id
stringlengths 11
17
| article_id
stringlengths 8
11
| path
stringlengths 11
60
| section_title
stringlengths 1
1.33k
| educational_score
float64 0
5.16
| domain
stringclasses 4
values | document_type
stringclasses 5
values | domain_scores
listlengths 0
3
| document_type_scores
listlengths 0
4
| text
stringlengths 1
110k
| authors
listlengths 0
8.02k
| article_url
stringlengths 3
63
| license_type
stringclasses 1
value | license_url
stringclasses 15
values | language
stringclasses 45
values | language_score
float64 0
1
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
PMC11276461_p19
|
PMC11276461
|
sec[2]/sec[2]/p[0]
|
3.3. Subgroup Analysis: Subjects Residing in Rural Areas
| 4.085938 |
biomedical
|
Study
|
[
0.99853515625,
0.0005249977111816406,
0.0008130073547363281
] |
[
0.99951171875,
0.0001819133758544922,
0.00019669532775878906,
0.00004380941390991211
] |
When considering subjects residing in rural areas, TMH utilization was significantly associated with age ( p < 0.001), sex ( p < 0.01), race ( p < 0.001), and primary insurance ( p = 0.01). Compared to the non-TMH cohort, the TMH cohort had lower odds of falling into the age groups 50–64 and over 65 when contrasted with the age group 18–34, with odds ratios of 0.54 (95% CI: 0.41–0.70) and 0.52 (95% CI: 0.36–0.74), respectively. Moreover, the TMH cohort had a higher proportion of females (71.35% vs. 61.81%), a higher proportion of White/Caucasian subjects (49.59% vs. 36.40%), and a lower proportion of Black/African American subjects (48.21% vs. 61.89%). In addition, the TMH cohort exhibited higher odds of using other insurance than Medicare, with ORs of 1.97 (95% CI: 1.25–3.09). Furthermore, the TMH cohort had higher proportions of subjects with household incomes of $42,000–$50,000 (37.79% vs. 35.70%) and $50,000 (12.82% vs. 10.76%), but a lower proportion with incomes less than $42,000 (49.39% vs. 53.54%) ( Table 3 ).
|
[
"Yunxi Zhang",
"Lincy S. Lal",
"Yueh-Yun Lin",
"J. Michael Swint",
"Ying Zhang",
"Richard L. Summers",
"Barbara F. Jones",
"Saurabh Chandra",
"Mark E. Ladner"
] |
https://doi.org/10.3390/ijerph21070819
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999995 |
PMC11276461_p20
|
PMC11276461
|
sec[2]/sec[2]/p[1]
|
3.3. Subgroup Analysis: Subjects Residing in Rural Areas
| 4.089844 |
biomedical
|
Study
|
[
0.9990234375,
0.0007038116455078125,
0.0004379749298095703
] |
[
0.99951171875,
0.0002605915069580078,
0.0002472400665283203,
0.00006473064422607422
] |
Regarding the unadjusted HCRU and medical expenditures, the TMH cohort residing in rural areas had significantly more mental and behavioral health-related outpatient visits (mean (SD): 0.39 (0.39) vs. 0.11 (0.28) PPPM; p < 0.001), inpatient admissions (mean (SD): 0.0020 (0.01) vs. 0.0019 (0.02) PPPM; p = 0.01), and medical expenditures (mean (SD): $26.71 (30.30) vs. $11.22 (37.31) PPPM; p < 0.001), but lower all-cause medical expenditures (mean (SD): $122.68 (167.04) vs. $152.70 (227.91) PPPM; p = 0.002) .
|
[
"Yunxi Zhang",
"Lincy S. Lal",
"Yueh-Yun Lin",
"J. Michael Swint",
"Ying Zhang",
"Richard L. Summers",
"Barbara F. Jones",
"Saurabh Chandra",
"Mark E. Ladner"
] |
https://doi.org/10.3390/ijerph21070819
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276461_p21
|
PMC11276461
|
sec[2]/sec[2]/p[2]
|
3.3. Subgroup Analysis: Subjects Residing in Rural Areas
| 3.984375 |
biomedical
|
Study
|
[
0.99853515625,
0.0005469322204589844,
0.0007348060607910156
] |
[
0.99951171875,
0.00026988983154296875,
0.00023818016052246094,
0.00004780292510986328
] |
After adjusting for all sociodemographic factors, TMH utilization among subjects residing in rural areas was estimated to be associated with a 205% increase in mental and behavioral health-related outpatient visits but a 19% decrease in all-cause medical expenditures (all p < 0.001) ( Table 4 ).
|
[
"Yunxi Zhang",
"Lincy S. Lal",
"Yueh-Yun Lin",
"J. Michael Swint",
"Ying Zhang",
"Richard L. Summers",
"Barbara F. Jones",
"Saurabh Chandra",
"Mark E. Ladner"
] |
https://doi.org/10.3390/ijerph21070819
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276461_p22
|
PMC11276461
|
sec[3]/sec[0]/p[0]
|
4.1. Principal Results
| 2.244141 |
biomedical
|
Study
|
[
0.9873046875,
0.0028057098388671875,
0.01000213623046875
] |
[
0.9951171875,
0.004276275634765625,
0.00026869773864746094,
0.0002872943878173828
] |
Our findings from this study shed light on the sociodemographic characteristics, HCRU, and medical expenditures associated with the utilization of TMH services at a medical center in Mississippi during the COVID-19 pandemic.
|
[
"Yunxi Zhang",
"Lincy S. Lal",
"Yueh-Yun Lin",
"J. Michael Swint",
"Ying Zhang",
"Richard L. Summers",
"Barbara F. Jones",
"Saurabh Chandra",
"Mark E. Ladner"
] |
https://doi.org/10.3390/ijerph21070819
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276461_p23
|
PMC11276461
|
sec[3]/sec[0]/p[1]
|
4.1. Principal Results
| 4.171875 |
biomedical
|
Study
|
[
0.9970703125,
0.0008730888366699219,
0.0018482208251953125
] |
[
0.998046875,
0.0003867149353027344,
0.0017232894897460938,
0.00006002187728881836
] |
Significant sociodemographic disparities were identified between the TMH and non-TMH cohorts. The TMH cohort had a higher proportion of younger subjects and females, suggesting the appeal and accessibility of TMH services to these groups, which aligns with the increasing acceptance and utilization of telehealth among these populations . Moreover, a higher proportion of White/Caucasian subjects in the TMH cohort indicates the potential accessibility of TMH services within this racial group, consistent with studies indicating lower technology usage for health management among older racial minorities . Furthermore, the higher proportion of subjects residing in rural areas in the TMH cohort demonstrates the crucial role of TMH services in addressing mental health needs among rural populations and its potential to overcome geographical barriers and improve mental healthcare access in underserved rural communities . The primary insurance disparities between TMH and non-TMH cohorts may reflect telehealth business models and insurance coverage policies. Additionally, the TMH cohort included a higher proportion of subjects with household incomes greater than $50,000, implying better access to TMH services in this group, potentially due to factors such as technology availability, insurance coverage, or financial resources. These findings highlight the importance of addressing sociodemographic disparities to achieve equitable access to TMH services, particularly among underserved populations, while considering digital health equity .
|
[
"Yunxi Zhang",
"Lincy S. Lal",
"Yueh-Yun Lin",
"J. Michael Swint",
"Ying Zhang",
"Richard L. Summers",
"Barbara F. Jones",
"Saurabh Chandra",
"Mark E. Ladner"
] |
https://doi.org/10.3390/ijerph21070819
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999999 |
PMC11276461_p24
|
PMC11276461
|
sec[3]/sec[0]/p[2]
|
4.1. Principal Results
| 4.167969 |
biomedical
|
Study
|
[
0.99755859375,
0.0014066696166992188,
0.0009570121765136719
] |
[
0.99853515625,
0.0003693103790283203,
0.0010957717895507812,
0.00009226799011230469
] |
In terms of HCRU and medical expenditures, the TMH cohort exhibited significantly higher mental and behavioral health-related outpatient visits, inpatient admissions, ED visits, and medical expenditures compared to the non-TMH cohort while experiencing decreased all-cause medical expenditures. After adjusting for sociodemographic factors, TMH utilization remained significantly associated with increased mental and behavioral health-related outpatient visits and medical expenditures. These findings suggest the vital role of TMH services in enhancing access to mental healthcare. Interestingly, TMH utilization was also associated with a decrease in all-cause medical expenditures. The improved access to mental healthcare through TMH services could potentially enhance mental and behavioral health and lifestyles, leading to better health status and consequently reducing overall medical expenditures . These results are consistent with previous studies showing cost savings associated with outpatient behavioral health treatment among populations covered by commercial insurance and those diagnosed with cancers . The observed reduction in all-cause medical expenditures underscores the potential economic benefits of TMH services, emphasizing the importance of addressing mental health needs to achieve better health outcomes and cost-efficiency. Future research may explore opportunities for integrating TMH and primary care services .
|
[
"Yunxi Zhang",
"Lincy S. Lal",
"Yueh-Yun Lin",
"J. Michael Swint",
"Ying Zhang",
"Richard L. Summers",
"Barbara F. Jones",
"Saurabh Chandra",
"Mark E. Ladner"
] |
https://doi.org/10.3390/ijerph21070819
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276461_p25
|
PMC11276461
|
sec[3]/sec[0]/p[3]
|
4.1. Principal Results
| 4.144531 |
biomedical
|
Study
|
[
0.998046875,
0.00099945068359375,
0.0009136199951171875
] |
[
0.99853515625,
0.00042247772216796875,
0.0011568069458007812,
0.00007021427154541016
] |
Furthermore, our subgroup analysis focusing on subjects residing in rural areas revealed significant associations between TMH utilization and sociodemographic factors, particularly age, sex, race, and primary insurance. This highlights the need for targeted efforts to improve access to TMH services among seniors and underserved racial groups, as well as Medicare and commercially insured populations, promoting equitable utilization of TMH resources . Moreover, the TMH cohort in rural areas demonstrated higher mental and behavioral health-related outpatient visits and medical expenditures but less mental and behavioral health-related inpatient admissions and all-cause medical expenditures. After adjusting for sociodemographic factors, the significant associations between TMH utilization and increased mental and behavioral health-related outpatient visits, as well as decreased all-cause medical expenditures, persisted. These findings further highlight the value of TMH in improving access to mental healthcare and reducing overall healthcare expenditures in rural communities. Efforts should be made to enhance access and utilization of TMH services among underserved rural populations while exploring strategies to improve the delivery and integration of TMH in rural healthcare systems .
|
[
"Yunxi Zhang",
"Lincy S. Lal",
"Yueh-Yun Lin",
"J. Michael Swint",
"Ying Zhang",
"Richard L. Summers",
"Barbara F. Jones",
"Saurabh Chandra",
"Mark E. Ladner"
] |
https://doi.org/10.3390/ijerph21070819
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276461_p26
|
PMC11276461
|
sec[3]/sec[1]/p[0]
|
4.2. Limitations
| 4.113281 |
biomedical
|
Study
|
[
0.99853515625,
0.0010461807250976562,
0.0005640983581542969
] |
[
0.9990234375,
0.0002073049545288086,
0.0005216598510742188,
0.00008541345596313477
] |
Several limitations should be considered when interpreting the findings of this study. First, though we adjusted for sociodemographic factors in this retrospective cohort study, unmeasured factors such as patient comorbidities may influence HCRU and medical expenditure outcomes. Consequently, we cannot definitively address the causal effects of TMH on these outcomes. Future studies employing causal inference methodologies are recommended. Second, the focus on TMH utilization within a single academic medical center in Mississippi during the COVID-19 pandemic may limit the generalizability to other geographic regions or time periods with different healthcare infrastructures, TMH implementation practices, and sociodemographic contexts. Although patients from all UMMC-affiliated sites were included, future studies should consider multiple academic centers or healthcare entities to validate the robustness of findings across different geographic regions and patient populations. Additionally, there is potential for missing data since patients may seek care at multiple institutions. This limitation was mitigated by limiting the study sample to patients who had evidence of regular care visits at UMMC and had evidence of insurance coverage for at least one visit. However, as the study sample consisted of insured patients who regularly seek healthcare from UMMC, it may not represent the entire population of Mississippi, such as those without any healthcare access, thereby limiting generalizability to uninsured populations. Future research should identify and address barriers faced by uninsured populations to provide a more comprehensive understanding of TMH utilization and its impact on HCRU and medical expenditures.
