Florida map
 
Best fit of Mitofsky data
 
Frequently Asked Questions
What Went Wrong in Ohio
Trends in Florida voting. A brief report.
Elections Research
Mission Statement
In Response to Douglas Jones' "Mark Sense Ballot Design: Human Factor Lessons from Florida Election 2000"
The letter Tom Paine eventually printed
From the New Deal to the Dirty Deal
 

Recommended websites

Carter: No doubt Gore won

uscountvotes.org

New Scientific Paper: Exit Poll Discrepancies

thealliancefordemocracy.org

Freepress.org
John Conyers' Blog
 
 
OliverDawshed@aol.com

 

Trends in Florida voting. A brief report.
By Oliver Dawshed
Revision 2.0   2/24/05
Typesetting errors in Figs. 3.1.b and 3.2.5.1 emended 10/22/05

 

Abstract. The present work uses data from 1992-2004 to develop standard methodology to understand voting trends. The focus was on crossover from U.S. Senator to President in Florida elections, but other races were considered. A normalized internal standard method was found to be most widely applicable.

The stimulus for this study was a controversy arising out of Election 2004 regarding the voting behavior of a number of rural north Florida counties. It was asserted that racial attitudes guide voting patterns, with the implication that residual white racism plays such a dominant role that other factors are negligible. In the present work, this is called the “Dixiecrat hypothesis.”  

In particular, it was asserted that significant numbers of people in some counties vote for the Democratic candidate for U.S. Senate but cross over to vote for the Republican presidential candidate. This assertion fails very basic tests. As judged by the Shapiro-Wilk test, the crossover distribution is essentially normal, with no unambiguous outliers. To the extent it deviates from normality, it does not seem to be consistently due to this effect alone. The most generous ad hoc interpretation might allow that perhaps five counties are in some way different.

Furthermore, the criterion suggested by proponents of the Dixiecrat hypothesis - disparity between registration and voting behavior - is found to be unsuitable. This criterion is characteristic of about half of Florida counties. It is also distributed in a bimodal fashion, making it difficult to treat by many statistical measures. Crossover, by contrast, is essentially normally distributed. 

It is always possible and never wise to construct post hoc analyses to support a viewpoint. The work presented here suggests that crossover behavior is extremely complex, not easily characterized by a phrase. Perhaps Dixiecrat voting does influence crossover. That conclusion, however, is not clear from the data, and if there is a trend or pattern, it is not evident in recent history.

 

1.0   Preface. It has been asserted [1] that voting behavior of certain rural, north Florida counties follows a clear, historical trend or pattern, such that crossover voting to the Republican presidential candidate by voters who vote for Democratic senator and other Democrats down the ballot is actually normal. Such crossover, known also as ticket splitting, is to be distinguished from differences between party registration and the observed vote, a phenomenon that is discussed in more detail in the text. 

Words like “clear,” “historical,” and “trend” (or “pattern”) have precise meanings to the statistically minded.  “Clear” implies “different by three standard deviations from the mean” on a single measurement and somewhat less with repeated measurements.  In distributional analysis, if a distribution is truly bimodal, one expects a failure of the Shapiro-Wilk or similar test of normality. “Historical” implies that the observation applies to at least two time points, and hopefully more.  And “trend” suggests that there is a non-zero slope to the graph of a variable vs. time, while “pattern” suggests that the slope is zero. 

There are, in fact, many potential explanations for unusual crossover voting. For example, a candidate in a given race might be much more popular in one locality than a candidate of the same party for another race. The assertion that north Florida Democratic voters cross over to vote Republican in the presidential race, a hypothesis called the “Dixiecrat hypothesis”, is just one possible explanation of many. In brief, this is a hypothesis that racial attitudes are the key factor in understanding Florida voting patterns, at least in some counties. As we have demonstrated, however, unusual crossover voting can also be indicative of electoral tampering. [2, 3]  This is neither a supposition nor a hypothesis, but a mathematical fact.

Racial attitudes are, of course, an issue in voting throughout the United States. It’s unclear why they should affect crossover, however, at least at the federal level. Presidents and U.S. Senators deal with a similar roster of issues relating to race. Furthermore, getting the proponents of the Dixiecrat hypothesis to provide criteria to specify which counties they believe should be called “Dixiecrat” has been frustrating.  Authors from Caltech and MIT, for example, cite 10 northern counties, based on discrepancies between registration and presidential voting patterns.[4] (see Box 1.1) But a Cornell academic has suggested similar criteria that would include seven counties, a very different list than proposed by Caltech/MIT. The lack of objective criteria is a warning sign that the analysis may be post hoc, fitting the observations to an idea, rather than letting the data speak for themselves. 

 

Box 1.1  Different definitions of “Dixiecrat” counties
1. Greatest discrepancy between registration and voting in 2004: Baker, Dixie, Calhoun, Franklin, Holmes, Lafayette, Liberty, Taylor, Washington, and Union.[4]
2. Democratic registration (1996) greater than 60%, but Democratic vote less than 45%: Baker, Bradford, Hardee, Holmes, Lafayette, Suwannee, and Walton.[1]

 

Proponents of the Dixiecrat hypothesis, quick to assert the pre-eminence of Dixiecrat voting as an explanatory variable,  have been very slow to concede that there might be other explanations for what they claim to see. One simple example is that north Florida crossover patterns at the county level could be somewhat different than those of the rest of the state because many of the Democrats who sought a seat in the US Senate, notably Bob Graham, Betty Castor, and Bill Nelson, served in Tallahassee before running for the Senate. Some of their Republican opponents were either from south Florida (Martinez) or served in Washington (McCollum). So, the Democrats may have enjoyed an advantage in name recognition, access to local media, and shorter travel distance from their homes to north Florida political events.

