Longitudinal Assessment of Mental Health Disorders and Comorbidities Across 4 Decades Among Participants in the Dunedin Birth Cohort Study.

Importance
Mental health professionals typically encounter patients at 1 point in patients' lives. This cross-sectional window understandably fosters focus on the current presenting diagnosis. Research programs, treatment protocols, specialist clinics, and specialist journals are oriented to presenting diagnoses, on the assumption that diagnosis informs about causes and prognosis. This study tests an alternative hypothesis: people with mental disorders experience many different kinds of disorders across diagnostic families, when followed for 4 decades.


Objective
To describe mental disorder life histories across the first half of the life course.


Design, Setting, and Participants
This cohort study involved participants born in New Zealand from 1972 to 1973 who were enrolled in the population-representative Dunedin Study. Participants were observed from birth to age 45 years (until April 2019). Data were analyzed from May 2019 to January 2020.


Main Outcomes and Measures
Diagnosed impairing disorders were assessed 9 times from ages 11 to 45 years. Brain function was assessed through neurocognitive examinations conducted at age 3 years, neuropsychological testing during childhood and adulthood, and midlife neuroimaging-based brain age.


Results
Of 1037 original participants (535 male [51.6%]), 1013 had mental health data available. The proportions of participants meeting the criteria for a mental disorder were as follows: 35% (346 of 975) at ages 11 to 15 years, 50% (473 of 941) at age 18 years, 51% (489 of 961) at age 21 years, 48% (472 of 977) at age 26 years, 46% (444 of 969) at age 32 years, 45% (429 of 955) at age 38 years, and 44% (407 of 927) at age 45 years. The onset of the disorder occurred by adolescence for 59% of participants (600 of 1013), eventually affecting 86% of the cohort (869 of 1013) by midlife. By age 45 years, 85% of participants (737 of 869) with a disorder had accumulated comorbid diagnoses. Participants with adolescent-onset disorders subsequently presented with disorders at more past-year assessments (r = 0.71; 95% CI, 0.68 to 0.74; P < .001) and met the criteria for more diverse disorders (r = 0.64; 95% CI, 0.60 to 0.67; P < .001). Confirmatory factor analysis summarizing mental disorder life histories across 4 decades identified a general factor of psychopathology, the p-factor. Longitudinal analyses showed that high p-factor scores (indicating extensive mental disorder life histories) were antedated by poor neurocognitive functioning at age 3 years (r = -0.18; 95% CI, -0.24 to -0.12; P < .001), were accompanied by childhood-to-adulthood cognitive decline (r = -0.11; 95% CI, -0.17 to -0.04; P < .001), and were associated with older brain age at midlife (r = 0.14; 95% CI, 0.07 to 0.20; P < .001).


Conclusions and Relevance
These findings suggest that mental disorder life histories shift among different successive disorders. Data from the present study, alongside nationwide data from Danish health registers, inform a life-course perspective on mental disorders. This perspective cautions against overreliance on diagnosis-specific research and clinical protocols.


