Do Elections Affect Immigration? Evidence From French Municipalities

IZA DP No. 15849 JANUARY 2023 Do Elections Affect Immigration? Evidence From French Municipalities* Using thirty years of municipal elections in France, we show that election results affect the share of immigrants across municipalities. In municipalities where a leftinstead of rightwing mayor has been elected, the share of immigrants in the population grows faster by 1.5 p.p. within six years after the elections, and by 3 p.p. within twelve years. To a large extent, these effects are driven by partisan differences in public housing constructions and changes in the composition of the population within existing public housing units. They also are associated with greater incumbency advantage, in line with a model of strategic partisan behavior. JEL Classification: D72, H4, H7, R38


Introduction
While many developed countries are currently facing an increase in both ethnic (Bickford and Massey, 1991;Cutler et al., 2008) and partisan segregation (Bishop, 2009;Brown and Enos, 2021;Brown et al., 2022), these two dynamics are thought to stem from distinct mechanisms. Ethnic segregation is often explained by social interaction models (Schelling, 2006;Card et al., 2008), and partisan segregation, by Tiebout sorting where people 'vote with their feet' into jurisdictions o↵ering their preferred bundle of public goods (Banzhaf and Walsh, 2008). However, if local policies impact the attractiveness of local jurisdictions in a way that varies along both ethnic and partisan lines, both dimensions of segregation may feed o↵ of one another.
In this paper, using data from French municipalities over three decades , we provide evidence that elections influence the share of immigrants in the population and that the main channel through which these changes operate is local public housing policies.
The existence of a large public housing sector, with substantial turnover, and where municipalities can influence construction and admission policies is a major di↵erence between the US and many European or Asian countries. By design, public housing can be targeted to specific constituents and is inherently linked with a residence. Using detailed information on beneficiaries, we show that election results influence the construction of new units and the share of immigrants within both newly constructed and preexisting public housing units. To a large extent, these partisan di↵erences appear specific to municipalities that already had a high share of immigrants in 1982, and, as such, received the bulk of immigrant inflows during the following decades.
To motivate the analysis, we discuss the various mechanisms that might explain why mayors would implement partisan policies that attract or repel immigrants. We highlight that if voters are partisan, mayors might target immigrants if doing so reshapes the electorate directly or indirectly in a favorable direction. Relative to natives, immigrants are overwhelmingly in favor of left-wing parties (Martiniello, 2006;Simmons et al., 2018;Brouard and Tiberj, 2011) and more than half of immigrants are allowed to vote as they have acquired French citizenship or originate from the European Union. 1 In addition, 1 Immigrants are defined by following the o cial international definition, whereby an immigrant is a implementing specific immigration policies might mobilize native supporters or a↵ect their share in the population. If those e↵ects are large, targeting immigrants will help mayors get reelected and even build political strongholds in the long run.
To identify the e↵ects of elections on local policies and the composition of the population, we follow Ferreira and Gyourko (2009) and use a regression discontinuity analysis (RDD) that exploits quasi-random variations in the political a liation of the mayor as a consequence of close elections. Our empirical investigation draws from a sample of 829 municipalities with more than 9,000 inhabitants in urban areas with more than 30 000 inhabitants. We observe these municipalities over six local elections that occurred approximately every six years from 1983 to 2014. Thus, our data allow us to investigate the e↵ects of elections on the composition of the population in the year of the next election, after 6 years, and also in the longer run, from 12 up to 18 years after the initial election, which corresponds to the time at the second and third election after the initial election, respectively.
We find that, after six years (by the time of the next election), the share of immigrants in the population has grown 1.5 p.p. faster in municipalities where a left-relative to a right-wing mayor has won. This e↵ect is large, as it corresponds to two-thirds of the average change in the share of immigrants between two elections in our sample of municipalities. We can rule out that this increase reflects general socioeconomic changes in the population, as the share of immigrants increases across all occupation groups. New residents with immigrant origins have socioeconomic characteristics similar to those of the natives they replace, even if immigrants are less likely to be in high-income occupations.
We also find that these di↵erences in the share of immigrants in the municipality increase over time, at least over two election cycles, albeit the estimates are more imprecise in the long run.
We then investigate how this large demographic e↵ect might translate into the political leaning of the electorate. The direct short-run e↵ect on the electorate is initially attenuated by the fact that the increase in immigration is only driven by non-European immigrants, half of whom cannot vote as they are not citizens. However, the share of immigrants who can vote increases over time, and most of the long-run e↵ects are driven foreign-born person who was not a French citizen at birth. by immigrants with voting rights. This result is suggestive of a direct impact of local elections on partisan segregation through immigration.
We then explore which systematic policy di↵erences might explain the e↵ects of elections on the share of immigrants across municipalities. Consistent with Ferreira and Gyourko (2009), we find small and imprecise e↵ects of elections on the two sources of municipal funding: the housing tax and the property tax. This rules out the hypothesis that di↵erences in immigrant inflows might be driven by a di↵erential increase in spending or taxes after the elections. In contrast, di↵erences in public housing construction and inflows of immigrants into existing and newly constructed municipal public housing explain much of the increase in the share of immigrants.
In practice, the election of a left-wing mayor is associated with a 2.0 p.p. larger increase in the share of households living in public housing in the short run, 6 years after the elections, and e↵ects after 12 years are, once again, twice as large. Overall, units constructed after the elections account for a third of the increase in the share of immigrants in public housing in the municipality in the short run. On top of this extensive margin, the election of a left-wing mayor is also associated with larger inflows of immigrants into existing public housing. Even if public housing only accounts, on average, for 25% of total housing in our sample, the increase in the share of immigrants who live in public housing accounts for half of the total e↵ect of elections on the share of immigrants in the population.
To understand the mechanism underlying our results, we investigate whether the effects of elections vary with municipalities' initial attractiveness to immigrants. Indeed, opening more public housing to immigrants would be useless in towns where few immigrants are likely to move. Using the fact that immigrants tend to locate in places where the share of immigrants is already high (see, e.g., Bartel, 1989;Jaeger et al., 2018), we proxy for a municipality's attractiveness to immigrants using its share of immigrants in the population in 1982 (12%), which is strongly correlated with future immigrant inflows.
Separate estimates for municipalities above and below this median show that our results are driven by municipalities that were already attractive to immigrants.
Finally, using the fact that the results are driven by municipalities that were more likely to attract immigrants, we assess whether these demographic changes are related to the incumbency advantage, defined here as the future electoral margin of the left following the close election of a left-wing mayor. We find substantial di↵erences in incumbency advantage depending on the initial share of immigrants. When the initial share of immigrants is high, the incumbency advantage is equal to 10 p.p. at the next election and 13 p.p. at the next-but-one election. In contrast, in municipalities with a low initial share of immigrants and for which we find little evidence of partisanship of local policies, the incumbency advantage is smaller and not statistically significant after one election. While these results are imprecise and should be interpreted with caution, they are compatible with a model in which mayors facing immigrant inflows are more likely to secure future electoral wins through partisan immigration policies.
Related literature -Our work first contributes to a large literature on the determinants of local policies. Building on Tiebout's (1956) seminal work, many theoretical analyses have emphasized that residential mobility and competition across local jurisdictions should prevent partisan behavior at the local level (Epple and Zelenitz, 1981;Calabrese et al., 2011). In contrast, recent empirical studies find substantial evidence of partisanship of local policies (Pettersson-Lidbom, 2008;Gerber and Hopkins, 2011;Meyersson, 2014;Dippel, 2019). In particular, our evidence of partisan di↵erences in local housing policy is closely related to Solé-Ollé and Viladecans-Marsal (2013), who find that left-wing parties in Spanish municipalities convert less land from rural to urban uses. Our findings suggest that, in France, left-wing municipalities tend to increase housing supply through public housing while we find no e↵ect on private housing. 2 The second strand of the literature to which this paper relates is the e↵ects of public housing on immigrants' location choice (see, among others, Verdugo, 2016). In many European and Asian countries, the public housing sector remains a key component of the housing market (Scanlon et al., 2014;Battiston et al., 2014;Chen et al., 2013;Chiu, 2013;Xu and Zhou, 2019) and recent European studies have shown that the concentration of immigrants in public housing has influenced their spatial segregation (Quillian and 2 Spain has very little public housing compared to France (2.5% against 20%), which explains why Solé-Ollé and Viladecans-Marsal (2013) do not explore this policy. Lagrange, 2016;Verdugo and Toma, 2018;Musterd and Deurloo, 1997). Our results are consistent with many qualitative studies that have described the discrimination of ethnic minorities by public housing programs across Europe. 3 Finally, this paper is also related to the growing literature exploring the relationship between immigration and elections. While a large body of work has shown that a local increase in the share of immigrants increases support for anti-immigration parties, 4 we focus on the reverse mechanism, whereby local election results a↵ect immigration inflows.
Our results are consistent with Bracco et al. (2018), who show that the election of a Lega Nord mayor in Italy discourages immigrants from moving into the municipality, and Slotwinski and Stutzer (2019), who highlight how local results of an anti-minaret referendum a↵ected the location choice of foreigners in Switzerland. We complement those two papers by providing evidence that, in a context where public housing allows mayors to target immigrants, local partisan policies explain part of the e↵ects of elections on the location choice immigrants.
The rest of the paper is organized as follows. To guide the empirical analysis, Section 1 discusses the theoretical conditions under which local elections may a↵ect immigration.
In Section 2, we describe the context of the study and the data. In Section 3, we detail our empirical strategy. Section 4 presents the results and Section 5 concludes.