|
[
"Yunxi Zhang",
"Lincy S. Lal",
"Yueh-Yun Lin",
"J. Michael Swint",
"Ying Zhang",
"Richard L. Summers",
"Barbara F. Jones",
"Saurabh Chandra",
"Mark E. Ladner"
] |
https://doi.org/10.3390/ijerph21070819
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276461_p27
|
PMC11276461
|
sec[4]/p[0]
|
5. Conclusions
| 4.148438 |
biomedical
|
Study
|
[
0.99853515625,
0.0009918212890625,
0.0006780624389648438
] |
[
0.9990234375,
0.0002377033233642578,
0.0004553794860839844,
0.00007659196853637695
] |
Our study contributes to the growing body of evidence supporting the importance of addressing sociodemographic disparities and promoting equitable access to TMH services. By investigating TMH utilization in Mississippi throughout the COVID-19 pandemic, our study highlights significant sociodemographic disparities between TMH and non-TMH cohorts, with younger patients, females, those residing in rural areas, and individuals with higher household incomes being more likely to utilize TMH services. A higher proportion of younger patients, females, and White/Caucasian patients in the TMH cohort was observed across all study subjects and within the subgroup of rural residents. These findings collectively suggest the need to ensure equitable access to TMH services across sociodemographic groups. This study also demonstrates the positive impact of TMH on mental and behavioral health-related outpatient visits and medical expenditures, suggesting its value in enhancing access to mental healthcare and reducing overall healthcare expenditures. Moreover, the subgroup analysis focusing on rural areas underscores the crucial role of TMH in addressing mental health needs among rural populations and providing accessible mental healthcare to patients in underserved rural communities.
|
[
"Yunxi Zhang",
"Lincy S. Lal",
"Yueh-Yun Lin",
"J. Michael Swint",
"Ying Zhang",
"Richard L. Summers",
"Barbara F. Jones",
"Saurabh Chandra",
"Mark E. Ladner"
] |
https://doi.org/10.3390/ijerph21070819
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p0
|
PMC11276473
|
sec[0]/p[0]
|
Introduction
| 3.773438 |
biomedical
|
Review
|
[
0.76904296875,
0.00325775146484375,
0.227783203125
] |
[
0.09002685546875,
0.06292724609375,
0.8466796875,
0.0004954338073730469
] |
Externalizing behaviors consist of aggressive, rule-breaking, destructive, and deceitful behaviors and are expressed in approximately a third of school-aged youth . For many children, externalizing behaviors begin during preschool or early school years. However, children are most at risk for externalizing behaviors during adolescence, a developmental period characterized by significant biological and psychological changes that coincide with increased social challenges . Adolescents displaying externalizing behaviors are at increased risk for academic underachievement, family dysfunction, legal system involvement, substance misuse, emotional distress, suicidality, teen pregnancy, and a host of health problems . Consequently, externalizing behaviors are associated with a high individual, family, and societal burden; they constitute a leading reason for referral to mental health services among adolescents and are a main cause of disability worldwide . Given the profound burden of externalizing behaviors, determining the factors that underlie these behaviors is an essential step for advancing prevention and treatment development.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p1
|
PMC11276473
|
sec[0]/sec[0]/p[0]
|
Neurocognition and Externalizing Behaviors
| 4.113281 |
biomedical
|
Study
|
[
0.99755859375,
0.00037479400634765625,
0.0018892288208007812
] |
[
0.9609375,
0.00043463706970214844,
0.0384521484375,
0.00011998414993286133
] |
A substantial body of research indicates that neurocognitive differences underlie the onset and maintenance of externalizing behaviors. Neurocognitive models of externalizing behaviors most commonly highlight the relevance of various components of executive functions, a term that encompasses cognitive processes related to the initiation planning, and regulation of behavior . Meta-analyses of cross-sectional and longitudinal data report medium-sized effects showing deficits in executive functioning, particularly within subcomponents of working memory and inhibition. However, neurocognition is a multicomponent construct and other components, such as learning, memory, and general decision-making, also appear deficient among adolescents showing externalizing behaviors . For example, Thompson and colleagues performed factor analysis on neurocognitive data from a large sample of 9–10-year-olds and found that low scores on the general cognitive ability and learning/memory functions, not just executive functioning, were associated with more externalizing behaviors.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p2
|
PMC11276473
|
sec[0]/sec[0]/p[1]
|
Neurocognition and Externalizing Behaviors
| 4.003906 |
biomedical
|
Review
|
[
0.9990234375,
0.00042366981506347656,
0.0007534027099609375
] |
[
0.48291015625,
0.004669189453125,
0.51220703125,
0.00044989585876464844
] |
Nonetheless, many studies in this area only include one or two subcomponents of neurocognition and few studies explore components of neurocognition beyond executive functioning. Further, much of the extant research relies on analytic methods (multiple regression, factor analysis) that struggle to capture complex, within-person, interactions among components of neurocognition, despite evidence that neurocognitive components work in concert within each person . Part of the challenge lies in finding effective analytical and conceptual frameworks to better describe variability in neurocognition, both within and across individuals .
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p3
|
PMC11276473
|
sec[0]/sec[0]/p[2]
|
Neurocognition and Externalizing Behaviors
| 4.0625 |
biomedical
|
Study
|
[
0.9951171875,
0.00017654895782470703,
0.00478363037109375
] |
[
0.9970703125,
0.0009002685546875,
0.0020427703857421875,
0.000043511390686035156
] |
Latent profile analysis (LPA) is a popular method for studying within-person variation. LPA explains observed variables by grouping participants into latent profiles, i.e. categories of people with similar characteristics. In contrast to factor analysis, LPA does not assume homogenous relationships among variables: it can represent groups of participants who perform neurocognitive tasks in different ways. For example, a recent study by Chaku and colleagues applied LPA to three neurocognitive variables representing different aspects of executive functioning (flanker task for inhibition, list sorting for working memory, and card sorting for cognitive flexibility). They identified four profiles, representing groups with overall high, medium, and low executive function, respectively, as well as one with specifically impaired inhibition. Adolescents belonging to the low executive functioning profile showed more externalizing behaviors than others.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p4
|
PMC11276473
|
sec[0]/sec[0]/p[3]
|
Neurocognition and Externalizing Behaviors
| 3.945313 |
biomedical
|
Study
|
[
0.99560546875,
0.00021648406982421875,
0.003936767578125
] |
[
0.9189453125,
0.061767578125,
0.0190582275390625,
0.00017189979553222656
] |
Despite its advantages for estimating within-person neurocognitive functioning, conventional LPA has a significant limitation: choosing the number of latent profiles involves a tradeoff between model fit and model interpretability . An LPA model with too many latent profiles will fit the data well (i.e. have a larger log-likelihood) by producing a fine-grained description of participants. However, this comes at the cost of interpretability: some profiles will be very similar. By comparison, an LPA model with too few latent profiles will have distinctive profiles and thus be easy to interpret. However, some participants will not be well described by any of the profiles, resulting in poorer model fit (log-likelihood).
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999995 |
PMC11276473_p5
|
PMC11276473
|
sec[0]/sec[0]/p[4]
|
Neurocognition and Externalizing Behaviors
| 3.570313 |
biomedical
|
Study
|
[
0.94970703125,
0.0003573894500732422,
0.04998779296875
] |
[
0.83203125,
0.1356201171875,
0.0318603515625,
0.0002875328063964844
] |
To determine the correct number of latent profiles, researchers fit multiple LPA models, each with a different number of profiles, and compare them using criteria such as model fit, classification uncertainty, a minimum number of people in each profile, statistical tests, and profile interpretability. However, these criteria often disagree, leaving model selection up to the researcher’s judgment . The number of profiles selected dramatically influences the interpretation of an LPA model and, by extension, how we understand the relationship between neurocognition and externalizing.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p6
|
PMC11276473
|
sec[0]/sec[1]/p[0]
|
Using Non-Parameteric Bayesian Inference to Balance the LPA Fit-Interpretability Tradeoff
| 4.078125 |
biomedical
|
Study
|
[
0.99951171875,
0.00021398067474365234,
0.0004286766052246094
] |
[
0.99951171875,
0.00033354759216308594,
0.0001970529556274414,
0.000040531158447265625
] |
In the present study, we sought to develop a non-parametric Bayesian LPA model to address limitations of the conventional LPA model and estimate within-person neurocognitive functioning. The Bayesian framework offers a solution by providing flexible probabilistic modeling approaches that rely less on assumptions and conventions and instead, allow us to use data-derived probability distributions to generate inferences .
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p7
|
PMC11276473
|
sec[0]/sec[1]/p[1]
|
Using Non-Parameteric Bayesian Inference to Balance the LPA Fit-Interpretability Tradeoff
| 4.03125 |
biomedical
|
Study
|
[
0.90478515625,
0.0004050731658935547,
0.09503173828125
] |
[
0.98046875,
0.018218994140625,
0.001148223876953125,
0.00010311603546142578
] |
Bayesian inference uses both the model likelihood and a prior distribution to estimate a model. The prior distribution describes the inherent probability of different model solutions. For LPA, Bayesian inference requires a prior distribution of participants’ latent profile membership. Existing Bayesian LPA methods use a prior that assumes a fixed number of latent profiles . They thus suffer from a similar drawback to conventional LPA: to decide the correct number of latent profiles, one must fit and compare models of different sizes. We avoid this problem by using a distribution called the Dirichlet process as the prior on profile membership. The resulting model (a Dirichlet process mixture) is very flexible: it allows for a large number of latent profiles, but places higher prior probability on model fits with fewer profiles. We developed a novel implementation of the Dirichlet mixture for application to latent profile analysis (DPM-LPA). DPM-LPA has the flexibility to infer a large number of latent profiles if needed, but unlike conventional LPA, DPM-LPA favors simpler model fits that produce a small number of distinct, non-redundant profiles .
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p8
|
PMC11276473
|
sec[0]/sec[1]/p[2]
|
Using Non-Parameteric Bayesian Inference to Balance the LPA Fit-Interpretability Tradeoff
| 4.105469 |
biomedical
|
Study
|
[
0.9990234375,
0.0002963542938232422,
0.0006728172302246094
] |
[
0.99951171875,
0.0001590251922607422,
0.00023639202117919922,
0.000041961669921875
] |
Here, the first goal was to compare DPM-LPA and conventional LPA. We fit both models to neurocognitive data from 11–12-year-olds included in the Adolescent Brain Cognitive Development Study℠ (ABCD Study®). Every two years participants complete the NIH Toolbox task battery and several other neurocognitive tasks during a magnetic resonance imaging session . We compared DPM-LPA and conventional LPA with respect to two metrics: profile similarity and the certainty with which participants were classified into profiles (entropy reduction). We hypothesized that DPM-LPA would perform better on both metrics, producing more interpretable profiles than conventional LPA. In addition, we compared DPM-LPA to conventional LPA and finite Bayesian LPA in a simulation study to determine how accurately these different methods could infer the true number of latent profiles. We hypothesized that DPM-LPA would perform at least as well as the other methods.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p9
|
PMC11276473
|
sec[0]/sec[1]/p[3]
|
Using Non-Parameteric Bayesian Inference to Balance the LPA Fit-Interpretability Tradeoff
| 4.035156 |
biomedical
|
Study
|
[
0.998046875,
0.0002644062042236328,
0.0014486312866210938
] |
[
0.99951171875,
0.00019943714141845703,
0.00013685226440429688,
0.000029742717742919922
] |
The second goal was to validate the DPM-LPA neurocognitive latent profiles in relation to externalizing behaviors. Participants complete measures of mental health symptoms and behaviors , including the Child Behavior Checklist (CBCL) measure of externalizing behaviors and the Positive and Negative Urgency Scales , which measure affective-based impulsive behavior. We used Bayesian methods to investigate the relationship among the neurocognitive profiles discovered by DPM-LPA and externalizing behaviors. We expected that profiles characterized by worse neurocognitive performance would have higher average levels of externalizing but given that we present a novel application of DPM-LPA, we did not have more specific hypotheses about the relationships among neurocognitive profiles and externalizing behaviors.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p10
|
PMC11276473
|
sec[1]/sec[0]/p[0]
|
Participants
| 3.623047 |
biomedical
|
Study
|
[
0.99755859375,
0.0005102157592773438,
0.00200653076171875
] |
[
0.99951171875,
0.0004551410675048828,
0.00012242794036865234,
0.00004285573959350586
] |
Participants were youth included in the ABCD Study Data Release 5.0 with a complete set of neurocognitive variables at the 2-year follow-up (T2; ages 11–12) . We compared participants with a complete set of neurocognitive variables to those with missing neurocognitive data with respect to sex, race/ethnicity, externalizing behaviors, and internalizing behaviors (see Supplemental Material for details). The only difference was with respect to race/ethnicity: the proportion of Black and Hispanic youth was lower among participants with complete neurocognitive data (Black: 11.0% vs. 18.2%, Hispanic: 18.9% vs. 21.4%). All parents or caregivers provided written informed consent and children provided verbal assent for participation in the study . See Garavan and colleagues for the ABCD Study baseline exclusion criteria.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p11
|
PMC11276473
|
sec[1]/sec[1]/p[0]
|
Assessments
| 2.193359 |
biomedical
|
Study
|
[
0.99560546875,
0.001277923583984375,
0.00296783447265625
] |
[
0.9921875,
0.00670623779296875,
0.0007371902465820312,
0.00036978721618652344
] |
The ABCD Study data collection involves biannual visits an extensive evaluation across neurocognition and behavior . We used data from the 2-year and 3-year follow-up visits across all 21 ABCD sites . Figure 2 provides correlations among all variables.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p12
|
PMC11276473
|
sec[1]/sec[1]/sec[0]/p[0]
|
Neurocognitive Data
| 3.816406 |
biomedical
|
Study
|
[
0.99951171875,
0.00020253658294677734,
0.00020575523376464844
] |
[
0.998046875,
0.00112152099609375,
0.0007205009460449219,
0.0000972747802734375
] |
We selected 11 neurocognitive measures ( Table 1a ) that spanned different domains of neurocognition, including executive functions, learning/memory, and general cognition . For all NIH toolbox tasks, we used the uncorrected standardized scores to summarize behavioral performance given that there has been some question about the use of the T-scores for neurocognitive assessment in adolescents .
