Coffee Drinking Is Widespread in the United States, but Usual Intake Varies by Key Demographic and Lifestyle Factors^1,2,3

Study design.

Starting in 1999, the NHANES became a continuous annual survey that collects data on the health, nutritional status, and health behaviors of the noninstitutionalized civilian resident population of the United States in a timely manner in order to address emerging public health issues and to provide objective estimates on health-related topics. A complex, multistage, probability sample design is used to select ∼12,000 persons/2-y cycle. Survey participants report demographic, socioeconomic, dietary, and health-related history during a household interview (29). In addition, they participate in a comprehensive health examination at a mobile examination center (MEC). The dietary interview component of NHANES, called What We Eat in America, is conducted by trained interviewers with the use of the USDA automated multi-pass method and can be used to estimate the types and amounts of foods and beverages consumed. To allow for estimates of usual intakes, two 24-h dietary recalls have been administered in each continuous NHANES cycle since 2002. The first 24-h dietary recall is conducted in the MEC, and the second is completed 3–10 d later by telephone. For NHANES 2003–2004 and 2005–2006, all English- or Spanish-speaking participants aged ≥2 y who completed at least one 24-h dietary recall interview were eligible to complete the FFQ component (previously referred to as the NHANES Food Propensity Questionnaire), which queried about frequency of consumption but not portion size and was developed to be used in combination with 24-h recall data to estimate usual food intakes (25). The FFQ component was mailed to eligible participants (n = 12,285), and the response rate was 92%.

The NHANES protocol was developed and reviewed to be in compliance with the Department of Health and Human Services’ Policy for Protection of Human Research Subjects and was approved by the National Center for Health Statistics’ Research Ethics Review Board. Federal laws protect all data (29). Our analyses used the publicly available NHANES data.

Study population.

To maximize sample size and to increase the reliability of usual coffee intake estimates for population subgroups, we used data from NHANES 2003–2004 and 2005–2006. Participants aged ≥20 y who had completed ≥1 reliable 24-h dietary recall (i.e., met the minimum criteria for quality and completeness established by the USDA’s Food Surveys Research Group) and reported frequency of coffee intake in the previous year on the FFQ were included in this analysis (n = 6219; Figure 1). The overall response rates for the household interview and health examination were 79% and 76% in 2003–2004 and 80% and 77% in 2005–2006, respectively. Of those who were examined in the 2003–2004 and 2005–2006 MEC, 92% and 84% completed reliable in-person and telephone-administered 24-h dietary recalls, respectively. Data from 3 additional cycles of NHANES (2007–2008, 2009–2010, and 2011–2012), which had 24-h dietary recall data but lacked FFQ data, were used to assess trends in population mean coffee consumption over time. Participants aged ≥20 y who completed a reliable in-person 24-h dietary recall were included in this analysis (n = 24,950). Similarly, for NHANES cycles 2003–2012, the mean response rates for the household interview and health examination across the 5 cycles were 78% and 75%, respectively. Of those who were examined in the MEC during this 10-y period, 92% completed a reliable in-person 24-h dietary recall.

Statistical analysis.

We estimated the amount of total reported coffee intake from all coffee-containing beverages, including decaffeinated and regular coffee made from ground or instant coffee and coffee mixes [USDA Food and Nutrient Database for Dietary Studies (FNDDS) 5.0 food code numbers 92100000–92193020]. We defined “coffee drinkers” as those who reported drinking coffee in the previous year on the FFQ. A small number of participants (n = 105) reported “none” to the FFQ item “How many cups of coffee, caffeinated or decaffeinated, did you drink?” yet reported drinking coffee on one or more 24-h dietary recall. We classified these participants as coffee drinkers and imputed a frequency value on the basis of the participant’s mean coffee intake from their two 24-h dietary recalls (coffee intake from the first recall was used if a second recall was not completed), as well as their age, sex, and racial/ethnic group according to the hot-deck method (30). The prevalence of coffee consumption is the estimated proportion of coffee drinkers, and the chi-square test of independence was used to evaluate whether the prevalence of coffee consumption differed by subgroups. We defined “acute coffee intakes” among coffee drinkers as the amount of coffee reported in the first day of the 24-h dietary recall and estimated the distribution of acute coffee intake as percentiles of intake per day. We defined “daily coffee drinkers” as coffee drinkers who reported consuming coffee ≥1 time/d in the previous 12 mo according to their FFQ. Among coffee drinkers, the distribution of usual (i.e., long-term average) coffee intake was estimated by using the NCI method, which has been described in detail elsewhere (28), to incorporate 2 d of dietary recall, an FFQ, and demographic data. Unpaired t tests were used to test for differences in mean usual intakes between each subgroup and the referent group for a given characteristic. The P values from these t tests comparing mean usual intakes are approximate because they are based on only the SEs of the means and do not account for the correlation between pairs of means resulting from the cluster sample design of NHANES and the model-based estimation.

