The Effect of the Minimum Wage on Poverty: Evidence from a Quasi-Experiment in Mexico

Abstract We analyze the effect on poverty of a significant increase in the minimum wage using a quasi-experimental situation in Mexico. In January 2019, the Mexican government announced an increase in the minimum wage: in most of the country it increased by 16 per cent, while in 43 municipalities along the U.S. border it increased by 100 per cent. Using household surveys and the official method for calculating labor income poverty, we implement difference-in-difference and synthetic control methodologies to estimate whether this policy affected poverty in Mexico. We find that poverty along the border decreased by 2.6–3.0 percentage points (11–13 per cent) due to the larger increase in the minimum wage. Poverty was reduced mainly by reducing the flow from non-poverty into poverty. However, poverty intensity increased, mainly because the policy did not affect the share of families without labor income among the poorest households.


Introduction
The relationship between the minimum wage and poverty has long been debated. Since its introduction in the U.S. Fair Labor Standards Act of 1938, the minimum wage has been promoted by its advocates as an antipoverty tool and as a mechanism to guarantee a 'living wage'. On the other hand, critics of this relationship, such as Stigler (1946), have argued that 'the connection between hourly wages and the standard of living of a family is remote and fuzzy' (Stigler, 1946, p. 363). Stigler's answer to the question 'Does such legislation diminish poverty?' (Stigler, 1946, p. 358) was a definitive 'No'. His argument had two bases: (1) the standard loss of employment effect, and (2) the fact that the concept of poverty was related to household or family resources, whereas the minimum wage was addressed to individuals. In his view, 'Unless the minimum wage varies with the amount of employment, number of earners, non-wage income, family size, and many other factors, it will be an inept device for combatting poverty even for those who succeed in retaining employment' (Stigler, 1946, p. 363). The debate has continued ever since, and it has been fueled by mixed empirical evidence. using information on the cost of the food basket and on labor income as reported in the ENOE. With this information, it estimates what is called 'labor income poverty', that is, the percentage of the population that cannot afford a food consumption basket with its labor income. This is the measure of poverty we use in this paper.
In addition to the substantial increase in the minimum wage, the federal government program for the northern border also included tax incentives. The value-added tax (VAT) was reduced from 16 to 8 per cent, and the corporate income tax (CIT) was reduced from 30 to 20 per cent for firms that complied with very specific requirements. 3 These incentives may have effects in opposite directions for our estimates. On the one hand, since they mitigate the potential negative effects of the minimum wage increase on economic activity, we may overestimate the actual effect of the minimum wage on poverty. On the other hand, since they may reduce inflationary pressures at the border, we may underestimate the effect on poverty if we use the same poverty line for the entire country. Empirical evidence on this and other minimum wage increases in Mexico has not shown a large negative effect on employment, which suggests that the former effect is less likely. 4 The second possibility, of underestimating the effect, is considered in our robustness estimates, as described below. 5 In our preferred estimates, we find that the substantial increase in the minimum wage reduced poverty at the northern border by 2.6 percentage points. The results are quite similar whether we contrast the border region with the rest of the country or only with the other northern municipalities. They are also robust to whether we use a regression framework or the synthetic control method. Given that the program cut the VAT in half and that there may be a pass-through effect on prices, we also estimate the effect of the program by adjusting the poverty line to prices in each region. This increases the magnitude of the effect on poverty to 3 percentage points. These results imply an elasticity of poverty to the minimum wage that ranges between À0.18 and À0.21, in the middle of those estimates that have been found in cases where an increase in the minimum wage reduces poverty.
We address two additional issues: (1) What is the channel through which the minimum wage affects poverty? and (2) What is its effect on poverty intensity? The first was analyzed using the rotating panel characteristic of the ENOE. We attempt to disentangle whether the decline in poverty comes from a decrease in the persistence of poverty, a decrease in the transition from non-poverty to poverty, or both. We find that the minimum wage increase did not affect the persistence of poverty, but it did decrease the flow from non-poverty into poverty. Here, the main channel of poverty reduction seems to be an increase in labor income for non-poor households, reducing the vulnerability of those households that were above the poverty threshold.
Finally, we analyze the effect of the minimum wage increase on poverty intensity. Interestingly, and consistent with the results from poverty transitions, we find that there is some evidence that poverty intensity increased, despite the reduction in poverty. It thus seems that the minimum wage increase was effective for those who were relatively close to the poverty threshold, but it was ineffective for the poorest segments of the population. This result may be related to Stigler's arguments: this type of policy may be ineffective for those families with no labor income, no attachment to the labor market, or who change their level of participation in the labor market. We believe that the results we have obtained may help to clarify different dimensions of the ongoing debate about minimum wage and poverty.

