The 1906 San Francisco Fire - Statistics

Figures Figure 1: The Area Burnt in the 1906 Fire in San Francisco, Hoag (1907) Figure 2: Location of Households in 1900, on Modern Map Note: Source: 1900 Census, Author’s calculations Figure 3: The Effect on Migration in 1910 (a) Dependent Variable: Living Outside of San Francisco (b) Dependent Variable: Living Outside of California (c) Dependent Variable: Living in Bay Area Note: Dots plot coefficients of a non-parametric regression of outcome on distance-to-border bins and the covariates in equation 3. The hollow dots are individuals in the donut hole which I exclude from the spatial RD regression. The solid lines show the fitted values given by the coefficients on the distance-to-border measure from the spatial RD specification in equation 3, controlling for race, marital status, home ownership and household size in 1900. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border. Observations weighted by inverse propensity weights (Bailey et al., 2020 and optimal balancing weights (Heuermann et al., 2019). Figure 4: The Effect on Occupation and Entrepreneurship in 1910 (a) Dependent Variable: Log Occupational Income (b) Dependent Variable: Self-Employment (c) Dependent Variable: Having Employees Note: Dots plot coefficients of a non-parametric regression of outcome on distance-to-border bins and the covariates in equation 3. The hollow dots are individuals in the donut hole which I exclude from the spatial RD regression. The solid lines show the fitted values given by the coefficients on the distance-to-border measure from the spatial RD specification in equation 3, controlling for race, marital status, home ownership and household size in 1900. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border. Observations weighted by inverse propensity weights (Bailey et al., 2020 and optimal balancing weights (Heuermann et al., 2019). Figure 5: Bandwidth Choice and Short-Run Treatment Effects (a) Dependent Variable: Migration (b) Dependent Variable: Log Occupational Income (c) Dependent Variable: Self-Employment (d) Dependent Variable: Having Employees Note: Dots plot coefficients o f a n on-parametric r egression o f o utcome o n d istance-to-border b ins a nd t he covariates in equation 3. The hollow dots are individuals in the donut hole which I exclude from the spatial RD regression. The solid lines show the fitted values given by the coefficients on the distance-to-border measure from the spatial RD specification in equation 3, controlling for race, marital status, home ownership and household size in 1900. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border. Observations weighted by inverse propensity weights (Bailey et al., 2020 and optimal balancing weights (Heuermann et al., 2019). Figure 6: Long-Run Effect on Migration Lives Outside of San Francisco .1 .05 0 -.05 1910 1920 1930 1940 RD Coefficients with 90% Confidence Intervals (a) Outcome: Living Outside of San Francisco Returned to San Francisco .1 .05 0 -.05 -.1 -.15 1920 1930 1940 RD Coefficients with 90% Confidence Intervals (b) Outcome: Migrant Returned to San Francisco Note: Figure shows the regression coefficients on a treatment indicator for the spatial regression discontinuity design and the 95% confidence interval. Standard errors allow for arbitrary spatial correlation of unknown form with a distance cutoff of 300m. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border, and who were between 10 to 35 years old in 1900. Observations weighted by inverse propensity weights (Bailey et al., 2020 and optimal balancing weights (Heuermann et al., 2019). Dependent variable in Panel (a) indicates whether individual lives outside of San Francisco, dependent variable in Panel (b) indicates residence in San Francisco for those who lived outside of San Francisco in 1910. Figure 7: Long-Run Effect on Occupational Income Log Occupational Income Score .05 0 -.05 -.1 -.15 1910 1920 1930 1940 RD Coefficients with 90% Confidence Intervals Occupational Downgrading Rel. to 1900 (a) Outcome: Log Occupational Income .1 .05 0 -.05 1910 1920 1930 1940 RD Coefficients with 90% Confidence Intervals (b) Outcome: Occupational Downgrading Note: Figure shows the regression coefficients on a treatment indicator for the spatial regression discontinuity design and the 95% confidence interval. Standard errors allow for arbitrary spatial correlation of unknown form with a distance cutoff of 300m. