In rural China, millions of children grow up while one or both parents work in distant cities. These "left-behind children" show markedly higher rates of depression, anxiety and self-injury than their peers. Schools are often the only place anyone could screen them, because trained psychologists are scarce, at roughly one per 28,000 people in rural areas. A screening model built for this setting faces a basic design question. Should it use one set of rules for every child, or separate rules for children whose parents are absent in different ways? Epidemiology suggests the risk factors really do differ across these groups. Machine learning says that splitting data into small groups invites overfitting. And the people who would use the tool care about something neither literature measures: whether the limited number of children a school counselor can actually see each term are the right ones.
We propose a procedure that makes the split-or-pool decision per subgroup, and makes it testably, on held-out data. There are three levels of specialization to choose from: one shared model, a shared model with group-specific calibration, or a separate model for that group. The choice for each group is judged by a single action-level quantity. That quantity is the share of that group's truly distressed children who land inside a fixed, school-wide referral budget. The deliverable is a screening tool, and also a falsifiable map of which migration typologies need their own model and which do not.
Core Claim
Caregiving arrangements produce different risk structures. In a five-region rural sample, Chang et al. found that remarried family structure (OR 3.55) and parental contact less than once every six months (OR 2.39) predicted suicide attempts only among left-behind children. Among non-left-behind children, by contrast, higher maternal education predicted higher risk. Separately, Negriff et al. showed that "parental closeness" was protective for community youth but a risk factor for child-welfare youth. That is the sign flip a shared linear model averages away.
These findings motivate subgroup modeling but do not settle it. The finer typologies that matter in practice have few positive cases. Examples include father-only absent, both absent with grandparent care, and remarried surrogate care. Many of the apparent differences may also be ordinary shifts in which kinds of children are in each group, rather than different feature-to-outcome rules.
Our claim is that the right level of specialization differs by subgroup and can be predicted in advance. Groups whose performance gap disappears once their feature distribution is reweighted should need at most recalibration. Groups whose gap persists should need their own model, but only when they have enough cases to support one. The test for this uses controlled evaluation from Understanding challenges to the interpretation of disaggregated evaluations, which we reuse as a method. That paper found racial gaps fully explained by covariate shift on ACSIncome but persisting on ACSPublicCoverage. We expect the same split to appear across migration typologies. If it does, the diagnostic can tell a deploying school system which subgroups to decouple before any model is fit.
Research Program
Data and outcome. The development data are the two large Chinese cohorts in the file. The Mianyang adolescent mental health cohort (Feng et al.) has 28,036 students, 5,114 of them left-behind. It includes bullying victimization, family function and peer relations, with self-injury and PHQ-9/GAD-7 symptoms. The five-region rural cohort (Chang et al.) has 13,952 adolescents, 6,034 of them left-behind. It includes neglect and abuse, negative life events, loneliness and 1-year suicide attempts. Both require data-access agreements with the original investigators; securing these is the first milestone.
Both are cross-sectional. We therefore frame the task honestly as case-finding: identifying children with current self-injury, or a recent suicide attempt, from information a school could collect without asking about the outcome directly. We do not frame it as forecasting. The instruments in the two cohorts differ, so we replicate the whole procedure in each cohort rather than pooling them.
Subgroups. These are built from parental-absence type (none, father only, mother only, both) crossed with caregiver arrangement where it is recorded (grandparent, other relative, remarried household). Cells too small to evaluate are merged by a pre-registered rule.
Utility. Let $B$ be the number of referrals a school can act on in a term, for example the top 10% of students. Let $R_B(s)$ be the $B$ students ranked highest by score function $s$ across the whole school. For subgroup $g$ with true-case set $C_g$, define the group's captured fraction
$$U_g(s) = \frac{|R_B(s) \cap C_g|}{|C_g|}.$$
Because the budget is shared across the school, scores from different subgroup models must be on a common probability scale. Calibration is therefore not a cosmetic step: a miscalibrated separate model can crowd out or starve its own group. This shared budget is what separates our setting from Fairness without Harm. That paper decouples classifiers per group under accuracy, using exact 0-1 loss integer programming (the TREE01 algorithm). Here, each group's benefit depends on every other group's scores.
Tier selection. For each subgroup we compare three tiers on held-out folds:
- the shared model;
- the shared model plus group-specific calibration and threshold;
- a separate model, optionally warm-started from the shared one.
We adopt TREE01's preference notions, but measure them with $U_g$ instead of accuracy. Rationality means a group's assigned tier captures at least as many of its cases as the shared model. Envy-freeness means it would not capture more using another group's model.
Moving a group to a more specialized tier requires a paired McNemar test on its true cases, each one captured or missed, to reject at a pre-set level. This is the same hold-out test TREE01 uses, applied to a different quantity. We then run controlled evaluation for each group and check whether the tier it predicts matches the tier the selection procedure picks.
Output for schools. For each referred child, the tool lists the tier-specific factors that drove the referral, mapped to actions a school can take. Examples are bullying victimization (an anti-bullying response) and infrequent contact with a migrant parent (arranging regular calls). This answers the clinician demand reported by Alexander et al. for "risk stratification leading to different actions," which 74.5% ranked as a top-two criterion. We deliberately do not build a language-model narrative layer like Kandala et al., which translates feature attributions into DSM-grounded text for psychiatrists. That would confound the modeling question with a presentation question.
Baselines
Naive references. The simplest reference refers students using a standard symptom cutoff, PHQ-9 ≥ 10 or GAD-7 ≥ 10 in Mianyang, filling the same budget $B$ in descending score order. A second rule refers left-behind children first, ordered by symptom score. This reflects current practice of targeting by status alone. Any model that cannot beat these on $U_g$ for left-behind subgroups has no reason to exist.