The Mexican context
Mexico is considered an upper middle-income country, with a GDP per capita of $19,800 USD (in constant 2017 PPP). Mexico's growth in recent decades has been lackluster, and it has underperformed in comparison both to its Latin American counterparts and to other middleincome countries (see Figure 1, panel A). In 2019, Mexican GDP stagnated (0 per cent growth), and in 2020 it suffered the largest contraction in 80 years (À8.2 per cent).
CONEVAL uses the income-expenditure survey (available every other year) to calculate official measures of poverty. Economic growth in Mexico has been disappointing for at least the past three decades, and Figure 1, panel B shows that the percent of the population in extreme poverty (those who cannot afford a food consumption basket) has not declined significantly throughout this period. The graph plots two different measures of poverty in Mexico as estimated by CONEVAL. The continuous line is the official measure for extreme income poverty, as estimated using the income-expenditure survey since 1992. The shorter line is an alternative measure of poverty, estimated using information from the quarterly occupational and employment surveys (ENOE) since 2005. Unlike the official measure, which uses all income sources to estimate the household income, the alternative measure uses only labor income, omitting sources such as transfers. For this reason, it reflects a much higher level of poverty than the official measure of extreme poverty. Its behavior, however, approximates the shape of the official measure very well. In this paper, we use the labor income poverty definition, since it is available on a quarterly basis.  Panel C shows the monthly minimum wage in constant Mexican pesos (July 2018). The real minimum wage fell during the 1980s and 1990s and remained constant in 2000-2014. The government began to increase the minimum wage in real terms by unifying its value across three different wage zones towards the highest value. Later, it increased the minimum wage above the rate of inflation. By the end of 2018, the newly-elected government decided to double the minimum wage in the 40 municipalities along the U.S. border as well as in three municipalities in Baja California, effective 1 January 2019. The measure was combined with a cut in the valueadded tax from 16 to 8 per cent and a cut in the corporate tax from 30 to 20 per cent. However, according to data from the Secretary of Finance, only a small number of firms meet the strict requirements to qualify for the corporate tax incentive. The government also increased the nominal minimum wage in the rest of the country by 16 per cent in 2019 and 20 per cent in 2020. In 2020, the minimum wage along the northern border was 51 per cent larger than in the rest of the country, and the average for 2019 and 2020 was 62 per cent larger. This differential trajectory is also shown in Figure 1, panel C.
Panel C also includes the poverty lines for a food consumption basket in urban and rural areas (the latter defined as localities with fewer than 2,500 inhabitants). The rural poverty line is lower than the urban one to reflect the different expenditures necessary for a food consumption basket. 6 The lines increase slightly over time, reflecting the fact that food prices have increased more rapidly than overall prices. Although the Mexican Constitution calls for the minimum wage to be sufficient for a household head to cover a family's basic needs, in practice the minimum wage has fallen short of this requirement for at least three decades. The real monthly minimum wage in 2000-2013, for example, was close to MXN $2,300, and the urban poverty line varied from MXN $1,200 to MXN $1,400 in the same period. A household of four members, with two of them earning one minimum wage each, would be considered extremely poor, since it had a per capita labor income of only MXN $1,150.
Finally, Panel D includes the ratio of the minimum wage to the average wage for full-time workers across a sample of countries, including Mexico, in 2018. The figure also includes data for the northern border region and the rest of Mexico in 2020. Before 2019, Mexico had one of the lowest ratios among various countries, including Latin American countries such as Argentina, Brazil, Colombia, Costa Rica, and Chile. The ratio was smaller only in Greece and in the U.S. However, the increases in the minimum wage observed in 2019 and 2020 are significant. In fact, in 2020 the Mexican border region had one of the largest ratios across the same sample of countries, as large as in Argentina and Colombia, with a value close to 60 per cent. As the minimum wage has also increased for the rest of the country, its ratio is now close to 40 per cent, similar to countries like Belgium, Brazil, Germany, Hungary, Latvia, Lithuania, the Netherlands, Poland, the Slovak Republic, and Turkey.

Data
We use information for 2016-2020 from the quarterly National Survey of Occupation and Employment (ENOE). This survey is carried out by the National Institute of Statistics and Geography (INEGI) and is the main source of information regarding the labor market in Mexico. It has a rotating panel design and each household in the survey is followed for five consecutive quarters. It includes information on employment status, labor income (no other type of income is included), and sociodemographic characteristics for each member of the household. CONEVAL uses this survey to calculate what it defines as 'labor income poverty' in Mexico, that is, the share of the population in poverty considering only labor income. The poverty line is defined using the monetary value of a food basket (the minimum well-being line).
The official estimates of poverty in Mexico are published every other year using a full-fledged income-expenditure survey. This measure is multidimensional and evaluates both social needs (access to health services, access to social security, education lags, access to food, housing quality and space, and access to basic housing services) and income needs (using income from all sources, not only from labor). As this measure is published only every other year, CONEVAL uses the quarterly ENOE and the labor income definition of poverty to estimate poverty trends in a more timely manner. By definition, labor income poverty is greater than official extreme poverty since it does not include sources of income other than labor income. However, since the main income source for most of the population comes from labor, labor income poverty follows the official measurement of poverty quite well (see Figure 1, panel B).
