Conclusions Although this study is preliminary, its results indicate that the relationship between poor nutrition and test scores may in fact be quite negative, strengthening the impetus for schools to consider policies that support students' healthy eating. In order to strengthen these findings and investigate possible mechanisms through which poor nutrition might affect test scores, there is a need for further research.

Results Standard ordinary least squares regression showed that test scores decreased as reported fast‐food consumption increased. In the propensity‐score‐matched analysis, which controlled for 25 student background characteristics, higher‐than‐average fast‐food consumption (‘four to six times in the last 7 days’ or more) was associated with significantly lower test scores in both reading (−11.15 points or 0.48 SD) and math (−11.13 points or 0.52 SD), even when teacher experience, school poverty level and school urbanicity were also included in the model.

Methods Data from the Food Consumption Questionnaire administered to approximately 12 000 fifth graders as part of the nationally representative Early Childhood Longitudinal Study‐Kindergarten Cohort was analysed using regression analysis. The analysis used propensity score matching to examine the relationship between students' reported fast‐food consumption and their test scores in reading and mathematics. Propensity score matching attempts to control for a host of background characteristics that might be correlated with both fast‐food consumption and test scores.

Background Children in the USA are experiencing obesity and overweight at epidemic rates. Schools have started to make policy decisions based on a popularly presumed connection between nutrition and academic achievement. This study aimed to determine whether such a relationship exists, and if so, its nature.

Data and methods This study uses the Early Childhood Longitudinal Study‐Kindergarten Cohort data collected by the National Center for Education Statistics (Tourangeau et al. 2006). The dataset includes extensive information on students' academic achievement, social functioning, food behaviours, physical health, day care, academic environment at home and attitudes towards school; parental income, education, employment, attitudes towards children's school, food sufficiency and involvement in children's schooling; teachers' levels of education and experience; and school attendance and poverty rates, urbanicity and other similar information. Surveys were administered to students regardless of learning disability status and were translated for non‐native speakers. Summary statistics The primary outcome variables examined in this study were the fifth grade Math and Reading Item Response Theory Scale Scores, which aim to measure students' skill levels. As shown in Table 1, scores on the reading test ranged from 58.23 to 181.22, with a mean of 141.52 (SD = 23.15), while math scores ranged from 46.97 to 150.94, with a mean of 115.24 (SD = 21.22). Additionally, as previous research would have us expect scores go up with parental income and race is correlated with test scores in this data. African‐American students had the lowest test scores in both reading and math, while white students had the highest reading scores, and Asian students had the highest math scores. Interestingly, no discernible relationships were found between students' body mass index and their scores, nor between body mass and fast‐food consumption, or between parents' income and fast‐food consumption. Table 1. Mean reading and math scores (n= 5571) Variable Mean (St. Dev.) Minimum Maximum Reading IRT Scale Score 141.52 (23.15) 58.23 181.22 Math IRT Scale Score 115.24 (21.22) 46.97 150.94 The primary independent variable of interest was consumption of fast food, which was categorically determined from the Food Consumption Questionnaire given to students in the fifth grade. The question read, ‘How many times in the last 7 days did you eat food from a fast‐food restaurant like McDonald's, Taco Bell, or Wendy's?’ and students were given the option of selecting ‘none’, ‘one to three times’, ‘four to six times’, ‘once a day’, ‘twice a day’, ‘three times a day’ or ‘four or more times a day’. The majority of students (54%) responded that they had consumed fast food one to three times in the past 7 days, but a surprisingly high almost 2% responded that they had eaten fast food four or more times per day. Table 2 shows the percentage of response in each category. Table 2. Reported fast‐food consumption Fast‐food consumption category Percentage Did not eat any in past 7 days 28.47 Ate one to three times in past 7 days 50.45 Ate four to six times in past 7 days 10.04 Ate once/day 5.89 Ate twice/day 2.08 Ate three times/day 0.90 Ate four or more times/day 2.18 Regression The first model used in this study attempts to control for the commonly accepted components of lowered academic achievement, socioeconomic factors such as race, parental income and education, and parental employment status, using ordinary least squares regression (OLS). Other independent variables of interest in the fully specified regression model included level of urbanicity of the student's school (31% urban, 42% suburban and 26% rural), his or her teacher's years of experience (mean 15.02 years), and percentage of students eligible for free and reduced‐price lunch at the school (mean 28.10%). Propensity score matching Simple OLS regression, although it gives us a good starting point, cannot control for the fact that it is possible that students who eat a lot of fast food are very different from their peers in ways not measured by the Early Childhood Longitudinal Study‐Kindergarten Cohort data. This selection bias problem plagues many non‐experimental designs. One way to correct for it is propensity score matching (Dehejia & Wahba 2002), which uses background characteristics of subjects grade to try to create equivalent, balanced groups that differ significantly only on the variable of interest (in this case, fast‐food consumption). The procedure creates groups that can be used in analysis as if they were treatment and control groups, with the group actually participating in the higher‐than‐average consumption of fast food being seen as the ‘treatment’ group, and the group whose likelihood of eating a lot of fast food is the same as the treatment group but who did not actually consume more than the average amount being seen as the ‘control’ group. In this way, propensity score matching allows us to approximate an experiment from what is merely observational data and to estimate the effects of the treatment on the treated. Students were matched on 25 independent variables thought to correspond with either fast‐food consumption or lower test scores. Parents' age, income, educational level and occupational status, as well as the number of hours per week that parents were employed, whether parents attended parent–teacher conferences, whether food security was ever a problem for the family, and how many hours per week each student spent in non‐parental care were all included. Also, each student's race, gender, age, grade level, attendance, number of books owned, interest and perceived skill in the subject (math or reading) being measured, other food behaviours (including eating sweet or sugary snacks not from fast‐food restaurants), country region of residence and body mass were included in the match process. Wherein i indexes students, j indexes schools, U is a measure of the urbanicity of the school a student attends, T measures the number of years the student's teacher has been teaching, L represents the percentage of students in the school eligible to receive free lunch, F indicates the student's fast‐food consumption, and the βs are coefficients to be interpreted. The dataset produced by the match process did not different significantly on any of the independent variables. A total of 2284 observations were available for use in the final regression estimation.

