Differences in associations between long-acting injectable and oral antipsychotic prescribing and healthcare utilisation in people with intellectual disability: a self-controlled case series (LOID study) Authors: Shek Ming Leung, Kyung Jin Lee, Yue Wei, Vincent K C Yan, Eunice Kehui Deng, Shoumitro Deb, David Castle, Caige Huang, Hei Hang Edmund Yiu, Simon Sai-Yu Lui, Vanessa W S Ng, Francisco Tsz Tsun Lai, Wai Ching Yan, Esther W Chan Published: 2026 ABSTRACT AbstractBackgroundLiterature guiding the use of long-acting injectable antipsychotics (LAIAs) versus oral antipsychotics (OAs) in people with intellectual disability (PWID) remains scarce.ObjectiveExamining the associations of LAIAs vs OAs with healthcare utilisation among PWID.MethodsThis self-controlled case series study used electronic health records from Hong Kong. We included adults with intellectual disability prescribed both LAIAs and OAs at different periods in 2004–2023. We compared within-individual LAIA and OA periods, adjusting for age, calendar year, season, inpatient stay and concomitant psychotropic use, using conditional Poisson regression to estimate incidence rate ratios (IRRs) for all-cause emergency department (ED) visits; all-cause hospitalisations; hospitalisations for psychiatric disorders, schizophrenia/bipolar disorder, behaviour-related problems, somatic disorders and cardiovascular diseases (CVDs) and diagnosis records of extrapyramidal symptoms (EPS); falls and fractures. Subgroup analyses were conducted by diagnosis of schizophrenia or bipolar disorder.FindingsWe identified 1675 PWID who had been prescribed both OAs and LAIAs at different periods during the study period. LAIA use was associated with lower risks of incident all-cause ED visits (adjusted IRR: 0.71, 95% CI 0.51 to 0.98), all-cause hospitalisations (0.91, 0.85–0.98); hospitalisations for psychiatric disorders (0.90, 0.81–0.99); for schizophrenia/bipolar disorder (0.86, 0.75–0.99) and EPS (0.34, 0.23–0.51), compared with OAs. In PWID with schizophrenia/bipolar disorder, LAIAs were associated with lower risks of all-cause hospitalisations, hospitalisations for psychiatric disorders, schizophrenia/bipolar disorder, behaviour-related problems and somatic disorders, and EPS, but a higher risk of hospitalisations for CVDs (with hypertensive disease being the most common), compared with OAs. In contrast, among PWID without such diagnoses, LAIA use was associated with higher risks of all-cause hospitalisations, hospitalisations for somatic disorders; a lower risk of incident all-cause ED visits and hospitalisations for CVDs, but no significant difference in hospitalisations for behaviour-related problems, compared with OAs.ConclusionsThese findings are consistent with potential benefit of LAIAs among PWID with schizophrenia or bipolar disorder, while raising concern about off-label use in those without these diagnoses.Clinical implicationsLAIAs may be considered in PWID with schizophrenia or bipolar disorder; off-label use in those without these diagnoses should be approached cautiously, and deprescribing may be considered where appropriate if these results are replicated.Keywords: Intellectual Disability, Schizophrenia, Bipolar and Related Disorders, Behavioral Medicine, Psychopharmacology WHAT IS ALREADY KNOWN ON THIS TOPIC Evidence on the effectiveness and safety of long-acting injectable antipsychotics (LAIAs) in people with intellectual disability (PWID) is very limited. Existing studies are small, short-term and mostly do not compare LAIAs with oral antipsychotics. It remains unclear whether LAIAs are beneficial or harmful in PWID, particularly according to the presence of schizophrenia or bipolar disorder. Evidence on the effectiveness and safety of long-acting injectable antipsychotics (LAIAs) in people with intellectual disability (PWID) is very limited. Existing studies are small, short-term and mostly do not compare LAIAs with oral antipsychotics. It remains unclear whether LAIAs are beneficial or harmful in PWID, particularly according to the presence of schizophrenia or bipolar disorder. WHAT THIS STUDY ADDS In this population-based self-controlled case series, LAIA use was associated with more favourable outcomes in PWID with schizophrenia or bipolar disorder, including fewer hospitalisations, psychiatric admissions and extrapyramidal symptoms, although cardiovascular hospitalisations were higher. In PWID without schizophrenia or bipolar disorder, LAIA use was associated with worse overall outcomes and no clear benefit for behaviour-related problems. This suggests that the effects of LAIAs differ substantially by underlying diagnosis. In this population-based self-controlled case series, LAIA use was associated with more favourable outcomes in PWID with schizophrenia or bipolar disorder, including fewer hospitalisations, psychiatric admissions and extrapyramidal symptoms, although cardiovascular hospitalisations were higher. In PWID without schizophrenia or bipolar disorder, LAIA use was associated with worse overall outcomes and no clear benefit for behaviour-related problems. This suggests that the effects of LAIAs differ substantially by underlying diagnosis. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY LAIAs may be considered in PWID with confirmed schizophrenia or bipolar disorder. In PWID without these diagnoses, LAIAs should be prescribed more cautiously, and deprescribing may be appropriate if these results are replicated in other studies. Future studies should be replicated in other regions and using alternative methods to help triangulate the findings. LAIAs may be considered in PWID with confirmed schizophrenia or bipolar disorder. In PWID without these diagnoses, LAIAs should be prescribed more cautiously, and deprescribing may be appropriate if these results are replicated in other studies. Future studies should be replicated in other regions and using alternative methods to help triangulate the findings. Background Intellectual disability (ID) encompasses a heterogeneous array of neurodevelopmental disorders characterised by significant limitations in both intellectual functioning (typically IQ<70) and adaptive behaviour. Approximately 1%–3% of the global population live with ID.1 People with ID (PWID) have a higher prevalence of psychiatric conditions, with 8.2% affected by psychotic disorders and 2.6% affected by bipolar disorder.2 Antipsychotics are a core treatment for schizophrenia and bipolar disorder. However, medication non-adherence remains challenging,3 particularly among PWID,4 due to impaired understanding, insight, organisation and other aspects of intellectual functioning. Treatment non-adherence can lead to increased hospitalisation, aggression and reduced quality of life.3 Long-acting injectable antipsychotics (LAIAs), also known as depot antipsychotics, are administered at intervals from weeks to months, offering greater convenience for patients and enhancing adherence. In addition, LAIAs have other advantages over oral antipsychotics (OAs), such as more stable plasma drug concentrations, lower relapse risk and adverse effects. Real-world evidence supports that LAIAs are associated with reduced risks of disease relapse, which may reduce hospitalisation and other healthcare utilisation in people with schizophrenia and bipolar disorder.5–7 The benefits are also evident in vulnerable populations, such as the elderly with schizophrenia8 and forensic patients with schizophrenia.9 However, research on LAIA use in PWID, another vulnerable group, remains limited. A study using data in the Czech Republic (n=62 636) found that one-third of PWID received antipsychotics, but only 19% were indicated for severe mental illness (SMI), implying off-label use in over 70%–80% of cases, raising significant safety concerns.10 In Hong Kong, 34% of community-dwelling PWID (n=239) were prescribed antipsychotics without a recorded SMI.11 This off-label prescribing targets challenging behaviours (eg, aggression, self-harm, property destruction), despite international guidelines recommending non-pharmacological interventions first and antipsychotics only as a last resort.12–14 Evidence supporting antipsychotics for challenging behaviours in PWID is limited and inconsistent,15 despite their frequent use in attempts to manage crisis-level behaviours and reduce healthcare use. Existing studies are small, short and often lack control groups, and two randomised placebo-controlled trials found placebo superior to antipsychotics.15–19 For LAIAs, two small uncontrolled studies reported behavioural improvement,16 17 but effectiveness and safety relative to OAs remain unclear. PWID are particularly vulnerable to certain adverse events associated with antipsychotics, notably extrapyramidal symptoms (EPS) and neuroleptic malignant syndrome,20 which may contribute to emergency visits and hospitalisations. Case reports have highlighted safety concerns with long-acting fluphenazine decanoate in PWID,21 while an Australian review cautioned against LAIA use in PWID with schizophrenia due to EPS concerns.22 Although two single-arm studies, one in paediatric patients with autism and ID (n=26) and one in adults with ID (n=34),16 17 reported acceptable LAIA tolerability, their small samples and lack of OA comparators limit conclusions and highlight the need for safety data comparing LAIAs with OAs. Objective To examine the associations of LAIAs versus OAs with healthcare utilisation outcomes related to effectiveness and safety among PWID, overall and by schizophrenia/bipolar disorder status. Methods Data source Data were obtained from the Clinical Data Analysis and Reporting System, the electronic health records database of the Hong Kong Hospital Authority (HA). The HA manages all public hospitals in Hong Kong, serving more than 7.5 million residents, accounting for 90% of all admissions in the territory.23 The database records diagnoses, procedures, dispensing records, laboratory results, demographics and admission records. It has been widely used to analyse outcomes in high-quality pharmacoepidemiologic studies.6 8 Self-controlled case series This study employed the self-controlled case series (SCCS) method, using within-individual Poisson regression to estimate incidence rate ratios (IRRs) by comparing outcome rates during LAIA versus OA periods. Only individuals who experienced the outcome of interest and had discrete exposure periods for both formulations were included, following the standard SCCS case-only design. This design is well suited to PWID, among whom switching between LAIAs and OAs is common, because each individual serves as their own control and all time-invariant characteristics (eg, sex, genetics, family history, ethnicity, socioeconomic status, baseline disease severity) are controlled by design, which is valuable when many important characteristics are not well captured in routine data. Time-varying covariates such as age, calendar year, season and concomitant medications were adjusted in the models. Patient identification The initial cohort was identified by the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes 317–319 for ID, recorded from 1 January 1993 to 31 December 2023. Cases were included if they had received at least one OA prescription, defined as British National Formulary (BNF) code 4.2.1 (excluding short-acting