Three thousand seventy-one secondary school children were asked to participate in a two-part questionnaire survey, and 2610 agreed. Two thousand thirty completed the questionnaires at T1, and 2307 completed them at T2. One thousand six hundred sixty completed the survey at both time-points, giving an attrition rate of 18.23%.

Information relating to dietary intake, caffeine consumption, demographics, and stress, anxiety, and depression levels in this sample has already been discussed in previous publications (19–21). However, it is considered useful to provide a quick overview of the demographic background of the sample again here. Table 1, therefore, includes information relating to which academy and year group the children belonged, whether they were male or female, the presence/absence of special educational needs (SEN) status, eligibility/ineligibility to receive free school meals (FSM), ethnicity, whether or not English was spoken as an additional language, and whether or not the children were looked after by a non-parental guardian. Whereas previous publications have simply included the information relating to T1 or T2, Table 1 also includes descriptive statistics for those who only participated at T1, only participated at T2, participated at both time-points, or did not participate at either.

The sex ratio was relatively balanced at both T1 and T2, and approximately 20% came from each year group of British secondary education (i.e., years 7–11). However, although similar numbers of children eligible to receive FSM were observed at both time-points, considerably more with a SEN status were present in the sample at T2 compared to T1. Most children who took part in the study were White, spoke English as their first language, and were not looked after by a non-parental guardian.

The Diet and Behaviour Scale [DABS; (19)] was used to provide a measure of both the frequency and amount of consumption of common food and drink products that might have effects on psychological outcomes. This scale consists of 29 items; 18 record frequency of consumption on a five-point scale (1 = never, 2 = once a month, 3 = once or twice a week, 4 = most days [3–6], 5 = every day), 11 record the amount consumed per week or per day. Four of these latter items were used to calculate weekly caffeine intake from energy drinks, cola, coffee, and tea. Richards et al. (19) reported the DABS to be associated with a four-factor structure consisting of junk food, caffeinated soft drinks/gum, healthy foods, and hot caffeinated beverages. Subscale scores for the junk food and healthy foods factors are used as covariates in the current study; the other two factors were not controlled for due to their sharing variance with the predictor variable.

Three items were used to provide an indication of exercise participation; frequency of mildly energetic, moderately energetic, and vigorous exercise were all recorded on a four-point scale (1 = three times a week or more, 2 = once or twice a week, 3 = about once to three times a month, 4 = never/hardly ever). Finally, participants were asked to state how good they considered their general health to have been over the previous 6 months (1 = very good, 2 = good, 3 = fair, 4 = bad, 5 = very bad). This last question was chosen because it has been suggested that, when examined as an outcome variable, the best way to operationalize health status is by using a global single-item measure (22). Furthermore, single-item measures of self-reported health status have been used in population studies for over half a century, can reduce time-costs associated with multi-item measures, and have been shown to be significantly and independently predictive of a number of specific health problems, mortality, use of health services, changes in functional status, and recovery from ill health [see Ref. (23) for an overview].

Teachers at three secondary schools administered pen and paper questionnaires to their students at both time-points in a classroom setting. Participants completed the questionnaires at the same time as those other students from their class who were taking part. Demographic information was then acquired directly from the School Information Management System (SIMS). These data, which were collected at both T1 and T2, included sex, age, school attended, school year, presence/absence of a SEN status, eligibility/ineligibility to receive FSM, ethnicity, whether English was spoken as an additional language, and whether or not the child was cared for by a non-parental guardian (do note however that sex and age were collected through SIMS at T1, but through questionnaire responses at T2).

Total weekly caffeine intake was estimated from the number of cans of energy drink and cola, and cups of tea and coffee consumed per week. As the first aim of the current paper was to identify demographic and lifestyle correlates of caffeine consumption in adolescents, this variable was used to explore such relationships by using Spearman’s correlations, between-subjects t-tests and one-way analysis of variance (ANOVA).

The second aim of the current paper was to investigate associations between caffeine intake and general health. To do this, general health was dichotomized into high and low: those who answered with “very good” or “good” (answers 1 and 2) made up the high general health group, and those who answered with “fair,” “bad,” or “very bad” (answers 3, 4, and 5) comprised the low general health group. Total weekly caffeine intake was categorized into six groups: 0 mg/week, 0.1–250 mg/week, 250.1–500 mg/week, 500.1–750 mg/week, 750.1–1000 mg/week, and >1000 mg/week. This approach was identical to that used by Richards and Smith (21) when examining associations between caffeine intake and stress, anxiety, and depression in this same sample of secondary school children. The method was chosen as it allows for the detection of dose-dependent relationships and can also be used to establish a level of consumption above which an effect occurs, rather than simply demonstrating whether the variables are correlated. Previous studies (3, 24) have used intake above the 90th percentile to indicate heavy caffeine consumption. It is, therefore, worth noting that the highest intake group included in the current study (>1000 mg/week) represented a similar level of intake, with 90th percentile consumption being 1031 mg/week and 1054.3 mg/week at T1 and T2, respectively.