The ENOE is representative at the national, urban, rural, state, and city levels. Our period of analysis goes from the first quarter of 2016 to the fourth quarter of 2020. The number of households included in each quarter is approximately 110,000, and the number of individuals at least 15 years old is approximately 300,000. Because of the COVID pandemic, the second quarter of 2020 is not fit for analysis, and the sample for the third quarter of 2020 is about 22 per cent smaller.
The ENOE provides information on the geographic location at the municipality level in Mexico. We define three relevant groups (see Map 1). First, the treatment group, or border group, includes households living in any of the 43 municipalities subject to the doubling of the minimum wage (40 that share a border with the United States plus 3 others in the nearby state of Baja California). The other two are comparison groups: households living in border states or Baja California Sur but that do not belong to the border group, and households in the rest of the country (including those in the first comparison group). We use these two alternative control groups for robustness purposes.
We use the same procedure as CONEVAL for the calculation of per capita household labor income and labor income poverty. There is, however, one issue with this information that needs to be addressed: in recent years, the percentage of workers who do not declare a specific Map 1. Map of Mexico. Notes: Authors' calculations. Northern border is the treatment group. We use two control groups: households living in municipalities in northern states including Baja California Sur, and households residing in municipalities located in the rest of the country (this includes those municipalities in the first control group).
The effect of the minimum wage on poverty 365 amount of labor income has increased; it is currently above 20 per cent, and this problem is particularly relevant in the upper tail of the distribution (Campos-Vazquez, 2013;Rodriguez-Oreggia & Lopez-Videla, 2015). If interviewees decline to answer the main income question, the ENOE offers them the possibility of instead indicating one of seven income ranges, all of which are expressed in terms of the minimum wage; CONEVAL then uses this response to impute labor income using the median value for each interval. It is unclear how precise this procedure is, and approximately 38 per cent of workers who refuse to answer the first question also refuse the second option. If there is at least one household member with missing labor income, the entire household is excluded from the labor income poverty calculation. We refer to this as the CONEVAL sample, and we use a hotdeck imputation procedure to analyze whether this procedure affects our results. The hotdeck procedure follows Campos-Vazquez (2013) and Rodriguez-Oreggia and Lopez-Videla (2015). The imputation is at the worker level, and we create groups by yearly quarter, seven age groups, gender, urban/rural, border, and five education groups (less than primary, at least primary, junior high school, high school, and college). In addition to these variables, the imputation uses the second question in the ENOE to assign an income within the interval selected. If the interviewee does not answer that question, we impute with the rest of the variables only. We create five imputations per worker, and then we use the average as a measure of labor income. If a worker with missing income does not have a match to impute income, we follow the CONEVAL procedure and drop the household's workers from the analysis. The CONEVAL procedure excludes approximately 15 per cent of households from its labor poverty calculation, but the hotdeck imputation procedure excludes only 6 per cent of the households in the period 2016-2020. We report results using both procedures.
Finally, another key question is whether we should use the same poverty line for the different regions. CONEVAL typically uses the same poverty lines for the entire country (although it maintains a separation between rural and urban areas). This made sense before 2019, when the value-added tax was the same throughout the country. However, with the reduction in VAT on the border from 16 to 8 per cent, inflation declined in that region relative to the rest of the country in 2019: Campos-Vazquez and Esquivel (2020) find that inflation there was 1.4-1.8 percentage points lower. If the poverty line is not adjusted, there may be an underestimation of the increase in poverty in the border region. However, the poverty line is constructed using only a food basket, mainly composed of non-processed food that is not taxed, so the impact from food prices could be small. 7 To assess this effect, we calculate poverty lines both for the border region and for the rest of the country. The CONEVAL poverty line uses generic products in the consumer price index: it calculates prices for corn, tortillas, bread, livestock, dairy products, fruits, vegetables, sugar, non-alcoholic beverages, and food consumed outside the household. 8 Quantities consumed are fixed over time and the price of each product varies each month. We follow the CONEVAL methodology but allow for different prices for the border and the rest of the country during the entire period of analysis (not only for 2019 and 2020). 9

Methods
The empirical strategy is twofold. First, we use a difference-in-difference design at the household level. Starting in January 2019, the northern border was affected by the doubling of the minimum wage and the halving of the VAT. Although the rest of the country also saw an increased minimum wage, at the northern border the increase in 2020 was 51 per cent. Hence, the treatment variable (T) is defined as those households living in the border region interacted with a dichotomous variable indicating whether the observation is from 2019 or 2020. The main regression specification is as follows: X is a vector of the following characteristics: rural status, family size, number of household members under 15 years old, number of members over 65 years old, and the age, years of schooling, gender, and marital status (married or living together) of the household head. We do not include other variables like formal employment status or number of hours worked, as these variables may be correlated with both the treatment and outcome variables. u t are fixed effects for each year-quarter and u region are fixed effects for the region and the state (northern border X state). The probability function is a probit. Time fixed effects control for shocks that affect both regions simultaneously (treatment and control), and regional fixed effects control for any unobserved difference in regions that does not vary over time. For instance, Tamaulipas, a northern border state, has two types of fixed effects, one for the municipalities at the border and the other for the rest of the state. Mexico City, on the other hand, only has one fixed effect. The fixed effects are similar for the other five border states (Baja California, Coahuila, Chihuahua, Nuevo Le on, and Sonora) and the rest of non-border states. The difference-in-difference estimate (h) is obtained by comparing poverty at the border before and after 2019 with a similar calculation for the control groups. Because the border region is wealthier than the rest of the country, it is not clear that the best control group should be the rest of the country. To make the analysis more transparent and robust, we use two comparison groups: (1) households living in northern border states (plus Baja California Sur), but not in border municipalities, and (2) households living in the rest of the country. As expected, households in northern states are more similar to households in the border region across different dimensions (see Descriptive Statistics subsection below).
