Studies on the influence of neighborhood environments (NE) (i.e., social and built environments (BE)) on brain health are still in their infancy but are growing rapidly and provide tentative evidence that our community environments may affect the brain throughout the lifespan [1
]. Greater neighborhood socioeconomic disadvantage has been associated with worse baseline cognition [4
], greater decline in cognition over time [6
], and total and regional brain volumes from magnetic resonance imaging (MRI) [7
] among older adults. Neighborhood racial/ethnic segregation has been linked to poorer cognitive health outcomes in middle and older aged individuals [8
]. In addition, measures of the BE such as greater land use mix (e.g., mix of retail and residential) [11
], access to retail destinations [12
], public transportation availability [13
], greater walkability [14
] (i.e., environment conducive to walking by providing multiple destinations and density of street connections), and greater greenness/park space access [15
] have been associated with various measures of brain health in older adults including diagnoses of Alzheimer’s disease and related disorders (ADRD). This body of research is typically rooted in the socioecological framework that posits that beyond individual level determinants of brain health, including age and genetics, there are likely higher-level social determinants of health (SDOH) operating, including neighborhood and community environments. SDOH affect environmental exposures and access to support, resources, and opportunities that ultimately affect a population’s morbidity and mortality.
Yet, even the most rigorous studies of the NE and brain health to date, such as those that employ population-based cohorts with longitudinal follow-up or natural experiments, can still be biased due to attrition (e.g., individuals with the outcome of interest dropping out at a higher rate), non-representativeness of the sample compared to the general population, and competing causes/residual confounding. Table 1
presents some of the methodological challenges unique to studying NEs and brain health, such as defining the neighborhood geographic boundary, capturing the neighborhood exposure, defining the neighborhood construct, typical reliance on studies of older adults, and the lag between development of pathology for ADRD and the diagnosis of dementia. We first present these broader issues to set the context for our more detailed focus on self-selection and reverse causation. Reverse causation occurs when the outcome precedes and results in the exposure (Figure 1
a). Cross-sectional studies are prone to this potential issue because temporal ordering of exposure and outcome cannot be established. This may also be an issue for studies on outcomes with long preclinical or subclinical periods. Reverse causation in which changes in brain health predict the neighborhood one resides in is an example of self-selection. However, self-selection can also bias NE-brain health studies not through reverse causation, but through confounding, in which the lifestyle or neighborhood preferences influence residential choices (Figure 1
b). Any association between the NE and brain health may instead be attributable to lifestyle behaviors and preferences that preceded neighborhood choice and that are also associated with the brain health outcome.
Studies of brain health including ADRD risk are particularly susceptible to reverse causality bias because older adults are likely to move to new neighborhoods following the development of cognitive or physical impairment including dementia [19
]. Thus, any resulting association between a particular NE and cognition or dementia risk may be due to these late life residential moves. In addition, ADRD neuropathology and subtler cognitive changes can occur a decade or more prior to a diagnosis of ADRD, further complicating matters. In these cases, researchers may presume that associations between NEs and ADRD risk were free of bias from neighborhood self-selection because residential environments were measured prior to disease development. However, ADRD may more subtly affect brain health and functioning up to two decades prior to a dementia diagnosis [20
]. Sometimes, more than a year can occur between full-blown dementia symptoms and the receipt of a diagnosis [22
]. Thus, residential moves in mid to late life may be in part due to these subtler and longer-running changes in brain health.
This paper aims to review methods for addressing reverse causation and self-selection in the context of the broader challenges of research on NEs and brain health. We provide a cursory overview of available methods to address reverse causation and self-selection, example studies that have employed these methods, and two case studies. The goal of the paper is not to provide specific instructions on how to conduct the outlined methods, but to provide a summary of the methods with useful references and considerations to guide future research on NEs and brain health.
