify several partnerships where this collaboration has worked successfully and empower attendees with access to this useful tool.

Download This Paper Open PDF in Browser Add Paper to My Library Share: Permalink Using these links will ensure access to this page indefinitely Copy URL Copy DOI

Social science research has shown that candidates with names indicative of certain races or genders often face discrimination in employment practices. Similarly, Large Language Models (LLMs) have demonstrated racial and gender biases in various applications. In this study, we utilize GPT-3.5-Turbo and Llama 3-70B-Instruct to simulate hiring decisions and salary recommendations for candidates with 320 first names that strongly signal their race and gender, across over 750,000 prompts. Our empirical results indicate a preference among these models for hiring candidates with White female-sounding names over other demographic groups across 40 occupations. Additionally, even among candidates with identical qualifications, salary recommendations vary by as much as 5 between different subgroups. A comparison with real-world labor data reveals inconsistent alignment with U.S. labor market characteristics, underscoring the necessity of risk investigation of LLM-powered systems.