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border, and who were between 10 to 35 years old in 1900. Observations weighted by inverse propensity weights (Bailey et al., 2020 and optimal balancing weights (Heuermann et al., 2019). Dependent variable in Panel (a) is log occupational income score, dependent variable in Panel (b) indicates whether an individual has a lower occupational income relative to 1900. Figure 8: Long-Run Effect on Entrepreneurship Self-Employment 0 -.05 -.1 -.15 -.2 1910 1920 1930 1940 RD Coefficients with 90% Confidence Intervals (a) Outcome: Self-Employment Has Employees 0 -.05 -.1 1910 1920 1930 1940 RD Coefficients with 90% Confidence Intervals (b) Outcome: Has Employees Note: Figure shows the regression coefficients on a treatment indicator for the spatial regression discontinuity design and the 95% confidence interval. Standard errors allow for arbitrary spatial correlation of unknown form with a distance cutoff of 300m. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border, and who were between 10 to 35 years old in 1900. Observations weighted by inverse propensity weights (Bailey et al., 2020 and optimal balancing weights (Heuermann et al., 2019). Dependent variable in Panel (a) indicates whether individual is self-employed, dependent variable in Panel (b) indicates whether an individual has employees. Tables Table 1: Differences in Pre-Treatment Characteristics (Measured in 1900) Age White Born in the US Born in California Mother Born in the US Father Born in the US Mother Born in California Father Born in California Married Number of Own Children Number of Own Children Below Age 5 Number of Siblings in Household Household Size Household Head Children in Labor Force Wife in Labor Force Kin Presence Middle Initial Length First Name Length Last Name Home Ownership Mortgage Log Occupational Income Score Months Unemployed in 1899 In Labor Force School Attendance Illiterate Observations Inside Razed District Standard Error 25492 25492 25492 25492 25492 25492 25492 25492 25492 25492 25492 25492 25492 25492 2491 8704 25492 25492 25492 25492 22800 5224 15971 20196 21839 25492 25492 0.422 -0.126 -0.110∗ -0.093∗∗∗ -0.060 -0.042 -0.014∗∗ -0.003 -0.078∗∗∗ -0.439∗∗∗ -0.065∗∗∗ -0.383∗∗∗ 0.778 -0.125∗∗∗ 0.042∗∗∗ 0.019 0.017 -0.075∗∗∗ -0.344∗ -0.368 -0.242∗∗∗ -0.096∗∗∗ -0.085∗∗∗ 0.264∗∗∗ 0.037∗∗ -0.077∗∗∗ 0.023∗ (0.601) (0.099) (0.057) (0.033) (0.039) (0.037) (0.006) (0.004) (0.025) (0.091) (0.024) (0.091) (0.709) (0.030) (0.014) (0.019) (0.034) (0.027) (0.204) (0.283) (0.031) (0.028) (0.023) (0.079) (0.017) (0.021) (0.013) Note: ∗ p < 0.1 ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border. Each row represents a separate regression of pre-treatment outcomes (measured in 1900) on treatment indicator, without additional controls. Implements inverse propensity weights (Bailey et al., 2020). Table 2: Placebo RD with Pre-Treatment Characteristics (Sample: Linked 1900-1910) Age White Born in the US Born in California Mother Born in the US Father Born in the US Mother Born in California Father Born in California Married Number of Own Children Number of Own Children Below Age 5 Number of Siblings in Household Household Size Household Head Children in Labor Force Wife in Labor Force Kin Presence Middle Initial Length First Name Length Last Name Home Ownership Mortgage Log Occupational Income Score Months Unemployed in 1899 In Labor Force School Attendance Illiterate Observations Inside Razed District Standard Error 24760 24760 24760 24760 24760 24760 24760 24760 24760 24760 24760 24760 24760 24760 2419 8442 24760 24760 24760 24760 22155 5142 15488 19578 21203 24760 24760 0.583 0.011 -0.032 -0.036 -0.031 -0.003 -0.016∗ 0.001 0.071∗∗∗ 0.099∗∗ 0.025 -0.263∗∗∗ -2.201∗∗∗ 0.104∗∗∗ 0.021 0.038 -0.071 0.022 -0.089 -0.171 -0.153∗∗∗ 0.007 -0.002 0.071 0.028 -0.019 -0.004 (0.514) (0.012) (0.030) (0.029) (0.037) (0.035) (0.009) (0.006) (0.014) (0.041) (0.017) (0.096) (0.725) (0.013) (0.024) (0.025) (0.062) (0.053) (0.077) (0.116) (0.046) (0.052) (0.033) (0.122) (0.025) (0.023) (0.006) Note: ∗ p < 0.1 ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border. Each row represents a separate placebo spatial regression discontinuity regression. Observations