CPG consumption in times of recession: novel evidence from matched administrative data

Brancatelli, Calogero; Inderst, Roman Article — Published Version CPG consumption in times of recession: novel evidence from matched administrative data Quantitative Marketing and Economics Provided in Cooperation with: Springer Nature Suggested Citation: Brancatelli, Calogero; Inderst, Roman (2025) : CPG consumption in times of recession: novel evidence from matched administrative data, Quantitative Marketing and Economics, ISSN 1573-711X, Springer US, New York, NY, Vol. 23, Iss. 2, pp. 265-289, https://doi.org/10.1007/s11129-025-09296-5 This Version is available at: https://hdl.handle.net/10419/323527 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by/4.0/ Vol.:(0123456789) Quantitative Marketing and Economics (2025) 23:265–289 https://doi.org/10.1007/s11129-025-09296-5 CPG consumption intimesofrecession: novel evidence frommatched administrative data CalogeroBrancatelli1· RomanInderst2 Received: 15 May 2023 / Accepted: 2 February 2025 / Published online: 10 March 2025 © The Author(s) 2025 Abstract We utilize a novel dataset that merges household administrative income and sociodemographic information from tax records with scanner data on CPG consumption. Our analysis reveals significant variation in household per-capita expenditures. However, we find only a modest economic relationship between CPG spending and income, even amidst substantial within-household income fluctuations during the Dutch "double-dip recession" from 2011 to 2018. This relationship remains small for households with low income and low liquidity and holds across both food and non-food expenditures. Keywords Income effects· Consumer-packaged goods· Administrative data JEL Classification D12· E21· E32· M30 This paper is the result of a collaboration with the AiMark foundation and Centraal Bureau voor de Statistiek (CBS) Netherlands. We thank both for their continuous support. We are also grateful for comments received from Thomas Otter and Michalis Haliassos and participants of the 2023 Bundesbank International Household Finance Conference. The authors benefitted from financial support from a long-term research grant by the Think Forward Initiative (TFI). During the study implementation, Brancatelli worked for Goethe University Frankfurt and is now working for the European Central Bank (ECB). The views expressed hereare our own and do not necessarily reflect those of the ECB or the Eurosystem. * Roman Inderst [email protected] Calogero Brancatelli Caloger[email protected] 1 European Central Bank, FrankfurtAmMain, Germany 2 Goethe University Frankfurt, FrankfurtAmMain, Germany 266 C.Brancatelli, R.Inderst 1 Introduction Consumption of consumer packaged goods (CPGs) constitutes a sizable fraction of all except the richest households’ consumption and disposable income (on average 9.95% in our sample). Analyzing how changes in income affect household CPG expenditure, for example over the business cycle, is thus of interest both to marketing scholars and practitioners as well as to macroeconomists and policymakers. Yet, quantifying this relationship with survey data proves challenging due to measurement errors in household-specific income and wealth components and a lack of granularity in expenditures. In this paper, we address those challenges by drawing on a novel data source that links individual households’ CPG expenditures from GfK’s scanner panel1 with total disposable income from administrative data, as collected by the Dutch statistical office (Centraal Bureau voor de Statistiek, CBS).As a main result, we find only an economically small relationship between CPG spending and income as a 10% change in disposable household income is related to only a 0.55% change in CPG consumption.2 Our empirical strategy exploits the granular panel structure of our dataset and draws on institutional specifics of grocery retailing in the Netherlands and pronounced household-specific income variations. Notably, the time frame of our analysis covers the period between 2011 and 2018 and thus the “second dip” (2011 to 2013) of the Netherlands’ double-dip recession following the financial crisis. Over this period, GDP growth dropped by 3.6 percentage points from the fourth quarter in 2011 to the third quarter in 2013 and unemployment rose by 3.3 percentage points. One-third of income changes in our data are negative. We also document considerable remaining heterogeneity between households even after stripping out (year-byyear) aggregate fluctuations. Thanks to this institutional setup, household-specific income variation can be considered as good as randomly assigned. We also note that our measure of income, namely household disposable income, is more comprehensive than main earner income alone, which is commonly used in the literature and typically (self-)reported in household scanner data. Moreover, relying on administrative data reduces the risk that results are downwards biased by measurement errors. Still, to test this more formally, we corroborate our finding of only an economically small relationship between household income and CPG expenditure using GfK self-reported figures. Our findings are robust to a series of different alternative specifications. First, our key finding of only an economically small relationship between household income and CPG expenditure holds for both food and non-food 1 Researchers from the U.S. may be more familiar with Nielsen’s household (or homescan) panel data. In many other countries, the panels provided by GfK or the sister organization Kantar furnish the same service. In the Netherlands, Nielsen does not have a household panel, making GfK’s panel the industry standard. 2 Though in our main regression, the income coefficient is statistically significant at the 1%-level, given its small size confidence intervals are relatively large. Still, in the reported case the upper boundary of the 95%-confidence interval corresponds to only a 0.74% increase in consumption following a 10% increase in income. 