SSIV Construction and Robustness Checks

B SSIV Construction and Robustness Checks B.1 Leveraging Push and Pull Factors for Identification The sorting patterns in Appendix Table A.3 highlight the importance of strategies for addressing threats to causal identification in our county-level analyses in Section 4 and elsewhere. To address residual sorting biases, we control directly for the potential confounders in that table. Insofar as these confounders are leading to bias, our estimates should be sensitive to their inclusion in Table 4, and they are not. This suggests that our SSIV provides robust and excludable sources of identifying variation through the combination of push-factor-based shifts from Appendix Table A.1 with the predetermined shares underlying chain migration pathways from Appendix Table A.3. This Appendix provides additional details and results related to this SSIV strategy. Shift-Share IV with Push Factors. Following Bazzi et al. (forthcoming) and building on Boustan (2010) and Derenoncourt (2022), we construct a SSIV for Southern white migrants outside the South in the early postbellum era. To do this, we use the linked Census records from Abramitzky et al. (2020) to approximate Southern white outmigration from Southern counties for Census periods 1870–1880 and 1880–1900. We then calculate for each Census period through τ ∈ {1880, 1900}: Southern white migrantso,τ  D  X # white men in o in τ −1 linked to d in τ = × Southern whiteso,τ−1 , # white men in o in τ −1 linked to Census τ d=1 (B.1) where o indicates Southern origin county, d indicates non-Southern destination county, and where the rightmost term, Southern whiteso,τ −1 , is based on the complete-count Census. In Bazzi et al. (forthcoming), we validate the accuracy of this approach for the mass Southern outflows during the Great Migration of the 20th century. Taking a similar approach here for the late 19th century, we see, in Appendix Figure B.1, that the estimated stocks of Southern white migrants from the linked Census closely approximate those based on the complete-count Census. We then predict decade-specific shifts using zero-stage regressions of equation (1) in the paper, which relates the outcome from equation (B.1) to origin-county push factors discussed in Appendix A: Southern white migrantsoτ = α + push0o,τ −1 βτ + φpopulationo,τ −1 + εoτ , where pusho,τ −1 is the vector of Southern county push factors. Columns 1–2 of Appendix Table A.1 shows estimates for the set of push factors selected by a LASSO algorithm from a set of linear predictors. As in Bazzi et al. (forthcoming), we also include square terms and cross-term interactions of each predictor for the purposes of constructing our predicted shifts. V We then sum these decade- and county-specific predictions, Southern white migrantsoτ , within Southd j,1870−1900 , which denote the preern sending states to construct the aggregate predicted shifts, ∆M dicted change in the number of whites from Southern state j living outside the South between 1870 and 1900. These shifts are interacted with the cross-sectional shares of white migrants from Southern origin state j living in non-Southern county c in the 1870 complete-count Census, which we denote πjc,1870 . Appendix Figure B.2 maps the overall population share of Southern-born whites in each B-1 county outside the South. Together, these predict the stock of Southern white migrants in 1900 as P d j,1870−1900 , which we scale by the 1870 county population to generate the Zc,1900 = Jj=1 πjc,1870 ∆M SSIV for % Southern W hitesc,1900 . Figure B.1: Validating the Linked-Sample Estimates of Southern White Migration from 1870–80 # Southern migrants (full-count census) 100 200 300 400 (a) Migration 0 ρ =.966 0 50 # Southern migrants (linked census) from 1880–1900 # Southern migrants (full-count census) 100 200 300 (b) Migration 100 0 ρ =.985 0 20 40 60 # Southern migrants (linked census) 80 Notes: Panel (a) plots the number of white men ages 10 and older from a Southern state (based on the BPL variable) in a non-Southern county using the 1880 full count Census against the number of white men from the same Southern state in the same non-Southern county who were successfully linked from 1870 to 1880 using the linked Census. Panel (b) plots the number of white men ages 20 and older from a Southern state (based on the BPL variable) in a non-Southern county using the 1900 full count Census against