Conceptual framework
Several mechanisms could explain why local elections influence the share of immigrants in a local jurisdiction. Firstly, election results might impact immigration because, independently of the policies implemented, they reveal the population's attitudes and indicate a more or less welcoming environment for immigrants (Bracco et al., 2018;Slotwinski and Stutzer, 2019). Such di↵erences in attitudes might discourage immigrants from settling in municipalities where the population is hostile. However, if elections merely play a signaling role, we should not observe a strong discontinuity following a close election associated with the victory of a given party.
If a discontinuity is observed as in our case, the changes in population are likely driven by systematic partisan di↵erences in local policies. Admittedly, these policies need not be targeted toward immigrants to have a larger e↵ect on their location choice. Borjas (1999) argues that in contrast to natives, immigrants might be more responsive to local policy di↵erences as they have already paid the fixed cost of migration into the host country and might not have strong local ties. This hypothesis implies, for example, that a policy change preferred by low-income residents from all origins might induce more substantial inflows of low-income immigrants relative to natives. However, if some low-income natives are mobile, we should observe at least some of them respond to the policy, which is not the case in our data.
In contrast to most local policies, a large literature discussed below suggests that the municipal provision of public housing units, for which we report strong evidence of partisanship, is targetable to specific subgroups in the population. In addition, immigrants (and, more specifically, non-European immigrants) are particularly attracted to public housing, to a large extent in response to discrimination in private housing (Acolin et al., 2016;Verdugo, 2016;Combes et al., 2018). As we document below, over the period, the share of non-European immigrant households among public housing inhabitants increased dramatically, reflecting both an increase in their share in the population and an increase in their participation in public housing. Therefore, mayors may try to influence the location of immigrants through partisan public housing policies.
Why would mayors want to attract or repel immigrants? If we abstract from ideological motives, a plausible explanation might involve electoral incentives. First, immigrants, and Non-European immigrants in particular, are overwhelmingly supportive of left-wing parties: recent survey data from Brouard and Tiberj (2011) reproduced in Figure S1 in the Supplementary Appendix show that they are three times more likely than natives to prefer the left. Unlike natives, their political preferences do not vary much with their occupation. Second, natives' attitudes toward immigrants are strongly partisan. Figure S2 in the Supplementary Appendix, reproduced from Piketty (2020), indicates that voters who believe that there are too many immigrants in France are 30 p.p. more likely to vote for the right. In contrast, voters favorable to immigration lean toward the left (see, also, Gethin et al., 2021). As a result, targeting immigrants might be used to indirectly a↵ect votes from non-immigrant voters.
The first possible mechanism for these indirect e↵ects on non-immigrant voters would be ideological signaling from the mayor to her supporters. As explained by , targeting immigrants may be used to send messages to native supporters and increase their electoral mobilization. For example, if a mayor knows that her supporters are hostile to immigrants, she may opt for an immigrant-hostile policy in order to increase their electoral mobilization and improve her reelection chances. A second related mechanism would be that local partisan policies change the composition of the local population.
In a seminal paper, Glaeser and Shleifer (2005) describe several instances where mayors of prominent US cities purposefully favored the group of voters from whom they enjoyed innate appeal (in their case, along ethnic lines) in order to push other groups to leave the city and secure a loyal electorate as a result.
In the Supplementary Appendix Section S1, we provide an extension of Glaeser and Shleifer's (2005) model adapted to the French situation where the targeted group, immigrants, will never account for the majority of the electorate. In such a case, targeting immigrants may still prove electorally rewarding if partisanship within the native population is unobservable to the mayor, but correlated with the propensity to tolerate immigrant neighbors. For example, suppose the supporters of the mayor are immigrant-friendly.
The mayor may then opt for an immigrant-friendly policy that increases not only the share of immigrants but also the share of partisan natives in the population. Indeed, a large literature has shown that some groups of natives may be quite sensitive to changes in the share of immigrants in their neighborhood (Saiz and Wachter, 2011). Concerns for such compositional amenities appear even more important toward immigrants from poorer countries, who are more likely to live in public housing (Card et al., 2012).
Close elections, which we use to identify a causal e↵ect, may be particularly relevant to study mayors' incentives to adopt partisan policies. Indeed, a close election signals lower reelection prospects, which should encourage mayors to announce a policy that will help them secure reelection. However, it may still be the case that the e↵ect of a local policy change is small compared to other determinants of immigrants' location choices.
In particular, immigrants may be quite unlikely to come and settle in municipalities that stand outside their traditional choice set. If mayors are aware of this impediment, they may not implement partisan immigration policies in municipalities that immigrants deem unattractive in the first place.
Overall, even though we will not be able to quantify the respective role played by each of the channels listed above, we can still test the following four predictions from models in which immigrants are strategically targeted: (i) election results will impact the magnitude of immigrant inflows into (and/or outflows from) the municipality; (ii) election results will trigger a shift toward local policies that can be targeted toward immigrants; however, (iii) the e↵ect of elections may be restricted to municipalities that are attractive enough to immigrants; and (iv) in the municipalities where the e↵ects of elections are observed, mayors will enjoy a greater incumbency advantage.

Context and data
We combine several French data sources to investigate whether municipal elections a↵ect the composition of the population and which systematic di↵erences in municipal policies can explain these e↵ects. We describe here the electoral setting and the data that we use.

Municipal elections
Municipal elections in France are held every 6 years to elect the municipal council that will elect the mayor. 5 The elections are based on lists, and voters can pick at most one list. If no list achieves an absolute majority in the first round, a second round is organized.
Only lists obtaining more than 10% of the vote in the first round can compete in this second round.
After the elections, the seats in the municipal council are allocated using a majority premium that gives a large absolute majority to the list that obtained the most votes in the final round. The electoral system stipulates that half of the seats are attributed to the list that obtained the most votes. The other half is distributed according to their vote share across all competing lists, including the list that obtained the most votes. Such an electoral system generates a large discontinuity between vote shares and the share of seats in the municipal council. For example, a list just above 50% of votes will get 75% of the seats. As a result, the list that obtained the most votes is guaranteed to govern the municipality, as no post-electoral coalitions among other lists could reach a majority and overturn the results. We exploit this feature in our RDD framework.
We use electoral data collected by the Center of Socio-Political Data of Sciences Po from electoral authorities. We exploit five elections that occurred in 1983, 1995, 2001, and 2008 in municipalities with more than 9 000 inhabitants in the election year. 6 Even if their administrative organization is di↵erent, the 20 districts ('arrondissements') of Paris and the eight and seven electoral sectors of Marseilles and Lyon are also included in the sample and considered as separate municipalities. Nevertheless, while districts control the allocation of public housing, we cannot examine the e↵ects of elections on local taxes as districts are not fiscally independent from each other, and we do not have specific information on their budget. In the robustness section, we show that our main results are virtually identical when these three municipalities are excluded from the sample.
We exclude from the sample rural municipalities and municipalities in small urban areas with less than 30 000 inhabitants, because the population changes that we are interested in are more likely to occur within a local labor market. 7 Using Census data, Table S1 in the Supplementary Appendix shows that this sample of municipalities includes more than 40% of the French population in the recent decades, but more than 60% of immigrants and public housing units. As a result, in 2014, the share of immigrants in the population of municipalities of our sample was 14.3% against only 5.2% in other mu-6 See Appendix A for details on the data. The results for previous elections were not systematically collected. Results for smaller municipalities were not collected before 2008. We do not include the 1989 elections in the sample because the closest subsequent census, which we use to assess changes in the composition of the population, took place in 1990, which is too early, and 1999, which is too late, as the 1995 elections occurred between the two. Table 12 shows that including the 1989 elections matched with the 1999 outcomes does not qualitatively change the results. 7 These municipalities represent only 20% of our electoral sample. We show in Table 12 that the results are broadly similar, but more imprecise when these municipalities are included.  Source: CDSP election data. Note: The table reports descriptive statistics of municipal elections from our estimation sample. The sample is restricted to municipalities with more than 9,000 inhabitants in urban areas with more than 30 000 inhabitants. We include elections with at least two lists in competition, and with a list from the Left and the Right ranked first and second. Column 2 shows the share of elections in which a second round occurred. Column 3 shows the share of elections with strictly more than two lists in competition. Column 4 shows the share of elections followed by a victory of the Left. Columns 5 to 7 document the share of elections with a margin of victory inferior to 10, 5, and 2.5 p.p. respectively. The margin of victory is defined as the di↵erence between the share of the vote between the lists ranked first and second.
nicipalities. Similarly, 22.2% of inhabitants lived in public housing in these municipalities against only 8.6% in other municipalities.
We categorize the lists in the final election round as either left-or right-wing using the classifications of French electoral authorities. Most of the time, this assignment is straightforward, as there were no national or local bipartisan coalitions over the period. 8 Table 1 summarizes the characteristics of municipal elections in our sample. While the sample is unbalanced as the electoral data was only collected for municipalities with a population above 9 000 inhabitants, the number of municipalities is quite similar across the elections, ranging from 608 to 805. Column 2 reveals that a second round occurs in 40 to 60% of elections, except in 1983 when electoral competition mainly involved two competing coalitions (Dupoirier et al., 1985). Column 3 shows that more than two lists compete in the final round in more than 50% of elections. Column 4 shows that the share of left-wing victories is below 50% in most elections, except in 2008.
To assess how the proportion of close elections that we use to identify a causal e↵ect varies over time, Columns 5 to 7 of Table 1 report the share of elections with margins of victory inferior to 10, 5, and 2.5 p.p. Close elections are frequent: the margin of victory is lower than 2.5 p.p. between 5% to 8% of elections. The share of close elections is quite similar across electoral years, albeit slightly higher in 1995.

Immigration
We assess the e↵ects of elections on the population with the 1975, 1982, 1990, 1999, 2008, and 2014 French censuses. While the ideal situation would be to observe the municipalities just before each election, in practice, the census years fall very close to the election years. We have access to individual-level data with a sampling rate of 25%, which allows us to precisely measure population changes, even in small municipalities. Municipalities are identified using a statistical identifier that is constant over time, and changes in municipal boundaries are extremely rare. 9 As is the case for most local authorities, mayors cannot use gerrymandering to change the border of municipalities.
To mitigate the possibility that di↵erences in household size might influence our measures, we use the head-of-household as the unit of analysis. We estimate all characteristics of municipalities, such as the share of immigrants or public housing inhabitants, using the head-of-household level. 10 Information on the country of birth and citizenship, distinguishing between native and naturalized citizens, is reported in the Census. However, there is no information on the ethnicity of descendants of immigrants. The Census data allow us to identify immigrants who can vote in local elections. This group includes immigrants with French citizenship and, since 1992, all immigrants from the European Union. We categorize a household as an immigrant if the head of household is a foreign-born person who was not a French citizen at birth.
9 Only three small municipalities of our sample merge with another municipality over the period and they are excluded from the sample after their merge.
10 Using the head-of-household is a standard practice to take into account the fact that immigrants might have native-born children, which would lower their share of the population (Bajari and Kahn, 2005;Cutler et al., 2008). Results using instead all the adult population are reported in Table 12 and are virtually identical.  : 1982, 1990, 1999, 2008 and 2014 census. Notes: Municipalities with more than 9,000 inhabitants in urban areas with more than 30 000 inhabitants. Panel A shows the share of immigrants and non-European immigrants in the population. Panel B shows the share of each occupation group in the population, interacted with national origin. Panel C shows the average rates of the property and housing tax rates. Panel D shows the probability of living in public housing for each group. Panel E reports di↵erences in turnover rates with respect to housing occupancy status. Turnover rates are defined as the share of current occupants arrived in their dwelling less than six years before. See Appendix A for details on their computation. Column 4 shows the average change of the group relative to the initial population of the municipality. Column 5 shows the associated standard deviation. All variables are measured at the household level.
In Table 2, we describe the evolution of immigration in the municipalities that form our sample over the 1982-2014 period. As shown in Panel A, between 1982 and 2014, the share of immigrants in the sample increased by 3.6 p.p. from 10.7% to 14.3%. However, while some municipalities received large inflows, immigrants never make up the majority of households. 11 An important characteristic of immigration over that period is that most of the increase in immigration was driven by non-European immigrants, whose share in the population more than doubled, from 5% to 12% at the end of the period.
Given that the census does not report any measure of income, we use occupations to capture changes in the socioeconomic composition of the municipality after the elections. We divide the population into three broad groups depending on the occupation of the head of household using the one-digit level of the French occupational classification: high-income occupations, which include highly-qualified workers such as managers and professionals, low-income occupations, and retirees. Panel B in Table 2 reports the evolution of the composition of the population across these three occupation groups interacted with immigration status. Over the period, the share of both high-income occupations and retirees increased dramatically. At the same time, while the proportion of immigrants increased across all occupation groups, their share increased more rapidly among the low-income group.

