DO SOCIODEMOGRAPHIC FACTORS IMPACT IN THE USE OF CASH IN RETAIL TRANSACTIONS

This paper presents the result of a study that assessed the use of cash in retail transactions in the city of São Paulo and the influence of individuals´ sociodemographic condition and transaction value on the likelihood of paying with banknotes. Respondents were selected by quota sampling process based on their sociodemographic profiles. The analyses were carried out using Pearson´s chi-squared test, as well as estimates from probit models. The results indicated that 56% of the respondents´ last purchases were settled in cash, and that both the transaction value and the buyer´s sociodemographic condition influence the likelihood of using banknotes.


Introduction
The presence of paper banknotes in our cultures has been questioned for at least half a century (Reisted, 1967). Predictions and expectations of cashless societies with the obsolescence of cheques and the appearance of purely electronic systems are not uncommon and have become assiduous since the appearance of the credit card in the 1950s (e.g., Bergsten, 1967;Caskey & Sellon, 1994;Szmigin & Foxall, 1999). Since then, plastic cards have conquered a significant market share in retail transactions in different nations (Capgemini & BNP Paribas, 2017), reaching important percentages both in developed (Esselink & Hernández, 2017) and developing countries (Committee on Payments and Market Infrastructures [CPMI], 2015). More recently, new forms of electronic transactions are being launched, with monetary authorities around the world talking seriously about central bank digital currencies (Auer, Cornelli, & Frost, 2020).
Although recent history shows the advance of alternative payment instruments, it does not seem difficult to admit that coins and banknotes are still the most used payment instrument in retail transactions in many countries (Krüger & Seitz, 2014;Rogoff, 2016;Wang & Wolman, 2016;Wheatley, 2017). Alfonso, Tombini, and Zampolli (2020) exalted the use of cash in Latin America, with preference over cashless payments occurring even among banked users. As a novelty, the authors described the instant payment instruments Pix and CoDi, in Brazil and Mexico, that threat to reduce cash usage considerably. Pix and CoDi were created by governments, thus "they feature an option that allows individuals to send and receive payments at no cost. Furthermore, merchants´ cost to receive payments is lowered to zero in CoDi and significantly reduced in Pix" (Alfonso, Tombini, & Zampolli, 2020, p.78). However, major changes in the cash dominance are yet to be seen. Some studies released by the Central Bank of Brazil (BCB) have confirmed the dominant position of cash in Brazil, in addition to presenting a positive correlation between the value of the transaction and the frequency of the use of banknotes (BCB, 2013(BCB, , 2018. In another Brazilian study, Fávero, Belfiore and Fouto (2006) found that individuals from less privileged sociodemographic category use more cash in their transactions than privileged ones. Generally, the Brazilian data confirms its alignment with results around the world.
Doubts may arise with the possibility of a relationship between an individual´s sociodemographic condition and the transaction´s value to make our perception of the use of cash short-sighted. It can be the case when the transaction value is the driving force towards the use of cash, rather than the agent´s income or sociodemographic condition. Although this scepticism seems dissipated because the consumption baskets of different individuals or countries potentially have a higher average value for the wealthier, there is a gap in the literature urging to be filled with combined considerations of both the sociodemographic condition of buyers and the value of transactions as determinants to use cash in retail transactions.
In this context, this study sought to evaluate, as a first objective, the participation of coins and banknotes as payment instruments in transactions declared by residents of São Paulo. As the first step in the analysis, I tested the similarity in the distribution of the instruments used by sociodemographic categories. Then, I evaluated the significance of the sociodemographic conditions and the value of the transaction over the probability of using coins and banknotes, as a second objective.
The data were collected applying a questionnaire with 188 residents of São Paulo, in 2018, inquiring about the payment instrument used in their last purchases, the value of the transaction, and other information for control. The segmentation of respondents into sociodemographic clusters was supported by an exploratory factor analysis.