The second methodology we use is a synthetic control analysis, following Abadie (2021) and Abadie, Diamond, and Hainmueller (2010). The border region is aggregated into one unit, and the rest of the households are aggregated at the state level. There are thus 32 units of analysis for the entire country: one for the border region and one for each of the remaining 31 states (the entire state of Baja California is part of the treatment group). The synthetic control method then hunts for trends in the states in the control group that are similar to the treated unit in the pre-treatment period. The method assigns weights (which are non-negative and total 1) to the states in the control group with the most similar trends. We construct different synthetic control groups by including lags of the dependent variable (up to half of the maximum, or six lags) and the mean of the characteristics in vector X during 2016. As in Abadie et al. (2010), the model with the minimum root mean squared prediction error (RMSPE) in the pre-treatment period is chosen. For statistical inference, this method uses a permutation test that constructs a synthetic comparison or placebo group for each state in the control group. If the increase in the minimum wage at the border really affected social outcomes, we should observe that it affected only the border region, not the placebos. The p-value of this test is then the number of times that the effect is larger in the placebos (from the permutation test), out of the total possible (31). For inference, we use both the standardized p-value adjusting for pretreatment match quality (using the RMSPE) and the p-value that is the proportion of placebos with the ratio of post-to pre-treatment RMSPE at least as large as the corresponding ratio for the treatment group (Galiani & Quistorff, 2017).
Finally, although we are mainly interested in poverty incidence, we also show results for per capita household income, for the poverty gap, and for the poverty gap squared (Foster, Greer, & Thorbecke, 1984) using multivariate linear regression instead of a probit. We use the natural logarithm of per capita income plus one to include those households with zero income. This allows us to compare how much average income increased in the treated regions as compared with the control regions. However, average income does not tell us whether the policy affected the extremely poor or the marginally poor. The poverty gap and the poverty gap squared, conditional on being poor, tell us more about how the minimum wage increase affected the intensity of poverty. 10 In particular, the poverty gap squared gives more weight to those households that are far from the poverty line than to those that are close to it.
The effect of the minimum wage on poverty 367 Finally, it is essential to keep in mind that we can only identify the effect on poverty of the increase in the minimum wage at the border in comparison to the smaller (but still substantial) increase in the minimum wage in the rest of the country. Thus, we cannot identify the total effect of the minimum wage increase on poverty in the country, but only the marginal effect coming from the additional increase at the border. Since it is likely that the main effect on poverty may occur closer to the poverty line, this factor may lead to an underestimation of the real effect of the minimum wage on poverty. Table 1 shows descriptive statistics for the sample obtained with the hotdeck procedure for missing incomes (descriptive statistics for the CONEVAL sample are in the Supplementary Materials, Table S1). It includes averages for the main characteristics at the household level for the third quarters of 2018 and 2020 for the treatment group (border) and the two control groups (northern states and rest of the country).