The above caffeine consumption groups were investigated cross-sectionally in relation to the dichotomous general health variable using Chi-square tests for linear association. Binary logistic regression (enter method) was then used to investigate these relationships further, while controlling for additional variance. Based on the methodology of Richards and Smith (21), the following covariates were entered into the models along with the predictor variables: diet (junk food and healthy foods DABS subscale scores; those for caffeinated soft drinks/gum and hot caffeinated beverages were not entered into the regression models due to them sharing considerable variance with caffeine consumption), demography (sex, school attended, school year, presence/absence of a SEN status, eligibility/ineligibility for FSM), lifestyle [average sleep hours, school attendance, exercise frequency; this latter variable was a factor analyzed score consisting of mild, moderate, and vigorous exercise, the details of which have been reported by Richards et al. (19)]. Note that although information relating to ethnicity, speaking English as an additional language, and being looked after by a non-parental guardian were recorded, they were not entered as covariates due to some of the groups being comprised of very small numbers of participants. Further to this, school year was included as a covariate rather than age. This was, first, to avoid multicollinearity (Spearman’s correlations determined that the two variables were very strongly correlated: T1 r[3013] = 0.972, p < 0.001; T2, r[2003] = 0.947, p < 0.001) and, second, because school year may capture additional variance, such as a child’s current Key Stage of secondary education. The analysis of caffeine and general health at T1 utilized covariates from T1, whereas the analysis from T2 used covariates from T2.

Health effects relating to caffeine could be attributable to trimethylpurinedione on the adenosine receptor or alternatively to other allelochemicals found in plants containing trimethylpurinedione. Therefore, it may be problematic simply to pool the effects of caffeine consumed from sources such as tea and coffee with those of cola and energy drinks. For this reason, it was deemed important to further investigate associations between general health and caffeine by separately examining each of the four sources of the substance measured by the DABS. Following the methodology used by Richards and Smith (21), participants were categorized as non-consumers, low consumers, or high consumers in the following manner: energy drinks (non, 0 mg/week; low, 0.1–133 mg/week; high, >133 mg/week), cola (non, 0 mg/week; low, 0.1–25 mg/week; high, >25 mg/week), coffee (non, 0 mg/week; low, 0.1–160 mg/week; high, >160 mg/week), and tea (non, 0 mg/week; low, 0.1–120 mg/week; high, >120 mg/week). Univariate level cross-sectional analyses were then conducted using Chi-square tests for linear association to help determine whether relationships between caffeine intake and general health might differ depending on the source from which the substance is obtained.

As two cross-sections of data were available, the third aim of the current paper was to investigate whether caffeine intake at T1 was predictive of general health outcomes 6 months later at T2. A Chi-square test for linear association was conducted, which was followed up with binary logistic regression. The same covariates (from T1) as used in the multivariate cross-sectional analyses were again included here.

A change-score analysis was conducted in order to address the fourth aim of the paper, to determine whether increasing caffeine consumption between the two time-points was associated with decreasing general health. This analysis proceeded with a Chi-square test at the univariate level, to examine whether increasing caffeine consumption (as opposed to not increasing caffeine consumption) was associated with decreasing (as opposed to not decreasing) in general health. This was then followed up with binary logistic regression, which again utilized the same covariates from T1 as used in the previous multivariate level analyses.

