Place and preference effects on the association between mental health and internal migration within Great Britain

&NA; Individuals with mental health needs are more likely to migrate than the general population, but the effects of migration preference and place of residence are often overlooked. These issues are addressed through the application of a novel origin and destination multilevel model to survey data. In comparison to those with good mental health, individuals with poor mental health are more likely to make undesired moves and this is moderated, but not explained by place of residence. Implications for understanding the mental health and migration relationship, and its impact on service provision are then proposed. HighlightsMental health needs are consistently associated with internal migration.Past research overlooks the importance of migration preference for this association.Mental health is associated with moving only among those who prefer not to migrate.This finding is independent of where individuals live.This displacement may have role in explaining geographies of mental health.


15
Poor physical health has been shown to be associated with low likelihoods of 16 internal (within-country, over any distance) migration in Europe (Westphal, 2016), 17 Northern America (Curtis et al., 2009) and Australia (Larson et al., 2004). Less 18 attention has been paid to the influence of mental health on migration behaviour. 19 In contrast to physical health, internal migrants are more likely to self-report 20 mental health problems than non-migrants (Larson et al., 2004;Tunstall et al., 21 2014). Extant research is primarily drawn from populations with severe and rare 22 mental health conditions (Harvey et al., 1996;Ngamini Ngui et al., 2013), 23 although analyses using instruments designed to measure common mental 24 disorders find similar associations between moving and mental health (Tunstall et 25 al., 2015). Although the mental health of internal migrants is well studied, the 26 majority of research compares the health of recent internal migrants to that of 27 non-movers, so it is unclear whether mental health affects the likelihood of 28 migration, or migration affects mental health. 29 2 The desire to migrate or stay (migration preference) and ability to meet this 30 preference may confound the relationship between mental health needs and high 31 rates of internal migration, and Great Britain (GB; England, Scotland and Wales) 32 provides an interesting case study to test this hypothesis. There is evidence of 33 undesired staying (i.e. not moving when one would like to) and undesired 34 migration (i.e. moving when one would not like to) among the population of GB 35 (Coulter and van Ham, 2013). Mental health needs are associated with a desire to 36 migrate regardless of whether an individual has recently moved, but not with 37 undesired migration. In addition, undesired staying and undesired migration are 38 associated with worsening mental health over time, after controlling for baseline 39 mental health (Woodhead et al., 2015). Mental health status may act as a barrier 40 to realising migration preferences, as mental health problems are associated with 41 relatively low levels of psychosocial resources, educational attainment, 42 employment and financial capital (Fryers et al., 2003;Weich & Lewis, 1998), all 43 factors that are drawn upon in the search for alternative residences (Lee, 1966). A 44 realistic estimation of the influence of mental health on internal migration must 45 control for interactions with migration preference, but this relationship is largely 46 overlooked in the literature. 47 In addition to ignoring mental health associations with migration preference, 48 place of residence effects are rarely accounted for in migration literature (Thomas 49 et al., 2015). Previous (origin) and current (destination) place of residence likely 50 moderates (i.e. affects the strength of) the association between mental health and 51 migration. Individuals with mental health needs have been found to migrate into 52 deprived and urban areas in GB shortly before the onset of severe mental health 53 problems (Harvey et al., 1996;Ngamini Ngui et al., 2013;Taylor, 1974). This has 54 been explained through the social selection or 'drift' theories, where the onset of 55 mental health problems leads to reductions in earning capacity or unemployment, 56 and then a reduced ability to remain in or move to affluent neighbourhoods (Lowe 57 et al., 2014). In the context of rising house prices and rental rates in GB over the 58 1990s and 2000s (Dorling, 2015), we might expect individuals with mental health 59 needs may be less able to afford to stay in desirable homes and neighbourhoods, 60 and less able to afford to move out of undesirable homes and neighbourhoods 61 (Smith & Easterlow, 2005), in comparison to the general population. Such place 62 moderation effects have been observed for physical health limitations, where the 63 overall positive association between good physical health and migration was 64 reversed in the Midlands of England in the 2011 Census (Wilding et al., 2016).
3 When place effects are explored, the characteristics of the place of residence 66 post-move (destination) are usually used. The dominance of destination effects is 67 challenged by established migration models such as the gravity model 68 (Flowerdew & Aitkin, 1982) and developments in multilevel modelling showing 69 that it is important to consider previous and current place of residence in 70 migration models (Thomas et al., 2015). Specifically, the association between 71 mental health and migration may differ for an area as an origin and destination 72 respectively, as in the 'drift' framework we would expect mental health to be 73 associated with moves into deprived urban areas (destination), but low rates of 74 moves out of these areas (origin). 75 In summary, individuals with poor mental health are more likely to become 76 internal migrants (over any distance) than the general population. This 77 association is confounded by migration preference, as those with poor mental 78 health are more likely to want to move, and wanting to move appears to be 79 harmful to mental health. The extant evidence fails to adequately account for the 80 potential moderation effect of place on this relationship, and there are theoretical 81 reasons for expecting the relationship between mental health and migration to 82 vary by area. The aims of this study are to test i) if poor mental health is 83 associated with internal migration ii) if the association between poor mental 84 health and internal migration differs between those who prefer to move, and 85 those who prefer to stay and iii) if the association between poor mental health 86 and internal migration varies by place of origin and destination. The rest of this 87 paper addresses these issues, using data from two major surveys, utilising a 88 cross-classified multilevel model to test whether mental health predicts internal 89 migration, and if this explained or moderated by origin, destination and 90 migration preference effects. 91 centroids were provided at a 100-metre resolution (Martin, 1993). Centroids later 139 became available at a 1-metre resolution (Rabe, 2009). Internal migrants are 140 defined as individuals whose grid reference at time t and t-1 differ by more than 141 100 metres, if the pair of grid references are identical or differ by 100 metres or 142 less then the observation is coded as a non-mover. A 100-metre cut-off is used as 143 this is the coarsest resolution for postcode grid references found in the postcode 144 directory over the study period, and it is assumed that postcode adjustments over 145 consecutive waves are unlikely to be of greater distances than 100 metres. 146