2. Methods to Address Reverse Causality and Self-Selection
The goal of research is typically to identify causal effects (e.g., what NE factors improve brain health and what factors hasten cognitive decline). These can then inform interventions or policy changes to improve wellbeing. However, often we deal with statistical associations and findings from observational studies that may be impacted by a number of biases including self-selection and reverse causation. Based on the idea of counterfactual models, i.e., that there are unobserved potential outcomes for each observed outcome, the goal of causal inference would be to a compare observed to counterfactual outcomes among the same individuals. Since this is not possible, we must substitute other individuals to approximate the counterfactual outcome. If the individuals in the two groups are not similar along other parameters, then this will introduce bias. The goal of many research strategies is to make exposed/unexposed groups similar along potentially biasing factors.
In this section, we describe various approaches that can be used to enhance causal inference in studies of NEs and brain health, such as propensity score weighting and natural experiments (Table 2
). Since few studies of the NE and brain health have incorporated these methods, we provide examples from the published literature on other health outcomes (e.g., physical activity).
2.1. Randomized Control Trial/Experiment
Randomized control trials (RCT) are considered the gold standard study design for estimating causal effects [26
]. RCTs employ randomization to help balance potential confounders across intervention groups and also establish clear temporal order of cause and effect. An example of a neighborhood-based RCT was the Moving to Opportunity trial in which individuals from high poverty neighborhoods in public housing were chosen randomly to move into either neighborhoods of high or low poverty. In one published study of the participants, families and children in that study were interviewed three years after their moves [27
]. Those who moved to neighborhoods of low poverty experienced less distress (parents) and anxiety/depressive symptoms (boys). However, RCTs may not be feasible for understanding neighborhood effects due to their costs, the ethical implications, and the difficulty of recruiting participants (or neighborhoods) for interventions. If conducted well, RCTs have low chance of bias by self-selection and reverse causation, but use of RCTs has been extremely rare in NE and health research due to the hurdles to implementation. Thus, we would not propose the use of a RCT given the difficulties for NE and health studies, but instead present it as the gold standard method with which other methods are generally compared.
2.2. Multivariable Regression—Covariate Adjustment
The most common approach to address self-selection is through covariate adjustment in multivariable regression models [28
]. Participant or neighborhood characteristics that are known or theorized to influence neighborhood choice are included as additional predictors in regression models. Such variables could include demographics (e.g., age, education, marital status, race/ethnicity) or measures that describe or relate to choice of neighborhood (e.g., preference for urban vs. rural or ranking of importance of neighborhood characteristics) [29
]. For example, a study of individuals from eight neighborhoods in Northern California examined cross-sectional associations between neighborhood BE characteristics and frequency of walking to the store, while controlling for self-reported neighborhood preferences when determining where to live (i.e., accessibility, safety, physical activity options, socializing, outdoor spaciousness, and attractiveness) [30
]. The authors found that BE characteristics including distance to destinations were significantly associated with walking to the store even after controlling for neighborhood self-selection characteristics.
Covariate adjustment in regression models may help capture key confounders, but rests on the assumption that confounders are accurately measured and included in the models. Capturing participant preferences and factors that accurately determine self-selection into neighborhoods may be challenging. As noted previously, measuring neighborhood preferences may not capture all facets of self-selection into neighborhoods because individuals may have difficulty articulating and prioritizing their reasons and preferences [31
]. Furthermore, adjustment in regression models does not easily account for potential reverse causality of brain health influencing neighborhood choice.
2.3. Multivariable Regression—Propensity Scores/Inverse Probability Weighting
Propensity scores employ a two-step process to control for confounding. First, probability (propensity) for exposure is modelled in a first stage and scores derived from predicted probabilities are created to include in the second stage, or primary analytic model. Various techniques for propensity scores have been developed, some of which include matching based on propensity score, and others the exclusion of outlier propensity scores. Several studies on the NE have used propensity scores [32
], as one advantage of this technique is that many covariates can be included when developing the propensity scores, and yet power will not be sacrificed in the full models. This may allow for better control of confounding by self-selection but still relies on similar assumptions as a traditional multivariable regression model.