267 CPG consumption intimesofrecession: novel evidence from… categories. Second, it also holds for households with low income or low savings. Third, our findings still apply when considering only negative income changes, which account for around one-third of observations in our sample. Despite the robustness of our results, our analysis reveals considerable differences in cross-sectional per-capita CPG spending per household, even though in the Netherlands, the retailing landscape is particularly homogeneous. As we document national pricing and assortment strategies of supermarket and discounter chains, we can rule out such supply side explanations for this variation. Our findings thus suggest that differences in household (per-capita) CPG expenditures may be largely a result of differences in taste and, consequently, households do not show strong adjustments in expenditures when facing changes in income. Our findings are relevant for marketing professionals in the CPG and retailing industry. In addition, they are of relevance for policy-oriented (macroeconomic) work. It suggests that in countries with a similar institutional and socioeconomic background as the Netherlands, CPG consumption is particularly stable across the business cycle, despite pronounced differences in per-capita household expenditures and often large price differences between products in the same category, such as between store brands and national brands. Our disaggregate analysis contributes to a large strand of literature on business cycle research in marketing (Dekimpe & Deleersnyder, 2018). While this literature has identified large effects for some categories such as consumer durables (e.g., Deelersnyder etal., 2004), generally effects are much less pronounced for necessary consumer goods (cf. Dutt & Padmanabhan, 2011, which uses aggregate data in a cross-country analysis of currency crises). Relying on within-household variations, we find only an economically small relationship. Dubé etal. (2018), whose approach we follow, find a small relationship between reported income and households’ private-label share in the US. Biswas etal. (2021) extend this to a joint analysis of how households vary their budget allocations across product categories as well as store formats and brand types, which allows them to derive more nuanced business implications, e.g., regarding the effects of a recession on non-food vs. food items sold in different store formats. Brancatelli etal. (2022) apply this methodology to the choice of store brands in the Netherlands, confirming an economically small relationship with income changes. The economically small relationship between store brand consumption and income change over the business cycle contrasts strongly with earlier findings using only aggregate data (e.g., Lamey etal., 2007). There is also sizable literature in macroeconomics, especially household finance, that seeks to measure the response of (households’) aggregate consumption to changes in income.3 There, changes in consumption are often measured 3 Still, a strand of the large literature studies the shape of consumer demand responses to income changes also for different consumption categories. Typically, food, which makes up the bulk of CPG expenditure, is found to be a normal good, so that spending increases with income, albeit its share in overall expenditure decreases. While we stipulate a linear relationship between log income and log expenditure, the literature has also analyzed more flexible functional forms (e.g., Banks etal., 1997). 268 C.Brancatelli, R.Inderst by surveys, based on diaries or interviews, rather than by recording actual expenditures.4 In such surveys also hypothetical questions can be asked, which allows for disentangling responses to anticipated vs. non-anticipated as well as temporary vs. permanent changes in income (see, for instance, Jappelli & Pistaferri, 2014 or Christelis etal., 2019). A key finding of this literature is that households would consume a surprisingly large fraction of only transitory income shocks, contradicting the “permanent income hypothesis”. These studies often stress the importance of household liquidity as a mediating factor. While research has also documented such excess sensitivity of consumption to real transitory income shocks (such as from the federal government shutdown in 2013, Baker & Yannelis, 2017, or from tax rebates in 2018, Misra & Surico, 2014), interestingly a study by Parker etal. (2013) documented only small effects for expenditures on food and non-durable items. Notably, our data does not allow us to distinguish between anticipated and non-anticipated income changes. Still, we corroborate our finding of only a small relationship with actual CPG spending by including lags. In contrast to this macroeconomic literature, we do not find that households with low income or savings react significantly more strongly. The remainder of this paper is organized as follows: In section “Data”, we introduce our dataset. In section “Stylized facts”, we present stylized facts about CPG consumption and retailing. Section “Estimation results” represents the results from our empirical analysis. We conclude in section “Conclusion”. Additional results are provided in an appendix. 2 Data We use and enrich a novel dataset introduced by Brancatelli et al. (2022) matching household scanner data from the Dutch GfK consumer panel with administrative data on wealth, income, and sociodemographic characteristics, as provided by CBS.5 Matching the datasets was possible as more than 5,000 households individually consented.6 Between 2011 and 2018 we are thus able to observe on the level of individual households both their purchases and actual, rather than only reported, information regarding income and wealth. Our level of aggregation is the household, rather than individuals both for the considered expenditures as well as for income. When using household scanner data alone, typically reported income instead only relates to the main earner. In the following subsections, we describe each of our two main datasets in detail. 4 Macroeconomic research has only recently turned towards the use of household scanner data, see for instance Kaplan and Schulhofer-Wohl (2017) for the measurement of inflation. A comprehensive review of the use of household scanner data in various strands of economic literature is provided by Dubois etal. (2022). 5 Household scanner data from the GfK consumer panel are provided by the AiMark foundation. 