eAppendix 1. Sample
Participants are members of the Dunedin Study, a longitudinal investigation of health and behavior in a complete birth cohort. The 1,037 (535[52%] male) participants were all individuals born between April 1972-March 1973 in Dunedin, New Zealand (NZ), who participated in the first assessment at age 3 years, representing 91% of participants who were eligible based on residence in the province 1 . The cohort represented the full range of socioeconomic status on NZ's South Island and in adulthood matches the NZ National Health and Nutrition Survey on key health indicators (e.g., BMI, smoking, GP visits) and matches the NZ Census of citizens the same age on educational attainment 2 . The cohort is primarily white (93%), matching South Island demographics. Assessments were carried out at birth and ages 3,5,7,9,11,13,15,18,21,26,32,38, and most recently, 45 years, when 94% of the 997 Study members still alive took part. At each assessment, each Study member is brought to the research unit for a full day of interviews and examinations. Written informed consent was obtained from cohort participants, and study protocols were approved by the institutional ethical review boards of the participating universities.
Beginning at age 11 years, Study members have been interviewed privately by health professionals about their mental health and psychiatric diagnoses have been made according to the Diagnostic and Statistical Manual of Mental Disorders (DSM). Pediatric neurocognitive examinations were carried out at age 3, neuropsychological testing was carried out in both childhood and adulthood, and neuroimaging was performed at age 45 when brain age was estimated.
Up to age 15, diagnoses were made according to DSM-III 7 ; at ages 18 and 21, according to DSM-III-R 8 ; at ages 26, 32, and 38, according to the DSM-IV 9 ; at age 45 according to the now-current DSM-V 10 (with the exception of substance-dependence disorders which were diagnosed according to DSM-IV, given that DSM-V dropped the distinction between abuse and dependence). This is a limiting factor of our research because diagnostic criteria for some, but not all, disorders have changed a bit over the course of the past 35 years. It is also reality; the length of the Dunedin Study means that Study members have lived through multiple versions of psychiatric nosologies. We do not have the ability to always match past interviews to current nosologies or current interviews to past nosologies. As such, our report about the natural history of mental health reflects the lived experiences of Study members.
To describe the longitudinal patterns of mental disorder we focused on three developmental parameters: age-ofonset, duration (number of phases during which diagnostic criteria were met), and diversity (number of disorder types whose criteria were met). Figure 2 in the Main Article shows that these three key developmental parameters of mental-disorder life-histories were inter-correlated: age-of-onset was correlated with the number of assessment phases during which diagnostic criteria were met (r=.71 [95%CI:.68,.74], p<.001), with meeting criteria for more different types of disorders (r=.64 [95%CI:.60,.67], p<.001), and number of assessment phases during which diagnostic criteria were met was correlated with meeting criteria for more different types of disorders (r=.83 [95%CI:.81,.85], p<.001). The Table on the next page shows these same data in a tabular form. Column 1 shows the assessment age at which participants first met diagnostic criteria for a mental health disorder. Columns 2 and 3 show the sequelae of early onset. Early onset was associated with a greater likelihood of meeting diagnostic criteria at more subsequent 12-month assessment windows, up to midlife (column 2) and with meeting criteria for more different types of psychiatric disorders in subsequent years, up to midlife (column 3).