Another analytic strategy to improve causal inference is to use weighting techniques based on propensity scores to account for potential biases. In some cases, adjustment may lead to overadjustment (e.g., if potential confounders are also potential mediators). In other cases, regression adjustment may lead to bias, such as adjusting for variables that are common effects of both NE and brain health (or their confounders) (aka collider-stratification or selection bias), or in cases of effect modification of confounders [34
]. Therefore, weighting techniques can be used to remove potential sources of bias not amenable to covariate adjustment techniques. Inverse probability weighting (IPW), inverse probability of treatment weights (IPTW), and marginal structural models use a two-stage process to first estimate the effect of important confounders on probability of treatment/exposure and then incorporate the inverse predicted probabilities into analytic models [36
]. Weights can be defined to account for time-varying confounding [37
], selection bias (e.g., modelling probability of selection instead of treatment) [38
], and missing data [39
]. Weights are used to calculate whether observations are either over-represented or under-represented when compared to a target population with no differences along potential confounders or compared to the original sample in the case of addressing selection/missing data. Observations that are under-represented are given increased weights (up-weighted) while observations that are over-represented are given decreased weights (e.g., down-weighted). This results in a pseudo-population balanced along potential confounder or selection variables [36
]. While useful in the case when confounding and selection processes are measured, this approach cannot account for unmeasured factors.
2.4. Longitudinal Study Design
Longitudinal study design is an often recommended method to reduce the chances of self-selection and reverse causation bias [29
]. Presumably, multi-year follow-up of individuals allows for the estimation of the effect of an exposure/treatment early on in the follow-up and longitudinal change in outcome or later development of disease (i.e., establishes temporality of exposure and outcome). Methods such as within-person analyses can account for time-invariant confounders that are measured and unmeasured [29
]. However, longitudinal designs alone do not fully address potential biases by self-selection and reverse causation because the previously mentioned complications of studies of NEs and brain health, namely the impact of neighborhood preferences that can change over time (i.e., time-varying) and the potential for preclinical and prodromal (undiagnosed brain disease) conditions to affect neighborhood choice/moves and thus the neighborhood characteristics under study.
2.5. Restricted/Stratified Sample
Restriction or stratification is another traditional method to account for potential confounding by self-selection or reverse causality. Studies can restrict analytic samples along some measurement (e.g., stated preference for location) or by limiting to a small area or certain neighborhoods. For instance, a study of 20–65 year olds in Belgium used a pre-existing questionnaire [40
] to assess neighborhood selection factors (i.e., house price, preference for city center, quietness, social/emotional reasons, walkability) [41
]. The authors then performed a sub-analysis to compare associations between an objective measure of neighborhood walkability and several outcome measures including walking for transportation in the full sample versus the restricted sample that reported high importance of neighborhood walkability characteristics. The associations the authors observed using the full sample remained significant and similar in the restricted sample.
Restriction/stratification may also be used to attempt to account for reverse causation. Individuals with cognitive impairment, low cognitive tests scores, or other biologic measures could be excluded to attempt to eliminate participants already experiencing brain changes that could affect neighborhood choice/characteristics. In longitudinal settings, this can also be applied such that cases of dementia within a certain period of baseline neighborhood measurements are excluded. The difficulty remains in that prodromal disease (e.g., undiagnosed of Alzheimer’s disease) can last many years and could thus still impact residential moves/neighborhood characteristics in the years leading up to diagnosis.
