key: cord-0316951-idt0ctz8
authors: Griffith, G. J.; Jones, K.
title: When does geography matter most? Age-specific geographical effects in the patterning of, and relationship between, mental wellbeing and mental illness.
date: 2020-07-17
journal: nan
DOI: 10.1101/2020.07.15.20152645
sha: d818f17687f69c3c3223c83e210ecd0c77b2ba38
doc_id: 316951
cord_uid: idt0ctz8

Mental illness and mental wellbeing are related but distinct constructs. Despite this, geographical enquiry often references the two as interchangeable indicators of mental health and assumes the relationship between the two is consistent across different geographical scales. Furthermore, the importance of geography in such research is commonly assumed to be static for all age groups, despite the large body of evidence demonstrating contextual effects in age-specific populations. We leverage simultaneous measurement of a mental illness and mental wellbeing metric from Understanding Society, a UK population-based survey, and employ bivariate, cross-classified multilevel modelling to characterise the relationship between geographical context and mental health. Results provide strong evidence for contextual effects for both responses before and after covariate adjustment, with weaker evidence for area-classification and PSU-level contextual effects for the GHQ-12 after covariate adjustment. Results support a two-continua model of mental health at the individual level, but indicates that consensual benefit may be achieved across both dimensions by intervening at household and regional levels. There is also some evidence of a greater contextual effects for mental wellbeing than for mental illness. Results highlight the potential of the household as a target for intervention design for consensual benefit across both constructs. Results highlight the increased importance of geographical context for older respondents across both responses. This research supports an area-based approach to improving both mental illness and mental wellbeing in older populations.

Mental illness and mental wellbeing are related but distinct constructs. Despite this, geographical 19 enquiry often references the two as interchangeable indicators of mental health and assumes the 20 relationship between the two is consistent across different geographical scales. Furthermore, the 21 importance of geography in such research is commonly assumed to be static for all age groups, 22

despite the large body of evidence demonstrating contextual effects in age-specific populations. We 23 leverage simultaneous measurement of a mental illness and mental wellbeing metric from 24

Understanding Society, a UK population-based survey, and employ bivariate, cross-classified 25 multilevel modelling to characterise the relationship between geographical context and mental 26 health. Results provide strong evidence for contextual effects for both responses before and after 27 covariate adjustment, with weaker evidence for area-classification and PSU-level contextual effects 28

for the GHQ-12 after covariate adjustment. Results support a two-continua model of mental health 29 at the individual level, but indicates that consensual benefit may be achieved across both 30 dimensions by intervening at household and regional levels. There is also some evidence of a greater 31 contextual effects for mental wellbeing than for mental illness. Results highlight the potential of the 32 household as a target for intervention design for consensual benefit across both constructs. Results 33 highlight the increased importance of geographical context for older respondents across both 34

responses. This research supports an area-based approach to improving both mental illness and 35 mental wellbeing in older populations. 36 Ambiguity around wellbeing and mental illness measures is a critical limitation of the literature as 50 the two constructs are distinct but correlated dimensions of mental health and cannot be assumed 51

to be equivalent (Keyes, 2002; Haworth et al., 2017) . Indeed, the relationship between mental illness 52 and wellbeing and their covariates is still debated in the psychological and social science literature 53 (Westerhof and Keyes, 2010; Kinderman et al., 2015) . Amongst adolescents cross-sectional evidence 54

shows different demographic prediction of wellbeing and mental illness (Patalay and Fitzsimons, 55 2016), and longitudinal evidence seems to suggest that despite complex sex patterning, predictors of 56 lower mental wellbeing broadly correlate with those of mental illness (Patalay and Fitzsimons, 2018) . 57

Whilst consensus on the dissimilarity of the constructs is emerging (Westerhof and Keyes, 2010; 58 Lamers et al., 2015) , assuming a relationship between constructs and any observed predictors relies 59 critically on assumptions of measurement validity for the measure of interest (Fried et al., 2016) . 60 Whether this much-scrutinised relationship between mental illness and mental wellbeing is 61 recapitulated in higher level spatial contexts has yet to be comprehensively explored. 62