As retrospective self-reporting of caffeine has been shown to be reliable [e.g., Ref. (25)], the DABS items that record the weekly number of cans of energy drink and cola, and cups of tea and coffee were used to calculate estimates of caffeine intake. The following values were assigned: can of energy drink (133 mg); can of cola (25 mg); cup of coffee (80 mg); and cup of tea (40 mg). The values for coffee, tea, and cola were chosen based on updated versions of those used by Brice and Smith (26), which were in turn based on those used by Barone and Roberts (3) and Scott et al. (27). The value for energy drinks was the mean caffeine content of the three products most commonly reported in the sample. The questionnaire did not consider differences in portion size, differentiate between instant and brewed coffee, or ask about the consumption of caffeine pills. The sum of caffeine intake from coffee, tea, energy drinks, and cola was calculated to provide an estimate of total weekly intake. The amount of caffeine consumed was found to be similar at both time-points: T1 M = 419.84 mg (SD = 526.76), T2 M = 421.77 (SD = 550). However, Ahluwalia and Herrick (16) have stated that studies of caffeine consumption typically report mean intake for either the whole sample or for all caffeine consumers, but not for both. It is, therefore, worth noting that the mean intake for those who used caffeinated products was markedly higher: T1 N = 1764, M = 471.25 (SD = 535.94), T2 N = 1973, M = 470.29 (SD = 560.79).

The current section aims to examine relationships between total weekly caffeine consumption and demographic and lifestyle variables. Significantly higher total weekly caffeine intake was observed in males compared with females, T1 t(1899.051) = 3.191, p = 0.001, T2 t(1948.766) = 4.262, p < 0.001, and in those who were eligible to receive FSM compared with those who were not, although this latter effect was only marginally significant at the first time-point: T1 t(1951) = 1.962, p = 0.05, T2, t(274.616) = 2.936, p = 0.004. Those with a SEN status were observed to consume significantly more caffeine than those without a SEN status at T2, t(2169) = 2.316, p = 0.021, although no such effect was detected at T1, t(1978) = 1.305, p = 0.192.

One-way between-subjects ANOVAs determined that caffeine consumption did not differ significantly between the three schools investigated, T1 F(2, 1979) = 2.044, p = 0.13, T2 F(2, 2199) = 0.741, p = 0.477, although Spearman’s correlations showed that intake was positively associated with school year, T1 r(1951) = 0.113, p < 0.001, T2 r(2146) = 0.221, p < 0.001. Caffeine intake was also negatively associated with average number of sleep hours, T1 r(1908) = −0.196, p < 0.001, T2, r(2115) = −0.199, p < 0.001, and school attendance, T1 r(1951) = −0.082, p < 0.001, T2 r(2150) = −0.157, p < 0.001, though no relationship was observed with exercise frequency, T1 r(1829) = 0.02, p = 0.381, T2, r(2055) = 0.016, p = 0.464.

Around 50% of the sample at both time-points reported their general health to be “good.” The majority of other respondents reported their health to be either “very good” or “fair,” and relatively few responded that their health was “bad” or “very bad” (for frequencies, see Table 2). For this reason, the measure was dichotomized into high and low, with those who answered “very good” or “good” (answers 1 and 2) making up the high general health group, and those who answered with “fair,” “bad,” or “very bad” (answers 3, 4, and 5) comprising the low general health group. This method allows for the risk of reporting less than ideal health to be assessed in relation to the amount of caffeine typically consumed.

Total weekly caffeine consumption varied considerably between participants, with mean intakes of 419.84 mg/week (SD = 526.76) and 421.77 mg/week (SD = 550) being observed at T1 and T2, respectively. Although these values are considerably lower than those reported in older populations [e.g., 1369.92 mg/week in US undergraduate students; (28)], caffeine consumption in a considerable proportion of participants is likely to have exceeded the recommendation of the EFSA NDA Panel (15) that children should avoid consuming >3 mg per kg of bodyweight per day. This in itself is concerning, as high consumption in young consumers has been associated with a range of undesirable outcomes, such as difficulty sleeping and morning tiredness (29), falling asleep at school (30), increased anger (31), violence and conduct disorder (32), and decreased school achievement (33).

The first aim of the current study was to identify demographic and lifestyle correlates of caffeine use in secondary school children. It was therefore interesting to note that caffeine consumption was higher in males than in females, and correlated negatively with average number of sleep hours. These findings reflect similar observations made in adult populations [e.g., Ref. (34, 35)]. In addition, children with a SEN status, and those eligible to receive FSM were also found to report elevated caffeine intake. A linear association was observed in which consumption increased throughout secondary school education, a trend that has previously been reported in children and adolescents (16). Caffeine intake was also negatively associated with school attendance, which may be a particular cause for concern considering the likely knock-on effects regarding academic attainment. Identifying such correlates in secondary school children is also useful because they can subsequently be controlled for when utilizing multivariate approaches to data analysis, thus reducing the chances of the true nature of effects being obscured by confounding factors.

