Skip to main content

View Diary: Ignoring Voter Suppression is Betraying Black Voters (317 comments)

Comment Preferences

  •  Voter suppression in Franklin county, Ohio (none)
    There is damning statistical evidence of voter suppression in predominantly Democratic precincts in Franklin county Ohio, at least linked text. I have done some statistical analyses of the precinct-level data available on this site, and found that even at a very conservative estimate, there is a systematic relationship between the proportion of votes cast for Kerry in a precinct, and the number of machines provided per active voter in the precinct. The greater the proportion of Kerry voters in a precinct, the fewer the number of voting machines were available per active voter. The chance of this relationship being due to random error (or random incompetence) is less than one in 10,000.  

    There is also solid statistical evidence to support the anecdotal reports that voter turnout was indeed suppressed by inadequate provision of machines.  The data suggest that after ceiling of around 200 voters per machine was reached, voter turnout started to drop significantly.  Again the chance of this being due to random effects is less than one in 10,000.  

    Put together, these findings are evidence that systematically inequitable provision of voting machines resulted in lower turnout, which in turn predominantly affected Kerry voters. So I then calculated the size of the effect in actual votes.  Even after allowing for the statistically significant lower turnout in pro-Kerry wards even when machine provision was adequate, I estimate that a total of around 18,500 votes were lost, of which about 66% would have gone to Kerry and 30% for Bush (based on the votes successfully cast in each precinct).  This would translate into a reduction in Bush's margin in Ohio of nearly 7000 votes from Franklin county alone.  

    This statistical evidence supports Richard Hayes Phillips allegation that "Elections Director Matt Damschroder deliberately withheld voting machines from predominantly black Democratic wards in Columbus, and dispersed some of the machines to affluent suburbs in Franklin County" linked text.  If I was a US citizen (I live in the UK) I would not consider further evidence was needed to take Damschroder to court for an illegal and successful attempt to influence the course of the election at the expense of the voting rights of African Americans in Columbus Ohio.  And if Phillips is also correct in suggesting that Blackwell and Bush were party to the strategy, can't someone impeach Bush?

    •  The author is familiar with queuing theory (none)
      ... and hence should be aware that uniform allocation of machines to voters is not optimal.

      Large, relatively urban precincts would achieve simlar service levels with lower machine/voter ratios than smaller, less urban precincts.

      Stratify the dataset by number of machines per location. This may sharpen the analysis (without rqeuiring a wholesale leap into a multiserver queuing models) ... or the suggested relationship may evaporate.

      The analysis also fails to state whether "precinct" and "voting location" are used interchangeably.

      •  Voter suppression in Franklin county, Ohio (2) (none)
        Good points.  Regarding machine numbers: all precincts had at least two machines, and the vast majority (85%) had three or four.  Only twenty seven precincts had only two machines, and only twenty eight had more than five. I stratified the sample by machine.  The relationship between percentage of Kerry votes and active voters per machine remained significant for precincts having between 3 and 4 machines.  For other precincts, numbers in each group were too small for analysis.  I repeated the whole analysis on the 85% of precincts with either 3 or 4 machines (432 and 234 respectively), with machine number as a factor. The results are very similar.  Believe me (or, don't believe me, try it yourself), I've done what I can to reduce the statistical significance of these effects - normalised the data, used non-parametric methods, got rid of high leverage data points, you name it.  Precincts with high proportions of Kerry voters got fewer voting machines per active voter than precincts with lower proportions of Kerry voters; and above the 200 mark (precincts with more than 200 active voters per machine), turnout relative to the number of active voters in the precinct started to drop.  It is also true that some crowded precincts appeared to have handled the crowding better (turnout less depressed): these are significantly more likely to be precincts with lower proportions of Kerry voters.  Note also that I am using "active voters" not "registered voters" in the analysis; "active voters" are those who have actually voted in the last two election cycles and are what the Board of Elections is supposed to base its estimate of turnout on.

        All analysis were done on precincts, not locations. I am aware some precincts shared locations.  I will try and look at this too. But these are huge effect sizes we are looking at; it would astonish me if these relationships were to be some kind of artefact of shared location.

