Factors affecting recent food price inflation in the United States – Adjemian – 2024 – Applied Economic Perspectives and Policy

Food prices in the United States are rising at a historic pace, and have been since mid-2021. According to the U.S. Bureau of Labor Statistics, over the year ending in August 2022, the price of food paid by urban Americans increased by 11.4% (BLS, 2022). As shown in Figure 1, that annual increase is the highest on record since April 1979. In fact, each of the year-over-year food price inflation observations from May 2022–December 2022 exceeded 10%. Figure 1 shows that, when grouped by decade, average food price inflation in the 2020s more closely resembles the steep food price increases of the 1970s and early 1980s, compared with the relatively tranquil intervening decades.

BLS’ inflation measurement is comprised of prices for a basket of product groups, weighted by their share of consumption for typical Americans. While the pace of overall and within-product-group inflation has varied since the onset of the pandemic, its extent is broad-based and robust. Figure 2 displays month-over-month contributions of major product categories to the monthly changes in U.S. food prices, including both food at home (FAH, i.e. grocery expenditures such as meats, cereals, and dairy) and food-away-from-home (FAFH, i.e. restaurant expenditures). Grocery prices spiked following lockdown orders in early 2020, given production disruptions at major meat-packing plants and surges in (precautionary) FAH demand. Prices again increased conspicuously during the summer of 2021, as the economy began to recover from the pandemic onset, and following several rounds of fiscal and monetary stimulus efforts by US lawmakers and regulators. Within grocery products, meat led price rises through much of 2021, but by January 2022 inflationary pressures broadened out significantly throughout all major food product groups. Figure A1 of the Appendix shows that while just 7.5% of food products experienced year-over-year inflation levels of more than 5% in January 2020, in July 2021 about a quarter of the product groups did, and by September 2022 that figure stood at more than 91%. Food inflation is therefore not confined to a subset of items hit by idiosyncratic supply disruptions, like cereals and wheat products affected by Russia’s invasion of Ukraine in early 2022,¹ or egg supply damaged by highly pathogenic avian influenza.^2,3 Rather, food price inflation in 2022 broadly entrenched throughout the U.S. food basket.

Of course, food is not the only product to exhibit notable recent price increases. To the contrary, in the wake of the pandemic, inflation is widespread across the United States and the globe. Figure 3 shows that the prices for food and nonfood (or energy) items—core items—alike increased at a faster pace after the onset of the pandemic.⁴ In the figure, food and core prices trend in the same direction over time, but food prices exhibit more volatility, likely due to more inelasticity in their supply and demand relative to, for example, industrial goods. Observers offered a range of reasons for why inflation increased after the pandemic onset, although they can be classified into two general perspectives (Furman, 2022), differing according to the side of the economy they support as most responsible for the rise in prices. Demand-driven explanations point to the swift rise in real economic output during the COVID rebound (CRS, 2021), magnified by the steps policy makers took to increase the monetary base and disposable income at a time of lockdowns and (partial) stockouts⁵; changes in food consumption behavior and reductions in mobility restrictions as the pandemic evolved also likely played a role. Supply-driven explanations focus on the unexpected reduction in productive capacity given the lockdowns themselves, the still-tight labor market that the recovery produced, and other supply chain issues like intermediate goods shortages and transportation bottlenecks that may have in part been driven by a COVID-compelled preference shift from services to goods, or disruptions caused by Russia’s invasion of Ukraine in early 2022.

Prices ration supply and demand, and economic theory predicts that they rise following positive shifts in demand, negative shifts in supply, or some combination of both. Each side of the market is likely responsible to some degree for post-pandemic food price inflation, which is especially concerning for lower income American households who are less capable of shielding their earnings and wealth (BIS, 2021). One important distinction that is generally cited about the cause of inflation, however, is that its persistence is an indication of its cause: lasting inflation tends to be demand-driven, while supply-driven inflation alleviates naturally as supply chains adjust (CRS, 2021). This paper seeks to understand the degree of association between several demand and supply factors with changes in food prices; we use retrospective time series techniques to estimate how food prices typically change following evolutions in factor-level series, and how much these shocks relate to the variation and historical realizations in U.S. food prices. The latter are accomplished with impulse response functions (IRFs), forecast error variance decompositions (FEVDs), and historical decompositions (HDs) from structural vector autoregressive models (SVARs).⁶ Given the monthly frequency of the data and the lack of strictly exogenous shocks, we cannot necessarily identify causal relationships; however, our results provide a first step in that direction, offering indicative evidence for the possible reasons behind recent food price rises.