Whilst definitions of mental health differ, there is a history dating back to Faris and Dunham (1939) 63 of research continuing to investigate the complex link between geographical context and mental 64 health ( Arcaya and Subramanian, 2017; Bambra, Smith and Pearce, 2019). It has been argued that 68

geographical processes influencing mental health should not be limited to a single spatial scale 69 (Ross, 2000; Pickett and Pearl, 2001) . Increasingly multilevel techniques have been adopted in 70 mental health research to allow multiple spatial scales to be investigated (Weich et al., 2003; 71 Propper Multilevel models are widely used to parameterise contextual and compositional geographical 73 effects in quantitative health geography (Duncan, Jones and Moon, 1995) . In this framework 74 outcomes are predicted using individual level variables nested within a spatial framework. 75

Contextual effects on a given outcome are thus are estimated net of the characteristics of the 76 individuals who live within them (Owen, Harris and Jones, 2016) . Functionally this means that, after 77 adjusting for individual covariates, contextual effects are assumed to be captured by common 78 unexplained variation at a higher structural level. A modelling approach which allows separation of 79 compositional and contextual effects is of particular relevance to epidemiological questions as 80

health-related processes are not produced within a single structural framework. Contextual and 81 compositional effects were found to exist net of one another in 22 of 33 previous research articles, 82 in a recent review of place-based health modelling (Schule & Bolte, 2015) . 83

The multilevel geographical approach is not without limitations. In a strictly hierarchical framework, 84 this also assumes consistent contextual effects exist solely between geographically adjacent areas. 85

Unexplained variation which is common across areas which not mutually nested within a higher level 86 grouping will be expressed at a lower structural level, as no common higher level structure exists to 87 capture these effects (Tranmer and Steel, 2001 CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted July 17, 2020. The final question of the GHQ-12 and the first of the SWE appear only 4 pages apart in the self-165 completion questionnaire and participating respondents are encouraged to complete it in the 166 presence of an interviewer, so measurement can be reasonably assumed contemporaneous. 167

The GHQ-12 was initially developed to detect psychiatric morbidity, but has since been extensively 168 validated and deployed as a population screening metric for broader mental health outcomes 169 (Goldberg, 1972 CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted July 17, 2020. both the GHQ-12 respondents and the SWE respondents is given below in Table 1 . 195 We take the postcode sectors constituting the Primary Sampling Units (PSUs) as a neighbourhood 200 measure. In order to explicitly investigate the importance of the characteristics of an area over and 201

above that of the broad region itself, we harmonise two separate measures of geographical context 202 with PSU and household-level data. Firstly, 22 higher-level geographical regions are constructed 203 using a formulation taken from Jones et al., (1992) to form the highest layer of our spatial hierarchy. 204

These reflect a breakdown of the standard UK governmental office regions into urban 205

"Metropolitan" areas, and more rural "Non-metropolitan" areas. Secondly, households are linked to 206

Lower Super Output Areas (LSOAs), allowing census based area-classification to be used to 207 categorise households into 52 different types of area. These area-classifications are taken from 208 geodemographic classifications taken from Vickers and Rees (see Vickers and Rees, 2006, for a full 209 description). As the model nesting is non-hierarchical this means two non-adjacent geographical 210 areas can belong to the same area-classification whilst being in different regions, and vice versa. 211

Simultaneous consideration of variance components at both levels is required to assess the relative 212 contribution of each level, necessitating an appropriately comprehensive methodological approach. 213 214 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. contextual effects for each outcome, but further allows us to estimate the covariance between 229 residuals, at each structural level within the model. Were the outcomes modelled separately we 230 would be able to compare the magnitude but not the direction of the unexplained variation in 231 residuals meaning even if the same structural level explained greatest variation, we could not know 232 the effects were consensual across outcomes. More simply, modelling both outcomes allows us to 233 see at what structural level the GHQ-12 and the SWE behave most similarly, which can inform the 234 appropriate spatial scale for intervention. If we see consistent effects across a specific structural 235 level, this may offer a promising level for intervention for consistent benefit across mental ill-health 236 and mental well-being. 237