6 Matching, based on household characteristics provided directly by GfK, took place in the secure environment of CBS and was undertaken by its statistical staff. All analyses with matched data took place in the secure environment of CBS. 269 CPG consumption intimesofrecession: novel evidence from… 2.1 Household scanner data We use household scanner data from the GfK consumer panel, which is the main provider of such information in the Netherlands. It covers household purchases for a wide range of CPGs, both food and non-food.7 To register purchases, households selected for the GfK consumer panel are equipped with an electronic home scanning device. Our unit of observation is thus a single product on the barcode level bought by an individual household on a certain day at a specific retail chain. For our analysis, we aggregate the data on a householdyear level. Matching with CBS data is only possible for those households that explicitly consented to the matching of their data.8 This restricts our unbalanced panel from originally 11,041 to 6,151 distinct households in 2018. Of these, we observe 89% over the entire period. Also, note that we are still able to draw on the full Dutch population from CBS and we use this information both to analyze the representativeness of our sample and to construct regression weights. We obtain from GfK also households’ self-reported (mean earner) income data, which is collected annually by GfK. We use this to corroborate our main findings based on CBS administrative income data. Panelists are asked to report average monthly income net non-financial income (including salary, income from selfemployment, pensions, and social benefits). Responses are provided at the end of a given year and GfK refreshes panel information in the first quarter of the subsequent year. We take this lag into account. We conduct the following data selection. We neglect households for which we observe less than six months of purchase data or less than three shopping trips on average per month. Next, we exclude transactions with unreasonably high or low prices by removing transactions with prices that are four times higher than the median or less than a quarter of the median of the same product. We identify the same product via barcodes and calculate median prices across households and months for each year. Overall, we lose less than 1% of total expenditures. To capture full years of transactions for each household, thereby avoiding seasonal biases as our level of observation consists of household-years, we exclude panel members that entered the panel after January of a particular year. We thereby lose 6.42% of households. 7 This includes fresh, frozen, refrigerated food, alcoholic and non-alcoholic beverages, health and beauty products as well as pet needs, cleaning, detergent products, and general merchandise at a wide variety of retail outlets. 8 In the European Union in May 2018 the General Data Protection Regulation (GDPR) came into force, which severely limits the creation, storage, and use of personal data. Due to the GDPR, GfK is obliged to ask for the consent of each household. In 2018 all households, which were in the panel at the time, had the opportunity to voluntarily agree to their household scanner data being linked to CBS data. 270 C.Brancatelli, R.Inderst 2.2 CBS data The Dutch Centraal Bureau voor de Statistiek (CBS), also referred to as Statistics Netherlands, is a governmental institution that collects statistical information in the Netherlands. It provides microdata to authorized institutions under strict conditions of confidentiality. CBS microdata data have been used previously in the literature, for instance in De Meijer etal. (2013), Lammers etal. (2013), and Raymond etal. (2010). CBS data include, in our case, households’ sociodemographic information and information on income and wealth. We use CBS information on the number of household members (at the same address), the number of children, the number of male or female persons, and the number of household members receiving an income. Changes in these variables may affect the required consumption basket and thereby expenditures. We thus account for the composition of a household through the inclusion of these variables in our regression analysis below. In the descriptive parts of this paper, we adjust income, wealth, and consumption figures for household size. We follow Attanasio and Pistaferri (2016) and define household size as 1+0.7(n−1)+0.3k , where n is the number of adults and k the number of children in the household.9 For instance, if a single-person household grows with the addition of an adult, the headcount increases from 1 to 1.7. Our key explanatory variable refers to disposable household income. CBS provides researchers with highly accurate income data from Dutch tax authorities. As some components of income, such as that from self-employment, are only available yearly and as this generally applies to net income after taxes, which is correctly calculated only at the end of the year, we use yearly income for our analysis. Disposable (household) income is defined as gross household income minus deductions, consisting of transfers and tax payments of all household members.10 CBS also provides detailed information about households’ financial wealth. Financial assets consist of deposit and current account balances and the market value of stocks and bonds as of 1st January of the following year. Dutch households need to declare this information in their annual income tax declaration to derive the base for the income tax on return on assets.11 As we explain below, the use of 9 We note that this does not fully correspond to scales referred to as the OECD (modified or equivalence) scales (see https:// www. oecd. org/ els/ soc/ OECDNoteEquiv alenc eScal es. pdf). We thank a referee for pointing this out. None of the insights from our descriptives would, however, be affected by changing the scale. 10 Gross household income comprises labor income, business income, capital income, social transfers, insurance benefits, allowances, and alimony from ex-spouses of all household members. Deducted transfer payments comprise income transfers, such as alimony to ex-spouses, as well as mandatory social transfers for social assistance and national insurance benefits. Taxes deducted from gross household income consist of all income tax and a wealth tax paid within a household. 11 Whenever no such information is provided in the income tax declaration, CBS uses reported data from the respective financial institutions to impute these values individually. As of 2016, this applies also to foreign financial assets that are in countries that are part of the OECD Common Reporting Standard adopted by the European Commission. 