Descriptive statistics
On average, households on the border and in northern states are wealthier than in the rest of the country. Households on the border are more similar to those in northern states than those on the border are to those in the rest of the country. In the regressions we control for household size, the number of household members under 15 years old and over 65 years old, rural status, and the age, gender, marital status, and years of schooling of the household head. Finally, the bottom part of the table shows the incidence of poverty in each region, the poverty gap and its square, and the percentage of households with zero labor income. The percentage of households considered poor is similar on the border and in northern states in 2018, before the doubling of the minimum wage (approximately 23 per cent). In part because of the pandemic, by the third quarter of 2020 the incidence of poverty increased to 25 per cent on the border and close to 30 per cent in the northern states. The incidence of poverty in the rest of the country increased by 6 percentage points between 2018 and 2020, similar to the northern states. Figure 2 plots the incidence of poverty throughout the period of analysis in the different regions, and shows that it started to decrease on the border in 2019, although it later increased because of the pandemic. This suggests that the minimum wage increase was effective in reducing poverty in 2019, and that the negative labor demand shock from the pandemic did not fully counteract this effect. Finally, Table 1 also includes the poverty gap and its square. In all regions, poverty intensity has increased, and on the northern border it has increased proportionally more than in the rest of the country. However, the share of households with zero income increased between 2018 and 2020 for all regions, and for the northern border it increased more in relation to the proportion of households in poverty. For example, of those border households in poverty, approximately 27 per cent (0.063/0.23) had zero income in 2018, and 38 per cent in 2020. In the other regions the increase is less than 11 percentage points. In the next section we analyze whether these trends hold in regression and synthetic control frameworks. Table 2 shows the difference-in-difference (DID) estimates from the increase in the minimum wage on the border with respect to the rest of the country and to the rest of the northern region.  Figure S1 shows the equivalent for the CONEVAL sample. Due to the COVID pandemic, the second quarter of 2020 is not fit for analysis.

Effects on poverty and per capita income
The effect of the minimum wage on poverty 369 The table includes results on poverty and household income per capita for both the CONEVAL and hotdeck samples. All estimates indicate that the minimum wage increase reduced poverty an average of 1.7-2.6 percentage points in 2019-2020. We prefer the hotdeck sample, as it includes more households and the results are more stable across comparison groups. The control group used does not seem to radically affect the results, except for the effect on per capita household income in the rest of the northern region. However, with the hotdeck sample the DID estimate on per capita household income is stable at around 6.0-6.6 per cent, although it is only significant at 5 or 10 per cent. Table 3 shows the difference between outcomes at the border and their synthetic controls for each quarter in 2019 and 2020. We also include the average effect of the 2019 and 2020 minimum wage increases, which should be comparable to the DID estimate in Table 2, and is indeed similar: a reduction of 2.6-2.8 percentage points in poverty (excluding 2020 from the calculation has a similar reduction). The effects are statistically significant for every quarter in 2019 and for the fourth quarter of 2020. Results in the first quarters of 2020 seem to have been affected both by the economic effects of the pandemic and by problems associated with information collection. Overall, however, the joint test for the whole period indicates a significant effect. This provides further confidence that the minimum wage increases did indeed reduce poverty incidence along the Mexican border. Finally, Figure 3 shows the effect on poverty for the border and for each of the placebos. As expected, there is no differential trend in the pretreatment period, and the effect of the minimum wage increase is substantial throughout 2019. We conclude from these results that the COVID-19 pandemic has not reversed the effects of the minimum wage increase on poverty (although the estimates are noisy for 2020). 11 Table 3 also includes the results for per capita household income. The minimum wage increase had on average a positive effect on household income in the first three quarters of 2019, but after that period the results become inconclusive. This is consistent with the DID estimate shown in Table 2 with marginal significant results. The fading effect on household income per capita is reminiscent of what Sotomayor (2021) finds for Brazil, although it can also be explained by the large income disruption in 2020 associated with the pandemic.
How large is the effect on poverty? The reduction of 2.6 percentage points (the average of the DID estimate using the hotdeck sample and the estimates with the synthetic control method) implies an 11 per cent reduction in poverty (Table 1 shows that the poverty incidence in 2018 was 23 per cent). In 2019 and 2020, the minimum wage was 61.5 per cent higher in the border region than in the rest of the country. Thus, the elasticity of poverty we find with respect to the minimum wage is À0.18. This elasticity could be over-or under-estimated for different reasons. On the one hand, it can be seen as an upper bound to the true effect because the reduction of taxes may have mitigated potential negative effects of the minimum wage increase. On the other hand, it may be biased downwards, since we are only estimating the marginal effects of the additional increase on the border, and not the total effect of the increase in the minimum wage.  The effect of the minimum wage on poverty 371 Nevertheless, this elasticity is similar to what has been found in previous studies of Brazil, Honduras, and Nicaragua (Sotomayor, 2021;Gindling & Terrell, 2010;Alaniz et al., 2011, respectively), and markedly lower than that found in Russia and the U.S. (Kapelyuk, 2015;Dube, 2019, respectively).

How was poverty reduced?