A simplified form of the outlined bivariate outcome specification, illustrating the specification of 238 random effects predictors and contextual effects, is provided below: 239 1 = 0 + ∑ 1 1 1 + ∑ 2 2 1 + ( 1 + 1 ) 240 2 = 3 + ∑ 4 1 1 + ∑ 5 2 1 + ( 2 + 2 ) 241 is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. . https://doi.org/10.1101/2020.07.15.20152645 doi: medRxiv preprint covariates are given by 1 and 2 , and are assumed to come from a joint Normal distribution with 251 a mean of zero and variance given by the associated variance-covariance matrix. The variance of the 252

Level-2 residuals for GHQ-12 is summarised by 1 2 , the between-household variation in GHQ-12 253 responses. The household-level covariance between responses is given by 1 2 , a positive value 254 would indicate that households with high average GHQ-12 values would also tend to exhibit high 255

average SWE values. Similarly, the variance of the Level-1 residuals for GHQ-12 and SWE are 256 summarised by 1 2 and 2 2 respectively, which represent the variation between individuals in the 257 same household. The individual-level covariance between responses is given by 1 2 , with a 258 positive value indicating that individuals who have high GHQ-12 scores tend to also have high SWE 259

scores. Model subscripts are elevated one character relative to traditional notation, as variation at 260 the sub-level (i) exists solely to determine the multivariate structure. 261

It is the variance and covariance terms which are of particular interest to the research questions we 262

are concerned with here. Using the variance components described above, the variance partitioning 263 coefficient (VPC) can be calculated for a given structural level using the formula below. The VPC 264

gives the degree of unexplained variation remaining in the model which is patterned at that 265 structural level. For more complex specifications, the substantive interpretation of level-specific 266 variance divided by total variance does not change, but the terms constituting these values will. The 267 modelled covariances at each given level can be standardised to give correlation coefficients, 268

indicating the degree of consistency between the GHQ-12 and SWE at a given structural level. 269

Level 2 VPC for 1 = CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. . https://doi.org/10.1101/2020.07.15.20152645 doi: medRxiv preprint

In order to assess the presence of geographical patterning for each response, an unadjusted 281 bivariate response model was fitted, with five cross-classified structural levels. In specifying this 282 model, we are estimating area-level effects without adjusting for demographics other than those 283

included in the generation of the area-classifications. Table 2 displays variance estimates for both  284 outcomes at all levels, with p-values under a null hypothesis of no contextual effect. Table 2  285 illustrates that there is strong evidence of non-zero contextual effects at all structural levels in an 286 unadjusted model for both outcomes. Testing for greater importance of higher hierarchical levels, 287

we find strong evidence that the contextual effect of PSU is greater than region for SWE responses 288 (p=<0.001), but not for GHQ-12 (p=0.143). There is also strong evidence that area classification is the 289 most important of all non-household, contextual levels for both responses (GHQ-12 p= <0.001, SWE 290 = <0.001). 291 CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. dose-response relationship for education in SWE scores, such that those whose highest qualification 324 is A-level or equivalent report lower wellbeing than those completing higher education (0.431, 325

p=<0.001) and those whose highest qualification is GCSE or equivalent report lower wellbeing again 326 (0.822, p=<0.001). There is no evidence for such a relationship for GHQ-12 responses, for which 327 those whose highest educational qualification is A-level do not respond significantly differently to 328 those completing higher education (0.022, p=0.256). Similarly there is strong evidence of higher 329 mental distress amongst those who complete A-level (0.230, p=0.002) and those with no formal 330 qualification (0.251, p=0.002) than those who complete higher education, however there is not 331 strong evidence of a difference between these two groups. 332 333 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. . https://doi.org/10.1101/2020.07.15.20152645 doi: medRxiv preprint Table 5 gives the level-specific variance estimates for the model displayed in Table 4 . There is still 340 strong evidence of non-zero contextual effects at all levels for the SWE responses, but evidence is 341 weaker for a non-zero contextual effect for GHQ-12 residuals at area-classification (0.015, p=0.051) 342

