Food insecurity, uncertainty in accessing enough food for an active, healthy life, affected 34 million individuals in the USA in 2021.1 Food insecurity is associated with worse clinical outcomes, along with greater healthcare use and cost.2,3,4,5,6,7,8,9,10 As a result, food insecurity is frequently the target of both clinic- and health insurance-based interventions, along with public health programs.11,12,13,14,15

Research regarding food insecurity interventions faces several challenges. First, funding is often predicated on reducing healthcare utilization or costs, or improving metrics used in pay-for-performance programs or public reporting. Outcomes less clearly linked to short-term “return on investment” may be understudied.16 For example, health-related quality of life and mental health are strongly associated with food insecurity17,18,19,20,21,22,23 and matter deeply to patients, yet may attract little interest from intervention funders.

A second issue relates to randomized clinical trials (RCTs). RCTs are an important method for demonstrating intervention effects, but have drawbacks that are especially salient for food insecurity interventions. These include expense, difficulty enrolling and retaining a representative sample (and in particular, ongoing issues with racial and ethnic inclusivity24,25), and ethical concerns about random allocation to study arms that may not provide support for basic needs (e.g., “usual care”).

An emerging methodological approach known as target trial emulation may help address these issues.26,27 Target trial emulation involves the use of observational data to emulate a “target trial” investigators would like to have conducted.28,29,30,31,32,33 Target trial emulation can help understand what intervention effects are possible. Furthermore, it can help inform how much could be spent for an intervention to be cost-effective, an important question for food insecurity research. Moreover, target trial emulation can complement RCTs by emphasizing external validity, in contrast to internal validity that RCTs emphasize.

Real-world interventions known to effectively treat food insecurity include food subsidies, delivery of food boxes, subsidized memberships to community-supported agriculture associations, medically tailored meals, and cash transfers.34,35,36,37,38,39 In this study, we emulated a target trial of an intervention that eliminates food insecurity among a nationally representative sample of US adults who screen positive for food insecurity. We examined outcomes that have been less commonly studied in food insecurity interventions—health utility, mental and physical health-related quality of life, psychological distress, and depressive symptoms. We hypothesized that an intervention that eliminated food insecurity would be associated with meaningful improvements for these outcomes.

METHODS

Data and Study Setting

We studied a nationally representative cohort of non-institutionalized US residents using the longitudinal data files and food insecurity files of the Medical Expenditure Panel Survey (MEPS, panel 21, 2016–2017).40 Included individuals completed MEPS interviews at multiple time points over the 2-year study period. These were the most recent years that measured food insecurity for the same participants over a 2-year period. This study was not considered human subjects research by the institutional review board at the University of North Carolina at Chapel Hill.

Eligibility Criteria

We employed eligibility criteria meant to mimic those used in food insecurity intervention trials, and use terminology analogous to that used in actual trials for that reason. The investigators conducting this study did not have any contact with the study participants. MEPS participants were considered “eligible” for this study if they were adults (age 18 years or greater at the time of initial MEPS interview) and screened positive for food insecurity on the “Hunger Vital Sign,” a validated food insecurity screening instrument.41,42 Adults who screened positive for food insecurity were considered “enrolled” in the emulated trial if they completed “baseline” assessment (i.e., the questionnaire items needed for the emulated trial, described below). “Baseline” visits occurred in 2016, with “follow-up” visits in 2017. Those who did not complete follow-up visits were considered “lost to follow-up.”

Food Insecurity

Food insecurity was assessed using the 10-item USDA Adult Food Security Survey Module, with a 30-day look-back period.43 Since the study only included adults, we treated household-level food security status as the food security status of the individual within that household. We dichotomized food security status as food secure (0 to 2 affirmative responses) or food insecure (3 or more affirmative responses), a standard categorization (sample size did not permit finer categorization).43

Outcomes

The primary outcome was the SF-6D (Short-Form Six Dimension) measure of health utility.44,45 For health utility scores, a global indicator of health, 0 represented the state of death and 1 represented a state of full health. One year of life spent in a state of full health would yield 1 quality-adjusted life-year (QALY). To help contextualize the magnitude of estimated effects on the health utility outcome, we express what the cost of a food insecurity intervention could be, to be cost-effective at a $100,000 per QALY threshold.46

Secondary outcomes were the mental component score (MCS) (minimum clinically important difference [MCID]: 1.947) and the physical component score (PCS) (MCID: 2.647) of the Veterans RAND 12-Item Health Survey48 (VR-12), the Kessler 649 (K6) (MCID: unknown) measure of non-specific psychological distress, and the Patient Health Questionnaire 2-item50 (PHQ2) (MCID: unknown) measure of depressive symptoms. The MCS and PCS health-related quality of life measures are designed to have a mean of 50 and a standard deviation of 10 in the US population, with lower scores indicating worse health-related quality of life. K6 scores range from 0 to 24, with lower scores indicating less distress49, and the PHQ2 measure of depressive symptoms ranges from 0 to 6 with lower scores indicating fewer depressive symptoms.17,50 Participants who completed assessment for some outcomes but not others were included in analyses for the outcomes for which data were available, and censored otherwise.