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p13
|
PMC11276473
|
sec[1]/sec[1]/sec[0]/sec[0]/p[0]
|
Executive functions
| 4.066406 |
biomedical
|
Study
|
[
0.9990234375,
0.0002143383026123047,
0.0007138252258300781
] |
[
0.9990234375,
0.0004649162292480469,
0.0002589225769042969,
0.00004398822784423828
] |
From the NIH Toolbox, we included Flanker (cognitive/attentional control) and Pattern Completion (processing speed). We also included a measure from the Stop Signal Task, which assesses response inhibition . On most trials, the participant’s objective is to quickly give one of two responses depending on the direction of an arrow. However, a subset of trials feature a stop signal, indicating that the response should be withheld. Analysis of response accuracy and response times allows computation of a stop signal reaction time (SSRT) that measures how well each participant can inhibit their responses; we used this measure (reverse coded by multiplying it by –1). Lastly, we included accuracy from the Emotional N-Back Test. The task includes 0-back blocks (respond to a stimulus seen at the beginning of the block) and 2-back blocks (respond to the stimulus seen two trials ago). The task includes both emotionally neutral stimuli (buildings) and emotionally charged stimuli (faces). We used accuracy across stimuli type during 2-back trials as a measure of working memory.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p14
|
PMC11276473
|
sec[1]/sec[1]/sec[0]/sec[1]/p[0]
|
Learning and Memory
| 3.984375 |
biomedical
|
Study
|
[
0.99853515625,
0.0001704692840576172,
0.0012369155883789062
] |
[
0.998046875,
0.0017290115356445312,
0.00030541419982910156,
0.00006538629531860352
] |
From the NIH Toolbox, we included Picture Sequence Memory (visuospatial sequencing/memory). To measure auditory memory, we used the number of correct words in the immediate recall test from the Rey Auditory Verbal Learning Task . Finally, in a second phase of the Emotional N-Back Test, a test of recognition memory was administered where participants are asked to discriminate between stimuli seen in the previous stage and new ones. We used d ′, a discriminability metric from signal detection theory to measure recognition memory.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p15
|
PMC11276473
|
sec[1]/sec[1]/sec[0]/sec[2]/p[0]
|
General Cognition
| 3.845703 |
biomedical
|
Study
|
[
0.9619140625,
0.0003714561462402344,
0.03753662109375
] |
[
0.99365234375,
0.005802154541015625,
0.0005712509155273438,
0.00007581710815429688
] |
From the NIH Toolbox, we included Picture Vocabulary (measuring language skills/verbal intellect) and Oral Reading Recognition (language skills/reading decoding). We also used the Little Man Task which measures visual-spatial processing. On each trial, participants see a rudimentary male figure rotated in various positions and must determine in which hand the figure is holding a briefcase. Performance is summarized by an efficiency score computed as percent correct responses divided by average response time. Lastly, to assess decision-making, we included the Game of Dice Task , which was adapted from the Iowa Gambling Task. On each trial, the participant chooses between a low-risk/low reward gamble and a high-risk/high-reward gamble. High-risk gambles are designed to be disadvantageous, so decision-making performance is measured by the number of low risk choices minus the number of high risk choices.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p16
|
PMC11276473
|
sec[1]/sec[1]/sec[0]/sec[2]/p[1]
|
General Cognition
| 4.027344 |
biomedical
|
Study
|
[
0.99951171875,
0.00018334388732910156,
0.0004036426544189453
] |
[
0.99951171875,
0.00022912025451660156,
0.00027108192443847656,
0.000039577484130859375
] |
This set of neurocognitive measures differed from those examined in previous work using the ABCD data . We analyzed data from when participants were 11–12 years old, whereas similar previous work looked at baseline data (9–10 years old). Over that period, the set of neurocognitive tasks administered to ABCD participants changed: the Card Sort (cognitive flexibility) and List Sort (working memory) tasks were dropped after the baseline assessment, and the Game of Dice Task (decision-making) was added starting at the two-year-follow-up timepoint. In addition, we used behavioral data from the Stop Signal and Emotional N-Back tasks. All neurocognitive variables were standardized by subtracting the mean and dividing by the standard deviation (i.e. z-scored) prior to analysis.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p17
|
PMC11276473
|
sec[1]/sec[1]/sec[1]/p[0]
|
Outcome Variables
| 4.074219 |
biomedical
|
Study
|
[
0.99658203125,
0.00033354759216308594,
0.003143310546875
] |
[
0.99951171875,
0.00016701221466064453,
0.00019979476928710938,
0.000027000904083251953
] |
We used the parent-report CBCL and the UPPS to measure externalizing behaviors ( Table 1b ). The CBCL provides a total externalizing score, and two subscales representing rule- breaking and aggressive behaviors. The CBCL was collected at the two-year-follow-up timepoint (concurrent with the neurocognitive data) and at the three-year-follow-up timepoint (one year after the neurocognitive data). We also performed a sensitivity analysis including the broad internalizing measure from the CBCL , which measures anxiety, depression/withdrawal, and somatic complaints. We opted to use the parent-report for these measured because the youth-report assessment for the ABCD Study protocol changed across the timepoints , especially in the externalizing measures. In addition, research has documented good concordance between youth- and parent-report measures of externalizing disorders, which tend to be observable behaviors . The UPPS scale provides two measures of impulsivity in affective circumstances (UPPS-Negative Urgency: the tendency to act rashly when experiencing negative emotions; UPPS-Positive Urgency: tendency to act rashly when experiencing positive emotion). The UPPS was only collected at the two-year-follow-up timepoint. All outcome variables were standardized by subtracting the mean and dividing by the standard deviation (i.e. z-scored).
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p18
|
PMC11276473
|
sec[1]/sec[2]/sec[0]/p[0]
|
Conventional Latent Profile Analysis (LPA)
| 4.121094 |
biomedical
|
Study
|
[
0.99658203125,
0.00022351741790771484,
0.003185272216796875
] |
[
0.99951171875,
0.00038170814514160156,
0.00014221668243408203,
0.000029981136322021484
] |
We used the tidyLPA and mclust packages in R to perform conventional LPA . We fit models with numbers of profiles ranging from 1 to 20. While it is rare to fit a conventional LPA model with 20 profiles, DPM-LPA can infer a large number of latent profiles if needed to describe the data, so including large conventional LPA models made for a fairer comparison. In mathematical terms, LPA is a finite mixture of multivariate Gaussians with a shared diagonal covariance matrix. The model starts by assuming that there is a fixed number of latent profiles ( T ) and that each individual belongs to a single latent profile. We use z i to represent person i ’s latent profile. For example, if person i belongs to latent profile 3, then z i = 3. Latent profiles can be arbitrarily relabeled, so it is convention to label the profile containing the largest number of participants as profile 1, the next largest profile as profile 2, etc. Each person has a set of m observed indicator variables that we denote x i = x 1, i , x 2, i , …, x m, i . In our study, these indicator variables are performance measures on the various neurocognitive tasks described in Table 1a . Each person’s indicator variables ( x i ) depend on the latent profile that person belongs to ( z i ). In particular, each indicator variable ( x j,i ) is assumed to have a normal distribution with a mean ( μ j,t ) that varies based on latent profile:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p19
|
PMC11276473
|
sec[1]/sec[2]/sec[0]/p[1]
|
Conventional Latent Profile Analysis (LPA)
| 2.591797 |
biomedical
|
Study
|
[
0.55126953125,
0.0007143020629882812,
0.447998046875
] |
[
0.57958984375,
0.4189453125,
0.0008869171142578125,
0.0005526542663574219
] |
Each profile is thus defined by its vector of means ( μ t = μ 1, t , μ 2, t , …, μ m,t ). Because all of our indicator variables are standardized (i.e. they have mean 0 and variance 1), it means that if μ j,t > 0 then people in profile t tend to have above average values of indicator variable x j , i , while if μ j,t < 0 they have below average values.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p20
|
PMC11276473
|
sec[1]/sec[2]/sec[0]/p[2]
|
Conventional Latent Profile Analysis (LPA)
| 4.058594 |
biomedical
|
Study
|
[
0.9755859375,
0.0003745555877685547,
0.0238800048828125
] |
[
0.98974609375,
0.009979248046875,
0.00039958953857421875,
0.00008660554885864258
] |
ξ j represents the precision, i.e. inverse variance, of indicator variable j : var ( x j , i ) = 1 ξ j \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\mathrm{var}}({x_{j,i}}) = {\textstyle{1 \over {{\xi _j}}}}\] \end{document} . Thus, a variable with high precision has low variance, and vice versa (parameterization in terms of precision makes computations more convenient). We assume that precision (variance) does not differ across latent profiles. This is largely a pragmatic assumption: if precision did vary across profiles, it would be difficult to estimate it for smaller profiles (those with fewer people). We also assume that the indicator variables ( x j,i ) do not have any covariance with each other, i.e., they are independent given latent profile membership ( z i ). This is important for making latent profile analysis easy to interpret: one only needs to know a participant’s latent profile to know the distribution of any of their indicator variables ( x j,i ). The model has the following parameters that must be inferred:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p21
|
PMC11276473
|
sec[1]/sec[2]/sec[0]/p[3]
|
Conventional Latent Profile Analysis (LPA)
| 3.21875 |
biomedical
|
Study
|
[
0.9580078125,
0.0005092620849609375,
0.0416259765625
] |
[
0.7392578125,
0.259033203125,
0.001018524169921875,
0.0004220008850097656
] |
This conventional LPA model is estimated using the expectation maximization (EM) algorithm to estimate these parameters by maximizing the model likelihood. This also produces estimates of z (profile membership) in the form of probability vectors ( ϕ i ):
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p22
|
PMC11276473
|
sec[1]/sec[2]/sec[0]/p[4]
|
Conventional Latent Profile Analysis (LPA)
| 4.035156 |
biomedical
|
Study
|
[
0.95751953125,
0.0003662109375,
0.04193115234375
] |
[
0.99658203125,
0.0024585723876953125,
0.0009832382202148438,
0.00005435943603515625
] |
The conventional LPA model assumes a fixed number of latent profiles ( T ). Thus to determine how many latent profiles are the data one must fit multiple models with different values of T and determine which best describes the data. We examine three standard criteria for comparing models to choose the correct number of profiles. First, the entropy reduction statistic quantifies how confident each model is about its classification of people into different latent profiles. We describe it in more detail below. A low entropy reduction statistic (less than about 0.8) suggests that the model cannot accurately assign participants to latent profiles . Second, the Akaike Information Criterion and Bayesian Information Criterion measure goodness of fit (log likelihood) penalized by model complexity (number of estimated parameters). Lower values indicate better model fit balanced by model complexity. Finally, the Bootstrap Likelihood Ratio Test provides a test of whether – for each number of profiles – adding an additional profile improves model fit.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p23
|
PMC11276473
|
sec[1]/sec[2]/sec[1]/p[0]
|
Dirichlet Process Mixture Latent Profile Analysis (DPM-LPA)
| 3.892578 |
biomedical
|
Other
|
[
0.87353515625,
0.0008511543273925781,
0.125732421875
] |
[
0.39306640625,
0.599609375,
0.007198333740234375,
0.0003612041473388672
] |
DPM-LPA is a non-parametric Bayesian form of LPA in which the number of latent profiles is not specified beforehand. Conventional LPA finds point estimates of model parameters by maximizing the likelihood of the data. In contrast, Bayesian models are fit by computing a probability distribution over their parameters (including participant-specific variables such as latent profile membership, z i ) rather than a single point estimate. This distribution, the posterior distribution, combines a prior distribution (representing what parameter values are probable before data are observed) and the likelihood via Bayes’ Rule:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p24
|
PMC11276473
|
sec[1]/sec[2]/sec[1]/p[1]
|
Dirichlet Process Mixture Latent Profile Analysis (DPM-LPA)
| 3.193359 |
biomedical
|
Other
|
[
0.90234375,
0.0007123947143554688,
0.0968017578125
] |
[
0.2266845703125,
0.76904296875,
0.003704071044921875,
0.00045013427734375
] |
Thus, in Bayesian statistics, parameter estimates are influenced by both the likelihood (as in conventional LPA) and the prior distribution. The prior distribution often leads to simpler model fits than would be obtained by maximum likelihood estimation .
