Financial assets and mental health over time
Study design and sample selection
The CLIMB study is a nationally representative longitudinal survey of U.S. adults aged 18 and older, fielded via the AmeriSpeak panel. The AmeriSpeak panel, a probability-based sample of U.S. households randomly selected from the NORC National Frame that covers approximately 97% of U.S. households, served as the sampling frame. Eligibility for participation in the CLIMB study was limited to English-speaking AmeriSpeak panelists who had completed a survey within the preceding 6 months. The CLIMB study is comprised of four survey waves fielded during the COVID-19 pandemic: Wave 1 (March 2020 – April 2020), Wave 2 (March 2021 – April 2021), Wave 3 (March 2022 – April 2022), and Wave 4 (March 2023–April 2023).
To be included in the analytic sample, first, we subset to participants who responded to Wave 1 and at least one additional survey wave (Wave 2, Wave 3, or Wave 4) (n = 1,297). Then we dropped participants who were missing all responses to the Patient Health Questionnaire-9 (PHQ-9), a measure of depressive symptoms43, and the Generalized Anxiety Disorder-7 (GAD-7), a measure of symptoms of anxiety44, at all waves that they participated in the CLIMB Study (n = 1). This led to wave specific samples of: Wave 1 (n = 1,296); Wave 2 (n = 1,182); Wave 3; (n = 1,093); Wave 4 (n = 938). Post-stratification weights were constructed and applied to ensure the alignment of the CLIMB sample with the U.S. adult population, via benchmarking to the Current Population Survey.
Outcome measures
To measure depressive symptoms, we used the 9-level version of the Patient Health Questionnaire (PHQ-9), which is scored on a range of 0–27, where a score of 10 or higher acts as a clinical threshold indicator of moderate to severe symptoms43. To measure symptoms of anxiety, we used the generalized anxiety disorder-7 (GAD-7), which is scored on a range of 0–21, and has a clinical threshold of 1044. Both measures have been clinically validated43,44, and are often used in primary care settings as screening tools ahead of formal diagnosis of depression and anxiety. Last, we documented the co-occurrence of elevated symptoms of depression and anxiety, which was defined by a positive screen for depression (PHQ-9 ≥ 10) and anxiety (GAD-7 ≥ 10) within the same person.
Covariates
Financial assets were measured categorically at the household level and included annual income (\(\:<\)$35,000, $35,000-$64,999, $65,000-$99,999, or \(\:\ge\:\)$100,000), accrued financial assets (\(\:<\)$5,000, $5,000-$34,999, $35,000-$64,999, $65,000-$99,999 or \(\:\ge\:\)$100,000), and total debts (no debt, $1 – $19,999, \(\:\ge\:\)$20,000), consistent with other publications45,46,47,48. Accrued financial assets refer to the total amount of funds across different accounts that adults have been able to save (i.e., funds left after expenses are deducted from income flows49). The question to determine total accrued financial assets in the CLIMB study was modified from a question asked in the National Health and Nutrition Examination Survey about total household savings: “We will now ask about household savings. By savings we mean money in all types of accounts, including cash, savings, or checking accounts, stocks, bonds, mutual funds, retirement funds (such as pensions, IRAs, 401Ks, etc.), and certificates of deposit. What category best represents how much money your household (including yourself) has in savings?”
Financial stress was defined as an indicator variable coded as 1 if respondent reported experiencing stress due to at least one of the following items in the previous 12 months: job loss, household job loss, financial problems, or difficulty paying rent. In sensitivity analyses we also deconstructed this indicator variable by each stressor.
To account for non-financial pandemic-related stress we constructed an indicator of social-emotional stress, defined by whether the respondent reported experiencing stress from loneliness, relationship problems, the death of someone close, or childcare problems. We used components of the stressors asked in all waves.
