Closed taransamarth closed 1 year ago
In Florida, per Art. III, sec. 20 of the state constitution, districts must:
We enforce a maximum population deviation of 0.5%.
Data for Florida comes from the ALARM Project's 2010 Redistricting Data Files. We obtain the 2010 Florida Congressional map from All About Redistricting.
We estimate CVAP populations with the cvap R package.
We sample 35,000 districting plans for the full state of Florida, thinned down to a set of 7,500. To appropriately district the entire state, we split the state into three regions, simulate two of the regions (North and the Miami area, as defined below) separately, and then simulate districts in the remainder of the state. In all simulations, we constrain county and municipality splits. Since some county populations are greater than the target population for one Congressional district, we create pseudocounties where needed.
Regional clustering: We split Florida into the following three regions:
We simulate the Miami metropolitan area and Northern Florida independently. Since each cluster has leftover population, we include a constraint to encourage unassigned precincts to be set along each cluster's boundary with Central Florida so those precincts can be assigned to contiguous districts in the final simulation step.
Simulating Miami: We simulate four SMC runs with 60,000 maps each for the Miami metropolitan area. To encourage Black and Hispanic opportunity districts, we apply Gibbs constraints in the simulation. We then subset down the plans to those where there exists one district with a Black voting-age population (BVAP) share of at least .4 and another district with a BVAP share of at least .25. From this set, we randomly sample 35,000 plans.
Simulating Northern Florida: We simulate two SMC runs with 40,000 maps each for Northern Florida. To encourage Black and Hispanic opportunity districts, we apply Gibbs constraints in the simulation. We then subset down the plans to those where at least one district has a BVAP share of .25 or greater. From this set, we randomly sample 35,000 plans.
Simulating Central Florida: Using the unassigned areas from the partial SMC simulations for Miami and Northern Florida, we simulate two SMC runs with 35,000 plans each for Central Florida. We apply Gibbs constraints to encourage Black and Hispanic opportunity districts. We then thin these maps down to the final set of 5,000.
SMC: 35,000 sampled plans of 7 districts on 2,312 units
`adapt_k_thresh`=0.985 • `seq_alpha`=0.65
`est_label_mult`=1 • `pop_temper`=0.02
Plan diversity 80% range: 0.63 to 0.91
R-hat values for summary statistics:
ndv hvap bvap hcvap bcvap dem16 dem18 dem20
1.014659 1.012807 1.008052 1.015582 1.006766 1.014170 1.019930 1.017734
Sampling diagnostics for SMC run 1 of 4 (60,000 samples)
Eff. samples (%) Acc. rate Log wgt. sd Max. unique Est. k
Split 1 40,573 (67.6%) 8.6% 0.24 37,802 (100%) 10
Split 2 30,484 (50.8%) 9.8% 0.57 33,174 ( 87%) 8
Split 3 30,251 (50.4%) 14.3% 0.67 30,247 ( 80%) 5
Split 4 24,266 (40.4%) 20.7% 0.61 29,689 ( 78%) 3
Split 5 20,783 (34.6%) 20.2% 0.67 28,168 ( 74%) 3
Split 6 3,436 (5.7%) 53.4% 0.74 26,032 ( 69%) 7
