How consumer behaviours changed in response to COVID-19 lockdown stringency measures: A case study of Walmart

Abstract

Walmart is a major player in the US retail sector and was one of the grocery corporations that bucked the trend of declining retail sales at the start of the COVID-19 pandemic in 2020. Particularly in the initial stages of the pandemic, governance priorities focussed on restricting the movement of people and closing non-essential retailers and service providers to slow the spread of the virus and keep people safe. This paper investigates the impact of non-pharmaceutical interventions, in the form of lockdown stringency measures, on consumer purchasing behaviours for essential goods over the onset of the pandemic. Focussing on both instore and online sales outcomes for Walmart in the US, we examine changes between pre-pandemic trends in two different sales outcomes, sales transactions and total spend, and trends in 2020. We then employ a series of multi-level regression models to estimate the impact that imposed stringency measures had on these sales outcomes, at both national and state level. Results indicate that nationally consumers were making fewer, larger physical shopping trips and huge increases in online sales was seen ubiquitously across the country. Novel and expansive insights from such a wide-spread retailer, such as Walmart, can help retailers, stakeholders and policy makers understand changing consumption trends to inform business strategies and resilience planning for the future. Furthermore, this study highlighted the value of examining spatial trends in sales outcomes and hopes to influence greater consideration of this in future research.

1. Introduction

The COVID-19 pandemic of 2020 caused widespread, and largely unforeseen, disruption to the retail sector across the western world. Particularly in the initial stages of the pandemic, governance priorities focussed on implementing stringency measures, such as restricting the movement of people and closing non-essential retailers and service providers to slow the spread of the virus and keep people safe. This, coupled with concerns regarding personal safety, led to substantial, and sudden, shifts in consumer spending and purchasing behaviours (Baker et al., 2020; Yang et al., 2020). Resultantly, the start of 2020 was a pivotal time for retail performance with widespread uncertainties for consumers passing on to retailers, stakeholders, and policy makers across the retail industry. In the US, the declaration of a national emergency on the 13^th March 2020 sparked the issuance of closure orders for many businesses deemed ‘non-essential’. Alongside this, a number of escalating non-pharmaceutical interventions were implemented by state governments until the 20^th April 2020 when the first easing of these restrictions was seen. As the largest retailer by revenue in the world (McGee, 2020) and with a store within 5 miles of more than half of the population (Joseph & Kuby, 2013), Walmart is a major player in the US retail sector. Furthermore, with ‘essential’ retailer status, Walmart was amongst the US grocery corporations that bucked the trend of declining retail sales during the periods of economic lockdown and was the only retailer to successfully generate net income growth during Q1 of 2020 (Ross, 2020). Much of this success has been widely attributed to growth in their online retailing provision during this time (NRF, 2021).

Impacts of the pandemic have altered and/or accelerated consumption behaviours and previously seen trends, ranging from what people bought, how often they shopped and the location or platform they chose for the purchase. Furthermore, economic fallouts have left the retail sector depleted in resilience and vulnerable to potential recession, the likes of which it has suffered substantially from in the past, with lasting effects on consumer confidence (Buck et al., 2020; Wrigley & Dolega, 2011). An understanding of how these behaviour changes precipitated, the key drivers, and the subsequent impacts on sales outcomes is likely to be crucial for informing the recovery of the sector in the aftermath of the pandemic. For retailers, this may be in the form of adapting business models to suit the changing expectations of consumer convenience, whilst policy makers or industry decision makers can use this knowledge to inform resilience planning and decision making for future periods of upheaval, both in the US and internationally. Although a recent occurrence and, by nature, a relatively new field of research, there is currently a growing body of COVID-related retail studies that are dedicated to building this understanding (e.g., Baker et al., 2020; Chenarides et al., 2020; Enoch et al., 2021). However, naturally, there are still many gaps in our understanding, particularly with regards to how non-pharmaceutical interventions impacted sales outcomes across different purchasing platforms and especially when considering the spatial and temporal variations.

In order to address this, this study aims to investigate the impact of lockdown stringency measures on consumer purchasing behaviours over the onset of the pandemic. This will be achieved using extensive sales data for Walmart in the US and a multi-level modelling approach to provide novel and robust statistical evidence for the trends in both instore and online sales outcomes. Due to the US’ state-based political system, and each state imposing their own stringency measures, the spatial variations in impact across different states are of particular interest. Principally we seek to:

• -Examine the spatial and temporal trends in instore and online sales outcomes, both volume and value, at Walmart.
• -Estimate the impact of increasing stringency measure on Walmart sales outcomes during 2020.
• -Analyse the extent of spatial variations in the impact of stringency measures across different states.

Henceforth this paper provides a comprehensive review of much of the existing literature and relevant debates surrounding the impact of the pandemic on the retail economy and consumption behaviours. The following sections outline the modelling methodologies and the data sources used for this analysis. Next, the results of the analysis pertaining to each objective will be presented and discussed alongside relevant literature. Ultimately, the key findings will be summarised along with a discussion of the limitations and implications of this research.

2. Literature review

2.1. The impact of COVID-19 on the retail economy

Disruption and behavioural shifts in response to the COVID-19 pandemic were widespread across western retail sectors. However, trends of change in consumer behaviours were not linear over time, customer demographic, or across purchasing platforms. Prior to the closure of non-essential retail businesses in the US, total spending across a wide range of categories increased by over 40% in the first half of March 2020, followed by an overall decrease of 25–30% after the announcement of a national emergency on March 13th (Baker et al., 2020). Chetty et al. (2020) find that a large proportion of spending decline came from the highest-income households, with a 49% reduction in spending compared to only 7% for the lowest-income households. Worldwide, the grocery sector notably deviated from overall trends of declining spending (Baker et al., 2020; Khaliq, 2021), whereas apparel and footwear firms were among some of the hardest hit, in some places losing up to 80% of turnover in the early months of the pandemic (Pilawa et al., 2022).