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p25
|
PMC11276473
|
sec[1]/sec[2]/sec[1]/p[2]
|
Dirichlet Process Mixture Latent Profile Analysis (DPM-LPA)
| 3.873047 |
biomedical
|
Study
|
[
0.98583984375,
0.0003466606140136719,
0.01364898681640625
] |
[
0.87451171875,
0.1241455078125,
0.0009851455688476562,
0.0002262592315673828
] |
DPM-LPA uses the same likelihood function as conventional LPA (Equation 1): each person’s vector of observed indicator variables ( x i ) follows a normal distribution with a mean vector ( μ ) that depends on the latent profile that person belongs to ( z i ) and with precisions ( ξ ) that are shared across profiles.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p26
|
PMC11276473
|
sec[1]/sec[2]/sec[1]/p[3]
|
Dirichlet Process Mixture Latent Profile Analysis (DPM-LPA)
| 1.96582 |
biomedical
|
Other
|
[
0.69482421875,
0.0013179779052734375,
0.303955078125
] |
[
0.30712890625,
0.689453125,
0.0025005340576171875,
0.0009365081787109375
] |
The prior distribution for DPM-LPA can be broken into two parts, the prior over latent profile membership, and the prior over means/precisions:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p27
|
PMC11276473
|
sec[1]/sec[2]/sec[1]/p[4]
|
Dirichlet Process Mixture Latent Profile Analysis (DPM-LPA)
| 1.87207 |
other
|
Other
|
[
0.281494140625,
0.0012121200561523438,
0.71728515625
] |
[
0.07659912109375,
0.921875,
0.0010585784912109375,
0.0006208419799804688
] |
The key to DPM-LPA is its prior on profile membership ( z i ), which is a distribution called the Dirichlet process . Using the Pólya Urn scheme representation , we can write it as:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999993 |
PMC11276473_p28
|
PMC11276473
|
sec[1]/sec[2]/sec[1]/p[5]
|
Dirichlet Process Mixture Latent Profile Analysis (DPM-LPA)
| 3.630859 |
biomedical
|
Study
|
[
0.85546875,
0.00047779083251953125,
0.1441650390625
] |
[
0.873046875,
0.1253662109375,
0.0011816024780273438,
0.00024235248565673828
] |
where α is a concentration parameter . This Dirichlet process prior distribution does not have a hard upper limit on the number of latent profiles. However, it favors simpler LPA solutions. The prior on z 1: n is high when participants are concentrated in a few latent profiles and low when they are spread across several large ones, with this preference varying as a function of α . A more complex model fit, with participants assigned to a greater number of latent profiles, will only be favored (have a higher posterior probability) if the resulting increase in likelihood by adding those profiles outweighs the decrease in the prior distribution.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p29
|
PMC11276473
|
sec[1]/sec[2]/sec[1]/p[6]
|
Dirichlet Process Mixture Latent Profile Analysis (DPM-LPA)
| 1.672852 |
other
|
Other
|
[
0.327392578125,
0.0023021697998046875,
0.67041015625
] |
[
0.01334381103515625,
0.9853515625,
0.000827789306640625,
0.00044655799865722656
] |
Means/precisions have a standard form of prior (called a conjugate prior) that keeps computations simple:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p30
|
PMC11276473
|
sec[1]/sec[2]/sec[1]/p[7]
|
Dirichlet Process Mixture Latent Profile Analysis (DPM-LPA)
| 1.755859 |
biomedical
|
Other
|
[
0.5888671875,
0.0027751922607421875,
0.408447265625
] |
[
0.266845703125,
0.7275390625,
0.003772735595703125,
0.0018281936645507812
] |
where T represents the number of distinct latent profiles implied by z 1: n .
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999995 |
PMC11276473_p31
|
PMC11276473
|
sec[1]/sec[2]/sec[1]/p[8]
|
Dirichlet Process Mixture Latent Profile Analysis (DPM-LPA)
| 3.714844 |
biomedical
|
Study
|
[
0.95654296875,
0.0002968311309814453,
0.042938232421875
] |
[
0.97705078125,
0.0223541259765625,
0.00043082237243652344,
0.00012105703353881836
] |
DPM-LPA is complex enough that its posterior distribution must be approximated. Markov chain Monte Carlo is one common approximation technique that produces a set of random samples from the posterior distribution . However, this is computationally expensive and thus not attractive for large datasets. Instead, we used an approximation method called mean field variational Bayes . We implemented this algorithm and supporting code in Python. Once the DPM-LPA model has been fit, it is easy to determine how many latent profiles describe the data. Most of the latent profiles estimated by the model will be empty, i.e. the total estimated probability of any participants belonging to them will be approximately zero. We used the simple rule that if a profile was not estimated as maximally probable for at least one participant, that profile was considered to be empty and excluded from further analysis and discussion.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p32
|
PMC11276473
|
sec[1]/sec[2]/sec[2]/p[0]
|
Entropy Reduction Statistic
| 3.908203 |
biomedical
|
Study
|
[
0.9482421875,
0.0003857612609863281,
0.05157470703125
] |
[
0.99755859375,
0.002124786376953125,
0.000362396240234375,
0.00005173683166503906
] |
We used a standard entropy reduction statistic to compare how confidently the conventional LPA and DPM-LPA models classify people into latent profiles. For each participant, the model estimates the probability of that person belonging to each latent profile: we call this ϕ i . For example, if for a certain participant ϕ i = [0.9, 0.07, 0.03] then the model estimates that the probability of this person belonging to profile 1 is 90%, to profile 2 is 7%, and to profile 3 is 3% (in this example there are three latent profiles). Intuitively, the model is fairly confident that the participant is in profile 1. If the probabilities were closer to being equal, say for example ϕ i = [0.4, 0.32, 0.28], then the model would be less confident about which profile the participant belonged to. The entropy of ϕ i quantifies this uncertainty:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p33
|
PMC11276473
|
sec[1]/sec[2]/sec[2]/p[1]
|
Entropy Reduction Statistic
| 2.060547 |
other
|
Study
|
[
0.327392578125,
0.0008640289306640625,
0.671875
] |
[
0.56494140625,
0.432373046875,
0.0016727447509765625,
0.0008139610290527344
] |
Consider the entropy of three hypothetical participants. In the first case ( ϕ 1 = [0.9, 0.07, 0.03]) entropy is low (entropy = 0.39), meaning that the model is fairly certain. In the second case ( ϕ 2 = [0.4, 0.32, 0.28]) entropy is high (entropy = 1.09), meaning that the model is not very certain. Maximum entropy occurs when all the values in ϕ i are equal, meaning intuitively that the model does not have any idea which profile the participant belongs to ( ϕ 3 = [0.33, 0.33, 0.33], entropy = 1.10 in a model with three profiles).
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p34
|
PMC11276473
|
sec[1]/sec[2]/sec[2]/p[2]
|
Entropy Reduction Statistic
| 4.03125 |
biomedical
|
Study
|
[
0.9775390625,
0.00039315223693847656,
0.022216796875
] |
[
0.95654296875,
0.0421142578125,
0.0010004043579101562,
0.00013637542724609375
] |
The proportional entropy reduction statistic quantifies the overall certainty of a model. It compares the total entropy of a model across participants ( − Σ i = 1 n Σ t = 1 T ϕ t , i log ( ϕ t , i ) \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ - \Sigma _{i = {\mathrm{1}}}^n\;\Sigma _{t = 1\;}^T{\phi _{t,i}}\,{\mathrm{log }}({\phi _{t,i}})\] \end{document} ) to the maximum possible entropy ( − n log ( 1 T ) \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ - n\,{\mathrm{log}}({\textstyle{1 \over T}})\] \end{document} ). Subtracting this ratio from 1 gives the following statistic:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999994 |
PMC11276473_p35
|
PMC11276473
|
sec[1]/sec[2]/sec[2]/p[3]
|
Entropy Reduction Statistic
| 1.814453 |
biomedical
|
Other
|
[
0.8095703125,
0.003543853759765625,
0.1866455078125
] |
[
0.0665283203125,
0.9296875,
0.0026264190673828125,
0.0009403228759765625
] |
Values can range from 1 (total confidence in classifying participants) down to 0 (the opposite).
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p36
|
PMC11276473
|
sec[1]/sec[2]/sec[3]/p[0]
|
Profile Distinctiveness (Mahalanobis Distance)
| 3.746094 |
biomedical
|
Study
|
[
0.96337890625,
0.00026535987854003906,
0.036376953125
] |
[
0.986328125,
0.0133514404296875,
0.0004513263702392578,
0.0000788569450378418
] |
Ideally, the latent profiles discovered by LPA should be very distinct from one another. This makes the profiles easier to interpret and reduces uncertainty about participants’ profile membership. We used the Mahalanobis distance to quantify the distinctiveness between each pair of profiles. This is a metric based on the Euclidean distance between two profiles’ mean vectors ( μ 1 and μ 2 ), scaled by the precision, i.e. inverse variance ( ξ ):
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999999 |
PMC11276473_p37
|
PMC11276473
|
sec[1]/sec[2]/sec[3]/p[1]
|
Profile Distinctiveness (Mahalanobis Distance)
| 2.425781 |
biomedical
|
Study
|
[
0.83056640625,
0.0007314682006835938,
0.16845703125
] |
[
0.93994140625,
0.058990478515625,
0.000644683837890625,
0.000286102294921875
] |
A distance of 0 indicates that two profiles are identical, while distance increases as profiles become more and more distinct. For each model (conventional LPA of varying sizes and DPM-LPA) we computed the minimum distance (to find the two most similar profiles in each model) as well as the mean distance across profile pairs.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p38
|
PMC11276473
|
sec[1]/sec[3]/p[0]
|
Simulations
| 3.154297 |
biomedical
|
Study
|
[
0.97802734375,
0.000537872314453125,
0.021453857421875
] |
[
0.998046875,
0.0013036727905273438,
0.00046944618225097656,
0.00007832050323486328
] |
We compared how well DPM-LPA, finite Bayesian LPA, and conventional LPA could infer the true number of latent profiles in simulated data. The simulation parameters were a subset of those used in a study by Tein, Coxe, and Cham examining various criteria for determining the number of profiles in conventional LPA. See the Supplemental Material for further details.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p39
|
PMC11276473
|
sec[1]/sec[4]/sec[0]/p[0]
|
Estimating Profile Means ( μ ( y ) \documentclass[10pt]{article}
\usepackage{wasysym}
\usepackage[substack]{amsmath}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage[mathscr]{eucal}
\usepackage{mathrsfs}
\usepackage{pmc}
\usepackage[Euler]{upgreek}
\pagestyle{empty}
\oddsidemargin -1.0in
\begin{document}
\[{\mu ^{(y)}}\]
\end{document} )
| 4.125 |
biomedical
|
Study
|
[
0.99853515625,
0.00022709369659423828,
0.0011138916015625
] |
[
0.99951171875,
0.00029540061950683594,
0.00018990039825439453,
0.00003409385681152344
] |
We used Bayesian methods to analyze the relationship between DPM-LPA latent profiles and outcome variables. For the sake of clarity, let us refer to the set of indicator variables (in our case neurocognitive measures) as x and the outcome as y . We assume that outcomes (e.g., CBCL externalizing) follow a normal distribution with a mean ( μ ( y ) \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\mu ^{(y)}}\] \end{document} ) that depends on the latent profile but a shared precision ( ξ y ):
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p40
|
PMC11276473
|
sec[1]/sec[4]/sec[0]/p[1]
|
Estimating Profile Means ( μ ( y ) \documentclass[10pt]{article}
\usepackage{wasysym}
\usepackage[substack]{amsmath}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage[mathscr]{eucal}
\usepackage{mathrsfs}
\usepackage{pmc}
\usepackage[Euler]{upgreek}
\pagestyle{empty}
\oddsidemargin -1.0in
\begin{document}
\[{\mu ^{(y)}}\]
\end{document} )
| 2.390625 |
biomedical
|
Study
|
[
0.9404296875,
0.000762939453125,
0.058563232421875
] |
[
0.666015625,
0.332275390625,
0.0009784698486328125,
0.0007691383361816406
] |
We used a normal-gamma conjugate prior distribution:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p41
|
PMC11276473
|
sec[1]/sec[4]/sec[0]/p[2]
|
Estimating Profile Means ( μ ( y ) \documentclass[10pt]{article}
\usepackage{wasysym}
\usepackage[substack]{amsmath}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage[mathscr]{eucal}
\usepackage{mathrsfs}
\usepackage{pmc}
\usepackage[Euler]{upgreek}
\pagestyle{empty}
\oddsidemargin -1.0in
\begin{document}
\[{\mu ^{(y)}}\]
\end{document} )
| 3.933594 |
biomedical
|
Study
|
[
0.9912109375,
0.00025343894958496094,
0.00846099853515625
] |
[
0.99755859375,
0.0023365020751953125,
0.00018918514251708984,
0.000052094459533691406
] |
If each participant’s true latent profile ( z i ) was known, then the formula for estimating the outcome means ( μ ( y ) \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\mu ^{(y)}}\] \end{document} ) for each latent profiles would be standard conjugate prior updates. However, we of course do not know z i , but only have an estimate from the DPM-LPA fit in the form of a probability vector ( ϕ i ). We followed a simple, commonly used procedure and assigned participants to their most probable profiles:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p42
|
PMC11276473
|
sec[1]/sec[4]/sec[0]/p[3]
|
Estimating Profile Means ( μ ( y ) \documentclass[10pt]{article}
\usepackage{wasysym}
\usepackage[substack]{amsmath}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage[mathscr]{eucal}
\usepackage{mathrsfs}
\usepackage{pmc}
\usepackage[Euler]{upgreek}
\pagestyle{empty}
\oddsidemargin -1.0in
\begin{document}
\[{\mu ^{(y)}}\]
\end{document} )
| 3.875 |
biomedical
|
Study
|
[
0.99267578125,
0.0002543926239013672,
0.00693511962890625
] |
[
0.99853515625,
0.0010881423950195312,
0.00017893314361572266,
0.000046253204345703125
] |
We then use these profile estimates ( z ^ i \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\hat z_i}\] \end{document} ) in the ordinary formulas to obtained approximate posterior distributions over each each outcome’s variance and profile means. This also gave us 95% posterior credible intervals over these parameters (the interval between the 2.5th and 97.5th percentiles of the posterior distribution).