Resample 1,095 (1.8%) NA% 6.75 6,124 ( 16%) NA *
Sampling diagnostics for SMC run 2 of 4 (60,000 samples)
Eff. samples (%) Acc. rate Log wgt. sd Max. unique Est. k
Split 1 40,580 (67.6%) 7.2% 0.24 37,986 (100%) 12
Split 2 31,035 (51.7%) 11.1% 0.56 33,154 ( 87%) 7
Split 3 30,939 (51.6%) 14.3% 0.67 30,339 ( 80%) 5
Split 4 27,997 (46.7%) 16.4% 0.60 29,771 ( 78%) 4
Split 5 20,032 (33.4%) 20.4% 0.66 28,388 ( 75%) 3
Split 6 3,600 (6.0%) 53.8% 0.74 26,074 ( 69%) 7
Resample 1,151 (1.9%) NA% 6.72 6,172 ( 16%) NA *
Sampling diagnostics for SMC run 3 of 4 (60,000 samples)
Eff. samples (%) Acc. rate Log wgt. sd Max. unique Est. k
Split 1 40,609 (67.7%) 8.6% 0.24 37,973 (100%) 10
Split 2 30,977 (51.6%) 13.0% 0.56 33,390 ( 88%) 6
Split 3 29,282 (48.8%) 17.5% 0.67 30,214 ( 80%) 4
Split 4 25,866 (43.1%) 20.6% 0.61 29,708 ( 78%) 3
Split 5 17,893 (29.8%) 25.9% 0.65 28,434 ( 75%) 2
Split 6 3,371 (5.6%) 53.9% 0.76 25,963 ( 68%) 7
Resample 1,009 (1.7%) NA% 6.72 6,057 ( 16%) NA *
Sampling diagnostics for SMC run 4 of 4 (60,000 samples)
Eff. samples (%) Acc. rate Log wgt. sd Max. unique Est. k
Split 1 40,598 (67.7%) 9.5% 0.24 37,841 (100%) 9
Split 2 31,106 (51.8%) 9.8% 0.56 33,226 ( 88%) 8
Split 3 30,367 (50.6%) 14.3% 0.67 30,022 ( 79%) 5
Split 4 26,764 (44.6%) 16.5% 0.62 29,747 ( 78%) 4
Split 5 20,939 (34.9%) 20.3% 0.68 28,121 ( 74%) 3
Split 6 3,518 (5.9%) 60.2% 0.73 25,924 ( 68%) 6
Resample 1,390 (2.3%) NA% 6.66 5,979 ( 16%) NA *
• Watch out for low effective samples, very low acceptance rates (less than 1%), large std. devs. of
the log weights (more than 3 or so), and low numbers of unique plans. R-hat values for summary
statistics should be between 1 and 1.05.
SMC: 35,000 sampled plans of 7 districts on 2,273 units
`adapt_k_thresh`=0.985 • `seq_alpha`=0.5
`est_label_mult`=1 • `pop_temper`=0
Plan diversity 80% range: 0.38 to 0.60
R-hat values for summary statistics:
hvap bvap hcvap bcvap dem16 dem18 dem20
1.000044 1.000348 1.000091 1.000374 1.000111 1.000022 1.000149
Sampling diagnostics for SMC run 1 of 2 (40,000 samples)
Eff. samples (%) Acc. rate Log wgt. sd Max. unique Est. k
Split 1 35,180 (88.0%) 13.8% 0.64 25,290 (100%) 7
Split 2 36,481 (91.2%) 13.0% 0.72 23,529 ( 93%) 9
Split 3 34,206 (85.5%) 12.8% 0.76 23,048 ( 91%) 6
Split 4 28,332 (70.8%) 14.3% 0.65 22,865 ( 90%) 4
Split 5 24,311 (60.8%) 15.8% 0.75 20,427 ( 81%) 3
Split 6 20,547 (51.4%) 15.8% 0.67 19,562 ( 77%) 4
Resample 15,624 (39.1%) NA% 10.02 17,022 ( 67%) NA
Sampling diagnostics for SMC run 2 of 2 (40,000 samples)
Eff. samples (%) Acc. rate Log wgt. sd Max. unique Est. k
Split 1 35,164 (87.9%) 13.7% 0.64 25,348 (100%) 7
Split 2 36,510 (91.3%) 19.3% 0.72 23,529 ( 93%) 6
Split 3 34,088 (85.2%) 18.7% 0.77 22,980 ( 91%) 4
Split 4 28,239 (70.6%) 14.3% 0.64 22,835 ( 90%) 4
Split 5 24,399 (61.0%) 15.6% 0.74 20,351 ( 80%) 3
Split 6 20,639 (51.6%) 15.7% 0.70 19,754 ( 78%) 4
Resample 13,986 (35.0%) NA% 10.02 17,045 ( 67%) NA
• Watch out for low effective samples, very low acceptance rates (less than 1%), large std. devs. of
the log weights (more than 3 or so), and low numbers of unique plans. R-hat values for summary
statistics should be between 1 and 1.05.