Age of first mental health diagnosis
Correction for observation window. It is possible that diversity of comorbid diagnoses could be a function of age ofonset, if individuals with older age-of-onset had fewer remaining waves of the study for diagnoses to be made. To correct for this, we calculated each individual's personal rate of diagnoses, by dividing their number of diagnoses by the 'n' of years between their onset age and the end of the study. This rate is referred to in the economics literature as a personal lambda. Next we re-estimated the association between age-of-onset and lifetime 'n' of total diagnoses. To perform this analysis, we had to omit those Study members who never met diagnostic criteria for a mental disorder and those who first met diagnostic criteria for a mental disorder at age 45 years. In the remaining subset of Study members (who had onset between 11 and 38 years), the association between age-of-onset and future n of disorders was r=.22 (95%CI: .17, .28), p<.001.
Additional details about mental disorder diagnoses in the Dunedin cohort. A reviewer inquired about the rates of schizophrenia and OCD in the Dunedin cohort.
The lifetime rate of schizophrenia in the Dunedin cohort is 3.7%. We have published our method of diagnosing schizophrenia in multiple publications over the past 2 decades 6,11 . It is believed that the prevalence of schizophrenia should be 1%, but, as we have discussed previously, there is a wide confidence interval around this 1% estimate. The Dunedin cohort's prevalence rate should be understood in the context of four methodological aspects of our study. First, our birth cohort, with its low attrition rate, allows us to count individuals with schizophrenia disorders overlooked by previous surveys. Individuals with psychotic disorders often decline to participate in surveys or die prematurely, and in addition surveys often exclude homeless or institutionalized individuals with psychosis. Our study assesses all of these groups missing from other surveys. Second, our cohort members are all from one city in the South Island of New Zealand. It is possible, given the known geographical variation in rates of schizophrenia, that the prevalence is somewhat elevated there. No comparable data exist to compare prevalence rates of schizophrenia in New Zealand to rates in other countries, but New Zealand has the highest prevalence of suicide worldwide and this fact could be consistent with a locally elevated prevalence of severe mental health conditions. Third, estimates of schizophrenia tend to be based on patients in clinical registers, but registers omit many community-dwellers whose disorder goes untreated. We note that over half of those diagnosed by the Dunedin Study were confirmed by receipt of treatment. By age 45, 2% of the cohort (N=20) met full DSM criteria for schizophrenia and had also been hospitalized for schizophrenia, according to our official New Zealand health system administrative record searches. However, an additional 1.7% (N=17) met all DSM criteria for schizophrenia, had auditory hallucinations by self-report (a criterion more strict than DSM), and suffered significant life impairment according to their informants. These 17 individuals had not, to our knowledge, been treated yet specifically for psychotic illness (those 20 treated and 17 not treated do not differ on cognitive status or symptom picture). Fourth, our research diagnoses did not make fine-grained distinctions among subtypes of psychotic disorders (e.g., schizophrenia versus schizoaffective psychotic disorder). Thus, the cohort members diagnosed with schizophrenia here might not be considered by all clinicians to have exclusively pure schizophrenia, which is what the oft-cited 1% lifetime prevalence rate is intended to reflect. Dunedin diagnoses of psychosis have been confirmed by consensus review by 2 psychiatrists.
The lifetime rate of OCD in the Dunedin cohort is 15%. It is believed that the prevalence rate should be around 2-3%. This belief is probably based on the NCS-R estimate. The NCS-R estimate is based on lifetime retrospective reports, which are known to undercount. It has been shown in multiple longitudinal cohort studies that lifetime prevalence rates in retrospective surveys are undercounted by at least half for many conditions; in fact, this literature is presented for readers in eAppendix 5 in our Supplement. How this discrepancy in lifetime prevalence between (a) one-off retrospective surveys and (b) cumulative prospective longitudinal studies happens can be easily seen in the Dunedin Study. The 12-month rates of OCD are presented in the table below, in gray. They range from 2% (in midlife adults) to 7% (in young people). OCD disorders are fairly stable in the cohort. That is, people who are diagnosed with OCD at one age are statistically more likely to be diagnosed with OCD at subsequent ages. This can be seen in the table below, by the transition matrix of odds ratios (transition ORs) which follows the simplex-like pattern one expects to see in longitudinal data. But there is also change over time (many people remit from OCD and new incident cases of OCD also accumulate). The result is that through multiple assessments, we end up with a total number of 150 who met diagnostic criteria for OCD at least once during several decades. Indeed, this is one important point of our report. When one takes a longitudinal life-course perspective on mental disorders rather than a cross-sectional snapshot, we see that lifetime mental disorders are much more prevalent than previously assumed. This is an important public health message (which we have discussed in the past specifically in reference to OCD 12 ).
Prevalence rates of OCD and transition odds ratios for OCD (ORs):

. Measuring Brain Function Across the Life Course: Age-3 Brain Health, Child and Adult Cognitive Functioning, Child-to-Adult Cognitive Decline, and Accelerated Brain Aging
Measuring age-3 brain health. At age 3 years, each child in the cohort participated in a 45-minute examination that included assessments of neurological soft signs, intelligence, receptive language, and motor skills, and afterwards the examiners (having no prior knowledge of the child) rated each child's behavior (all described in the Table  below). Using this information, we created a summary factor score via confirmatory factor analysis which we termed brain health, a global index of the child's early neurocognitive status 13 . The model fit the data well, χ 2 (N=1035, df=5) = 6.459, p = .2641, CFI = .999, TLI = .997, RMSEA = .017. Factor scores were output and standardized to a Mean = 0 and SD= 1.

