> Letter to Howard Dean about the DNC election report and the evidence of vote fraud in Ohio
2 July 2005, 00:03
Thanks for your thoughtful critique of our work in the DNC Ohio 2004
report.
All I will say about your overall argument that the report does not
take into account everything that is known about the 2004 election in
Ohio is that that is correct. The report did not have in its scope to
try to gather evidence regarding all of the administrative decisions
and processes involved in the election. So, for example, we did not
seek to gather documents regarding the handling of voter registrations
coming in shortly before the election, even though we expected that
mishandling of those would be related to voting problems. Indeed, it
seems from what we did find regarding especially provisional ballots
that mishandling of voter registrations was a problem. But resource
constraints drove the decision to limit the scope of the data
collection in that way. We were and are mindful that the study we did
was neither the only one nor the last. Others with advantages such as
subpoena power and a longer time horizon we hoped would be able to
push more insistently into such questions. Also, even though the
report includes descriptions of election-day problems and voter
protection efforts in parts of the state (especially Cuyahoga County),
it does not attempt a census of all the reports of that kind that have
been produced. The DNC report was not framed as a synthesis but as an
effort to get analytically sharp answers about specific questions.
Here are some comments regarding each of your numbered points.
"1. Failed, unlawful recount, and lack of cooperation from the
Secretary of State": We used a data file containing election returns
and information about voter registration and turnout in precincts
that, as I understand it, came from the Secretary of State’s office.
I did not personally acquire the data. As the DNC report states, Eric
Greenwald had that job. I believe that that file was supposed to
reflect the latest version of all the information, including revisions
based on recounts in precincts where those occurred. We went through
several versions of that data file, as our analysis turned up problems
in it. There were horrendous problems with data from Lucas County,
which traced to records being incorrectly combined in the file from
the SoS (I determined that by comparing the SoS data to a canvass file
we received from Lucas County). A few other counties apparently had
similar problems. Lucas is the only one I worked out all the details
for myself, before receiving revised data from the SoS. Up to the
penultimate version of the file that was used to produce the results
appearing in the DNC report, there were many outliers for Butler
County that I was told stemmed from a permuted records problem like
the one in the Lucas County data. I did not have time to verify that
with information other than the SoS data, as the deadline was
approaching and I needed to redo all of the statistical analysis to
use the corrected data. The number of oddities and outliers in the
analysis of the Kerry-Bush vote split fell considerably between the
penultimate and the final version of the data used for the DNC report.
The files we used to run the analysis reported in the DNC report,
except for information about precinct racial composition (which the
DNC does not wish to release), are available from my website, at http://macht.arts.cornell.edu/wrm1/Ohio2004/OhioDNC/
"2. Implausibly low voter turnout in Cleveland, Cuyahoga County": We
found many anomalous results in Cuyahoga County. The significant
negative coefficient for the effect of voting machine provision on
voter turnout (precinct report Table 3) is striking but small.
I did not emphasize it because it does not appear to be associated
with large practical declines in voter turnout, but you are right to
call attention to the anomaly. Cuyahoga was also unusual in the
analysis of the Kerry-Bush vote split. Only in Cuyahoga was a higher
proportion voting for the anti-gay marriage amendment associated with
a higher share of votes for Kerry, and there were unusally many
outlier precincts in Cuyahoga (precinct report Tables 34 and 35). We
remark in the DNC report that the results observed for Cuyahoga County
warrant further investigation.
"3. Voter suppression through insufficient machine allocation -
Franklin County": You state the following:
"So, what does the DNC report have to say about this? It says that
those who decided to leave the polls early because of long lines were
split evenly between Bush and Kerry voters."
I don’t believe such a claim occurs anywhere in the DNC report. The
Voter Experience Survey (Section III) finds that across the state
between 2 and 3 percent of voters left the polls due to long lines and
did not return. The precinct analysis of the effect of voting machine
provision on voter turnout produces a comparable estimate. The sample
size in the Voter Experience Survey is too small to make a reliable
estimate of the partisan breakdown of the voters who said they left
the polls due to long lines.
"4. Anomalies in southwestern Ohio": You mention that the DNC report
does not present results specifically for Clermont, Butler, and Warren
Counties. The results for all counties are available in the file
checkpres2simnlN.Rout that is included in the file DNCreplic1.zip,
downloadable from my website mentioned above. Here are the results
for the three counties you mention. I’m also including results for
Miami County, which is the topic of your point 5.