Principally, a large gap in how retailers had to adapt emerged between essential and non-essential retail businesses. Under imposed lockdown restrictions, non-essential retailers were forced to close and faced significant declines of instore sales, often for unknown and varied lengths of time (Berg et al., 2020). Many were compelled to reshape their product offering in order to survive economically, including selling facemasks or hand sanitisers (Roggeveen & Sethuraman, 2020) or adopt new and innovative ways to reach customers (Heinonen & Strandvik, 2020). Primarily, this occurred by diversifying the services on offer (Pilawa et al., 2022) and creating an online platform or hybrid, Click & Collect style, service became crucial for retailers without a strong existing online presence (He et al., 2021). On the other hand, essential retailers, who were allowed to stay open, faced challenges in the form of increasing demand for products as well as delivery or contact-free collection services. Subsequent challenges such as pressure on supply chain and logistics management had to be dealt with all whilst keeping the physical retail environment safe for both consumers and store workers (Roggeveen & Sethuraman, 2020). Measures such as limiting customer numbers in store (Shumsky et al., 2021), implementing queuing or one-way systems (Budd et al., 2021; Shumsky et al., 2021) and restricting the number of products customers could purchase (Showrav et al., 2021) were common for many essential retailers.

The implementation of lockdown or social distancing measures, coupled with changing work-life patterns, exposed the fragility of traditional retail centres (Cilliers et al., 2021) and global supply chains (Sharma et al., 2021). High retail unit vacancy rates and significant reductions in footfall were two of the key issues that many cities around the world were faced with (Lashgari & Shahab, 2022). In an attempt to stabilise the retail economy and keep businesses afloat, governments brought in a number of economic measures. In the UK, this included business rate holidays and grants that were largely aimed at retail businesses (JLL, 2020), and the Furlough scheme which allowed struggling businesses to retain staff and reduce job losses (Pope & Shearer, 2021). Similarly, federal government in the US implemented the Paycheck Protection Program, providing low-interest loans to small businesses to enable them to retain employees (Fairlie & Fossen, 2022) and $300bn worth of one-time stimulus payments in April 2020 (Li et al., 2021). These policies, themselves, had subsequent impacts on consumer spending behaviours, with stimulus payments leading to a sharp, but short-term, increase in spending (Baker et al., 2020; Chetty et al., 2020).

2.2. Changing consumer behaviours

In response to unprecedented circumstances and economic shocks, consumers adopt new short-term behaviours (Goldberg, 2020) and in the context of the pandemic, these behaviours were shaped by two salient factors: preventive measures (i.e. stringency measures) and self-protective mentalities (i.e. personal risk appraisals) (Wang et al., 2022). In many places, consumers altered the location of where they chose to shop, tending to avoid larger retailers and crowded high street centres (Lashgari & Shahab, 2022) in favour of smaller stores (Li et al., 2020), local retail centres (Ballantyne et al., 2022) and out-of-town based shopping alternatives (Sundström et al., 2021). A study using mobile phone records for retail foot traffic in busy commuter zones in the US found a significant fall in consumer traffic even before legal restrictions were enacted (Goolsbee & Syverson, 2020). Similarly, Enoch et al. (2021) report that footfall figures in English town centres fell by 57–75% during the pandemic, with wide spatial variation between centres of different size and offering, likely indicative of a significant drop in the number of transactions occurring and the value of consumption in these centres. Finally, when surveying Italian consumers, Principato, Secondi, Cicatiello, and Mattia (2022) found that food purchases from supermarkets and street markets declined during periods of restrictions in favour of ‘proximity’ shops and online channels.

Perhaps the most significant, and profound, behavioural change in response to the pandemic was the mass shift to e-commerce platforms, particularly during periods of lockdown restrictions, considerably accelerating the pre-existing growth of the e-retail sector (Nanda et al., 2021; OECD, 2020). As a proportion of total retail, the US Census Bureau (2021) reports a jump from 11.8% in Q1 of 2020 to 16.1% in Q2 of 2020 for e-commerce in the US, representing a considerable 44.5% increase from Q2 in 2019. This was a common trend seen internationally and, in the UK, by comparison, online sales rose from 20% of total spending in February 2020 to a peak at over 33% in May 2020 (ONS, 2021). All these factors have contributed to a major upheaval in the retail industry during 2020 and 2021, with almost a decade worth of changes happening within just a few months (Hutton & Rhodes, 2021). These increases were likely driven by a combination of necessity, given that non-essential stores were forced to close, and personal safety concern while the virus was spreading (Das et al., 2021). Grashuis et al. (2020) studied the influence of positive case numbers on grocery shopping preferences and found that when COVID-19 was actively spreading there was a shift to online platforms as customers preferred not to shop inside the store. This shift is further supported by Chenarides et al. (2020) who found significant increases in the popularity of grocery delivery and pick-up services amongst the consumers surveyed at the start of the pandemic. With regards to non-essential consumption, Sayyida et al. (2021) found that there was also a change in how seasoned online consumers used the platforms with less flexibility for ‘showrooming’ or ‘webrooming’ behaviours while physical stores were inaccessible.

In addition to where consumers shopped, there were also notable changes in what they bought and how much they spent during the pandemic. One key behaviour that likely contributed to the increase in spending early on was widespread panic buying and stockpiling which can be common in times of economic uncertainty (Martin-Neuninger & Ruby, 2020). Even in advance of restrictions being imposed, panic buying behaviours led to spending trends that significantly deviated from usual (Pantano et al., 2020), particularly for a whole range of basic products, such as non-perishable food and alcohol, toilet roll, hand sanitiser and soap (Martin-Neuninger & Ruby, 2020; Islam et al., 2021). Wang et al. (2020) show that female, high income and high education level were all characteristics that led to greater scale food reservation behaviours. Finally, studies examining more regular consumption behaviours during the pandemic have noted changes in frequency of trips and a range of basket characteristics, including value, size and product mix. Using sales from a Northern Irish retailer, Boyle et al. (2022) found that declines in total sales early on in 2020 were driven by declining transaction counts and that, once a lockdown was announced, average basket sizes increased notably. Typically, detailed sales data from retailers is difficult to acquire so the majority of research done in this area has used survey results of self-reported behaviours from individual consumers or households (e.g. Chenarides et al., 2020; Dou et al., 2020). Both of these studies, indicate that the majority of consumers, in the US and further afield, reduced the number of trips to the grocery store during the pandemic but tended to buy more than they usually would. Significant changes in the frequency of Spanish food shopping trips after the onset of the pandemic have been found (Laguna et al., 2020), whilst Principato et al. (2022) noted an increase in average weekly expenditure on groceries during restrictions.