Municipal policies
We focus on two main measures of municipal policies that can be observed in available data sets.
Taxes -We use data from the Recensement deséléments d'imposition, an administrative database collected by the French Ministry of Finance that contains the total amount collected and the tax base for each municipality and tax from 1982 to 2014. A limitation of this dataset is that it does not contain any information on how the budget is spent. We assess whether elections a↵ect the rate of the two main municipal taxes levied on households: the housing tax and the property tax, both based on net rental value. Consistent with Charlot et al. (2008), Panel C in Table 2 shows that these two tax rates increased regularly over the period.
Public housing -The second crucial local policy that we examine concerns municipal public housing. The census reports whether a dwelling belongs to the public or the private housing sector, which allows us to investigate if elections influence the supply of public housing or the composition of the population in public relative to private housing. The census also reports information on each housing unit's construction year, allowing us to isolate public housing units constructed after the elections. Unfortunately, the 1975 census does not contain information on public housing, which limits the scope of some of the placebo tests designed to assess the plausibility of our identification strategy.
When a municipality builds new public housing units, construction works do not directly a↵ect the municipal budget, as they are financed through state subsidies and longterm low-interest loans (Driant, 2011). Public housing is easily accessible to a large share of the population, as more than 55% of households are eligible, including documented immigrants. 12 After the construction, rent proceeds are used to reimburse the loan and manage the properties. Default risk is negligible: should a tenant become insolvent, rents would be directly paid for to local public housing authorities by social security in the form of housing benefits.
While eligibility rules are defined at the national level, public housing units are constructed, managed, and allocated by municipalities through local housing authorities -see Appendix B for details. Reflecting this discretionary power, the press frequently reports local examples of nepotism and corruption. 13 The 2000 SRU law -The contested law on "solidarité et renouvellement urbain" (SRU) illustrates the possible conflict between municipalities and higher levels of government over public housing. The law was implemented under socialist Prime Minister Lionel Jospin (1997)(1998)(1999)(2000)(2001)(2002) in December 2001. The law aims at inducing municipalities with more than 3500 inhabitants to construct public housing, in order to reach a share of 20% of public housing among all housing units (this share was raised to 25% in 2013 under socialist president François Hollande). 14 Municipalities below the 20% threshold that do not construct enough social housing have to pay fines proportional to their gap in percentages (Chapelle et al., 2022).
Such a law might have changed the behavior of municipalities as the incentives to construct public housing are no longer purely political but also financial. If anything, we thus expect to see less partisanship in public housing construction after its implemen- Characteristics of Public housing -Like many European and Asian countries, a large share of the population lives in public housing in France. Panel D of Table 2 shows that more than 21% of households live in public housing in our sample over the period.
The table also shows that Non-European immigrants are twice as likely as natives to live in public housing as more than 43% of them live in public housing. Large concentrations of immigrants in public housing are also observed in other Western-European countries (Musterd and Deurloo, 1997;Verdugo and Toma, 2018).
An important question is whether tenant mobility in public housing is su cient to influence the composition of the population between elections. Panel E of Table 2 reports adjusted residential turnover rates in public housing based on the share of households who arrived in their housing unit between two municipal elections (6 years). Consistent with Pan Ké Shon (2010), who also finds high levels of mobility in deprived public housing estates, turnover rates in public housing units are large, ranging between 41% and 57% between elections.

Public housing, immigration and politics
Public housing, immigration, and the interplay between the two have long been divisive topics in French public debate. 15 As shown in panel A of Figure 1, the election of a leftwing mayor is followed by substantially higher growth rates of the immigrant population.
This di↵erence increases over time, consistent with evidence that immigrants tend to locate in places where the share of immigrants is already high.
Regarding public housing, Foucault and Nadeau (2018) recently showed that the share of inhabitants in public housing is a strong predictor of the electoral trajectories of French municipalities, with a higher share of public housing being favorable to the left. In accordance, Panel B in Figure 1 shows that the construction of public housing, as measured by the increase in the share of households living in public housing, is higher after the election of a left-wing mayor. The di↵erence is larger after earlier elections, but it remains statistically significant until 2008, after the passing of the SRU law. The rest of the paper investigates whether part of the relationships depicted in Figure 1 are causal.

Empirical model
To assess the causal e↵ect of municipal elections, an ideal experiment would be to randomly assign a left-or a right-wing mayor to a sample of municipalities. We approximate such an experiment with a regression discontinuity design that compares municipalities in which a left-wing mayor was closely elected with municipalities that closely elected a right-wing mayor. Di↵erences between the two identify the local treatment e↵ect of electing a left-instead of a right-wing mayor after a close election.
15 Right-wing presidents such as Jacques Chirac (in 1991) or Nicolas Sarkozy (in 2005) were famous for their anti-immigrant discourse while visiting public housing. In contrast, newly-elected socialist president François Mitterrand voted a law in October 1981 granting immigrants the possibility to be represented in local public housing agencies.  where the left has won more votes relative to municipalities where the right has won more votes. Panel B: Di↵erence in the growth rate (in %) of households living in public housing between municipalities where the left has won more votes relative to municipalities where the right has won more votes. 95% confidence intervals are depicted by the black segments. Sample of municipalities described in section 2. The match between election and census data is described in Appendix A. Sources: 1983Sources: , 1995Sources: , 2001Sources: , and 2008Sources: elections and 1982Sources: , 1990Sources: , 1999Sources: , 2008, and 2014 census.

The RDD design
Following Lee (2008), our running variable X it is the vote margin of the left. It is defined as the di↵erence in the share of votes between the better-ranked left-wing list and the better-ranked right-wing list in municipality i in election year t in the final round. It is thus positive when the left has won and negative otherwise. 16 With treatment defined as a left victory, the assignment variable D it is a dummy equal to one when the election was won by the left, and thus X it > 0, and zero otherwise. Our empirical model is given by: where g i⌧,t = (G i⌧ G it )/L it measures the growth in households from group G in municipality i between election year ⌧ > t and year t relative to the initial population of the municipality L it . As the changes in g i⌧,t are relative to the initial population L it , our dependent variable is not a↵ected by potentially endogenous changes in the total population.
Our parameter of interest is 1 which captures the local average treatment e↵ect (LATE) at the threshold X = 0 of electing a left-instead of a right-wing mayor after a close election. As discussed in Section 1, these local treatment e↵ects are of particular interest: a close election might send a signal to mayors that the reelection ahead will be di cult.
One limitation of g i⌧,t is that if the population of the group grows at the same rate as the population of the municipality, the share of the group in the municipality will remain unchanged. To take this issue into account, we also present results using changes in shares defined as s i⌧,t = G i⌧ /L i⌧ G it /L it that adjust the size of group G to the contemporaneous level of the population between elections.
Following Cattaneo et al. (2020), we non-parametrically estimate the model using a local linear function with a triangular kernel, and inference is based on their robustbias correction method. To allow for comparability across outcomes, we use in our main specification a bandwidth of 15% and a polynomial of order 2. Such bandwidth is close to 16 For municipalities with no left-right competition, we assign a zero share of votes to the missing party. In this case, the margin is thus equal to the share of votes of the list ranked first, with a positive sign for a left-wing list and a negative sign for a right-wing list. the optimal bandwidths proposed by Calonico et al. (2014) and estimated for the change in the share of immigrants or inhabitants in public housing. To assess the robustness of the results to the bandwidth choice, we also report results using smaller or larger bandwidth, in addition to results obtained with the optimal bandwidth. We also report results using alternative polynomial orders and global polynomials as robustness checks.

Validity checks
Whether our RDD estimates capture a causal e↵ect depends on whether municipalities close to each side of the threshold are comparable and had similar characteristics before the elections. 17 To assess this hypothesis, Table 3 follows Fack and Grenet (2015) by reporting several "placebo" RDD estimates that test whether the outcome of close elections is associated with past levels or changes in the characteristics of the municipality as measured before the elections of interest.
In Panel A, we consider di↵erences in the origins of the population using the share of European and non-European immigrants, the share of workers in low-income occupations interacted with national origins, the share of immigrants with voting rights, and the share of immigrants living in public housing. 18 For most outcomes, we find little correlation between these variables' past share or growth and the results of future close elections.
Two exceptions are the past share of European immigrants and immigrants in low-income occupations that appear to correlate with a future victory of the left. Reassuringly, the corresponding estimates using the past growth of these variables instead are statistically insignificant.
In Panel B, we consider the past population size and growth and alternative measures of the housing status of inhabitants through the share of homeowners and households in public housing, di↵erences in the rates of property and housing taxes, and the share of workers in the manufacturing sector. Overall, we find little correlation between the outcome of close elections and these variables' predetermined levels or changes except for the growth of homeowners and property tax rates, which are significant at the 10% level.  Sources : 1983, 1995, 2001 and 2008 municipal elections, 1981, 1988, 2002 and 2007 presidential elections and 1975, 1982, 1990, 1999 and 2008 census. Notes: The table shows placebo RDD estimates on the indicated predetermined outcomes. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth and a quadratic polynomial. The predetermined outcomes are defined using the change of the population from the group between elections divided by the initial population of the municipality in panels A1 and B1 or using the predetermined share of the group in the municipality population in panels A2 and B2. Outcomes are measured at the household level. Panel A and B show results for predetermined characteristics of the population using predetermined changes in panel 1 and levels in panel 2. Panel C uses as a dependent variable the characteristics of the municipal elections in the first three columns, the previous margin of the left in the fourth column, and the past results in the municipality for the most recent presidential election. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth and a quadratic polynomial. We    Sources: 1982Sources: , 1990Sources: , 1999Sources: , 2008 and 2014 censuses. Notes: The table reports of averages of the share of immigrants, public housing inhabitants, blue-collar workers and managers within municipalities in our sample. The first line reports these averages for the whole sample. Other lines use averages from elections with margin victories inferior in absolute value to 10%, 5% and 2.5%. The margin of victory is the difference between the share of votes between the lists ranked first and second in the final round. : 1983, 1995, 2001 and 2008 elections and 1982, 1990, 1999 and 2008 census. Notes: The table reports the average values of the share of immigrants, public housing inhabitants, low-income and highincome occupations in the population of municipalities in our sample. The first line reports these averages for the entire sample. Other lines report averages from elections with margins of victory lower than 10%, 5% and 2.5%. The margin of victory is the di↵erence between the share of votes between the list ranked first and the list ranked second in the final round. Column 1 reports the total population but the share of each group in other columns is measured at the household level.