Several demographic variables were constructed, including a categorical age variable (18–29, 30–44, 45–59, \(\:\ge\:\)60 years), a categorical race and ethnicity variable (non-Hispanic White, non-Hispanic Asian, non-Hispanic Black, non-Hispanic other, or Hispanic), an indicator for respondent sex (male, female), categorical variables of respondent marital status (married/living with partner, divorced/separated, never married) and education status (high school/General Educational Development (GED) or less, some college or vocational/associates degree, bachelor’s degree, or graduate/professional degree), an indicator for respondent employment status, the respondent’s region of residence (Midwest, Northeast, South, or West), health insurance (health insurance paid for by an employer or a union; health insurance you or your family pays for yourself; Medicaid; Medicare; no health care insurance; or some other kind of health insurance) and household size measured as the number of individuals living in the respondent’s home.
Analysis
First, to describe the sample, we computed population-weighted proportions and unweighted frequencies for all covariates across each survey wave; aggregate frequencies and weighted proportions were also computed. We estimated the unadjusted relationship between each covariate and our three mental health indicators, respectively. We also mapped the frequency of positive screen for symptoms of depression, anxiety, and their co-occurrence at each wave in an Euler plot (Supplemental Fig. 1).
Second, to describe how outcomes evolved over the course of the pandemic and how they varied by key covariates, we computed the regression adjusted population-weighted prevalence of exceeding the clinical thresholds of the PHQ-9, GAD-7, and both by survey year and stratified by accrued financial assets. Weighted, adjusted probabilities were taken from multivariable logistic regressions.
Third, to estimate the contemporaneous association between accrued financial assets and the odds of exceeding clinical thresholds of the PHQ-9, GAD-7, or both, we used weighted logistic regression models with year fixed effects. Incorporating year fixed effects accounts for time-varying factors across the study period such as macroeconomic trends, and standard errors were clustered at the individual level to account for repeated measurement of individuals in our pooled cross-sectional sample. All models were adjusted for multiple covariates including contemporaneous age, race and ethnicity, sex, marital status, education status, employment status, region of residence, debt, income, health insurance, number of people in the household, and presence of pandemic-related past-year financial and social-emotional stress. Our estimation approach was selected for its robustness to the incidental parameters problem (relative to a model including individual fixed effects) and because it has been well established in the literature50,51,52. Our adjustment variable selection was guided by prior literature demonstrating important confounders17, and empirically corroborated by LASSO regression. We were primarily interested in the change in cross-sectional relationships between accrued financial assets and mental health in our study period, accounting for year-to-year variation but not within person changes across time.
To test for changes in the association of accrued financial assets with mental health over time, we conducted a longitudinal analysis that used the subsample of individuals who participated in each of the four survey waves (n = 824). We present the coefficients from the time interactions with accrued financial assets in the main text and for the whole model in the Supplemental Materials. All longitudinal analyses were estimated as hierarchical linear models (HLM).
We hypothesized (1) that the association between accrued financial assets and mental health would be stronger than the association between income and mental health given that wealth may confer more mental security than income, and (2) that the relationship between accrued financial assets and mental health would strengthen over time from 2020 to 2023.
To test the robustness of our results, several alternative model specifications and estimation techniques were considered. As assets are likely correlated, additional models of the relationship between PHQ-9, GAD-7, both PHQ-9 and GAD-7, and household income, accrued financial assets, and household debts, respectively, were also estimated to test sensitivity to multicollinearity. We also estimated a model in which the covariates “Financial Stress” and “Social-Emotional Stress” were decomposed into their constituent parts. To test sensitivity to continuous instead of binary operationalization of our mental health outcomes, we estimated additional models via ordinary least squares that use continuous measures of the PHQ-9 (range: 0–27) and GAD-7 (range: 0–21), respectively.
All analyses were conducted in R version 4.1.0, and all hypothesis testing was two-sided at a significance level of 0.05. Analyses were conducted in accordance with Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for reporting observational studies. All methods were performed in accordance with the relevant guidelines and regulations. The collection of data through the CLIMB study was deemed exempt by the Institutional Review Board (IRB) at NORC at the University of Chicago (IRB Protocol Number# 23-03-1219). As secondary analysis of de-identified data was performed in the study, the need of ethics approval was waived by the IRBs at Boston University Medical Campus (under IRB# H-39986) and Johns Hopkins Bloomberg School of Public Health (under IRB# 25544). Informed consent was obtained from all subjects to participate in the study.
link
