Mental Health in Context

Effect Heterogeneity and Mechanism in the Relationship between Neighborhood Disadvantage and Adolescent Depression and Anxiety

Kara Rudolph

thank you, tom for that kind introduction and thanks to you all for coming today. my name is kara rudolph ;and i'm going to be talking about my dissertation which is titled [read]

Agenda

Motivation Dissertation overview Presentation of results Discussion

Motivation: Anxiety and Depression

National Comorbidity Survey Replication Adolescent Supplment (NCS-A). Merikangas et al., 2010

My research so far has been motivated by the goal of working to improve the MH of teens in disadvantaged neighborhoods. This graph is from a US nationally representative ...dataset that I used for this dissertation. IT shows the cumulative prevelance of different categories of disorders by age. For this talk, I'm going to be focusing on adolescent depression (comprises most of what's shown in green) and anxiety (which is shown in red) nearly 12% of adolescents had ever had major depressive disorder and nearly 32% ever had an anxiety disorder....

Motivation: Anxiety and Depression

Missed school and sociodevelopmental opportunities $\rightarrow$

Education
Employment
Earnings
Family life

These disorders are not typically isolated, time-limited events, but can have consequence that extend over the lifecourse For example, they cause children and adolescents to Miss school and sociodevelopmental opportunities $\rightarrow$ which studies have shown negatively affects educational achievement (Breslau et al., 2008), future employment (Kessler et al., 2006), earnings (Kessler et al., 2008), family life (Kessler et al., 1998)

Motivation: Anxiety and Depression

Depression: 500 million disability days/year, cost of 36B
Anxiety: 700 million disability days/year
WHO disease burden rankings

MHdisorders in childhood also Increase risk of disorders in adulthood discorders in adulthood have been shown to have signiifcant $\rightarrow$ economic costs [read] in fact, depression and anxiety are the leading contributors to years lost to disability in the WHO disease burden rankings

Motivation: Importance of Place

So, it's clear that MH is imporatnat, but why neighborhood? well, intuitively, we understand that living in a neighborhood in east bmore may be a different experience than living in a neihgborhood in roland park. For example, i taught creative writing to a class of public HS students in DC for a year. theyhad to deal wiht violence on a fairly regular basi-- one of their classmates was tragically even shot and killed. its hard to imagine a classroom of kids from one of the surrounding wealthy suburbs having a similar experience. so living in a disadvantaged neighborhood is thought to entail exposure to more sources stress--e.g., violence, lack of opportunity for work and for education, low-collective efficacy--that act to dysregulate the stress response system within individuals. dysregulation of the stress response system has been linked to depression and anxiety disorders

Motivation: Research Gaps

Inconsistent results in the literature

There has been a lot of research into the relationship between neighborhood factors like disdavatnage and mental disorders--- in the past 10 years in particular--but results have been inocnsistent. I think some of these inconsistencies could be due to methodologic limitations that threaten inference

Motivation: Research Gaps

Inconsistent results in the literature
- Confounding (neighborhood assignment not random)

1. [read]

Motivation: Research Gaps

Inconsistent results in the literature
- Confounding (neighborhood assignment not random)
- Positivity violations (economic and racial segregation of neighborhoods)

2. positivity violations represent another methodologic challenge. Pos viols means that for a particular set of covariate values, assignment of neighborhood is completely defined--there is no randonmess. I'll get into this more later, but one example of where this can come about in neihgborhood research is due to the economic and racial segregation of neihgborhoods in this country. For example...

Motivation: Research Gaps

Inconsistent results in the literature
- Confounding (neighborhood assignment not random)
- Positivity violations (economic and racial segregation of neighborhoods)
- Lack of generalizability (neighborhood studies are typically conducted among a particular sub-population)

3. A 3rd challenge that could contribute to inconsistent results is a Lack of generalizability. means that results estimated in some study samples may not be valid for other samples or target populations. [read] In this dissertation, we addressed each of these methodologic challenges--as least in part

Dissertation Overview

Here is the simplified dag from earlier. In aim 1, we looked at the relationship between ND and prevalent D and A and assessed the extent to which level of urbanicity may modify this effect.