Outside of a handful of articles, the recent agricultural economics literature is mostly silent on the question of why U.S. food prices move the way they do. This is not especially surprising, given the stability in domestic prices over the last several decades evident from Figure 1; periods of high food prices seem to invite academic inquiry into the topic. Much of the extant work is tied to a food “crisis” period, tends to be descriptive in nature, is focused on commodity rather than retail or customer prices, and examines international rather than domestic prices. For example, in a series of reports prepared for the Farm Foundation, Abbott, Hurt, and Tyner (2008, 2009, 2011) document and offer detailed explanations behind the runup in agricultural commodity prices toward the end of the 2010s. Although comprehensive in identifying potential reasons, the authors do not attempt to model the question, empirically. Hochman (2014) quantifies the causes of observed high prices during that period but is focused on international commodity prices, rather than the prices US consumers pay for food, which include additional components like the costs of processing, packaging, transportation to retail markets, and advertising.

Three important exceptions—each focused on recent price runups—deserve mention. Baek and Koo (2010) note that the all-food Consumer Price Index (CPI) increased during the second half of 2007,⁷ and model the relationship between it and the Bureau of Labor Statistics (BLS) indices for prices received by the producers of farm commodities and energy products, and an index for the real effective exchange rate; they find that food prices in the first decade of the 2000s are related to energy prices and the exchange rate in the short and long run. Lambert and Miljkovic (2010) apply a similar approach as Baek and Koo (2010) to a longer time series—this time focused on the U.S. food and beverage CPI, but use an expanded set of variables: in addition to food prices, farm and energy costs, they include data on wages and consumer income. They find farm prices and wages to be most significant in determining food price changes. Irz, Niemi, and Liu (2013) likewise model retail food prices (in Finland) as a function of the costs of agricultural raw materials, energy, and labor. They do not directly consider the effect of income due to data limitations, but try to proxy for it by adding a trend to their model. The authors find that the Finnish food prices they study are mainly determined by agricultural prices, while energy and wage costs play significant but less robust roles.

Other empirical work not explicitly associated with food price “crisis” periods studies the general effect of various factors on commodity prices. Dorfman and Lastrapes (1996) use Bayesian time series methods to show that a positive shock to the U.S. money supply increases both domestic crop and livestock prices in the short run. Awokuse (2005) combines directed acyclic graphs and time series techniques, and shows that macroeconomic variables affect agricultural commodity prices to a large degree, although he estimates that the money supply has a relatively small impact on their variation. These authors use different empirical techniques and model different variables over timeframes that only partially overlap. Moreover, the rate of money supply increase was for many decades fairly steady—limiting the variability that researchers could exploit.

The U.S. government’s policy response to the pandemic, both monetary and fiscal, represents a significant departure from that regime (CRS, 2021). According to the Committee for a Responsible Federal Budget (CRFB), the federal government has so far provided Americans with $10.9 trillion in various forms of relief, just over half the entire (nominal) size of the domestic economy in 2019 (BEA, 2020). Official U.S. government figures roughly agree with that assessment. The Pandemic Response Accountability Committee (PRAC), created by the Coronavirus Aid, Relief, and Economic Security (CARES) Act estimates fiscal support of various programs at $5.2 trillion (2022). Monetary relief was equally substantial. Between February and June 2020, the U.S. Federal Reserve’s (Fed’s) balance sheet (assets) increased by over $3 trillion; it grew by an additional $1.8 trillion through April, 2022 (Fed, 2022)—together, the Fed’s assets more than doubled over that timeframe.⁸ Because the Fed purchases those assets from sellers with currency, an increase in balance sheet assets implies an increase in the domestic money supply. The U.S. M2 monetary aggregate, which represents money and “near money” such as money market, accounts increased by 40% over the same period.⁹

