Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 111 of the 111 references it cites were found in a registry or an index.

Big data approaches to decomposing heterogeneity across the autism spectrum

Open access
Lombardo, Michael V., Lai, Meng-Chuan, Baron-Cohen, Simon · 2019 · Molecular Psychiatry

Published version of a preprint

Abstract

Autism is a diagnostic label based on behavior. While the diagnostic criteria attempt to maximize clinical consensus, it also masks a wide degree of heterogeneity between and within individuals at multiple levels of analysis. Understanding this multi-level heterogeneity is of high clinical and translational importance. Here we present organizing principles to frame research examining multi-level heterogeneity in autism. Theoretical concepts such as 'spectrum' or 'autisms' reflect non-mutually exclusive explanations regarding continuous/dimensional or categorical/qualitative variation between and within individuals. However, common practices of small sample size studies and case-control models are suboptimal for tackling heterogeneity. Big data are an important ingredient for furthering our understanding of heterogeneity in autism. In addition to being 'feature-rich', big data should be both 'broad' (i.e., large sample size) and 'deep' (i.e., multiple levels of data collected on the same individuals). These characteristics increase the likelihood that the study results are more generalizable and facilitate evaluation of the utility of different models of heterogeneity. A model's utility can be measured by its ability to explain clinically or mechanistically important phenomena, and also by explaining how variability manifests across different levels of analysis. The directionality for explaining variability across levels can be bottom-up or top-down, and should include the importance of development for characterizing changes within individuals. While progress can be made with 'supervised' models built upon a priori or theoretically predicted distinctions or dimensions of importance, it will become increasingly important to complement such work with unsupervised data-driven discoveries that leverage unknown and multivariate distinctions within big data. A better understanding of how to model heterogeneity between autistic people will facilitate progress towards precision medicine for symptoms that cause suffering, and person-centered support.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41380-018-0321-0