The results revealed that 56% of the interviewees´ last transactions were paid with coins or banknotes, suggesting the prevalence of this instrument at the time. Furthermore, the study´s results corroborated the influence of individuals´ sociodemographic condition on the likelihood of using paper money on a retail purchase, controlling for the transaction value. In other words, it supports the idea that people from a less privileged sociodemographic category tend to use relatively more coins and banknotes than the most privileged ones, regardless of the transaction value.
The paper was divided into six sections, including this brief introduction. The second section describes the study approaches to assess the use of cash. The third section explicitly presents the two hypotheses to be tested. The fourth describes the methodological aspects applied in the study, and the fifth presents its results. The sixth and final section is dedicated to the concluding remarks.

Literature Review
The literature that evaluates cash usage´s driving forces points out that both the specifics of the transactions and the characteristics of the agents can impact the likelihood a retail transaction is paid with coins or banknotes. Cash is more common in low-value transactions (Humphrey, Pulley, & Vesala, 1996;Klee, 2008;Wang & Wolman, 2016) and in some types of commerce (Hayashi & Klee, 2003); and some authors suggest a negative impact of buyers´ incomes on the use of coins and banknotes (e.g., Carow & Staten, 1999).
Interestingly, some of the factors driving the use of cash in retail transactions can be directly associated to merchants´ propensity to accept other payment instruments. At major urban cities, plastic cards are the leading alternatives. In online shopping, they have increased significantly in the last years, with a further push during the COVID-19 pandemic for several products. However, driven by the use of cash, this paper remains focused on only physical payment transactions, such as those taken place at brickand-mortar shops or street vendors.
Credit and debit cards are amongst payment instruments that were most widespread in urban populations at the end of the 20th century. The advancement of their embedded technology from magnetic stripes to contactless quality with the incorporation of smart chips with radio frequency identification (RFID) explain the phenomenon (CPMI, 2015).
The propensity for merchants to accept card payments can be driven by the possibility of discriminating non-card instruments with discounts or surcharges, and the rates and fees of point-of-sale terminals and cash advances. Against card acceptance, the possibility of tax evasion or the use of cash for illicit activities due to their low traceability should increase the demand for banknotes. Humphrey (2004) and Rogoff (2014 and2016) cited these advantages of cash to advocate for restrictions on the issuance of high-value notes by governments, in a gradual process of eliminating paper money.
Plastic cards can have some other benefits over paper money and other payment instruments (Guedes, Curi, Arashiro, Ribeiro, & Mazetto, 2011). Practicality is an apparent advantage. There is evidence that the use of credit cards increases the probability of consumption (Braga, Isabella & Mazzon, 2013) and increases the average purchase ticket (Prelec & Simester, 2001;Soman, 2001), serving as an alternative to loans term (Demirgüç-Kunt & Klapper, 2012). Merely recording transactions makes cards convenient for small businesses without organized bookkeeping. In violent urban centres, the attributes of safety and practicality can be relevant and generate benefits for both cardholders and merchants. Cards are also associated with financial inclusion by their use in the promotion and distribution of social benefits (Demirgüç-Kunt & Klapper, 2012). On the other hand, advantages related to the use of cards in day-today transactions do not come free of charge. Many cardholders do not avoid paying annual fee to issuers.
Another aspect related to the use of cash is the impact of support infrastructure, such as ATMs. There is abundant evidence that banknote use increases with the number of ATMs in an economy, while the balances in banknotes held at home or wallets are reduced (Boeschtoen & Fase, 1992;and Humphrey, Pulley, & Vesala, 1996). From the banks´ point of view, economy of scale is relevant in the ATM supply, although it does not necessarily have an effect on profitability (Humphrey, 2004).
As the benefits of ATMs to economies and service providers are related to scale, it is usual to find joint operations in different countries, in order to dilute unit costs and not to compete in cash distribution services. TecBan is a Brazilian company owned by the main banking institutions in the country and provider of Banco24Horas ATMs throughout the country that fits this idea. Along the same line, Snellman and Virén (2006) advocated that the concentration of ATM networks in associations or a smaller number of companies generate incentives to reduce the number of terminals, negatively influencing the demand for cash. Although ATMs facilitate payments thus stimulating the economy, it also spurs cash use by incentivising withdrawing.