271 CPG consumption intimesofrecession: novel evidence from… financial wealth data allows us to analyze whether households with smaller savings react more strongly to income changes as often suggested by the literature. Through CBS we also have access to the granular data on the income of all Dutch households. We use this both to check for the representativeness of our sample and to perform a weighted analysis. We comment below on the representativeness and, in this context, also provide descriptive statistics for socio-demographic characteristics, including income and financial wealth, both for our sample and the population (cf. also Table10 in the appendix). 2.3 Representativeness ofmatched data Via CBS, we have access to household-level information for the whole Dutch population and can compare this to our sample. In Table1, we first provide detailed information on our main explanatory variable of interest, disposable income, which we have deflated and standardized for household size. Table1 shows that the mean, median, and key percentiles of the income variable in our sample are similar to those obtained from the population as a whole. Table10 in the appendix provides, in addition, key statistical descriptives for all relevant sociodemographic variables, including income and financial wealth, and for both endpoints of our considered period from 2011 to 2018. Once more, we report figures Table 1 Summary statistics of disposable income in sample and population Table1 shows summary statistics for disposable income per standardized household member. Values are expressed in constant 2018 Euro. Household size is standardized using an OECD scale defined as 1 + 0.7(n1) + 0.3k, where n is the number of adults and k the number of children in the household Disposable per capita income 2011 2018 Mean—S 23,379.33 24,037.68 MeanPop 23,532.35 24,852.26 P10—S 14,736.30 14,683.00 P10—Pop 12,164.27 12,513.91 P25—S 17,660.22 17,795.22 P25—Pop 15,925.80 16,552.00 P50—S 21,993.73 22,571.00 P50—Pop 21,001.24 22,349.00 P75—S 27,269.93 28,410.00 P75—Pop 27,768.59 29,635.88 P90—S 33,986.74 35,008.50 P90—Pop 36,471.79 38,164.17 N—S 3,110 5,451 N—Pop 7,544,704 7,973,137 272 C.Brancatelli, R.Inderst both for our sample and for the whole population. Overall, this confirms that along these dimensions our sample is largely comparable to the whole population. Notably, when looking at similarities along the distribution, we find that the largest differences prevail at the 10th and 90th percentiles. In particular, this may reflect well-known problems in the under-representativeness of commercial panels at the tails of the income distribution. To alleviate concerns that this may bias the effects of interest in our analysis, we exploit the access to CBS registry data to construct sampling weights based on the procedure described in Brancatelli etal. (2022). For our main regression and robustness checks we provide separate results including those weights.12 3 Stylized facts 3.1 CPG consumption andretailing Figure1 illustrates that household expenditures for CPGs represent a sizable component of total disposable income, namely between 9.5% and 11.5% over time. The products under consideration are divided into two main segments: food and non-food items. The non-food segment includes, for example, detergents, cleaning products, and pet supplies. Food items consist of beverages as well as various types of fresh and refrigerated foods. Our data allow us to break down total CPG expenditures into food and non-food components, which we distinguish between at various points in the analysis. Overall, food consumption accounts for 82.1% of total CPG expenditure. Table 2 depicts the distribution of CPG (food and non-food) spending levels pooled across all household-year observations, where, again, household size has been standardized using the OECD scale. Values are expressed in constant 2018 Euro (using as a deflator the Dutch Harmonized Index of Consumer Prices, HICP). The median CPG consumption level per year and standardized household members is 1,968 euros, which masks substantial heterogeneity. Remarkably, total CPG consumption per standardized household member at the 90th percentile is three times larger than consumption at the 10th percentile. This heterogeneity remains evident even when examining variations within relatively narrow brackets of household disposable income (e.g., quintiles of 12 Specifically, we construct year-specific adjustment weights using post-stratification. We follow Bethlehem and Biffignandi (2011) and start by stratifying our regression samples into strata using disposable income from the CBS household income registry INHATAB. For each year, we first split the population into income deciles to create 10 strata of equal size. Next, we compute for each stratum a correction weight as the ratio between the population percentage of households over the sample percentage of households. Finally, we multiply the correction weight by the inverse probability of each household being in the sample. Since the true (and household-specific) probability of being sampled is unknown and since we assume that households within the stratum are homogeneous concerning their inclusion probability, we assign an equal inclusion probability to each household within the stratum. We impute the stratum-specific inclusion probability with the ratio of households within the stratum over the total number of households in the population. 279 CPG consumption intimesofrecession: novel evidence from… We interpret these results to be consistent with the view that households may adjust consumption also with some delay to an initial income change, for instance, due to a change in perceptions or due to revised expectations about future income. 