Cites

111 of 111 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Approaches to decomposing heterogeneity in autism. a A population of interest is shown, and autism cases are colored in green, pink, and blue. The different colors are meant to represent different autism subtypes. In b we show the impact of ignoring heterogeneity on effect size. With a typical case–control model, we ignore these possible subtype distinctions and compare autism to controls on some dependent variable. In this example scenario there is no clear case–control difference but the autism group shows higher variability (indicated by the larger error bars). An approach towards decomposing heterogeneity might be to construct a stratified model whereby we model the subtype labels instead of one autism label, and then re-examine differences on the hypothetical dependent variable of interest. In this example, the autism subtypes show contradictory effects. These effects are masked in the case–control model as the averaging cancels out the interesting different effects across the subgroups. c Heterogeneity is shown in autism as multi-level phenomena. This panel also visualizes the difference between broad versus deep big data characteristics and labels the top-down versus bottom-up approaches to understanding heterogeneity in this multi-level context. Finally, this panel also shows how development is another important dimension of heterogeneity to consider at each level of analysis (i.e., ‘chronogeneity’). In this example, chronogeneity is represented by different trajectories for different types of autism individuals
Fig. 1 Approaches to decomposing heterogeneity in autism. a A population of interest is shown, and autism cases are colored in green, pink, and blue. The different colors are meant to represent different autism subtypes. In b we show the impact of ignoring heterogeneity on effect size. With a typical case–control model, we ignore these possible subtype distinctions and compare autism to controls on some dependent variable. In this example scenario there is no clear case–control difference but the autism group shows higher variability (indicated by the larger error bars). An approach towards decomposing heterogeneity might be to construct a stratified model whereby we model the subtype labels instead of one autism label, and then re-examine differences on the hypothetical dependent variable of interest. In this example, the autism subtypes show contradictory effects. These effects are masked in the case–control model as the averaging cancels out the interesting different effects across the subgroups. c Heterogeneity is shown in autism as multi-level phenomena. This panel also visualizes the difference between broad versus deep big data characteristics and labels the top-down versus bottom-up approaches to understanding heterogeneity in this multi-level context. Finally, this panel also shows how development is another important dimension of heterogeneity to consider at each level of analysis (i.e., ‘chronogeneity’). In this example, chronogeneity is represented by different trajectories for different types of autism individualsCC BY · PubMed Central
Fig. 2 Case–control vs stratified model example with adult autism and mentalizing ability. This figure reports data from Lombardo et al. [25] on two independent datasets of adults with autism and performance on an advanced mentalizing test, the Reading the Mind in the Eyes Test (RMET). a (Discovery), b (Replication) Case–control differentiation and the standardized effect size for each dataset are shown. c–f RMET scores and standardized effect sizes from the same two datasets after unsupervised data-driven stratification into five distinct autism subgroups and four distinct TD subgroups. Autism subgroups 1–2 are highly impaired on the RMET, while autism subgroups 3–5 are completely overlapping in RMET scores with the TD population
Fig. 2 Case–control vs stratified model example with adult autism and mentalizing ability. This figure reports data from Lombardo et al. [25] on two independent datasets of adults with autism and performance on an advanced mentalizing test, the Reading the Mind in the Eyes Test (RMET). a (Discovery), b (Replication) Case–control differentiation and the standardized effect size for each dataset are shown. c–f RMET scores and standardized effect sizes from the same two datasets after unsupervised data-driven stratification into five distinct autism subgroups and four distinct TD subgroups. Autism subgroups 1–2 are highly impaired on the RMET, while autism subgroups 3–5 are completely overlapping in RMET scores with the TD populationCC BY · PubMed Central
Fig. 3 Simulation of sample effect size estimates at different sample sizes and across a range of true population effects for a hypothetical case–control study. In this simulation we set the population effect size to a range of different values, from very small (e.g., d = 0.1) to very large (e.g., d > 1.0) (panels a–e show simulation results when effect size ranges from d = 0.1 to d = 0.9 in steps of 0.2). We then simulated data from two populations (cases and controls), each with n = 10,000,000, that had a case–control difference at these population effect sizes. Next, we simulated 10,000 experiments where we randomly sampled from these populations different sample sizes (n = 20, n = 50, n = 100, n = 200, n = 1000, n = 2000) and computed the sample effect size estimate (standardized effect size, Cohen’s d) for the case–control difference. These histograms (gray) show how variable the sample effect size estimates are (black lines show 95% confidence intervals) relative to the true population effect size (green line). Visually, it is quite apparent how small sample sizes (e.g., n = 20) have wildly varying sample effect size estimates and that this variability is consistent irrespective of what the true population effect size is. Overlaid on each gray histogram are red histograms that show the distribution of sample effect size estimates where the hypothesis test (e.g., independent samples t-test) passes statistical significance at p < 0.05. The rightward shift in this red distribution relative to the true population effect size (green line) illustrates the phenomenon of effect size inflation. The problem is much more pronounced at small sample sizes and when true population effects are smaller. We then computed what is the average effect size inflation for this red distribution and plotted this average effect size inflation as a percentage increase relative to the true population effect in (f). Each line in panel f refers to simulations with different sample sizes. This plot directly quantifies the degree of effect size inflation across a range of true population effects and across a range of sample sizes. The code for implementing and reproducing these simulations is available at https://github.com/mvlombardo/effectsizesim
Fig. 3 Simulation of sample effect size estimates at different sample sizes and across a range of true population effects for a hypothetical case–control study. In this simulation we set the population effect size to a range of different values, from very small (e.g., d = 0.1) to very large (e.g., d > 1.0) (panels a–e show simulation results when effect size ranges from d = 0.1 to d = 0.9 in steps of 0.2). We then simulated data from two populations (cases and controls), each with n = 10,000,000, that had a case–control difference at these population effect sizes. Next, we simulated 10,000 experiments where we randomly sampled from these populations different sample sizes (n = 20, n = 50, n = 100, n = 200, n = 1000, n = 2000) and computed the sample effect size estimate (standardized effect size, Cohen’s d) for the case–control difference. These histograms (gray) show how variable the sample effect size estimates are (black lines show 95% confidence intervals) relative to the true population effect size (green line). Visually, it is quite apparent how small sample sizes (e.g., n = 20) have wildly varying sample effect size estimates and that this variability is consistent irrespective of what the true population effect size is. Overlaid on each gray histogram are red histograms that show the distribution of sample effect size estimates where the hypothesis test (e.g., independent samples t-test) passes statistical significance at p < 0.05. The rightward shift in this red distribution relative to the true population effect size (green line) illustrates the phenomenon of effect size inflation. The problem is much more pronounced at small sample sizes and when true population effects are smaller. We then computed what is the average effect size inflation for this red distribution and plotted this average effect size inflation as a percentage increase relative to the true population effect in (f). Each line in panel f refers to simulations with different sample sizes. This plot directly quantifies the degree of effect size inflation across a range of true population effects and across a range of sample sizes. The code for implementing and reproducing these simulations is available at https://github.com/mvlombardo/effectsizesimCC BY · PubMed Central
Fig. 4 Simulation showing sampling variability and bias of enrichment of specific strata in small sample size studies. In this simulation we generated a control population (n = 1,000,000) with a mean of 0 and a standard deviation of 1 on a hypothetical dependent variable (DV). We then generated an autism population (n = 1,000,000) with 5 different autism subtypes each with a prevalence of 20% (e.g., n = 200,000 for each subtype). These subtypes vary from the control population in effect size in units of 0.5 standard deviations, ranging from −1 to 1. This was done to simulate heterogeneity in the autism population that is reflective of very different types of effects. For example, the autism subtype 5 shows a pronounced increased response on the DV, whereas autism subtype 1 shows a pronounced decreased response on the DV. Across 10,000 simulated experiments, we then randomly sampled from the autism population sample sizes of n = 20, n = 200, and n = 2000, and computed the sample prevalence of each autism subtype. The ideal result without any bias would be sample prevalence rates of around 20% for each subtype. This 20% sample prevalence is approached at n = 2000, and to some extent at n = 200. However, small sample sizes such as n = 20 shows large variability in sample prevalence rates of the subtypes and this can markedly bias the results of a case–control comparison. The code for implementing and reproducing these simulations is available at https://github.com/mvlombardo/effectsizesim
Fig. 4 Simulation showing sampling variability and bias of enrichment of specific strata in small sample size studies. In this simulation we generated a control population (n = 1,000,000) with a mean of 0 and a standard deviation of 1 on a hypothetical dependent variable (DV). We then generated an autism population (n = 1,000,000) with 5 different autism subtypes each with a prevalence of 20% (e.g., n = 200,000 for each subtype). These subtypes vary from the control population in effect size in units of 0.5 standard deviations, ranging from −1 to 1. This was done to simulate heterogeneity in the autism population that is reflective of very different types of effects. For example, the autism subtype 5 shows a pronounced increased response on the DV, whereas autism subtype 1 shows a pronounced decreased response on the DV. Across 10,000 simulated experiments, we then randomly sampled from the autism population sample sizes of n = 20, n = 200, and n = 2000, and computed the sample prevalence of each autism subtype. The ideal result without any bias would be sample prevalence rates of around 20% for each subtype. This 20% sample prevalence is approached at n = 2000, and to some extent at n = 200. However, small sample sizes such as n = 20 shows large variability in sample prevalence rates of the subtypes and this can markedly bias the results of a case–control comparison. The code for implementing and reproducing these simulations is available at https://github.com/mvlombardo/effectsizesimCC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 111 references