immediate-release injections) and at least one LAIA prescription, defined as BNF code 4.2.2 and at least one of the corresponding outcome events. The observation period began on the latest of: 1 January 2004; the date of the first diagnosis of ID; or the date of turning 18 years old. The end of observation was defined as the earliest of: 31 December 2023; or the date of death. Patients were excluded if age, sex, date of birth or date of death data were missing; if they died before 1 January 2004; if they had no follow-up, defined as the observation start date being the same as the observation end date; if they were younger than 18 years during the observation period; or if they were lost to follow-up before 1 January 2004. For SCCS, the set of cases analysed for each outcome could differ because only patients with outcome events (see later) are included in the corresponding analyses. Exposures The observation period was categorised into four mutually exclusive exposure periods (online supplemental eFigure 1): no treatment, OAs-only, LAIAs-only and concurrent use of both. Exposure was defined at the formulation level rather than by individual antipsychotic agent; thus, LAIA and OA periods represented time exposed to any LAIA or any OA of the same formulation class, respectively. Because the study was formulated at the class level, changes between individual agents within the same formulation category were not separately modelled. Concurrent-use periods likely reflect cross-titration, oral supplementation or greater illness severity and thus represent a clinically complex treatment phase. They were modelled separately and do not materially affect the main LAIA-only versus OA-only comparison. Exposure periods for antipsychotics were determined from prescription start and end dates, adding five half-lives of the OA (half-life range, 8–94 hours) to the end date to account for residual effects (online supplemental eTable 1). In addition to full treatment period analyses, we conducted analyses comparing only the subsequent treatment periods, defined as periods beyond the first 90 days of each treatment episode (online supplemental eFigure 1). This analysis focused on the maintenance phase, accounted for residual effects of previous treatments,6 8 and helped reduce confounding by indication arising from risk modification related to the indication for starting treatment, especially in the initial period. Pre-exposure time bias in SCCS arises from event-dependent exposure, when the occurrence of the event influences the likelihood or timing of subsequent exposure. In our data, we included a 30-day pre-exposure period in the model as separate exposure periods to adjust for this bias. Outcomes Primary outcomes included all-cause emergency department (ED) visits, all-cause hospitalisations and hospitalisations for psychiatric disorders, schizophrenia/bipolar disorder and behaviour-related problems. All-cause ED visits and hospitalisations were included to capture overall healthcare burden and the balance between potential benefits and harms of antipsychotic treatment. Secondary outcomes included hospitalisations for somatic disorders generally and cardiovascular diseases (CVDs) specifically, and diagnosis records of EPS, falls and fractures, all defined in online supplemental eTable 2. Hospitalisations for behaviour-related problems were defined as admissions with diagnoses indicating disruptive, aggressive or conduct-related behaviours; self-injury or suicidality; behavioural disturbance; personality or impulse-control disorders; or other behavioural problems. Since the coding algorithm for behaviour-related problems was not validated previously, we also did a post-hoc sensitivity analysis with a narrower definition. Statistical analyses Conditional Poisson regression was used to calculate the adjusted incidence rate ratios (aIRRs) and the corresponding 95% CIs of outcomes between LAIAs-only and OAs-only periods. Time-varying covariates specified in online supplemental eTable 3—including age, season, calendar year, inpatient stay (for analyses of records of EPS, falls and fractures), and concurrent use of mood stabilisers, antidepressants, benzodiazepines or Z-drugs and antiparkinsonian drugs—were adjusted. For the analyses of ED visits and hospitalisations, the outcome was defined as the start of a new ED visit or inpatient admission. Once an admission had begun, inpatient days after the day of admission were modelled as a separate period for this outcome, because a new ED visit or hospital admission could not begin until discharge. Sensitivity analyses First, we restricted analyses to the first event to assess independence of recurrent events. Second, we excluded patients who died during follow-up to evaluate event-dependent censoring (early termination of follow-up due to death, which prevents observation of later events). Third, we excluded ‘as-needed’ (p.r.n.) OA prescriptions to test for exposure misclassification, as these prescriptions may lead to potentially lower actual exposure. Fourth, because clozapine is a highly effective OA reserved for treatment-resistant schizophrenia/bipolar disorder, potentially making OA periods appear more effective or increasing outcome risk due to confounding, we excluded individuals with clozapine prescriptions. Finally, we analysed hospitalisations for upper respiratory tract infection as a negative control outcome to detect residual bias, including monitoring bias, based on prior evidence of no difference between LAIAs and OAs.6 8 We repeated analyses without the pre-exposure periods and tested 60-day and 90-day pre-exposure periods to confirm that including 30-day