weighted by inverse propensity weights (Bailey et al., 2020) and optimal balancing weights (Heuermann et al., 2019), running variable is distance to the border. Dependent variables measured in 1900. No individual level controls. Table 3: The Short-Run Effect of the 1906 San Francisco Fire, Spatial Regression Discontinuity Design Lives Outside of SF in 1910 Lives Outside of CA in 1910 Migrant Lives in Bay Area in 1910 Log Occupational Income in 1910 Difference Log Occupational Income 1900-1910 Occupational Downgrading 1900-1910 Self-Employment in 1910 Has Employees in 1910 Wife’s LFP in 1910 Child’s LFP (Age 10-15) in 1910 Child’s School Attendance (Age 6-15) in 1910 Got Married 1906-1910 Children Born 1906-1910 Observations Inside Razed Area Standard Error 24760 24760 24760 18413 11755 16746 17933 17933 12951 3610 7274 24760 24760 0.063∗∗∗ 0.016 0.046∗∗∗ -0.074∗∗ -0.044 0.063∗∗ -0.100∗∗∗ -0.053∗∗ 0.029∗∗ 0.020 -0.012 -0.019∗∗ -0.025 (0.022) (0.018) (0.014) (0.030) (0.039) (0.027) (0.028) (0.024) (0.012) (0.019) (0.021) (0.009) (0.018) Note: ∗ p < 0.1 ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border. Each row represents a separate spatial regression discontinuity regression, controlling for race, marital status, home ownership and household size in 1900. Observations weighted by inverse propensity weights (Bailey et al., 2020) and optimal balancing weights (Heuermann et al., 2019), running variable is distance to the border. Table 4: Robustness of Main Short-Run Results (Part I) Inside Razed Area Observations Inside Razed Area Observations Log Occupational Income (2) SelfEmployment (3) Has Employees (4) (a) Baseline Specification 0.063 -0.074∗∗ (0.022) (0.030) 24,760 18,413 -0.100∗∗∗ (0.028) 17,933 -0.053∗∗ (0.024) 17,933 (b) No Pre-Treatment Controls 0.070∗∗∗ -0.073∗∗ (0.021) (0.031) 24,760 18,413 -0.104∗∗∗ (0.028) 17,933 -0.058∗∗ (0.025) 17,933 Lives Outside San Francisco (1) ∗∗∗ (c) Add Controls: Occscore, Number of Children, LFP, Born US, Age Inside Razed Area 0.064∗∗∗ -0.070∗∗ -0.092∗∗∗ -0.050∗∗ (0.022) (0.029) (0.025) (0.024) Observations 24,760 18,413 17,933 17,933 (d) Add Controls: Lagged Dependent Variable Inside Razed Area 0.063∗∗∗ -0.074∗∗ -0.085∗∗∗ (0.022) (0.030) (0.021) Observations 24,760 18,413 17,933 -0.049∗∗ (0.021) 17,933 (e) No Inverse Probability Weights 0.065∗∗∗ -0.072∗∗ -0.101∗∗∗ (0.022) (0.030) (0.027) 24,760 18,413 17,933 -0.056∗∗ (0.023) 17,933 (f) No Optimal Balancing Weights 0.037∗ -0.059∗∗ -0.045∗ (0.019) (0.025) (0.023) 18,028 13,427 13,047 -0.034∗ (0.018) 13,047 Inside Razed Area Observations Inside Razed Area Observations Note: ∗ p < 0.1 ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Baseline specification in Panel (a) is spatial regression discontinuity regression, controlling for race, marital status, home ownership and household size in 1900. Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border. Observations weighted by inverse propensity weights (Bailey et al., 2020) and optimal balancing weights (Heuermann et al., 2019), running variable is distance to the border. Panel (b) omits pre-treatment controls from the regression, Panel (c) additionally includes 1900 values of occscore, number of children, LFP, born US, and age, Panel (d) controls for the lagged value of the dependent variable. Panel (e) does not use inverse propensity weights, Panel (f) does not use optimal balancing weights and excludes individuals along the Eastern part of the fire border. Table 5: Robustness of Main Short-Run Results (Part II) Inside Razed Area Observations Inside Razed Area Observations Inside Razed Area Observations Inside Razed Area Observations Inside Razed Area Observations Inside Razed Area Observations Log Occupational Income (2) SelfEmployment (3) Has Employees (4) (a) Baseline Specification 0.063 -0.074∗∗ (0.022) (0.030) 24,760 18,413 -0.100∗∗∗ (0.028) 17,933 -0.053∗∗ (0.024) 17,933 (b) Latitude and Longitude 0.069∗∗∗ -0.044∗∗ (0.018) (0.022) 24,760 18,413 -0.070∗∗∗ (0.021) 17,933 -0.045∗∗ (0.018) 17,933 (c) Include the Donut Hole Observations 0.046∗∗ -0.061∗∗∗ -0.086∗∗∗ (0.018) (0.022) (0.021) 27,486 20,429 19,898 -0.046∗∗ (0.019) 19,898 Lives Outside San Francisco (1) ∗∗∗ (d) Restrict to Whites 0.063 -0.066∗∗ (0.021) (0.029) 23,059 17,238 -0.098∗∗∗ (0.027) 16,764 -0.053∗∗ (0.025) 16,764 (e) Age 20 - 50 in 1910 0.054 -0.078∗∗ (0.026) (0.034) 19,209 14,545 -0.114∗∗∗ (0.026) 14,444 -0.062∗∗∗ (0.024) 14,444 ∗∗∗ ∗∗ (f) Alternative Linking Method 0.045∗∗ -0.075∗∗∗ -0.095∗∗∗ (0.018) (0.026) (0.030) 34,512 25,610 24,874 -0.035∗ (0.021) 24,874 Note: ∗ p < 0.1 ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Baseline specification in Panel (a) is spatial regression discontinuity regression, controlling for race, marital status, home ownership and household size in 1900. Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border. Observations weighted by inverse propensity weights (Bailey et al., 2020) and optimal balancing weights (Heuermann et al., 2019), running variable is distance to the border. Panel (b) uses longitude and latitude as the running variables. Panel (c) includes individuals who live less than 100m away from the border. Panel (d) restricts the sample to whites, and Panel (e) restricts the sample to individuals who are between 20 and 50 years old in 1910. Panel (f) implements baseline specification on a sample linked with Price et al. (2021) crosswalks. Table 6: Effect of Fire in 1910 by 1900 Home Ownership Inside Razed Area Home Ownership × Inside Observations Lives Outside San Francisco (1) Log Occupational Income (2) SelfEmployment (3) Has Employees (4) 0.079∗∗∗ (0.026) -0.051 (0.058) 22,155 -0.075∗∗∗ (0.027) -0.023 (0.042) 16,472 -0.098∗∗∗ (0.033) 0.035 (0.030) 16,056 -0.054∗∗ (0.027) -0.020 (0.032) 16,056 Note: ∗ p < 0.1 ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border. Each column represents a separate spatial regression discontinuity regression, controlling for race, marital status, home ownership and household size in 1900. Treatment indicator, running variable, and controls are interacted with 1900 home ownership indicator. Observations weighted by inverse propensity weights (Bailey et al., 2020) and optimal balancing weights (Heuermann et al., 2019), running variable is distance to the border. Table 7: Effect of Fire and Kin Presence Unaffected Kin Affected Kin Unaffected Kin × Inside Affected Kin × Inside Inside Razed Area Observations Lives Outside San Francisco (1) Log Occupational Income (2) SelfEmployment (3) Has Employees (4) -0.021∗∗∗ (0.004) -0.020 (0.021) -0.014 (0.032) 0.004 (0.023) 0.075∗∗∗ (0.024) 15,529 0.014∗∗ (0.006) 0.138∗∗∗ (0.041) 0.006 (0.018) -0.138∗∗∗ (0.039) -0.096∗∗∗ (0.033) 11,471 0.022∗∗∗ (0.004) 0.103∗∗ (0.049) 0.030 (0.023) -0.108∗∗ (0.050) -0.101∗∗∗ (0.031) 11,183 0.030∗∗∗ (0.004) -0.001 (0.025) 0.026 (0.017) 0.004 (0.026) -0.059∗∗ (0.024) 11,183 Note: ∗ p < 0.1 ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to individuals linked from 1900 to 1910, with 1900 household location within 100m to 1600m distance to the fire border. Each column represents a separate spatial regression discontinuity regression, controlling for race, marital status, home ownership and household size in 1900. out in Treatment indicator, running variable, and controls are interacted with Rirs and Rirs which measure presence of unaffected and affected kin, respectively. Observations weighted by inverse propensity weights (Bailey et al., 2020) and optimal balancing weights (Heuermann et al., 2019), running variable is distance to the border. Table 8: Effect of Fire on Children’s Generation in 1940 Inside Razed Area Inside × Before Observations Years of Schooling (1) High School Degree (2) -0.521∗∗ (0.240) 0.008 (0.150) 13,119 -0.039 (0.034) -0.029 (0.033) 13,119 College SelfAttendance Employment (3) (4) -0.121∗∗∗ (0.042) 0.069∗ (0.039) 13,119 -0.005 (0.024) -0.018 (0.025) 11,684 Has Employees (5) -0.007 (0.007) 0.005 (0.016) 11,684 Note: ∗ p < 0.1 ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to sons of 1900 San Francisco residents, with 1900 household location within 100m to 1600m distance to the fire border. Spatial regression discontinuity regression controls for father’s for race, marital status, home ownership, household size in 1900, and child’s age in 1940. All covariates interacted with indicator for child born before 1906. Outcomes measured in 1940. Observations weighted by inverse propensity weights (Bailey et al., 2020) and optimal balancing weights (Heuermann et al., 2019), running variable is distance to the border. A Census Variables Table A1: Definition of Census Variables Variable IPUMS Variable Definition Available Years Number of children NCHILD Number of own children in household. Number of own children under age 5 in household. Number of own siblings in household. 