Assessment of the Role of IQ in Associations Between Population Density and Deprivation and Nonaffective Psychosis

This cohort study investigates whether lower IQ contributed to the association between population density or deprivation and nonaffective psychosis and whether these associations were stronger in people with lower IQ among Swedish military conscripts.

The IQ tests for Swedish conscripts have been described in detail previously. 1 The Swedish military service conscription examination involves a medical assessment, including general intelligence, with IQ scores summed using four scales to assess verbal, logical, spatial, and technical abilities. Four tests were done measuring verbal, logical, spatial, and technical abilities. Each test gave a score ranging from 1-9 and these tests were summed to give a total general intelligence score.
The verbal IQ test lasted seven minutes and had 40 questions. Participants were given rows of five words and told to underline the odd one out.
The spatial test took 40 minutes. Participants were shown a focal geometric shape. They were then four other shapes of different sizes and orientations and asked which of these shapes could be used to form the focal geometric shape they had originally seen.
The logic test comprised tests of general knowledge and intelligence and had 40 items and took 12 minutes. Participants were presented with a combination of shapes and letters. They were asked questions such as "put a line through the square under the longest word." There were also asked questions on general knowledge and mathematics.
The technical test required a basic knowledge of physics and mechanical ability. For We adjusted for paternal age given prior evidence of a strong positive association with offspring psychosis, 2 and theoretical grounds that it may be associated with our exposures.
We used a binary variable for parental history of severe mental illness, which was coded 1 if either parent had ever been diagnosed with non-affective psychosis or bipolar disorder since 1973, when the National Patient Register began to collect data on psychiatric admissions (ICD-9 codes: 295.x, 296.x; ICD-10 codes: F20-31). 3 Parental time in education used the following categories: (less than 9 years compulsory education; 9 years compulsory education; secondary education; post-secondary education for less than 2 years; post-secondary education for 2+ years; doctorate education) Family disposable income when the offspring was born was based on information in the LISA on total family income including wages, welfare benefits, other social subsidies, and pensions. We categorized family disposable income into quintiles, relative to all other people in Sweden in the given year to account for inflation. 4 Migrant status (Swedish-born to two Swedish-born parents, or Swedish-born to at least one foreign-born parent) was coded from the Register of the Total population and the Multigeneration register.
For each participant we obtained their parents time in education as recorded in the National School Register, and classified this as less than 9 years compulsory education, 9 years compulsory education, secondary education, post-secondary education for less than 2 years, post-secondary education for 2+ years or doctorate education. Where available, paternal IQ (from earlier conscript data) was also recorded.

Missing data
To establish whether missing data were associated with our exposures and outcome (and might therefore introduce bias), we compared the characteristics of those with and without complete data. We also conducted a sensitivity analysis using multiple imputation with chained equations (MICE). We assumed data were missing at random and imputed 20 datasets, combined using Rubin's rules. To impute missing data, we used conscripts' educational attainment at age 16, all characteristics described and several auxiliary variables

Associations of population density and deprivation with non-affective psychosis
We used multilevel logistic regression with individuals (level 1) clustered in SAMS regions (level 2), and a random intercept at the SAMS level. First, we ran separate univariable models for associations between each exposure and non-affective psychosis. Second, we ran a bivariable model including population density and deprivation. We then adjusted this model for potential confounders.

Population attributable fraction
We calculated the population attributable fraction (PAF) for the main exposures, deprivation and population density. In reality these exposure are continuous and it would be arbitrary to create a category for whether people were exposed or unexposed. Consistent with a previous study, we therefore considered this a theoretical exercise and calculated the PAF as an estimate of the proportion of non-affective psychoses that could be prevented if we could identify and remove all factors that lead to increased incidence associated with deprivation and population density. 3

Modification by IQ
We tested for interactions first, so that any interaction terms significant at an alpha level of <.05 could be modelled in subsequent mediation analyses. We first added conscript IQ to the fully-adjusted model described above. Next, we sequentially fitted two interaction terms to test whether the associations between population density and deprivation at birth respectively, and subsequent risk of non-affective psychosis, were modified by IQ. Effect modification was assessed via likelihood ratio test (LRT). We conducted power analysis via simulations of these two interactions on non-affective psychosis, based on our analytic sample size for the (i) observed interaction effect sizes (odds ratios (OR)), and ORs of (ii) 1.10 and (iii) 1.20 (see