and PSU level (0.100, p=0.051). Region and area-classification level variance are considerably smaller 343 than seen in the unadjusted model, suggesting that a large portion of the contextual effects seen in 344 the unadjusted model was explained by within-group compositional effects. Consistent with the 345 unadjusted model, there is still strong evidence of PSU-level variation being greater than the region 346 level variance for both the GHQ-12 (0.100, 0.043, p=0.006) and the SWE (0.155, 0.050, p=<0.001). 347 CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. . https://doi.org/10.1101/2020.07.15.20152645 doi: medRxiv preprint assumed that the relative importance of geography is consistent for all respondents. Here, we allow 381 modelled age coefficients to vary by area-classification and region-level geography, allowing the 382 estimation of context-specific age effects. More simply, this allows the estimation of for which age 383 group geography matters most. 384 Figure 2 gives unadjusted variance relationships for both responses across the region and area-385 classification levels over the range of ages in USoc. Unadjusted variance estimates can be seen in 386

Supplementary Table 2 . Although previous results do not indicate strong regional differences in 387 either outcome (Table 2) , when this is allowed to vary by age group there is clear age-heterogeneity 388 in regional contextual effect . Figure 2suggests that both region and area-classification matter more 389 for the oldest and youngest age groups. Unadjusted variance estimates are greater for the SWE 390 responses than for the GHQ-12. . Region level VPC estimates appear more important than area-391 classification Variance patterning seems to display a consistent but exaggerated pattern to that in 392 Table 2 , with SWE showing greater contextual effects than the GHQ-12, and greater region-level 393 contextual effects seen for both responses. Having established the importance of age for contextual effects in unadjusted estimates, we now 399 consider age-sensitive contextual effects net of covariates. Figure 3 gives the same output as Figure  400 2 but adjusted for covariates (with estimates in Supplementary Table 3 ). The contextual effect of 401 area-classification has heavily declined, in line with VPCs in Table 6 . However, there is still evidence 402 of a greater importance of geographical context captured at the region-level for the older and 403 younger participants. The range of region residuals net of age is around 0.5 (on a 0-28 scale) for the 404 most extreme regions on either score (see Supplementary Figure 1 ). The importance of geographical 405

context for the elderly and young has diminished after adjusting for covariates, suggesting that age-406 specific contextual effects may have been driven at least partially by the differential demographic 407 patterning of differently aged groups. The importance of context is still lowest for middle aged 408 respondents, indicating geographical consistency in the midlife distress peak. 409 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. Table 2 for household and individual level, region does not account for variation 416 in the magnitude that between-individual or between-household differences levels do. However, 417

there is a clear suggestion that region may be a non-trivial source of variation for low and high aged 418 respondents, net of the covariates included in the model. 419

[ Figure 4 about here] 420 knowing all individuals' GHQ-12 scores, and knowing they hypothetically resided in the same 439 household, same PSU, same region and same area-classification, would allow the characterisation of 440 29% of the variation in SWE response. If we treat the responses as is common in the literature, 441

namely that they are capturing mental wellbeing and mental illness, this correlation strongly 442 supports a two-continua model of mental health with mental wellbeing and mental illness being 443 distinct but correlated constructs (Keyes, 2002) . In unadjusted models, the highest correlation exists 444 at area-classification level (0.98), suggesting that in absence of detailed demographic information, 445

interventions aiming to consensually improve wellbeing and mental illness may be more 446 appropriately targeted at areas based on their characteristics, rather than based on broad 447 geographical location. 448