Covariates

We accounted for several covariates that may confound the relationship between food insecurity and study outcomes. We classified these covariates as time-invariant or time-varying. We used baseline (2016) measurements of time-invariant covariates, and both 2016 and 2017 measurements of time-varying covariates. Time-invariant covariates were age, gender, education, English as the language of interview, and self-reported race and ethnicity (categorized within MEPS as non-Hispanic Black, non-Hispanic White, non-Hispanic Asian, non-Hispanic Other [which in MEPS includes individuals who identify as multi-racial], and Hispanic). These were the finest categories feasible with the available data. We included race and ethnicity variables in our analyses as indicators of potentially experiencing both structural and interpersonal racism that can confound the relationship between food insecurity and study outcomes. Time-varying covariates were household income expressed as a percentage of the federal poverty guideline, region of residence, health insurance, and family size.

Statistical Analysis

Target trial emulation has two key phases—study design, in which an analytic cohort is constructed to be analogous to a sample in an actual trial, and analysis, in which models are applied to observed data, and then used to estimate outcomes under counterfactual scenarios (Fig. 1).51 For analysis, we used targeted minimum-loss estimation (TMLE) to quantify the association between food insecurity and study outcomes, while accounting for confounding (eFigure 1) and loss to follow-up.52,53 TMLE has been previously described54 (and we provide additional details in the Technical Appendix), but the intuition is that one models the relationship between food insecurity, the covariates, and the outcome (similar to the parametric g-formula) and the relationship between the covariates and food insecurity (similar to propensity score methods) and accounts for possibly informative censoring resulting from loss to follow-up (similar to inverse probability of censoring weighting). These models together form a doubly robust estimator of the impact of food insecurity on study outcomes. Because there is a known bidirectional relationship between food insecurity and health6, using TMLE is appropriate as it avoids the bias that conventional longitudinal regression methods can have in the presence of exposure and outcome feedback.55

Figure 1
figure 1

Representation of the target trial emulation approach. This figure depicts relevant aspects of the target trial emulation methodological approach, separated into the design phase and the analysis phase. MEPS Medical Expenditure Panel Survey, MCS mental component score, PCS the physical component score of the Veterans Rand 12-Item Health Survey, K6 the Kessler 6 psychological distress scale, PHQ-2 Patient Health Questionnaire 2-item measure of depressive symptoms.

Full size image

The models incorporated the abovementioned covariates and the baseline outcome (e.g., 2016 health utility in analyses that used 2017 health utility as the outcome), to help account for unmeasured time-invariant confounding. We conducted TMLE analyses both with generalized linear regression models alone (primary analyses) and with a “SuperLearner” machine learning approach that uses cross-validation to find the best-performing weighted combination of a set of estimation algorithms.56

After fitting these models on the observed data, we used them to estimate and contrast outcomes under different, possibly counterfactual, distributions of food insecurity.57 This provided an estimate of the “treatment effect” a hypothetical intervention might offer. We estimated treatment effects of food insecurity elimination intervention in two scenarios.

Scenario 1 (primary scenario) contrasted outcomes under the distribution of food insecurity observed in 2017 with the outcomes the models estimated would be observed had all participants been food secure in 2017. The estimand targeted is an average treatment effect (ATE) if all participants had adhered to their “assigned” treatment and none had been lost to follow-up. The rationale for scenario 1 was to provide a benchmark treatment effect for an intervention that eliminated food insecurity, compared with a “status quo” or usual care condition (i.e., some individuals remained food insecure and some become food secure owing to background programs).

Scenario 2 targeted an ATE contrasting outcomes estimated to have occurred had everyone been food secure and completed follow-up with outcomes estimated to have occurred had everyone been food insecure and completed follow-up. Scenario 2 was analogous to an RCT in which some individuals were assigned to be food insecure and some were assigned to be food secure, and all participants adhered perfectly to their assigned groups.

As the experience of food insecurity can result from forms of oppression such as racism and sexism, we conducted subgroup analyses to examine whether the estimated effect of food insecurity elimination varied across categories of individuals defined by race and ethnicity, and gender. Furthermore, because food insecurity screening instruments have moderate specificity, but are more commonly used in clinical settings than full assessments owing to lower respondent burden, we conducted subgroup analyses to determine if estimated intervention effects varied by whether an individual who screened positive was ultimately categorized as experiencing food insecurity on full assessment or not.