We want to understand how poverty incidence was reduced on the border, since the increase in the minimum wage may have affected the flows into and out of poverty. Following Alaniz et al. (2011), Kapelyuk (2015, and Neumark and Wascher (2010), we calculate the probability that a household is poor in period t, given that it was already poor in period t À 1, as well as the probability that a household is poor in period t, given that it was not poor in period t À 1. The first exercise provides information on what is called poverty persistence or the probability of staying in or moving out of poverty, while the second one refers to the probability of falling into poverty. The ENOE is a rotating panel in which a household is followed for five quarters. This design allows us to calculate short-run poverty flows from one quarter in year t À 1 to the same quarter in year t. Empirical evidence so far has shown mixed results. For example, Alaniz et al. (2011) find that an increase in the minimum wage in Nicaragua mainly affected the probability of moving out of poverty. Neumark and Wascher (2010) also find an increase in this probability, but it was outweighed by an increase in the probability of falling into poverty. Table 4 shows the results for poverty persistence and for the probability of falling into poverty. The minimum wage increase did not affect the probability of staying in poverty. Although the estimates are positive, they are highly imprecise. However, the minimum wage increase seems to have influenced the flow into poverty by the non-poor. Using the hotdeck sample, the probability of becoming poor declined between 1.9 and 2.8 percentage points. As the base probability is 23 per cent, the effect in relative terms is 8-12 per cent. The evidence based on these transition probabilities suggests that the minimum wage increase benefitted those above the poverty line with labor market attachments more than those at the bottom of the distribution. Table 4. Effects on the probability of poverty in period t given that the household is poor or non-poor in t À 1 Pr(poor in t j poor in t À 1) Pr(poor in t j non-poor in t À 1) Notes: Authors' calculations. The regressions report the marginal effect from a probit. Robust and clustered standard errors in brackets (at the state x border level). Panel A uses the households in the CONEVAL sample to calculate poverty, and panel B uses the sample obtained from the hotdeck procedure. Regressions restrict to households either poor or non-poor in period t À 1. Regression controls for year-quarter fixed effects, state x border fixed effects (both in period t), and for rural status, family size, number of members less than 15 years old, number of members over 65 years old, and the age, years of schooling, gender, and marital status (married or living together) of the head of household in period t À 1.

Rest of the country
Of course, this result should be seen in the context of the pandemic and its negative economic impacts. As already noted, one outcome associated with the pandemic was a generalized increase in poverty levels in Mexico (see Table 1). Therefore, another way to interpret these results is to say that the increase in the minimum wage strengthened the resilience of those above the poverty line and helped them not to fall into poverty as much as they did in other parts of the country. 12 To analyze how the increase in the minimum wage affected different quantiles of the distribution of per capita household income, we estimate an unconditional quantile regression (Firpo, Fortin, & Lemieux, 2009), as in Dube (2019) andDel Carpio et al. (2019). This allows us to corroborate which households are the most affected by the increase in the minimum wage. Figure 4 presents these results. Below quantile 10, there is no effect, or it is noisy due to households with zero income. However, for most quantiles below the median there is hardly any effect. The increase in the minimum wage had a positive and significant effect on household incomes only in the upper half of the distribution. The figure is also consistent with the estimates provided in Table 4. The flow into poverty decreased because labor income increased for non-poor households, and poverty persistence did not decrease, due to a null effect of the policy at the bottom of the distribution. 13 A key question arises as to why the minimum wage does not affect the bottom part of the distribution. As shown in the supplementary materials ( Figure S5), formal sector attachment in low-income households is low. Table 1 shows that 6-8 per cent of households have no labor income, and other poor households have no formal sector income. Unlike Argentina, Bolivia, Brazil, Colombia, Paraguay, Peru, and Uruguay (Organisation for Economic Co-operation and Development [OECD], 2021), Mexico has a large percentage of informal workers below the poverty line, as a share of the total number of informal workers in the economy. The minimum wage thus has no effect at the bottom of the distribution.
As to why the effects seem persistent in the upper half of the distribution, we offer two arguments. First, the minimum wage in Mexico and in other Latin American countries may work as a lighthouse and reference for higher wages in the economy (Campos-Vazquez & Esquivel, 2020;Lemos, 2009). In this case, minimum wage workers are not the only ones affected: higher earners also receive an increase. As there are more formal sector workers in higher-income than The effect of the minimum wage on poverty 373 in low-income households, the increase in the minimum wage may benefit them more. We rule out the possibility that the positive effect is a result of the substitution of low-income with highincome workers, as a previous study, using administrative data, found that this did not occur (Campos-Vazquez & Esquivel, 2020). Second, as mentioned in the introduction, the minimum wage increase was accompanied by a decrease in taxes (mainly VAT). It is plausible, then, that in addition to the lighthouse effect, the policy provided a greater benefit to households in the upper half of the distribution, especially those with a greater attachment to the formal sector.
Determining which channel is more important for the upper half of the distribution is a task for future research.