2.3. Literature gaps

It is evident that the pandemic has led to major changes in both instore and online sales, influenced strongly by a convolution of policy intervention, consumer behaviour changes and how retail businesses have adapted to these changes. Although, by nature, a relatively new field of study, there is a growing body of research on the impact of many aspects of the pandemic on consumer behaviours with a distinct leaning towards examining consumer spending across the whole sector. Where more specific consumer purchasing behaviours have been examined, much of the existing research centres around grocery retailing and, in the absence of detailed sales data, relies heavily on consumer surveying and self-reported behaviours with inherent limitations. Current understanding of the pandemic's impact on sales outcomes is incomplete, with little research on the specific impact of increasing stringency measures. Spatial and temporal differences in consumer behaviour across a country have also been under-examined.

3. Methods

3.1. Statistical modelling approach

This study employs a multi-level modelling approach to examine the impact of varying lockdown stringency measures on transaction counts (volume) and total spend (value) outcomes for both instore and online sales. This allowed modelling at three different scales: Level 1 representing individual day-county observations; Level 2 representing grouped observations at the state level and Level 3 representing stringency measure level within states. Transaction counts and total spend were our dependent variables. Different models were used for each of the sales outcomes reflecting their distinctive nature. A linear mixed-effect regression is used to model our continuous spend. A negative binomial generalised linear mixed-effect regression is used to model our transaction count data. A negative binomial distribution is used to handle the over-dispersed nature of transaction counts (see Rowe, 2021).

3.2. Model development

Four separate nested random intercept models of three levels were estimated for each individual outcome: (1) Instore transactions, (2) Online transactions, (3) Instore spend and (4) Online spend. Intercepts are allowed to vary across stringency categories within individual states to capture variations in the extent of stringency by state. A stringency index is used to capture variations in the degree of stringency over time. As outlined in Table 1 , stringency index values, ranging from 0 to 100, were re-classified into 5 treatment groups using equal interval cut points: 0 falls into group 0, 1–25 into group 1, 26–50 into group 2, 51–75 into group 3 and 76+ into group 4. Additional explanatory variables in the model consist of a range of temporal, demographic, store characteristic and weather variables, each obtained from the supplementary datasets outlined in section 4. Following Rowe et al. (2022), a natural spline, with 1 knot, was incorporated to account for seasonality and temporal autocorrelation patterns.

Table 1.

Descriptions of the variables used in the study.

Dataset	Variables	Geographical level	Processing description
Oxford COVID-19 Government Response Tracker	Stringency index	State	Index grouped using equal intervals (0; 1–25; 26–50; 51–75; 76–100).
Safegraph Core Places	Location	County	Store densities (per 1000 km2) calculated by aggregated counts of Walmart stores per land area in each county.
American Community Survey	Male population	County	Percentage male population out of the total population of the county.
	Median age		Median age in the county.
	Median income		Median income of the county per $1000.
	Unemployed population		Percentage unemployed population out of the total population of the county.
	Owned tenure population		Percentage of population who own their domestic property out of the total population of the county.
Weather Source Climatology	Temperature	County	5 × 5km grided cells overlaid onto county boundaries. Spatial join algorithm used to calculate average values for each county.
	Precipitation
NOAA Daily Summaries	Temperature	County	Weather station points with 0.4° buffer overlaid onto county boundary. Spatial join algorithm used to calculate average values for each county.
	Precipitation

The formulation for the nested random intercept model is as such:

$$Y_{ijk}=f\left({\beta }_0+{\beta }_1{X}_{1ijk}+\sum _{p=1}^{p+1}{\beta }_{pi}+c_k+s_{jk}+{\epsilon }_{ijk}\right)$$ (1)

$$c_k\sim \mathcal{N}\left(0,{\sigma }_c^2\right),s_{jk}\sim \mathcal{N}\left(0,{\sigma }_s^2\right),{\epsilon }_{ijk}\sim \mathcal{N}\left(0,{\sigma }^2\right)$$ (2)

where Y represents sales outcomes for individual county, $i$, at day, t, within each state, $j\text{,}$ during state-specific stringency level $k$ as nested random effects. ${\beta }_0$ represents the fixed intercept. ${\beta }_1$ represents the slope coefficients capturing the relationship between a business outcome and a vector of attributes $X$ (see Section 3.2). $\sum _{p=1}^{p+1}{\beta }_{pi}$ represents a random natural spline slope at 1 knot point that varies across states and capture systematic temporal patterns in sales outcomes. $c_k$ is the state specific random intercept; $s_{jk}$ is the state-stringency specific random intercept; and ${\epsilon }_{ijk}$ denote the individual error terms. Random terms are assumed to be normally distributed. Random intercepts estimate deviances from fixed intercepts, capturing the existing heterogeneity in the differentiated effects of stringency measures on sales outcomes across states.