Dissertation Overview

If there is a relationship between ND and D and A, it's been hypothesized to operate through dysregulation of the stress response system. So in Aim 2 we move slightly upstream and estimated the assocation between ND and 1 component of the stress response syste, salivary cortisol levels

Dissertation Overview

In aim 3, we address a problem from Aim 2--that is that we were only able to estimate the association between ND and cortisol among the subsample of participatns for whom we had cortisl data for. however, what we really care about is getting an idea for what the association may be in the US population of adolescents. Now, the association may not be the same if there are multiple effect modifiers which also influence selection into the sub-sample. Broadly, the goal here is to generalize a treatment effect from a survey subsample. we evaluate a few ways to do this in Aim 3

NCS-A

Nationally representative, cross-sectional sample of U.S. adolescents 13-17 years old (N=10,123)
Sources of information relevant to this dissertation:
- Face-to-face, computer-assisted interviews with the adolescent.
  - Composite International Diagnostic Interview (CIDI) modified for adolescents.
- Saliva collected before and after each interview.
- Self-administered questionnaire to parents or parent surrogate of the adolescent.
- Geocoded residence.

For this dissertation, we use data from the NCSA. read first bullet The NCS-A collects information from a variety of sources, but i'll just go through the sources that are relevant to this disseration read through CIDI = Composite international diagnostic interview, corresponds to DSM-IV and ICD-10 diagnoses. saliva was collected before and after each interview for the cortisol samples.

Aim 1

Again, in Aim 1 we examined the pieces that are highlighted in red. We looked at the association between ND and prevalent D and A and assessed possibile effect modification by urbanicity.

Aim 1: Background

Gaps

Inconsistent results
Hypothesis: effect modification by urbanicity may contribute to inconsistent results.

This aim was motivated by the Inconsistent results I found when reviewing the literature on the relationship between neighborhood disadvantage-adolescent D and A. 1. I noticed that Studies... 2. This may be because living in a disadvantaged neighborhood in an urban area entails exposure to a diffrent set of stressors... for ex, violence, lack of green space, crowding that have also been assoc with higher risk of dep/anx 3. so our hypothesis was that urbanicity may be an effect modifier. specifically, that The association between neighborhood disadvantage and adolescent depression/anxiety may be greater in urban versus non-urban neighborhoods.

Aim 1: Background

Study

NCS-A well-suited to evaluate this question.
- Nationally representative survey of U.S. adolescent mental health (DSM-IV mental health diagnoses)
- Geocoded residences, samples include mix of urban, suburban, and rural.

I thought that the NCSA was particularly well-suited to evaluate this question. Because of the sampling scheme, it draws adolescents from neighborhoods across a range of urbanicities. In addition, addresses are geocoded to CTs, so we can obtain contextual information from Census data. The definition of mental disorders is partciularly strong, because the instrument that was used--the CIDI--maps on to DSM-IV diagnoses.

Aim 1: Analysis

Multiple imputation by chained equations
Propensity score subclassification, excluding those without similar counterparts (4.7%)
Survey design-based, weighted analysis

Here is an outline of our analysis approach. [first 2 bullets] This was to address posible positivity violations. That's what I'm showing in the figure on the right. describe fig Say what PS is by excluding those without similar counterparts, we are essentially chopping off the tails. s such that for a given propensity score, there are participants in both categories of neihgborhood. meaning that for a given propensity socre, there is some randomness in neighborhood assingment.

Aim 1: Analysis

Multiple imputation by chained equations
Propensity score subclassification, excluding those without similar counterparts (4.7%)
Survey design-based, weighted analysis

PS subclassification also addresses Addresses confounding stemming from non-random assignment of neighborhood residence. I'm showing a one-dimensional summary of how well this worked on teh right. blue dots are pre-subclass.red dots are post-subclas.all of our potential confounding variables are on the y axis and the standaridized mean difference is on the x axis. so this is showing how balanced the two neighborhood groups are for each of the confounding variables. We can see by the blue dots that before PS subclassification, many of the variables were imbalanced. Afterwards, though, the red dots show us that the subclassification was successful in balancing these variables. Within each of the PS subclasses, for each imputed dataset, we then conducted Survey design-based, weighted analysis to maintain the national representativeness of the sample.

Aim 1: Results

Log odds ratios of prevalent depression or anxiety in disadvantaged versus non-disadvantaged neighborhoods.

here are our results from Aim 1. y axis shows the log odds ratio of having prevalent Dep or Anxiety comparing adolescents living in Dis Neigh to nonDis neihg. x axis estimates this log OR for each the 3 urbanicities: non-urban, urban fringe and urban center. see a dose response relationship where the log OR increases by level of urbanicity.

Aim 1: Results

Log odds ratios of prevalent depression or anxiety in disadvantaged versus non-disadvantaged neighborhoods.

for adolescents in urban centers, living in a disadvantaged neighborhood is associated with have 59% (1.25, 2.03) greater odds of having a prevalent anxiety/depression as compared to living in a non-disadvantaged neighborhood

Aim 1: Results

Log odds ratios of prevalent depression or anxiety in disadvantaged versus non-disadvantaged neighborhoods.

the interaction between urban center and neighborhood dis is also significant. Specifically, the association between neighborhood disadvantage and prevalent anxiety/depression is more than twice as large (ratio of odds ratios: 2.08, 95% CI: 1.23, 3.55) for adolescents living in urban centers versus non-urban areas

Aim 2

slow down As a reminder, in Aim 2 we examined the components of our conceptual model that are highlighed in red. We moved upsteam and estimated the association between ND and one aspect of the stress response system: salivary cortisol.

