Greenhouse gas emissions and energy use associated with production of individual self-selected US diets

We conducted a systematic search in Web of Science and Google Scholar databases. Search terms included combinations of 'LCA' and 'life cycle' with 'food'. Further refined searches targeted individual underrepresented foods. In addition, collected citations were cross-referenced with the extensive review by Clune et al [22] and relevant additional citations were included. The literature review was limited to reports available in the public domain. Articles and reports written in English and published in 2005–2016 that applied LCA methods to one or more food products and provided primary (i.e. not cited from elsewhere) mid-point impact assessment of GHGE and/or CED were reviewed and inventoried. Peer reviewed journal articles as well as thoroughly documented reports from governmental and non-governmental organizations (including theses) were considered. Additional details about our methodology as well as the full listing of references included in our database, dataFIELD⁵, is provided in supporting information available at stacks.iop.org/ERL/13/044004/mmedia. For database consistency, mid-point indicator values were adjusted to a functional unit of 'kg of food,' with meat and fish/seafood adjusted to 'kg of edible boneless weight'. See supporting information for details on conversion factors used.

To link environmental impacts to the 332 commodity foods in FCID, we followed a four-step process (see table 1). First, we used data from original research on specific foods inventoried in the literature review, as described above. The mean, standard deviation, minimum and maximum values for CED and GHGE at farm gate and at processor gate were calculated for each specific food, and then matched to the FCID. Studies of heated greenhouse vegetable production or those of beef from dairy herds were not included in our averages because information on market share of these production methods is unavailable or unreliable. Second, if we did not have an original research report on an FCID food, we turned to reports with previously-compiled food LCA data to supply environmental impacts [23–29]. These resources contained data not captured in the literature review, perhaps due to non-English language reports or proprietary sources. Overall, for stage 1 and 2 of the linkage process, CED matches were made for 35% of the food commodities, and GHGE matches for 47%. Third, remaining FCID foods were populated with values from similar foods as proxies. Specifically, we took an average of either CED or GHGE values from existing entries within a specific food grouping (e.g. berries, brassicas, brassica greens, citrus, fresh herbs, grains, other greens, nuts, roots, dried spices, other tree fruit, tropical fruit) to proxy for a specific food item in that same grouping that was lacking data. Failing this approach, other proxies of foods with similar form were then assigned. These assignments were based on similarities of specific crops in their botany and, most importantly, production methods, as determined by the expertise of our research team. Values that were assigned from other foods in the database in this third stage accounted for 50% of CED values and 39% of GHGE values. Fourth, the FCID dataset includes minimally processed forms of fruits and vegetables (e.g. strawberry juice, dried apples). Where direct LCA matches were not available for these forms, we applied a mass conversion factor, gathered from nutritional databases [30, 31], to the base fruit or vegetable in order to approximate the agricultural production burdens of these processed forms. This stage accounted for the remaining 15% of CED and 15% of GHGE values for FCID foods. For juices, vinegar and maple syrup, additional sources were used to develop valid estimates. These additions are detailed in supporting information.

Because of the inconsistency in full life cycle boundary conditions across the literature review entries, cradle-to-farm gate impact factors were chosen for the vast majority of foods. This choice is further supported by the fact that these commodity foods, in many cases, become ingredients in processed, as-consumed foods, and inclusion of life cycle stages downstream from the farm gate would not necessarily reflect impacts of the actual foods consumed. The exceptions to this farm gate boundary condition are foods within the FCID listing that require processing: flours, refined sugars, vegetable oils, etc supporting information contains additional details on these boundary condition choices, as well as an environmentally extended input-output based estimate of the cumulative food processing impacts excluded in this analysis.

Variability is expected in the LCA data gathered for a given food type, both due to differences in agro-climatic conditions and production practices, as well as LCA methodological approaches such as allocation choice. To characterize this variability and estimate its influence on the impacts of diet, we calculated a 95% confidence interval around the average impact for each food, based on the observations for that food (or related foods) that we found in the literature. If there were too few observations for a given food to calculate a confidence interval, we used the confidence interval for a related food or group of foods. We used the lower and upper bounds of this confidence interval in subsequent calculations of diet-level variability. Supporting information also contains details on this method.