Sources
In Panel C, we test whether the outcomes of the close elections are correlated with the characteristics of the same municipal elections, as measured by the share of registered voters, the abstention rate, and the number of lists in competition. We find no evidence that these characteristics are associated with the outcome of close elections. In addition, Column 4 shows little association between the outcome of close elections and the margin of the left in the previous municipal election.
If our identification hypothesis is valid, the outcome of close elections should also not be correlated with local political attitudes. To proxy these attitudes, we use the share of votes in the first round of past presidential elections for the far-right and leftwing candidates in columns 5 and 6, respectively. 19 Once again, we find no statistically significant correlation.
Following McCrary (2008) and Cattaneo et al. (2018), we investigate whether the forcing variable is continuous near the threshold, which is a test for the manipulation of the forcing variable. To check the smoothness of the density of the vote share near the threshold visually, Figure S3 in the Supplementary Appendix reports the histogram of the forcing variable. In addition, Figure S4 reports the outcomes of a local linear regression, estimated separately on each side of the cuto↵, to test the null hypothesis that the discontinuity at the cuto↵ is zero. While the share of close victories for the left appears slightly larger, the di↵erence is not statistically significant and we cannot reject the null of no density jump around the threshold (p-value = 0.18). These results are consistent with earlier evidence from Eggers et al. (2015) and Lippmann (2018) for France and other countries.
Finally, as the causal e↵ects are identified from close elections, an important question to interpret the results is whether close elections occur in municipalities that are systematically di↵erent from others. To investigate this issue,

Results
We first present results on how elections influence the composition of the population in terms of national origins and socioeconomic status. Then, to understand the mechanisms underlying the observed changes, we turn to the e↵ects of elections on municipal policies, focusing on local taxes and the construction and allocation of public housing. Finally, we estimate whether the e↵ects of elections depend on the initial share of immigrants and study the impact of elections on future electoral outcomes.

E↵ects on the population
Origins and socioeconomic status -In Table 5, we consider 'short-run' changes in the population, observed around the next election year, six years after the election of interest. We begin in Column 1 in Panel A by assessing whether elections influence the growth of the population in the municipality and decompose this growth in panel B and C  Sources : 1983, 1995, 2001, and 2008 elections and 1982, 1990, 1999, 2008, and 2014 census. The sample includes municipalities with more than 9,000 inhabitants in urban areas with more than 30 000 inhabitants. Notes: The table shows RDD estimates of the e↵ect of the victory of a left-relative to a right-wing mayor in the municipal election. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth and a quadratic polynomial. Panel A: in column 1, the dependent variable is the change in log total population. In columns 2 to 4, the dependent variable is the change in the indicated occupation group relative to the initial population (columns 2 to 4) Panel B: in column 1, the dependent variable is the change in immigrants relative to the initial population. In columns 2 to 4, the dependent variable is the change in households that are in the occupation group and are also immigrants Panel C: in column 1, the dependent variable is the change in natives relative to the initial population. In columns 2 to 4, the dependent variable is the change in households that are in the occupation group and are also natives. Changes are measured over a six-year period using the closest census to the municipal elections.
Outcomes are measured at the household level. Robust-bias corrected standard errors are reported in parenthesis. (*) and (**) denote statistical significance at 5% and 1% levels.
into the contribution of immigrants and natives, respectively. While the estimated e↵ects for natives are small and very imprecise, we find a large and statistically significant e↵ect of elections on the growth of immigrant households, in line with the first prediction laid out in Section 1. Our estimates indicate that six years after the elections, the number of immigrant households grows by 1.5 p.p. more rapidly in municipalities that closely elected a left-wing mayor than municipalities that closely elected a right-wing mayor.
Quantitatively, this e↵ect corresponds to more than 10% of the share of immigrants in our sample.
An important question is whether these e↵ects are specific to immigrants or reflect more general changes in the composition of the population along socioeconomic lines. As immigrants are more likely to work in low-income occupations, as suggested earlier in Table 2, these results might reflect that left-wing municipalities attract more low-income households from all origins, not specifically immigrants. To distinguish the e↵ects on economic status from the e↵ects of origin, other columns in Table 5 decompose the growth of the population into three broad occupation groups: high-and low-income occupations and retirees. We interact these groups with immigration status in panels B and C to assess the specific role of origins.
When we do not consider the national origin, as in Panel A, the coe cients are very imprecise, and there is no evidence that elections a↵ect any occupation group except for the high-income group, which increases in response to the election of a left-wing mayor.
In contrast, in Panel B where immigrant status and occupations are interacted, we find a higher growth of immigrants from each occupation group. This growth can be observed not only in the low-but also in the high-income occupation group, albeit to a lower extent.
For natives, Panel C suggests the e↵ects of a left-wing mayor on low-income natives is close to zero while there is some evidence of a decrease in the number of retirees and an increase in native workers in high-income occupations. However, the estimated coe cients for any subgroup of natives are not statistically significant. Overall, an essential lesson from Table   2 is that the e↵ects of elections on the population appear specific to immigrants and not associated with more general changes along socio-economic lines. In particular, we find no evidence that they are related to an increase in the share of low-income households in left-wing municipalities.
The role of inflows and outflows -For any group G, the growth g i⌧,t between elections ⌧ and t is given by the di↵erence between the inflow rate of new inhabitants from the group, denoted I ⌧,t /L t , and the outflow rate of former inhabitants, This simple decomposition implies that the larger growth of immigrants into left-wing municipalities can result from two mechanisms: first, larger inflows of immigrants into left-wing municipalities, reflecting that left-wing municipalities are more attractive to immigrants; second, larger outflows from right-wing municipalities, reflecting that more immigrants are leaving right-wing municipalities.
To assess their contribution, Panel A in Table 6 reports estimates of the e↵ects of elections on inflows and outflows of immigrants from the municipality. The results sug-  Sources : 1983, 1995, 2001, and 2008 elections and 1982, 1990, 1999, 2008, and 2014 census. Notes: The table shows RDD estimates of the e↵ect of the victory of a left-relative to a right-wing mayor in the municipal election. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth, and a quadratic polynomial. Column 1 uses as a dependent variable the growth of immigrants and natives relative to the initial population. This growth is decomposed by the inflow and outflow rates in columns 2 and 3. Column 4 uses as a dependent variable the changes in the share of immigrants (panel A) and natives (panel B) in the population. Robust-bias corrected standard errors are reported in parenthesis. (*) and (**) denote statistical significance at 5% and 1% level.
gest that most of the larger growth of immigrants in left-wing municipalities is explained by their larger inflows into these municipalities relative to right-wing municipalities. In contrast, there is little evidence of larger outflows of immigrants from right-wing municipalities. On the contrary, the estimated coe cient for outflows is positive, albeit not statistically significant, which indicates that immigrant outflows from left-wing municipalities also tend to be larger. In sum, our results are not explained by more immigrants leaving right-wing municipalities but instead by left-wing municipalities attracting larger inflows of immigrants.
Panel B reports the corresponding decomposition for natives. Overall, there is no strong evidence that elections lead to large native outflows from left-wing municipalities.
If anything, electing a left-wing mayor is associated with a small positive e↵ect on natives' inflow and a small negative e↵ect on natives' outflow. However, both coe cients are close to zero, and the estimates are very imprecise. As previously, most of the population responses are specific to immigrants.
Changes in population share -So far, we have defined our dependent variable using changes in the population from each group divided by the initial total population to circumvent the possible impact of overall population growth. However, the change in the relative share of a group in the population is also an important outcome. This change captures the overall changes in composition associated with the election. To investigate this issue, column 4 in Table 6 reports results using changes in the share of the group s y,i⌧ = G i⌧ /L i⌧ G it /L it instead of using the growth relative to the initial population. The estimated e↵ects are also statistically significant but lower by a third relative to column 1. These results suggest that larger overall population growth in leftwing municipalities attenuates the impact of immigrant inflows on their share in the population.
Long-run e↵ects -Are the e↵ects of elections on immigrants persistent over time? In practice, because of ethnic networks, immigrants locate where the share of immigrants is already high. If the initial e↵ects of elections are large enough, they may further increase municipalities' relative attractiveness to immigrants in the long run.
To investigate this issue, we report in panel A of Table 7 estimates on the relative growth of immigrants in the municipality after two and three terms, which correspond to 12 and 18 years after the initial election. We find that the e↵ects of the initial elections appear not only to persist after two elections but also to increase over time. After 12 years, the close election of a left-wing mayor is associated with a 3 p.p. larger increase in the share of immigrants relative to a right-wing municipality. Overall, the coe cient is twice as large after 12 relative to 6 years. After 18 years, the coe cient remains as large as after 12 years but it is imprecise and not statistically significant.
Separate estimates by election -To assess whether the e↵ects of elections vary over time, we report in Table S3 7: RDD estimates -Growth in immigrants, short-and long-run e↵ects  Sources: 1983Sources: , 1995Sources: , 2001Sources: , and 2008Sources: elections and 1982Sources: , 1990Sources: , 1999Sources: , 2008Sources: , and 2014 census. Notes: The table shows RDD estimates of the e↵ect of the victory of a left-relative to a right-wing mayor in the municipal election on the changes in the number of immigrant households relative to the initial municipality population. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth, and a quadratic polynomial. Estimates are reported using changes over 1, 2, and 3 elections in columns 1, 2, and 3, respectively. In Panel A, the dependent variable is the changes in the number of immigrant households relative to the initial population. In Panel B, the dependent variable is the changes in the number of non-European immigrants relative to the initial population. In Panel C, the dependent variable is the changes in the number of immigrant households with voting rights relative to the initial population. Changes are measured over 6 years after the elections in column 1, 12 years in column 2, and 18 years in column 3. Robust-bias corrected standard errors are reported in parenthesis. (*) and (**) denote statistical significance at 5% and 1% levels.
also positive and large for the 2001 election after two terms.
E↵ects by origins and voting rights -As immigrants might not be not allowed to vote, an important question is how much the previous changes a↵ect the composition of the electorate in the short and longer run. To investigate this issue, we report in Panel B of Table 7 separate estimates for immigrants from non-European origins who need to become citizens to vote in local elections, in contrast with most immigrants from European origins. The estimates clearly indicate that our results are driven by immigrants from non-European origins as there are few di↵erences between the estimates on overall immigrants relative to non-European immigrants.
Yet, as most immigrants in France acquire citizenship over time (Fougère and Safi, 2009), the longer-run e↵ect of these inflows on the electorate might still not be negligible. Figure 2: RDD plots -E↵ects on the share of immigrants in the population Sources: 1983Sources: , 1989Sources: , 1995Sources: , 2001Sources: and 2008Sources: elections and 1975Sources: , 1982Sources: , 1990Sources: , 1999Sources: , 2008 and 2014 census. Note: the graphs report quantile-spaced bins that capture averages from the same number of observations for each treatment group (Calonico et al., 2015). The solid lines represent a first-order polynomial. The vertical lines capture the discontinuity point at zero. Panel A represents changes in the share of immigrants in the population using changes observed before the election over 6 years and Panels B to D represent changes in the share of immigrants 6, 12 and 18 years after the election. Outcomes are measured at the household level.
We assess this hypothesis in Panel C by showing estimates using the growth in the share of immigrants with voting rights as a dependent variable. The results suggest that, over the period, about half of the growth in the share of immigrants is driven by immigrants with voting rights. As discussed earlier, these changes are directly politically favorable to elected mayors given that immigrants are much more likely to have a strong preference for the left.
Graphical evidence -We graphically illustrate our main results in Figure 2. 20 Consistent with the aforementioned evidence on the relationship between political orientation, origin and attitudes toward immigration, Graph A shows that the evolution of the share 20 To reflect the support of the running variable, the graphs report quantile-spaced bins that capture averages from the same number of observations for each treatment group (Calonico et al., 2015). For visual clarity, we restrict the sample to margins of victory lower than 50% (95% of our sample) and use a first-order polynomial. of immigrants is positively correlated with the relative margin of the left. However, after the election, a discontinuity at the threshold is visually clear (Graph B). Graphs depicting the evolution over more than one election show the same pattern, albeit with higher variability (Graphs C and D). In contrast, there is no visible discontinuity in 'placebo estimates' of the e↵ect of elections on the evolution of the share of immigrants before the election (Graph A).