Analyses incorporated representativeness weights provided by MEPS. We estimated efficient influence curve-based standard errors and included a robustness check for the primary outcome by clustering observations at the sampling stratum level (Technical Appendix). Analyses were conducted in SAS 9.4 and R 4.0.2 using the lmtp software package.57,58 A p-value less than 0.05 for the primary outcome was taken to indicate statistical significance.

RESULTS

In the MEPS dataset, 2,035 adults were eligible for inclusion (Fig. 2). Of these, 2,013, representing 32 million non-institutionalized adults in the USA, completed “enrollment” and were included in the overall study sample. The mean age in the study sample was 44.0 years (SD: 16.5), and 56.0% were women (Table 1).

Figure 2
figure 2

Study flow diagram.

Full size image
Table 1 Demographics
Full size table

Among those who screened positive for food insecurity, 1,368 (68.6%) reported food insecurity in 2016 as defined by the full food security survey module. In 2017, 790 (39.4%) participants were food insecure. The mean health utility in 2016 was 0.727 (SD: 0.160). Twenty-seven individuals were lost to follow-up and censored for all outcome analyses.

Scenario 1

Under the status quo, the mean health utility in 2017 was 0.819 (95% confidence interval [CI] CI 0.812 to 0.826). Had all individuals been food secure in 2017, we estimated that mean health utility would have been 0.827 (95% CI 0.818 to 0.836). Relative to the status quo, we estimated that food insecurity elimination would improve health utility by 80 QALYs per 100,000 person-years, or 0.008 QALYs per person per year (95% CI 0.002 to 0.014, p = 0.005) (Table 2 and Fig. 3). This suggests that if such a health utility difference was sustained for 1 year, after which both groups experienced identical health utility, food insecurity elimination would be cost-effective at the $100,000 per QALY threshold if it cost $800 (95% CI $200 to $1400) or less per person, relative to the status quo.

Table 2 Estimated Effects of a Food Insecurity Elimination Intervention
Full size table
Figure 3
figure 3

Forest plot of standardized mean differences. To help visualize the magnitude of the estimated effects across outcomes measured on different scales, we converted estimates to standardized mean differences (SMD) and plotted the point estimate (circle or square) with 95% confidence intervals (whiskers). Scenario 1 indicates a comparison between a condition in which all participants are made food secure via a hypothetical intervention and a condition in which food insecurity is distributed as observed under the status quo. Scenario 2 indicates a comparison between a condition in which all participants are made food secure via a hypothetical intervention and a condition in which all participants are made food insecure via a hypothetical intervention. These estimates of SMD are used only for visualization, not inference. SF-6D Short-Form Six Dimension health utility index, MCS mental component score, PCS the physical component score of the Veterans Rand 12-Item Health Survey, K6 the Kessler 6 psychological distress scale, PHQ2 Patient Health Questionnaire 2-item measure of depressive symptoms.

Full size image

We also estimated that, relative to the status quo, food insecurity elimination would have benefits for mental health (difference in MCS scores [95% CI]: 0.55 [0.14 to 0.96]), physical health (difference in PCS scores: 0.44 [0.06 to 0.82]), psychological distress (difference in K6 scores: − 0.30 [− 0.51 to − 0.09]), and depressive symptoms (difference in PHQ-2 scores: − 0.13 [− 0.20 to − 0.07]).

Scenario 2

If all individuals had been food insecure in 2017, we estimated that mean health utility would have been 0.804 (95% CI 0.793 to 0.815). Relative to this, we estimated that food insecurity elimination would improve health utility by 230 QALYs per 100,000 person-years, or 0.023 per person per year (95% CI 0.010 to 0.036, p < 0.001). This suggests that if such a health utility difference was sustained for 1 year, after which both groups experienced identical health utility, the hypothetical intervention that eliminated food insecurity would be cost-effective at a $100,000 per QALY threshold if it cost $2300 (95% CI $1000 to $3600) or less per person, relative to a condition in which everyone was food insecure.

We also estimated that, relative to experiencing food insecurity, food insecurity elimination would have benefits for mental health (difference in MCS scores [95% CI]: 1.54 [0.54 to 2.54]), physical health (difference in PCS scores: 1.35 [0.46 to 2.25]), psychological distress (difference in K6 scores: − 0.83 [− 1.30 to − 0.35]), and depressive symptoms (difference in PHQ-2 scores: − 0.36 [− 0.53 to − 0.19]).

