Study population

The Wages Study is a prospective cohort study. Recruitment methods and inclusion criteria are described in detail elsewhere^{(Reference Shanafelt, Sadeghzadeh and Chapman13)}. In January 2018, the Wages Study began following a cohort of 974 low-wage workers (those earning ≤ $11·50 an hour at baseline) in Minneapolis (n 495) and low-wage workers in a comparison city with no minimum wage increase (Raleigh, North Carolina, n 479). The study aims to follow this cohort throughout 4·5-years of implementation of the Minneapolis minimum wage ordinance (1 January 2018–1 July 2022). Recruitment and baseline data collection occurred from January to October 2018. Of note, the baseline data collection period (hereon referred to as Wave 1) was extended from the original completion date of July 2018 to October 2018 due to challenges in recruitment. Details of this are discussed elsewhere^{(Reference Shanafelt, Sadeghzadeh and Chapman13)}. Wave 2 data collection occurred during the summer and fall of 2019. Data will be collected again in the summers of 2020 (Wave 3), 2021 (Wave 4) and 2022 (Wave 5).

This study described in this manuscript uses the first two waves of longitudinal data from the currently ongoing Wages Study (n 655, as 319 participants of the originally recruited 974 participants were lost to follow-up at Wave 2). After exclusions, data from 603 Wages participants at Waves 1 and 2 were available for the study’s first set of analyses (Fig. 3), and 540 Wages participants were available for the second set of analyses (Fig. 4). The study was approved by the institutional review boards of the University of Minnesota and the University of North Carolina at Chapel Hill, and participants gave written informed consent to participate.

Fig. 3

Flow chart for Wages participant exclusion in the present study’s policy analyses

Fig. 4

Flow chart for Wages participant exclusion in the present study’s hourly wage analyses

Hourly wage assessment

Wages participants attend one data collection appointment each year in which wages are verified and a computer-based survey is administered. Participants are asked to bring a recent pay stub or other document from their primary employer to verify their hourly wage at the annual data collection appointment. At Wave 1, 75·67 % of participants verified their hourly wage (737/974). At Wave 2, 81·37 % of the 655 participants who returned for a follow-up appointment verified their hourly wage (533/655). All other participants self-reported their hourly wage.

Dietary assessment

To assess dietary intake, the computer-administered survey included 22 questions from the validated 26-item National Cancer Institute’s (NCI) Dietary Screener Questionnaire (DSQ)^{(14,Reference Thompson, Midthune and Kahle15)} . The primary investigators for the Wages research team excluded four DSQ items from the computer survey (milk, cheese, red meat and processed meat) because the research team wanted to keep only the most relevant questions in the computer survey to minimise participant survey fatigue.

For the current analysis, we used the DSQ frequency data to estimate participants’ daily frequency of intake of three different food groups to be used as the study’s dependent variables: F&V, whole-grain-rich foods (in which the first ingredient is a whole grain) and foods high in added sugars (> 5 grams of sugar per serving). These food groups were created because all three are associated with weight gain (a lower risk for F&V^{(Reference Nour, Lutze and Grech16)} and whole-grain-rich foods^{(Reference Maki, Palacios and Koecher17)} and a higher risk for foods high in added sugars^{(Reference Faruque, Tong and Lacmanovic18,Reference Malik, Schulze and Hu19)} ) and chronic disease risk (a reduced risk for F&V^{(Reference Boeing, Bechthold and Bub20)} and whole-grain-rich foods^{(Reference Seal and Brownlee21)} and an increased risk for foods high in added sugars^{(Reference Khan and Sievenpiper22)}) in previously conducted scientific literature. The five-gram cut-off was chosen for foods high in added sugars because the daily value of added sugars is 50 grams per day based on a 2000 kilocalorie per day diet⁽²³⁾, and the Food and Drug Administration considers a food to be a ‘good’ source of a nutrient if it contains 10–19 % of the daily value⁽²⁴⁾. Thus, we designated a food as being high in added sugar if it contained more than 10 % daily value for sugar (greater than 5 grams). Supplemental Table 1 displays the foods from the DSQ that contribute to each food group.

To create the food group-dependent variables, the research team first classified all foods from the DSQ as to whether they belonged, or not, in each of the three food groups. Foods could belong to more than one food group. We then converted participants’ responses to the DSQ into daily frequencies for each food (e.g. if a participant responded that he/she consumed popcorn ‘2–3 times last month,’ we divided 2·5 by 30 and assigned that participant a value of 0·083 for their popcorn consumption variable). Finally, we created three new variables for each participant in the data set. The first variable was the sum of the daily frequencies for all F&V foods, the second variable was the sum of the daily frequencies for all whole-grain-rich foods and the third was the sum of the daily frequencies for all foods high in added sugars.