Municipal policies
Next, we turn to the e↵ects of elections on municipal policies to understand which systematic di↵erences in policies may explain the larger immigrant inflows in left-relative to right-wing municipalities.
Taxes -A first possibility is that these population changes result from di↵erences in local taxes and spending between left-and right-wing municipalities. If immigrants prefer higher levels of public spending, which might be more likely in left-wing municipalities, systematic di↵erences in local fiscal policies might have attracted them into left-wing municipalities. To test this hypothesis, we report in Table 8 estimates of the e↵ects of elections on the two main municipal taxes: the housing tax in Panel A, which a↵ects all residents, including renters, and the property tax in Panel B, which only a↵ects homeowners, who tend to be more a✏uent. 21 We find no significant e↵ect of elections on local tax rates: all estimates are small (below 1 p.p.) and statistically insignificant. In the long run, point estimates even turn negative for the housing tax rate. Overall, these results are inconsistent with the hypothesis that left-wing municipalities might increase taxes. An important limitation of these results is that we do not have detailed data on how the municipal budget is spent.
As a result, while it is clear that taxes and spending did not increase systematically in left-relative to right-wing municipalities after a close election, we cannot rule out a reallocation of the budget toward policies preferred by immigrants.   1983, 1995, 20011983, 1995,20011983, 1995Sources: 1983, 1995, 2001elections and 1982, 1988, 1994, 2000, 2007 Recensement des Elments d'Imposition. Notes: The table shows RDD estimates of the e↵ect of the victory of a left-relative to a right-wing mayor. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth, and a quadratic polynomial. Panel A uses as a dependent variable the changes in the housing tax rate. Panel B uses as a dependent variable the changes in the property tax rate. Changes are measured over six years after the elections in column 1, 12 years in column 2, and 18 years in column 3. Robust-bias corrected standard errors are reported in parenthesis. (*) and (**) denote statistical significance at 5% and 1% levels.
Public housing -As discussed in Section 2, by design, public housing can be used to directly target future voters, and public housing disproportionately attracted non-European immigrants over the period. We investigate in Table 9 1983, 1995, 2001, 2008 1983, 1995, 2001 1983, 1995 xx Sources: 1983Sources: , 1995Sources: , 2001Sources: , and 2008Sources: elections and 1982Sources: , 1990Sources: , 1999Sources: , 2008, and 2014 census. Notes: The table shows RDD estimates of the e↵ect of the victory of a left-relative to a right-wing mayor in the municipal election. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth, and a quadratic polynomial. Panel A: the dependent variable is the change in the number of households in public housing relative to the initial municipality population. Panel B: the dependent variable is the change in the number of households in private housing relative to the initial municipality population. Panel C: the dependent variable is the change in the number of immigrants in public housing relative to the initial population. Panel D: the dependent variable is the share of immigrants in recent public housing, constructed after the indicated election relative to the initial municipality population. Panel E: the dependent variable is the change in the number of non-European immigrant households in public housing relative to the initial municipality population. Changes are measured over 6 years after the elections in column 1, 12 years in column 2, and 18 years in column 3. Outcomes are measured at the household level. Robust-bias corrected standard errors are reported in parenthesis. (*) and (**) denote statistical significance at 5% and 1% levels. Table S3 in the Supplementary Appendix. The e↵ects of elections on public housing are broadly consistent across elections, albeit dramatically larger for the 1983 election relative to other elections, in particular the 1995 election.

Separate estimates by elections are reported in Panel B of
In panel C, we assess how changes in the population in public housing contributed to the relative increase in the share of immigrants in left-wing municipalities. If di↵erences in inflows of immigrants into public housing are a key driver of our results, they should account for a substantial share of the overall growth identified earlier. Consistent with Figure 3: RDD plots -E↵ects on the share of households living in public housing in the population Sources: 1983Sources: , 1989Sources: , 1995Sources: , 2001Sources: and 2008Sources: elections and 1982Sources: , 1990Sources: , 1999Sources: , 2008 and 2014 census. Note: the graphs report quantile-spaced bins that capture averages from the same number of observations for each treatment group (Calonico et al., 2015). The solid lines represent a first-order polynomial. The vertical lines capture the discontinuity point at zero. Panel A represents changes in the share of households living in public housing in the population using changes observed before the election over 6 years and Panels B to D represent changes in the share of households living in public housing 6, 12 and 18 years after the election.
this hypothesis, we find that, even though public housing only accounts for 25% of total housing units on average, about half of the overall increase in the share of immigrants reflects an increase in the share of immigrants living in public housing.
Panel D isolates the contribution of inflows into recently constructed public housing in the overall growth in immigrants living in public housing. Even if most of the growth of immigrants in public housing is driven by an increase in existing units, the contribution of inflows into units constructed following the elections is not negligible. Overall, immigrants living in public housing units constructed after the elections account for about one-third of the total increase in the short run and close to half after twelve years. Finally, Panel E confirms that most of the e↵ect of elections is concentrated on non-European immigrants, who are much more likely to live in public housing. Figure 3 represents graphically the impact of elections on the share of households living in public housing. There is a clear discontinuity associated with the victory of a left-wing mayor, both in the short and long run. As shown in Figure S5 in the Supplementary Appendix, a large fraction of this discontinuity reflects the e↵ects of close elections on the share of immigrants living in public housing. Overall, the e↵ects of elections on public housing are consistent with the second prediction laid out in Section 1. Table S5 in the Supplementary Appendix, we investigate whether the SRU law implemented after 2001 influenced the e↵ects of elections on public housing constructions. We estimate the model separately for municipalities facing financial penalties if they do not construct more public housing and other municipalities. Consistent with the hypothesis that the SRU law made construction decisions less political, point estimates of the e↵ects of elections on public housing are twice as low in municipalities a↵ected by the SRU law. However, these results must be interpreted with caution as our estimates are not significantly di↵erent from zero, nor from each other.

Heterogeneity by initial immigration share
An important question is whether the e↵ects of elections vary with the initial attractiveness of the municipality to immigrants. Indeed, in line with the extensive literature on the determinants of immigrants' location choice showing that immigrants tend to locate where many immigrants are already living, local immigrant growth is strongly autocorrelated at the municipal level. 23 As a result, in France, as in other countries, immigrants concentrated in a subset of municipalities that received disproportionately larger immigrant inflows in recent decades (Ortega and Verdugo, 2021).
Given that immigrants are much more likely to live in public housing, mayors of municipalities with already many immigrants might anticipate that new public housing construction will dramatically increase immigrant inflows. This e↵ect might exacerbate partisan behavior, as it should incentivize right-wing mayors to avoid constructing public housing and, conversely, left-wing mayors to launch new construction programs.
To investigate this issue, we approximate municipalities' initial attractiveness to immi-23 Over the period, the autocorrelation between censuses is equal to 0.53, spanning between 0.43 in 1975-1982/1982-1990 and 0.68 in 1990-1999/1999-2008. 1983, 1995, 2001, 2008 1983, 1995, 2001 1983, 1995 (1)  1983, 1995, 20011983, 1995, 20011983, 1995Sources: 1983, 1995, 2001elections and 1982, 1990, 1999, and 2014 census. Notes: The table shows RDD estimates of the e↵ect of the victory of a left-relative to a right-wing mayor in the municipal election. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth, and a quadratic polynomial. Panel A: the dependent variable is the change in the number of immigrants relative to the initial municipality population. Panel B: the dependent variable is the change in the number of households in public housing relative to the initial municipality population. Columns 1, 3, and 5 report estimates of the model on municipalities with an above-median share of immigrants in the population in 1982 (12%). Columns 2, 4, and 6 report estimates of the model using municipalities below the median. Changes are measured over 6 years after the elections in column 1, 12 years in column 2, and 18 years in column 3. Outcomes are measured at the household level. Robust-bias corrected standard errors are reported in parenthesis. Robust-bias corrected standard errors are reported in parenthesis. (*) and (**) denote statistical significance at 5% and 1% levels.
grants with their median share of immigrants in 1982 (12%). Over the 1982-2014 period, the initial share of immigrants is a strong predictor of future immigrant inflows: the share of immigrants increases on average by 5 p.p. in municipalities above the median, but only by 3 p.p. in the other municipalities. In our empirical analysis, we make use of this feature to explore the potentially heterogeneous e↵ects of elections on immigration. Panel B shows that the e↵ect of elections on public housing construction also tends to be larger in high-immigration municipalities. However, the di↵erences are much less dramatic than for immigrants. Our estimates suggest that elections increase by 2.6 p.p.
the share of households in public housing in high-immigration municipalities, against 1.5 p.p. in low-immigration municipalities. Interestingly, the estimates are also more precise for high-immigration municipalities as the coe cient is not statistically significant for low-immigration municipalities.
Overall, di↵erences in immigrant inflows associated with the party elected appear specific to municipalities that were already attractive to immigrants at the beginning of the period, in line with the third prediction laid out in Section 1. This finding suggests that our results reflect partisan di↵erences between mayors who were facing larger potential immigrant inflows in their municipalities. 24