That level of stimulus is unprecedented in historical terms. Indeed, as a share of the GDP, the federal budget deficit grew to 14.9% in 2020, the first time that percentage increase reached double digits since the Second World War (CRS, 2021). It is also far larger than the last stimulus effort the U.S. government undertook. In response to the financial crisis of the late 2000s, the federal government’s fiscal response through the American Recovery and Reinvestment Act amounted to $831 billion (CBO, 2012). As nominal GDP for the United States in 2007—the last calendar year before a cascade of bankruptcies—totaled $14.5 trillion, that earlier round of stimulus was about 6% of the size of the economy. On the other hand, as a share of pre-crisis GDP, the fiscal stimulus response to the pandemic, at 25% of GDP, is roughly five times larger than its response to the financial downturn. In addition to sharply increasing the money supply, the federal stimulus raised personal incomes—an unusual observation during an economic downturn since incomes generally correlate positively with the business cycle—especially through its direct economic impact payments and enhanced unemployment benefits (CRS, 2020). Whether or not the pandemic-era stimulus contributed to economy-wide inflation is an open empirical question, although recent results suggest that it did (see, e.g., de Soyres, Santacreu, & Young, 2022, and Jordà, Liu, Nechio, & Rivera-Reyes, 2022).

Another aspect to consider regarding food price pressure is the changing nature of food purchases observed over the course of the pandemic. As stay-at-home orders prevented Americans from visiting restaurants, they shifted spending from FAFH to FAH; full-service establishments saw the most dramatic declines, although limited-service eateries were also affected (Okrent & Zeballos, 2022). Total expenditures on food dropped steeply, too (USDA ERS, 2023). However, by March of 2021, U.S. expenditures on FAFH recovered to pre-pandemic levels, and have continued to rise (and outpace FAH expenditures) following pre-pandemic trends. That same month, total food expenditures outpaced the previous high observed in December, 2019, and have since then also increased.

Ongoing empirical work seeks to explain how much of the observed economy-wide inflation since the pandemic by either supply and demand shocks. Shapiro (2022) introduces an approach to disentangle supply and demand shocks at the PCE category level, using a straightforward binary identification strategy,¹⁰ while Di Giovanni, Kaelmli-Ozcan, Silva, and Yildrim (2022) specify a global input–output model and use it to estimate which side of the market contributed to CPI price rises. Both authors show that demand- and supply-driven shocks are responsible, with the demand side explaining around half of the recent inflation. Neither study explores food price inflation in much depth.

To study factors associated with U.S. food prices, and what accounts for their typical and recently observed patterns, we gather comprehensive data related to the same demand and supply-side factors explored by authors of related studies (e.g., the size of the money supply, consumer income, and prices for food inputs like agricultural commodities, energy, and labor), as well as an index that represents global supply chain stress, in light of the impact that it may have on the availability of food. Given their relationship to food price inflation in Figure 3, we also include core prices in our analysis. By doing so, in effect we measure how these other factors contribute to observed food price inflation while controlling for the influence of core inflation. For robustness, we verify that our headline results maintain when altering the order of our analysis: allowing either demand or supply variables explanatory primacy.

Using innovation accounting of structural vector autoregressive (SVAR) results, three of the four identification strategies we use agree that, on average, increases in the money supply, income, wages, supply chain stress, and the prices for energy, transport, and farm products tend to lead U.S. food prices and are associated with its unexpected variation. Core price rises tend to predict food price increases a few months out, but have the opposite effect in the short and longer term; yet, their ability to explain the unexpected variation in food prices is somewhat limited—in three of the four models we estimate, their changes account for less than 10 percent of the forecast error variation in food prices up to 20 months out. For the food price spike observed in the wake of the pandemic onset, our results indicate that while both sides of the market are responsible for food price shocks, unexpected changes in the demand-side factors we model and especially the money supply are more notable than they have been in the recent past—even after including core prices (which may themselves be affected by changes to the money supply and income) directly in the model. The monthly frequency of our data, the correlation in the variables we model, and lack of true exogenous shock series do not permit us to make causal statements about our results. Why changes to demand factors have grown more prominent in explaining the path of food prices is therefore not precisely known. It may be that while government stimulus in response to the pandemic supported household well-being and expenditures at a time of elevated uncertainty and supply chain disruption, it may have also contributed to notable food price inflation by boosting demand as lockdowns and disruptions eased. On the other hand, the rising prominence of demand factors may also be related to the rapid release of pent-up demand or preference changes for a higher intensity of food purchases associated with the lifting of pandemic lockdowns.