4.3 Potentially asymmetric responses Various studies have documented asymmetries between upand downswings in the economy or likewise income increases and decreases. For instance, Deelersnyder etal. (2004) document this over the business cycle for durable goods. With regards to CPG, there is also some evidence regarding asymmetric responses to consumers’ share of private labels (Lamey etal., 2007). In macroeconomics, based on survey questions various studies find that households react more strongly to negative changes (Christelis etal., 2019). Such asymmetries may derive from the way households gain or lose trust in the economic climate. Also, it has been suggested that consumers may first use additional financial slack to pay back debts during recovery Table 4 Main regression analysis Table4 presents the estimation results of the baseline regression. The dependent variable is the log of total CPG expenditures per year. Column 1 presents results including only the log of disposable income, next to household and year fixed effects. Column 2 adds demand side variables, column 4 includes the lag of log disposable income. Column 5 adds supply-side controls to the specification in column 2. Standard errors, clustered at the household level, are displayed in parentheses. Significance levels are defined as follows: * p < 0.10, ** p < 0.05, *** p < 0.01. All models are estimated using a constant (1) (2) (3) (4) (5) Log(CPG) Log(CPG) Log(CPG) Log(CPG) Log(CPG) Log(HH income) 0.134*** 0.060*** 0.058*** 0.048*** 0.060*** (0.012) (0.010) (0.011) (0.011) (0.010) L.Log(HH income) 0.047*** (0.009) Share of females in HH −0.110*** −0.108*** −0.090** −0.109*** (0.036) (0.038) (0.038) (0.036) Share of children in HH 0.126*** 0.138*** 0.128*** 0.125*** (0.030) (0.034) (0.033) (0.030) Share of income earners in HH 0.007 0.003 −0.003 0.007 (0.018) (0.022) (0.020) (0.018) HH members 0.082*** 0.078*** 0.075*** 0.081*** (0.007) (0.008) (0.007) (0.007) Household fixed effects Yes Yes Yes Yes Yes Year fixed effects Yes Yes Yes Yes Yes Supply-side controls No No No No Yes Sample weights No No Yes No No Observations 30,543 30,543 30,543 25,309 30,543 R20.900 0.903 0.901 0.917 0.903 Adjusted R20.880 0.884 0.881 0.898 0.884 280 C.Brancatelli, R.Inderst or that differences are accounted for by loss aversion or a precautionary savings motive (see also below for a specific analysis of consumers’ financial situation). We define an income change by calculating the year-on-year percentage change of deflated disposable income for each household. Next, we divide the sample by separating negative rates from those being positive or zero. Overall, 36% of all measured income changes are negative. As reported in Table5, the coefficient remains economically small also for negative income changes: The coefficient’s size in column 2 implies that a 10% decrease in disposable income is associated with a 0.81% decrease in CPG consumption. We note also that while the reported coefficient is slightly larger for negative income changes than for positive income changes, the respective confidence intervals are again relatively large. When we estimate a single regression with a dummy denoting negative income changes, the respective coefficient is not statistically significant (even on a 10%-level). Table 5 Negative vs. positive income changes Table5 presents the estimation results separated by negative (columns 1 through 2) and positive (columns 3 through 4) income changes. The dependent variable is the log of total CPG expenditures per year. Column 1 (respectively 3) presents results including only the log of disposable income, next to household and year fixed effects. Column 2 (respectively 4) adds demand side variables. Standard errors, clustered at the household level, are displayed in parentheses. Significance levels are defined as follows: * p < 0.10, ** p < 0.05, *** p < 0.01. All models are estimated using a constant Negative income changes Positive income changes (1) (2) (3) (4) Log(CPG) Log(CPG) Log(CPG) Log(CPG) Log(HH income) 0.162*** 0.085*** 0.139*** 0.063*** (0.024) (0.023) (0.016) (0.015) L.Log(HH income) Share of females in HH −0.128** −0.111** (0.058) (0.045) Share of children in HH 0.079 0.142*** (0.069) (0.034) Share of income earners in HH −0.004 0.022 (0.035) (0.022) HH members 0.086*** 0.079*** (0.012) (0.009) Household fixed effects Yes Yes Yes Yes Year fixed effects Yes Yes Yes Yes Supply-side controls No No No No Observations 10,994 10,994 19,549 19,549 R20.932 0.935 0.910 0.912 Adjusted R20.890 0.894 0.877 0.881 281 CPG consumption intimesofrecession: novel evidence from… 4.4 Using GfK self‑reported income data We next conduct our main regression using now GfK self-reported income. As shown in Table 6, the income coefficient remains small even with the updated measure: a coefficient of 0.079 indicates that a 10% change in disposable income is associated with a 0.76% change in CPG consumption. While this aligns with our previous findings, a detailed comparison between coefficients derived from administrative versus reported income data warrants further examination. Using GfK self-reported income, one might expect a weaker relationship with expenditures for several reasons: (1) GfK data only captures the income of the household’s main earner, (2) it is recorded in 200-Euro brackets, potentially limiting variability, and (3) it may contain reporting inaccuracies, including reluctance Table 6 Analysis using GfK self-reported income. (GfK’s income figures are also censored at the bottom and the top, with respective categories of below 700 EUR and above 4.100 EUR, though only ca. 6% of households fall into these categories. As we associate each household with the mean of the respective bracket, we lose those in the highest band. In addition, for around 7% of household years income is missing, which together explains the drop in observations.) Table 6 presents the estimation results of the baseline regression using self-reported income. The dependent variable is the log of total CPG expenditures per year. Column 1 presents results including only the log of disposable income, next to household and year fixed effects. Column 2 adds demand-side variables, column 3 includes the lag of log disposable income, and column 5 supply-side controls to the specification in column 2. Standard errors, clustered at the household level, are displayed in parentheses. Significance levels are defined as follows: * p < 0.10, ** p < 0.05, *** p < 0.01. All models are estimated using a constant (1) (2) (3) (4) Log(CPG) Log(CPG) Log(CPG) Log(CPG) Log(HH income) 0.114*** 0.079*** 0.075*** 0.079*** (0.014) (0.013) (0.015) (0.013) L. Log(HH income) 0.002 (0.015) Share of females in HH −0.106*** −0.086** −0.105*** (0.037) (0.040) (0.037) Share of children in HH 0.081*** 0.069** 0.080** (0.031) (0.035) (0.031) Share of income earners in HH 0.020 0.014 0.019 (0.019) (0.021) (0.019) HH members 0.094*** 0.090*** 0.094*** (0.007) (0.007) (0.007) Household fixed effects Yes Yes Yes Yes Year fixed effects Yes Yes Yes Yes Supply-side controls No No No Yes Observations 28,328 28,312 23,233 28,312 R20.901 0.905 0.918 0.905 Adjusted R20.880 0.885 0.899 0.885 282 C.Brancatelli, R.Inderst among panelists to update their reported income. In Brancatelli etal. (2022), we suggested that these factors likely contributed to the weaker relationship between self-reported income and households’ private label shares. However, self-reported income may better capture households’ perceived economic well-being, possibly indicating a stronger link to consumption. Our current findings on total CPG consumption support this perspective, and we revisit this issue in our concluding remarks. 