Population changes and incumbency advantage
Do larger immigrant inflows and public housing construction a↵ect future electoral outcomes? On the one hand, to the extent that they are associated with partisan behavior, larger population changes after the elections might lead non-partisan voters to punish the incumbent mayor in the next election (Ferejohn, 1986;Ashworth, 2012). On the other hand, such a policy might yield electoral gains if it reflects strategic partisan behavior from the mayor. As discussed earlier, the policy di↵erences could mobilize voters to support the incumbent mayor in addition to producing electorally favorable population changes.
To investigate which e↵ect dominates, we report in Table 11 the e↵ects of the close election of a left-wing mayor on the vote margin of the left in future municipal elections.
In Panel B, we allow the e↵ects to vary with the share of immigrants in 1982, which is strongly associated with partisan behavior, as argued in the previous section.
We find important di↵erences in incumbency advantage between these two groups of municipalities. In municipalities with an above-median immigrant share in 1982, the 24 The two groups of municipalities also vary along other dimensions. Indeed, municipalities above the median tend to be larger, have fewer workers in high-income occupations, a higher unemployment rate, and a higher share of households in public housing. In the Supplementary Appendix in Table S4, we report estimates where we include the predetermined values of all these characteristics as control variables. These estimates are very similar to our baseline results.  1983, 1995, 2001, 2008 1983, 1995, 2001 1983, 1995 (1) (2) (3) (4) (5) (6) At next election (6 years) At next-but-one election (12 years) At next-but-two election (18 years 1983, 1995, 20011983, 1995, 20011983, 1995Sources: 1983, 1989, 1995, 2001elections and 1982, 1990, 1999 census. Notes: The table shows RDD estimates of the e↵ect of the victory of a left-relative to a right-wing mayor in the municipal election. The dependent variable is the di↵erence in percentage points between the total number of votes won by the left and the total number of votes won by the right. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth, and a quadratic polynomial. Panel A reports estimates of the model in the full sample. Panel B reports estimates of the model on municipalities with an above or below-median share of immigrants in the population in 1982 (12%). Robust-bias corrected standard errors are reported in parenthesis. Robust-bias corrected standard errors are reported in parenthesis. (*) and (**) denote statistical significance at 5% and 1% levels.
estimated incumbency advantage for the next election is 10 p.p. In contrast, the incumbency advantage in other municipalities is twice as low and not statistically significant.
The di↵erence is even larger after two elections: in municipalities with an above-median immigrant share, we obtain an even higher incumbency advantage of 13 p.p. At the same time, the point estimate drops to 0.6 and is very imprecise for municipalities with few immigrants. This discrepancy suggests that the close election of a left-wing mayor in a municipality with a large initial immigrant population might trigger the creation of a political stronghold, consistent with the last prediction laid out in Section 1. However, as suggested by Columns 5 and 6, di↵erences between the two groups of municipalities dampen after two electoral cycles, even if the point estimate appears higher in highimmigration municipalities. Figure 4: RDD plots -E↵ects on future vote margin of the left: heterogeneity Sources: 1983Sources: , 1989Sources: , 1995Sources: , 2001Sources: , 2008Sources: and 2014Sources: elections and 1982 census. Note: the graphs report quantile-spaced bins that capture averages from the same number of observations for each treatment group (Calonico et al., 2015). The solid lines represent a first-order polynomial. The vertical lines capture the discontinuity point at zero. Panel A represents the vote margin of the left during the next election after 6 years. Panel B represents the vote margin of the left during the next-but-one election after 12 years. In Graphs A1 and B1, the sample is restricted to municipalities with an above-median share of immigrants in 1982. In Graphs A2 and B2, the sample is restricted to municipalities with a below-median share of immigrants in 1982. For visual clarity, we restrict the range of the forcing variable between -20 and 20 percentage points. Figure 4 represents graphically these results. In municipalities with an above-median share of immigrants, there is a clear discontinuity in the future vote margin of the left associated with the previous close victory of a left-wing mayor, both at the next and the next-but-one elections. In contrast, in the other municipalities, the incumbency advantage appears smaller and more imprecise at the next election. The estimate is virtually zero at the next-but-one election.

Robustness
Other samples and specifications -We document in Table 12 the robustness of our main results using alternatively as a dependent variable the growth of immigrant-headed households in Panels A1 and B1 and of households in public housing in Panels A2 and B2.
First, we assess in Panel A the sensitivity of our estimates to alternative samples. In Column 2, we include in our sample the 1989 election, which was not included previously as the Census data do not allow us to observe the outcomes 6 years after this election, unlike other elections. In Column 3, we include municipalities in smaller urban areas with less than 30 000 inhabitants where few immigrants live and which contain little public housing. In Column 4, we exclude the districts of Paris, Lyon, and Marseilles, which are not fiscally independent from the central municipality. In Column 5, we exclude elections in which a regionalist or independent list, which we somewhat arbitrarily classified as right-wing, competed in the final round. In Column 6, we exclude the 5% smallest and 5% largest municipalities. In Column 7, we exclude municipalities that are not observed at each election in our baseline sample. Overall, the results are in line with previous estimates, albeit more imprecise with the 1989 election and quantitatively lower when smaller urban areas are included in the sample. Table 12 investigates the robustness of our results to alternative construction of the data or specifications of the statistical model. In column 1, instead of constructing our dependent variables at the household level, the number of immigrants and inhabitants in public housing is measured using the population of all adults of more than 18 years of age. In practice, the estimates are una↵ected.

Panel B of
In column 2, instead of using our dependent variables in changes, we use the levels of these variables as observed 6 years after the election. Column 3 reports estimates    : 1983, 1989, 1995, 2001, 2008 and 2014 elections and 1982, 1990, 1999, 2008 and 2014 census. Unless otherwise specified, the sample includes municipalities with more than 9,000 inhabitants in urban areas with more than 30 000 inhabitants. Notes: The table shows RDD estimates of the e↵ect of the victory of a left-relative to a right-wing mayor. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth and a quadratic polynomial. In Panels A1 and A2, the dependent variable is the changes in the number of immigrants relative to the initial population. In Panels B1 and B2, the dependent variable is the change in the number of households in public housing relative to the initial population. Panel A: Column 1 reproduces our baseline estimates. Column 2 shows estimates on a sample including the 1989 elections. Column 3 shows estimates on a sample including municipalities located in small urban areas. Column 4 shows estimates on a sample excluding the municipalities (but not the urban areas) of Paris, Lyon and Marseilles. Column 5 excludes elections where regionalist lists were competing. Column 6 excludes municipalities among the 5% smallest or largest in terms of population. Column 7 excludes municipalities that are not observed for every election in our baseline sample. Panel B: Column 1 redefines the outcome variables using all adult individuals in the population instead of using the head-of-household. Column 2 presents the results of a specification where the dependent variables are defined in levels, using the share of immigrants and of households in public housing after the elections as a dependent variable. Column 3 includes additional control variables: the log of the initial population, the share of high-income occupations, the share of unemployed workers and the share of households in public housing in 1982, the number of lists in competition, the registration rate and the participation rate. Columns 4 and 5 estimate separate models for elections with 2 and 3 lists or more in competition. Columns 6 and 7 estimate separate models for elections with a single round and two rounds. Robust-bias corrected standard errors are reported in parenthesis. (*) and (**) denote statistical significance at 5% and 1% level.
In other columns, we allow the estimates to vary with the characteristics of the election.
Columns 4 and 5 test whether the results di↵er between elections with only 2 lists in the final round and elections with at least 3 lists in competition. Alternatively, columns 6 and 7 test whether the results depend on whether a second round had to be organized. Across these estimates, the results are very similar to our baseline results even if estimates are more precise when a second round is organized.
Alternative bandwidth choice and polynomial order -In Table S6 in the Sup-plementary Appendix, we document how our results depend on the choice of bandwidth.
While we used a fixed 15 p.p. bandwidth in the paper, we provide results using the data-driven procedure of Calonico et al. (2014), which automatically selects the bandwidth using a mean square error optimal bandwidth choice. We also report results for bandwidths of 5 p.p., 10 p.p., and 20 p.p. Overall, our estimates are not very sensitive to the choice of a specific bandwidth. If anything, point estimates obtained using the largest 20 p.p. bandwidth are lower than our baseline estimates. However, the results remain broadly similar and statistically significant in all three specifications.
We also report in Table S6 estimates using alternative polynomial orders. Instead of order 2 as in the rest of the paper, we experiment with polynomials of order 0, 1, and 3, as robustness checks. The estimates obtained using higher-order polynomials tend to be larger but are qualitatively similar.
Global polynomial estimates -In Table S7 in the Supplementary Appendix, we report estimates obtained using global instead of local polynomials as in the rest of the paper. While local estimators and low-order polynomials might be preferable (Gelman and Imbens, 2019), our results are not very sensitive to using this method with di↵erent polynomial orders. If anything, the point estimates tend to be smaller with that method.
The reported Akaike information criterion indicates that zero is the preferred polynomial order when the dependent variable is public housing and two for immigration. However, the results are statistically insignificant in the quadratic specification for the changes in the number of immigrants.