Regardless of recognized imperfections and difficulties in analysing the use of different payment instruments (Krüger & Seitz, 2012), many authors conclude that electronic devices generate lower social costs than physical instruments such as cash (Hancock & Humphrey, 1997;Humphrey, Willesson, Lindblom, & Bergendahl, 2003;Valverde, Humphrey, & Del Paso, 2004;Krüger & Seitz, 2012). In comparison to some other payment alternatives, banknotes have already been associated with smaller average purchasing values (Hirschman, 1979;Feinberg, 1986) and lower willingness to spend (Prelec & Simester, 2001), mainly but not exclusively, because cash does not serve as an option for credit, like cheques or cards (Demirgüç-Kunt & Klapper, 2012). As a concern exacerbated by the recent COVID-19 pandemic, and previously studied by some authors, paper banknotes can be an important vector in the transmission of viruses (Thomas et al. 2008;Auer, Cornelli, & Frost, 2020). So, why banknotes are massively used in many countries? Similarly, is the overuse of banknotes in the country a symptom of some dysfunction in the payment instruments market?
It seems clear that any dysfunction in the payment instrument market could or should be explained not only in the instruments per se, but also in other elements of a national economy. In Brazil, for example, where the economy shows a high level of informality (e.g., Schneider, 2004), the large share of conventional instruments such as paper banknotes does not seem to be an anomaly. Nevertheless, the share of cash used in retail transactions should be contemplated for further inferences.
Assessing the use of banknotes in an economy requires segregating the focus onto supply and demand. On the supply side, the production of banknotes and coins are quantified and accessible. Despite the advantage of data availability, this focus restricts the assessment of the effective use of banknotes in a society to its production. On the demand side, research is sparser though more granular and largely supported by data from families and individuals. The paper follows with comments on the two approaches.

Studies from the supply side perspective
From a supply perspective, the studies that investigate cash as a payment instrument focus on banknote production and circulation, largely benefiting from the fact that the data is known to the monetary authorities. The concept of monetary base is usually employed in this approach, excluding the cash held by banking institutions. Nevertheless, the data associated with the money supply that highlight patterns for the amount of cash in circulation or kept by households, also present some well-known deficiencies.
One notable gap in the supply approach is the level of data aggregation. Considering the ratio between the paper money held by the public and the gross domestic product of a country, as in Rogoff (2016), it is possible to estimate the levels of participation of banknotes in different nations, which vary, in most cases, between 3% and 20%. Although the data are comparable across nations, information on how much of that share is used in retail transactions is not provided.
Furthermore, the presence of a strong currency, such as the dollar, euro or Swiss franc, which holdings by foreigners make statistics more challenging, seems to influence the number of banknotes in circulation.
In the US, it is estimated that 40% to 70% of the dollars issued and held by the public are outside the country (Porter & Judson, 1996).
Additionally, Rogoff (2016) argued that a nation´s inflation influences the use of cash in an economy, endorsed by the evidence that countries with a history of high inflation such as Argentina, Nigeria and Brazil show low values of local currency per capita in their economies. In such cases, the loss of the reserve value of the notes makes the population turn to other alternative assets.
For the reasons above mentioned, evaluations focused on the supply side do not provide a conclusive view on the use of banknotes in economies, despite its importance. Therefore, more attention is attributed to the demand side approach.

Studies from the demand perspective
From a demand perspective, well-known studies that evaluate the possession or use of paper money focus on the characteristics of users or transactions. They can be divided into two major methodological groups. The first includes sample surveys of individuals or families in which respondents are explicitly asked about the transactions they carried out and the payment instruments used to settle them. In Brazil, the Family Budget Survey (POF) of the Brazilian Institute of Geography and Statistics (IBGE) is an example of this approach, allowing the differentiation between the share of credit cards and other instruments. However, POF does not distinguish instruments such as banknotes, cheques and debit cards in different categories, segregating only credit cards from all other payment instruments grouped under the name "cash purchase".