The FCID database contains a recipe file that links foods as reported by NHANES respondents to ingredients in the form of commodities. For example, the recipe for 100 grams of 'lasagna with meat,' which is one of 17 lasagna dishes reported by respondents, contains gram quantities of commodities, including wheat flour, milk, beef, tomato, etc. We linked impacts from dataFIELD to these FCID commodities and adjusted for recipe quantities and amounts eaten in order to assign impacts for each food consumed during the 24 hour recall day as reported by each respondent. See supporting information for a complete listing of FCID foods and impacts. In some cases, when there was sufficient LCA literature to describe the impact of processed foods (specifically: cheese, yogurt, tofu, beer, carbonated drinks, and liquor) we linked directly from dataFIELD to NHANES, without using the FCID recipes. For alcoholic beverages, we created our own recipe file for linking from FCID to NHANES. The impacts of edible losses were calculated for the amount of each commodity consumed using loss factors from the USDA's Loss-Adjusted Food Availability (LAFA) data series [32]. Commodity items were assigned retail (edible food lost at outlets such as supermarkets and restaurants) and consumer (cooking losses and uneaten food) loss factors from the matching LAFA commodity. If there was not a direct match, the food was assigned loss factors for something similar (e.g. apple juice factors for apricot juice) or for an average of similar items (e.g. an average of loss factors for blueberries, raspberries, and strawberries for huckleberries). After assigning impacts to each food consumed, we summed impacts over the entire day for each individual. All analyses accounted for the NHANES sampling weights and survey design parameters.

Our comprehensive literature review resulted in 1645 entries (combinations of food types and production scenarios) from 321 unique sources (listed in supporting information). System boundaries varied across the LCA studies inventoried: while nearly all entries considered some form of agricultural production, 51% accounted for processing beyond farm gate, 19% followed products through to retail/regional distribution hubs, and 6% included some form of use (consumption) phase. Supporting information contains additional information on the distribution of entries by publication type and geographic origin of production.

Environmental impacts were assigned for the 332 unique food commodity forms in the Food Commodities Intake Database (FCID), and for seven additional foods linked directly to NHANES. Table 2 shows the distribution by literature review entries broken down by broad FCID food groups, with meat, fruit, vegetables and dairy accounting for more than half of the entries. Table 2 also shows the number of FCID foods in each of these groups, as well as the percentage of foods in these groups requiring proxy values. While 55% of foods required proxy in calculating GHGE for the total diet, these foods accounted for only 3% of total impact. This is because the foods requiring proxies tend to be low impact and less frequently consumed foods. For example, the meats group contributed 57% of dietary GHGE (see table 4), but only 0.1% of this group's impact came from proxies. There were a number of proxies used in the legumes group, accounting for 27% of the GHGE from this group. However, legumes contributed only 0.3% of total dietary GHGE (table 4), so proxied legumes account for only 0.09% of total dietary GHGE.

The NHANES dietary intake survey is representative of the US population. Thus, linking dataFIELD to the individual, self-selected diets from NHANES offers a way to estimate the distribution of diet-related impacts across the population on a given day. Table 3 summarizes these results at the distribution mean for the total population on both a per day basis as well as normalized to 1000 kilocalories (kcal) dietary intake. Tables 3 also demonstrates the contribution of food losses to the environmental impact of diet.

Zoom In Zoom Out Reset image size

Figure 1 provides the distribution of diet-related GHGE per person per day across the self-selected diets from NHANES. An analogous figure for CED is included in supporting information. While the distribution in figure 1 shows a sharp rise to a peak in the distribution curve at around 2.2 kg CO₂ eq person⁻¹ day⁻¹, there is also a long tail on the distribution (truncated in figure 1: truncated tail represents 1.3% of total impacts). The 20% of diets with the highest carbon footprint account for 45.5% of the total diet-related emissions. Also shown in figure 1 is a range of distributions representing the influence of variability in emissions due to food production methods and LCA modeling.

The distribution of GHGE by food group for all 1 day diets as shown in table 4 is quite typical of Western dietary patterns, with the dominant impacts coming from meats and dairy. An analogous table for CED is included in supporting information. For the total population, 80.6% of the meats group GHGE comes from beef, 9.5% from poultry, 8.5% from pork, with other meats making up the remaining 1.5%. Of interest is the relatively high (5.9% of GHGE, 16.0% of CED) contribution from beverages (tap and bottled water, carbonated drinks, coffee, tea, juices, beer, wine and spirits). Beverages have not always been identified as a separate food group in past diet impact assessments, but were the third most impactful group in our analysis. Across the total population, fruit and vegetable juices make up 33% of the GHGE in the beverages group, followed by coffee, beer, carbonated drinks, cocktails, and wine at 20%, 19%, 9.6%, 8.9% and 7.0%, respectively. Bottled water, tap water and tea contribute less than 2% each.