2.6. Quasi-Experiments and Natural Experiments
Quasi-experimental and other causal inference approaches have become increasingly popular as methods that use observational data to approximate randomized designs [42
]. When RCTs are not feasible, there may be “natural experiments” that occur which can be leveraged to estimate causal effects. A natural experiment is when an intervention has occurred but the circumstances the lead to the intervention were not controlled by the researchers [43
]. Some examples are specific policy changes or laws that affect the NE, development projects in neighborhoods, natural events, or other quasi-experimental factors. Natural experiments are posited to not be influenced by participant characteristics, which can help eliminate bias from unmeasured or mismeasured confounders. Natural experiments also establish temporal order of events. Estimates from natural experiments are thus not as susceptible to self-selection or reverse causality and may provide an unbiased estimate of the causal effects of NE on brain health. Various analytic approaches can be used in natural experiments, such as interrupted time series, pre-post designs, instrumental variable analyses [44
], regression discontinuity analyses [42
], and difference-in-difference approaches [46
Example natural experiments in the NE and health literature include studies focused on new BE amenities, such as rail lines, parks, or supermarkets [47
]. While some of these studies suggest that BE additions have had positive impacts on health, the evidence has been mixed on whether changes to the BE have had positive effects on health outcomes such as physical activity, diet, and obesity [50
]. The limitations of these types of natural BE experiments are as follows. (1) The individuals studied before and after the BE change may be different (e.g., measuring change in health outcomes at census tract level does not account for movers). (2) A change to the BE is not the same as a change in exposure to that BE characteristic (e.g., a new supermarket or rail line is not necessarily used by the individuals) [51
]. (3) Competing factors associated with the change in health are not necessarily captured (e.g., new neighborhood gyms). (4) The change in health outcome needs to be measured on the appropriate time scale (e.g., detection of changes in brain health requires a longer follow-up than changes in physical activity). (5) Population-based samples remain difficult to obtain. (6) It is difficult to find neighborhoods that are good controls/comparison groups (i.e., differ from intervention comparison in no important way related to the health outcome other than the intervention itself). Further, (7) environment changes are often known after the fact, and it is challenging to collect pre-change health data for a pre-post study [29
Given these limitations, some researchers have suggested a reorientation of these natural experiments to focus on a comparison of individuals who move versus those who do not move, to examine how the change in BE due to moves affects health outcomes [51
]. For example, one study examined changes in body mass index and physical activity levels following residential relocation among participants in the Multi-Ethnic Study of Atherosclerosis [52
]. The authors found that individuals who moved to more walkable neighborhoods (i.e., 10-point higher Walk Score) compared to their prior residence increased transport walking by 16.04 min per week (95% confidence interval (CI) = 5.13, 29.96).
Despite positive findings such as the above example, a systematic review found that the majority of natural experiment studies of the impact of the BE on physical activity had moderate to high risk of bias (e.g., did not control for important confounders, inadequate control sites) [53
]. In addition, the previously observed impacts of a residential move to health outcomes such as physical activity may not be readily observed in studies of brain health outcomes. One can imagine a comparison of movers and non-movers could be confounded by unmeasured characteristics that affect mobility and residential choice and thus confound the associations. Regardless of the type of natural experiment (change of BE due to additions/renovations or moves), the expected impact of a BE change on cognition or ADRD incidence may not be large and immediate enough to rule out competing causes.
2.6.1. Pre-Post Designs, Difference-In-Difference
One common quasi-experimental approach is to leverage longitudinal data and examine before and after health effects resulting from neighborhood changes, reducing concerns for reverse causality. These studies typically rely on individual data to compare before and after effects. Studies have employed pre-post analyses to examine changes in health due to natural experiments such as the construction of light rail [47
]. However, these types of studies may still be susceptible to self-selection bias and not all studies have identified a control/comparison group [29
], which limits causal inference.
Difference-in-difference (DID) is another approach that can capitalize on the use of longitudinal data. DID is a quasi-experimental method to estimate the effect of a specific intervention or NE/BE change by comparing the difference in outcomes over time between a population that received the intervention/exposure and a population that did not (control) [43
]. For instance, DID has been used to estimate the association between neighborhood condition and weight changes while accounting for potential self-selection due to movers and non-movers [54
]. The method has also been used to estimate the effect of neighborhood investment on physical activity and body mass index adjusting for general changes over time [55
]. In addition, a DID approach has been used to examine self-selection [56
], by estimating how baseline health may relate to changes in NE characteristics after the baseline time period. DID is useful when longitudinal data are available but rests on the assumption that the time-varying trends in the treated/exposed group are parallel to the control group in the absence of the intervention [43
]. Ultimately, there still may be potential bias, such as that due to unmeasured time-varying confounding factors.
2.6.2. Instrumental Variables
Instrumental variable analyses [44
] have emerged as a popular approach to estimate causal effects of exposures/treatments. Assumptions that an instrumental variable must meet are: 1. The instrument must be associated with the outcome. 2. The instrument must affect the outcome only through the exposure. 3. The instrument is not associated with confounders of exposure and outcome. 4. Estimated effects of instruments are the same across differing levels of the instrument (monotonicity). Instrumental variables are expected to be randomly determined and thus in theory allow for an unbiased estimate of the exposure and outcome association.