4.5 Analysis ofpotential heterogeneity Our finding of only a modest relationship between income measures and CPG consumption may conceal effect heterogeneity within the sample, suggesting potentially larger coefficients for certain subgroups. To investigate this possibility, we analyze Table 7 Analysis of potential heterogeneity Table7 presents the estimation results of modified baseline regressions. All columns use the same set of control variables as in baseline Table4. Column (1) repeats the baseline from Table4. In column (2) the dependent variable is the log of total CPG food expenditures per year. In column (3) the dependent variable is the log of total CPG non-food expenditures per year. Column 4 presents results including an interaction term of Log(income) and a dummy variable for low household income. Column 5 presents results including an interaction term of Log(income) and a dummy variable for low household financial wealth. Standard errors, clustered at the household level, are displayed in parentheses. Significance levels are defined as follows: * p < 0.10, ** p < 0.05, *** p < 0.01. All models are estimated using a constant (1) (2) (3) (4) (5) Log(CPG) Log(Food) Log(Non-food) Log(CPG) Log(CPG) Log(HH income) 0.060*** 0.060*** 0.058*** 0.069*** 0.062*** (0.010) (0.010) (0.020) (0.013) (0.011) Low Income x Log(HH income) −0.024 (0.021) Low liquidity x Log(HH income) −0.010 (0.027) Share of females in HH −0.110*** −0.133*** 0.060 −0.109*** −0.110*** (0.036) (0.036) (0.059) (0.036) (0.036) Share of children in HH 0.126*** 0.131*** 0.109** 0.128*** 0.126*** (0.030) (0.031) (0.052) (0.030) (0.030) Share of income earners in HH 0.007 −0.000 0.034 0.007 0.007 (0.018) (0.018) (0.034) (0.018) (0.018) HH members 0.082*** 0.086*** 0.066*** 0.081*** 0.082*** (0.007) (0.007) (0.011) (0.007) (0.007) Household fixed effects Yes Yes Yes Yes Yes Year fixed effects Yes Yes Yes Yes Yes Supply side controls No No No No No Observations 30,543 30,543 30,540 30,543 30,543 R20.903 0.906 0.862 0.903 0.903 Adjusted R20.884 0.887 0.834 0.884 0.884 283 CPG consumption intimesofrecession: novel evidence from… potential heterogeneity and present the results in Table7. Each column shows the outcome of our baseline analysis (repeated in column 1) when the regression is applied to subsamples. Column 2 focuses solely on food categories, and column 3 on non-food categories. The effect size remains similar and thus economically modest across both food and non-food categories. As suggested by macroeconomic research (e.g., Baker, 201818; Jappelli & Pistaferri, 2010), households under tighter financial constraints may react more strongly to income changes. To examine if this applies to CPG consumption in the Netherlands, we classify a household as low-income if its deflated standardized income falls within the lowest quintile. Similarly, we define a low-liquidity household based on low financial assets, using data from CBS. Columns 4 and 5 of Table7 show results when introducing an interaction term for low income or low liquidity. The statistically insignificant coefficients of the interaction term do not suggest a stronger response among these groups. Although the confidence intervals are broad, the values remain modest even at the upper bounds of standard confidence intervals. 5 Conclusion In this paper, we combine household scanner data with administrative income data and explore CPG consumption patterns of households in times of recession. Our key finding is an economically small relationship between household income and expenditures for CPGs. Our empirical results rest on a series of unique institutional factors that considerably mitigate estimation challenges. First, our data cover the second episode of the Dutch double-dip recession following the financial crisis which ended in 2018. As noted above, this ensures that around one-third of yearly income changes are negative. Second, we document national pricing and assortment strategies in the Netherlands which allows to rule out a series of supply side factors. We also document that in the Netherlands, households have comparable access to discounters or store brands and that such offers are considerably cheaper. Third, administrative data allows us to minimize measurement error for our two key variables so that we can rule out such concerns as an explanation for our estimated small coefficient. Access to administrative data also enables us to learn about households’ financial savings, so we can explore whether households respond more pronouncedly to income changes when they are more likely to be financially constrained. Disaggregate consumption data in turn allows us to analyze effects within food and non-food categories. Yet, our key finding of an economically small relationship remains robust across various model specifications and sample splits that cater to these factors. In sum, we are confident that the (albeit small) estimated coefficient indeed captures behavioral changes. Despite this, we show also that households’ (per-capita) CPG expenditures vary widely, even within income brackets despite controlling for common over-time 18 From a marketing science perspective, see, for instance, Orhun and Palazzolo (2019). 