Aim 2: Background

Gaps

Few studies
Small sample size
Racial/ethnic homogeneity
Geographic homogeneity

In Examining the association between ND and cortisol, we were motivated by several gaps in the literature. First, there has been surprisingly little research into this realtionship in adolescets. In fact, in my literature review, I only found 3 studies. However, each of thes 3 studies were Limited by small sample size and racial/ethnic and geographic homogeneity. for example the biggest study only included 163 teens (163, 100, 79) 2/3 studiies were only of african americans and all 3 of the studies only sampled from single urban areas (flint, St Louis, Philadelphia)

Aim 2: Background

Study

Largest sample to date (2,490 have cortisol measurements).
Racially and ethnically diverse.
Participants drawn from different regions of the U.S., urbanicities. Again, I thought the NCS-A was well-suited to address these gaps. [read]

Aim 2: Background

Relationship between neighborhood sources of stress and cortisol levels is complex

Hypothalamic-pituitary-adrenal (HPA) axis dimension
Blunting
Time of day

(Do et al., 2011.)

Before I get into the specifics of our analysis, I want to convey that the stress response system is extremely complex, and so Examining the relationship between neighborhood sources of stress and cortisol involves requires stratgies to manage this complexity. 1. numerous systems involved in stress response. cortisol is a component of one of these systems called the ... The relationship between ND and cortisol likely depends on the dimension measured. the two (1) resting, unprovoked levels and (2) and reactivity to an acute stressor 2. Seoncd, the relationship between ND and cortisol also likely depense on the History of exposure to stress. With moderate amounts of stress, resting cortisol levels increase and reactivity also increases. But among those who have been exposed to really severe and persistant stress---children who are abused---the cortisol levels and reactivity actually are blunted---look more similar to unstressed children.This is called blunting, because the idea is that after repeated or prolonged activation of the stress response system, this response wears out. 3. time of day. Graph: diurnal rhythm. explain CAR. and the relationship between enighborhood sources of stress may differ depending on time of day. explain graph

Aim 2: Analysis

Challenge Approach Not at-risk for a cortisol response Restrictive exclusion criteria Blunting Sensitivity analysis Heterogeneity by CAR/ post-CAR Restricted to post-CAR Diurnal rhythm Propensity score matching with Mahalanobis matching Other strong predictors Propensity score matching with exact matching given this complexity, cortisol is ideally measured under very strict laboratory protocols, but this isn't practical in a large, nationally representative sample like the NCSA we were left with the challenge: How to construct a controlled analysis? (first part: not at risk for cortisol being influenced by environmental sources of stress) when say blunting talk about possible washout effect strong predictors like sample time, etc. combined ps matching with matching on the subset of these especially important predictors. explain.

Aim 2: Analysis

Cortisol outcomes

Pre-interview
Post-interview
Slope over the course of the interview

There were 3 cortisol outcomes that we examined in this analysis. [read the 3] These don't correspond to specific HPA axis dimensions, since cortisol wasn't measured under controlled conditions with standardized protocols, so I'm going to take a second to describe what these outcomes may reflect. Pre invterview levels likely reflect any activity the adolescent was engaged in prior to the interview, foods, as well as anticipation of the new situation of being interviewred in her home by a researcher for ta survey on mental health. Something out of the ordinary! And relevant, because the HPA access is particularly sensitive to novelty. The interview was specifically designed not to be stressful, so adoelscents probably became pretty bored over the course of the interview. The post-interview levels may relfect that as well as the fact that now the adolescent has been sitting down for 2 and a half hours. We also looked at slope over the course of the interview-- post - pre/ interview time. this can be thought of as a measure of recovery from this novel interview situation.

Aim 2: Analysis

Multiple imputation by chained equations
Regression outcome analyses on matched, imputed datasets
- Gamma regression for point-in-time measures
- Linear regression for cortisol rate

Here again is an outline of our analysis approach. [MICE] We then performed PS matching with replacement combined with matching on the subset of especially important cortisol predictors as described on the previous slide. [describe figure] Then, we performed regression outcome analyses on the [read] This approach can be thought of as being double robust, because the potential confounders are included in both the PS model and outcome model. so if one model is incorrectly specified but the other is correctly specified, we should still get a consistent estimate