These are robust binomial logistic regression model estimates,
matching those reported for three counties in precinct report Table
34. Variable "dsenlogit04" is the logit of the vote for Fingerhut,
"I1logit" is the logit of the vote Yes on Issue 1, and "VANHAAprop" is
the proportion African American.
Number of Observations: 157
Number of observations with at least one zero weight: 0
The only unexpected result is the negative estimate for the
coefficient for VANHAAprop in Warren County, but that estimate is not
statistically significant. I decided to highlight only statistically
significant deviations from the expected coefficient signs. There are
no outlier precincts in any of these four counties.
You write the following.
"Also consider the fact that part of the reason for Bush’s excess
vote margin in the three counties was an extra-ordinarily large
increase in voter registration from 2000, including a 30% increase in
Warren County. Yet, according to the DNC report, an increase in voter
registration was supposed to favor Kerry in 2004."
In fact we observe that larger increases in registration from 2002 to
2004 mostly went with higher proportions of votes for Kerry, but
larger increases in voter turnout from 2002 to 2004 mostly went with
higher proportions of votes for Bush. Regarding the increase in
turnout, the Summary of Principal Findings in the precinct report
states, "Increases in voter turnout above the rates expected based on
the 2002 general election were strongly associated with the proportion
voting Yes on Issue 1 (opposing gay marriage)." Other evidence in the
report shows that where registration increased sharply during 2004, so
did the proportion of voters forced to cast a provisional ballot (see
the provisional ballot survey conducted in Cuyahoga County, Sections
IV and V, and the analysis of data from Franklin County, at the end of
Section VI). So part of what happened was that Republican GOTV
produced better results than Democratic GOTV did, and part of what
happened is that Democratic registration efforts were thwarted by
inadequate (to say the least) election administration.
"5. Late vote surge in Miami County": For analysis of the Kerry-Bush
vote split in Miami County see the preceding item. We lacked data
about the number of voting machines in each precinct for Miami County,
so data from that county are not included in the precinct report
analysis that compared turnout in 2002 to turnout in 2004 (precinct
report Tables 6 and 8). For analysis that includes Miami County see
the file checkprecturnout2mN.Rout included in the file DNCreplic1.zip,
downloadable from my website mentioned above. Here are results from
that file for "Opt Central" counties. Variable "vlogit02" is the
logit of turnout in 2002 and "I1logit04" is the logit of the vote Yes
on Issue 1.
Number of Observations: 593
Number of observations with at least one zero weight: 4
county place
4337 GEAUGA HUNTSBURG TOWNSHIP PRECINCT B
7951 MIAMI CONCORD TOWNSHIP SOUTH EAST PRECINCT
7952 MIAMI CONCORD TOWNSHIP SOUTH PRECINCT
7953 MIAMI CONCORD TOWNSHIP SOUTH WEST PRECINCT
There are four outlier precincts, three of which are from Miami
County. Two of them have substantially higher turnout than expected
based on 2002 and support for Issue 1 and one has substantially lower
turnout.
"6. Vote switching in Mahoning County": In the analysis that related
the Kerry-Bush vote split to the vote for governor in 2002, no
Mahoning precinct is an outlier (precinct report Tables 30—33). The
analysis that compares the Kerry-Bush vote split to other 2004 votes
find nothing unusual in parameters estimated using the precinct data
from Mahoning. Here are the results for Mahoning from
checkpres2simnlN.Rout.
Number of Observations: 311
Number of observations with at least one zero weight: 0
county SPC precinct std.resid
7331 MAHONING ACV YOUNGSTOWN CITY SECOND WARD PRECINCT L 3.672540
7333 MAHONING ADK YOUNGSTOWN CITY THIRD WARD PRECINCT H -3.164071
7536 MAHONING AQV YOUNGSTOWN CITY SIXTH WARD PRECINCT I -3.921959
Even though there are no outlier precincts in Mahoning, there are
three precincts in Mahoning County that have notably unususal results.
The summaries printed in checkpres2simnlN.Rout list every precinct
that has a studentized residual with magnitude greater than 3.0 (a
precinct that’s an outlier has a residual of magnitude greater than
4.0). You can see the three precincts that had Kerry-Bush vote split
results that were that unusual, relative to the estimated model. In
two of the precincts the vote for Kerry was unusually low and in one
the vote for Kerry was unsually high.
"7. As yet uncounted ballots": As you observe, we did extensive
analysis of the residual vote. Unavailable data prevented us from
doing anything regarding uncounted provisional ballots.