1900 - 1940 Number of children NCHLT5 under age 5 Number of siblings NSIBS Age Marital status AGE MARST Duration of marital DURMARR status Married Based MARST Household size 1900 - 1940 1900 - 1940 Age in years. 1900 - 1940 Married (spouse present), mar- 1900 - 1940 ried (spouse absent), separated, divorced, widowed, never married. Duration of current marital status in 1900, 1910 years. on Equals 1 if person is married 1900 - 1940 (spouse present or absent), and 0 otherwise. Number of individuals in househld. 1900 - 1940 Race White RACE Based on RACE Born in US Based on BPL Born in California Based on BPL Mother born in US Based on MBPL Mother born in Cali- Based on MBPL fornia Father born in US Based on FBPL Father born in Cali- Based on FBPL fornia White, black, Chinese, Japanese. 1900 - 1940 Equals 1 if person is white, and 0 1900 - 1940 otherwise. Equals 1 if person was born in US, 1900 - 1940 and 0 otherwise. Equals 1 if person was born in Cali- 1900 - 1940 fornia, and 0 otherwise. Equals 1 if person’s mother was 1900 - 1940 born in US, and 0 otherwise. Equals 1 if person’s mother was 1900 - 1940 born in California, and 0 otherwise. Equals 1 if person’s father was born 1900 - 1940 in US, and 0 otherwise. Equals 1 if person’s father was born 1900 - 1940 in California, and 0 otherwise. Continued on next page Variable Table A1 – continued from previous page IPUMS Variable Definition Available Years School attendance SCHOOL 1900 - 1940 Literacy Years of education Respondent attended school during a specified period. Based on LIT Equals 1 if person know to read and write, and zero otherwise. Based on vari- Years of completed education. ables EDUC and HIGHGRADE 1900 - 1930 1940 Wife’s labor force Based on LABparticipation FORCE Indicates whether a person’s core- 1900 - 1940 siding wife participated in the labor force. Self employed Based on Equals 1 if person was recorded as 1910 - 1940 CLASSWKR an ”employer” or ”working on own account”, and zero otherwise. Employer Based on Equals 1 if person was recorded as 1910 - 1940 CLASSWKR an ”employer”, and zero otherwise. Months unem- MOUNEMP Number of months unemployed, 1900 ployed out of the last 12 months. Occupational in- OCCSCORE See section 3.1. 1900 - 1940 come score Home ownership OWNERSHP Mortgage MORTGAGE Middle initial Length first name Length middle name Length surname Indicates whether home was rented or owned. Indicates whether an owneroccupied housing unit was owned free and clear or was encumbered by a mortgage, loan, or other type of debt. 1900 - 1940 Indicates whether individual has a middle name or middle initial. Length of first name. Length of middle name. Length of surname. 1900 - 1940 1900 - 1920 1900 - 1940 1900 - 1940 1900 - 1940 B Enumeration Districts A Census enumeration district is a geographical unit that was intended to be covered by one census taker during the census period. Since these districts covered approximately the same number of households, they varied in size from just a few city blocks in urban areas up to a whole county in rural areas. Aside from the street name and street number, the enumeration district is the smallest geographical level that is reported in the Census. There are 319 enumeration districts in San Francisco for the 1900 Census. Of these, 11 consist entirely of hospitals, prisons, asylums or military bases, which I exclude from the analysis. The remaining 308 districts have about 1,100 inhabitants on average. I leverage the information on enumeration districts to localize each 1900 San Francisco resident within the city and with respect to the razed area. While I use the street name and street number reported by each household to infer the exact location of the dwelling, I check the validity of this localization using the enumeration district. I impute residence location coordinates with enumeration district centroid coordinates for 24 percent of all 1900 households. Census enumeration district maps for 1900 are not available in a GIS-usable format, so I employ images of the paper maps retrieved from the genealogy platform FamilySearch. The maps themselves show 1898 election districts; the 1900 Census enumeration district borders and numbers are later handwritten