Table 4 also shows how the contribution by food group differs between lower-impact (1st quintile) and higher-impact (5th quintile) diets. For the higher-impact diets, meats account for 70% of total diet GHGE, whereas they only account for 27% with lower-impact diets. In part, this has to do with the makeup of the meats group in each of the two quintiles. Whereas poultry is the largest contributor in the first quintile (55% of meats group GHGE), beef contributes 91% of meat GHGE in the fifth quintile diets. Although for some food groups the percent contributions in table 4 decline from the 1st to the 5th quintiles (e.g. dairy % contribution to GHGE drops from 28% to 11%), absolute impacts for both GHGE and CED increase for all food groups between the first and fifth quintile. This is largely because diets in the 5th quintile have greater amounts of these foods than in the 1st quintile, as suggested by the overall caloric intakes. In terms of overall impact, increases in beef intake account for 72% of the absolute increase in GHGE between the diets of the first and fifth quintile.

GHGE associated with the fifth quintile are 7.9 times those of the first quintile. Total caloric intake is an important factor in ranking impacts per day, as the consumption of more food calories typically translates into greater environmental impact. The fifth quintile consumes, on average, 2.25 times the kilocalories of the first quintile. However, even when impacts are normalized by caloric intake, GHGE of the fifth quintile are five times those of the first quintile. Although this distribution of dietary data from NHANES is representative of diets in the US on any given day, there are a couple of caveats. First, self-reported diets typically understate actual intakes. Second, a distribution of one-day diets is more disperse than a 'usual intake' distribution. The low caloric intake reported for the first quintile is, in part, a result of both of these issues. Although we do not have a way to calculate underreporting in this sample, we do know that 26% of respondents in the first quintile reported that their consumption on the recall day was much less than usual.

The literature review that underlies development of dataFIELD found that LCA studies that can be used to link to dietary choices have increased significantly in recent years, but data gaps still exist for many food types. This is consistent with other recent reviews (e.g. [22]). A scan of the foods requiring proxy assignments in table S4 (supporting information) offers a sense of current data gaps and a target for LCA practitioners interested in filling such gaps. In addition, many foods important to evaluation of healthful diets with low impact—nuts, legumes, meat substitutes—are poorly represented in the literature and deserve additional attention. Geographical representation is biased toward Europe. As has been customary in the diet-LCA literature, our main estimates for diet-level impacts are based on average LCA values applied to each food consumed. However, unlike other studies, we have addressed variability due to production practice, geography, or LCA method by calculating upper and lower bounds of impacts for each food and carrying these estimates through to diet-level impacts. As the NHANES dietary recall data does not specify production methods or geographical origin, we cannot be more precise in assigning impacts from LCA studies to foods eaten by NHANES respondents. However, geographical specificity becomes increasingly important with other impact categories such as water use, eutrophication, or land use. Although currently available data in these categories are limited, we plan to expand our database to water and land use impacts, specific to the US food market, in a future iteration.

Zoom In Zoom Out Reset image size

It has become common practice in diet impact studies to assign proxy foods as approximations in the case of missing data, but to our knowledge, this paper is the first attempt to quantify the contributions from those proxy assignments. Table 2 indicates that proxy foods contribute 3% to the average diet GHGE, and 8% to CED. Proxy assignments are made based on foods with similar production characteristics. However, even if we assume that all of our proxy estimates are in error by a factor of 2 (i.e. all proxy impact factors are doubled), the mean diet-level impacts would still only increase by 2.6% for GHGE and 8.2% for CED.

This study demonstrates the disproportionate impacts that can be caused by some types of self-selected diets. Figure 2 displays the cumulative emissions of these diets when ranked in order of GHGE per person per day. GHGE associated with the fifth quintile of diets are nearly eight times that of the first quintile and three times that of the third (middle) quintile. If the top quintile of diets (representing 44.6 million Americans on a given day⁶) shifted such that their associated GHGE were aligned with the mean impact, this would represent a one-day reduction in GHGE of 0.27 million metric tons CO₂ eq. (mmt), equivalent to eliminating 661 million average passenger vehicle miles⁷ on a given day.