periods was sufficient to address the pre-exposure time bias. Subgroup analyses Subgroup analyses were conducted by sex and schizophrenia/bipolar disorder status. Patients were classified as ‘with schizophrenia/bipolar disorder’ from the first diagnosis onward and as ‘without schizophrenia/bipolar disorder’ if they never received either diagnosis. We also conducted a sensitivity analysis restricted to person-time before or without a schizophrenia/bipolar disorder diagnosis, including up to 30 days before first diagnosis among those who developed the condition during follow-up. Post-hoc sensitivity analyses further stratified schizophrenia and bipolar disorder separately, and ID severity and compared LAIA and OA periods for first-generation and second-generation antipsychotics (FGA and SGA) in subgroups with and without schizophrenia/bipolar disorder. Interaction analyses were used to estimate relative incidence ratios, assessing effect modification by sex and by schizophrenia/bipolar disorder status. E-values, the minimum risk ratio strength an unmeasured confounder must have with both exposure and outcome to nullify an observed association in observational research, were calculated to assess the robustness of the results to unmeasured confounding. Findings with p values <0.05 were considered statistically significant. Analyses were conducted in R V.4.4.1 using the survival package. Two independent investigators (SML and KJL) performed and crosschecked the analyses. The study followed the Strengthening the Reporting of Observational studies in Epidemiology (STROBE) guidelines. Role of the funding source The funder had no role in the study design, data collection, analysis, interpretation, nor writing of the manuscript. Findings Among 29 036 PWID identified in our database, 24 382 remained after data cleaning (age range: 18.0–103.5 years), with 1675 who received both OAs and LAIAs (online supplemental eFigure 2). Online supplemental eTable 4 shows the distribution of severity of ID. A total of 1675 PWID entered the observation period at a mean (SD) age of 35.7 (12.0) years (range: 18.0–76.0 years); 718 (42.9%) were female (online supplemental eTable 5). During a mean (SD) follow-up of 15.7 (5.4) years, 304 (18.1%) died at a mean (SD) age of 56.2 (12.0) years. Mean (SD) exposure durations were 7.6 (6.6) years for OAs-only, 1.5 (3.2) years for LAIAs-only and 5.0 (5.9) years for concurrent use. Characteristics of subgroups with and without schizophrenia/bipolar disorder are shown in online supplemental eTable 6. In the subgroup of schizophrenia/bipolar disorder (n=1205), 1117 PWID entered the subgroup with diagnostic records of schizophrenia, and 134 PWID entered the subgroup with diagnostic records of bipolar disorder. Online supplemental eTable 7 shows relatively higher LAIA than OA doses, with a numerically larger difference in the subgroup without schizophrenia/bipolar disorder than in the subgroup with schizophrenia/bipolar disorder. Online supplemental eTable 8 presents the number of events, person-years and incidence rates per 100 person-years for the OA-only and LAIA-only periods as well as the main analysis results, while descriptive data for the subgroup with and without schizophrenia/bipolar disorder and concurrent OA and LAIA periods are shown in online supplemental eTables 9 and 10. Because the first-event-only sensitivity analyses showed a lower risk of incident all-cause ED visits but a higher risk of recurrent all-cause ED visits during LAIA periods, suggesting event dependence (where an initial event alters the risk of subsequent events), violating assumption of SCCS, we, therefore, replaced recurrent all-cause ED visits with incident all-cause ED visits as the primary outcome in our main model to avoid event dependence. During the full treatment period, LAIAs were associated with lower risks of incident all-cause ED visits (aIRR=0.71, 95% CI 0.51 to 0.98), all-cause hospitalisations (aIRR=0.91, 95% CI 0.85 to 0.98), hospitalisations for psychiatric disorders generally (aIRR=0.90, 95% CI 0.81 to 0.99), schizophrenia/bipolar disorder specifically (aIRR=0.86, 95% CI 0.75 to 0.99) and records of EPS (aIRR=0.34, 95% CI 0.23 to 0.51) compared with OAs. No significant differences were observed for hospitalisations for behaviour-related problems, somatic disorders and CVDs, or records of falls and fractures. These results remained consistent when restricting to the subsequent treatment period, with additionally a lower risk of hospitalisations for somatic disorders (aIRR=0.86, 95% CI 0.78 to 0.96) associated with the use of LAIAs (online supplemental eTable 8). Figure 1 shows the results of subgroup analyses. After sex stratification, most results were generally consistent with the main analyses. Notable exceptions included hospitalisations for behaviour-related problems, somatic disorders and CVDs. Among females, LAIAs were associated with a higher risk of hospitalisations for behaviour-related problems (aIRR=1.31, 95% CI 1.02 to 1.68), but the opposite was recorded in males (aIRR=0.81, 95% CI 0.67 to 0.98). LAIAs were associated with a lower risk of hospitalisations for somatic disorders only among females (aIRR=0.78, 95% CI 0.68 to 0.90) and a higher risk of hospitalisations for CVDs only among males (aIRR=1.73, 95% CI 1.19 to 2.50) (figure 1). Online supplemental eTable 11 indicates that sex significantly modifies the associations for these outcomes. Sex-stratified results in the subsequent treatment period were generally consistent with the full-period analyses (online supplemental eFigure 3). The most common diagnoses for behaviour-related hospitalisations were mild and moderate ID with