You state, "The facts that Kerry won the Ohio exit poll by a
statistically significant 4.2%...." We decided not to do anything
with the exit poll data once it became clear we would not be able to
find out the precincts they used (we tried pretty hard). That meant
we would have no new evidence to add to the extensive public
controversy on the topic. The DNC study therefore takes no position
regarding the exit polls. My own opinion, based on closely following
the extensive controversy and talking to various people, is that the
exit polls in Ohio had a Democratic bias. But that’s just my opinion.
Walter Mebane
* - * - * - * - * - * - * - * - * - * - * - * - * - * - * - * - * - *
Walter R. Mebane, Jr. email: wrm1@cornell.edu
Professor office voice: 607/255-3868
Department of Government cell: 607/592-0546
Cornell University fax: 607/255-4530
217 White Hall WWW: http://macht.arts.cornell.edu/wrm1/
Ithaca, NY 14853-7901
* - * - * - * - * - * - * - * - * - * - * - * - * - * - * - * - * - *
Thanks for your thoughtful critique of our work in the DNC Ohio 2004
report.
All I will say about your overall argument that the report does not
take into account everything that is known about the 2004 election in
Ohio is that that is correct. The report did not have in its scope to
try to gather evidence regarding all of the administrative decisions
and processes involved in the election. So, for example, we did not
seek to gather documents regarding the handling of voter registrations
coming in shortly before the election, even though we expected that
mishandling of those would be related to voting problems. Indeed, it
seems from what we did find regarding especially provisional ballots
that mishandling of voter registrations was a problem. But resource
constraints drove the decision to limit the scope of the data
collection in that way. We were and are mindful that the study we did
was neither the only one nor the last. Others with advantages such as
subpoena power and a longer time horizon we hoped would be able to
push more insistently into such questions. Also, even though the
report includes descriptions of election-day problems and voter
protection efforts in parts of the state (especially Cuyahoga County),
it does not attempt a census of all the reports of that kind that have
been produced. The DNC report was not framed as a synthesis but as an
effort to get analytically sharp answers about specific questions.
Here are some comments regarding each of your numbered points.
"1. Failed, unlawful recount, and lack of cooperation from the
Secretary of State": We used a data file containing election returns
and information about voter registration and turnout in precincts
that, as I understand it, came from the Secretary of State’s office.
I did not personally acquire the data. As the DNC report states, Eric
Greenwald had that job. I believe that that file was supposed to
reflect the latest version of all the information, including revisions
based on recounts in precincts where those occurred. We went through
several versions of that data file, as our analysis turned up problems
in it. There were horrendous problems with data from Lucas County,
which traced to records being incorrectly combined in the file from
the SoS (I determined that by comparing the SoS data to a canvass file
we received from Lucas County). A few other counties apparently had
similar problems. Lucas is the only one I worked out all the details
for myself, before receiving revised data from the SoS. Up to the
penultimate version of the file that was used to produce the results
appearing in the DNC report, there were many outliers for Butler
County that I was told stemmed from a permuted records problem like
the one in the Lucas County data. I did not have time to verify that
with information other than the SoS data, as the deadline was
approaching and I needed to redo all of the statistical analysis to
use the corrected data. The number of oddities and outliers in the
analysis of the Kerry-Bush vote split fell considerably between the
penultimate and the final version of the data used for the DNC report.
The files we used to run the analysis reported in the DNC report,
except for information about precinct racial composition (which the
DNC does not wish to release), are available from my website, at
http://macht.arts.cornell.edu/wrm1/Ohio2004/OhioDNC/
"2. Implausibly low voter turnout in Cleveland, Cuyahoga County": We
found many anomalous results in Cuyahoga County. The significant
negative coefficient for the effect of voting machine provision on
voter turnout (precinct report Table 3) is striking but small.
I did not emphasize it because it does not appear to be associated
with large practical declines in voter turnout, but you are right to
call attention to the anomaly. Cuyahoga was also unusual in the
analysis of the Kerry-Bush vote split. Only in Cuyahoga was a higher
proportion voting for the anti-gay marriage amendment associated with
a higher share of votes for Kerry, and there were unusally many
outlier precincts in Cuyahoga (precinct report Tables 34 and 35). We
remark in the DNC report that the results observed for Cuyahoga County
warrant further investigation.
"3. Voter suppression through insufficient machine allocation -
Franklin County": You state the following:
"So, what does the DNC report have to say about this? It says that
those who decided to leave the polls early because of long lines were
split evenly between Bush and Kerry voters."