4. Data

4.1. Walmart sales data

Transaction and spending data for Walmart purchases across all US stores and Walmart.com was produced by Facteus and made openly available through the SafeGraph COVID-19 Data Consortium, more recently known as the Placekey Community. Facteus' Transactions Data includes the instore and online daily transaction and total spend data at Walmart. This transaction data comes from 3 primary sources: (1) challenger banks such as Simple and N26, (2) payroll cards, and (3) government cards. Challenger banks are typically newer, smaller banks that exist online rather than having a physical ‘brick-and-mortar’ presence. Payroll cards enable employers to automatically load employee wages as an alternative to paper checks or direct debits. These cards are predominantly aimed at part-time or seasonal workers, or those who do not have a personal bank account (Martindale & Tetreault, 2014). Similarly, government cards do not require the holder to have a personal bank account and are used by state governments to distribute social security funds or alimony receipts. As a result, coverage of transactions is predominantly lower-income or young consumers (Yang et al., 2020) so the results of this analysis cannot be assumed to be representative of the entire Walmart customer base. However, this presents an interesting opportunity to study the behaviours of consumers who may have been the hardest hit by the financial implications of the pandemic (Bateman & Ross, 2021) as Kar et al. (2021) find a significant difference in store visits during the pandemic with income level. Both instore and online datasets provide the number of transactions and spending totals, per day, aggregated to the zip code of the bank card holder from 1^st January 2017 to 17^th April 2020. For the purpose of analysis, the unit of spend data is $1000.

4.2. Supplementary data

In addition to the transaction data, supplementary datasets were sourced to serve as explanatory variables for the analysis.

• -A core component of this analysis, the Government stringency index, recording the strictness of lockdown measures, was procured through the Oxford COVID-19 Government Response Tracker (OxCGRT) project. This index ranges from 0 (no stringency measures) to 100 (heaviest stringency measures).
• -Walmart store location data was used to calculate aggregated counts and store densities per land area to serve as a proxy for store accessibility in the absence of store-level information. This was provided by the SafeGraph Core Places dataset which provides coordinate locations and addresses for over 6.3 million points of interest in the US. 4661 Walmart store locations were extracted for this purpose.
• -Demographic and socio-economic variables were procured from the 2019 American Community Survey (ACS) through the ‘tidycensus’ R package.
• -Weather conditions have been frequently found to be influential factors in the decision-making process of consumers (see Rose & Dolega, 2021). As the study area for this investigation spans multiple climate zones, it was not deemed pertinent to directly compare the weather conditions across disparate parts of the US. One of the key reasons for this is that respective populations are likely to be acclimatised to different conditions and therefore what may be considered an extreme in one location may be the norm for another. To control for this, two different weather datasets were used: the daily climate averages (across 2005–2020) and the observed prevailing conditions in 2020. Anomalies were calculated using the difference between average and observed daily conditions.

Daily climate averages were provided by the Weather Source Climatology dataset and procured through CARTO's Spatial Data Catalog and CARTOFrames. In 5 × 5km grided cells covering the 48 contiguous states, daily weather means for air temperature and precipitation were provided for each day of the year based on conditions over a 15-year period, 2005–2020. To prepare this data for the study, these 5 km cells were overlaid onto county boundaries and a spatial join algorithm was used to calculate daily average values for each county.

Observed conditions were sourced through the NOAA online data portal. The Daily Summaries datasets provides air temperature and precipitation values recorded at over 26,000 weather stations across the 48 contiguous states. Each station point location was given a 0.4° buffer (approx. 44 km) and in a similar method to the 5 × 5km grid cells, these buffers were overlaid onto the county boundaries and average values were calculated for each county.

Further details on the variables used, geographical granularity and the processing of these datasets can be found in Table 1.

5. Results

5.1. Examining year-on-year changes

We first analysed sales transactions and total spend data to identify changes in their historical trends during the COVID-19 pandemic in 2020. Using the data aggregated at national level, a natural cubic spline was fitted using 7-day recurring breakpoints. These time series plots for national sales trends are shown in Fig. 1 a–d with the observed data for 2020 shown in red.

Fig. 1.

Walmart sales trends Jan 2017–April 2020 for a) instore transactions, b) instore spend (per $1000), c) online transactions and d) online spend (per $1000). Trend line based on values shown in black (2017–2019) and extrapolated forward for expected 2020 trend. Data in red shows the observed 2020 values.

Examining instore data, a notable decline in observed transactions can be seen from the projected trend. Even in January and February, before much of the pandemic impact and where sales are typically lowest, the distinct year-on-year increase that is evident prior to 2020 cannot be seen. Through March and April there is little increase in average transaction counts but there is a small peak immediately prior to the announcement of restrictions, followed by a relatively strong decline. The same deviation from expected trend is not immediately evident in the instore spend data, however. The observed data follows previous year trends closely in much of 2020 with a similar magnitude yearly peak in mid-March. At the end of the series in mid-April, there is evidence of a peak in spending starting to appear. With regards to the online sales data, transaction counts and total spend show very similar 2020 trends. Both show a distinct peak around mid-March, in keeping with trends from previous years, and then an even larger peak into April that deviates substantially from the expected trend.

At a state level, Fig. 2 a and b allows transaction counts in 2019 and 2020 to be directly compared, using the period January through April for both years. In this instance, a loess smoothing function was applied to the highly variable time series data to highlight key trends. When looking at instore transactions in Fig. 2a, we identified four distinctive patterns. The first involves 2020 trends mirroring those of the previous year, but at a lower magnitude, such is the case in Washington, Oregon and California. The second pattern identified is a similar decline in 2020 transaction counts, however the two trends have very different characteristics in terms of variability, as exemplified by Montana and Wyoming. Thirdly, and much less commonly, some states show an increase in sales in 2020 from the previous year. Ohio and West Virginia are key examples of this, with the former showing a distinct deviation from mid-January onwards. Finally, in some states, there is almost no distinct difference between the two years, particularly from January to March. Examples of this can be seen in Alabama and Georgia.

Fig. 2.

State-level sales trends, Jan–April 2019 vs 2020 for a) instore and b) online transaction counts.

Despite these wide variations, in almost all cases, sales transactions seem to show an upward trend going into April 2020, approximately two weeks after widespread transmission of the virus started and the imposition of lockdown stringency measures. In some states, this means that 2020 counts jump above the previous year's figures (e.g. Wisconsin, Illinois); others bounce back closer to pre-COVID levels after a period of decline (e.g. Washington, Oregon); and some states are much slower to recover (e.g. California, Colorado). Furthermore, the states of Maine and Florida appear to be exceptions to this ubiquitous upward trend as transaction trends continue to decline into April. Although this is only a small snapshot of the pandemic-related trends, this may provide vital indications as to the resilience of various markets across the US.