significant impaired behaviour requiring attention/treatment (online supplemental eTable 12). Results of subgroup analyses during the full treatment period. Adjusted for pre-exposure time bias by including 30-day pre-exposure periods in model, also adjusted for age, season, year, inpatient hospitalisation (for analyses of diagnosis records of EPS, falls and fractures), antiparkinsonian drugs, antidepressants, benzodiazepines or z-drugs, mood stabilisers. Incident all-cause ED visits analysis is replacing the all-cause ED visits analysis as the primary model due to risk of event dependence. aIRR, adjusted incidence rate ratio; ED, emergency department; EPS, extrapyramidal symptoms; LAIAs, long-acting injectable antipsychotics; OAs, oral antipsychotics. For PWID with schizophrenia/bipolar disorder, LAIAs showed a broadly favourable pattern: lower risks of all-cause hospitalisations, psychiatric, schizophrenia/bipolar, behaviour-related and somatic hospitalisations, and recorded EPS, but a higher risk of CVD hospitalisations (figure 1, online supplemental eFigure 3). The most common diagnosis for CVD hospitalisations in both LAIA and OA periods was hypertensive disease (online supplemental eTable 13). For PWID without schizophrenia/bipolar disorder, although LAIA periods were associated with lower risks of incident all-cause ED visits and hospitalisations for CVDs, LAIA periods were generally associated with a less favourable pattern, including higher all-cause hospitalisations and higher hospitalisations for somatic disorders, but no clear improvement in psychiatric or behaviour-related hospitalisations. Online supplemental eTable 11 indicates that schizophrenia/bipolar disorder status significantly modifies the associations for all outcomes except incident all-cause ED visits, falls and fractures. These findings were consistent in subsequent treatment period analyses, with an additional increase in hospitalisations for psychiatric disorders (online supplemental eTable 13). The subgroup with and without schizophrenia/bipolar disorder had similar baseline rates of CVD diagnoses (online supplemental eTable 14). The model without 30-day pre-exposure periods showed diverse outcomes (online supplemental eTable 15), evidencing pre-exposure time bias before the adjustment. Analyses using 60-day and 90-day pre-exposure period yielded results consistent with the main analyses (online supplemental eTable 16), indicating that including 30-day pre-exposure periods in the model corrected the bias. Online supplemental eTable 17 presents the results of analyses stratified by schizophrenia, bipolar disorder and severity of ID. Some estimates were unstable in the bipolar subgroup and in the moderate and severe/profound ID groups because of smaller sample sizes and fewer events. Results stratified by ID severity were consistent with the main analysis. Results stratified by schizophrenia and bipolar disorder were also generally consistent with the subgroup analysis when they were combined, except that the schizophrenia subgroup showed no significant difference (p=0.052) in hospitalisations for schizophrenia during LAIA periods. Besides event dependence in all-cause ED visits, other sensitivity analyses produced broadly consistent results, with a few exceptions. Excluding PWID who died during follow-up altered the risk of fractures during subsequent LAIA periods (online supplemental eTable 18) and the risks of falls and fractures in the subgroups (online supplemental eFigure 4), prompting further diagnostic sensitivity analyses for event-dependent censoring for these outcomes. Online supplemental eTable 19 shows that censoring was not a significant effect modifier of the exposure effect, online supplemental eTable 20 indicates that the most common cause of death was respiratory conditions, with no deaths directly attributable to falls or fractures, and online supplemental eFigure 5 shows that the distributions of intervals from event to end of observation were similar for censored and uncensored cases. Taken together, these findings suggest that falls and fractures are unlikely to have caused substantial event-dependent censoring. Online supplemental eTable 21 presents the results of a sensitivity analysis using a narrower definition of hospitalisations for behaviour-related problems, which yielded smaller event counts. The overall patterns were similar to those in the main analysis, although the associations became non-significant across all subgroups. Online supplemental eTable 22 shows analyses by antipsychotic generation. In the subgroup with schizophrenia/bipolar disorder, results for FGA-LAIAs were consistent with the main subgroup analyses across outcomes. While for SGA-LAIAs, only all-cause and psychiatric hospitalisations remained significantly lower than during SGA-OA periods. In contrast, in the subgroup without schizophrenia/bipolar disorder, all associations for FGA-LAIAs versus FGA-OAs were non-significant except for cardiovascular hospitalisations, whereas SGA-LAIAs were associated with a generally less favourable pattern, including higher risks of all-cause, psychiatric, behaviour-related and somatic hospitalisations, and higher rates of EPS, compared with SGA-OAs. In the negative control outcome analysis, no significant difference was observed (online supplemental eTable 18). E-value analyses showed that our results were unlikely to be explained by an unmeasured confounder (online supplemental eTable 23). Discussion This is, to our knowledge, the first study to evaluate the long-term associations of LAIAs versus OAs with effectiveness and safety outcomes in PWID, addressing a critical evidence gap. Marked