Discussion
An important question is the external validity of our results outside the context of close elections. For left-wing municipalities, Figure 2 shows that the partisan di↵erences in immigration growth appear to be fairly local and specific to close elections. Far from the threshold, immigrant growth appears lower not only in municipalities that received higher right-wing votes but also in those receiving the largest left-wing votes. The fact that close elections might trigger partisan behavior is consistent with a mechanism involving electoral incentives, because closely-elected mayors may be more concerned by their future prospects, and more likely to engage in strategic behavior to maximize their reelection chances.
In addition, elections further away from the cuto↵ may also lead to di↵erent policies because they involve di↵erent campaigning strategies. Bernhardt et al. (2020) have recently shown that candidates having a large popularity advantage may not adopt extreme positions favored by their core supporters, but, instead, target their opponents' moderate supporters. This pattern is consistent with a form of 'policy reversal', whereby betterelected mayors become more moderate (Cukierman and Tommasi, 1998;Moen and Riis, 2010).
Di↵erences in local context might also explain why the e↵ect of close elections appears local. In practice, municipalities in which the left or the right have a sizable electoral advantage are quite specific. The largest margins of the left correlate with a high share of low-income households and unemployed inhabitants, which could diminish the attractiveness of the municipality to immigrants. In contrast, and as illustrated earlier in Table 4, municipalities in which close elections occur are similar to the average municipality in our sample. This similarity suggests that our results may be deemed relevant for the entire population of municipalities.

Conclusion
This paper shows that the results of French municipal elections have influenced the share of immigrants across municipalities in the short and long run. We highlight that the main municipal policy associated with these changes is public housing, and we uncover systematic partisan di↵erences in how mayors used public housing. Relative to a rightwing mayor, the close election of a left-wing mayor resulted in more construction of public housing and more public housing tenants of immigrant origin. Even though we cannot formally establish the reason for these partisan di↵erences, we find suggestive evidence that electoral incentives may have played a role and that, as a result, immigration contributed to building local political strongholds. Our findings open a fruitful avenue for future research by shedding light on a possible feedback loop between ethnic and partisan segregation.
An important limitation of this study is that we cannot observe all dimensions of local policies. As a result, we cannot rule out that other unobserved systematic policy di↵erences between left-and right-wing mayors also increased demand from immigrants to live in public housing in left-wing municipalities and discouraged immigrants from locating in right-wing municipalities. In particular, given that we lack information on applications to public housing, we cannot disentangle the e↵ects of elections on the local demand for public housing from their identified e↵ects on supply.
Outside the French context, our results are relevant for designing housing programs that aim to improve the housing conditions of the poorest part of the population. While out of favor in North America, public housing remains popular in Europe (Scanlon et al., 2014), Asia (Chen et al., 2013;Chiu, 2013;Xu and Zhou, 2019) and, increasingly, South America (Krause et al., 2013). Public housing may be an e cient policy tool against high levels of market-driven residential segregation, be it income-based or origin-based (Quillian and Lagrange, 2016;Verdugo and Toma, 2018). However, the influence of local authorities suggests that local political considerations may hinder their e↵ectiveness.
In light of our results, one might want to advocate a stricter separation between local authorities and public housing agencies. One way to achieve such a separation is to delegate control over larger jurisdictions, such as inter-municipal cooperation units, and push for a higher level of concentration in the public housing sector. However, larger public housing agencies may also be more disconnected from the specificity of local housing supply and demand. In France, recent legislation seems to indicate that the government is trying to meet this double challenge. 25 25 The ALUR law, voted in 2015, gave inter-municipal cooperation units authority over local housing policy. In addition, it created the National Agency for Public Housing Control (ANCOLS), which implements random audits of public housing agencies to ensure that the most deserving applicants are served first, regardless of their other characteristics. As stipulated by the ELAN Law, voted in 2018, small public housing agencies (managing less than 12,000 dwellings) have been compelled to join a consortium since January 2021.

A Data
Municipal election data -The data come from the Center of Socio-Political Data (Centre de Données Socio-Politiques), which is a research center from  The data include elections results in 1983, 1989, and 1995 for municipalities with more than 9,000 inhabitants and results from the 2001 and 2008 elections for municipalities with more than 3,500 inhabitants. We drop Corsica, French Guyana, and Overseas territories from the sample because these regions host many candidates who are a liated with a regionalist and independentist party that cannot easily be categorized as right-wing or left-wing. We Census data -We use restricted access data obtained from the Centre d'Accès Sécurisé Distant (CASD) which is a secure access data center. We have access to confidential individual files for a very large sample extract (25%). For all years, we use the supplement files (exploitation complémentaire), which contain detailed variables on occupations and education.
Before the 2000s, the French census was based on a complete enumeration. Since 2002, the annual census releases have been based on a rolling sample in which municipalities with more than 10,000 inhabitants are divided into five rotation groups surveyed every five years. As a result, the new census reflects 5-year averages around the census year (Durr, 2005;Desplanques and Rogers, 2008). For municipalities with populations between 9 000 and 10 000, the data are only collected once every five years. These municipalities account for less than 5% of our sample.
We use the occupational classification at the one-digit level and the labor force status from the census to decompose the share of natives and immigrants in the population into three groups: 1) high-income occupations, which correspond to managers and professionals; 2) low-income occupations, which include blue-collar and service sector employees; 3) retirees. Software -All RDD estimates and graphs have been obtained using the Rdrobust package on Stata (Calonico et al., 2017).

B Mayors and public housing in France
From the beginning of public housing programs in the late XIX th Century to the late 1970s, the housing market in France was managed in a very centralized manner. Local authorities had limited oversight and their power was restricted to situations of emergency, for instance when dilapidated buildings posed an immediate threat to security. However, the situation changed radically during the following three decades, which make up the period under study in this paper. Between the decentralization law of 1982 and the law on inter-municipal cooperation of 2010, mayors played an essential role in local real estate policies, ranging from long-term urban planning to daily decisions on the delivery of building permits. 27 In particular, mayors had the power to initiate new public housing programs (and, in some particular situations, to engage the privatization of some public housing units). They were also co-responsible for the management of the local public housing stock, alongside 27 This situation has changed again with the rise of inter-municipal cooperation units (EPCI), which are meant to address the challenges raised by the very large number of municipalities in France (Tricaud, 2021). As of today, EPCIs are in charge of private and public housing policies, both in terms of the development of future programs and the management of existing ones. Mayors elect a president of the EPCI, who is often, in practice, the mayor of the largest municipality. local public housing agencies (the "organismes HLM "). Local elected o cials were de jure board members of these agencies.
The distribution of public housing was decided by a committee in charge of allocating vacant dwellings (the "commission d'attribution"). This committee was composed of representatives from the agency, the municipal council, and the ministry of housing and was supposed to discuss the merits of the di↵erent applicants and propose a ranking. In practice, the list of applicants was not public and the committee was not required to justify its decisions. In addition, the committee was often chaired by the mayor or one of her delegates in small municipalities. Finally, as is still the case today, a fraction (up to 20%) of the stock of public housing was considered "reserved" for the municipality, and exempt from the scrutiny of the committee.  and Verdugo (2022). Section S1 lays out a theoretical model based on Glaeser and Shleifer (2005)

S1
Shaping the electorate through immigration: a model S1.1 Voting framework The population is composed of immigrants and natives. Consistent with our data, immigrants are always a minority. For simplicity and consistent with the empirical finding that elections do not a↵ect the size of municipalities, total population is fixed. We assume that favoring immigrants entails a transfer of resources away from natives. For example, it may be that the latter lose access to municipal public housing because its supply is limited. Denote as q the cost paid by each native when the mayor favors immigrants and as ⇡(q) > 1 the ratio of the number of natives to the number of immigrants in the municipality, which is decreasing in q. Each immigrant receives ⇡(q)q when each native experiences a loss q. 28 We also assume that voters are partisans such that the native population is split between pro-immigrant, who mildly favor the mayor, and anti-immigrant, who disfavor the mayor. However, mayors cannot directly observe natives' preferences, which implies that redistribution can only target immigrants or natives. We denote as I(q), O(q) and S(q) the respective population shares of immigrants, anti-immigrant natives and proimmigrant natives associated with a given q. By definition, I(q) = 1/(1 + ⇡(q)) and O(q) + S(q) = ⇡(q)/(1 + ⇡(q)). As in Glaeser and Shleifer (2005), voters' preferences for the incumbent depend on three components: idiosyncratic support for the candidate, group membership and past policies. The idiosyncratic component of preferences is captured by assuming that each voter receives utility j from supporting the incumbent against the opponent, where j is symmetrically distributed around zero with single-peaked density f (j) and cumulative distribution F (j). The group membership component is captured by assuming that immigrants receive utility of v 0 /2 > 0 if a left-wing mayor is elected and v 0 /2 if a right-wing mayor is elected. 29 For pro-immigrant natives who lean mildly toward the left, the election of a left-wing mayor provides utility of v 00 /2 2 [0, v 0 /2] and that of a right-wing mayor, v 00 /2. In contrast, anti-immigrant natives receive utility v 0 /2 (resp., v 0 /2) if a right-wing (resp., left-wing) leader is elected. Finally, voters are also influenced by the incumbent's past policies. Higher redistribution toward immigrants increases votes from this group but might decrease votes from natives. Immigrants obtain utility of v 1 (⇡(q)q) if the incumbent left-wing mayor is reelected, where v 1 (0) = 0 and v 0 1 (·) > 0. The two groups of natives obtain utility of v 1 ( q) if the incumbent is reelected.
Assuming these three components are additive and utility is linear, the left incumbent who implements policy q > 0 will then face a vote share of among anti-immigrant natives, of P I (q) = (1 F [ v 0 v 1 (⇡(q)q)]) among immigrants and of P S (q) = (1 F [ v 00 v 1 ( q)]) among pro-immigrant natives. The total share of votes is given by P (q) = P g2{O,I,S} g(q)P g (q). The mayor will set a strictly positive level of q if it helps her being reelected, that is, if P 0 (0) > 0. For consistency with our empirical strategy based on close elections, we assume that the initial situation is such that when q = 0, the election is a tie and P (0) = 1/2. 30

S1.2 Urban framework
We embed this voting process in a simple urban framework featuring job-unrelated mobility within a single urban area. 31 The urban area is made of two municipalities of fixed size ↵ and 1 ↵. Ex ante, size is the only exogenous di↵erence between the two municipalities and it has no impact on utility. 32 Therefore, the initial composition of the population is the same in both municipalities. In particular, the initial share of natives in each munici-29 These preferences are independent of past policies and are best thought of as representing a pure political taste derived from history.
32 For example, the larger municipality may also be the more central, so that commuting costs and housing costs o↵set each other. pality is the same as in the urban area as a whole and given by ⇡(0)/(1 + ⇡(0)). We focus on the municipality of size ↵.
Locational preferences are partly idiosyncratic. This element of preferences is captured by assuming that, independently of group membership, each voter receives utility ⇠ from living in the municipality of interest, compared to the other one. This variable ⇠ is distributed according to a distribution`(⇠), with cumulative L(⇠) and hazard h(⇠) and we define the cuto↵ value ⇠ 0 such that ↵ = 1 L(⇠ 0 ). On top of their intrinsic preference ⇠, residents also care about the provision of public services: natives will assign any policy change q > 0 a value q, while immigrants will value it ⇡(q)q. Finally, anti-immigrant natives also derive negative utility from having more immigrant neighbors than they would have if they lived in the alternative municipality. This feature is represented by a function V (⇡(q)), with V (⇡(0)) = 0 and V 0 (·)  0. Assuming that utility is linear and additive in these three components, the residential utility of the three groups is given by: In order to recover the resulting spatial political equilibrium, we assume perfect sorting and free mobility. The timing is as follows: (i) the mayor announces the policy q; (ii) voters adjust their location; (iii) the expected vote share is computed.