Contrary to the spatial variation seen for instore transactions, online transaction trends (Fig. 2b) present a near-universal pattern across the country. In most states, there is very little divergence in year-on-year trend until mid-March when there is a distinct take off in online sales transactions, coinciding with the declaration of a national emergency. Despite this, there are a number of states that may be considered early-adopters of pandemic-related online purchasing behaviours. Idaho, West Virginia, Ohio, Vermont and New Hampshire all show distinct increasing deviations in advance of, and perhaps in preparation for, this mid-March spike. Another notable variation across the states is in the trajectory of the increases seen. In some states, particularly in the north eastern states, there is constant exponential increase in online transactions from the initial period of increase whereas in other states there is a slowing of this trend around the start of April before it continues again exponentially.

Fig. 3 allows us to further examine these changes at a higher spatial resolution using the percentage change in transaction counts between January to April 2019 and the same period in 2020. At county level, the overall patterns reflect much of what was seen previously with regards to the general trend of decreasing instore transactions and increasing online transactions. In addition, it highlights some of the complexities that were masked at state and national level, emphasising the value of modelling these changes at the lowest possible spatial scale.

Fig. 3.

Percentage change in a) instore and b) online transaction counts between 2019 and 2020 at US county level.

Clear spatial and temporal changes appear in the sales trends presented here, distinguishing 2020 as an exceptional year when compared to previous observations. While some systematic changes in transactions exist, the extent of changes differ spatially across both state and county. Several key divergence points appear to correspond with notable non-pharmaceutical intervention events related to COVID-19 in 2020. Going forward, this analysis will look specifically at non-pharmaceutical interventions in the form of stringency measures imposed by national and state-level governments to examine the extent to which these influenced the four sales outcomes explore here.

5.2. Association between sales outcomes and stringency measures

To address and explore the relationship between sales outcomes and lockdown stringency, four separate multi-level models were estimated, as described in Section 2.2: one model for each sales outcome, instore and online (Table 2 ). Results capture the relationship between sales outcomes and a range of covariates across the country and state-level variations. Standard error values for each variable coefficient are shown in brackets below. We first analyse the results for transaction counts.

Table 2.

Transaction and spend model fixed effects.

	Instore transactions	Online transactions	Instore spend	Online spend
(Intercept)	14.8607 ***	11.0590 ***	71.2836 ***	2.6905 ***
	(0.0819)	(0.1415)	(1.7063)	(0.00925)
Day of year (ns)	0.4767 ***	2.5126 ***	12.3825 ***	1.2156 ***
	(0.0253)	(0.0348)	(0.5897)	(0.0265)
Weekend	−0.0967 ***	−0.4413 ***	−1.9950 ***	−0.1237 ***
	(0.0041)	(0.0058)	(0.0961)	(0.0043)
Public Holiday	−0.2178 ***	−0.2336 ***	−2.6045 ***	−0.0600 ***
	(0.0106)	(0.0151)	(0.2464)	(0.0110)
% Male population	−0.1214 ***	−0.1432 ***	−0.7136 ***	−0.0281 ***
	(0.0009)	(0.0015)	(0.0201)	(0.0009)
Median age	−0.0716 ***	−0.0456 ***	0.0617 ***	0.0110 ***
	(0.0006)	(0.0008)	(0.0119)	(0.0005)
Median income	0.0247 ***	0.0278 ***	0.2581 ***	0.0098 ***
	(0.0002)	(0.0003)	(0.0043)	(0.0002)
% Unemployed population	0.0273 ***	0.0465 ***	0.0398	0.0123 ***
	(0.0012)	(0.0015)	(0.0220)	(0.0010)
% Owned tenure	−0.0563 ***	−0.0654 ***	−0.7402 ***	−0.0366 ***
	(0.0004)	(0.0005)	(0.0079)	(0.0004)
Store density	0.1895 ***	0.1265 ***	1.3078 ***	0.0374 ***
	(0.0014)	(0.0013)	(0.0131)	(0.0006)
Temperature anomaly	−0.0036 ***	−0.0023 ***	−0.0402 ***	−0.0019 ***
	(0.0002)	(0.0003)	(0.0048)	(0.0002)
Precipitation anomaly	0.0481 ***	−0.0393 ***	0.3842 **	−0.0119
	(0.0061)	(0.0084)	(0.1464)	(0.0066)
Observations	319,788	319,788	319,788	319,788
No. groups: stringency level	5	5	5	5
No. groups: state:stringency level	230	230	230	230
Variance: stringency level	0.0000	0.0546	6.0507	0.0232
Variance: state:stringency level	0.9265	0.6779	48.0154	0.2050
R2m	0.3938	0.3397	0.1442	0.1413
R2c	0.6062	0.4868	0.2172	0.2820

***p < 0.001; **p < 0.01; *p < 0.05 (Standard error values for each variable coefficient shown in brackets below).

5.2.1. Sales transactions

Table 2 reports variance estimates. These estimates indicate the extent of variability that can be explained at each level of our model. When accounting for variations in stringency level at national scale, variance estimates of virtually zero and 0.06 for instore and online transactions, respectively, indicate that practically no variability is explained in the models. However, these figures rise to 0.94 and 0.68 when stringency measures vary by state, pointing to the importance of accounting for state-specific variations in the impact of COVID interventions on sales transactions. Furthermore, the pseudo r² values, r² _m and r² _c, show a significant increase in variance explained in the outcome between a fixed effects model and our full mixed-effects model for both instore and online transactions.

Regarding the covariates, our transaction model estimates in Table 2 indicate that the strongest associations occur between sales transactions and temporal variations in transitions relating to Day of year, Weekend and Public Holiday. The influence of the weekend is much greater for online transactions than instore. Other notable results include relatively strong associations with store density and the percentage of males in the area, each positive and negative respectively. Temperature and precipitation anomalies show a significant association with sales, but these are relatively weak overall. Warmer temperatures appear to negatively affect transaction counts, but increased precipitation has differing effects for both instore and online. Overall, standard error rates stay relatively low throughout.