Strongest competitor. This is a single pooled gradient-boosted tree model that is given the typology indicators as features, followed by subgroup-specific thresholds. The threshold step is the policy layer from the sepsis threshold study, which fixed one model and tuned operating points across 16 diagnostic categories. That study reported diagnostic odds ratio gains of up to 68%. Its savings were $4.6 billion overall, $700 million of it over a uniform 80% sensitivity threshold.
This competitor is strong because trees can split on the typology indicator and learn sign-flipped effects without explicit decoupling. Group thresholds then absorb calibration differences. It will get the same features, folds and hyperparameter search budget as every tier of our method; we use the same model family for the separate-model tier so that model capacity is matched.
There is no reported machine-learning result on Lacuna for either Chinese cohort. The nearest reported figures come from sex-stratified random forests in the French i-Share study, with AUC 0.84 (girls) and 0.82 (boys) and PPV 40% and 36%, and from Negriff's subgroup models, with AUC 0.80 and 0.85. These are context, not targets, since the populations differ.
What counts as winning. Beating this baseline means a statistically significant improvement in $U_g$ for at least one left-behind subgroup, with no subgroup significantly worse, at matched total budget.
Ablations. We include four, each isolating one ingredient:
- all groups forced to the separate-model tier, which is what TREE01-style full decoupling would do here;
- tiers selected by 0-1 accuracy instead of $U_g$;
- tiers selected without the significance test;
- tiers assigned by the controlled-evaluation diagnostic alone, with no outcome-based selection.
Why not just…?
Why not just use one pooled model with a left-behind indicator? For linear models, Negriff's sign flip and Chang's divergent odds ratios argue against it. For boosted trees the argument is much weaker, and this is the real threat. If the pooled-trees-plus-thresholds baseline matches our tiered procedure on every subgroup, we will report that as the answer: decoupling is unnecessary at these sample sizes. In that case the remaining contribution is the diagnostic showing why.
Why not just recalibrate thresholds per group, as in the sepsis study? That study kept the model fixed, so it can fix covariate shift but not a reversed feature effect. Recalibration is one of our three tiers rather than a rival. The experiment measures how many subgroups need anything more, and we would not be surprised if the answer is few.
Why not just give teachers the PHQ-9 and GAD-7? We include this as the naive baseline. Two strands of evidence suggest it underperforms for this group. PsyDraw reports low questionnaire engagement among young rural children, and Chen reports that grandparent caregivers miss internalizing symptoms. These suggest that self-report alone underidentifies the left-behind children who suppress distress. Whether contextual features add capture within a fixed budget is an empirical question that $U_g$ answers directly.
Why not just prompt a multimodal language model, as PsyDraw does? PsyDraw's single pipeline reached 71.03% overall agreement with school psychology teachers. Agreement fell to 58.89% on "Warning" cases, exactly the children that matter. It also applies one pipeline to all children. Our procedure could treat its output as one more input feature. The 290-drawing sample is too small to stratify by typology, so we leave that combination to future data.
Proposed Experiments
Stage one (retrospective, both cohorts). Using stratified cross-validation with school-level grouping, we compute $U_g$ for every tier and baseline at budgets of 5%, 10% and 20%. We also report subgroup calibration error and the count of rationality and envy-freeness violations on held-out data.
There are three decisive checks:
- Gains. Does the tiered procedure beat pooled-trees-plus-thresholds for any left-behind subgroup without harming others?
- Diagnostic. Does controlled evaluation predict the selected tier more often than chance? This is the claim that would make the method transferable.
- Sample-size boundary. In the five-region cohort the suicide-attempt rate is 3.75% among left-behind children, so typology cells hold few cases. The procedure should choose pooling there, while it chooses decoupling more often for the more common self-injury outcome in Mianyang. A procedure that decouples small groups regardless of case count is mis-specified.
Stage two (prospective silent trial, conditional on stage one). The prospective test follows the DEPLOYR design. That paper ran models silently in live use and saw AUROC fall from 0.76 to 0.69 on one model. In partner schools, the frozen stage-one tool scores students at the start of a term, and nobody acts on the scores. Children are assessed at term end by school psychology teachers, and $U_g$ is recomputed.
The decisive question is whether tier assignments that won retrospectively keep their advantage, or whether retrospective-to-prospective decay is larger for separate-model tiers. A larger decay for specialized tiers would be the signature of small-sample overfitting.
References
- A comparative analysis of suicide attempts in left-behind children and non-left-behind children in rural China
- Using machine learning to determine the shared and unique risk factors for marijuana use among child-welfare versus community adolescents
- Understanding challenges to the interpretation of disaggregated evaluations of algorithmic fairness
- Association between bullying victimization and mental health problems among Chinese left-behind children: a cross-sectional study from the adolescence mental health promotion cohort
- Fairness without Harm: Decoupled Classifiers with Preference Guarantees
- Perspective Toward Machine Learning Implementation in Pediatric Medicine: Mixed Methods Study
- From Explainability to Action: A Generative Operational Framework for Integrating XAI in Clinical Mental Health Screening
- Optimizing the Implementation of Clinical Predictive Models to Minimize National Costs: Sepsis Case Study
- A machine learning approach for predicting suicidal thoughts and behaviours among college students
- PsyDraw: A Multi-Agent Multimodal System for Mental Health Screening in Left-Behind Children
- Analysis of the Influences of Family Education on Children's Mental Health
- DEPLOYR: a technical framework for deploying custom real-time machine learning models into the electronic medical record