Measure/Test Description Neurologic soft signs
At age three years, each child was examined by a pediatric neurologist for neurologic signs, including assessment of motility, passive movements, reflexes, facial musculature, strabismus, nystagmus, foot posture, and gait, based on procedures described by Touwen & Prechtl 14 . Peabody Picture Vocabulary Test Intelligence was assessed at age three with the Peabody Picture Vocabulary test 15 .

Receptive Language
Receptive language was assessed at age three using the Reynell Developmental Language Scales (25) .

Motor Development
Motor development was assessed at age three years with the Bailey Motor Scales 16 .

Lack of Control
Following the testing, each examiner rated the child's lack of control in the testing session, yielding a behavioral style factor, labeled Lack of Control 17 , which characterized children who at age three years were labile, had low frustration tolerance, lacked reserve, were resistant, restless, impulsive, required attention, and lacked persistence in reaching goals.
Measuring cognitive functioning and cognitive decline. The Wechsler Intelligence Scale for Children -Revised (WISC-R) 18 was individually administered at ages 7, 9, and 11 years. IQ scores for the three ages were averaged 19 .
The Wechsler Adult Intelligence Scale-IV (WAIS-IV) 20 was individually administered at age 45 years.
We measured cognitive decline by studying IQ scores at midlife after controlling for IQ scores in childhood. (As a sensitivity analysis, in addition to analyzing residualized change we also analyzed difference (change) scores, and obtained the same substantive and statistically-significant results.) We focus on change in the overall IQ given evidence that age-related slopes are correlated across all cognitive functions, suggesting that research on cognitive decline may be best focused on a highly reliable summary index, rather than focused on individual functions 21 .
Measuring accelerated structural brain aging. At age 45 years, brain images were acquired from Study members using a Siemens Skyra 3T equipped with a 64-channel head/neck coil. We estimated Brain Age with a publicly available algorithm 22 which uses information about cortical anatomy and whole-brain functional connectivity to estimate the age of a person's brain relative to their chronological age. The algorithm has been shown to predict chronological age in multiple independent samples, although it has a documented tendency to underestimate chronological age by approximately 3 years among adults between chronological ages 44-46 (and for this reason we standardized the scores to the mean chronological age of the Dunedin Study members at the time of their scanning in the Phase-45 assessment) 23 . Deviations of predicted brain age upwards of chronological age are presumed to reflect accelerated brain aging.