Fig. 4, Fig. 5 display the random intercept effects results from our transaction models capturing the average national effect of stringency measures on transactions and the geographic variability of this effect across individual states. Random intercept effects are deviations from the national average i.e. the regression intercept. Zero on the y axis of these figures represents the model intercept estimate reported in Table 2. The particularly small values of change in Fig. 3a suggest that there is virtually no statistically significant change in national average instore transaction counts moving through the different stringency levels. Conversely, Fig. 3b reveals a steep trend of declining online transactions, particularly moving from stringency levels 0 to 2 and through levels 3 and 4. Large confidence intervals for stringency level 4 reflects the high extent of variability across states over time.

Fig. 4.

Random intercept estimates for a) instore transactions and b) online transactions across increasing levels of lockdown stringency with 95% confidence intervals.

Fig. 5.

State-level random intercept estimates for instore and online transactions across increasing levels of lockdown stringency with 95% confidence intervals.

Fig. 5 reports state-specific random estimates capturing the effect of stringency measures on transactions in individual states. These estimates are reported as deviations from the regression intercept. Compared to national level trends, state-specific random estimates for instore transactions show large variability in transactions across stringency levels, pointing to large variations in the impact of stringency measures across states. In many states we see sustained declines at every level of increasing stringency (e.g. Florida, Michigan, Montana) however there is very little evidence of sustained increases in transaction count intercepts. Ohio stands out as the only state where the intercept for the highest stringency level is greater than at level 0, before any stringency interventions were made.

Across the board, there is high variability across the intercept values for online transactions with most states showing fluctuations between increasing and decreasing intercept values as stringency increases. There are only a few exceptions where a consistent trend is identified across the stringency levels, Ohio with a sustained increase and Arkansas and New Mexico with sustained decreases.

When comparing instore and online transaction trends, a set of distinctive patterns emerge. In some states, the trends appear to mirror each other indicating that both sets of transactions have been impacted in a similar manner by increasing stringencies. The clearest examples of this can be seen for North and South Dakota, both showing similar deviations from the national average intercept value for instore and online transactions. Other states show divergent or convergent trends across the two types of transactions, in each case showing increasing online transactions and decreasing instore transactions. Pennsylvania, New Jersey, and Connecticut are three such examples of diverging trends where online transactions are seen to increase at a slower rate than the instore transaction decline. This may indicate a shift from one platform to another but that online transactions are not fully replacing the loss in instore transactions. Virginia is an example of converging trends but in this instance, online transactions seem to increase at a similar rate to the decrease in instore transactions.

Finally, we can identify where the greatest shifts in the relationship between stringency level and transaction counts occurs. With respect to instore transactions, the greatest increases or decreases in intercept value are commonly seen at low levels of stringency (between levels 0 and 1), perhaps most evident for Washington or South Dakota. Conversely, the biggest changes for online transactions are seen at higher levels of stringency, either in the jump between levels 2 and 3 or 3 and 4, as exemplified by Tennessee or Louisiana.

5.2.2. Total spend

We now turn our attention to the results from our total spend models (Table 2). Larger variance estimates at state level than at national level, 48.02 vs 6.05 for instore and 0.21 vs 0.02 for online, indicate much greater variability explained when accounting for the impact of stringency measures across individual states. Additionally, the pseudo r² values indicate that including the random effects improves the level of variance explained and the fit of the model.

When examining the model covariates, similarly to transactions, the strongest associations with both instore and online spend are seen with the temporal variables of Day of year, Weekend and Public holiday. However, weaker relative coefficients for public holidays with online spend indicates that these have less of a negative influence on how much people spend online than weekends whereas for instore spend the opposite is true. Store density has one of the strongest associations with instore spend but, although statistically significant, store density does not have as strong a relationship with online spend. Furthermore, no significant relationship was found between instore spend and unemployment rate. Temperature shows significant negative but relatively weak associations for both instore and online spend. For precipitation, however, there is a relatively strong and significant influence on instore spend but the relationship was not significant with regards to online spend. Instore spend shows the strongest relationship with both temperature and precipitation across all the models.

Unlike for transactions, both instore and online spend random effects in Fig. 6 a and b show significant variations across stringency levels. Instore spend shows a continuing declining trend in average spend as stringency become stricter. This trend of decline is mirrored for the online spend estimates with a notable difference: a random estimate around zero for level 4 stringency measures indicating that average online spend was similar to the national average online spend without stringency measures in place. Though, wider confidence intervals indicate high levels of uncertainty for both estimates at stringency level 4.

Fig. 6.

Random intercept estimates for a) instore spend and b) online spend across increasing levels of lockdown stringency with 95% confidence intervals.

We examine the impact of stringency measures on retail spend across states by analysing our state-specific random effects estimates. The scale of the change in the online intercepts is much smaller than that of the instore intercepts and this is true of almost every state. For comparison sake, the intercept values and corresponding confidence intervals for online spend have been multiplied by 10 in Fig. 7 .

Fig. 7.

State-level random intercept estimates for instore and online spend across increasing levels of lockdown stringency with 95% confidence intervals. Online values multiplied by 10 to enable comparison of trends.

Conversely to what was seen for transactions where there was high variability, there is much more consistency in the state-level trends in total spend. Again, we can identify some distinctive patterns. The first is that there is much more mirroring of trends across the two purchasing platforms with the clearest examples of this being North Dakota, South Dakota and Iowa. In all three examples, both instore and online intercepts show consistent rates of decline as increased stringency measures are introduced. Perhaps the most notable deviation from this trend is California where we see a general trend of decreasing instore spend and increasing online spend. The second pattern identified is the distinct change in trend around the second level of stringency, specifically. Wisconsin, Illinois and Rhode Island are amongst the clearest examples of this trend occurring, each with a distinct minimum or maximum point at stringency level 2. California shows a similar spike but at level 3.