Mental health 147
The 12-item General Health Questionnaire (GHQ) is used to measure mental 148 health status in this analysis. The GHQ was designed to measure the risk of 149 common mental disorders in observational studies (Goldberg, 1978). Each item 150 has four possible answers in a Likert scale design. Responses in the two lower 151 categories are coded as 0 for each item, and the two higher categories are coded 152 as 1. This coding system is known as the 'GHQ method' (Hankins, 2008). The 153 sum of item scores is calculated (with a minimum of 0 and maximum of 12); 154 sums of 3 or more are considered to be indicative of poor mental health, and 155 sums less than 3 are indicative of good mental health (Shelton & Herrick, 2009). 156 The 12-item GHQ has been shown to be a strong predictor of common mental 157 disorders in a range of contexts, and is robust to gender, age and educational 158 differences in reporting of symptoms (Goldberg et al., 1997). In line with past 159 research, individuals with poor mental health (as measured by high GHQ scores) 160 are expected to be more likely to move than those with good mental health 161 (Larson et al., 2004), and this association will differ in strength between those 162 who prefer to move, and those who prefer to stay (Woodhead et al., 2015). between those who respond with 'don't know' and 'stay here' (Coulter & Scott,195 2014; Woodhead et al., 2015). The 'don't know' preference category is separated 196 in this analysis to control for ambiguity in preference, as there are complex 197 7 processes involved in shaping migration preferences which have implications for 198 later mobility (Lu, 1998

Equation 1 Model structure 239
In this framework, migration is predicted at occasion i for individual j living in 240 destination LA k at t and origin LA l at t-1. y * is the estimate for the predicted 241 probability of moving according to the cumulative distribution, such that when 242 y * =0 the predicted probability is 50%. Values of y * greater than zero indicate a 243 greater than 50% probability of moving, and the opposite is true for values less 244 than zero. 0 is a fixed constant, is the vector of covariates outlined in table 245 1 which are measured at time t-1, ℎ ℎ is a fixed effect associated with 246 having poor mental health at time t-1, is a fixed effect 247 associated with migration preference at time t-1, and interaction terms between 248 mental health and migration preference are included. The interaction terms 249 estimate the additional effect of having poor mental health on the probability of 250 migration for those who prefer to move or don't know their preference. 251 The individual-specific random intercept is given by the parameter 0 . 252 The destination-specific random intercept is given by the parameter 253 0 ( +1) , and an additional slope for individuals with poor mental health 254 at time t-1 is given by the parameter 1 ( ) ; these two parameters are 255 also estimated at the origin level ( 0 ( −1) & 1 ( −1) ). The random effects 256 approach is used, where the random effects ( ) are assumed to be normally 257 distributed, have a mean of zero and a constant variance. The variance of each 258 parameter ( 2 ) and the covariance between intercepts and slopes 259 In order to answer the third research question (whether the association between 275 poor mental health and internal migration varies by place of origin and 276 destination), the ratio for the probability of migration by mental health is 277 calculated by each LA as an origin (the probability of future migration) and 278 destination (the probability of having moved). The predicted probability of 279 migration for the population with good and poor mental health in each origin LA 280 is calculated using the random intercept ( + 0 ) for the former, the intercept 281 and slope ( + 0 + 1 ) for the latter. The ratio of probabilities for the 282 population in poor mental health, relative to the population in good mental health 283 is then calculated (termed the 'mental health migration ratio') and this ratio is 284 compared over the percentage of the population with good mental health 285 predicted to move. This process is repeated for each destination LA ( + 0 ) 286 and ( + 0 + 1 ). 287