heterogeneity by psychiatric diagnosis was observed. Among PWID with schizophrenia/bipolar disorder—for whom antipsychotics are clinically indicated—LAIAs were associated with more favourable outcomes, consistent with better effectiveness and safety. These associations may in part reflect improved adherence with LAIAs and more stable serum drug concentrations. The findings align with real-world evidence in other populations5 6 8 and are consistent with potential benefit of LAIAs in PWID with schizophrenia/bipolar disorder. The lower hospitalisations for behaviour-related problems observed in PWID with schizophrenia/bipolar disorder when using LAIAs are consistent with links between disease relapse and aggression.24 Conversely, in PWID without schizophrenia/bipolar disorder—who often receive off-label antipsychotics for behaviour management—LAIAs were associated with increased all-cause hospitalisations and somatic-disorder hospitalisations, without clear improvement in behaviour-related hospitalisations. This pattern suggests that excess admissions may be driven mainly by somatic disorders, possibly related to higher effective antipsychotic exposure during LAIA periods. The lack of behavioural benefit may be theoretically compatible with a predominance of more planned or goal-directed forms of aggression, which have been proposed to be less responsive to psychotropics25; however, our data did not include aggression subtype measures and cannot test this hypothesis. Overall, the adherence advantages of LAIAs did not translate into measurable off-label behavioural benefit in PWID without schizophrenia/bipolar disorder. Because this was an observational SCCS study, the observed associations may still reflect time-varying confounding by indication, disease severity, behavioural instability, service contact, non-pharmacological interventions, body weight and medication adherence. The findings should, therefore, be interpreted as associations rather than evidence of causality. LAIAs may have been initiated in response to clinical deterioration, relapse, worsening behavioural disturbance, poor adherence, increased service contact or changes in caregiving and residential support. These factors could increase the risk of subsequent hospitalisations, psychiatric admissions and behaviour-related admissions during LAIA periods, potentially underestimating any true protective association of LAIAs, particularly in PWID with schizophrenia/bipolar disorder. Conversely, LAIA prescribing may coincide with improved follow-up, enhanced caregiver support, residential placement changes or non-pharmacological interventions, and with higher effective antipsychotic exposure than OAs, which could exaggerate apparent benefits. Thus, residual confounding could bias estimates in either direction. In PWID without schizophrenia/bipolar disorder, similar time-varying confounding mechanisms may also have been present. LAIAs may have been initiated during periods of behavioural escalation, poorer adherence or greater clinical concern, which would tend to increase adverse events during LAIA periods, while dose-related differences and co-occurring non-pharmacological interventions may have favoured the LAIA period and biased estimates towards benefit. However, despite these potentially favourable factors, we still observed higher hospitalisations without clear behavioural benefit. Therefore, even if LAIAs conferred some true benefit in this subgroup, it appears insufficient to offset the underlying confounding and the excess adverse outcomes, unlike in the subgroup with schizophrenia/bipolar disorder. These mechanisms may differ between people with and without schizophrenia/bipolar disorder, so the diagnostic subgroup differences we observed probably reflect a combination of true treatment effects and residual confounding by indication and patient selection. More favourable outcomes with LAIAs in the schizophrenia/bipolar subgroup may partly reflect both genuine treatment benefits and residual confounding, whereas less favourable outcomes in those without these diagnoses may partly reflect preferential use of LAIAs in the most complex or unstable cases. Previous meta-analyses of randomised controlled trials in schizophrenia have reported relative risks (RR) of approximately 0.70–0.82 in favour of LAIAs,26 27 similar to our IRR of 0.80 for LAIA versus OA periods for hospitalisations for schizophrenia/bipolar disorder and IRR of 0.76 for psychiatric hospitalisations in subgroup with schizophrenia/bipolar disorder, suggesting that residual confounding in this subgroup may be limited. However, because comparable trials are lacking for other outcomes and for people without schizophrenia/bipolar disorder, the extent of residual confounding in those contexts remains uncertain. Stratified analyses of schizophrenia and bipolar disorder separately suggest that the reduction in hospitalisations for schizophrenia/bipolar disorder in the combined subgroup may be partly driven by a stronger association with lower bipolar-disorder admissions in the bipolar subgroup. The lack of significant reduction in risk of hospitalisations for schizophrenia may reflect residual time-varying confounding by indication, such as initiation during greater clinical instability, relapse or behavioural escalation, whereas the bipolar subgroup may have been less affected by these factors. Together, these findings suggest heterogeneity across diagnostic subgroups and should be interpreted cautiously. The significant IRR of 0.78 for psychiatric hospitalisations in the schizophrenia subgroup is broadly consistent with estimates from meta-analyses of randomised controlled