Results Fast‐food consumption was associated with lower test scores in both OLS and propensity‐score analyses. In Table 3, Model 1 shows that eating fast food once a day in the last week corresponded to −18.25 test points, and three times a day in the last week was associated with −33.88 points. Although these coefficients are significant, they get smaller with the inclusion of control variables. Column 3 shows a relationship between fast food and test scores of −9.29 points for those eating fast food four to six times in the last week, −18.15 points for those eating fast food once a day, and −28.90 points for those who ate fast food three times a day. Table 3. Ordinary least squares regression estimates: reading Model 1 Model 2 Model 3 Ate fast food four to six times in past 7 days −8.88*** (1.56) −9.17*** (1.74) −9.29*** (1.89) Ate fast food once/day −18.25*** (1.42) −18.64*** (1.51) −18.15*** (1.51) Ate fast food three times/day −33.88*** (4.23) −31.29*** (4.53) −28.90*** (4.49) # of years teacher has taught 0.15** (0.05) 0.17** (0.05) % free lunch‐eligible students −0.32 (1.28) −0.11 (1.30) City 4.91*** (1.28) Rural −5.03* (1.99) Constant 140.35*** (0.64) 137.27*** (1.11) 139.56*** (1.21) Observations 11 042 8539 8170 R 2 0.09 0.10 0.11 F 63.25 44.11 34.86 In Table 4, Model 1 shows very similar results for math scores, with students who ate fast food four to six times in the previous week predicted to have scores −7.66 points lower than those in the comparison group (one to three times in the last week), those eating fast food once a day scoring −17.36 points lower, and three times per day corresponding to −29.03 points lower on math test scores. When all the controls were included, Model 3, these numbers stayed significant. Eating fast food four to six times in the last week corresponded to −8.33 test score points, one fast‐food consumption per day was related to −16.42 points, and three times per day correlated with 26.50 points lower than the comparison group. Table 4. Ordinary least squares regression estimates: math Model 1 Model 2 Model 3 Ate fast food four to six times in past 7 days −7.66*** (1.57) −7.95*** (1.79) −8.33*** (1.95) Ate fast food once/day −17.36*** (1.37) −17.00*** (1.67) −16.42*** (1.69) Ate fast food three times/day −29.03*** (2.70) −28.31*** (2.84) −26.50*** (2.79) # of years teacher has taught 0.08 (0.04) 0.08 (0.04) % free‐lunch eligible students 0.63 (1.22) 0.52 (1.19) City 3.93** (1.30) Rural −4.5* (2.15) Constant 114.73*** (0.64) 112.66*** (1.00) 114.68*** (1.13) Observations 11055 8550 8181 R 2 0.09 0.09 0.10 F 76.93 48.27 42.84 Once data were matched for student background characteristics, a second set of regression equations was estimated, which yielded similar results. For the purposes of this regression, because fast‐food consumption was used as a ‘treatment’, categorical responses had to be re‐coded into binary groups, with 0 representing fast‐food consumption at or below average (responses ‘did not eat fast food in the past 7 days’ and ‘one to three times in the past 7 days’), and 1 representing any category above those (‘four to six times in the last 7 days’ to ‘four or more times per day’). The results for both reading and math scores demonstrated patterns similar to what emerged in the earlier regressions without student‐level controls. In reading (Table 5), students who ate more fast food than average were expected to score −11.15 points lower than those who did not, controlling for the teacher's years of experience, the percentage of students at the school eligible for free lunch, and the urbanicity of the area. This result was statistically significant. In math (Table 6), students reporting more than average consumption of fast food scored 11.13 points lower. Table 5. Ordinary least squares regression matched estimates: reading Model 1 Model 2 Model 3 Fast food −10.48*** (2.03) −11.46*** (1.92) −11.15*** (1.88) # of years teacher has taught 0.12 (0.08) 0.13 (0.08) % free‐lunch eligible students −3.18 (1.62) −2.37 (1.62) Suburban 10.97*** (2.70) Urban 9.74** (3.03) Constant 144.71*** (1.18) 143.76*** (1.64) 135.31*** (2.97) Observations 2251 2213 2213 R 2 0.03 0.04 0.08 F 26.64 13.56 12.61 Table 6. Ordinary least squares regression matched estimates: math Model 1 Model 2 Model 3 Fast food −10.97*** (1.78) −11.43*** (1.78) −11.13*** (1.76) # of years teacher has taught 0.01 (0.08) 0.02 (0.08) % free‐lunch eligible students 0.38 (1.64) 0.73 (1.55) Suburb 9.09*** (2.60) Rural 5.69 (3.07) Constant 118.65*** (1.10) 118.82*** (1.22) 112.51*** (2.38) Observations 2253 2215 2215 R 2 0.04 0.04 0.07 F 37.86 14.10 12.75