284 C.Brancatelli, R.Inderst variation via year fixed effects. Households’ limited reaction to negative income shocks may thus be largely due to consumption preferences, which may not have been sufficiently “tested” in the analyzed recession environment. This supports earlier research by Brancatelli etal. (2022), which documents that in the Netherlands also the relationship between income and households’ consumption of store brands is economically small. While not reported there, we find such a small relationship also for households’ propensity to frequent discounters (cf. Biswas etal., 2021 for an analysis of households’ choice of store formats in the US). Casual evidence from market research suggests that the recent surge in inflation, which not only affects (relative) prices but also real income and wealth, may have had a larger impact and may indeed have tilted CPG choices. Contrasting this episode with that of the recession documented in this paper (and with it our finding of only an economically small adjustment of households) could be an avenue for future research with the matched data.19 Finally, our data enables a comparison between findings based on administrative income data and those using self-reported income data. As discussed earlier, Brancatelli etal. (2022) find that, for households’ private-label share, the income coefficient was significantly attenuated when relying on self-reported data. That study explored potential reasons for this attenuation, including the income bracketing applied by both GfK and Nielsen. By leveraging both self-reported and administrative income data, we verified that findings from Dubé et al. (2018), which used Nielsen selfreported income, are not solely due to measurement error in self-reported income. The relationship between income and households’ private-label share remains economically modest regardless of data type. Our current analysis corroborates this for total CPG consumption, showing that the relationship between income and consumption remains modest both with administrative and self-reported income data. While the small effect sizes and relatively broad confidence intervals prevent us from examining in depth why one income measure might produce a more pronounced relationship than the other, we propose that future research should address this question. Such investigation could enhance our understanding of the limitations associated with self-reported income data. Finally, our data allows us to contrast findings with administrative income data to those with self-reported income data. As discussed above, Brancatelli etal. (2022) find that for households’ private-label share, the coefficient was much attenuated when relying on self-reported data. The paper tried to disentangle various reasons for this, including the bracketing of income reports, as practiced by both GfK and Nielsen. Using both self-reported and administrative data allowed us to confirm 19 The matching of administrative and household scanner data for the present analysis ended in 2018, given also the focus on the Netherlands’ past recession. In contrast, while CPG supply was severely disrupted in the first phase of the COVID-19 crisis, contrary to other countries income that was not derived from self-employment was largely unaffected in the Netherlands, given extensive public support programs to firms and employees. 285 CPG consumption intimesofrecession: novel evidence from… that the findings in Dubé etal. (2018), which uses Nielsen self-reported income, are not due to measurement errors in self-reported income: The relationship between income and households’ private-label share remains economically small in either case. Our present analysis confirms this for total CPG consumption, as again the relationship remains economically small both with administrative income data and with self-reported income. While the small effects, together with relatively large confidence intervals, do not allow us to dig deeper into whether and why the relationship with one measure of income is more pronounced, we suggest addressing this issue in future work. This should allow a better understanding of potential limitations when working with self-reported income. Appendix Sociodemographic characteristics: Overall GfK data vs. regression data We report key summary statistics of panelists in GfK data compared to only those who consented to the matching with CBS. Table8 Table 8 Summary statistics of panelists in GfK compared to consenting households in 2011 Table8 shows key summary statistics for the year 2011 for sociodemographic variables obtained from GfK data for both all GfK households and households consenting to the matching with CBS. Panel A shows summary statistics for age, panel B shows summary statistics for household size, panel C focuses on social class, and panel D on the income of the household head. The table is based on appendix Table14 in Brancatelli etal. (2022) Panel A: Age Panel B: Household size Panel C: Social class Panel D: Income class All Consent All Consent All Consent All Consent Mean 6.94 7.40 2.49 2.40 4.53 4.61 14.35 14.72 SD 3.02 2.53 1.24 1.20 1.64 1.61 6.25 6.00 P5 1.00 1.00 1.00 1.00 1.00 1.00 3.00 3.00 P25 6.00 6.00 2.00 2.00 4.00 4.00 9.00 11.00 P50 7.00 8.00 2.00 2.00 5.00 5.00 16.00 16.00 P75 9.00 9.00 4.00 3.00 6.00 6.00 20.00 20.00 P90 11.00 11.00 4.00 4.00 6.00 6.00 21.00 21.00 P95 11.00 11.00 5.00 5.00 6.00 6.00 22.00 22.00 286 C.Brancatelli, R.Inderst National pricing We describe how retailers in the Netherlands practice national pricing. To illustrate that national pricing existed throughout the entire period for Dutch retailers, we analyze price variation across regions for six major retailers in the Netherlands. For this purpose, we remove products with prices that are fourfold higher than the median or smaller than a quarter of the median price. Furthermore, we also disregard barcodes artificially introduced by GfK to classify unpackaged fresh food products, which usually do not have a GTIN. From this set of products, we identify and keep the 500 most popular products per retailer per year and separately estimate a price regression for each retailer and year from 2011 to 2018. In our regression, the dependent variable is the log of a household shopping tripspecific price of any given product. The independent variables comprise a set of retailer-specific barcode dummies, monthly time fixed effects, and three regional fixed effects.20 Throughout the analysis, we use North Netherlands as the basis. Overall, we estimate price regressions for six retailers and eight years21 and extract the set of regional fixed effects for each regression. Summary statistics are represented in Table9, showing that across all retailers and years, the fixed effects are considerably small, with more than 90% being smaller than 1%. We finally note that even if a given retailer practiced regional price differentiation not captured by this approach, this would not risk confounding our interpretation of the income coefficient as long as price differences do not change over time. Table 9 Summary statistics of regional fixed effects across all years and retailers Table9 depicts summary statistics of regional fixed effects from independent yearly regressions for six major retailers based on the household scanner data of the GfK consumer panel between 2011 and 2018. The table is based on Table4 in Brancatelli etal. (2022) Mean SD Skew P5 P10 P25 Median P75 P90 P95 N Regional FE 0.0017 0.0042 1.1826 -0.0031 -0.0022 -0.0009 0.0008 0.0031 0.0078 0.010 132 20 Household scanner data do not contain geo information about individual shopping trips. Therefore, we rely on household location (postcode) to construct regional fixed effects. Regions are derived from the official NUTS 1 codes of the Netherlands: North Netherlands, East Netherlands, West Netherlands, South Netherlands. 