the number of white men from the same Southern state in the same non-Southern county who were successfully linked from 1880 to 1900 using the linked Census. B-2 Table B.1: Push Factors for Southern Former Slaveholder Migrants, 1870–1900 Southern Former Slaveholder Migrant Outflows (1) (2) (3) Dependent Variable: % Black population Manufacturing output per capita % from slaveholding households (1860) % from Confederate Army households (1860) Any Civil War battles in county? Tobacco county Agricultural potential Population size in sending counties -0.682∗∗∗ (0.074) 0.008 (0.054) 0.791∗∗∗ (0.135) -1.000∗∗∗ (0.119) 4.823 (3.892) 4.591∗∗ (2.254) -48.735∗∗∗ (8.555) 0.004∗∗∗ (0.000) -1.067∗∗∗ (0.131) 0.109 (0.130) 1.022∗∗∗ (0.238) -1.456∗∗∗ (0.214) 6.803 (6.131) 13.511∗∗∗ (4.408) -17.640 (22.373) 0.003∗∗∗ (0.000) -11.775∗∗ (4.835) -1.095∗∗∗ (0.147) 0.128 (0.095) 0.815∗∗∗ (0.136) -1.221∗∗∗ (0.125) 5.635 (3.595) 9.176∗∗∗ (2.645) -33.621∗∗ (13.385) 0.003∗∗∗ (0.000) -4.981∗ (2.714) -0.351 (0.292) -0.019 (0.053) 0.253∗ (0.149) 1870–1880 1880–1900 1,019 25.23 0.561 1,017 35.12 0.366 1870–1900 Yes 2,036 30.17 0.423 Cotton county Average manufacturing wages % Vote share for Breckinridge (1860) % Slaves (1860) Period Year FE Observations Outcome mean Adj. R2 Notes: Estimates of a version of equation (1), which regresses a measure of Southern white former slaveholder migrant outflows from Southern counties on various observable origin county characteristics. These outflow measures from each origin county to non-Southern counties are based on linked Census records, which track (white male) migrants across Census periods for decades following the Civil War from 1870– 1900. All columns use a linear LASSO specification to select optimal sets of covariates for each Census period. Column 3 pools data across years and estimates a version with year fixed effects. Sample counties include those in the twelve former Confederate states plus Oklahoma. Robust standard errors in parentheses. Significance levels are denoted by * p < 0.10, ** p < 0.05, *** p < 0.01. To ensure sufficient orthogonal variation for identifying the distinct effects of (i) Southern white migrants overall and of (ii) former slaveholder migrants (e.g., as in Table 8), we construct a separate SSIV for the latter, using an alternative version of equation (B.1) above based on linked Southern former slaveholders in Southern counties, combined with linked outmigration flows of former slaveholders from each Southern county o to all non-Southern counties: slaveholder migrantso,τ =  D  X # slaveholders in o in τ −1 linked to d in τ # slaveholders in o in τ −1 linked to τ   # slaveholders in o in τ −1 linked to 1860 × # white men in o in τ −1 linked to 1860 d=1 × Southern whiteso,τ−1 , B-3 (B.2) where d indicates non-Southern destination counties and the last term, Southern whiteso,τ −1 , is based on the complete-count Census in the previous Census period. As above, we predict decade-specific shifts using zeroth stage regressions, which relate the measure produced from equation (B.2) to the origin-county push factors discussed in Appendix A, again using a LASSO algorithm selecting from all linear, square, and cross-term interactions of those push factors (see columns 1–2 of Appendix Table B.1 for zeroth-stage estimates from a simple linear approach). The output from these flexible LASSO regressions are then summed within Southern sending states to construct aggregate predicted shifts for 1870–1900. Finally, these shifts are interacted with cross-sectional shares of migrants, from Southern origin state j linked to non-Southern county c in the 1870 complete-count Census, and then scaled by 1870 county population to yield an SSIV for former slaveholder migrants in 1900. Figure B.2: Mapping Southern White Migrants Outside the South in 1870 Notes: This figure maps county-level population shares of white individuals born in the South and residing outside the South in 1870, according to the complete-count Census. B-4 B.2 SSIV Robustness Checks Table B.2: Further Restricting Confederate Leader Name Inputs in Table 2 Dependent Variable: Leader Name Inputs: (1) % Southern Whites, 1900 Estimator State FE Baseline controls 1870 share control Sorting controls Observations Outcome mean F-statistic Anderson-Rubin, p-val KP Underident., p-val Any Confederate Memorials ≥ 2 Word Name Inputs ≥ 3 Word Name Inputs (2) (3) (4) (5) (6) 0.015∗∗∗ (0.006) 