Other surveys in Brazil evaluate the possession and use of banknotes and coins, with the BCB sponsoring some of them (BCB 2013(BCB , 2018. The most recent ones endorsed the relevance of money in the Brazilian economy. One evidence of the importance of banknotes was given by the statistics showing that 37% of the wages in the country were paid in cash in 2018 (BCB, 2018).
In the US, both the Diary of Consumer Payment Choice (DCPC) and the Survey of Consumer Payment Choice (SCPC) allow inferences about the number of payments made with banknotes and coins, in addition to the Survey of Consumer Finances (SCF). The first two are organised by the Federal Reserve of Boston and the last one by the Fed Board. The DCPC has proved itself to be an improvement over the SCPC in two aspects (Greene, O´Brian, & Schuh, 2017). First, instead of using the respondent´s memory to remember past transactions, as the SCPC does, the DCPC asks respondents to keep notes on current expenses each day. Additionally, instead of tracking only the quantities of transactions carried out with each payment instrument, the DCPC keeps a record of the values of these transactions, in a similar manner to the Brazilian IBGE´s POF. Despite the DCPC´s richness of details, methodological changes that have occurred between editions make comparative assessments difficult over time. With an increased comprehensive focus on household consumption, the SCF -held every three years -is also adopted in studies about payment instruments.
The second approach or methodological line commonly employed to evaluate the use of banknotes and coins retains data from transactions, collected directly from the establishments´ cash registers, as in Klee (2008) and Wang and Wolman (2016). In addition to the advantage of extracting factual payment instrument used, data from retailers provide additional statistics such as the settlement speed associated with each instrument, at the cash desk, generating useful information for better understanding of people´s use of banknotes.
Studies that apply one of the two approaches show, in general, that cash is associated with both the particularities of transactions and the characteristics or sociodemographic factors of individuals. Regarding the characteristics of transactions, the use of banknotes is more common in low-value purchases (Humphrey et al. 1996;Bounie & François, 2006;Klee, 2008;BCB, 2013;Wang & Wolman, 2016), being the instrument most used in some type of businesses, such as restaurants and gas stations 2021 20 (Hayashi & Klee, 2003). Moreover, paper money provides the shortest time for settlement between competing instruments (Caskey & Sellon, 1994;Klee, 2008).
Under a socioeconomic focus, some published results have endorsed that individuals of lower wealth or lower sociodemographic strata tend to use more cash proportional to their spending levels than rich individuals or those from higher sociodemographic strata (Kennickell & Kwast, 1997;Carow & Staten, 1999). Such a conclusion on the level of wealth or income, however, does not yet seem to be definitive. With data from the 1998 edition of SCF, Stavins (2001) found that net wealth negatively affects the likelihood of using electronic instruments. Finally, with a plausible explanation based on the security issue, it has already been pointed out that women tend to use proportionally less banknotes than men (Bounie & François, 2006).
The hypothesis that sociodemographic factors are associated with the payment instrument choice was tested in Brazil by Fávero et al. (2006), who rejected the null hypothesis of indifference between groups of consumers with three sociodemographic profiles at the level of significance of 10%. In other words, according to the authors less privileged sociodemographic groups tend to use banknotes in their transactions to a greater extent than more privileged sociodemographic groups. However, the results were not conclusive in this respect as it was not possible to reject the same hypothesis at the 5% level of significance. Additionally, as the authors only contemplated low, medium-low and medium income groups and did not control the values of the transactions, this research sought to complement the knowledge on the topic, asking the respondents about the instruments used in their latest acquisition and the value of their purchase, in order to allow statistical control of the average values of transactions by the buyer´s social subgroup.