In brief, the NCI method uses a 2-part model: the first part modeled the person-specific probability of consuming coffee on a particular day and the second part modeled the person-specific amount of coffee typically consumed on the days that coffee was consumed. This 2-part modeling allows for correlation between the person-specific random effects included in each of the parts of the model (28). The NCI-developed SAS macros, MIXTRAN version 2.1 and DISTRIB version 2.1 (31), were used to generate variable estimates and predicted intakes after covariate adjustment for the frequency of coffee consumption as reported on the FFQ (<1 cup/mo, 1–3 cups/mo, 1 cup/wk, 2–4 cups/wk, 5–6 cups/wk, 1 cup/d, 2–3 cups/d, 4–5 cups/d, or ≥6 cups/d) (26), day of week of the 24-h dietary recall [weekday (Monday–Thursday) or weekend (Friday–Sunday)], sequence of dietary recall (first or second), age group (20 to <30, 30 to <40, 40 to <50, 50 to <60, or ≥60 y), sex, race/ethnicity (non-Hispanic white, non-Hispanic black, or Hispanic or other race), education (less than a high school education or high school graduate), annual household income (<$55,000 or ≥$55,000), smoking status (never, former, or current), alcohol drinking (nondrinker or <14 or ≥14 g ethanol/d; 14 g ethanol = 1 US standard drink), and general health condition (excellent or very good, good, or fair or poor). Indicator variables, specifying whether a categorical covariate value was missing, were used to retain observations in the analyses with missing covariate data. A Monte Carlo method with 100 simulations/individual was used to estimate the distribution of usual coffee intake. The analyses were weighted by FFQ sample weights that were derived from the NHANES 2003–2004 and 2005–2006 sample weights and modified to account for FFQ nonresponse. To take into account the complex stratified hierarchical cluster sample design of the NHANES when estimating SEs, we used balanced repeated replication variance estimation with 32 replicates for the combined 2 cycles of NHANES; this work used the computational resources of the NIH’s High-Performance Computing Biowulf cluster.

To evaluate trends in population mean intakes, we considered coffee intake from the first 24-h dietary recall in relation to the NHANES survey cycle (2003–2004 through 2011–2012). Sample weighted multiple linear regression models were adjusted for age, sex, and race/ethnicity and predicted marginal means for coffee intake were used to estimate adjusted population mean coffee intakes and SEs (32). The predicted margins were calculated by using linear regression models with main effects for the subgroup variables age, sex, and race/ethnicity and an interaction between NHANES survey cycle as a categorical variable and each main effect. By using a Wald F test, P values for heterogeneity of trends between the levels of the subgroup variable of interest and NHANES survey cycle as a continuous variable were computed. All of the analyses were conducted by using SAS (version 9.3; SAS Institute) or SUDAAN (version 11.0; RTI International) software.

Prevalence of coffee consumption.

The estimated proportion of coffee drinkers was 74.7%, representing 154.4 million US adults aged ≥20 y according to data collected during 2003–2006 in a nationally representative sample (Table 1). The prevalence of coffee drinking did not differ by sex or general health condition or by education or household income (all P > 0.05) but differed considerably by age (P < 0.001): 59% of US adults who were between the age of 20 and 30 y were coffee drinkers compared with 86.4% of US adults who were aged ≥60 y. The prevalence of coffee drinking also differed by race/ethnicity (P < 0.001) and smoking status (P < 0.001). More than 75% of non-Hispanic whites and Hispanics/other races were coffee drinkers but only 61.4% of non-Hispanic blacks were coffee drinkers, and >80% of former and current smokers were coffee drinkers but only 67.2% of never smokers were coffee drinkers.