0.025∗∗∗ (0.008) 0.021∗∗∗ (0.007) 0.011∗ (0.006) 0.027∗∗∗ (0.009) 0.022∗∗∗ (0.008) IV Yes Yes IV Yes Yes Yes IV Yes Yes IV Yes Yes Yes 1,701 0.15 30.6 0.01 0.00 1,701 0.15 15.0 0.00 0.00 IV Yes Yes Yes Yes 1,699 0.15 15.3 0.02 0.00 1,701 0.10 30.6 0.07 0.00 1,701 0.10 15.0 0.00 0.00 IV Yes Yes Yes Yes 1,699 0.10 15.3 0.01 0.00 Notes: Regressions of Confederate memorials (coded as any Confederacy leader-inspired monuments as well as matched place names, street names or school names) on the share of Southern whites in 1900 in non-Southern counties (sample mean of 2.2%). Excluded Southern counties are those belonging to states of the former Confederacy and Oklahoma. Columns 1–3 identify Confederate location names using only the subset of Confederate leader name inputs with at least two words, while columns 4–6 restrict further to those with at least three words (including initials). All columns instrument the share of Southern whites using a shift-share instrument based on the 1870 cross-sectional distribution of Southern whites and the predicted change in the Southern white population living outside the South from 1870 to 1900. The latter is generated via a set of flexible LASSO regressions (see equation (1)). Baseline controls include log county population in 1870 and log county area (in square miles). Additional sorting controls include cotton, tobacco, and overall agricultural potential, Black shares in 1870, slave shares in 1860, Union Army enlistment and mortality rates, Breckinridge vote shares in 1860, and dummies for on the frontier in 1860 and never on the frontier, based on Bazzi et al. (2020). All continuous baseline and sorting controls are entered flexibly using quadratic terms. All regressions control for the share of Southern whites in 1870 and include state fixed effects. Standard errors are clustered at the 60×60 square-mile grid cell level, following the approach of Bester et al. (2011). Significance levels are denoted by * p < 0.10, ** p < 0.05, *** p < 0.01. Table B.3: Breaking Out Confederate Memorialization in Table 2 Dependent Variable: % Southern Whites, 1900 Estimator State FE Baseline controls 1870 share control Sorting controls Observations Outcome mean F-statistic Anderson-Rubin, p-val KP Underident., p-val Any Confederate Monuments (1) (2) Any Confederate Place Names (3) (4) Any Confederate Street Names (5) (6) 0.023∗∗∗ (0.008) 0.021∗∗∗ (0.008) 0.014∗∗ (0.007) 0.012∗ (0.007) 0.025∗∗ (0.010) 0.024∗∗ (0.010) IV Yes Yes Yes IV Yes Yes Yes Yes 1,699 0.07 15.3 0.01 0.00 IV Yes Yes Yes IV Yes Yes Yes Yes 1,699 0.09 15.3 0.15 0.00 IV Yes Yes Yes IV Yes Yes Yes Yes 1,699 0.16 15.3 0.01 0.00 1,701 0.07 15.0 0.00 0.00 1,701 0.09 15.0 0.10 0.00 1,701 0.16 15.0 0.00 0.00 Notes: Regressions of (i) Confederate monuments, (ii) Confederate leader-inspired place names, and (iii) Confederate leader-inspired street names on the share of Southern whites in 1900 in non-Southern counties (sample mean of 2.2%). Excluded Southern counties are those belonging to states of the former Confederacy and Oklahoma. All columns instrument the share of Southern whites using a shift-share instrument based on the 1870 cross-sectional distribution of Southern whites and the predicted change in the Southern white population living outside the South from 1870 to 1900. The latter is generated via a set of flexible LASSO regressions (see equation (1)). Baseline controls include log county population in 1870 and log county area (in square miles). Additional sorting controls include cotton, tobacco, and overall agricultural potential, Black shares in 1870, slave shares in 1860, Union Army enlistment and mortality rates, Breckinridge vote shares in 1860, and dummies for on the frontier in 1860 and never on the frontier, based on Bazzi et al. (2020). All continuous baseline and sorting controls are entered flexibly using quadratic terms. All regressions control for the share of Southern whites in 1870 and include state fixed effects. Standard errors are clustered at the 60×60 square-mile grid cell level, following the approach of Bester et al. (2011). Significance levels are denoted by * p < 0.10, ** p < 0.05, *** p < 0.01. B-5 Table B.4: Placebo: Using White Lynchings in Table 3 Any Lynchings of Whites (2) (3) Dependent Variable: (1) % Southern Whites, 1900 Estimator State FE Baseline controls 1870 share control Sorting controls Observations Outcome mean Adj. R2 F-statistic