MODELING APPROACH AND DATA

We use time series methods to estimate the role that various factors play in the determination of domestic food prices. Specifically, we estimate a set of covariance stationary SVARs with data chosen according to economic theory (see, e.g., Irz et al., 2013) for relevant supply-side (wages, energy price, a supply chain pressure index, the price of transportation, and the price of farm commodities) and demand-side variables (the money supply and per-capita disposable income). These variables are deliberately chosen because they have in most cases been considered by previous researchers as potential food price determinants, and are meant to represent either the factors that affect the production and availability of food, or the demand for it on the part of consumers. For example, food processing firms pay wages to their employees and must cover those costs in the price they charge to downstream buyers, while per-capita consumer income helps determine the American food budget and demand for food purchases. Two exceptions to this rule are the supply chain pressure index, a new data series we added to represent the prominent transportation bottlenecks observed during the pandemic period, and core prices. As shown in Figure 3, the latter trend in the same direction as food prices and serve as a key target for monetary policy choices, so their presence in the model helps us identify how various factors correlate to food price inflation while controlling for the influence of core prices.

Throughout, although we model the way that each of these variables affects the path of food prices, we routinely lump the supply-side variables together, and do so on the demand side, as well. We do this to explain how each side of the market contributes to food price changes, and also to place our results in the context of other recent work (e.g., Shapiro, 2022) investigating the nature of inflation in the wider economy. We note that because the demand-side factors we model—money supply and consumer income—were both affected by stimulus efforts following pandemic onset, they offer a proxy for the study of whether and how that stimulus might have contributed to the rise in food prices. We further note that many other factors explain changes in money supply and income, as well, and that the correlation in the variables we model at the monthly frequency as well as the lack of a series of observable exogenous shocks prevent us from identifying causality; as a result, it is not possible to distinguish the true cause of food price changes through our analysis. Yet our results can be interpreted indicatively, and invite further investigation.

While reduced-form VARs state each variable as a function of its own past values, as well as the past values of the other variables in the model, they cannot be used to establish precise relationships because their error terms are correlated; in contrast, researchers can use SVARs to estimate the impact of unexpected changes to a given variable on the other variables in the model. More specifically, we interpret the SVAR results (i.e., the outcome of the structural shocks) in this paper as indicative of an associative relationship between the variables in question—a necessary (but not sufficient) condition for establishing causality. Identification of these relationships within SVARs depends on assumptions (i.e., restrictions) made by the econometrician. Perhaps the most popular classical method, termed Cholesky identification, uses theory to order variables in terms of the primacy of their relationships in the short run: certain shocks have contemporaneous relationships with some or all of the modeled variables (see, e.g., Kilian, 2009), while others do not.¹¹ Yet Cholesky-identified models may be overly restrictive if contemporaneous shocks do not follow a recursive pattern (Uhlig, 2017). Several alternative techniques achieve identification of structural shocks via data-driven methods, which rely on the statistical properties of the data (Lange, Dalheimer, Herwartz, & Maxand, 2021). To ensure that our results are robust to other patterns of structural impacts, in addition to Cholesky decomposition we apply three data-driven identification techniques that exploit the non-Gaussianity of certain variables. Ultimately, we produce qualitatively similar results across three separate identification schemes, with one relative outlier.

The variables in $y_t$ are transformed via natural logarithm (except for GSCPI, which is an index that contains negative values), first-differencing, and de-seasoning.¹⁴ We further standardize each variable by subtracting its mean and dividing by its standard deviation; although GSCPI is standardized by its authors, we adjust it for the time period under study.¹⁵ Following Kilian and Lütkepohl (2017), we specify the reduced-form model with a conservative, 12-month lag order (avoiding issues that can arise with data-based selection of lag order), suppressing deterministic terms for explication:

$y_t=A_1{y}_{t-1}+\dots +A_{12}{y}_{t-12}+w_t$, where $A_i$, $i=1\dots 12$ are coefficient matrices and the reduced-form stochastic residuals $w_t$ are serially uncorrelated white noise. The mutually uncorrelated structural shocks ${\epsilon }_t$ (which are used to generate impulse responses, variance decompositions, and historical decompositions) are recovered through the identification of the nonsingular matrix $B$, with ${\epsilon }_t=B^{-1}{w}_t$ (this relation connects the structural shocks of interest to the realizations of the variables we model). The covariance matrix of reduced-form residuals ${\sum }_w$ is therefore related to the (diagonal) structural shock matrix ${\sum }_{\epsilon }$ through the decomposition matrix $B$, with ${\sum }_w=B{\sum }_{\epsilon }{B}^{\prime }$. Because that relation holds for any matrix that accomplishes the decomposition, restrictions are required to identify a unique matrix $B$. For this paper, we do so following Lange et al. (2021), who demonstrate six data-driven identification alternatives: three that exploit shock heteroskedasticity, and three applicable to data that exhibit non-Gaussianity.

Although our data do not support those identification techniques based on covariance changes,¹⁶ Table 1 shows evidence that, according to Shapiro–Wilk and Jarque–Bera tests each variable (with the possible exception of one, GSCPI) is distributed non-normally over our period of observation. This condition permits full identification of the $B$ matrix (Hewartz, Lange, & Maxand, 2021). Those same authors describe how the decomposition matrix can be estimated by (i) specifying a log-likelihood function assuming a standardized Student’s t-distribution for each component (termed non-Gaussian maximum likelihood, or NGML), or nonparametric implementation via minimizing a dependence criterion: either (ii) the distance covariance (DC) or the (iii) Cramér–von Mises distance (CVM). Through simulation analysis, Hewartz et al. (2021) demonstrate that—at least in their stylized scenarios—each of these three data-driven methods generates results robust to heteroskedasticity and structural shock distribution (as long as it is non-Gaussian) of unknown form.

Under all four methods of identification, we use innovation accounting to understand how each variable in the data contributes to the realization of food prices in the United States, both on average and also during the most recent period of high inflation. Impulse response functions (IRFs) display the average path of food prices, over time, of an unexpected (one standard deviation) rise in a modeled variable. We trace out impulse responses using our SVAR results as well as local projections (LPs) (Jordà, 2005).¹⁷ LPs may offer some advantages, depending on the horizon under consideration (Brugnolini, 2018); Plagborg-Møller and Wolf (2021) show that both methods estimate the same IRFs in the population under stationarity and an unrestricted lag structure, but have different finite-sample properties. Local projections are also less sensitive to specification errors since they are estimated via horizon-specific regressions, while VAR misspecifications are compounded at each horizon (Ramey, 2016). Forecast error variance decompositions (FEVDs) measure the fraction of the error in forecasting a given variable attributable to shocks to the modeled variables, on average; they signify the importance of each variable in explaining observed variations of a given variable over time. Re-writing $y_t$ as a function of past structural shocks, the $h$-step ahead forecast error is the sum of the shocks experienced over the horizon, and the FEVD indicates the share each variable makes in the mean squared prediction error generated by the model (Kilian & Lütkepohl, 2017). Finally, historical decompositions (HDs) document how shocks to the variables in the system predict observed changes in a given variable (from its forecasted value); they can be used to examine the effect of shocks at different time periods. HDs are approximated by combining the recovered structural errors with the IRFs to estimate the cumulative contribution each variable makes to the variable of interest (in our case, food prices) at any point in time (Kilian & Lütkepohl, 2017). As HDs permit comparison of the relative importance of the modeled variables over time, we use them to understand how changes to the various supply and demand factors relate to recent food price increases.

RESULTS AND DISCUSSION

Figure 4 displays the time path of the transformed variables in our analysis. U.S. food prices increase over time, but do so at a faster pace after pandemic onset, in early 2020. Core prices follow a similar path, but with less variability. Wages paid to US workers trace a likewise increase. The money supply and per-capita income do too, but each exhibit sharp increases that correspond to policy shocks meant to offset the effects of the pandemic (CRS, 2020; CRS, 2021). Every other variable in the data rises more sharply in the wake of the COVID shock, but only GSCPI rises in early 2020 as, for example, port backups soared in response to the Coronavirus. Energy, farm, and transport prices each fell at that time, as aggregate demand in the United States shifted inward. We use innovation accounting to document how each of these variables relates to U.S. food prices and their short-to-medium run variation.