SMC: 5,000 sampled plans of 27 districts on 9,435 units
`adapt_k_thresh`=0.985 • `seq_alpha`=0.5
`est_label_mult`=1 • `pop_temper`=0
Plan diversity 80% range: 0.55 to 0.74
R-hat values for summary statistics:
pop_overlap total_vap total_cvap plan_dev comp_edge comp_polsby
1.004815 1.005773 1.003861 1.004369 1.002839 1.018513
pop_white pop_black pop_hisp pop_aian pop_asian pop_nhpi
1.009303 1.001235 1.002639 1.005529 1.003806 1.008631
pop_other pop_two vap_white vap_black vap_hisp vap_aian
1.004640 1.002148 1.008781 1.001248 1.002214 1.003881
vap_asian vap_nhpi vap_other vap_two cvap_white cvap_black
1.004518 1.005871 1.003650 1.004882 1.011704 1.002023
cvap_hisp cvap_asian cvap_aian cvap_nhpi cvap_two cvap_other
1.001592 1.009570 1.000306 1.016361 1.002289 1.004326
pre_16_rep_tru pre_16_dem_cli pre_20_rep_tru pre_20_dem_bid uss_16_rep_rub uss_16_dem_mur
1.005602 1.006264 1.004550 1.005938 1.004472 1.006337
uss_18_rep_sco uss_18_dem_nel gov_18_rep_des gov_18_dem_gil atg_18_rep_moo atg_18_dem_sha
1.004963 1.006274 1.005555 1.006227 1.005603 1.006086
adv_16 adv_18 adv_20 arv_16 arv_18 arv_20
1.006295 1.006244 1.005938 1.004653 1.005771 1.004550
county_splits muni_splits ndv nrv ndshare e_dvs
1.001248 1.000740 1.006285 1.005382 1.007339 1.007269
e_dem pbias egap
1.000128 1.003726 1.000083
Sampling diagnostics for SMC run 1 of 2 (35,000 samples)
Eff. samples (%) Acc. rate Log wgt. sd Max. unique Est. k
Split 1 33,348 (95.3%) 19.4% 0.43 22,131 (100%) 8
Split 2 31,655 (90.4%) 28.7% 0.49 21,941 ( 99%) 5
Split 3 30,865 (88.2%) 27.1% 0.57 21,444 ( 97%) 5
Split 4 30,227 (86.4%) 31.1% 0.63 21,359 ( 97%) 4
Split 5 29,500 (84.3%) 24.3% 0.67 21,155 ( 96%) 5
Split 6 29,226 (83.5%) 27.7% 0.71 20,928 ( 95%) 4
Split 7 28,801 (82.3%) 32.2% 0.73 20,878 ( 94%) 3
Split 8 28,711 (82.0%) 29.6% 0.74 20,684 ( 93%) 3
Split 9 28,901 (82.6%) 34.1% 0.75 20,578 ( 93%) 2
Split 10 28,746 (82.1%) 14.8% 0.76 20,164 ( 91%) 5
Split 11 28,987 (82.8%) 19.8% 0.74 19,594 ( 89%) 3
Split 12 28,794 (82.3%) 22.0% 0.73 19,084 ( 86%) 2
Split 13 29,110 (83.2%) 12.8% 0.72 18,195 ( 82%) 3
Split 14 29,293 (83.7%) 2.3% 0.70 16,690 ( 75%) 6
Resample 20,778 (59.4%) NA% 0.70 19,405 ( 88%) NA
Sampling diagnostics for SMC run 2 of 2 (35,000 samples)
Eff. samples (%) Acc. rate Log wgt. sd Max. unique Est. k
Split 1 33,323 (95.2%) 17.3% 0.43 22,261 (101%) 9
Split 2 31,684 (90.5%) 28.4% 0.49 21,841 ( 99%) 5
Split 3 31,028 (88.7%) 23.2% 0.56 21,402 ( 97%) 6
Split 4 30,469 (87.1%) 31.1% 0.62 21,304 ( 96%) 4
Split 5 29,715 (84.9%) 20.6% 0.66 21,226 ( 96%) 6
Split 6 29,138 (83.3%) 19.5% 0.71 20,937 ( 95%) 6
Split 7 28,898 (82.6%) 25.8% 0.73 20,885 ( 94%) 4
Split 8 28,641 (81.8%) 29.7% 0.75 20,713 ( 94%) 3
Split 9 28,399 (81.1%) 34.8% 0.77 20,379 ( 92%) 2
Split 10 28,580 (81.7%) 18.3% 0.77 20,033 ( 91%) 4
Split 11 28,776 (82.2%) 19.9% 0.76 19,618 ( 89%) 3
Split 12 28,865 (82.5%) 21.9% 0.74 19,242 ( 87%) 2
Split 13 28,890 (82.5%) 17.6% 0.74 18,116 ( 82%) 2
Split 14 29,114 (83.2%) 6.5% 0.71 16,505 ( 75%) 2
Resample 20,381 (58.2%) NA% 0.71 19,208 ( 87%) NA
• Watch out for low effective samples, very low acceptance rates (less than 1%), large std. devs. of