Effects on poverty intensity
So far, we have shown that the increase in the minimum wage reduced poverty on Mexico's northern border. However, we also want to know whether it had an effect on poverty intensity: whether the minimum wage increase affected the marginally poor or the poorest. We now calculate the indicator known as the poverty gap and its square, and we perform similar regressions as before. To avoid the confounding effect of the decrease in poverty incidence in these calculations, we estimate our regressions using information only for households already in poverty. Table 5 shows the main results for the different samples and comparison groups. The DID estimates imply an increase in the intensity of poverty due to the minimum wage increase. The effects range from 0.19 to 0.30 for the poverty gap and from 0.25 to 0.39 for its square. In all cases the estimates are statistically significant at conventional levels. Using the average value for the poverty gap and its square in 2018, the DID estimates imply an increase from 5 to 10 per cent in poverty intensity. Table 6 uses the synthetic control method to calculate the effects on both the poverty gap and its square. Although the average effects are similar to the DID estimates, the results are less precise and in general they are not statistically significant, but in both estimations poverty intensity and its square both decreased.
However, if we perform the regression analysis only on poor households with positive labor income, there seems to be no effect on poverty intensity (  Material). It thus seems that the increase in poverty intensity is due to families with no labor income. If the minimum wage does not increase labor market participation, then it is irrelevant for poor non-working families (Card & Krueger, 1995;Burkhauser & Sabia, 2007). This suggests that other public policies are needed to address poverty among those households without labor market attachments.

Results obtained by adjusting the poverty line in each region
All of the previous results were estimated using the same poverty line for the northern border and for other regions. However, in addition to the substantial increase in the minimum wage there was also a reduction in VAT on the border, from 16 to 8 per cent, which could have led to a pass-through effect of lower inflation (Mariscal & Werner, 2018). Alternatively, the increase in the minimum wage may have increased inflation as producers passed higher labor costs through to final prices (MaCurdy, 2015). Since these two effects act in opposite directions, it is an empirical matter whether these policies affected inflation or price levels at the border. Previous evidence shows that the VAT effect dominated, as overall inflation decreased more on the border than in the rest of the country (Campos-Vazquez & Esquivel, 2020). This differential pattern in inflation may affect the monetary value of the consumption basket used to estimate the poverty line in each region. If inflation of the poverty line is substantially lower on the border, the effect on poverty may be underestimated (in absolute terms). On the other hand, since poverty lines are estimated using a food basket, and most food products do not incur VAT, it is possible that the reduction in VAT may not have significantly affected the poverty line on the border. 14 Table 7 shows the main results for poverty incidence and the poverty gap and its square, adjusting the poverty line by region. The previous analysis indicated an average poverty reduction of around 2.6 percentage points in poverty incidence. Adjusting for a different poverty line in each region, the effect now is 3 percentage points, and using the synthetic control method it is 3.4 percentage points (Table S7 in the Supplementary Material). This implies that poverty was reduced by approximately 13-15 per cent, as opposed to the 11 per cent in the previous analysis, with a lower bound in the elasticity of poverty with respect to the minimum wage of À0.21. In terms of poverty intensity, the results are barely affected. Although overall inflation was lower on the northern border than in the rest of the country in 2019, it did not greatly affect the products included in the food basket to calculate the poverty line. Therefore, the most important effect from the adjustment in the poverty lines seems to be that the effect on poverty incidence is slightly greater (an increase from 2.6 to 3 percentage points).

Conclusions
We analyzed the effect of a significant increase in the minimum wage on poverty in Mexico using a quasi-experimental situation. We exploit the regional variation introduced by a 2019 differential increase of the minimum wage in Mexico in two regions. Our results show that poverty along the northern border declined between 2.6 and 3 percentage points due to the larger increase in the minimum wage in that region. These results are robust to alternative empirical methodologies (difference-in-difference and synthetic control), different samples, and different control groups. These results imply an elasticity of poverty to the minimum wage that ranges between À0.18 and À0.21, which is comparable to what has been found in previous studies. Our analysis also sheds light on the mechanism that allowed the reduction in poverty along the border: poverty was reduced mainly by reducing the flow from non-poverty into poverty; poverty persistence was not affected in a significant way. Considering that these results were obtained in a period in which the economy was severely hit by a pandemic, it is not surprising that the mechanism that prevailed was the increased resilience for those households with incomes above the poverty line, which are those with a greater formal sector attachment. It is likely that this result is driven by the specific context in which it was obtained.