I don’t believe such a claim occurs anywhere in the DNC report. The
Voter Experience Survey (Section III) finds that across the state
between 2 and 3 percent of voters left the polls due to long lines and
did not return. The precinct analysis of the effect of voting machine
provision on voter turnout produces a comparable estimate. The sample
size in the Voter Experience Survey is too small to make a reliable
estimate of the partisan breakdown of the voters who said they left
the polls due to long lines.
"4. Anomalies in southwestern Ohio": You mention that the DNC report
does not present results specifically for Clermont, Butler, and Warren
Counties. The results for all counties are available in the file
checkpres2simnlN.Rout that is included in the file DNCreplic1.zip,
downloadable from my website mentioned above. Here are the results
for the three counties you mention. I’m also including results for
Miami County, which is the topic of your point 5.
These are robust binomial logistic regression model estimates,
matching those reported for three counties in precinct report Table
34. Variable "dsenlogit04" is the logit of the vote for Fingerhut,
"I1logit" is the logit of the vote Yes on Issue 1, and "VANHAAprop" is
the proportion African American.
county 9 : BUTLER
Choice 1 : kerry04 Estimates and SE:
Est SE.Sand t.val.Sand
(Intercept) 0.407 0.0201 20.3
dsenlogit04 0.931 0.0198 47.1
I1logit -0.203 0.0189 -10.7
VANHAAprop 1.340 0.1150 11.7
LQD sigma: 1.115918
TANH sigma: 1.037107
Number of Observations: 288
Number of observations with at least one zero weight: 0
county 13 : CLERMONT
Choice 1 : kerry04 Estimates and SE:
Est SE.Sand t.val.Sand
(Intercept) 0.241 0.0368 6.54
dsenlogit04 0.841 0.0263 32.00
I1logit -0.158 0.0326 -4.84
VANHAAprop 1.790 0.6040 2.96
LQD sigma: 0.9000457
TANH sigma: 0.827516
Number of Observations: 191
Number of observations with at least one zero weight: 0
county 55 : MIAMI
Choice 1 : kerry04 Estimates and SE:
Est SE.Sand t.val.Sand
(Intercept) 0.538 0.0488 11.00
dsenlogit04 0.923 0.0402 23.00
I1logit -0.223 0.0510 -4.38
VANHAAprop 0.603 0.4200 1.44
LQD sigma: 1.083368
TANH sigma: 1.011146
Number of Observations: 82
Number of observations with at least one zero weight: 0
county 83 : WARREN
Choice 1 : kerry04 Estimates and SE:
Est SE.Sand t.val.Sand
(Intercept) 0.451 0.0488 9.24
dsenlogit04 0.855 0.0326 26.20
I1logit -0.331 0.0414 -8.00
VANHAAprop -0.121 0.0899 -1.35
LQD sigma: 1.259254
TANH sigma: 1.152096
Number of Observations: 157
Number of observations with at least one zero weight: 0
The only unexpected result is the negative estimate for the
coefficient for VANHAAprop in Warren County, but that estimate is not
statistically significant. I decided to highlight only statistically
significant deviations from the expected coefficient signs. There are
no outlier precincts in any of these four counties.
You write the following.
"Also consider the fact that part of the reason for Bush’s excess
vote margin in the three counties was an extra-ordinarily large
increase in voter registration from 2000, including a 30% increase in
Warren County. Yet, according to the DNC report, an increase in voter
registration was supposed to favor Kerry in 2004."
In fact we observe that larger increases in registration from 2002 to
2004 mostly went with higher proportions of votes for Kerry, but
larger increases in voter turnout from 2002 to 2004 mostly went with
higher proportions of votes for Bush. Regarding the increase in
turnout, the Summary of Principal Findings in the precinct report
states, "Increases in voter turnout above the rates expected based on
the 2002 general election were strongly associated with the proportion
voting Yes on Issue 1 (opposing gay marriage)." Other evidence in the
report shows that where registration increased sharply during 2004, so
did the proportion of voters forced to cast a provisional ballot (see
the provisional ballot survey conducted in Cuyahoga County, Sections
IV and V, and the analysis of data from Franklin County, at the end of
Section VI). So part of what happened was that Republican GOTV
produced better results than Democratic GOTV did, and part of what
happened is that Democratic registration efforts were thwarted by
inadequate (to say the least) election administration.