In order to test whether this higher-level patterning exists for compositional or truly contextual 449 reasons the model estimates were adjusted for a number of individual and household level 450

predictors. The patterning of responses by demographic characteristics was broadly similar to that 451 found in previous work, specifically the patterning of wellbeing responses by sex (Patalay and 452 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. . https://doi.org/10.1101/2020.07.15.20152645 doi: medRxiv preprint Fitzsimons, 2018). After adjusting for demographic characteristics, there was strong evidence of non-453

zero contextual effects at all levels for the SWE, and strong evidence for non-zero contextual effects 454 at the region level for the GHQ-12. Evidence was less strong for area classification and PSU level 455 contextual effects for the GHQ-12. After covariate adjustment there is similar magnitude of regional 456 contextual effect, but there is stronger evidence for consistent regional contextual effects across 457

measures to evidence effects amongst older participants due to sample size, however there is evidence of 470 greater contextual effects amongst elderly respondents. A clearer representation of the relative 471

importance of this variance for adjusted models is given in Figure 4 , where region-level variation can 472 be seen to account for less than 0.25% of variation in middle aged respondents but approaches 4% 473

in the older respondents. Region-specific residuals are presented net of covariates in the 474 supplementary materials, with Supplementary Figure 1 illustrating some interesting patterns, such 475 as the consensually poor mental health experience in the urban west-midlands and rural northwest, 476 and the positive mental health experience in Scotland once individual controls are included, 477 however discussion of these is beyond the scope of this paper. 478

There are several limitations to this work. Firstly, this work is explicitly carried out using the first 479

wave of USoc as it provides a unique insight into the response patterning of first-time respondents 480 to two complex mental health outcomes without concerns about possible retest effects. However 481 this means the research is cross-sectional, making it difficult to unpack the age and mental health 482 relationships discussed here as we cannot assume individuals longitudinally track these age profiles, 483

nor that there is no specific cohort effect producing this pattern net of respondent age (Bell, 2014) . . 484

The associations presented here may highlight mechanisms for further study in the longitudinal 485 spatial epidemiological framework advanced by contemporary health geographers (Green, Arcaya 486 and Subramanian, 2017; Morris, Manley and Sabel, 2018).Secondly, the results are conditional on 487 the structure in which it is specified. Although we incorporate more contextual complexity than 488 previous studies, explicit choices in the selection of structural variables were still required (Owen, 489 Harris and Jones, 2016). This may induce spurious patterning over levels included in the model in the 490 absence of the levels at which variance truly lies (Tranmer and Steel, 2001) . The non-hierarchical 491 structure imposed via area-classification attempts to address this, and behaved as anticipated given 492 individual covariates, however there are many more potential structural compositions which could 493 be hypothesised and investigated. Moreover, the selection of included covariates also imposes 494 structure on the conclusions. For instance, we consider education, job classification and housing 495 tenure to capture the three main dimensions of socio-economic position, but many 496 parameterisations exist and this choice may influence observed effects (e.g. Galobardes et al., 2006; 497 Weich, Twigg and Lewis, 2006) although less so for older respondents (Darin-Mattsson, Fors and 498

Kåreholt, 2017). We have largely considered unexplained variation at a given level as indicative of 499 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. . https://doi.org/10.1101/2020.07.15.20152645 doi: medRxiv preprint contextual effects; however, we cannot be certain this is contextual rather than simply a result of 500 unmeasured compositional effects which we have not controlled for. Further research should 501 explore the effect on the relationship between mental illness and wellbeing net of a wider set of 502 covariates, and under the imposition of different geographical structures in complex cross-classified 503 frameworks (Leckie, 2019) . 504

Thirdly, non-response to the SWE and GHQ-12 questionnaires is unlikely to be truly random. This is 505 evidenced in lower response rates for the self-completion questionnaires in the USoc EMBS (69.6%) 506

versus the GPS (87.4%). Although demographic differences between the 40452 partial GHQ-12 507 respondents and the 37836 SWE respondents were slight, if non-response is patterned with respect 508

to unobserved variables then spurious relationships may be introduced between predictors and 509 outcomes (Munafò et al., 2018) . As such, caution must be exercised when generalising findings to a 510 target population if it is suspected that selection processes are likely to differ between observed 511 respondents and the target population (Griffith et al., 2020) . 512