Research works printed but not linked: 1

  1. Loth E, Ahmad J, Mason L, Crawley DV, Hayward HL, San Jose Caceres A, et al. Identifying cross-domain cognitive subtypes among children, adolescents, and adults with autism spectrum disorders. San Francisco: International Society for Autism Research; 2017.

Cited by

See all 546

Metadata

Authors
Lombardo, M. V., Lai, M.-C., & Baron-Cohen, S.
Affiliations
  • University of Cambridge (GB)
  • Centre for Addiction and Mental Health (CA)
Corresponding author
Lombardo, Michael V.
Venue
Molecular Psychiatry
Published
2019-01-07
Volume
24
Issue
10
Pages
1435-1450
ISSN
  • 1359-4184 (ISSN-L)
  • 1476-5578
OpenAlex
W2951588804
Cited by
537
References
111
Citations by year
2026: 75 · 2025: 65 · 2024: 84 · 2023: 77 · 2022: 75 · 2021: 74 · 2020: 64 · 2019: 20
FWCI
7.38
Type
journal article
Language
en
Open access
Licence
  • cc-by
  • creativecommons.org/licenses/by/4.0 (tdm)
Locations
  • Molecular Psychiatry (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Apollo (University of Cambridge) (submittedVersion, cc-by)
  • Apollo (cc-by)
Funders
EC | European Research Council (755816)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Autism Spectrum Disorder Research
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Genetics and Neurodevelopmental Disorders
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Mental Health Research Topics
Keywords
autism spectrum disorder, autism heterogeneity, big data, dimensional versus categorical models, longitudinal development, precision medicine, person-centred support, sample size
MeSH terms
Autism Spectrum Disorder, Big Data, Humans, Longitudinal Studies, Case-Control Studies, Genetic Heterogeneity

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Big data approaches to decomposing heterogeneity across the autism spectrum. https://app.cochat.ai/papers/10.1038/s41380-018-0321-0
The paper
Lombardo, M. V., Lai, M.-C., & Baron-Cohen, S. (2019). Big data approaches to decomposing heterogeneity across the autism spectrum. Molecular Psychiatry. https://doi.org/10.1038/s41380-018-0321-0

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.