Research Hypothesis
Based on the previous discussion, the first hypothesis of this study can be expressed as: • H1: The frequency distribution of payment instruments used in the last transactions of individuals in São Paulo is similar between different sociodemographic categories.
Some evidence pointed out earlier (e.g., Carow & Staten, 1999;Fávero, Belfiore, & Fouto, 2006) suggested that people from less privileged categories use cash proportionally more than people from more privileged categories, understood here by the constructed sociodemographic groups. Therefore, I expect the rejection of hypothesis H1.
As the use of more paper money by individuals of a lower sociodemographic stratum can occur due to transactions of lower average values and not due to their sociodemographic condition per se, the second hypothesis to be tested is: • H2: There is no influence of the individual´s sociodemographic category on the probability of using paper bills when controlled by the value of the transactions.
For the tests, I do not expect the rejection of the variable´s parameter representative of the sociodemographic category. Intuitively, the projected results are in line with the idea that people from less privileged areas, denoted in this paper as "poor", use banknotes differently from people in more privileged areas, regardless of the transactions´ price. Therefore, parameters from both transaction value and sociodemographic category parameters are expected to be different from null, rejecting H2.

Description of the research and methodology
To achieve the proposed objectives, a sample survey was conducted with residents of the city of São Paulo. For each individual, I asked the payment instrument used in the last transaction along with its value. The age, education level and gender of each individual at the time of the interview were also captured. The questionnaire was applied in person in high-traffic areas between February 15th and March 2nd, 2018, in different locations of São Paulo. For the data collection, ten districts were selected, two for each of the five subgroups.
The process of selecting respondents was a quota sampling with the sample structured as an equivalent proportion of the population for five established subgroups, which does not allow precise inferences about the population because it is non-probabilistic, but it admits testing whether people from the lower social strata tend to use notes more than people from the higher social strata, controlling for the values of transactions.
In the sampling process, a factor analysis was applied to stratify the target population, i.e., residents of the 96 administrative districts that make up the city of São Paulo. An indicator was constructed from the normalised values of the six variables presented below, with sources indicated in parentheses: The exploratory factor analysis (EFA) allows identifying dimensions of common variability in a set of phenomena and generating observable factors. Thus, I sought to summarise the six sociodemographic characteristics presented in a smaller number of factors, possibly a single one, with the support of an EFA. Despite the criticisms related to the subjectivity implicit in the decisions necessary to conduct a factor analysis (Henson & Roberts, 2006), the choices made were preferably over usual alternatives in studies alike.
The analysis of the correlation matrix between the variables was performed by applying the factor extractions by the principal component analysis, according to the Kaiser criterion, i.e., retaining the factors with eigenvalues greater than 1 and performing the rotation orthogonal to the matrix using the Varimax method. The adequacy of the data regarding the degree of partial correlation was assessed with the Kaiser-Meyer-Olkin test (KMO), requiring 0.7 as a minimum limit to consider the data adequate. The hypothesis of inadequacy of the constructed model was evaluated with the Bartlett test.
The data suggested the formation of a single explanatory factor of 92.5% of the observed variance, with a KMO of 0.78 and no single variable showing an indicator below 0.71. Bartlett´s test rejected the null hypothesis of variables not correlated with a p-value of 0.001. The weights of each variable in the single built factor are shown in Table 1. When applying the weights to the original variables, the factor indicator (FI) varied between -1.56 and 2.21 for the 96 administrative districts of the city. Based on the FI for each administrative district, the classification proposed by Fávero et al. (2006) and applied in this study is presented in Table 2. Although those authors used only the first three categories in their study, all the five groups were considered here. The average data for all variables and the FI by district are presented in the Appendix I. For the purpose of calculating the required sample size, the question about the payment method can be seen as dichotomous, considering whether the interviewee used banknotes or not in his or her last purchase. As the proportion of the response in the population was not known, the sample size was calculated considering an ad hoc proportion of 50%, with a confidence level of 90% and sampling error of 6%. The unusual value for the sampling error was established because of the available resources. Thus, for a population of 10,679,760 people, the required sample was 188 individuals, distributed among the groups as described in Table 3. Due to the fact that the application of the questionnaires extrapolated the number of individuals for some subgroups in the collection process, i.e., 304 individuals were interviewed, the surplus was removed randomly among respondents in each subgroup, keeping the desired proportion among them For the assessment of hypothesis H1 on the frequency distribution of the instruments among the subgroups, a contingency table was constructed with subgroups and payment alternatives, applying Pearson´s chi-squared test (χ 2 ).