The univariate cross-sectional analyses uncovered negative linear associations between total weekly caffeine intake and general health. These appeared to reflect increased occurrence of low general health in very high (>1000 mg/week) caffeine consumers, though an additional protective effect was observed in the 0 mg/week group at T2. However, although the association between caffeine intake and general health remained significant at T2 after additional dietary, demographic, and lifestyle covariates had been controlled for, the effect at T1 disappeared entirely. It is, therefore, not straightforward to interpret these findings. Further analysis of the individual sources of caffeine did, however, suggest that these negative associations relied on the consumption of caffeine from cola and energy drinks rather than from tea or coffee.

Although univariate associations between caffeine intake and general health existed cross-sectionally, total weekly consumption at T1 was not associated with general health at T2. In addition, no association was observed between changes in caffeine intake and changes in general health between the two time-points. These findings, therefore, suggest that the initial associations observed cross-sectionally are unlikely to have a causal basis. However, a possible explanation for these null findings is that, as mentioned above, caffeine and general health only remained significantly associated at the multivariate level in one cross-section of data. As the longitudinal analyses necessarily relied upon data from both cross-sections, this somewhat surprising finding might, therefore, have confounded the results.

The findings from the current study suggest that associations between caffeine intake and health in adolescents likely differ from those observed in adults. For instance, although an umbrella analysis conducted by Grosso et al. (12) reported that coffee and caffeine intake are typically associated with health benefits in adults, findings from the current study showed that high caffeine consumption was (at least at the cross-sectional level) associated with low general health, and that caffeine intake from coffee was not associated with health at all. The observation that caffeine consumed from cola and energy drinks appeared to explain the negative associations is also in line with the findings of a recent review article by Visram et al. (36). This paper showed that energy drink consumption in children and young people was associated with higher risk of health-damaging behaviors and physical health symptoms, such as headaches, stomachaches, insomnia, and hyperactivity, and suggested that such effects might be dose-dependent. Further research (both in relation to cola as well as energy drinks) is, therefore, required to better understand the nature of these effects.

Although an obvious advantage of the current research was that the samples examined were large, they were not entirely representative of the schools from which they came in regards to certain aspects of demography [representativeness of the samples has been discussed in previous publications; for T1, see Ref. (19); for T2, see Ref. (21)]. In addition, they related to a very specific population, both in terms of demography (e.g., age, socioeconomic background, ethnicity), and geographical location. For these reasons it is suggested that further research should aim to investigate the effects of caffeine on the health of adolescents in more representative samples. Further to this, adolescents commonly gain in size and weight during a 6-month period, and these variables were not controlled for in the current study.

An additional issue with the current study is that it relied upon a self-reported measure of general health. Although such measures can be reliable, they have rarely been used in samples of children, and even less so in relation to food/nutritional consumption. It is also possible that children might have misinterpreted the term “general health.” For instance, children might have considered small/medium injuries when responding to this measure, such as those picked up playing sports, which are unlikely to be related to diet. It is, therefore, questionable whether children and adolescents are able to answer such a question in a meaningful way.

The methodological nature of the study presented here meant that the longitudinal analyses faced a number of limitations not initially incurred at the cross-sectional level. One such problem is that only 6 months separated the collection of the two cross-sections of data. This close temporal proximity may have made it impossible to detect long-term effects of caffeine use over time. Future research could, therefore, address this issue by conducting longitudinal studies that leave a greater amount of time between initial data collection and subsequent follow-up.

Another limitation of the current study is that the samples at T1 and T2 differed more than was initially expected (which might also provide an explanation as to why the multivariate cross-sectional association between caffeine and general health was significant only at the latter time-point). In particular, the percentage of pupils with a SEN status was considerably higher at T2 (29.2%) compared with T1 (21.8%). This specifically reflected increases in the percentages of children with a SEN status present in Academies 2 and 3: Academy 2 T1 SEN = 17.1%, T2 SEN = 25.5%; Academy 3 T1 SEN = 26.1%, T2 SEN = 40.8% (the proportion of pupils with a SEN status at Academy 1 being 25.2% at both time-points). The large increase in percentage of participants with a SEN status from Academy 3 may also have reflected the observation that the number of pupils from this academy who were present in the sample increased by 223 between the two time-points; this was in stark contrast to Academies 1 and 2, which gained only 17 and 12 pupils, respectively. A likely explanation for this is that teachers at Academy 3 did not administer questionnaires to all classes at T1. In addition, some pupils may not have realized that they had consented to take part until the second time-point. This latter observation could therefore explain the relatively higher response rate at T2 (88.4%) compared with T1 (77.8%).