the log weights (more than 3 or so), and low numbers of unique plans. R-hat values for summary
statistics should be between 1 and 1.05.
TODO
lines from the template code have been removedenforce_style()
to format my coderedist_map
and redist_plans
objects, and summary statistics) have been edited@christopherkenny
@taransamarth, congrats. @CoryMcCartan, @tylersimko, and I are all in agreement. Good to go!
Redistricting requirements
In Florida, per Art. III, sec. 20 of the state constitution, districts must:
Algorithmic Constraints
We enforce a maximum population deviation of 0.5%.
Data Sources
Data for Florida comes from the ALARM Project's 2010 Redistricting Data Files. We obtain the 2010 Florida Congressional map from All About Redistricting.
Pre-processing Notes
We estimate CVAP populations with the cvap R package.
Simulation Notes
We sample 35,000 districting plans for the state of Florida, thinned down to a set of 7,500. To appropriately district the entire state, we split the state into three regions, simulate two of the regions (North and South, as defined below) separately, and then simulate districts in the remainder of the state. In all simulations, we constrain county and municipality splits. Since some county populations are greater than the target population for one Congressional district, we create pseudocounties where needed.
Regional clustering: We split Florida into the following three regions:
We simulate the Miami metropolitan area and Northern Florida independently. Since each cluster has leftover population, we include a constraint to encourage unassigned precincts to be set along each cluster's boundary with Central Florida so those precincts can be assigned to contiguous districts in the final simulation step.
Simulating Miami: We simulate 60,000 maps for the Miami metropolitan area. To encourage Black and Hispanic opportunity districts, we apply Gibbs constraints in the simulation. We then subset down the plans to those where there exists one district with a Black voting-age population (BVAP) share of at least .4 and another district with a BVAP share of at least .25. From this set, we randomly sample 35,000 plans.
Simulating Northern Florida: We simulate 40,000 maps for Northern Florida. To encourage Black and Hispanic opportunity districts, we apply Gibbs constraints in the simulation. We then subset down the plans to those where at least one district has a BVAP share of .25 or greater. From this set, we randomly sample 35,000 plans.
Simulating Central Florida: Using the unassigned areas from the partial SMC simulations for Miami and Northern Florida, we simulate 35,000 plans for Central Florida. We apply Gibbs constraints to encourage Black and Hispanic opportunity districts. We then thin these 35,000 maps down to the final set of 7,500.
Validation
Miami / South Florida cluster
*North Florida cluster
Full map
VRA compliance
VAP histograms
CVAP histograms
Checklist
TODO
lines from the template code have been removedenforce_style()
to format my coderedist_map
andredist_plans
objects, and summary statistics) have been edited@christopherkenny