Impulse responses

IRFs trace the expected path over time, on average, in an outcome variable following an unexpected change in another modeled variable. Each pane in Figure 5 shows our estimates of the path in domestic food prices (in standard deviations) after a single standard deviation shock to the impulsed variable. Different series in the chart represent the mean IRFs for all identification/estimation techniques we employed: recursive Cholesky SVAR (black solid), NGML (purple dash), DC (orange dash), CVM (blue dot-dash), and LP (black dotted). Green shading represents the 90% confidence band around the LP IRF.¹⁸

Across identification techniques, the IRFs are quite similar in most cases, although the CVM model results sometimes stand apart. For example, each signifies that a single standard deviation rise in the (change in) money supply, wages, and supply chain pressure will precede an increase in the domestic food price, in the short run. Income has a positive relationship in the very near term. The IRFs also show that core price rises predict food price increases a few months out, but have the opposite effect in the short and medium term. Likewise, all but the CVM models predict—as expected—that food prices are positively related to their own shocks. In addition, all models predict that a rise in transport prices and farm prices predicts a food price increase in the short-to-medium run. In other cases, nearly all models agree on the path of each IRF. Of the IRFs in Figure 5, only energy shocks do not appear connected to food price changes during the period of observation, on average. As shown in Figure A2, the alternative ordering (which places supply- ahead of demand-oriented factors) produces very similar IRF results, with only the CVM model again acting as somewhat of an outlier.

Forecast error variance decompositions

Figure 6 presents FEVDs for all four decompositions; like IRFs, FEVDs represent the average situation—in this case error variance decomposition—over the period of observation. FEVDs in the figure portray the contribution of unexpected changes in each modeled variable to the n-period-ahead unexpected variation (forecast error) in food prices, with each panel representing a 20-month ahead horizon. In all four panels, time t shocks to the other modeled variables explain a large fraction of the variation observed in food prices. After 4 months, they together explain upwards of 50% (and in some cases, substantially more) across identification methods. Unexpected shocks to food prices themselves, unaccounted for by the other variables in the system, capture the residual explanatory power.¹⁹

In our baseline specification, for both the recursive and data-driven methods, average food price error variance is mostly explained by shocks to supply-side factors (e.g., the cost of inputs like wages, and prices for energy, transport, farm inputs, and supply chain backups). These range in significance depending on the identification method, but the price of energy, transport, and farm inputs stand out. For three of the models, shocks to the demand-side factors (the money supply and consumer income) together represent around 25% of the short- and long-run variation in food price variation, while changes in core prices account for a smaller amount.²⁰ The CVM model is once again an outlier, though, as it attributes between 35% and 40% of food price variation to changes in demand factors beginning at the second–period ahead horizon, and associates around 20% of that variation to unexpected changes in core prices. Our alternative, supply-side-first specification shown in Figure A3 likewise attributes somewhat more of the variation in food prices to supply-oriented factors: by the 20-month ahead horizon, all four identification methods assign between 70% and 80% the variation that way. Demand factors take up between 10% and 15%, and unexpected changes in core prices make up the remaining 5%–10%.

Historical decompositions

While IRFs and FEVDs document how, on average, food price changes and their variation are related to changes in the supply and demand factors that we model, HDs portray the contribution that changes to those factors make to the time path of food prices over the period of observation.²¹ Reviewing those associations across the time dimension is helpful in understanding why observed food price changes may have occurred. Our specific interest is in the most recent runup in food prices, beginning with the COVID shock in early 2020 and continuing through the present.

The panels in Figure 7 show HDs generated by each of the four SVARs we estimated. All panels plot as a red line the shock of interest: the standardized change in the (logged and de-seasoned) domestic food price. The colors in the HD charts represent the estimated involvement each of the modeled variables has in the change in food prices; for each month, their vertical sum is equivalent to the red line, the observed change in U.S. food prices. For the most part, the HD results match the FEVD findings, as expected. Changes in supply-side factors like prices for energy, farm products, labor, and transport, as well as supply chain stress each make up a large share of the HD charts.²² Demand-side factors, represented by the money supply and per-capita income, tend to associate more with food prices once they began a sustained and dramatic increase in 2021. Differences across the HD panels generally match the differences in the FEVDs. Relative to the other identification techniques, the money supply is more strongly associated with food price variation according to DC identification (Figure 6c), so more purple appears in its HD (Figure 7c).