288
The first aim of this analysis was to test if poor mental health is associated with 289 internal migration. In the cross-tabulation (table 2) the overall between-wave 290 migration percentage is 9.2%, the percentage for the population with good mental 291 health is lower than this average (8.5%) and it is higher than average among the 292 population with poor mental health (11.3%). There is significant evidence for this 293 association, according to the chi-square statistic ( 2 = 330.9 df = 1, p<.01 Access datasets. Good mental health is defined as General Health Questionnaire summary scores of 0-2, and poor is a score between 3 and 12. Author's own calculations. Table 2 Tabulation of mental health and migration status 296 Table 3 shows the results for a CCM, predicting the probability of migration by 297 mental health and migration preference, accounting for all control variables. The 298 inclusion of the two interaction terms between mental health and migration 299 preference led to a 31 unit decrease in the DIC, suggesting that the interaction 300 terms improve the overall model fit (results not shown). Holding all other factors 301 constant, those with poor mental health are more likely to move (an increase in 302 the z-score probability of moving of 0.162, 95% credible interval 0.125 -0.199) 303 13 than those with good mental health. Expressed as percentages, 11.3% of those 304 with poor mental health are predicted to move, compared to 8.5% of those with 305  Table 3 Cross-classified probit model predicting the probability of moving 308 between survey waves 309

Interaction effects 310
The second aim of this analysis was to test if the association between poor 311 mental health and internal migration differs between those who prefer to move, 312 and those who prefer to stay. The interaction terms between mental health and 313 migration preference in table 2 represent the additional change in the z-score for 314 the probability of migration among those with poor mental health within that 315 specific migration preference group. As both interaction terms are negative, this 316 indicates that the association between mental health and migration is less 317 positive among those who prefer to move or don't know their migration 318 preference, compared to those who prefer to stay. The probabilities of migration 319 by mental health and migration preference are then calculated in MLwiN's 320 prediction window, with simulated 95% confidence intervals (figure 2). This figure  321 displays that mental health is associated with migration only among those who 322 prefer to stay, providing evidence of confounding. Non-response analysis 347 Non-response (not participating in a survey wave) and attrition (permanent non-348 response) have the potential to affect the generalizability of findings from panel 349 survey data, if population subgroups are particularly likely to not respond 350 (Mostafa & Wiggins, 2015). In the BHPS and USoc, however, there is no prior 351 evidence that GHQ scores are associated with non-response, although internal 352 migration and preferring to move are (Lynn et al., 2012;Uhrig, 2008). As a result, 353 non-response is unlikely to affect estimates of the association between mental 354 health and migration in this analysis, unless there is a relationship between 355 mental health, migration preference and non-response. In our own analysis 356 (results not shown), those who prefer to stay and have a low GHQ score more 357 likely to respond in the following survey wave (95% CI 91.5%-91.8%) than those 358 who prefer to move and have a high GHQ score (95% CI 87.6%-89.6%). As a result, 359 selective attrition may explain the lack of difference in migration probabilities 360 between those who prefer to move and have high and low GHQ scores. 361

362
This analysis set out to test three research questions: i) if poor mental health is 363 associated with internal migration; ii) if the association between poor mental 364 health and internal migration differs between those who prefer to move, and 365 those who prefer to stay and iii) if the association between mental health and 366 internal migration varies by place of origin and destination. The findings for each 367 research question are discussed in turn. 368 In the cross-tabulation (table 2), poor mental health was associated with a greater 369 probability of migration, and this association persisted after controlling for 370 potential confounders in the probit model (table 3). This finding corroborates 371 with previous research indicating that common predictors of migration do not 372 explain the association between mental health and internal migration (Tunstall et 373 al, 2014). 374 The overall effect of poor mental health appears to differ by migration preference 375 in this analysis (research aim 2), however. Through interaction terms (figure 2), 376 we find that mental health is only associated with migration among those who 377 prefer not to move (displacement), not for those who prefer to move (desired 378 migration) or those who do not know their migration preference. There are 379 several plausible mechanisms behind the elevated probability of undesired 380 migration among the population in poor mental health shown here, the 381 identification of which lie outside the scope of this paper. Drawing on the place 382 utility framework (Lee, 1966), individuals in poor mental health may be drawn 383 away from areas where they prefer to stay in order to gain greater access to 384 healthcare (Moorin et al, 2006), or in order to escape discrimination (Lewis et al, 385 1992). Alternatively, those with poor mental health may be being priced out of 386 desirable homes through rising rental rates (Dorling, 2015). Quantitative analyses 387 can inform on what is happening and where, but complementary person-focused 388 research is needed to understand why such processes occur. Collaborative work 389 with mental health needs groups is required to assess the challenges related to 390 retaining residence faced by those with mental health needs to further 391 understand the elevated rates of undesired migration among this group. 392 The third research aim explored whether the association between poor mental 393 health and migration varied by place. The ratio for the probability of moving 394 between those in poor and good mental health was consistently positive in all 395