For hypothesis H2, an econometric evaluation was used with the sectional data collected, applying probit models, in which the conditional probability of payment in cash, given the sociodemographic classification, can be seen as: where the cash payment option is represented by the cash binary variable, which values 1 when the last transaction made by the respondent was paid with cash and 0 when another instrument was used, i.e., credit card, debit card, cheque, prepaid card, store card or voucher; poor is a dummy variable that indicates whether an individual resides in districts belonging to sociodemographic subgroup I or II, as classified in Table 2 by the denomination "low" or "medium-low", respectively; and ε is a stochastic term. In the probit model, G represents the normal cumulative distribution function.
One of the challenges of the model presented in equation 1 is the absence of other relevant variables to explain the use of banknotes, such as the value of the transaction (Humphrey, Pulley, & Vesala, 1996;Bounie & François, 2006;Klee, 2008;Wang & Wolman, 2016). Furthermore, it has been pointed out that women use less cash than men (Bounie & François, 2006). The lack of these variables would imply a non-consistent parameter γ1. Therefore, they must be included in the estimates.
Additionally, due to the fact that the questionnaire applied (in the Appendix II) captured the years of education and the respondents´ age, such information were also considered in the equation to be estimated: where X represents the set of explanatory variables; cash, poor and ε are as before; value is the value of the transaction declared by the respondent, in reais (R$); gender is a binary variable equal to 1 if the respondent is a man and 0 if she is a woman; age is the number of complete years of the respondent; and educ is the respondent´s level of education measured by the number of years in school. Descriptive statistics for the collected data are presented in Table 4. The correlations between the variables for the constructed sample are shown in Table 5. As a point of attention, the presence of binary variables requires the application of specific correlation statistics for some pairs. In particular, bisserial and tetrachoric correlations were used. As a highlight, the cash variable has significant and not negligible correlations with value, poor and educ. It should be observed that education influences the variable poor because the average years of study was used in the construction of sociodemographic classifications. Thus, to mitigate collinearity between educ and poor, I considered the variable educ as an instrument in a two-stage setup. In this configuration, the equations to be estimated can be represented as: [second stage] P(cash = 1| X2) = G(γ0 + γ1poor + γ2value + γ3gender + ε2) (4) where all variables are as before; X1 contains educ and age, X2 contains poor, value and gender, and ε1 and ε2 are stochastic terms. For the two-stage setup, two difficulties occurred. The first arose as both endogenous variables, i.e., cash and poor, are binary. Such a case is configured in the so-called forbidden regression, which discourages the use of a two-stage least squares estimator (2SLS) for non-linear models resulting in inconsistent standard errors. In these conditions, the procedure of maximum joint likelihood was followed, as described in Baum (2006).
The second difficulty referred to the fact that the instrument chosen, i.e., educ, does not qualify properly as a valid one. Although the significant correlation between educ and poor does not suggest a null parameter β1 in the first stage (equation 3), eventual educ covariance with errors in the second stage (ε2 in equation 4) may show the instrument´s weakness. Intuitively, the educational level can contain information about the way an individual settles his or her retail transactions in addition to determine the sociodemographic group he belongs. For this problem, it was preferred to exacerbate this criticism to the questionnaire´s design and to maintain the proposed estimates.

Results
The results of the interviews revealed that money was the instrument most frequently used in the respondents´ declared purchases, employed in 56% of the transactions (106 times out of 188); followed by the debit card, with about a quarter of participation in total (46 times out of 188).