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p43
|
PMC11276473
|
sec[1]/sec[4]/sec[1]/p[0]
|
Comparing Profile Means
| 3.976563 |
biomedical
|
Study
|
[
0.9970703125,
0.00019502639770507812,
0.00286865234375
] |
[
0.998046875,
0.001834869384765625,
0.00025916099548339844,
0.000051915645599365234
] |
We used a Bayesian analogue of the analysis of variance (ANOVA) to test if any of the profiles differed from each other, and if they did, we followed up with Bayesian post-hoc tests to determine which profiles had different means. The Bayesian ANOVA compares a null hypothesis H 0 (which assumes that all profiles have equal means) to the alternative hypothesis H 1 that profile means differ:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p44
|
PMC11276473
|
sec[1]/sec[4]/sec[1]/p[1]
|
Comparing Profile Means
| 3.519531 |
biomedical
|
Study
|
[
0.99267578125,
0.00023889541625976562,
0.007328033447265625
] |
[
0.98095703125,
0.01812744140625,
0.0007386207580566406,
0.00013828277587890625
] |
We compared these two hypotheses using the Bayes factor ( BF 10 ), which is defined as the ratio between the evidence for the two hypotheses:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p45
|
PMC11276473
|
sec[1]/sec[4]/sec[1]/p[2]
|
Comparing Profile Means
| 3.435547 |
biomedical
|
Study
|
[
0.947265625,
0.0005311965942382812,
0.05218505859375
] |
[
0.7958984375,
0.2021484375,
0.0015544891357421875,
0.00042819976806640625
] |
Where the evidence for either hypothesis is computed as the likelihood of the observed data ( y 1 , y 2 , …, y n ) averaged across the prior distribution:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p46
|
PMC11276473
|
sec[1]/sec[4]/sec[1]/p[3]
|
Comparing Profile Means
| 2.318359 |
biomedical
|
Other
|
[
0.83837890625,
0.0007572174072265625,
0.16064453125
] |
[
0.451171875,
0.54638671875,
0.0019359588623046875,
0.0007715225219726562
] |
Bayes factors greater than 1 show support for H 1 , Bayes factors less than 1 show support for H 0 , and Bayes factors close to 1 show similar amounts of evidence for both hypotheses.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p47
|
PMC11276473
|
sec[1]/sec[4]/sec[1]/p[4]
|
Comparing Profile Means
| 2.060547 |
biomedical
|
Study
|
[
0.93994140625,
0.0009360313415527344,
0.059112548828125
] |
[
0.75732421875,
0.239501953125,
0.0020904541015625,
0.000995635986328125
] |
Because Bayes factors can vary so widely in size, we also reported them in base-10 logarithmic scale:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999999 |
PMC11276473_p48
|
PMC11276473
|
sec[1]/sec[4]/sec[1]/p[5]
|
Comparing Profile Means
| 3.302734 |
biomedical
|
Study
|
[
0.958984375,
0.00039887428283691406,
0.040374755859375
] |
[
0.8642578125,
0.134521484375,
0.0009579658508300781,
0.0003285408020019531
] |
On this logarithmic scale, positive values indicate support for H 1 , negative values indicate support for H 0 , and 0 indicates no support for either hypothesis. By convention , if –0.5 < log 10 ( BF 10 ) < 0.5 then the result is considered inconclusive: there is not substantial evidence for either hypothesis. We used the BayesFactor package in R for these calculations.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p49
|
PMC11276473
|
sec[1]/sec[4]/sec[1]/p[6]
|
Comparing Profile Means
| 3.839844 |
biomedical
|
Study
|
[
0.95849609375,
0.0003707408905029297,
0.0413818359375
] |
[
0.92626953125,
0.072021484375,
0.0017671585083007812,
0.00015223026275634766
] |
One way to accomplish post-hoc comparisons would be to use a Bayesian analogue of the t-test to test for equality between each pair of profile means. However, this approach suffers from two problems. First, each Bayesian t-test would only use the data from the two latent profiles being compared. This omits data from the other profiles, which is useful for estimating the outcome variance. Also, when there are more than three or four profiles, reporting and interpreting the results of all pairwise tests would be cumbersome.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999999 |
PMC11276473_p50
|
PMC11276473
|
sec[1]/sec[4]/sec[1]/p[7]
|
Comparing Profile Means
| 3.392578 |
biomedical
|
Study
|
[
0.92333984375,
0.00039458274841308594,
0.07635498046875
] |
[
0.91455078125,
0.084228515625,
0.0008606910705566406,
0.00021338462829589844
] |
Instead, we used a different post-hoc analysis method that arrives at a similar result to pairwise comparisons without suffering the same drawbacks . Each pairwise comparison can have two conclusions: the means are either equal or unequal. What results from the post-hoc analysis is a partitioning of the latent profiles into two or more sets, with equal profile means within each set and different means between sets. For a given number of latent profiles, we can enumerate all possible partitions (and thus all possible post-hoc analysis results) through a mathematical algorithm . For example, with three latent profiles the post-hoc analysis could result in the following possible partitions:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p51
|
PMC11276473
|
sec[1]/sec[4]/sec[1]/p[8]
|
Comparing Profile Means
| 1.370117 |
other
|
Other
|
[
0.09722900390625,
0.0011110305786132812,
0.90185546875
] |
[
0.029052734375,
0.96923828125,
0.001068115234375,
0.0006594657897949219
] |
where the notation {1, 2}, {3} indicates that profiles 1 and 2 form a set with one shared mean, while profile 3 has a different mean and thus forms another set by itself.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999999 |
PMC11276473_p52
|
PMC11276473
|
sec[1]/sec[4]/sec[1]/p[9]
|
Comparing Profile Means
| 3.140625 |
biomedical
|
Study
|
[
0.9482421875,
0.0005660057067871094,
0.051025390625
] |
[
0.619140625,
0.37841796875,
0.0019025802612304688,
0.0004572868347167969
] |
Thus, instead of conducting pairwise Bayesian t-tests, we can enumerate all possible partitions of the latent profiles and decide which one describes the data best by comparing their model evidences (the calculation is the same as the Bayesian ANOVA).
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p53
|
PMC11276473
|
sec[1]/sec[4]/sec[2]/p[0]
|
Effect Sizes
| 4.058594 |
biomedical
|
Study
|
[
0.99951171875,
0.00017702579498291016,
0.00033664703369140625
] |
[
0.99951171875,
0.00028514862060546875,
0.000263214111328125,
0.00004887580871582031
] |
We also computed a simple effect size measure for the relationship between the neurocognitive latent profiles and outcome variables. This was an r 2 statistic similar to the one used with linear regression. First, we computed each individual’s predicted value ( y ^ i \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\hat z_i}\] \end{document} ) based on the fitted DPM-LPA model:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p54
|
PMC11276473
|
sec[1]/sec[4]/sec[2]/p[1]
|
Effect Sizes
| 2.376953 |
biomedical
|
Study
|
[
0.9658203125,
0.0009503364562988281,
0.033172607421875
] |
[
0.88623046875,
0.1107177734375,
0.002288818359375,
0.0007696151733398438
] |
Then we computed r 2 as the proportional reduction in error:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999995 |
PMC11276473_p55
|
PMC11276473
|
sec[1]/sec[4]/sec[2]/p[2]
|
Effect Sizes
| 1.905273 |
other
|
Other
|
[
0.483154296875,
0.001468658447265625,
0.51513671875
] |
[
0.19970703125,
0.79736328125,
0.0018749237060546875,
0.000972747802734375
] |
where y ¯ \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\bar y\] \end{document} is the overall mean of y .
