Using Classification Trees to Identify Multiple Patterns of Characteristics That Result in U.S. Urban and Rural Households Having Very Low Food Security

Year: 2016

Research Center: Center for Regional Development, Purdue University

Investigator: Frongillo, Edward A., Maryah S. Fram, and Seul Ki Choi

Institution: University of South Carolina

Project Contact:
Edward A. Frongillo
University of South Carolina
Department of Health Promotion, Education, and Behavior
915 Greene Street
Columbia, SC 29208
Phone: 803-777-4792


In 2015, about 6.3 million U.S. households (5.0 percent) experienced very low food security (VLFS). A household has VLFS when one or more household members has reduced food intake or disrupted eating patterns because there is not enough money for food. VLFS is associated with a range of serious negative health, developmental, and psychosocial consequences. Thus, identifying households at risk for VLFS and reaching them with effective intervention is important.

Previous studies focused on predicting VLFS in households using traditional variable-focused analytic methods, although VLFS cannot be explained in the simple way assumed by these methods. Classification and regression trees (CART) is a powerful analytic method for making prediction from data and is well suited to identifying VLFS households, taking into account their complex situations.

This study applied CART analysis to identify patterns of characteristics that distinguish VLFS households in the United States from those with greater food security. The study hypothesized that many different processes can lead households to experience VLFS. Therefore, it expected to show multiple patterns of characteristics reflecting different paths through which households with different characteristics have VLFS.

The study used data from the 2011-14 National Health Interview Survey (NHIS), the 2005-12 National Health and Nutrition Examination Survey (NHANES), and the 2002-12 Current Population Survey (CPS) to consider a wide range of variables that may help predict VLFS. Survey respondents ages 18 years or older and with household income of less than 300 percent of the Federal poverty level (FPL) were included in the analysis. CPS and NHIS data were analyzed separately for households with children (at least one member under age 18), adult-only households (only adults between ages18 and 64), and older-adult households (those without any children and with at least one adult age 65 years or older). CPS participants were further categorized based on residential area (metropolitan and nonmetropolitan). NHANES data were categorized into individuals ages 18-64 and those age 65 or older. The 10-item U.S. Adult Food Security Scale was used to measure household food security during the last 30 days for NHIS and CPS and during the last 12 months for NHANES and CPS. Variables in several domains, including sociodemographic characteristics, health status and behaviors, healthcare access and utilization, and participation in governmental assistance programs and food assistance programs, were selected as predictors, according to availability in each survey. The three data sources were analyzed separately using classification tree analysis. All variables were used to build binary split trees using SAS Enterprise Miner 13.2.

For all types of households from NHIS, having unmet medical needs (did not get medical care when it was needed) was the first predictor to distinguish VLFS households. Among households with children, this was the only distinguishing predictor (VLFS: 23.3 percent among households with unmet medical needs; 6.5 percent among households without unmet medical needs). In older-adult households, those that had unmet medical needs and received Supplemental Nutrition Assistance Program (SNAP) benefits for two or more months out of the last 12 months had the highest VLFS prevalence (27.4 percent). Adult-only households showed complex paths predicting the experience of VLFS. Households with unmet medical needs, with a household member limited by memory problem, and receiving SNAP benefits for two or more months out of the last 12 months had the highest prevalence of VLFS (60.0 percent). Only 5.0 percent of households that did not have unmet medical needs and did not receive SNAP benefits experienced VLFS.

For NHANES participants ages 18-64, the highest prevalence of VLFS (31.0 percent) was in households that received SNAP benefits in the last 12 months and included a household member who experienced moderate or severe depressive symptoms in the last two weeks. Individuals in households that did not receive SNAP benefits had the lowest prevalence of VLFS (8.3 percent). For individuals age 65 or older, household income was the single predictor to distinguish VLFS households. Prevalence of VLFS was 9.0 percent among individuals with household income less than 125 percent of the FPL versus 1.9 percent among individuals with household income between 125 and 299 percent of the FPL.

For CPS participants, households with children who did not receive free or reduced-cost school lunch had VLFS prevalence of 3.1 percent. Among those where children received free or reduced-cost school lunch and all adult members were unemployed, VLFS prevalence was 15.2 percent. For older-adult households, those that received SNAP benefits during the last 12 months and included a disabled person had the highest VLFS prevalence (9.1 percent). Households that did not receive SNAP benefits and had some interest income had the lowest VLFS prevalence (0.6 percent). For adult-only households, the only variable distinguishing households experiencing VLFS was receipt of SNAP benefits during the last 12 months (23.7 percent VLFS versus 5.6 percent for households not receiving SNAP). There was no difference in classification trees for food security between households living in metropolitan and nonmetropolitan areas. Classification trees for food security during the last 12 months were similar to trees for food security during the last 30 days.

Multiple characteristics across multiple domains predicted VLFS households. VLFS is about more than just food. Health status, depression, disability, employment, income from interest, and other factors predicted the experience of VLFS. Identifying and responding to the needs of households with VLFS will require nuanced attention to a range of social, economic, health, and contextual challenges. The results point to several promising directions for identifying VLFS households using a combinations of characteristics. Households with VLFS may need supports (for example, better access to physical and mental health services) along with better access to food. Multiple domains should be used to identify VLFS households and to inform policies and programs that can address VLFS households’ needs.