This shift—which could be done by changing foods, reducing calories, or some combination of these two—would be represented graphically in figure 2 by removing the section of the curve above the average emission diet line for the fifth quintile. Current economy-wide US net emissions (based on 2015 data [34]) are 1023 mmt above the target levels in year 2025, as submitted to the U.N. Framework Convention on Climate Change (UNFCCC) [35]. The hypothetical diet shift described above, if implemented every day of the year and met by equivalent shifts in domestic production, would account for 9.6% of remaining reductions necessary to meet the target. (see supporting information for the emission reduction calculations.) Even if high emission diets (arbitrarily defined here as >25 kg CO₂ eq. person⁻¹ day⁻¹; the truncated tail extending above the representation in figure 2) are excluded from the estimate based on a presumption that they are either atypical or that such individuals are unlikely to shift diets, moving the remainder of the high quintile (GHGE >6.9 but <25 kg CO₂ eq. person⁻¹ day⁻¹) to the mean GHGE still accomplishes 9% of the reductions necessary for the US to meet the UNFCCC target. See supporting information for a parallel discussion on the cumulative impacts of food losses. Our estimates of reductions are likely to be somewhat exaggerated because a distribution of 1 day diets is known to be more dispersed than a distribution of usual diets [36]. This is one of the limitations of using NHANES. Since it is based on the 24 hour diet recall tool, it also tends to underestimate total energy intake, although this is true of all self-reported diet instruments [37]. In fact, 24 hour recalls provide more details about foods consumed and tend to be less biased than food frequency questionnaires [38]. Moreover, NHANES provides the only ongoing nationally representative source for information about individuals' diets. Our analysis highlights the importance of looking at individual behaviors rather than just population means, since there is clearly a wide range of impacts being caused by self-selected diets.

Table 5 offers a comparison of the results from this study with other reported estimates of the impacts of the US diet, as well as self-selected diets in other countries. When excluding studies that include broader boundary conditions, there is strong agreement across results, with a coefficient of variation of 3% for GHGE with US diets only, and 7% across all diets. Self-reported diet surveys carry a well-known under-reporting bias [39, 40] whereas food balance based estimates (production + imports—exports—non-food uses ± changes in stock) are often considered to be overestimates [41, 42]. A more refined food type characterization and a more exhaustive literature review were utilized in this study in comparison to that of Heller and Keoleian [11]. While beverages have not always been delineated as such in previous studies of diet impacts [11, 12, 15, 19, 43–44], we find them to be important contributors. This finding is further strengthened by the fact that packaging and use phases are not included within the boundary conditions of our estimates. Packaging often represents a hotspot in the life cycle impacts of beverages [45–49], and use phase activities (heating water, brewing coffee) can be important for hot beverages [50, 51].

The boundary conditions for the current study are cradle to farm gate for most food commodities, and include processing for the collection of FCID foods that are minimally processed ingredients (flours, oils, juices, etc). As such, our reported values should be considered underestimates of actual impacts associated with food consumption in the US as they include the production impacts of processed food ingredients, but not the impacts of processing itself. Using the US Environmentally Extended Input Output model developed by US EPA [52] and an approach detailed in supporting information, we estimate that food processing not captured in our bottom-up estimates amounts to 15% of the total cradle to processor gate (including agricultural production sectors) GHGE. Packaging materials represent an additional 6%. Inclusion of these missing food processing and packaging contributions would raise our estimates by ~27%, although it is important to note that these input-output based approximations are made for the food and agricultural sectors in aggregate, and will not apply evenly across different food types or for specific diets (i.e. they apply only at the mean).

In figure 1, we demonstrate the influence that food impact factor variability, as represented by our literature review, has on the diet-level impacts of the US population. Based on this estimated variability, the GHGE for the mean of the population ranges from 3.8–5.6 kg CO₂ eq. person⁻¹ day⁻¹, or ±19% of the value based on average impact factors. Variability of food production systems across geographies and production methods is expected. In most cases, the granularity of available LCA data is not sufficient to reasonably and consistently differentiate between these food production variables. On top of this, methodological choices within LCAs, such as how impacts are allocated between co-products, introduce an additional level of variability between studies that cannot be effectively disaggregated from production variability. It is important to keep in mind, however, that even if such environmental impact data were complete, the corresponding information in diet databases is not available. NHANES represents the best information on diet—both in the aggregate and in its diversity across the population—available for the US Yet, it does not (currently) contain information on the methods of production for food sources (e.g. was a tomato grown in a heated greenhouse? Was it organically grown?) or geographic origin of production (California? Michigan? Chile?). To further refine these estimates would require more information on foods in the NHANES survey, as well as better LCA data on food production variability.