eAppendix 4. Modeling the Structure of Psychopathology
We have previously described the structure of psychopathology up to age 38 years 13 ; here we extend these models to include the age 45 data.
We used symptom data from the 6 adult assessments, carried out at ages 18 Using Confirmatory Factor Analysis (CFA), we tested two standard models that are frequently used to examine the structure of psychopathology 25 : (a) a correlated-factors model and (b) a hierarchical or bifactor model. Data analysis syntax appears in the last section of this supplement. In CFA, latent continuous factors are hypothesized to account for the pattern of covariance among observed variables. Our CFAs were run as multitrait-multimethod models. In these models, observed variables represented each of the disorders with a symptom scale at each assessment age (e.g., alcohol dependence was measured with a symptom scale at ages 18, 21, 26, 32, 38, and 45). Each model also included method/state factors designed to pull out assessment-related variance (e.g., assessment-specific interviewer effects, assessment-specific study member mood effects) that was uncorrelated with the psychopathology factors of interest. Because symptom-level data are ordinal and have highly skewed distributions, we used polychoric correlations when testing our models. Polychoric correlations provide estimates of the Pearson correlation by mapping thresholds to underlying normally distributed continuous latent variables that are assumed to give rise to the observed ordinal variables. All CFA analyses were performed in MPlus version 8.3 26 using the weighted least squares means and variance adjusted (WLSMV) algorithm. 27 We assessed how well each model fit the data using the chi-square value, the comparative fit index (CFI), the Tucker-Lewis index (TLI), and the root-mean-square error of approximation (RMSEA The high lifetime diagnosis rates in the Dunedin Study may come as a surprise to some readers. But, in fact, these rates are in line with data from other epidemiological studies, and with rates reported in other prospectivelongitudinal studies. Specifically, multiple longitudinal-epidemiological studies from different countries converge on (a) finding that by age 15-16, approximately 35% of children meet criteria for a mental disorder and (b) that up to midlife the vast majority of people will have experienced a mental disorder in their lifetime.
Costello et al. 29 reported the cumulative prevalence of child and adolescent psychiatric disorder in a cohort of 1,420 North Carolina children who were first assessed when they were 9 to 13 years old and assessed annually thereafter. By age 16, 36.7% had received research diagnoses of at least one psychiatric disorder. In a cohort of 447 children from two upstate New York counties who were assessed for psychiatric disorder when they were 9 to 14 years of age and again at 12 to 16 years 30 39% had been diagnosed with at least one psychiatric disorder by age 16. Among Dunedin Study participants, who were assessed for psychiatric disorder at 11, 13, and 15 years of age, we find that 35.4% met criteria for at least one psychiatric disorder by age 15. The consistency in these estimates of the cumulative prevalence of child and adolescent psychiatric disorder is striking given that the samples come from two countries (the United States and New Zealand), two regions of the United States (the rural south versus northeastern United States), and involve different historical cohorts of children (born in the 1960s and 1970s in New York and New Zealand and born in the 1980s in North Carolina).
By midlife, lifetime rates continue to accumulate. Prospective studies support the contention that retrospective and single-wave, cross-sectional studies underestimate the burden of disease in the population over time. As shown below, the lifetime rates that we report are in keeping with rates reported in all other prospective-longitudinal studies. First, we show data from studies of Scandinavian national registries which report that the lifetime prevalence of registered mental-disorder treatment is 33%. However, because many people with disorder are not treated, this is a lower bound. Second, we show data from cross-sectional surveys, such as the U.S. National Comorbidity Survey (NCS-R), that ask people to report retrospectively about their lifetime experience with mental disorders. These estimate lifetime prevalence near 50%. However, individuals with disorders resulting in homelessness, institutionalization, and survey refusal are missed in such surveys, and respondents' retrospective reports are documented to be biased by recall failure. Thus, 50% is an undercount. Third, we show data from the Dunedin Study and four other prospective birth cohort studies. These studies, begun decades ago, count cases irrespective of treatment, minimize recall failure, and gradually build participants' trust; these studies report that the vast majority of people experience a mental disorder at some point in their 31

eAppendix 6. Does Anyone Have Just One Exclusive Diagnosis?
The table provides the data graphed in Figure 3

eAppendix 8. Sequential Comorbidity
The first figure summarizes the sequential comorbidity of Internalizing, Externalizing, and Thought disorders. Participants with a disorder in any of the three diagnostic families at one specific age were at significantly higher risk for both other diagnostic families at subsequent ages. The Risk Ratios in black depict the continuity of the same disorders (e.g., "What is the risk of people with an Internalizing disorder at age 15 or at age 18, or at age 21, etc., presenting with a subsequent Internalizing disorder at later phases?"). The Risk Ratios in red depict sequential comorbidity (e.g., "What is the risk of people with an Internalizing disorder at age 15, or at age 18, or at age 21, etc., presenting with a subsequent Externalizing disorder at later phases?"). Average risk ratios across ages were calculated with a Generalized Estimating Equation (GEE) that nested individuals within time.
The next figure shows the risk of presenting with a specific disorder at subsequent assessment waves given a specific disorder at an earlier assessment wave. The Risk Ratios on the diagonal depict the continuity of the same disorder; the off-diagonal Risk Ratios depict sequential comorbidity from the row diagnoses to the column diagnoses. Average Risk Ratios across assessment phases were calculated using Generalized Estimating Equations (GEE) that nested individuals within time. The overall impression in this figure is one of a uniform positive manifold: Individuals who meet criteria for one disorder are significantly more likely to subsequently meet criteria for the same disorder (along the diagonal) but also different disorders. Of the 196 risk ratios estimated, 183 (93%) were positive (only four risk ratios were <= 1.0 and nine could not be estimated given that models would not converge; these nine mostly involved eating disorders and mania, which had the lowest prevalence rates in the Study). The figure makes clear that longitudinal "cross-family" patterns are not confined to particular pairings, but are ubiquitous.