Although the dominant influence of the supply-oriented factors is ever-present, three of the HDs we estimate (Cholesky, NGML, and DC) highlight the association between demand-oriented factors and changes in the path of food prices after the onset of the pandemic. All three attribute a substantial share of the persistent food price rises since 2021 to contemporaneous sharp increases in M2, and all indicate that money supply increases have strong explanatory power for persistent food price rises since 2021. The Cholesky model also assigns (shocks that affect) core prices an important role in determining the price of food; it makes up a substantial share of Figure 7a. As core prices increased following the pandemic shock, food prices increased along with them.²³ For the other three techniques, whose identification is data driven and less sensitive to variable ordering, changes in core prices are more limited in their contribution to the monthly changes in food prices. Two of the identification techniques that estimate the HDs in Figure 7 (NGML and DC) indicate that income shocks also contributed to observed food price inflation since early 2020. However, the CVM-identified HD in Figure 7d—again serving as an outlier—reports a less prominent role for the demand-oriented factors in food price inflation.

Figure 8 displays the percent of the absolute shocks to food prices explained by the absolute shocks to our demand factors (after accounting for the movement of core prices). Three of the identification methods we use (Cholesky, NGML, and DC) agree that during pre-pandemic periods, represented by the 5-year value between 2015 and 2019, demand factors explain a smaller portion of absolute food price changes than they do in the 2021–2022 period. As expected based on Figures 7 and 8, changes in supply factors provide the strongest contribution to the path of food prices. However, even after including changes in core prices directly, the importance of demand shocks in recent food price determination is stronger over the last couple years.²⁴ Because the demand factors we model were affected by government stimulus efforts, Figure 8 provides indicative evidence that stimulus through monetary and fiscal policy may have contributed to observed food price rises.

Figure A4 in the Appendix displays similar HD results to Figure 7 for all the identification methods we estimate, except that it assigns more influence to the supply-oriented factors on food prices. This is to be expected because the results are generated using our alternate model, which places those variables before demand factors in the ordering. Likewise, Figure A5 places lower weight on the association between absolute demand innovations and food prices, although the pattern it depicts is similar to Figure 8: the same three models agree that demand-oriented factors become more important in contributing to the realizations of food prices beginning with the pandemic onset.

CONCLUSIONS

Food prices in the United States are rising at a faster pace than they have in over 40 years. Observers have offered a range of explanations for why this is happening, from COVID-driven supply disruptions and lockdowns, to Russia’s invasion of Ukraine and the stress it placed on international commodity markets, to the stimulus U.S. policy makers put in place to assist in the economic recovery from the pandemic. But so far, no academic work has attempted to quantitatively differentiate among these explanations.

We use retrospective time series techniques to decompose the BLS food price CPI, in order to understand the way different supply– and demand-oriented determinants contribute to food price changes in normal times, as well as how they may have affected the sharp rises observed in the wake of the pandemic. Due to data and modeling limitations, we are unable to establish causal relationships, so our findings should be interpreted indicatively. Most of our models agree that, on average, positive shocks to the money supply, disposable income, wages, energy prices, supply chain stress, transport prices, and farm prices tend to lead to increases in U.S. food prices and affect its unexpected variation. Core prices also play a role, although their effects vary with the time horizon considered.

Across identification schemes, the supply-side factors we model make up the dominant portion of contributions to food price changes over time. Yet, beginning with the onset of the pandemic, the demand factors in our models (the money supply and per-capita U.S. income, leaving aside core prices, which are also potentially affected by these variables but modeled explicitly) grow in importance in the contribution they make to realized food price changes—by about 20% relative to the previous 5-year period (on average). Because these demand-side factors were affected by monetary and fiscal stimulus programs, which supported economic activity during and after the initial pandemic-driven contraction, our results suggest that stimulus may be partially responsible, among other factors, for observed food price inflation. Other potential explanations for a growing role of demand factors in food prices include the rapid release of pent-up demand or preference changes generated by the lifting of pandemic lockdowns. Our findings invite further research to investigate precisely why recent rising food prices appear more sensitive to demand pressures.