1. Reproducing the stated factor

definition                              n   sum f / sum m
all rows, sum/sum               3,761,792          0.7335
both sides nonzero              1,553,602          0.7636
verdict = isomorphic              679,283          0.7601
male>=11 and female>=8            388,946          0.7568
male>=11 only                     559,472          0.6820
female>=8 only                    556,286          0.8025

None of these lands on 0.581. The published number is stated without a derivation on the dataset website, and this table is aggregated by cell type, so the definitions may differ.

2. Does the factor drift with connection strength?

   Thresholding both sides, as the paper describes for noise removal:
     male weight >=    11   n= 44,759   f/m = 1.3419
     male weight >=    15   n= 45,372   f/m = 1.1199
     male weight >=    20   n= 64,972   f/m = 0.9515
     male weight >=    30   n= 75,077   f/m = 0.8257
     male weight >=    50   n= 55,187   f/m = 0.7679
     male weight >=    80   n= 41,712   f/m = 0.7434
     male weight >=   130   n= 25,429   f/m = 0.7414
     male weight >=   200   n= 20,098   f/m = 0.7399
     male weight >=   350   n=  9,304   f/m = 0.7283
     male weight >=   600   n=  3,811   f/m = 0.7147
     male weight >=  1000   n=  3,225   f/m = 0.6943
   spread 0.6943 to 1.3419 = 1.93x, and monotonic.
   Read naively this says one factor cannot work.

3. Control: threshold one side only

     male weight >=    11   n=108,325   f/m = 0.6930
     male weight >=    15   n= 84,479   f/m = 0.6861
     male weight >=    20   n= 97,129   f/m = 0.6808
     male weight >=    30   n= 93,724   f/m = 0.6784
     male weight >=    50   n= 63,114   f/m = 0.6758
     male weight >=    80   n= 46,074   f/m = 0.6742
     male weight >=   130   n= 27,588   f/m = 0.6833
     male weight >=   200   n= 21,728   f/m = 0.6841
     male weight >=   350   n=  9,981   f/m = 0.6798
     male weight >=   600   n=  4,036   f/m = 0.6753
     male weight >=  1000   n=  3,294   f/m = 0.6883
   spread 0.6742 to 0.6930 = 1.03x. Flat.

4. Simulation: true constant ratio plus counting noise

   ground truth ratio used: 0.6820
     bin    simulated       real
      11       0.8254     1.3419
      15       0.7240     1.1199
      20       0.6868     0.9515
      30       0.6831     0.8257
      50       0.6823     0.7679
      80       0.6825     0.7434
     130       0.6812     0.7414
     200       0.6812     0.7399
     350       0.6822     0.7283
     600       0.6822     0.7147
    1000       0.6824     0.6943

   A constant ratio plus Poisson noise reproduces the shape of the
   drift. The drift in step 2 is an artefact of selecting on both
   axes at once, not evidence against a single factor.
   The real drift is steeper than Poisson, so the counts are
   overdispersed relative to independent sampling.

wrote D:\fly-research\results\c1_scaling_factor.json