Covariate assessment

We collected data on demographic, economic and health-related factors, including age (continuous), sex (male, female and non-binary), race (White alone, Black or African American alone, Asian alone, Native Hawaiian or Pacific Islander alone, Native American or Alaskan native alone, more than one race, or other race), ethnicity (Hispanic/Latino and non-Hispanic/Latino), marital status (married or single), birthplace (born in the USA, born abroad to American parents or born abroad), whether or not a participant was a food service worker (as food service employees are often provided meals on the job^{(Reference Gupta, Coates and Dorrian25)}, which may impact their dietary intake), educational attainment (less than high school, some high school, high school diploma, associate/technical degree, some college, or Bachelor’s degree or higher), number of adults living in the household (one, two, three, four, or five or more), number of children living in the household (one, two, three, four, or five or more), pregnancy status (pregnant and not pregnant), smoking status (current smoker, quit less than 12 months ago, quit more than 12 months ago or never smoker), health insurance status (insured all year (any type of health insurance), uninsured for at least part of the year or uninsured all year), BMI (continuous), the timing (in weeks) of the participant’s data collection appointment relative to the minimum wage increase, number of jobs worked (one job worked or more than one job worked) and the amount received in Supplemental Food and Nutrition Assistance Program (SNAP) benefits (I do not receive any SNAP benefits, $1–$25, $26–$50, $51–$75, $76–$100, $101–$150, $151–$250, $251–$500, $501–$750, more than $750).

Statistical analysis

The research team performed two sets of difference-in-difference (DID) analyses to address the present study’s aims. The first DID analysis examined whether living in a city with a mandated minimum wage increase was associated with changes in daily frequency of F&V consumption (model one), whole-grain-rich foods (model two) and foods high in added sugars (model three). This analysis categorised participants by city of residence (0 = Raleigh and 1 = Minneapolis) when assessing the exposure in the DID models. We term these the ‘policy’ analyses. The second analysis examined whether changes in individual hourly wage were associated with changes in daily frequency of F&V consumption (model four), whole-grain-rich foods (model five) and foods high in added sugars (model six). This analysis used Wages participants’ hourly wage as the exposure variable in the DID models. We term these the ‘hourly wage’ analyses.

The research team chose to conduct two sets of DID analyses for several reasons. First, the policy analysis examines the association between the change in policy and the change in outcome, acknowledging that some Wages participants in Minneapolis may not experience a wage increase from Wave 1 to Wave 2 (due to the possibility that some businesses may not be compliant in implementing the minimum wage ordinance, or due to the possibility that wage changes may not be linear and positive over time among low-wage workers, particularly if job changes, job losses or a reduction in hours worked occur). This is important as it estimates the impact of the ordinance under ‘real-world’ conditions of adherence. Additionally, participants in Raleigh may also experience wage increases (due to job promotions, raises, etc.). Lastly, it accounts for any impact of simply living in an area with a mandated wage increase on dietary outcomes (e.g. changing attitudes, beliefs and norms)^{(Reference Kinzig, Ehrlich and Alston26)}. However, the research team also wanted to run models in which participants’ individual hourly wage served as the primary predictor variable. This is important to test, as increasing hourly wage is the key mechanism by which a minimum wage ordinance may increase household income and therefore improve dietary intake (Fig. 5). Prior to conducting these DID analyses, the research team examined the parallel trends assumption^{(Reference Zhou, Taber and Arcona27)} using BRFSS SMART data⁽²⁸⁾ (Fig. 6) and found that current trends in dietary intake between the two cities did not differ meaningfully when comparing data from residents with incomes less than $35 000 per year from 2005 to 2015. All analyses were conducted in Stata/IC (version 16.0, 2019, StataCorp LLC).

Fig. 5

Hypothesised causal pathway for the relationship between a minimum wage policy and changes in dietary intake

Fig. 6

Dietary intake by metropolitan statistical area 2005–2015 BRFSS SMART data – incomes < $35 000 per year

Analysis 1 – Policy analysis

In the first DID analytic approach, the Wave 1 Wages data were designated as the pre-policy period, while Wave 2 served as the post-policy period. The treatment group consisted of the Minneapolis participants and the control group the Raleigh participants. A product term for these two variables provided the DID estimate (city * time period). For all models, analyses adjusted for the covariates listed above, except for amount received in SNAP benefits, as we thought this variable could be a potential mediator in the relationship between a minimum wage increase and dietary intake.

The DID models were estimated using negative binomial regression, as the outcomes were over-dispersed count data. Likelihood ratio tests that the dispersion parameter was equal to zero revealed that negative binomial models were a better fit than Poisson models^{(Reference Cameron and Trivedi29–Reference Johnson and Kotz31)}. Data were analysed with longitudinal regression analysis (generalised estimating equations (GEE) with clustering by the individual), using the Huber – White sandwich estimator of variance and an autoregressive correlation matrix in order to account for repeated measures within individuals^{(Reference Diggle, Liang and Zeger32)}. Sensitivity analyses were performed, and results did not change when alternative correlation matrices were specified.