The use of instrumental variables has been infrequent thus far in neighborhood environment and health studies [58
]. Example instruments in these previous studies of NEs and diet and obesity included: (a) distance to arterial roads and non-residential zones, (b) distance to nearest highway, and (c) buildable land available for fast-food restaurants within a half-mile of the participant’s residence. These instruments were chosen to try to allow for causal inference between neighborhood characteristics and health outcomes independent of self-selection factors. To our knowledge, instruments have not been employed in studies of NEs and brain health. Thus, we explore this method to investigate reverse causality in the second case study presented in Section 3
Overall, natural experiment designs rest on additional assumptions that may be untestable, such as the assumption that the neighborhood intervention/instrumental variable is independent of factors that also influence brain health (e.g., no confounding). These study designs are still susceptible to other sources of bias such as selection bias (e.g., due to selective survival) and competing risk of death. Despite the outlined difficulties of conducting natural experiments, if designed well, they may provide the best hope at causal inference among the discussed methods. Incorporating multiple approaches that are not as susceptible to the same biases can help lead to triangulation of evidence. Comparing natural/quasi-experimental approaches to regression estimates may help understand the potential biasing effects of unmeasured attributes of self-selection and reverse causation.
In this paper, we outlined the most common methods used and proposed to address self-selection and reverse causation in studies of NEs and health. We discussed the strengths and pitfalls of using these methods when they are extended to studies of brain health and ADRD, and we provide a qualitative summary of the possibility for each method to successfully account for self-selection and reverse causation and thus improve causal inference (Table 2
). Cross-sectional analyses of NEs and brain health remain more common than longitudinal or quasi-experimental studies and have a high chance of bias without careful consideration of self-selection and reverse causality, as well as the other unique methodological considerations that must be weighed when studying neighborhood characteristics and brain health outcomes (Table 1
Our case studies demonstrated four of the methods outlined in this paper, namely: (1) adjustment for self-selection as a covariate; (2) restriction/stratification by the self-selection variable; (3) inverse probability weighting (from propensity scores) based on the probability of living in a neighborhood; and (4) instrumental variable analysis to investigate the potential for reverse causation (e.g., brain health leads to the neighborhood characteristic). The results from our case studies reiterate that methodological choices can have an impact on study findings and that studies of NEs and brain health would ideally conduct sensitivity analyses, whenever possible, to determine if results hold using different methods. Consistent findings across the methods may provide support for a true association, whereas inconsistent findings may help explain potential differences in associations observed between similar studies previously published. Authors should openly discuss their data and methodological limitations. If data or resources are not available to employ more rigorous methods, the improved/aspirational methods should also be acknowledged in the limitations sections of papers.
Few studies of NE and brain health have yet to incorporate the methods discussed in this review. Thus, we have provided more established methods used in public health and social science fields that can assist in addressing causality when confronted with potential self-selection or reverse causation. Methods borrowed from related disciplines (e.g., geography, statistics, or highly specialized medical fields) may further improve our ability to establish causality and should be investigated in future methodological studies. In addition, a mixed methods approach [82
], where researchers use both quantitative and qualitative approaches (e.g., focus groups, interviews), may be fruitful in capturing key information to elucidate the causal timing of exposure and outcome (i.e., addressing reverse causality) and neighborhood preference and residential moves due to factors related to brain health (i.e., addressing self-selection).
The next wave of NE and brain health papers need to advance beyond the traditional cross-sectional and longitudinal studies designs with multivariable regression analyses to strengthen causal inference. Changes to NEs can be costly, time consuming, and it can be difficult to garner political and public support. Proposed changes to NEs based on methodologically unsound health studies could result in unintended harms to population health that may be avoidable with more rigorously conducted studies. Neighborhood interventions, policies, and programs to improve health behaviors, exposures, and reduce the risk of ADRD will be best supported by minimally biased studies that have a critical eye on causal inference.