Anderson-Rubin, p-val KP Underident., p-val (4) 0.000 (0.001) 0.001 (0.003) 0.001 (0.005) 0.001 (0.005) OLS Yes Yes IV Yes Yes IV Yes Yes Yes 1,701 0.02 0.00 1,701 0.02 1,701 0.02 IV Yes Yes Yes Yes 1,699 0.02 30.6 0.78 0.00 15.0 0.80 0.00 15.3 0.85 0.00 Notes: Regressions of any recorded lynchings of whites on the share of Southern whites in 1900 in non-Southern counties (sample mean of 2.2%). Excluded Southern counties are those belonging to states of the former Confederacy and Oklahoma. Columns 2–4 instrument the share of Southern whites using a shift-share instrument based on the 1870 cross-sectional distribution of Southern whites and the predicted change in the Southern white population living outside the South from 1870 to 1900. The latter is generated via a set of flexible LASSO regressions (see equation (1)). Baseline controls include log county population in 1870 and log county area (in square miles). Additional sorting controls in column 4 include cotton, tobacco, and overall agricultural potential, Black shares in 1870, slave shares in 1860, Union Army enlistment and mortality rates, Breckinridge vote shares in 1860, and dummies for on the frontier in 1860 and never on the frontier, based on Bazzi et al. (2020). All continuous baseline and sorting controls are entered flexibly using quadratic terms. Columns 3–4 control for the share of Southern whites in 1870, and all regressions include state fixed effects. Standard errors are clustered at the 60×60 square-mile grid cell level, following the approach of Bester et al. (2011). Significance levels are denoted by * p < 0.10, ** p < 0.05, *** p < 0.01. Figure B.3: Sensitivity to Sample Changes (a) Dropping individual sending states (b) Dropping individual receiving states Arizona California Colorado Connecticut Delaware District of Columbia Idaho Illinois Indiana Iowa Kansas Kentucky Maine Maryland Massachusetts Michigan Minnesota Missouri Montana Nebraska Nevada New Hampshire New Jersey New Mexico New York North Dakota Ohio Oregon Pennsylvania Rhode Island South Dakota Utah Vermont Washington West Virginia Wisconsin Wyoming Alabama Arkansas Florida Georgia Louisiana Mississippi North Carolina Oklahoma South Carolina Tennessee Texas Virginia 0 .2 .4 .6 0 .1 .2 .3 .4 Notes: Coefficients from IV regressions of Confederate cultural activity after 1900 (score 0–4) on the share of Southern whites in 1900 in non-Southern counties. All regressions include the set of controls from columns 3 and 6 in Tables 2 and 3. Estimates are compared to one with all states included, which is reported in the solid vertical red line in the respective graphs. Panel a excludes Southern sending states one-by-one when constructing the 1900 share of Southern whites living outside the South in a given non-Southern county c as well as the instrumental variable, with the excluded sending state reported on the vertical axis. The instrumental variables regressions instrument the share of Southern whites using a shift-share instrument based on the 1870 cross-sectional distribution of Southern whites and the predicted change in the Southern white population living outside the South from 1870 to 1900. Panel b excludes receiving states one-by-one where the excluded state is reported on the vertical axis. The dashed red line marks zero. Error bars represent 90% confidence intervals. Standard errors are clustered at the 60×60 square-mile grid cell level, following the approach of Bester et al. (2011). B-6 Figure B.4: SSIV Using Random Placebo Shifts Placebo estimate coefficient 20 10 0 -10 -20 0 200 400 600 800 1000 Simulation Notes: Coefficients from IV regressions of Confederate cultural activity after 1900 (score 0–4) on the share of Southern whites in 1900 in all non-Southern counties. The share of Southern whites in 1900 is instrumented using a shift-share instrument based on the 1870 crosssectional distribution of Southern whites and a randomly generated shift. The random shift was generated based on a normal distribution with mean zero and variance five as in Adao et al. (2019). The figure shows the coefficients and 95% confidence intervals from instrumental variables regressions where the instrument was generated with 1,000 random shifts. All regressions include the set of controls from columns 3 and 6 in Tables 2 and 3. Standard errors are clustered at the 60×60 square-mile grid cell level, following the approach of Bester et al. (2011). B-7