"5. Late vote surge in Miami County": For analysis of the Kerry-Bush
vote split in Miami County see the preceding item. We lacked data
about the number of voting machines in each precinct for Miami County,
so data from that county are not included in the precinct report
analysis that compared turnout in 2002 to turnout in 2004 (precinct
report Tables 6 and 8). For analysis that includes Miami County see
the file checkprecturnout2mN.Rout included in the file DNCreplic1.zip,
downloadable from my website mentioned above. Here are results from
that file for "Opt Central" counties. Variable "vlogit02" is the
logit of turnout in 2002 and "I1logit04" is the logit of the vote Yes
on Issue 1.
[1] "Opt Central"
Choice 1 : votescast04 Estimates and SE:
Est SE.Sand t.val.Sand
(Intercept) 1.040 0.0179 58.00
vlogit02 0.671 0.0155 43.40
I1logit04 0.194 0.0303 6.42
LQD sigma: 2.340966
TANH sigma: 2.136888
Number of Observations: 593
Number of observations with at least one zero weight: 4
county place
4337 GEAUGA HUNTSBURG TOWNSHIP PRECINCT B
7951 MIAMI CONCORD TOWNSHIP SOUTH EAST PRECINCT
7952 MIAMI CONCORD TOWNSHIP SOUTH PRECINCT
7953 MIAMI CONCORD TOWNSHIP SOUTH WEST PRECINCT
GEAUGA & ACA & -5.14 \
MIAMI & ABX & -6.95 \
MIAMI & ABY & 5.02 \
MIAMI & ABZ & 5.83 \
There are four outlier precincts, three of which are from Miami
County. Two of them have substantially higher turnout than expected
based on 2002 and support for Issue 1 and one has substantially lower
turnout.
"6. Vote switching in Mahoning County": In the analysis that related
the Kerry-Bush vote split to the vote for governor in 2002, no
Mahoning precinct is an outlier (precinct report Tables 30—33). The
analysis that compares the Kerry-Bush vote split to other 2004 votes
find nothing unusual in parameters estimated using the precinct data
from Mahoning. Here are the results for Mahoning from
checkpres2simnlN.Rout.
county 50 : MAHONING
Choice 1 : kerry04 Estimates and SE:
Est SE.Sand t.val.Sand
(Intercept) 0.870 0.0194 44.90
dsenlogit04 1.120 0.0269 41.50
I1logit -0.298 0.0458 -6.51
VANHAAprop 1.040 0.0813 12.80
LQD sigma: 1.255465
TANH sigma: 1.179296
Number of Observations: 311
Number of observations with at least one zero weight: 0
county SPC precinct std.resid
7331 MAHONING ACV YOUNGSTOWN CITY SECOND WARD PRECINCT L 3.672540
7333 MAHONING ADK YOUNGSTOWN CITY THIRD WARD PRECINCT H -3.164071
7536 MAHONING AQV YOUNGSTOWN CITY SIXTH WARD PRECINCT I -3.921959
Even though there are no outlier precincts in Mahoning, there are
three precincts in Mahoning County that have notably unususal results.
The summaries printed in checkpres2simnlN.Rout list every precinct
that has a studentized residual with magnitude greater than 3.0 (a
precinct that’s an outlier has a residual of magnitude greater than
4.0). You can see the three precincts that had Kerry-Bush vote split
results that were that unusual, relative to the estimated model. In
two of the precincts the vote for Kerry was unusually low and in one
the vote for Kerry was unsually high.
"7. As yet uncounted ballots": As you observe, we did extensive
analysis of the residual vote. Unavailable data prevented us from
doing anything regarding uncounted provisional ballots.
You state, "The facts that Kerry won the Ohio exit poll by a
statistically significant 4.2%...." We decided not to do anything
with the exit poll data once it became clear we would not be able to
find out the precincts they used (we tried pretty hard). That meant
we would have no new evidence to add to the extensive public
controversy on the topic. The DNC study therefore takes no position
regarding the exit polls. My own opinion, based on closely following
the extensive controversy and talking to various people, is that the
exit polls in Ohio had a Democratic bias. But that’s just my opinion.
Walter Mebane
* - * - * - * - * - * - * - * - * - * - * - * - * - * - * - * - * - *
Walter R. Mebane, Jr. email: wrm1@cornell.edu
Professor office voice: 607/255-3868
Department of Government cell: 607/592-0546
Cornell University fax: 607/255-4530
217 White Hall WWW: http://macht.arts.cornell.edu/wrm1/
Ithaca, NY 14853-7901
* - * - * - * - * - * - * - * - * - * - * - * - * - * - * - * - * - *