        •  Good work so far (none)
          I would suggest screening out shared-location precincts. I assume you've also controlled for poll voters vs absentees etc.

          Within the standalone 3-machine precinct population (and neighboring machine-count classes, for context), construct the cumulative density function of active voter counts. Then break down and noramlize the corresponding high-Bush and high-Kerry components.

          If machines were deliberately allocated preferentially to Republican precincts, the allocation probably follows some fairly simple rule of thumb with visible abrupt junctures. If, on the other hand, machines were allocated preferentially but unintentionally (as an artifact of unidentified intervening variables, for instance), we'd expect the effect to appear as a more incrementally graduated relationship.

          •  OK : (none)
            I have not controlled for poll voters vs absentees.  I assumed the figures I had were for poll voters only.  The number of provisional ballots issued will presumably also affect crowding. I will check the data source.

            Within the population of stand-alone 3-machine precincts (N=158), the correlation between percent Kerry and active voters per machine is still positive and significant (Spearman's rho=0.52m p<.000001. I then plotted (for this population) a cumulative percentage of precincts along an active voters/machine axis, for precincts in which Kerry won more votes than than Bush (N=97)and those in Bush won more votes than Kerry (N=61).   The Kerry plot is virtually linear from start to finish.  The Bush plot rises much more steeply reaching the 95% mark at around 265 voters per machine, then levelling off.  The maximum for Bush precincts is 284 active voters per machine. 40% of Kerry precincts have more than 265 active voters per machine, rising to a maximum of 378, with no levelling in sight.

            There is no obvious algorithm, except that voters per machine seem to have been capped in Bush precincts but not in Kerry precincts.

            I do not think that this is evidence that no deliberate strategy was planned.  Removing precincts with only 2 machines from the analysis was a good move, as providing two machines as a minimum even in low density suburban precincts (likely to be more Republican) is clearly a defensible strategy.  But we have now excluded this as a confound.  Similarly, crowding might have been a confound in multi-precinct locations.  But we have now excluded that as well. Of course there may still be unidentified intervening variables, but at this point I think the onus should be on Damshroder to defend his machine allocation strategy.  A spearman's rho of .52 at a probability of less than .000001 is one hell of a correlation. If it isn't systematic disenfranchisement of Democrat voters in Columbus I'd like to know what it is.

            •  Getting warmer! 'Capped' is just the kind of ... (none)
              ... quirk we need. Natural populations don't normally jump up and cap themselves. This is progress.

              Now look for those "missing" high-count GOP precincts. They should appear among the 4-machine data.

              This should yield a set of "poster child" precincts for closer scrutiny and disclosure -- one smallish set of D 3-seaters with too many voters, and another smallish set of R 4-seaters with simlar voter headcounts.

              This narrows the question from "how did you assign machines?" to "why did you give precinct X 3 machines and precinct Y 4 machines?".

              (Other explanations are possible, for instance part of the county may have been re-mapped more recently than the rest.)

            •  P.S. I thought we had 432 3-seaters (none)
              So most of these were in multi-precinct sites? That muddies the water a bit.
              •  Voter suppression in Franklin county, Ohio (3) (none)
                I'll keep working on these data, taking different strata one at a time.  So far, any homogeneous population of precincts I take that is big enough to analyse shows the same effect.  I have also looked at the top and bottom quartiles of Kerry per cent precincts (hope this makes sense).  There just seems to be an upper limit on the active-voter-per-machine-numbers in pro-Bush precincts that simply isn't there in the pro-Kerry precincts. The only legitimate explanation I can imagine is that somehow Damschroder genuninely thought that urban precincts would (as you suggest) process voters more efficiently (twice as efficiently?  with a complicated ballot?  In areas likely to have generally greater levels of educational disadvantage?) as suburban areas.  This seems improbable to me, but if true, he was massively and demonstrably wrong.
              •  Probably the last analysis (none)
                I've done this every which way now. I've looked at machines per location as well as machines per precinct.  It doesn't matter how you do it, the Bush precincts always get more machines per active voter in the precinct.