Subgroup Analyses

The estimated effect of a hypothetical food insecurity elimination intervention was similar, for all outcomes, across categories of individuals defined by race and ethnicity or gender (eTable 1), although estimates were more uncertain given smaller sample sizes. However, since individuals who experience oppression, such as non-Hispanic black and Hispanic individuals, and women, frequently experience food insecurity, such an intervention may disproportionately benefit categories of individuals that experience oppression, even if the estimated effect of the intervention is similar at the level of the individual. Point estimates of estimated intervention effects were generally larger for those who were categorized as food insecure in 2016 by the full assessment, compared with those who screened positive but were not food insecure on full assessment. However, the differences in point estimates were small in magnitude and confidence intervals generally overlapped.

Robustness Checks

Robustness checks using SuperLearner for model estimation yielded results that were similar to the main analyses (eTable 2), as were results using different methods of standard error estimation.

DISCUSSION

In this target trial emulation study using longitudinal, nationally representative data, we estimated that food insecurity elimination would lead to modest improvements in health utility, mental and physical health-related quality of life, psychological distress, and depressive symptoms. Furthermore, analyses found little variation in estimated intervention effects by categories of race and ethnicity or gender.

This study found that food insecurity interventions that cost between $200 and $3600 per person per year may be cost-effective under various scenarios. Real-world interventions known to address food insecurity do have costs within this range. For example, in 2017, the mean per-person SNAP benefit was about $1440 per year, and the WIC program spent about $800 per beneficiary per year.59,60 Furthermore, a randomized trial of a community-supported agriculture intervention in the same period found meaningful food insecurity reductions with a $300 per year subsidy.61 Moreover, framing results in this way helps researchers compare food insecurity interventions to other interventions they may consider in similar populations, such as case management.62 For secondary outcomes of MCS and PCS scores, food insecurity elimination was associated with a statistically significant improvement, but its magnitude was small, below commonly used minimum clinically important difference thresholds. The magnitude of improvement for K6 and PHQ-2 scores is difficult to contextualize, since there is no commonly used minimum clinically important difference threshold for these outcomes.

The results of this study should be understood in the context of prior work. Cross-sectional studies have found that food insecurity is associated with worse health-related quality of life, and with depressive symptoms.18,19,20,22,23 A smaller body of longitudinal and interventional work has suggested that health-related quality of life and mental health may improve when food insecurity improves.17,36,37 The current study adds to these findings by helping to quantify the magnitude of changes that might be expected with food insecurity interventions. Furthermore, the study results highlight the economic precarity of many individuals in the USA. Although interventions to address food insecurity once it occurs may help improve health, it is also important to consider the overall context of distributive institutions in which poverty occurs in the USA, and what policy approaches might prevent food insecurity from occurring.63,64

These findings have important implications and suggest directions for future work. First, given the potential for improvement in health utility, health-related quality of life, and mental health, such outcomes should be examined in future RCTs of food insecurity interventions. It is important to move beyond narrow considerations of healthcare utilization and cost, and into more holistic consideration of how food insecurity interventions may benefit patients more broadly. Second, the analyses regarding estimated intervention effects between those who were categorized as food insecure on full assessment and those who screened positive but were not food insecure on full assessment suggest trade-offs relevant for intervention planning. Because intervention effects relate to eliminating food insecurity, two key considerations for participant selection are the likelihood of experiencing food insecurity in the absence of the intervention, and the presence of alternative food insecurity interventions. In situations where individuals are less likely to experience food insecurity or more likely to have food insecurity addressed through other channels, an intervention is likely to offer less population-level benefit, and vice versa.

The results of this study should be understood in the context of several limitations. First, this is an observational study, and unmeasured confounding could bias results. The study design accounts for many time-invariant and time-varying measured confounders, and using repeated measures within individuals helps protect against unmeasured time-invariant confounding, but unmeasured time-varying confounding cannot be excluded. Second, this study examined the potential impact of a hypothetical food insecurity intervention. An actual trial would need to test specific interventions. Fortunately, several food insecurity interventions are known to reduce food insecurity in real-world settings.34,35,36,37,38,39 Third, to provide a benchmark for actual trials, the scenarios studied here hypothesized an intervention that fully addressed food insecurity in the sample to which it was applied. An actual intervention that was less effective at addressing food insecurity would be expected to be proportionately less effective at improving study outcomes.

CONCLUSIONS

This target trial emulation estimated that a hypothetical intervention to address food insecurity would produce modest improvements in health utility, health-related quality of life, and mental health. As interventions to address health-related social needs become more widespread, it is important not to narrowly focus on issues of healthcare use and cost, but instead to investigate the full potential of these interventions to improve the many different aspects of health.