Analysis 2 – Hourly wage analysis

The second analytic approach was identical to the first analytic approach, except continuous DID models were specified and Wages participants’ hourly wages, rather than city of residence, were used to calculate the DID indicator (hourly wage * time period). The researchers adjusted for the same covariates as done in the first set of analyses, except city was added as an additional covariate.

Sensitivity analyses

Given the study’s high attrition rate and thus the possibility of selection bias and biased parameter estimates^{(Reference Munafò, Tilling and Taylor33)}, we examined differential attrition by baseline measures of age, sex, race, ethnicity, educational attainment and SNAP usage among those who returned for a Wave 2 appointment (n 655) v. those who did not (n 319) using t tests and chi-square tests (Tables 3 and 4). We also conducted sensitivity analyses for both the ‘policy’ analyses and the ‘hourly wage’ analyses using inverse probability-of-censoring weights. Inverse probability-of-censoring weight inversely weights regression analyses by the probability of participation (determined based on a logistic regression model for probability of participation given past history covariates and outcomes)^{(Reference Robins, Rotnitzky and Zhao34–37)}. This inflates the impact of underrepresented subjects, so we can observe associations that would have been observed if all subjects had stayed in the Wages Study at Wave 2 (assuming the models are correctly specified)^(40–43).

To perform inverse probability-of-censoring weight, we first fit a logistic regression model to estimate the probability of not returning at Wave 2 based on baseline characteristics of age, sex, race, ethnicity, educational attainment, birthplace, marital status, number of children living in the household, SNAP usage, hourly wage, job type and whether the participant lived in Minneapolis or Raleigh. We then used weights derived from this model to re-estimate the six DID regression models and the associations that would have been observed if all subjects from Wave 1 had remained in the study at Wave 2. The six weighted DID models used 1/P as weights. All covariates used in the weighted DID models were identical to the covariates in the original unweighted DID models.

Post hoc analyses

In the event that results were not as hypothesised and a minimum wage policy change was not associated with changes in dietary intake between the cities, the research team decided to conduct post hoc analyses to understand why. We hypothesised that if results were null, perhaps the first phase of change in minimum wage policy did not translate to higher hourly wages or higher household income between the two cities. We therefore decided to examine the following post hoc research questions: RQ1) on average, did the hourly wage significantly change between the cities from Wave 1 to Wave 2?; RQ2) were changes in hourly wage associated with changes in household income?; RQ3) on average, did household income categories significantly change between the cities from Wave 1 to Wave 2?; RQ4) was the policy associated with changes in hourly wage between the cities from Wave 1 to Wave 2?; and RQ5) was the policy associated with changes in household income categories between the cities from Wave 1 to Wave 2? Because only 1 year had passed between Waves 1 and 2, we did not inflation-adjust hourly wages in the post hoc analyses. Figure 5 displays how we thought the change in policy would lead to change in wages and outcomes, and what relationship each set of post hoc analyses tested.

To address post hoc RQ1 (did the hourly wage significantly change between the cities from Wave 1 to Wave 2?), the research team performed data tabulations to examine the average wages and change in wages among Raleigh and Minneapolis participants at Waves 1 and 2. Additionally, we estimated unadjusted DID regressions using longitudinal regression analysis (GEE with clustering by the individual), and using Huber–White sandwich estimator of variance and an autoregressive correlation matrix to adjust for the within-subject correlation.

Given that household income was an ordinal variable in our data set, the research team addressed RQ2 (were changes in hourly wage associated with changes in household income?) by estimating a multinomial logistic regression model. A multinomial logistic regression model was estimated rather than an ordinal logistic regression model because the proportional odds assumption was tested and violated. Standard errors were clustered at the level of the individual.

Again, because household income was an ordinal variable, the research team addressed RQ3 (did household income categories significantly change between the cities from Wave 1 to Wave 2?) by estimating a DID ordinal logistic regression model with standard errors clustered at the level of the individual. The same DID product term was used as described in RQ1. The proportional odds assumption was tested and held.

The same DID model from RQ1 was used for RQ4 (was the policy associated with changes in hourly wage between the cities from Wave 1 to Wave 2?). However, the following covariates were added: race, sex, age, education level, job classification and the number of job trainings completed during the past 12 months. These covariates were selected because they are associated with both hourly wages and living in a particular area in existing economic literature^(38,39) .

To address RQ5 (was the policy associated with changes in household income categories between the cities from Wave 1 to Wave 2?), the research team estimated a multinomial logistic DID regression model because the proportional odds assumption was again violated. Standard errors were clustered at the level of the individual. The following covariates were included in the models: race, sex, age, number of adults living in the household, marital status, education level, job classification and the number of job trainings completing during the last 12 months^(38,39) .