When looking at where the greatest shifts in the relationship between stringency and spend occur, the largest gradients in instore spend are seen most commonly between levels 0 and 1 (e.g. Ohio, Georgia) or 2 and 3 (e.g. Michigan, Nevada). The steepest changes for online spend are typically seen at the highest level implemented (e.g. Pennsylvania, Maryland).

6. Discussion

The results presented in this study have provided empirical evidence regarding the magnitude and the nature of the economic impact that the shock of a major pandemic and nonpharmaceutical interventions, in particular closures of non-essential retail stores, had on consumer behaviours, and subsequently on retail sales outcomes in the US over the onset of the COVID-19 pandemic.

6.1. Year-on-year changes

This study presents evidence that in-store transactions declined during the onset to the pandemic while spending remained steady, indicating that Walmart customers likely made fewer but larger physical shopping trips, reducing the number of transactions but maintaining overall spending. There are several potential reasons for this type of behaviour change under the types of restrictions that were seen. For example, significant declines in everyday mobility rates (Cronin & Evans, 2020) and increased proportions of the population working from home (Bick et al., 2021) are likely to have lessened the occurrence of smaller, top-up shops which may occur on route to work or during the working day. Instead, these purchases may be incorporated into a larger weekly shop or avoided altogether. Alternatively, these effects may be indicative of stockpiling behaviours that were seen worldwide (Ahmadi et al., 2022) amidst the panic of unknown impacts of the virus on factors such as global trade and supply chains (Hobbs, 2020) and a decline in consumer confidence across the Western world (Teresiene et al., 2021). These behaviours led to the mass emptying of supermarket shelves and imposed purchasing limits for several household items (Hobbs, 2020). In the US, mass panic buying behaviours were observed as early as the end of February (Yang et al., 2020), even before rapid community transmission of the virus was identified. Finally, the peak in overall spend that appears towards the end of the time series in Fig. 4b corresponds to the date that the first round of stimulus checks was issued on April 13th. Baker et al. (2020) found that this issuance caused a significant increase in consumer spending across the board, but that this impact was temporary and not sustained long-term.

Importantly, online sales outcomes experienced a significant increase which coincided with the declaration of a national emergency in the United States, indicating that it was likely the imposition of stringency measures, or even the imminent probability of stringency measures (Cronin & Evans, 2020), that sparked the increase in online purchasing. Differing from what was seen in instore sales, when the online sales outcomes did deviate from the expected trend, both transactions and spend showed very similar timeframes and magnitudes of increase. This is likely to indicate that it was the increased volume of transactions led to the increase in total spend as new consumers embraced online sales platforms at a level not previously seen (Richards & Rickard, 2020). One potential source of these new online consumers is those who are generally less digitally-savvy who were previously reticent to adopt online purchasing habits (Liu et al., 2020) or vulnerable populations who have turned to online consumerism due to the level of safety it affords (Pantano et al., 2020). Our results also support the trends of increased demand for online grocery delivery at the start of the pandemic that were widely reported worldwide (Eriksson & Stenius, 2020) leading to groceries being the retail sector with the largest online increase in demand (Li et al., 2020).

Furthermore, we revealed wide variation in instore trends across states, which may reflect differing rates of cases, deaths and overall perceived threat of the virus in those regions. These factors correlate strongly with levels of stringency early on in the pandemic so were not included in the model to avoid issues with multicollinearity. Personal health concerns play a significant role in the reduction of movement and social interaction even before official restrictions were implemented (Cronin & Evans, 2020). There is also evidence of political and partisan influence in behaviours during lockdown and general concern towards the virus (Baker et al., 2020) which may well play a role in the variations seen here. The near-ubiquitous feature of an upward trend in transactions at the start of April corresponds almost exactly to the signing of the CARES (Coronavirus Aid, Relief and Economic Security) Act, which included, amongst many other relief packages, the promise of up to $1200 direct payments per adult for most American households (Carroll et al., 2020). The upturn in sales transactions seen here may be a response to a short-term relief of economic pressures when these policies were signed into law. Conversely, with regards to online sales, there is very little spatial variation between states. The consistent and significant increase from the 2019 transaction counts occurring simultaneously across states supports the finding that, unlike instore behaviours, this was a direct impact of national-level rather than state-specific policy.

6.2. Association between sales and increasing stringency levels

Two fundamental observations were made from our random effects estimates. Firstly, instore transactions effectively stayed constant across all stringency levels whilst spend showed a consistent trend of decline as the stringency level increased. Previously this study has shown lower transaction counts in 2020 than previous years but the stability of these lowered transaction counts throughout 2020, regardless of stringency level, may be resultant from the nature of the goods sold by Walmart. As an essential retailer, many consumers will have relied on regular shops for household groceries and other necessities even if this was at a decreased frequency from previous years. Conversely, the decrease in total spend found here supports the widespread evidence that suggests US consumers have been saving at record levels during the pandemic (Bauer et al., 2020). The high levels of economic uncertainty afforded to the public by the pandemic and onset of the resulting lockdowns prompted an increase in personal saving rates from 7% to 33% between January and April 2020 (Pisani-Ferry, 2020), amounting in the largest decline in consumer spending for over 3 decades (Corcoran & Waddell, 2020). Secondly, online transactions and spend both show declining transaction and spend figures during periods of moderate stringency followed by an uptick at higher levels. Our results suggest that an increase in online sales was sparked by the imposition of stringency measures, particularly during the highest stringency levels. Contrary to findings of consistent increases in online purchasing throughout Q1 and Q2 of 2020 (US Census Bureau, 2021), this investigation found that both online transactions and spend at Walmart were in decline during the lowest levels of stringency.