trials,26 27 whereas the IRR of 0.45 in the bipolar subgroup is lower than the reported relative risk of 0.58 in a previous meta-analysis of RCTs,28 suggesting that residual confounding in our bipolar subgroup may bias the association in favour of LAIA periods. In this study, LAIA prescriptions were generally at higher doses than OA prescriptions. Among patients with schizophrenia/bipolar disorder, this higher effective exposure may partly explain the lower rates of psychiatric, schizophrenia/bipolar-related and behaviour-related hospitalisations, and therefore fewer all-cause hospitalisations, although it does not readily explain the lower somatic-disorder hospitalisations or EPS. Among patients without schizophrenia/bipolar disorder, higher LAIA doses may partly explain the higher rates of somatic-disorder hospitalisations, and therefore more all-cause hospitalisations, but not the lower rates of CVD hospitalisations. The lower recorded EPS rate during LAIA periods in PWID was consistent with a previous study in patients with schizophrenia.6 EPS, however, may be under-recorded in routine data, and the lower rates during LAIA periods may partly reflect under-recognition or differential recording. LAIAs are often initiated after an oral tolerability trial, which may bias recording towards fewer EPS events. At the same time, LAIA prescriptions in our study involved higher doses than OA prescriptions, which would be expected to increase EPS risk. In the schizophrenia/bipolar subgroup, the apparent protective effect may, therefore, reflect the balance of these competing factors, with confounding and tolerability screening potentially outweighing the effect of higher LAIA exposure. In contrast, in the subgroup without schizophrenia/bipolar disorder, the larger dose difference between LAIAs and OAs may have outweighed the tolerability-trial effect, which may explain the absence of a significant difference in EPS risk in that subgroup. LAIAs were associated with lower somatic-disorder hospitalisations but higher CVD hospitalisations in PWID with schizophrenia/bipolar disorder, with the opposite pattern in those without these diagnoses, despite similar baseline CVD diagnoses. This subgroup-specific cardiovascular pattern was driven mainly by differences between FGA-LAIAs and OAs. The divergent associations may reflect competing mechanisms and confounding. LAIA periods may reduce some adverse events by preventing treatment interruption and discontinuation symptoms, the tolerability-trial effect, and may coincide with better care engagement and closer follow-up, which could lower outcome risk. At the same time, improved adherence and higher prescribed doses may increase actual antipsychotic exposure and some adverse outcomes, and closer monitoring may increase detection and recording of hospitalisations. These mechanisms are likely present in both diagnostic subgroups, but their balance may differ. In patients with schizophrenia/bipolar disorder, unmeasured postdiagnosis lifestyle changes may act as effect modifiers,29 30 interacting with higher LAIA exposure to influence cardiovascular risk. In patients without schizophrenia/bipolar disorder, LAIA use may have been initiated during behavioural escalation, potentially related to underlying somatic illness, which could confound associations with somatic hospitalisations. Because lifestyle factors, reasons for LAIA prescribing, and whether behavioural worsening was driven by somatic illness were not captured, these explanations remain speculative. Several shared mechanisms may explain the less favourable pattern observed with SGA-LAIAs relative to FGA-LAIAs, seen in both diagnostic subgroups. First, confounding by indication related to clinical familiarity and severity is plausible: physicians may be more comfortable initiating FGA-LAIAs given their longer track record, reserving SGA-LAIAs, which are also costlier in the public setting, for more difficult or complex cases. Second, confounding by indication related to tolerability may also contribute: because SGAs are perceived as having fewer adverse effects, particularly EPS, patients who experienced adverse events on FGAs may be preferentially switched to SGA-LAIAs, so the SGA-LAIA-treated group may include a higher proportion of patients already prone to adverse events. Third, as a newer option, SGA-LAIA initiation may more often coincide with clinical escalation or deterioration, again reflecting confounding by indication, whereas FGA-LAIA use may more often represent stable, legacy maintenance therapy with comparatively fewer events. This study has several strengths. The inclusion of a wide range of antipsychotics reduces the influence of drug-specific factors and allows the analyses to focus on formulation-level associations, enhancing generalisability. Adjustment for key time-varying covariates further helps reduce residual confounding. The SCCS design, described in the Methods section, improves internal validity by controlling for all time-invariant confounders and accommodating key time-varying covariates. Still, there are limitations. Because schizophrenia and bipolar disorder were identified from routinely recorded diagnostic codes, some under-recording of psychiatric diagnoses cannot be completely excluded, and a degree of misclassification of off-label antipsychotic use is possible. However, the proportion of antipsychotic users with recorded schizophrenia/bipolar disorder in our study was comparable to that reported in other population-based studies of adults with IDs,10 31 and the prevalence of schizophrenia/bipolar disorder among PWID in our dataset was similar to estimates from other regions,2 