                So I gave Damschroder one last benefit of the doubt.  The poor guy is short of machines (which he shouldn't be of course)so not only does he  consider "active" voters when it comes to allocating machines, he also looks at turnout in 2000.  If a precinct had a low turnout in 2000, they don't get need as many machines in 2004. And as the Democratic precincts tend to have lower turnout in turns of votes per active voter as well as votes per registered voter - it might be a confound.

                Well it isn't.  Turnout in 2000 is a significant predictor of machine allocation, so it looks as though they did take it into account.  But the way people voted is still a significant predictor of how many machines they got.  Kerry precincts got fewer machines, even after allowing for turnout in 2002.  This remains true when only precincts with 3 or 4 machines are considered.

                I can keep running regressions for ever, but it is very difficult to avoid the conclusion that the way people were likely to vote was a significant factor in the decision as to how many machines would be provided.

                I'll try and post this stuff as a diary unless anyone down this end of the thread can see a potential flaw.  

              •  Last analysis (none)
                OK, it was either stupidity or subtle.

                I think I got the algorithm they used.  They didn't use active voters, they based machine allocation on turnout in 2000. They allocated a machine for every 180 (is my guess) voters, with a minimum of two per precinct. They were actually slightly more generous to the Kerry precincts on this basis (but not by much).

                Except that because 180 voters per machine means that the machines are operating at just about maximum capacity, it didn't amount to generosity, it amounted to a cap on turnout at 2000 levels.  That's why it looks like a cap.  

                •  Good work! Seriously! One small step for ... (none)
                  ... listening to the data first and jumping to conclusions later.

                  Your surmise could probably be confirmed directly by asking the Franklin Co elections chief. It puts a set of waiting time issues in sharper focus, it identifies where Franklin was and was not sufficiently diligent, should provide cautionary context for future cycles, and it might possibly provide foundation for a potent complaint with respect to this year's cycle.

                  To what extent should county officials be expected to anticipate departures from turnout history, and how much "headroom" should they allow on general principles?

                  Waiting times and discouraged voters were likely affected by other factors (uniformity of arrivals, and fraction of high-service-time "customers" such as provisional balloters, first time voters, disabled voters and low literacy voters).

                  And let me apologize here for referring to your work as "crap" (where I first encountered it cited in excerpt, couched in extravagant claims, and defended by dismissing sharp questions).

                  •  crap (none)
                    I don't know when or where you called what crap, or where you saw what, or why you thought it was mine....

                    But it might have been.

                    I am fairly confident of this algorithm, and will try to find out if it was what was used, and whether it is legal.  It exactly predicts the number of machines used in 64% of precincts, and a further 25% of precincts had one more machine than predicted by the algorithm. The remaining 19 percent includes 4% where the 2000 turnout data is missing in the dataset. This leaves only 15% of precincts that are special in some way and some of these were very large.

                    I think there is still a burning issue here though.  This does not invalidate my finding (yet) that turnout was suppressed after a certain throughput was reached, nor that this suppression (whether accidental or culpably negligent) disproportionately affected Kerry voters.  It says that a) not enough machines were available county wide and that b) the precincts in which there was most scope for increased turnout (where the difference between "active voters" and voters in 2000 was greatest, ie in strongly Dem precincts) the crowding, and thus the suppression of turnout, was greatest. Any democratic system MUST be designed to make sure there is no ceiling to the number of potential voters in a district who can actually vote. I stand by my estimate of a net Kerry loss of around 7000 votes in Franklin county.

                    •  So '2000 voter' and 'active voter' stix ... (none)
                      ... appear to diverge with respect to their predictions of 2004 turnout, and that divergence differs for Kerry votes and Bush voters.

                      I agree that multi-hour waits (or even hour-long waits) are intolerable -- whether or not they differentially disadvantage particular candidates or demographics. Your model may provide decent estimates of discouraged votes.

    •  thank you, I'm glad to read this (none)
      Any chance you could post this work as a diary? This particular thread is so quiet now, I don't think many will see it down here.

      I see you're in touch with Cliff Arnebeck, so I know the lawyers will see your analysis.

Subscribe or Donate to support Daily Kos.

Click here for the mobile view of the site