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999995 |
PMC11276473_p56
|
PMC11276473
|
sec[2]/sec[0]/p[0]
|
Simulations
| 3.960938 |
biomedical
|
Study
|
[
0.916015625,
0.0004930496215820312,
0.08331298828125
] |
[
0.99560546875,
0.0033435821533203125,
0.0008330345153808594,
0.00006628036499023438
] |
DPM-LPA has similar performance to conventional LPA and superior performance to finite Bayesian LPA . DPM-LPA detected the correct number of latent profiles more often than conventional LPA (using either the BIC or BLRT) at a small sample size of n = 250 (DPM-LPA: 92% correct, conventional LPA with BIC: 82% correct, conventional LPA with BLRT: 72% correct). DPM-LPA was equally accurate to conventional LPA using the BLRT at larger sample sizes , although the BIC was slightly more accurate (BIC: 98% correct, DPM-LPA/BLRT: 92–94% correct). Finite Bayesian LPA was less accurate than either DPM-LPA or conventional LPA, with the difference increasing at larger sample sizes.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p57
|
PMC11276473
|
sec[2]/sec[1]/p[0]
|
Selecting the Number of Profiles in Conventional LPA
| 3.894531 |
biomedical
|
Study
|
[
0.94287109375,
0.0004401206970214844,
0.056549072265625
] |
[
0.99853515625,
0.0007524490356445312,
0.0007638931274414062,
0.000048279762268066406
] |
Commonly used criteria (entropy reduction statistic, AIC, BIC, bootstrap likelihood ratio test) led to contradictory conclusions about the correct number of latent profiles in our sample. The entropy reduction statistic became worse (i.e., decreased) as the number of profiles increased , falling to low levels (around 0.7) for models with more than 4 profiles. This suggested that the correct number of profiles is small. However, AIC, BIC, and the bootstrap likelihood test led to the opposite conclusion. AIC and BIC improved as the number of profiles increased , showing that adding more profiles improved model fit. Similarly, the bootstrap likelihood ratio test reported a statistically significant improvement in model fit ( p < 0.01) with each profile added up to 16 profiles.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999995 |
PMC11276473_p58
|
PMC11276473
|
sec[2]/sec[2]/p[0]
|
Comparing Conventional LPA and DPM-LPA: Entropy Reduction Statistic
| 2.177734 |
biomedical
|
Study
|
[
0.79736328125,
0.0010395050048828125,
0.20166015625
] |
[
0.9609375,
0.037689208984375,
0.0009551048278808594,
0.00031495094299316406
] |
DPM-LPA had a better entropy reduction statistic than any conventional LPA model . In other words, DPM-LPA was more confident than conventional LPA in its classification of participants into latent profiles. This may be due to the relative distinctiveness of DPM-LPA’s profiles, as described above.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999995 |
PMC11276473_p59
|
PMC11276473
|
sec[2]/sec[3]/p[0]
|
Comparing Conventional LPA and DPM-LPA: Profile Distinctiveness (Mahalanobis Distance)
| 3.410156 |
biomedical
|
Study
|
[
0.5595703125,
0.0008072853088378906,
0.439697265625
] |
[
0.9892578125,
0.00980377197265625,
0.0009732246398925781,
0.00014102458953857422
] |
In conventional LPA, smaller models (2, 3, or 4 profiles) had a higher minimum pairwise distance between profiles than the larger ones . In contrast, mean pairwise distance between profiles was lower for the very small conventional LPA models (2 or 3 profiles) than the larger ones . In general, DPM-LPA was comparable or better than conventional LPA models with 5 or more profiles on both measures of distinctiveness (minimum and mean pairwise distance). Small conventional LPA models (2, 3, or 4 profiles) had better minimum distance than DPM-LPA, but worse mean distance. DPM-LPA (which discovered 9 latent profiles) was superior to the 9-profile conventional LPA model on both measures.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p60
|
PMC11276473
|
sec[2]/sec[4]/p[0]
|
Description of Conventional LPA Latent Profiles
| 3.882813 |
biomedical
|
Study
|
[
0.99853515625,
0.00016260147094726562,
0.001499176025390625
] |
[
0.9990234375,
0.0008072853088378906,
0.00017559528350830078,
0.00004017353057861328
] |
We selected the 4-profile model because its entropy reduction statistic was acceptable (0.78); larger models had unacceptably low entropy reduction statistics (0.67–0.72), despite their superior BIC/AIC and support from the BLRT. Based on these criteria, a four profile solution was selected : Profile 1 (56% of the sample) represented average neurocognition, Profile 2 (21% of the sample) represented above average neurocognition, Profile 3 (21% of the sample) represented below average neurocognition, and Profile 4 (2% of the sample) represented below average neurocognition with exceptionally low working memory.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p61
|
PMC11276473
|
sec[2]/sec[5]/p[0]
|
Description of DPM-LPA Latent Profiles
| 4.167969 |
biomedical
|
Study
|
[
0.99853515625,
0.0002944469451904297,
0.0012960433959960938
] |
[
0.99951171875,
0.0002372264862060547,
0.0003783702850341797,
0.000038683414459228516
] |
DPM-LPA identified nine latent profiles. Figure 5 shows plots of the estimated means for each profile. Profile 1 represented average performance across neurocognitive components and encompassed 58% of the sample. Profile 2 (12% of the sample) represented above average performance across most neurocognitive components. Profile 3 (10% of the sample) represented below average performance across most neurocognitive components, with particularly low vocabulary and reading decoding but average response inhibition. Participants in Profile 4 (9% of the sample) had broadly above average neurocognitive function, with particularly good vocabulary and reading decoding. Participants in Profile 5 (5% of the sample) had below average working memory, recognition memory, and response inhibition. Profile 6 (3% of the sample) was characterized by low cognitive/attentional control. Profile 7 (2% of the sample) was characterized by above average processing speed, cognitive/attentional control, and spatial processing. Profiles 8 and 9 each contained 1% of the sample. Profile 8 showed generally poor neurocognitive functioning, while participants in Profile 9 had average performance on most tasks but exceptionally low working memory. Overall, profiles 2, 4, and 7 represented different patterns of above average neurocognitive function, while profiles 3, 5, 6, 8, and 9 represented different patterns of below average neurocognitive function. Importantly, each profile represented a unique distribution of neurocognitive scores.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p62
|
PMC11276473
|
sec[2]/sec[6]/p[0]
|
Neurocognitive Latent Profiles and Psychopathology
| 3.316406 |
biomedical
|
Study
|
[
0.990234375,
0.00037932395935058594,
0.00958251953125
] |
[
0.9990234375,
0.0006394386291503906,
0.00035190582275390625,
0.00005042552947998047
] |
See Figure 6 for plots of estimated profile means for all outcome variables and Table 2 for a summary of results. For comparison with DPM-LPA, we conducted a similar analysis using the 4-profile conventional LPA model (Supplemental Table 2). Results using conventional LPA were broadly similar to the DPM-LPA results reported below, but the DPM-LPA analysis provided a more nuanced description of the relationship between neurocognition and externalizing behaviors.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p63
|
PMC11276473
|
sec[2]/sec[6]/sec[0]/sec[0]/p[0]
|
Externalizing, Rule-Breaking, and Aggression
| 3.882813 |
biomedical
|
Study
|
[
0.8994140625,
0.0007333755493164062,
0.10003662109375
] |
[
0.99853515625,
0.0008473396301269531,
0.0004715919494628906,
0.000053822994232177734
] |
Overall externalizing behaviors at the two-year follow-up differed across latent profiles ( log 10 ( B F 10 ) = 3.51 , B F 10 = 3.22 × 10 3 , r 2 = 0.0090 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\mathrm{lo}}{{\mathrm{g}}_{10}}(B{F_{10}}) = 3.51,\,\,B{F_{10}} = 3.22\,\, \times \,\,{10^3},\,\,{r^2} = 0.0090\] \end{document} ). Post-hoc tests showed that profiles 6 and 7 had equal means to profile 1 (average neurocognition), profiles 2 and 4 were related to lower levels of externalizing, while profiles 3, 5, 8, and 9 were related to higher levels of externalizing.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p64
|
PMC11276473
|
sec[2]/sec[6]/sec[0]/sec[0]/p[1]
|
Externalizing, Rule-Breaking, and Aggression
| 4.132813 |
biomedical
|
Study
|
[
0.98095703125,
0.0004489421844482422,
0.01837158203125
] |
[
0.99951171875,
0.0003981590270996094,
0.00024056434631347656,
0.0000368952751159668
] |
The rule-breaking and aggression subscales also differed across latent profiles (rule-breaking: log 10 ( BF 10 ) = 11.59, BF 10 = 3.87 × 10 11 , r 2 = 0.0167; aggression: log 10 ( BF 10 ) = 1.22, BF 10 = 16.70, r 2 = 0.0070). However, the relationship with neurocognition was much stronger for rule-breaking than aggression . Post-hoc tests for rule-breaking revealed that profile 7 (posterior mean estimate: μ ^ 7 ( y ) = − 0.05 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _7^{(y)} = - 0.05\] \end{document} ) had the same mean as profile 1 (average neurocognition, μ ^ 1 ( y ) = − 0.06 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _1^{(y)} = - 0.06\] \end{document} ), profiles 2 ( μ ^ 2 ( y ) = − 0.19 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _2^{(y)} = - 0.19\] \end{document} ) and 4 ( μ ^ 4 ( y ) = − 0.19 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _4^{(y)} = - 0.19\] \end{document} ) had a lower mean level of rule-breaking, and profiles 3 ( μ ^ 3 ( y ) = 0.19 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _3^{(y)} = 0.19\] \end{document} ), 5 ( μ ^ 5 ( y ) = 0.14 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _5^{(y)} = 0.14\] \end{document} ), 6 ( μ ^ 6 ( y ) = 0.08 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _6^{(y)} = 0.08\] \end{document} ), 8 ( μ ^ 8 ( y ) = 0.25 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _8^{(y)} = 0.25\] \end{document} ), and 9 ( μ ^ 9 ( y ) = 0.34 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _9^{(y)} = 0.34\] \end{document} ) showed higher mean levels of rule-breaking. For aggression, profiles 6 ( μ ^ 6 ( y ) = 0.01 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _6^{(y)} = 0.01\] \end{document} ) and 7 ( μ ^ 7 ( y ) = − 0.05 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _7^{(y)} = - 0.05\] \end{document} ) had the same mean as profile 1 ( μ ^ 1 ( y ) = − 0.03 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _1^{(y)} = - 0.03\] \end{document} ), profiles 2 ( μ ^ 2 ( y ) = 0.13 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _2^{(y)} = 0.13\] \end{document} ) and 4 ( μ ^ 4 ( y ) = − 0.13 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _4^{(y)} = - 0.13\] \end{document} ) showed lower mean levels of aggression, profiles 3 ( μ ^ 3 ( y ) = 0.06 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _3^{(y)} = 0.06\] \end{document} ), 5 ( μ ^ 5 ( y ) = 0.13 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _5^{(y)} = 0.13\] \end{document} ), and 8 ( μ ^ 8 ( y ) = 0.08 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _8^{(y)} = 0.08\] \end{document} ) showed higher mean aggression, and profile 9 ( μ ^ 9 ( y ) = 0.44 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _9^{(y)} = 0.44\] \end{document} ) had the highest mean level of aggression.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999994 |
PMC11276473_p65
|
PMC11276473
|
sec[2]/sec[6]/sec[0]/sec[0]/p[2]
|
Externalizing, Rule-Breaking, and Aggression
| 3.939453 |
biomedical
|
Study
|
[
0.88427734375,
0.0008549690246582031,
0.11474609375
] |
[
0.9990234375,
0.0006866455078125,
0.0004696846008300781,
0.00005620718002319336
] |
At the three-year follow-up, the latent profiles were still related to overall externalizing behaviors (log 10 ( BF 10 ) = 2.11, BF 10 = 1.30 × 10 2 , r 2 = 0.0079) and rule-breaking (log 10 ( BF 10 ) = 8.01, BF 10 = 1.03 × 10 8 , r 2 = 0.0140), but not aggression (log 10 ( BF 10 ) = –0.85, BF 10 = 0.14). Post-hoc results for rule-breaking were the same as those for the two-year follow-up described above. However, the post-hoc results for overall externalizing changed at the three-year follow-up. Profiles 1 ( μ ^ 1 ( y ) = − 0.02 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _1^{(y)} = - 0.02\] \end{document} ) and 3 ( μ ^ 3 ( y ) = 0.05 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _3^{(y)} = 0.05\] \end{document} ) shared the same mean externalizing, profiles 2 ( μ ^ 2 ( y ) = − 0.12 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _2^{(y)} = - 0.12\] \end{document} ), 4 ( μ ^ 4 ( y ) = − 0.08 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _4^{(y)} = - 0.08\] \end{document} ), and 7 ( μ ^ 7 ( y ) = − 0.07 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _7^{(y)} = - 0.07\] \end{document} ) had a lower mean externalizing, and profiles 5 ( μ ^ 5 ( y ) = 0.26 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _5^{(y)} = 0.26\] \end{document} ), 6 ( μ ^ 6 ( y ) = 0.13 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _6^{(y)} = 0.13\] \end{document} ), 8 ( μ ^ 8 ( y ) = 0.19 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _8^{(y)} = 0.19\] \end{document} ), and 9 ( μ ^ 9 ( y ) = 0.28 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _9^{(y)} = 0.28\] \end{document} ) had a higher mean externalizing.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p66
|
PMC11276473
|
sec[2]/sec[6]/sec[0]/sec[1]/p[0]
|
Positive and Negative Urgency
| 4.117188 |
biomedical
|
Study
|
[
0.98583984375,
0.0004813671112060547,
0.0134429931640625
] |
[
0.99951171875,
0.0004208087921142578,
0.0002624988555908203,
0.000038504600524902344
] |
Positive urgency differed across neurocognitive latent profiles (log 10 ( BF 10 ) = 18.87, BF 10 = 7.46 × 10 18 , r 2 = 0.0230). Post-hoc tests showed that profile 7 ( μ ^ 7 ( y ) = − 0.03 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _7^{(y)} = - 0.03\] \end{document} ) had the same positive urgency mean as profile 1 ( μ ^ 1 ( y ) = − 0.06 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _1^{(y)} = - 0.06\] \end{document} ), profile 2 ( μ ^ 2 ( y ) = − 0.21 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _2^{(y)} = - 0.21\] \end{document} ) and profile 4 ( μ ^ 4 ( y ) = − 0.25 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _4^{(y)} = - 0.25\] \end{document} ) had a lower positive urgency mean, profile 3 ( μ ^ 3 ( y ) = 0.14 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _3^{(y)} = 0.14\] \end{document} ) and profile 5 ( μ ^ 5 ( y ) = 0.12 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _5^{(y)} = 0.12\] \end{document} ) had a higher positive urgency mean, while profiles 6 ( μ ^ 6 ( y ) = 0.33 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _6^{(y)} = 0.33\] \end{document} ), 8 ( μ ^ 8 ( y ) = 0.46 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _8^{(y)} = 0.46\] \end{document} ), and 9 ( μ ^ 9 ( y ) = 0.53 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\hat \mu _9^{(y)} = 0.53\] \end{document} ) had the highest positive urgency mean. The Bayes factor testing the relationship between neurocognitive profiles and negative urgency was indecisive (log 10 ( BF 10 ) = –0.47, BF 10 = 0.34, r 2 = 0.0052); if there is a relationship, it is weaker than for positive urgency .