An additional interesting result is that poverty intensity increased on Mexico's northern border despite the reduction in poverty. This situation occurred because the increase in the minimum wage did not affect the share of families without labor income or increase labor income among the poorest households. Therefore, this result provides some support to the concerns expressed by the early critics of the minimum wage as an antipoverty tool. Finally, some caveats are in order. While our results are robust in different dimensions, it is likely that our estimates are biased in one or another direction. The increase in the minimum wage was accompanied by some tax incentives, and it is possible that these helped to mitigate the negative effects on employment of the increased minimum wage. These tax incentives may have provided greater benefit to households in the upper half of the distribution, where there is greater formal sector attachment. In this sense, our results could have led to an overestimation of the absolute effect on poverty. Based on previous empirical evidence, however, we believe that this effect is relatively small. On the other hand, since the analysis is based on a differential increase in the minimum wage in two regions, and the increase in real terms for the control group was also substantial (close to 30 per cent), it is possible that an important share of the effect on poverty of the minimum wage had been captured by the control group, therefore leading to an underestimation of the effect. In general, however, we believe that the quasi-experimental nature of our study provides a useful setting for analyzing the effects of the minimum wage on poverty.
before the request and generate at least 90 per cent of its income within the region. Firms must also demonstrate no involvement with those issuing false invoices. Hence, less than 3 per cent of firms qualified for the program . 4. Campos-Vazquez et al. (2020) find a small negative short-run effect on employment from the northern border program, though it is statistically insignificant. Bouchot (2018) and Campos-Vazquez, Esquivel, and Santillan (2017) use the equalization of minimum wage zones in 2012 as a natural experiment and find no effect on employment (either formal or informal). See Bosch and Manacorda (2010) for an analysis of the decline in the real value of the minimum wage in the 1990s and its effect on inequality. 5. Even if there are no employment effects, it is not clear that the program can reduce poverty. First, the minimum wage is compulsory only in the formal sector of the economy. In Mexico, around 37 per cent of workers have social security benefits and among wage workers only 55 per cent were formal in 2015-2018. Second, the cut in VAT may not stimulate the local economy, or it may provide relatively more benefit to producers than consumers. Third, inflation from the increase in the minimum wage could erode some of the gains in nominal income. However, Campos-Vazquez and Esquivel (2020) find that the combined measures reduce inflation, which suggests that the effect on poverty may be slightly understated if poverty lines are not adjusted for this effect. 6. For example, some products affect only the urban food basket (like some meat and dairy products) or the rural food basket (like grain corn, fresh milk, and other dairy products). The quantities of products in each basket also vary. For instance, the rural basket includes 218 g of tortillas per day and the urban basket only 155 g. 7. For instance, Campos-Vazquez and Esquivel (2020) find small negative effects on fruits, vegetables, and livestock products (around À0.4 per cent). 8. The consumer price index is obtained from prices in 55 cities. Five of these cities are on the northern border: Tijuana and Mexicali, Baja California; Ciudad Ju arez, Chihuahua; Ciudad Acuña, Coahuila; and Matamoros, Tamaulipas. INEGI calculates price indexes for generic products at the national level using non-public data on the weights of each product across cities. The only public data are the weights for each city or for each generic product, but not their interaction. We construct poverty lines with the public data for the period of analysis, 2016-2020. That is, we use generic products for each city and then aggregate them into regions using the corresponding weight in the consumer price index for each city. 9. CONEVAL (2019) explains how the poverty lines are constructed, and it publishes the poverty lines each month at https://www.coneval.org.mx/Medicion/MP/Paginas/Lineas-de-bienestar-y-canasta-basica.aspx. 10. The poverty gap is just the difference between the poverty line and per capita household income over the poverty line. The regression is estimated conditional on being poor as to avoid the confounding effect of poverty incidence. 11. In the Supplementary Material (Table S2 and Figure S2) we describe the selected model with the synthetic control method for each sample. We also plot the effect from the best five models out of the 22 models we estimate. The range of effects overlaps the multivariate regression results and the main synthetic control results. There is clear evidence in these models that the minimum wage increase reduced poverty. The Supplementary Material also shows that the difference-in-difference estimates are robust to excluding 2020 from the main sample. 12. Table S3 in the Supplementary Material includes the results using the synthetic control method. Tables S5, S6 and S7 adjust the poverty line for different prices across regions. Results are similar to those shown here. The synthetic control method indicates that the effect of the program on the probability of falling into poverty is primarily from the first quarters of 2019, which implies either a quick fading away of the impact (Sotomayor 2021), greater noise due to the pandemic, or more uncertainty due to lower sample sizes. There are no estimates in which the results for poverty persistence are affected. 13. Figure S3 in the Supplementary Material shows that results restricting the sample up to 2019 are similar to Figure 4. Moreover, Figures S4 and S5 show that there is no effect likely due to low formal sector coverage at the bottom of the distribution. 14. Figure S6 in the Supplementary Materials plots the evolution of the poverty line in each region using the third quarter of 2018 as a base period. After the start of the program in 2019, the inflation of the poverty line is slightly less on the border (in the second and third quarters for the rural areas and in the second quarter for the urban areas), but in general it shows similar inflation patterns in each region with respect to the base period. Before the start of the program, however, inflation of the poverty line was greater in the rest of the country than in the border region. In any case, effects coming from differentiated price or inflation patterns seem to be rather small.