To Subsequent Diagnosis
From    37 . For diagnoses of schizophrenia and mania, we have had independent reviews by clinicians who achieved interrater agreement on cases 6 . Unreliability tends to be magnified in diagnostic data, where subthresholds are ignored. For example, a person with X number of symptoms is said to meet diagnostic criteria, but a person with X-1 does not, and such qualitative thresholds contribute to apparent unreliability. What impact does unreliability have? Unreliability in measurement underestimates associations between different disorders over time. To document this, we re-estimated our analyses which show longitudinal shifting of mental disorders across time. Our analyses of sequential comorbidity used GEE (General Estimating Equations) to estimate the risk that people with a specific disorder at one wave will present with a different disorder at subsequent waves, and to estimate the likelihood that this shifting occurs across Internalizing, Externalizing, and Thought disorders. We re-estimated these models using Latent Makov analyses (LM). LM is a form of error-in-variable model, which takes account of unreliability (one can think of it as a "correction" which tells us what the estimated association is expected to be if one could measure X and Y with perfect reliability of 100% As is apparent by comparing the GEE and Latent Markov estimates, unreliability does not affect the pattern of findings. The estimates were all significant in the GEE models, and they are, if anything, stronger in the Latent Markov models. We continue to see significant changes between disorders across repeated assessments years apart. This shows that diagnostic unreliability is not the culprit in why we observe shifting among different successive disorders over the life course. Moreover, whereas some of our analyses use qualitatitve diagnostic thresholds, our confirmatory factor analyses rely on quantitative symptom-level information (see eAppendix 4).

eAppendix 10. The Ebb and Flow of Mental Disorders Among Participants Who Received Inpatient Mental-Health Services.
Over the course of the Study, 83 participants received inpatient services. The Sankey chart highlights the 83 inpatient mental-disorder life-histories embedded within the mental-disorder life-histories of the entire cohort, which are shown in gray background (see Figure 4 in the Main Article). (Note: it is possible to follow groups across contiguous assessments, not across the entire panel.) The Main Article, and the scatterplots above, report the associations between p-factor scores and measures of brain function. These correlations are shown in the shaded column of the next  The results show that each of the three psychopathology factors was antedated by age-3 brain dysfunction, accompanied by child-to-adult cognitive decline, and associated with older brain-age at midlife. There was no specificity. This nonspecificity is understandable in light of the life-history evidence that people's diagnosis changes frequently, which is summarized parsimoniously in p.  Age18 BY adhd18b* alc18 mar18 CD18 mde18 gad18 fear18 anrx18 bul18 ocd18; Age21 BY alc21* mar21 smk21 CD21 mde21 gad21 fear21 anrx21 bul21 ocd21 man21 scz21; Age26 BY alc26* mar26 drg26 smk26 CD26 mde26 gad26 fear26 anrx26 bul21 ptsd26 ocd26 man26 scz26; Age32 BY alc32* mar32 drg32 CD32 mde32 gad32 fear32 ptsd32 ocd32 man32 scz32; Age38 BY adhd38b* alc38 mar38 drg38 smk38 CD38 mde38 gad38 fear38 ptsd38 ocd38 man38 scz38; Age45 BY adhd45b* alc45 mar45 drg45 smk45 CD45 mde45 gad45 fear45 ptsd45 ocd45 man45 scz45;