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p67
|
PMC11276473
|
sec[2]/sec[6]/sec[0]/sec[2]/p[0]
|
Interim Summary
| 4.105469 |
biomedical
|
Study
|
[
0.96923828125,
0.0005750656127929688,
0.0300140380859375
] |
[
0.99853515625,
0.0004696846008300781,
0.0009870529174804688,
0.000041425228118896484
] |
Altogether, profiles 2 and 4 seem to capture a group of adolescents typified by above average neurocognitive performance and good behavioral regulation capacities across the two-and-three-year data (consistently had the lowest levels of overall externalizing behaviors, rule-breaking, and aggression). In contrast, profiles 3, 5, 6, 8 and 9 showed below average performance across most neurocognitive measures, and overall had higher than average levels of externalizing behaviors. Profile 3 included adolescents who struggled the most with tasks that rely on vocabulary and reading skills and display higher than average levels of externalizing, rule-breaking, and aggression (although their level of externalizing declined to the same level as profile 1, the average group, at the three-year follow-up). Profile 5 captured adolescents who showed relatively poorer executive functions (working memory and response inhibition), which related to their externalizing behaviors. Adolescents in profile 6 showed a particular pattern of neurocognitive scores that indicates that they found it challenging to engage cognitive/attentional control, which was related to higher scores on rule-breaking and positive urgency but not aggression. Finally, profiles 8 and 9 appear to capture the most severe cases of general neurocognitive problems, particularly in terms of working memory, and these groups showed higher levels externalizing behaviors across measures and time.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p68
|
PMC11276473
|
sec[2]/sec[6]/sec[1]/p[0]
|
Internalizing
| 3.904297 |
biomedical
|
Study
|
[
0.998046875,
0.00032448768615722656,
0.0016965866088867188
] |
[
0.99951171875,
0.0002753734588623047,
0.00017333030700683594,
0.00003647804260253906
] |
Sensitivity analysis showed that the neurocognitive profiles were not related to the CBCL internalizing scale at either the two-year-follow-up (log 10 ( BF 10 ) = –1.57, BF 10 = 0.03) or the three-year-follow-up (log 10 ( BF 10 ) = –0.62, BF 10 = 0.24). This suggest that the differences in neurocognitive patterns within-persons are specific to externalizing-related outcomes .
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p69
|
PMC11276473
|
sec[3]/p[0]
|
Discussion
| 4.085938 |
biomedical
|
Study
|
[
0.998046875,
0.00022745132446289062,
0.0016937255859375
] |
[
0.99951171875,
0.00021398067474365234,
0.00022780895233154297,
0.000028073787689208984
] |
The overarching goal of this work was to develop and test a novel approach to study the relationship between individual differences in neurocognitive functioning and externalizing behaviors in adolescents. Our findings highlight the advantages and power of a novel non-parametric Bayesian approach to LPA, called DPM-LPA, for untangling subgroups of adolescents based on an extensive and diverse set of neurocognitive metrics, even when groups overlap in the patterns of observable externalizing behaviors.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p70
|
PMC11276473
|
sec[3]/sec[0]/p[0]
|
Comparing LPA Methods
| 3.744141 |
biomedical
|
Study
|
[
0.93701171875,
0.0004253387451171875,
0.0626220703125
] |
[
0.9970703125,
0.0018281936645507812,
0.000885009765625,
0.00006020069122314453
] |
Consistent with previous work , different selection criteria gave different answers about the correct number of profiles in conventional LPA . Model fit (measured by AIC and the bootstrap likelihood ratio test) improved by increasing the number of profiles, but this came at the cost of having highly similar, redundant profiles and hence lower classification certainty (measured by the entropy reduction statistic). This illustrates the tradeoff in conventional LPA between model fit and model interpretability.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p71
|
PMC11276473
|
sec[3]/sec[0]/p[1]
|
Comparing LPA Methods
| 4.042969 |
biomedical
|
Study
|
[
0.93994140625,
0.0005030632019042969,
0.059478759765625
] |
[
0.998046875,
0.0012798309326171875,
0.0004930496215820312,
0.000048995018005371094
] |
In contrast, DPM-LPA automatically decided how many latent profiles to include. Compared to conventional LPA models with equal flexibility, DPM-LPA produced less similar profiles and classified participants into profiles with greater certainty . This can be explained by DPM-LPA’s non-parametric Bayesian inference process. Fitting conventional LPA only involves maximizing the likelihood, which rewards the fitting algorithm for using all available latent profiles to classify people even if some of them end up being very similar . However, DPM-LPA takes both the likelihood and prior distribution into account. The prior favors a small number of profiles that actually contain participants, leaving unneeded profiles empty. Thus, DPM-LPA has the flexibility to infer a large number of latent profiles if it needs to, but does not tend to infer redundant, highly similar profiles, in the way that conventional LPA does. Simulations provide further support for DPM-LPA (see Supplemental Material): it compares well with conventional LPA in terms of detecting the correct number of latent profiles, and may be the superior method at small sample sizes ( n = 250).
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p72
|
PMC11276473
|
sec[3]/sec[1]/p[0]
|
Neurocognitive Latent Profiles and Externalizing Behaviors
| 4.085938 |
biomedical
|
Study
|
[
0.99951171875,
0.0002237558364868164,
0.0004849433898925781
] |
[
0.99951171875,
0.0001722574234008789,
0.0003859996795654297,
0.0000400543212890625
] |
In addition to methodological superiority over conventional LPA, DPM-LPA also provides a general approach to generate unique insights into how the same externalizing behaviors can be associated with differences in patterns of neurocognition across subgroups. Broadly speaking, in the present study, the profiles showing more externalizing behaviors were typified by below average neurocognitive functioning, and those characterized by above average neurocognitive performance showed the opposite pattern. However, most importantly, we were able to further pinpoint subgroup-specific patterns reflecting issues with subcomponents of executive functions (working memory, inhibition) and language processing in the latent profiles showing relatively higher levels of externalizing.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999995 |
PMC11276473_p73
|
PMC11276473
|
sec[3]/sec[1]/p[1]
|
Neurocognitive Latent Profiles and Externalizing Behaviors
| 4.121094 |
biomedical
|
Study
|
[
0.99951171875,
0.00024962425231933594,
0.0004067420959472656
] |
[
0.99853515625,
0.0001977682113647461,
0.0009832382202148438,
0.000057756900787353516
] |
The presence of subgroups predominantly showing reduced working memory capacity is in line with the proposal that executive function plays a key role in the development of reduced behavioral regulation during childhood . Notably, we also identified subgroups with equally worse performance on two subcomponents of executive functions, working memory and inhibition (profile 5), and two out of three subgroups showing poorer working memory also presented with worse inhibition (profiles 8 and 9). This pattern suggests a possible interaction between these two subcomponents of executive functions, which converges with prior evidence in adults indicating that cognitive inhibition accounts for part of the variability in age-related working memory . In sum, these findings further highlight the sensitivity of the DPM-LPA for detecting latent profiles based on fine-grained differences in neurocognitive functions, perhaps even providing sufficient sensitivity to disentangle subgroups in which the interaction of different subcomponents of neurocognitive functions could turn out to be a key etiological factor. Future studies could explore this possibility further.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p74
|
PMC11276473
|
sec[3]/sec[1]/p[2]
|
Neurocognitive Latent Profiles and Externalizing Behaviors
| 3.994141 |
biomedical
|
Study
|
[
0.98486328125,
0.0004940032958984375,
0.01447296142578125
] |
[
0.99853515625,
0.0006937980651855469,
0.0007696151733398438,
0.00003987550735473633
] |
Our findings also highlight the existence of a group of adolescents with externalizing proneess and below average general neurocognitive functioning that especially struggle with language development (profile 3). Prior work has shown that language skills play a particularly important role in the development of externalizing behaviors, with one study indicating that having poorer language abilities is a temporally stable within-person predictor of having more externalizing problems later in childhood . The importance of identifying which adolescents present with poor language abilities becomes particularly salient in the context of intervening on externalizing behaviors, as many interventions are delivered verbally and thus place a burden on individuals’ language skills. It may be necessary to support these individuals through alternative strategies that take into consideration their neurocognitive difficulties.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p75
|
PMC11276473
|
sec[3]/sec[1]/p[3]
|
Neurocognitive Latent Profiles and Externalizing Behaviors
| 4.140625 |
biomedical
|
Study
|
[
0.9990234375,
0.00028395652770996094,
0.0007319450378417969
] |
[
0.99951171875,
0.0001424551010131836,
0.0004305839538574219,
0.000044345855712890625
] |
Importantly, our results suggest that the association between neurocognition and externalizing behaviors varies depending on the subgroup of neurocognition and subtype of externalizing behaviors. For example, we saw divergence in the association between variability in neurocognition and engagement in impulsive behavior when feeling positive versus negative emotions. Below average neurocognitive profiles all related to higher positive urgency (albeit to different degrees). There was little evidence of differences in neurocognition related to negative urgency. It is possible that for this age-group, emotion-relevant impulsivity, particularly positive emotions, relates to neurocognitive vulnerability. It also is possible that the association between neurocognitive difficulties and negative urgency emerges at more extreme levels of impulsive behavior . Far more work is needed to examine if neurocognitive associations differentially relate to positive and negative urgency . As another example, we saw a lack of temporal stability in the association between neurocognitive profiles and aggressive behavior measured at timepoints two (age 11–12) and three (age 12–13). This finding is consistent with developmental research which points out that aggression and rule-breaking represent different dimensions within the context of broad externalizing behaviors, each with a separate pattern of (neuro)biological correlates and developmental timeline . While aggression decreases as children age, the tendency to break rules increases over time and into adolescence . Follow-up studies in future waves of the ABCD data collection could try to elucidate the developmental changes in neurocognition and behavioral outcomes.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p76
|
PMC11276473
|
sec[3]/sec[2]/p[0]
|
Limitations
| 3.226563 |
biomedical
|
Study
|
[
0.99755859375,
0.00031447410583496094,
0.001918792724609375
] |
[
0.99853515625,
0.0009160041809082031,
0.00034236907958984375,
0.00006568431854248047
] |
This study had several limitations which can be addressed in future work. First, based on the ABCD two-year follow-up neurocognitive data we find that DPM-LPA is superior to conventional LPA. Examining the robustness of this claim with other datasets, varying in sample type, measures of neurocognition, and the number of variables will be an important next step.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p77
|
PMC11276473
|
sec[3]/sec[2]/p[1]
|
Limitations
| 3.857422 |
biomedical
|
Study
|
[
0.99853515625,
0.00016260147094726562,
0.00128936767578125
] |
[
0.9951171875,
0.0020427703857421875,
0.002559661865234375,
0.00006890296936035156
] |
Second, while the neurocognitive battery used at the two-year-follow-up is quite extensive, it did not tap all neurocognitive functions (e.g., cognitive flexibility) or include tasks with more ecologically-valid stimuli . Therefore, the pattern of neurocognitive functions may shift when considering additional cognitive and task-related factors. Related, adolescence is a time of rapid neurocognitive and behavioral development. Therefore, we may obtain very different neurocognitive latent profiles at an older or younger age. Future work should investigate the stability or change in neurocognitive latent profiles and their relationship to externalizing behaviors.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276473_p78
|
PMC11276473
|
sec[3]/sec[2]/p[2]
|
Limitations
| 1.621094 |
other
|
Other
|
[
0.1724853515625,
0.000865936279296875,
0.82666015625
] |
[
0.058502197265625,
0.93896484375,
0.0019407272338867188,
0.0005803108215332031
] |
Third, our DPM-LPA software is limited in several ways and can be improved upon in later versions. The software can currently only model indicator variables using normal distributions. However, many psychological applications of LPA involve binary or categorical indicators, or indeed a combination of variable types. Implementing greater flexibility in the allowable distributions of indicators would improve the applicability of DPM-LPA .