Finally, with any investigation into self-rated mental health, there is measurement error inherent in 513 response variables. Responses do not capture noiseless indication of mental health, but also 514 constructs such as capacity to articulate mental health, which has been suggested to be a factor in 515 the divergence in male and female responses in self-reporting and deaths by suicide (Yong, 2006; 516 Moore et al., 2013; Rodrigo et al., 2019) . As such, structural differences which appear to indicate 517 patterning of mental health experience could truly reflect patterning of response tendency, such as 518

stoicism. It is difficult to entirely account for this in further work, but establishing measurement 519 invariance between geographical regions is a reasonable starting point for research aiming to inform 520 group-based policy intervention (Byrne and Watkins, 2003 ; L. Milfont and Fischer, 2010) .Evaluation 521 of measurement invariance across geographical groups of survey respondents as standard practice 522 would greatly strengthen inference drawn from such studies of mental health outcomes (Griffith and  523 Jones, 2019). 524

This study has demonstrated the value of bivariate cross-classified models for investigating 527 structural similarities between complex mental health constructs at a series of conceptual scales. 528

There is strong evidence of non-zero contextual effects for both responses both before and after 529 covariate adjustment. The findings add to a body of evidence illustrating substantive dissimilarity 530 between mental illness and mental wellbeing and we advocate for explicit characterisation and 531

consideration of the dimension of mental health under investigation. These findings support 532 evidence of the potential of the household as a key level at which to target policy intervention, given 533 consistency between mental wellbeing and mental illness measures within a household. 534

Furthermore, the evidence suggests that wellbeing may be patterned more strongly than mental 535 illness with respect to household, local and area context. We demonstrate an age varying 536 importance of regional geography which suggests the greatest potential for intervention lies in 537 targeting the consistently poorer mental health of middle-aged respondents across the UK. 538

Consistency in midlife mental health distress is also consistent with the suggestion that those 539 experiencing the greatest mental distress are the most geographically consistent. There appears is 540 great potential for area-based interventions among older populations. Finally, our results suggest 541 that although there exists a midlife peak in distress and a U-shaped wellbeing curve, the magnitude 542 of this curve may not hold consistently across geographical locations, especially for the elderly. 543 544 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. . https://doi.org/10.1101/2020.07.15.20152645 doi: medRxiv preprint CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted July 17, 2020. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted July 17, 2020. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted July 17, 2020. . https://doi.org/10.1101/2020.07.15.20152645 doi: medRxiv preprint

Area variations in health: A spatial multilevel modeling approach', Health 546 and Place

Scaling up : The politics of health and place

Life-course and cohort trajectories of mental health in the UK, 1991-2008 -A 551 multilevel age-period-cohort analysis', Social Science and Medicine

Design of the Understanding Society Ethnic Minority Boost Sample Ethnic 554 Minority Boost Sample

Is Happiness U-shaped Everywhere? Age and Subjective Well-being in 556 132 Countries

Social Isolation and Loneliness in Older Adults-A Mental Health/Public Health 559 Challenge

Calibrating well-being, quality of life and common mental 561 disorder items: psychometric epidemiology in public mental health research

UKHLS Wave 1 Technical Report

MCMC Estimation in MLwiN, v3.03. Centre for Multilevel Modelling

The issue of measurement invariance revisited

The social distribution of psychological distress and depression in 571 older adults

Longitudinal Evidence for a Midlife Nadir in 573 Human Well-being: Results from Four Data Sets

Locating Economic Risks for Adolescent Mental and Behavioral Health : 576 Poverty and Affluence in Families, Neighborhoods, and Schools