suggesting that substantial under-recording is unlikely. The large number of outcomes and subgroup analyses increases the risk of chance findings. Our analytic sample comprised individuals with ID who had used both OAs and LAIAs at different times during follow-up; therefore, the findings reflect within-individual differences between these treatment periods and may not be generalisable to individuals who received only OAs or only LAIAs. The sample was enriched for mild ID, so the results may not apply to people with moderate, severe or profound ID. Our main outcomes were hospital admissions and ED visits, which capture only the more severe end of the harm spectrum; milder but clinically important deterioration in behaviour or functioning without admission as well as other forms of healthcare use such as outpatient visits, psychotherapy, behavioural and cognitive therapies may not be reflected. Nevertheless, for comparative analysis, the relative differences in rates between LAIAs and OAs remain meaningful even if not all clinical events are captured, and they provide a useful parameter for evaluating the overall benefit–risk balance. Moreover, admission thresholds depend on local practice patterns, service availability, culture and policy, which vary across settings; so our findings may not be fully generalisable to settings with different models of inpatient admission. Also, actual adherence was not recorded, and exposure end dates may have differed from prescription end dates if treatment stopped early by patient or prescriber. Finally, the sample size and duration of LAIA exposure periods for SGAs were smaller than those of other exposure periods, so results may not be generalisable to SGA-LAIA versus SGA-OA comparisons. Future studies should include dosage, markers of clinical severity and care context, such as prior hospitalisation, outpatient psychiatric contact, residential care status, caregiver support and behavioural interventions, and should be replicated in other regions and using alternative methods to help triangulate the findings. Given the limited sample size, we did not stratify by individual antipsychotic agent, but drug-specific analyses would be valuable in future work. Clinical implications Our findings can help inform targeted clinical guidance for LAIA use in PWID across diagnostic subgroups. For PWID with schizophrenia/bipolar disorder—where guidelines support the use of LAIAs,32 33 but lack specific recommendations for this vulnerable population—our data help contextualise prior safety concerns by showing more favourable associations for several effectiveness and safety measures compared with OAs. Conversely, for PWID without schizophrenia/bipolar where guidelines advise cautious and short-term antipsychotic use for behaviours without specifically mentioning use of LAIAs, we observed no clear benefit for behaviour-related hospitalisations and higher rates of all-cause hospitalisations, which raises concern about routine off-label use, suggesting clinicians may wish to consider deprescribing where clinically appropriate. For PWID continuing antipsychotics for behaviours, clinicians should conduct scheduled reviews at least every 6 months, assessing behavioural ratings, physical health, adverse effects and the use of non-pharmacological interventions. Deprescribing should be considered regularly. Suitable candidates include those with prolonged antipsychotic use and no clear evidence of benefit from the current medication, with withdrawal undertaken gradually and one medicine at a time. Decisions should be made within a multidisciplinary team, including pharmacists, physicians, nurses, social workers and caregivers, with a documented plan specifying review intervals, non-pharmacological supports and criteria for reinitiating medication if behaviour deteriorates. Where deprescribing is not pursued, the rationale for continuing treatment should be recorded, alongside ongoing regular review.12 14 Conclusions Our findings suggest that LAIAs may be beneficial for PWID with schizophrenia/bipolar disorder, although some associations may be influenced by residual confounding and should be replicated in future studies. However, the lack of improvement in psychiatric and behavioural outcomes in PWID without schizophrenia/bipolar disorder remains concerning, especially given the more favourable pattern seen in the schizophrenia/bipolar disorder subgroup, which serves as an internal reference subgroup with a similar confounding structure. This makes confounding alone an unlikely explanation. Taken together, the results do not support routine off-label use of LAIAs for behavioural management in PWID and deprescribing may be considered in appropriate clinical settings if these results are replicated in other studies. Supplementary material Footnotes Funding: Excellent Young Scientists Fund of the NSFC of China (Project number: 82022079). Provenance and peer review: Not commissioned; externally peer-reviewed. Patient consent for publication: Not applicable. Ethics approval: The study was approved by the Institutional Review Board of the University of Hong Kong/Hospital Authority Hong Kong West Cluster (reference number: UW 20-113). Participants gave informed consent to participate in the study before taking part. Data availability free text: The data cannot be shared with others due to restrictions imposed by the data custodians. However, the analytic code can be made available upon reasonable request. Data availability statement Data are available upon reasonable request. References Associated Data Supplementary Materials Data Availability Statement Data are available upon reasonable request.