S1.3 Shaping the electorate
Setting q > 0 will trigger population responses, which will translate into a new nativeimmigrant ratio ⇡(q) and a new political make-up in the native population. We formalize how the share of anti-immigrant natives O(q) responds to a change in q by a function 1+⇡ (0) ⌘ , with G 0 (·) > 0 and G(⇡(0)) = 1.
This setting allows us to write Proposition 1: Proposition 1. A positive value of q raises support for the mayor if and only if: with population responses defined by: Proof. See Appendix S1.4.
Equation S4 shows that, as in Glaeser and Shleifer (2005), the politician will choose to favor immigrants when group preferences (v 0 ) are important and when the vote losses associated with discrimination (v 0 1 (0)) are small, that is, if past choices have little direct e↵ect on voters' behavior. A distinct feature of our model is the group of pro-immigrant natives that acts as a disciplining device for the incumbent. When those have a weak intrinsic preference for the incumbent (v 00 ! 0), reshaping is less likely to increase votes.
More specific to our purpose, condition S4 also illustrates why the usefulness of electoral reshaping e↵ectively depends on two population responses: on the extensive margin ( ⇡ 0 (0)/⇡(0) is large) when setting q > 0 triggers demographic reshu✏ing and on the intensive margin (G 0 (⇡(0)) is large) when setting q > 0 triggers political reshu✏ing among the native population. Even a very moderate demographic reshu✏ing may have large political consequences if anti-immigrant natives are very prone to reacting to immigrant inflows by leaving the municipality. 33
Share of immigrants -We first consider a simple case where changes in q may only a↵ect the ethnic composition of the municipality between immigrants I and natives N . With V (⇡(q)) = 0, we can define a single utility function for natives U N (·) such that U N (⇠, q) = U S (⇠, q) = ⇠ q. The measure of ⇠ in the two groups is given by 33 Note that a more complete model could assume that immigrants also derive utility from having more immigrant neighbors. This feature could be represented by a function W (⇡(q)), with W (⇡(0)) = 0 and W 0 (·) 0. In that case, Equation S5 would become ⇡ 0 (0)/⇡(0) = (1 + ⇡(0)) · L 0 (⇠ 0 )/[1 L(⇠ 0 ) ⇡(0)L 0 (⇠ 0 )W 0 (⇡(0))], which means that the policy shift would trigger greater demographic reshu✏ing. Dynamics could also be introduced through random utility shocks, as in Ouazad (2015). We abstract from these extensions for the sake of simplicity. 34 A Mathematica code is available upon request.
Composition of the native population -We now turn to the more general case where V (⇡(q)) 6 = 0 if q > 0. Implementing q > 0 will increase the share of immigrants to the extent that U I (⇠ I (q), q) = U S (⇠ S (q), q) = U O (⇠ O (q), q). This means that ⇠ S (q) ⇠ I (q) = q(1 + ⇡(q)) and ⇠ O (q) ⇠ S (q) = V (⇡(q)). Then we can use the market-clearing relationships: 1+⇡ (0) ⌘ , we get: Di↵erentiate this expression, simplify using V (⇡(0)) = 0 and equation S5 and isolate G 0 (⇡(0)) to get equation S6.  Note: In 1988, the left vote was 31 points higher among voters believing that there are not too many immigrants in France; in 2012, this gap was equal to 40 points. Electoral surveys, nationwide sample. Source: Piketty (2020, Figure S14.19d). Figure S3: Discontinuity in the forcing variable. Vote histogram Sources: 1983Sources: , 1995Sources: , 2001Sources: and 2008Sources: elections and 1982Sources: , 1990Sources: , 1999Sources: , 2008 and 2014 census. Note: Histogram of the vote margin for the left defined as the di↵erence between the share of votes for the better ranked left-and right wing lists in competition in the final round of the election. Elections with a vote margin for the left superior to zero have been won by the left. When no right-wing is competing, the vote margin is defined to be one, and minus one when no left-wing list is competing. Figure S4: Discontinuity in the forcing variable. Kernel plot Sources: 1983Sources: , 1995Sources: , 2001Sources: and 2008Sources: elections and 1982Sources: , 1990Sources: , 1999Sources: , 2008 Figure S5: RDD plots -E↵ects on the share of immigrants living in public housing in the population Sources: 1983Sources: , 1995Sources: , 2001Sources: and 2008Sources: elections and 1982Sources: , 1990Sources: , 1999Sources: , 2008 and 2014 census. Note: the graphs report quantile-spaced bins that capture averages from the same number of observations for each treatment group (Calonico et al., 2015). The solid lines represent a first-order polynomial. The vertical lines capture the discontinuity point at zero. Panel A represents changes in the share of immigrants in public housing in the population using changes observed before the election over 6 years and Panels B to D represent changes in the share of immigrants in public housing 6, 12 and 18 years after the election. Outcomes are measured at the household level. Figure S6: RDD plots -E↵ects on future vote margin of the left Sources: 1983Sources: , 1989Sources: , 1995Sources: , 2001Sources: , 2008 and 2014 elections. Note: the graphs report quantile-spaced bins that capture averages from the same number of observations for each treatment group (Calonico et al., 2015). The solid lines represent a first-order polynomial. The vertical lines capture the discontinuity point at zero. Panel A represents the vote margin of the left during the previous election and Panels B to D represent the vote margin of the left during the next, next-but-one and next-but-two election. Figure S7: RDD plots -E↵ects on future vote margin of the left -Zooming in Sources: 1983Sources: , 1989Sources: , 1995Sources: , 2001Sources: , 2008 and 2014 elections. Note: the graphs report quantile-spaced bins that capture averages from the same number of observations for each treatment group (Calonico et al., 2015). The solid lines represent a first-order polynomial. The vertical lines capture the discontinuity point at zero. Panel A represents the vote margin of the left during the previous election and Panels B to D represent the vote margin of the left during the next, next-but-one and next-but-two election. For visual clarity, we restrict the range of the forcing variable between -20 and 20 percentage points. Sources : 1982, 1990, 1999, 2008 and 2014 census. Notes: Panel A shows the share of the population, of immigrants and of public housing units in municipalities with more than 9 000 inhabitants located in urban areas with more than 30 000 inhabitants. Panel B compares the share of immigrants and of inhabitants in public housing between municipalities with more than 9 000 inhabitants located in urban areas with more than 30 000 inhabitants and other municipalities.    Sources : 1983, 1989, 1995, 2001 and 2008 elections and 1982, 1990, 1999, 2008 and 2014 census. Notes:

S3 Additional Tables
The table shows RDD estimates of the e↵ect of the victory of a left-relative to a right-wing mayor in the municipal election. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth and a quadratic polynomial. Each cell report a separate estimate for di↵erent elections in our sample using changes over alternative time horizons. In panel A, the dependent variable is the change in the number of immigrant households in the population over the initial population of the municipality. In panel B, the dependent variable is the change in the number of households in public housing over the initial population of the municipality. All outcomes are measured at the household level. Robust-bias corrected standard errors are reported in parenthesis. (*) and (**) denote statistical significance at 5% and 1% level.  1983, 1995, 20011983, 1995, 20011983, 1995Sources: 1983, 1995, 2001elections and 1982, 1990, 1999 and 2014 census. Notes :  The table reproduces estimates from Table 10 (Panels A and B) and from Panel B in Table 11 (Panel C), controlling for the same covariates as in Column 3 of Panel B in Table 12. Robust-bias corrected standard errors are reported in parenthesis. Robust-bias corrected standard errors are reported in parenthesis. (*) and (**) denote statistical significance at 5% and 1% level. All Sources: 2001 and 2008 elections and 2008 and 2014 census. The sample includes municipalities with more than 9,000 inhabitants. Notes: The table shows RDD estimates of the e↵ect of the victory of a left-relative to a right-wing mayor in the municipal election. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth and a quadratic polynomial. The dependent variable is the change in the number of households in public housing over the initial population of the municipality. Column 2 report estimation results for municipalities a↵ected by the SRU law. Column 3 report estimation results for municipalities not a↵ected by the SRU law. All outcomes are measured at the household level. Robust-bias corrected standard errors are reported in parenthesis. Sources : 1983, 1989, 1995, 2001 and 2008 elections and 1982, 1990, 1999, 2008 and 2014 census. Notes: The table shows RDD estimates of the e↵ect of the victory of a left-relative to a right-wing mayor in the municipal election. Estimates are obtained using local linear regressions with a triangular kernel, a 15% bandwidth and a quadratic polynomial. In panel A, the dependent variable is the change in the number of immigrant households in the population over the initial population of the municipality. In panel B, the dependent variable is the change in the number of households in public housing over the initial population of the municipality. All outcomes are measured at the household level. Robust-bias corrected standard errors are reported in parenthesis. (*) and (**) denote statistical significance at 5% and 1% level.  Sources: 1983Sources: , 1989Sources: , 1995Sources: , 2001Sources: and 2008Sources: elections and 1982Sources: , 1990Sources: , 1999Sources: , 2008Sources: and 2014. Notes: The table shows RDD estimates from global two-sided polynomials with the indicated polynomial order. All outcomes are measured at the household level. Robust standard errors are reported in parenthesis. (*) and (**) denote statistical significance at 5% and 1% level. AIC: Akaike information criterion.