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276473_p79
|
PMC11276473
|
sec[3]/sec[2]/p[3]
|
Limitations
| 3.875 |
biomedical
|
Study
|
[
0.9716796875,
0.0003020763397216797,
0.0281982421875
] |
[
0.9990234375,
0.0005793571472167969,
0.00023496150970458984,
0.0000355839729309082
] |
Finally, our simulations comparing DPM-LPA to conventional LPA and finite Bayesian LPA were limited by the empirical focus of the current study. Simulations for evaluating statistical methods require substantial computational resources, and are often full studies in and of themselves . We only simulated one possible scenario (10 indicator variables, 5 latent profiles, equal number of participants in each latent profile, profile means separated by 1.5 standard deviations) out of many possible ones. We also only generated 50 simulated data sets in each condition: more replications could improve our estimates of each model/method’s accuracy. Further, detecting the correct number of latent profiles is not the only criterion by which to compare methods: accuracy of assigning participants to the correct latent profiles and accuracy of estimating profile means also are important. Conducting extensive simulations that address these limitations would be a useful direction for future research.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p80
|
PMC11276473
|
sec[4]/p[0]
|
Conclusions
| 4.097656 |
biomedical
|
Study
|
[
0.998046875,
0.00023651123046875,
0.0015544891357421875
] |
[
0.99951171875,
0.00024437904357910156,
0.00035858154296875,
0.0000374913215637207
] |
This paper presents a novel implementation of DPM-LPA, a non-parametric Bayesian approach to conducting LPA, which can detect latent profiles from relatively large amounts of indicator variables and offers a possible solution for dealing with the tradeoff between interpretability and model fit that plagues conventional LPA. We showed that DPM-LPA can be used to better understand how externalizing behaviors seen across individuals relate to differences in neurocognitive subcomponents by detecting latent subgroups of individuals with similar neurocognition. Our study marks a step towards addressing the challenge of finding novel ways to use data on neurocognitive functioning to better describe the individual . Such advances are particularly relevant given that current call for person-centered diagnostics and tailored interventions for individuals showing these costly behaviors.
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276473_p81
|
PMC11276473
|
sec[5]/p[0]
|
Additional File
| 1.016602 |
biomedical
|
Other
|
[
0.64453125,
0.0027484893798828125,
0.352783203125
] |
[
0.0248870849609375,
0.97216796875,
0.0019216537475585938,
0.0008902549743652344
] |
The additional file for this article can be found as follows:
|
[
"Sam Paskewitz",
"Inti A. Brazil",
"Ilker Yildirim",
"Sonia Ruiz",
"Arielle Baskin-Sommers"
] |
https://doi.org/10.5334/cpsy.112
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276485_p0
|
PMC11276485
|
sec[0]/p[0]
|
1. Introduction
| 2.662109 |
biomedical
|
Other
|
[
0.99658203125,
0.0013065338134765625,
0.0020008087158203125
] |
[
0.08148193359375,
0.89697265625,
0.0139923095703125,
0.00774383544921875
] |
One of the first scientific descriptions of hypohidrotic ectodermal dysplasia was provided in 1875 by Charles Darwin, who studied reports on a four-generation Hindu family from India with ten affected male relatives. These studies enabled him to deduce fundamental principles of X-linked recessive inheritance .
|
[
"Stefan J. Rietmann",
"Noëlle Cochet-Faivre",
"Helene Dropsy",
"Vidhya Jagannathan",
"Lucie Chevallier",
"Tosso Leeb"
] |
https://doi.org/10.3390/genes15070854
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276485_p1
|
PMC11276485
|
sec[0]/p[1]
|
1. Introduction
| 5.007813 |
biomedical
|
Study
|
[
0.998046875,
0.0015058517456054688,
0.0006313323974609375
] |
[
0.962890625,
0.0042877197265625,
0.0307159423828125,
0.002216339111328125
] |
The clinical phenotype in humans involves sparse or absent hair, abnormal dentition characterized by partially missing teeth and remaining teeth exhibiting a distinctive pointed morphology, as well as a deficiency in various glands, notably sweat glands, resulting in heat intolerance . This phenotype has been interchangeably termed hypohidrotic ectodermal dysplasia (HED), anhidrotic ectodermal dysplasia, or Christ–Siemens–Touraine syndrome . The vast majority of human HED patients carry loss-of-function variants in the X-chromosomal EDA gene encoding ectodysplasin A . Ectodysplasin A is a homotrimeric type II transmembrane protein with an intracellular N-terminus, a single transmembrane domain, an extracellular short collagen-like domain that mediates triple helix formation and trimerization and a C-terminal signaling domain that has sequence homology to tumor necrosis factor (TNF) . Alternative splicing gives rise to two alternative transcripts from the ~400 kb EDA gene, which encode two protein isoforms termed EDA-A1 and EDA-A2 that differ by the presence or absence of two amino acids in the TNF signaling domain and bind to two different receptors . The physiological functions of the shorter EDA-A2 isoform and its receptor are largely unknown. Expression of the longer EDA-A1 isoform during fetal development prompts the formation of many different ectodermal appendages, such as hair follicles, tooth buds or sweat glands. The signaling cascade involves extracellular proteolytic cleavage of the membrane-bound EDA-A1 by furin proteases to release a paracrine trimeric signaling molecule. The released soluble fragment can bind with its TNF signaling domain to the ectodysplasin A receptor (EDAR) on target cells. Activated EDAR recruits an intracellular adaptor protein termed EDAR associated via death domain (EDARADD), and the complex activates NFκB signaling to modulate the expression of target genes .
|
[
"Stefan J. Rietmann",
"Noëlle Cochet-Faivre",
"Helene Dropsy",
"Vidhya Jagannathan",
"Lucie Chevallier",
"Tosso Leeb"
] |
https://doi.org/10.3390/genes15070854
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276485_p2
|
PMC11276485
|
sec[0]/p[2]
|
1. Introduction
| 4.148438 |
biomedical
|
Study
|
[
0.99951171875,
0.00019669532775878906,
0.00017631053924560547
] |
[
0.91064453125,
0.0406494140625,
0.0479736328125,
0.0008873939514160156
] |
Loss-of-function of EDA , EDAR or EDARADD leads to identical clinical phenotypes in human patients . However, the vast majority of human patients are due to EDA variants, and this specific form of the condition is termed X-linked hypohidrotic ectodermal dysplasia . The Leiden Open Variation Database currently lists 164 pathogenic or likely pathogenic variants . Autosomal recessive or dominant inheritance is seen in very rare forms of HED due to variants in EDAR or EDARADD .
|
[
"Stefan J. Rietmann",
"Noëlle Cochet-Faivre",
"Helene Dropsy",
"Vidhya Jagannathan",
"Lucie Chevallier",
"Tosso Leeb"
] |
https://doi.org/10.3390/genes15070854
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999996 |
PMC11276485_p3
|
PMC11276485
|
sec[0]/p[3]
|
1. Introduction
| 2.546875 |
biomedical
|
Other
|
[
0.998046875,
0.0004515647888183594,
0.0012760162353515625
] |
[
0.2392578125,
0.7470703125,
0.01122283935546875,
0.00255584716796875
] |
EDA variants causing hypohidrotic ectodermal dysplasia were also reported in mice , dogs and cattle . EDA -deficient dogs have been successfully used as animal models for therapeutic trials that are now ongoing in human patients .
|
[
"Stefan J. Rietmann",
"Noëlle Cochet-Faivre",
"Helene Dropsy",
"Vidhya Jagannathan",
"Lucie Chevallier",
"Tosso Leeb"
] |
https://doi.org/10.3390/genes15070854
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276485_p4
|
PMC11276485
|
sec[0]/p[4]
|
1. Introduction
| 3.207031 |
biomedical
|
Study
|
[
0.99609375,
0.003536224365234375,
0.000568389892578125
] |
[
0.97265625,
0.0171966552734375,
0.0008511543273925781,
0.009490966796875
] |
This study was prompted by the presentation of a male cat with clinical signs resembling hypohidrotic ectodermal dysplasia in humans and other species. The aim of our study was to provide a detailed characterization of the clinical phenotype together with an investigation of the underlying causative genetic variant.
|
[
"Stefan J. Rietmann",
"Noëlle Cochet-Faivre",
"Helene Dropsy",
"Vidhya Jagannathan",
"Lucie Chevallier",
"Tosso Leeb"
] |
https://doi.org/10.3390/genes15070854
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276485_p5
|
PMC11276485
|
sec[1]/sec[0]/p[0]
|
2.1. Ethics Statement
| 1.75293 |
biomedical
|
Study
|
[
0.9921875,
0.002063751220703125,
0.005542755126953125
] |
[
0.7490234375,
0.2449951171875,
0.00234222412109375,
0.0036773681640625
] |
The affected and all 96 control cats in this study were privately owned. Blood samples were collected with the consent of the owners. The diagnostic work-up of the index case did not constitute an animal experiment in the legal sense. The collection of blood samples from control animals was approved by the Cantonal Committee for Animal Experiments . All animal experiments were conducted in accordance with local laws and regulations.
|
[
"Stefan J. Rietmann",
"Noëlle Cochet-Faivre",
"Helene Dropsy",
"Vidhya Jagannathan",
"Lucie Chevallier",
"Tosso Leeb"
] |
https://doi.org/10.3390/genes15070854
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276485_p6
|
PMC11276485
|
sec[1]/sec[1]/p[0]
|
2.2. Clinical Examination
| 1.509766 |
biomedical
|
Clinical case
|
[
0.7236328125,
0.263671875,
0.0125885009765625
] |
[
0.042877197265625,
0.411865234375,
0.0068359375,
0.53857421875
] |
The affected cat underwent regular general and dermatologic examinations. Regular monitoring of atypical dermatitis and chronic calicivirus was carried out for two years. Hematological and biochemical check-ups, urinary analyses and abdominal ultrasound scans were regularly carried out as part of the animal’s clinical follow-up.
|
[
"Stefan J. Rietmann",
"Noëlle Cochet-Faivre",
"Helene Dropsy",
"Vidhya Jagannathan",
"Lucie Chevallier",
"Tosso Leeb"
] |
https://doi.org/10.3390/genes15070854
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999997 |
PMC11276485_p7
|
PMC11276485
|
sec[1]/sec[1]/p[1]
|
2.2. Clinical Examination
| 2.091797 |
biomedical
|
Study
|
[
0.994140625,
0.0021610260009765625,
0.0038318634033203125
] |
[
0.90283203125,
0.09283447265625,
0.0012102127075195312,
0.002960205078125
] |
The 96 control cats represented population controls without any reports of combined tooth and hair abnormalities. As the phenotype in the affected cat was very striking, we considered all 96 control cats as clinically unaffected.
|
[
"Stefan J. Rietmann",
"Noëlle Cochet-Faivre",
"Helene Dropsy",
"Vidhya Jagannathan",
"Lucie Chevallier",
"Tosso Leeb"
] |
https://doi.org/10.3390/genes15070854
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999998 |
PMC11276485_p8
|
PMC11276485
|
sec[1]/sec[2]/p[0]
|
2.3. DNA Isolation and Whole-Genome Sequencing
| 4.171875 |
biomedical
|
Study
|
[
0.99951171875,
0.0003211498260498047,
0.00019407272338867188
] |
[
0.99951171875,
0.0002951622009277344,
0.00025916099548339844,
0.00010102987289428711
] |
Genomic DNA was isolated from EDTA blood on a Maxwell RSC 16 or 48 instrument using the Maxwell RSC Whole Blood DNA Kit (Promega, Dübendorf, Switzerland). A PCR-free library with ~400 bp insert size was prepared from genomic DNA of the affected cat. The library was sequenced with 2 × 150 bp paired-end chemistry at 20× coverage on an Illumina NovaSeq 6000 instrument (Illumina, San Diego, CA, USA). The raw reads in fastq files were processed into a binary alignment map (bam-file) with respect to the F.catus_Fca126_mat1.0 genome reference assembly . Subsequently, single-nucleotide variants and small indels were called as described before . The accession numbers of the sequence data were deposited in the European Nucleotide Archive and are listed in Table S1 . Functional effects of the called variants were predicted with the SnpEff version 4.3t software together with NCBI annotation release 105 for the F.catus_Fca126_mat1.0 genome reference assembly.
|
[
"Stefan J. Rietmann",
"Noëlle Cochet-Faivre",
"Helene Dropsy",
"Vidhya Jagannathan",
"Lucie Chevallier",
"Tosso Leeb"
] |
https://doi.org/10.3390/genes15070854
|
N/A
|
https://creativecommons.org/licenses/by/4.0/
|
en
| 0.999995 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.