The contingency tables for the evaluation of the first hypothesis (H1) are shown in Tables 6 and 7, for groups I to V and the poor variable, respectively. The division by the six payment instruments considered does not show rejection to the null hypothesis of equality in the distributions by the five groupings, according to Pearson´s chi-squared test (χ 2 ), pointing to a result contrary to that of Carow and Staten (1999), who suggested a negative impact of buyers´ incomes on the use of banknotes.  Instrument  I  II  III  IV  V  1  0  Coin or Banknote  34  30  16  16  12  64  44  Debit card  11  6  9  13  7  17  29  Credit card  5  7  7  2  4  12  13  Cheque  0  0  0  1  0  0  1  Store card  1  0  0  1  0  1  1  Voucher  2  3 1 0 0 5 1 p-values of Pearson´s χ 2 (0.301) (0.074) Source: Elaborated by the author However, considering the grouping provided by the poor variable, the hypothesis of equal distributions is rejected at the 0.10 level (χ 2 (4) = 4.541), in favour of the intuition that the instruments used by individuals in the less privileged geographical areas of the city (i.e., poor = 1) are distinct from those used by others.
The dichotomous specification of payment in cash (cash = 1) or with another instrument (cash = 0) showed similar results rejecting the null hypothesis of equality in the distributions between the two groups, in line with the idea that individuals from the less privileged geographical area of the capital use proportionately as much paper money as residents of other sociodemographic background, at the 0.05 level (χ 2 (1) = 4.434). Overall, the results for the grouping category poor rejected the first hypothesis (H1) of the study, suggesting that people from a better sociodemographic background uses cash less frequently than unprivileged ones.
The analysis of the second hypothesis (H2) was constructed in a way that allows to control both transaction characteristics and individual sociodemographic background. The estimations were performed with a gradual addition of dependent variables on the model of equation 1 (versions 1A, 1B and 1C), until reaching the model of equation 2. Finally, the estimates were performed with the poor variable as instrument in a two-stage setting (equation 4). The main results of the different models are in Table 8.  The results exhibit that the transaction value has a negative impact on the probability of paying with cash at a significance level of 5% or better in all models. For the estimation of equation 2, the introduction of the educ variable reduces the significance of the poor variable´s parameter, failing to refute the null hypothesis at 10%. A caveat that must be made is that the discrete variable educ has a different amplitude from the dummy variable poor, partially explaining the loss of significance of poor when educ was introduced.
The two-stage estimation proved to be adequate, showing significant parameters for the first stage (not shown), and finding that educ presented high correlation with the variable poor, and low correlation with the errors of equation 4´s estimation. As an endorsement to the two-stage model, the parameter ρ that measures the correlation of ε1 and ε2 errors in equations 3 and 4 were significantly different from zero (χ 2 = -1.13, df = 1, p = 0.006), suggesting that the two equations are associated and should be estimated together. The general F-test presents the assessment that all parameters are null and was rejected at the 0.01 level (χ 2 = 87.89, df = 6, p = 0.000). Overall, the treatment of the poor variable`s potential endogeneity by the application of a two-stage model was adequate.
The equation`s 4 results showed significant parameters at the 0.1 level for both value and poor, indicating that not only the value of the transaction influences the likelihood of using money but also the sociodemographic condition of consumers. Thus, the results are in line with the intuition that less privilege people tend to pay more with cash, but at certain value of the good or service traded the payment mode may shift to a cashless alternative, given that the transaction value influences the probability of using cash as well.

Final considerations
Despite the disadvantages of coins and banknotes over other payment instruments, international evidence suggests that they continue to dominate retail transactions in many countries (Krüger & Seitz, 2014;Rogoff, 2016;Wang & Wolman, 2016;Wheatley, 2017;BCB, 2018). The results of the present study, which focused on residents of the city of São Paulo, were in line with this general view and indicated that 56% of the respondents' last transactions were settled with cash. Debit and credit cards were the second and third choices, respectively, being used by half of the privileged respondents (i.e., poor = 0) and by a third of the unprivileged ones (i.e., poor = 1). Although plastic cards have conquered a position among the most important payment choices, the results of this study show that they still lagged behind coins and banknotes in 2018.