Different indicators of socioeconomic status and 579 their relative importance as determinants of health in old age', International Journal for Equity in 580 Health

Public Mental Health Priorities: 582 Investing in the Evidence', London: Department of Health

Health 2040 -Better Health 584 Within Reach

Including a measure of mental wellbeing in the ONS Wellbeing 586 Framework -Measuring National Wellbeing Technical Advisory Group Paper

Psychiatric morbidity: a multilevel approach to regional 588 variations in the UK

Mental disorders in urban areas: an ecological study of 592 schizophrenia and other psychoses

Ageing in the margins: Expectations of and 594 struggles for "a good place to grow old" among low-income older Minnesotans

Do labour market status transitions predict changes in psychological well-597 being?

Effect of neighbourhood deprivation and social cohesion on mental health 600 inequality : a multilevel population-based longitudinal study

Neighbourhood identi fi cation and mental health : How social identi fi cation 603 moderates the relationship between socioeconomic disadvantage and health

Measuring Depression Over Time . . . or not ? Lack of Unidimensionality and 606 Longitudinal Measurement Invariance in Four Common Rating Scales of Depression

Indicators of socioeconomic position

Housing and health inequalities: A synthesis of systematic reviews of 611 interventions aimed at different pathways linking housing and health', Health and Place

Is the apparent U-shape of well-being over the life course a result of inappropriate 614 use of control variables?

The detection of psychiatric illness by questionnaire: A technique for the 618 identification and assessment of non-psychotic psychiatric illness

Moving to opportunity and mental health: 620 Exploring the spatial context of neighborhood effects

Using Internal Migration to Estimate the 623

Causal Effect of Neighborhood Socioeconomic Context on Health : A Longitudinal Analysis

Using Internal Migration to Estimate the Causal Effect of Neighborhood Socioeconomic Context on 625 Health : A Longitu

Collider bias undermines our understanding of COVID-19 disease risk and 627 severity', medRxiv

Understanding the population structure of the GHQ-12 : 629 Methodological considerations in dimensionally complex measurement outcomes

Understanding the genetic and environmental specificity and overlap 632 between well-being and internalizing symptoms in adolescence

Disparities in the Geography of Mental Health: Implications for Social Work

People, Places and Regions: Exploring the Use of 637

Multi-Level Modelling in the Analysis of Electoral Data

Ethnic variations in mental health among 10 -640 15-year-olds living in England and Wales: The impact of neighbourhood characteristics and parental 641 behaviour', Health & Place

Age group differences in psychological distress: the role of psychosocial risk 643 factors that vary with age

The Mental Health Continuum: From Languishing to Flourishing in Life

Causal and mediating factors for anxiety, depression and well-being

Association between socioeconomic status and the development of 651 mental and physical health conditions in adulthood: a multi-cohort study

Testing measurement invariance across groups: applications in 654 cross-cultural research

The bidirectional relation between positive mental health and 657 psychopathology in a longitudinal representative panel study

Cross-classified multilevel models

Depression and socio-economic risk factors : 7-year longitudinal population 661 study AUTHOR ' S PROOF Depression and socio-economic risk factors : 7-year longitudinal 662 population study

Understanding Society: The UK Household Longitudinal Study Waves 1-664 5 Quality Profile

Are neighbourhood characteristics associated with 666 depressive symptoms? A review of evidence

Understanding Society -The UK household longitudinal study: Wave 1

Mental health and wellbeing in England: Adult Psychiatric Morbidity 673 Survey 2014. A survey carried out for NHS Digital by NatCen Social Research and the Department of 674 Health Sciences

The contribution of housing and neighbourhood conditions to 676 educational inequalities in non-communicable diseases in Europe: findings from the European Social 677 Survey (2014) special module on the social determinants of health

Troubling stoicism: Sociocultural influences and applications to health and 680 illness behaviour', Health:: An Interdisciplinary Journal for the Social Study of Health, Illness and 681

Residential mobility: Towards progress in mobility 683 health research

An empirical comparison of well-being measures used in UK

Collider scope: When selection bias can substantially influence observed 688 associations