The estimated proportion of daily coffee drinkers was 49.0%, representing 101.1 million US adults (Table 2). The prevalence of daily coffee drinking differed by sex (P = 0.03), but otherwise patterns were similar to those for overall coffee drinking.

Usual coffee intake.

On average, the usual coffee intake among coffee drinkers in the United States is 14.1 ± 0.5 fluid ounces (fl oz)/d or ∼417 mL/d (1 fl oz = 29.6 mL) (Table 1). Approximately 63%, 35%, and 9% of coffee drinkers consumed ≥8, ≥16, and ≥32 fl oz/d, respectively (data not shown in Table 1). Among coffee drinkers, the 50th, 75th, and 95th percentiles of usual intakes were 11.4 ± 0.4, 19.7 ± 0.6, and 39.0 ± 1.8 fl oz/d (337 ± 9, 583 ± 18, and 1154 ± 53.3 mL/d), respectively (Table 1). Approximately 83% of US adult coffee drinkers consumed less than three 8-fl-oz cups of coffee/d and 95% consumed <5 cups; therefore, 12% of coffee drinkers consumed in the moderate, 3–5 cups/d, range (data not shown in Table 1), assuming 8 fl oz (237 mL) represents 1 standard US cup.

Patterns observed for the prevalence of coffee drinking extended to our estimates of daily consumption among coffee drinkers. No considerable differences in the distribution of usual intakes were observed by educational attainment, income, or self-reported general health condition (Table 1); however, on average, men drank significantly more than women (P < 0.001). For example, among coffee drinkers, 40% and 11% of men but only 29% and 6% of women consumed ≥16 and ≥32 fl oz/d, respectively (data not shown in Table 1). Usual coffee intakes also varied by age. For example, coffee drinkers aged 20–30 y [5.1 ± 0.5 fl oz/d (151 ± 15 mL/d)] had a significantly lower mean usual intake than those aged 50–60 y [18.0 ± 1.2 fl oz/d (533 ± 36 mL/d); P < 0.001]. The distribution of usual coffee consumption also varied by race/ethnicity, with significantly higher intakes among non-Hispanic whites [16.0 ± 0.6 fl oz/d (474 ± 18 mL/d)] than among non-Hispanic blacks [7.3 ± 0.5 fl oz/d (216 ± 15 mL/d); P < 0.001] and Hispanics or other races [9.3 ± 0.4 fl oz/d (276 ± 12 mL/d); P < 0.001]. Finally, the distribution of usual coffee intake differed by smoking and alcohol drinking status. Coffee drinkers who were never smokers had lower usual coffee intakes than former (P < 0.001) or current (P < 0.001) smokers. Those who did not consume alcohol had usual intakes similar to those who consumed alcohol infrequently (i.e., <14 g ethanol/d; P = 0.79) but lower usual intakes than frequent alcohol drinkers (i.e., ≥14 g ethanol/d; P = 0.04).

Acute compared with usual coffee intakes.

We compared the prevalence and intake amount as estimated by using 1 d of recall (i.e., acute intakes) with usual intakes as estimated by the NCI method. The population mean acute intake was 11.6 ± 0.5 fl oz/d (343 ± 15 mL/d; Supplemental Table 1), which, as expected, was similar to the estimate of 10.5 ± 0.3 fl oz/d (311 ± 9 mL/d) obtained by using the NCI method to integrate all available dietary data (data not shown in table). Among coffee drinkers, the median usual intake amount was also similar to the median acute intake; however, larger differences in estimates were observed at the tail-end of the distribution: the estimated 95th percentile of acute intake was >20% larger than the corresponding percentile of usual intake. In addition, the prevalence of coffee consumption was substantially underestimated by using only the first day of dietary recall (51.8%; data not shown in table) compared with the use of both days of recall plus an FFQ (74.7%; Table 1).

Trends in coffee intake.

Estimates of population mean coffee intakes remained consistent over the 10-y period (2003–2012) examined (P-trend = 0.09). We observed no significant differences in trends in population mean intakes between adults across categories for age (P-heterogeneity = 0.10), sex (P-heterogeneity = 0.30), or race/ethnicity (P-heterogeneity = 0.79) (Table 3).