Supplemental Materials for full details).
Mediation by IQ Figure 1 shows our hypothesized mediation model. To determine whether population density or deprivation at birth were associated with IQ scores at age 18 years (Path A, Figure 1) we used multilevel linear regression. We tested whether IQ was associated with non-affective psychosis (Path B, Figure 1) using multilevel logistic regressions. Path C was assessed above.
All models were run before and after adjusting for confounders. including paternal IQ in a sensitivity analysis. We formally tested for mediation using the "potential outcomes" framework (see below), 32 a class of causal mediation analysis fitted using parametric mediation models. 30,31

Sensitivity analyses
Consistent with a previous study 22 and to reduce the possibility that IQ at age 18 captured a prodromal effect of non-affective psychosis, we re-ran analyses of the association between IQ and non-affective psychosis after excluding participants diagnosed with non-affective psychosis within two years of conscription. We also conducted sensitivity analyses to additionally adjust for paternal IQ, which was only available on a subset of participants.. Sensitivity analyses with missing data replaced by multiple imputation were reported for all associations except the causal mediation analysis, since multiple imputation approaches in this context are not yet routinely available. All analyses were done in Stata version 15. eMethods 2. Mediation study, we calculate the 'total effect' of population density/deprivation on non-affective psychosis, which comprises a 'direct effect' and an 'indirect effect' (the latter is an estimate of how much of the association between population density/deprivation and non-affective psychosis is mediated by IQ). The total effect estimates the 'average causal effect (ACE)' or change in outcome status if everyone moved from unexposed to exposed (E{Y(1) -Y(0)}).
We estimated the 'controlled' direct effect, which compares outcomes under exposure levels A=1 versus A=0, fixing M=M(0), where A=exposure and M=mediator. By fixing M to 0, the controlled direct effect estimates what the association between exposure and outcome would be if the mediator had the same value in those exposed and unexposed (i.e. if changes in the mediator were not induced by exposure). The natural indirect effect compares outcomes under the counterfactual scenarios where M=M(1) versus M=M(0), fixing A=1 (everyone is exposed, and the outcome is compared between those who do and do not experience the mediator). We expressed indirect effects as a percentage of the total effect by log transforming the odds ratios and dividing the indirect effect by the total effect.

eMethods 3. Auxiliary Variables Included in Multiple Imputation Models
We obtained data on maternal smoking during pregnancy, obstetric complications (OC) and early life infections in infancy (0-12 months) from the MBR and NPR. Maternal smoking during pregnancy was recorded at the first antenatal visit and recoded as "none", "1-9 cigarettes per day" or "10 or more cigarettes per day". We considered the following major obstetric complications, based on evidence from a previous systematic review that they were associated with future risk of schizophrenia, 5 or evidence that they were associated with childhood cognitive function 6 : maternal diabetes (yes/no), non-elective Caesarian section (yes/no), any maternal hypertension disorders during pregnancy (including pre-eclampsia) (yes/no), small for gestational age, 1 minute Apgar score less than 7 (rated 0-10), asphyxia © 2020 Lewis G et al. JAMA Psychiatry.
(yes/no) and congenital malformations (yes/no). Full details and ICD codes are given in Supplemental Table 8. We identified early life infections resulting in hospitalization in the first year of life, as recorded in the NPR. We included the same set of infections as used in a major previous investigation from our group (see Supplemental Table 9 for ICD codes), 7 and coded participants according to the number of times they received a diagnosis for one of these infections in the NPR, categorized as 0, 1, 2 or 3+ diagnoses. We also included the number of small area (SAMS) residential moves in childhood and adolescence as auxiliary data, which has been previously associated with risk of non-affective psychosis in this sample, and may inform missing covariate data in this study. Full details of the derivation of these variables are described elsewhere, 4  for correlations between population density, deprivation and IQ and were fitted in a generalized linear model with a binomial distribution and logit link function. Alpha was set to 0.05 and 500 Monte Carlo replications per simulation. Power simulations were fitted using the user-written powersim command in Stata, based on the full methodology described by Luedicke. 12 While we were underpowered to detect interaction effects of the magnitudes observed (Supplemental Table 12), we had over 98% power to detect interaction odds ratios greater than 1.05 and 100% power for effect sizes greater than 1.10 (Supplemental Table 12).