The dominant takeaway from the state level analysis is the diversity of trends in sales outcomes across stringency levels. This is likely to reflect different convoluting factors, some of which were accounted for in the modelling process of this investigation, however there is still some variability that persists after all the factors considered here. In addition to variations in COVID-19 cases and the perceived threat in different areas, each state imposed different levels of restrictions under their own time scales and in line with the political sway of the area. Political leaning and partisanship seem to have a significant impact on shaping the perception of risk associated with COVID-19 (Barrios & Hochberg, 2020). Republican-led counties or states were overall slower to implement stay-at-home policies (Brodeur et al., 2021) and, of the seven states that never issued orders of this nature, all voted Republican in the 2020 US election. This top-down polarisation and general resistance to act is likely to filter through to influence public compliance to the measures and impact consumer habits diversely in different areas. Furthermore, Democrats were found to have greater concern over money worries because of the pandemic (Bruine de Bruin et al., 2020) however, Baker et al. (2020) find little-to-no evidence of this partisanship split impacting spending levels in their study on consumption. Although not directly within the scope of this investigation, given the high levels of variability across all states, there is no immediate evidence presented here to contradict the findings of Baker et al. (2020).

Finally, of key importance is the stage at which the biggest changes in each sales outcome were recorded. With regards to instore sales, although still very early in the overall scale of the pandemic and lockdown disruptions, it is expected that the onset of lockdown fatigue plays a significant role here. Lockdown fatigue has been found as a widespread issue across the US and has almost certainly led to reductions in the efficacy of lockdown measures (Goldstein et al., 2021). In addition to the natural fatigue of staying at home for long periods of time (Amanzio et al., 2021), uncertainty over the longevity of these non-pharmaceutical interventions are also likely to have played a part in consumer decision making. If consumers initially thought that restrictions would only last a matter of weeks, they may have held off from shopping, or made short-term habitual changes, as long as they were able to. These changes may have proved to be unsustainable as it became evident that stringency measures would not be lifted as soon as hoped. The uptake in online sales and the timings of this have already been discussed, but similarly to instore behaviours, it is possible that online purchasing of non-essential items rose as consumers realised that non-essential stores would not be re-opening as quickly as they had expected or hoped. Unfortunately, it has not been possible to examine product breakdowns within these trends but it is expected that this would enable much deeper insight into why the results found here may have occurred.

7. Conclusions

There is little doubt that the COVID-19 pandemic has had considerable and potentially permanent impacts on the retail industry at every level. Understanding the nature and magnitude of these impacts at different stages and spatial levels throughout the pandemic is essential for stakeholders and policy makers to negotiate and manage these changes to the sector effectively. This paper has employed a multi-level modelling approach to explore the impact of the US lockdown on retail outcomes for Walmart over the onset of the pandemic, considering national, state and county level trends. Fundamentally, we have shown that the imposition of lockdown stringencies had measurable impacts on both transaction and spend outcomes that differed distinctly across purchasing platforms, as well as spatially and temporally. There is a clear indication that nationally consumers were making fewer, larger physical shopping trips with high levels of variations in these trends at state level. With regards to online sales, a huge increase in transactions and spend was seen immediately following the announcement of a national emergency in the US. Unlike instore trends, this was seen almost ubiquitously across the country. Nationally there was no effect on instore transactions across stringency levels but when stratified by state, again high levels of variation were recorded in both instore and online transactions. In certain states there was evidence of a direct switch in purchasing platform but the increase of online transactions didn't fully replace the loss of instore transactions. The random effects for total spend also showed high levels of spatial variation but the dominant state-level trend was that of mirroring effects in both instore and online spend.

By adding to a growing and vital body of literature, the implications of these findings can help understand how different retail outcomes were impacted by unprecedented stringency measures across the US and at different spatial scales. With access to data from a retailer as wide-spread and dominant as Walmart in the US retail landscape, this research gives important and novel insights into how consumer spending behaviours altered during various stages of lockdown stringencies at the onset of the pandemic. As arguably, the findings from this research pertain to the most uncertain phase of the pandemic, they could help retailers become more resilient and policy makers respond more effectively to future pandemics and economic shocks and adapt to new trends in consumption behaviour. Furthermore, the diverse range of products offered by Walmart, including groceries, implies that these findings could be applicable to retailers operating in various sectors.

Although designed to be comprehensive in approach, there are inherent limitations to such a study as this. Firstly, as previously recognised, the data used here is sourced from predominantly ‘challenger banks’ alongside payroll and government cards, meaning that consumer representation is skewed towards younger and lower-income demographics (Martindale & Tetreault, 2014; Yang et al., 2020). Subsequently, the findings of this investigation cannot be deemed to be wholly representative of the Walmart consumer base, but we provide unique insights into the grocery retail behaviours of consumers that may have been the hardest hit financially during this period. Furthermore, this study focuses on one case study retailer, Walmart, although this limitation is somewhat mitigated by their widespread presence across the US, they are the largest physical retailer in the world, and they specialise in a wide range of general merchandise goods in addition to groceries. As mentioned in the paper, additional information could be found by examining at the impact of factors such as the public perception of risk posed to them by COVID-19 and political partisanship in different areas, however this information was either not available or beyond the scope of investigation. Furthermore, there is undoubtedly huge amounts of information to be gained from studying the sales outcomes of specific product types, however this was not possible given the data available.

As such, there is huge scope for further research in this area. Although Walmart has a giant presence in the US, both physically and online, this study covers just one case study retailer over the very onset of the pandemic before stringencies started to be loosened in the US. The subsequent stages of the pandemic are likely to have very different impacts on retail outcomes which will be of great interest as we try to understand the full impact of the pandemic on the retail industry. Additionally, this research can be developed further by examining different product types, considering consumer demographics, and studying trends at lower spatial scales.

CRediT author statement

Natalie Rose: Conceptualisation, Methodology, Software, Formal analysis, Writing – Original draft, Visualisation.

Francisco Rowe: Conceptualisation, Methodology, Writing – reviewing and editing, Supervision.

Les Dolega: Conceptualisation, Methodology, Writing – reviewing and editing, Supervision.

Funding

CDRC (ES/L011840/1), CDT (ES/P000401/1).

Competing interests

None.

Handling Editor: Dr. Y.D. Wei