21 In sum, we estimate 44 regression specifications: Until 2014, we estimate the regression for six major retailers. Since two major retailers merged in 2014, we only estimate the regression for five retailers from 2015 onward. 287 CPG consumption intimesofrecession: novel evidence from… Socio‑demographics forhouseholds inregression data Table10 Table 10 Summary statistics of key socio-demographic variables Table10 shows summary statistics for both the analysis sample and the population for the following socio-demographic variables for the years 2011 and 2018: disposable income per standardized household member, financial assets per standardized household member, the number of household members, the number of females, the number of children and the number of income earners in a household. Columns abbreviated with S. show the corresponding statistics for the sample and columns abbreviated with Pop. Show corresponding population figures Note that household size for disposable income and financial assets is standardized using an OECD scale (defined as 1 + 0.7(n1) + 0.3k, where n is the number of adults and k the number of children. Euro values are expressed in constant 2018 Euro Year Label m—S m—Pop p10—S p10—Pop p25—S p25—Pop p50—S p50—Pop p75—S p75—Pop p90—S p90—Pop 2011 Disposable income p.c 23,379.33 23,532.35 14,736.30 12,164.27 17,660.22 15,925.80 21,993.73 21,001.24 27,269.93 27,768.59 33,986.74 36,471.79 2018 Disposable income p.c 24,037.68 24,852.26 14,683.00 12,513.91 17,795.22 16,552.00 22,571.00 22,349.00 28,410.00 29,635.88 35,008.50 38,164.17 2011 Financial assets p.c 29,624.24 37,910.28 1,060.96 344.06 3,333.90 1,943.71 11,589.33 9,032.25 29,338.02 26,900.76 66,394.65 76,083.19 2018 Financial assets p.c 27,996.16 36,221.45 715.00 346.15 2,677.65 1,790.77 9,814.00 8,731.18 25,378.06 26,476.00 58,580.00 70,196.00 2011 Household members 2.46 2.18 1.00 1.00 2.00 1.00 2.00 2.00 3.00 3.00 4.00 4.00 2018 Household members 2.40 2.12 1.00 1.00 1.00 1.00 2.00 2.00 3.00 3.00 4.00 4.00 2011 Number of children 0.49 0.43 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 2.00 2.00 2018 Number of children 0.49 0.39 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 2.00 2.00 2011 Number of females 1.26 1.10 1.00 0.00 1.00 1.00 1.00 1.00 2.00 1.00 2.00 2.00 2018 Number of females 1.25 1.07 1.00 0.00 1.00 1.00 1.00 1.00 2.00 1.00 2.00 2.00 2011 Number of inc. earners 1.94 1.72 1.00 1.00 1.00 1.00 2.00 2.00 2.00 2.00 3.00 3.00 2018 Number of inc. earners 1.92 1.73 1.00 1.00 1.00 1.00 2.00 2.00 2.00 2.00 3.00 3.00 288 C.Brancatelli, R.Inderst Funding Open Access funding enabled and organized by Projekt DEAL. One author was partially funded by a long-term research grant from the Think Forward Initiative (TFI, grant 3902030019). Declarations Competing interests None of the authors has competing interests. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/ licenses/by/4.0/. References Allcott, H., Diamond, R., Dubé, J. P., Handbury, J., Rahkovsky, I., & Schnell, M. (2019). Food deserts and the causes of nutritional inequality. The Quarterly Journal of Economics, 134(4), 1793–1844. Attanasio, O. P., & Pistaferri, L. (2016). Consumption inequality. Journal of Economic Perspectives, 30(2), 3–28. Baker, S. R. (2018). Debt and the response to household income shocks: Validation and application of linked financial account data. Journal of Political Economy, 126(4), 1504–1557. Baker, S. R., & Yannelis, C. (2017). Income changes and consumption: Evidence from the 2013 federal government shutdown. Review of Economic Dynamics, 23, 99–124. Banks, J., Blundell, R., & Lewbel, A. (1997). Quadratic Engel curves and consumer demand. Review of Economics and Statistics, 79(4), 527–539. Bethlehem, J., & Biffignandi, S. (2011). Handbook of web surveys. Wiley. Biswas, S., Chintagunta, P.K., & Dhar, S.K. (2021). Do households’ budget allocations vary with economic factors? Evidence from Nielsen data. Brancatelli, C., Fritzsche, A., Inderst, R., & Otter, T. (2022). Measuring income and wealth effects on private-label demand with matched administrative data. Marketing Science, 41(3), 637–656. Christelis, D., Georgarakos, D., Jappelli, T., Pistaferri, L., & Van Rooij, M. (2019). Asymmetric consumption effects of transitory income shocks. The Economic Journal, 129(622), 2322–2341. De Meijer, C., O’Donnell, O., Koopmanschap, M., & Van Doorslaer, E. (2013). Health expenditure growth: Looking beyond the average through decomposition of the full distribution. Journal of Health Economics, 32(1), 88–105. Deelersnyder, B., Dekimpe, M. G., Savary, M., & Parker, P. M. (2004). Weathering tight economic times: The sales evolution of consumer durables over the business cycle. Quantitative Marketing and Economics, 2(4), 347–383. Dekimpe, M. G., & Deleersnyder, B. (2018). Business cycle research in marketing: A review and research agenda. Journal of the Academy of Marketing Science, 46(1), 31–58. Dubé, J. P., Hitsch, G. J., & Rossi, P. E. (2018). Income and wealth effects on private-label demand: Evidence from the Great Recession. Marketing Science, 37(1), 22–53. Dubois, P., Griffith, R., & O’Connell, M. (2022). The use of scanner data for economics research. Annual Review of Economics, 14, 723–745. Dutt, P., & Padmanabhan, V. (2011). Crisis and consumption smoothing. Marketing Science, 30(3), 491–512. IPLC (2016). Retaining consumers tempted by the discount model (Research Report). IRI (2016). Special report: Private label in western economies (Research Report). Jappelli, T., & Pistaferri, L. (2010). The Consumption response to income changes. Annual Review of. Econonomics, 2, 479–506.