additions. Figure A3 shows a typical example. The maps are relatively well preserved although in some cases damage to the paper obscures parts of the map. I first georeference the photographs using ArcMap’s georeferencing tool. Then I create a polygon layer (i.e. a machine-readable district boundary map) by hand which includes all 308 enumeration districts of San Francisco. From this newly created boundary map I derive each district’s size, centroid coordinates, location relative to the razed area (inside, outside, intersect) and the centroid’s shortest distance to the fire area boundary. C Geolocation of 1900 Household Addresses There are 87,130 unique households who report living in San Francisco county in the 1900 Census. Three percent (2,606) of the households do not report a street name and the street number is missing for 522 additional households. The original street name strings have been cleaned by IPUMS and are therefore mostly complete and legible. In a first step, I geolocate all available addresses (street name and street number if available) using the geocoding tool in ArcMap. The software returns coordinates that correspond to the correct enumeration district for 59.3 percent of all households. For the remaining households, I update the street names that have changed between 1900 and today. In many cases, only parts of the streets have been renamed. This first step will therefore identify the location of households residing in the non-renamed parts of renamed steets. In the second step, I repeat the geolocation on the updated street names. An additional 1,090 addresses can now be geolocated in the correct enumeration district. This brings the total share of household addresses can be geolocated directly to 62.8 percent. Third, I impute coordinates of neighbors’ addresses if they reside in the same enumeration district as the household with the missing coordinates, and I recover the neighbors’ coordinates such that they correspond to the correct enumeration district in steps 1 or 2. I identify neighbors by the household’s serial number (Census variable SERIAL). Serial numbers are assigned by the order in which the Census takers enumerated households (Nelson, 2020). Therefore, an adjacent serial number usually corresponds to a household residing in the same building or the building next door. In some cases Census takers might have ”jumped” within the enumeration district depending on street layout. I assume that Census takers tried to minimized their walking distance and that jumps rare. Inspection of the addresses in the order of serial numbers confirms that Census takers tended to walk from house to house without skipping buildings. For the imputation I proceed via a simple algorithm: 1. Assign the coordinates of one the immediately neighbors (starting with the neighbor identified by serial number + 1, then serial number - 1) if there is an exact match on the street name. Imputed addresses: 4,455. 2. Impute coordinates for all missing households with an exact match for enumeration district and address (street name number) using coordinates obtained from step 1. Imputed addresses: 1,223. 3. Assign coordinates of one of the immediately adjacent neighbors regardless of street name match. Assign coordinates of one of the two-doors-away neighbors (serial number +/- 2) if there is a match for the stemmed street name37 . Imputed addresses: 5,512. The final step assigns the remaining 23.9 percent of households (20,103) the coordinates of the geometric centroid of their enumeration district. The average enumeration district is 0.63 square kilometers (about 156 acres), with enumeration districts further away from the center (and the razed area) being considerably larger. Among all enumeration districts within 1km of the fire border, the average size is only 0.14 square kilometers (34 acres). Imputing enumeration district centroid coordinates is therefore a decent approximation. I exclude households in enumeration districts that are intersected by the fire border from this imputation. This pertains to 5,001 households, so that the final sample of households includes coordinates for 94 percent of the observations (79,048). Figure 2 shows all households locations, enumeration district borders and geographic centroids, and the border of the razed district. D Weights D.1 Inverse