Work Text:
Chapter 4: Methodology
Chapter Text
Data
We analyzed data from the 2022 U.S. Survey of Consumer Finances (SCF), a comprehensive national survey conducted every three years that captures detailed information about how American families manage their finances, along with their demographic characteristics and economic attitudes. The 2022 SCF was particularly valuable for this research due to its inclusion of COVID-19-specific questions addressing health impacts, financial hardships, and government assistance received during the pandemic. The SCF employs multiple imputation techniques to handle missing data, generating five complete datasets (implicates) for each respondent. Our analytical sample consisted of 4,595 households, resulting in 22,975 observations across the five implicates. All analyses incorporated population weights to ensure national representativeness, with the weighted sample representing approximately 131 million U.S. households.
Measures
The dependent variable, precautionary savings, was operationalized as a binary measure following established methodologies (Fisher & Anong, 2012; Fisher, 2010). Households were coded as “1” if they reported saving for unemployment, illness, medical expenses, emergencies, or liquidity needs, and “0” otherwise. This approach captures intentional precautionary saving motives distinct from other savings goals.
We constructed three key measures based on the Health Belief Model framework. To capture perceived susceptibility, we identified households where either the respondent or their spouse/partner had received a COVID-19 diagnosis. Perceived severity was operationalized as financial hardship related to medical expenses during the pandemic, with respondents coded as “1” if they reported delayed or unpaid medical bills. Cues to action were measured through receipt of government assistance, including stimulus payments, unemployment benefits, or PPP loans. Financial knowledge, the moderator variable, was constructed as a score (0–3) based on responses to three validated financial literacy questions assessing understanding of interest, inflation, and risk diversification (Angrisani et al., 2020).
Control variables included race/ethnicity (Non-Hispanic White as reference, plus Non-Hispanic Black, Hispanic, and Other Non-Hispanic categories), gender/marital status (married/partnered, single female, single male), age (continuous), education level, employment status, presence of children under 18, and health insurance coverage. Economic factors were addressed through log-transformed income and inverse hyperbolic s ine-transformed networth to accommodate skewed distributions and negative values.
Analytical Approach
We employed hierarchical logistic regression analysis to examine factors associated with precautionary saving behavior. Model 1 established baseline relationships using only sociodemographic and economic control variables. Model 2 added the three HBM components (perceived susceptibility, perceived severity, and cues to action) while maintaining demographic controls. Model 3 extended this framework by incorporating financial knowledge and its interaction with perceived severity to test for moderation effects as hypothesized in H4.
All analyses were conducted using Stata software, with the multiple imputation structure of the SCF addressed through Rubin’s rules for combining results across implicates. Model comparisons were performed using Wald tests for nest ed models to assess the significance of added variables. Diagnostic tests, including examination of variance inflation factors (VIF), confirmed the absence of problematic multicollinearity (all VIF < 5). The use of population weights throughout the analysis ensured that results remained representative of U.S. households.
Model 1: ln(P(Y=1)/P(Y=0)) = β₀ + β1(Race) + β2(Age) + β3(Gender/Marital_Status) + β4(Dependents) + β5(Employment) + β6(Education) + β7(ln_Income) + β8(ihs_NetWorth) + β9(Has_Insurance)+ε Model 2: ln(P(Y=1)/P(Y=0)) = β₀ + β₁(COVID) + β₂(Financial_Hardship) + β₃(Govt_Assistance) + β₄(Race) + β₅(Age) + β₆(Gender/Marital_Status) + β₇(Dependents) + β₈(Employment) + β₉(Education) + β₁₀(ln_Income) + β₁₁(ihs_NetWorth) + β₁₂(Has_Insurance)+ε Model 3: ln(P(Y=1)/P(Y=0)) = β₀ + β₁(COVID_Diagnosed) + β₂(Financial_Hardship) + β₃(Govt_Assistance) + β₄(Race) + β₅(Age) + β₆(Gender_Marital_Status) + β₇(Children) + β₈(Employment_Status) + β₉(Education) + β₁₀(ln_Income) + β₁₁(ihs_NetWorth) + β₁₂(Has_Insurance) + β₁₃(Financial_Knowledge) + β₁₄(Financial_Hardship × Financial_Knowledge) + ε