Under examination: Multilevel models, geography and 690 health research

Correlates of Mental Illness and Wellbeing in Children

Development and predictors of mental ill-health and wellbeing 694 from childhood to adolescence', Social Psychiatry and Psychiatric Epidemiology

Multilevel analyses of neighbourhood socioeconomic context and 697 health outcomes: a critical review

Local neighbourhood and mental health: evidence from the UK.', Social 702 science & medicine

Method effects associated with negatively and positively worded items 704 item General Health on the 12--Questionnaire ( GHQ-12 ): results from a sectional survey with a 705 representative sample of Catalonian workers

Neighborhood disadvantage and adult depression

Suicide statistics report: Latest Statistics for the UK and Republic of Ireland

Samaritans Suicide Statsistics Report

Geographical gerontology: Perspectives, 713 concepts, approaches

Geographies of ageing : Progress and 715 possibilities after two decades of change

Trajectories of Neighborhood Cohesion in'

Neighborhood Characteristics at Birth and Positive and Negative Psychotic 719 Symptoms in Adolescence: Findings From the ALSPAC Birth Cohort

Bayesian measures of model complexity and fit

Internal construct validity of the Warwick-Edinburgh Mental Well-725 being Scale (WEMWBS): a Rasch analysis using data from the Scottish Health Education Population 726 Survey

Socioeconomic 728 gradients and mental health: implications for public health

Socioeconomic 731 gradients and mental health: Implications for public health

The Warwick-Edinburgh Mental Well-being Scale (WEMWBS): development 734 and UK validation.', Health and quality of life outcomes

Towards a mentally flourishing Scotland: Policy and action plan 736

Ignoring a level in a multilevel model: evidence from UK census 738 data

Life Beyond 65 : Changing Spatial Patterns of Survival at 740 Older Ages in the United States

Geographic Variation in the Prevalence of Common Mental Disorders in 742 Britain: A Multilevel Investigation

Mental well-being and mental illness: Findings from the adult psychiatric 745 morbidity survey for England

Rural/non-rural differences in rates of common mental 748 disorders in Britain: prospective multilevel cohort study

The stability of the factor structure of the General Health Questionnaire

Mental Illness and Mental Health : The Two Continua 753 Model Across the Lifespan

Place and preference effects on the association between 755 mental health and internal migration within Great Britain', Health & Place

Supplementary Table 3: Random part estimates for final model with covariates and complex age term (500,000 iterations)

Supplementary Figure 1: Region level residuals for final model adjusted for covariates and complex age term

Top to bottom, left to right. Fig 1a indicates regional residuals for the average respondent for

SWE. Fig. 1b gives regional residuals for GHQ-12 age slope and GHQ-12 intercept. Figure 1c gives regional residuals for

Figure 1d gives regional residuals for SWE age slope and GHQ-12 intercept. Figure 1e 792 gives regional residuals for SWE age slope and SWE intercept. Figure 1f gives regional residuals for SWE age slope

Supplementary Figure 1a illustrates the consensually poor mental health response in the West 795

the consensually good mental health 796 response in the Rest of Scotland (Scotland excluding Strathclyde and East/Central Scotland), and the 797 average mental health response in West Yorkshire. Figures 1b, 1c, 1d and 1e highlight positive 798 covariance (evidenced in Supplementary Table 3) between response intercept and response age-799 slope, which indicates that areas with poorer mental health overall (on a given metric) tend to see 800 steeper slopes of mental health decline with age. However, as seen in Supplementary Table 3 these 801 are imprecisely estimated and as with complex variance functions in general

Owen for helpful comments on an earlier version of this manuscript. The authors declare no conflict 817 of interest. This publication is the work of the authors and Gareth Griffith will serve as guarantor for 818 the contents of this paper. 819 820. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted July 17, 2020. . https://doi.org/10.1101/2020.07.15.20152645 doi: medRxiv preprint