Discussion This study found statistically significant relationships between higher‐than‐average consumption of fast food and lowered test scores in math and reading, lending plausibility to the idea that fast‐food consumption is related to lowered academic achievement. However, this study is preliminary and there are several limitations to consider when interpreting the results. First, the Food Consumption Questionnaire is self‐reported data from fifth graders, and no items on the parent or school questionnaires allowed for triangulation of the information. It is possible that fifth graders do not reliably report their food consumption habits. Also, by lumping together students who ate fast food one to three times in a week, the questionnaire does not allow for a potentially important distinction between students who enjoy a once‐weekly fast‐food treat and those who eat fast food on 3 days each week. Because of attrition and movement of students, the sample size decreased considerably over the course of the study, from 22 666 in the kindergarten year to 12 029 by fifth grade. Children whose parents refused to cooperate were excluded from the fifth grade study (Tourangeau et al. 2006). It is possible that students who moved or whose parents would not participate in the study activities were more (or less) likely to have unhealthy eating habits. Finally, although propensity score matching attempts to control many factors potentially related to food behaviours and test scores, there are limits to what can be used because of the data that was collected. However, these limitations do not detract from what are significant findings. In addition, effect sizes are large: scores in both reading and math are nearly one half of one standard deviation lower for students eating more fast food. That these findings were robust to multiple specifications indicates, at the very least, a need for further study. The reasons for the correlations reported herein are not known, but several hypotheses arise which are worthy of future study. One, it is possible that the types of foods served at fast‐food restaurants cause cognitive difficulties that result in lower test scores. Although little research is available in the medical literature that points to one specific biological mechanism through which fast food could affect test scores, there is mounting evidence that consumption of fats and food additives can affect cognition. Lower‐fat diets correlate with less mental decline in older patients (e.g. Féart et al. 2009), while rats fed high‐fat diets demonstrate cognitive impairment in both short‐ and long‐term measures (Murray et al. 2009). Using data from the Third National Health and Nutrition Survey 1988–1994, Zhang and colleagues (2005) found an association between cholesterol consumption and lowered cognitive functioning in 3666 children ages 6–16. Additionally, there is evidence that artificial additives such as food colouring may increase hyperactivity in children (e.g. Bateman et al. 2004). Schab and Trinh (2004) conducted a meta‐analysis of 15 clinical trials on the effects of artificial food colouring on hyperactivity and found ‘evidence that neurobehavioral toxicity may characterise a variety of widely distributed chemicals’ (p. 423). Such hyperactivity leads to discipline and concentration problems in the classroom and could be another mechanism through which fast‐food consumption would affect test scores. Alternatively, it is possible that the propensity to eat fast food is correlated with unobserved characteristics such as parental involvement in homework, which would also affect test scores. The robustness of these finding indicate that further study is needed, including a better measure of exactly what food behaviours students engaged in and how often. If student‐reported food behaviours could be corroborated with parent‐ or teacher‐level data, and multiple imputation could be used to correct for the possible bias of data missing not completely at random, a stronger causal link could be hypothesised. Another interesting research direction would look at the timing of fast‐food consumption and academic outcomes, for example, fast food eaten immediately before test‐taking might have a different effect than fast food eaten days before a test. If they can be replicated, the findings of this study have uncertain policy implications. Indeed, this study raises more question than it answers. Because the fast‐food behaviours examined herein take place off school campuses, what can be done to prevent students from purchasing these foods? Recent legislative efforts in San Francisco to take toys out of Happy Meals, however, indicate that there are possibilities. For schools, continued investment in school nutrition plans and curricula designed to make students and parents aware of the academic consequences of their food choices, as well as significantly expanded parental outreach programmes and lobbying of governmental agencies for assistance, would be positive steps to take. Additionally, further attempts to bolster school funding so that schools are not forced to rely on contracts with outside fast‐food vendors would help to send students the right messages about healthy eating. Additional study that brings the medical and educational fields together will help further uncover the relationship between fast food and test scores, which could be is of crucial importance for the health and academic success of our students.

Key message • American children are experiencing an epidemic of obesity, likely related to poor nutrition.

• Little is known about the connection between poor nutrition and academic achievement.

• This study found that children who eat more fast food have significantly lower scores in both math and reading.

• More research is needed to understand this connection and how it should inform school nutrition policies.