Considering the fact that previous studies focusing on payment options in retail transactions are scarce and a yet smaller number considered coins or banknotes, this study pursued new evidence regarding the use of cash in Brazil, aiming to only physical payment transactions such as those that took place at brickand-mortar shops or street vendors. The exclusion of online acquisitions was purposeful and should not be viewed as a limitation. On the contrary, it helps to shed light on the most important payment choice in the city of São Paulo.
Even though the paper employed a small sample in its evaluations, it contemplated the effects of transactions' values and buyers' characteristics, and the results suggested that the probability of using banknotes in a retail transaction is influenced by both. The results substantiated the argument that regardless of the price of a good or service being bought, the sociodemographic stratum of a person influences the likelihood of paying in cash. Nevertheless, the results also suggested that some goods and services, due to their prices, are more inclined to be paid in cash, even by rich individuals. Tipping valet drivers, waiters and other service providers, together with buying from street vendors and purchasing small stationery items, candies and cigarettes, is considered among the low-value candidates.
The results can be interesting for several stakeholders in retail transactions. Focusing on cashless payments, Szmigin and Foxall (1999) claimed that a better understanding of the differences among the consumers is important for more effective segmentation; the study provides lessons to retailers, marketing professionals and researchers. Furthermore, the non-cash payment instrument suppliers, such as acquirers, card issuers or other players with innovative payment technologies, can benefit from the data and results disclosed. Some inferences can be made using the knowledge that almost 40% of the Brazilians did not own a smartphone at the time this study was taken; according to the Pew Research Center (2019), these were mostly women and old citizens. For example, it would not make much sense to provide a smartphone app-only access to pay for goods or services that are focused on the bottom of the sociodemographic pyramid.
Several new payment instruments that challenge coins and banknotes have been launched around the world. Pix is an example. Created by the Brazilian government, it provides instant money transfer via smartphones. Although privileged residents may adhere to Pix for at least some of their transactionswhether currently paid with cash or an alternative instrument such as plastic cards or cheques -lessprivileged persons may encounter a device barrier. Despite the fact that this statement is a mere conjecture, the Pew's research clearly sends a signal that differentiates the city of São Paulo from other major urban cities in developed countries, where smartphone penetration is more than 99% in some cases. Thus, the confirmation that less-privileged residents of São Paulo use more cash than more privileged ones raises a flag towards the exclusion of those who do not own smartphones in the paymentcompetitive landscape.
Like any empirical study, this also comes with criticism. When the respondents' years of study (i.e., variable educ) was adopted as an instrument for the sociodemographic category, I assumed a low covariance between educ and the probability to pay with banknotes after taking out the effects of sociodemographic category (poor), the value of the transaction (value) and the gender of buyers (gender). However, the assumed low covariance is not necessarily true although sample statistics presented small values. In addition, the study did not contemplate the supply side of the transaction, such as merchants' characteristics or characteristics of the goods or services purchased, which can provide valuable information too. Finally, the data collection period lasted less than a month. Thus, the data may be influenced by the flow of wages, which are usually paid on a monthly basis in the city. As another consequence of the same deficiency, the work did not show the growing and apparently consensual trend in the use of electronic instruments (Dodgson, Gann, Wladawsky-Berger, Sultan, & Goerge, 2015).
Lastly, after completing the study, the theme was stimulated by the COVID-19 crisis. The data collection was carried out before the pandemic, but a simplified methodology can be put to the test and prompt a similar and comparable assessment afterwards. Do the residents of São Paulo continue to use as much paper money after the pandemic? Did the difference in the use of money between distinct sociodemographic classes change after the pandemic? These questions remain open to be answered in another study.