Propensity Weights Following Bailey et al. (2020), I construct inverse propensity weights (IPW) in three simple steps. First, I estimate a probit model for the dependent variable that indicates whether in individual was linked from 1900 to the year of interest, for example 1910. The covariates of this model are – Indicator for middle initial, length of first, middle, last names, – quadratic in age and month of birth, – presence of siblings, number of siblings. Second, I predict the conditional probability of being linked, pi , from the results of the probit. Third, I construct the weight as IP Wi = 37 1 − pi pi I stem street names by retaining the first word (e.g. ”Octavia” in ”Octavia Street” or ”Octavia St”.) (6) D.2 Optimal Balancing Weights Following Heuermann et al. (2019), I reweight observations around each border segment according to OBWb = nb,inside × nb,outside nb,inside + nb,outside (7) where nb,inside and nb,outside are the number of observations in each border segment b, inside and outside the razed area, respectively. E Historical Maps Figure A1: The Course of the 1906 Fire in San Francisco Figure A2: San Francisco Reconstruction Map, Punnet Brothers (1908) Figure A3: Photo of a Map Showing 1900 San Francisco Enumeration Districts 46 - 61 F Additional Results Table A2: Placebo RD with Pre-Treatment Characteristics (Sample: Linked at Least Once) Age White Born in the US Born in California Mother Born in the US Father Born in the US Mother Born in California Father Born in California Married Number of Own Children Number of Own Children Below Age 5 Number of Siblings in Household Household Size Household Head Children in Labor Force Wife in Labor Force Kin Presence Middle Initial Length First Name Length Last Name Home Ownership Mortgage Log Occupational Income Score Months Unemployed in 1899 In Labor Force School Attendance Illiterate Observations Inside Razed District Standard Error 37176 37176 37176 37176 37176 37176 37176 37176 37176 37176 37176 37176 37176 37176 3224 11927 37176 37176 37176 37176 32917 7328 23396 29524 31867 37176 37176 -0.005 0.011 -0.036 -0.027 -0.031 0.006 -0.012 0.002 0.043∗∗∗ 0.057 0.022∗ -0.167∗ -1.264∗∗ 0.076∗∗∗ 0.033 0.022 -0.090∗ 0.025 -0.079 -0.195∗ -0.144∗∗∗ -0.014 -0.002 0.088 0.018 -0.008 -0.001 (0.453) (0.010) (0.031) (0.031) (0.035) (0.032) (0.009) (0.006) (0.011) (0.045) (0.012) (0.087) (0.605) (0.016) (0.023) (0.024) (0.049) (0.051) (0.082) (0.106) (0.045) (0.049) (0.025) (0.121) (0.025) (0.019) (0.005) Note: ∗ p < 0.1 ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to individuals linked from 1900 to at least one of the years 1910, 1920, 1930, or 1940, with 1900 household location within 100m to 1600m distance to the fire border. Each row represents a separate placebo spatial regression discontinuity regression. Observations weighted by inverse propensity weights (Bailey et al., 2020) and optimal balancing weights (Heuermann et al., 2019), running variable is distance to the border. Dependent variables measured in 1900. No individual level controls. Table A3: Effect of Fire on Children’s Generation in 1940, Additional Results Inside Razed Area Inside × Before Observations Migrated (SF) (1) Migrated (CA) (2) Urban (3) Log Occscore (4) 0.021 (0.016) 0.008 (0.008) 13,282 -0.003 (0.026) -0.004 (0.040) 13,282 0.013 (0.011) -0.012 (0.017) 13,282 -0.042 (0.030) 0.022 (0.035) 11,722 Log Wages (5) -0.027 (0.075) -0.010 (0.071) 9,669 Note: ∗ p < 0.1 ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors allow for arbitrary spatial clustering with a distance cutoff of 0.4 km (in parentheses). Sample restricted to sons of 1900 San Francisco residents, with 1900 household location within 100m to 1600m distance to the fire border. Spatial regression discontinuity regression controls for father’s for race, marital status, home ownership, household size in 1900, and child’s age in 1940. All covariates interacted with indicator for child born before 1906. Outcomes measured in 1940, wages only include non-selfemployed individuals. Observations weighted by inverse propensity weights (Bailey et al., 2020) and optimal balancing weights (Heuermann et al., 2019), running variable is distance to the border. Figure A4: Location of Households in 1900 (Spatial RD Sample Manually Selected), on Modern Map Note: Source: 1900 Census, Author’s calculations