At the precinct level, however, very few reasons suffice to explain unusual crossover. Regional advantages, such as access to local media and convenient travel for campaigning cannot be explanatory. Within a county, demographic differences are also substantially compressed. Although housing segregation can certainly follow racial and economic lines, one might expect genuine segregation to manifest itself in the form of clusters of precincts with high crossover and others with low, not as a smooth, linear, normal distribution. So, one is left in a quandary if the pattern observed is the latter.

Nor does the Dixiecrat hypothesis explain why the pattern of crossover within a county should change substantially and unpredictably from election to election, nor can it explain why the unusual crossover is observed in locations far outside of rural north Florida, many of which meet none of the criteria proposed for qualifying as Dixiecrat. Finally, the Dixiecrat hypothesis cannot explain why certain counties show crossover ratios greater than 1, representing voters who vote Republican at the senatorial level but cross over to vote Democratic at the presidential level.

The present work investigates whether crossover in north Florida, or in any Florida county, can be considered to be a clear, historical trend (or pattern). Means of analyzing voting patterns involving crossover are explored and objective statistical criteria by which to fairly assess the question are presented.

 

2.0   Introduction. The present work investigates whether crossover in north Florida, or in any Florida county, can be considered to be a clear, historical trend (or pattern). Longitudinal studies of voting behavior are notoriously difficult, so discerning trends (or patterns) in voting is particularly challenging. The candidate roster changes, as do perceptions of individual candidates. Consider that in 1992, an incumbent Republican president faced a Democratic challenger in the face of a strong third party challenge.  In 1996, an incumbent Democratic president faced a Republican challenger in the face of a weak third party challenge.  In 2000, neither presidential candidate enjoyed the benefits of incumbency, nor was there a serious third party challenge. In 2004, a Republican incumbent faced a Democratic challenger with no serious third party challenge. While some voters will come out to vote for the candidate of their chosen party no matter what, many may stay home or switch between parties depending on the circumstances.

Factors other than those inherent in the candidates also affect voting. Different segments of the electorate are energized or indifferent in a given year. The demographics of a region may change. Voting machine technology evolves, altering the number of miscast or spoiled ballots. A minor factor, but one that complicates analysis, is the fact that more votes are typically cast for the candidates at the top of the ticket than for candidates down the ballot. For these and other reasons, accurately apprehending trends in voting behavior is a remarkably difficult task. 

There are many ways to try to control for the principal factors. One of the simplest is the use of percentages, which controls for changes in the size of the electorate.  A further refinement is normalization, in which all data are displayed on a fixed scale, typically zero to 1.  This is done by determining the minimum and the maximum and calculating for each data point the expression:


 
(value – minimum)/(maximum – minimum) eq. 2.0.1
   

So, if the presidential vote percentage for a candidate in a state ranges from 20% to 70%, and a particular county has a value of 45%, this becomes (45-20)/(70-20) = 0.5.  This approach has the valuable effect of diminishing the influence of third party candidates on the resultant statistics.  If a third party takes either a given number of percentage points or a proportional fraction of a party’s vote in each county, then both numerator and denominator of eq. 2.0.1 stay the same.  

A third means of controlling for factors involves an internal standard. As described in earlier work, the ratio of votes between two candidates on the same ticket should be more or less constant across a state.  If a presidential candidate receives 90,000 votes and a senatorial candidate receives 100,000, the crossover ratio is 90,000/100,000 = 0.9. Deviations from the statewide average crossover ratio in one county may signal that one candidate or the other has a local preference, either positive or negative.   

Finally, it is often helpful to convert data into Z-scores.  The Z-score expresses the distance of a single observation from the mean in units of standard deviations. 

 

3.0   Results



3.1     Categorization of Florida counties according to presidential preference

When Democratic presidential performance (expressed as a percent of total vote) in Florida counties for the elections 1992-2004 is normalized according to equation 2.0.1, the slope of change taken, and the slopes rank-ordered, 11 counties are found to show increasing relative Democratic strength by more than one standard deviation, but less than two, from the mean.a These are St. Lucie, Martin, Orange, Osceola, Monroe, Indian River, Seminole, Palm Beach, Brevard, Sarasota, and Pinellas.  Thirteen counties show the opposite trend.  Twelve are more than one standard deviation from the mean (Washington, Nassau, Levy, Columbia, Sumter, Holmes, Baker, Suwannee, Bradford, Union, Lafayette, and Gilchrist), while Dixie County is more than two standard deviations from the mean. 

When Republican presidential performance is taken through the same procedure, a different list is obtained.  Relative Republican strength increased by more than one standard deviation in Okeechobee, Glades, Putnam, Union, Calhoun, Taylor, Sumter, Walton, Suwannee, Levy, and Lafayette and by more than two standard deviations in Gilchrist and Dixie.  Republican strength waned by more than one standard deviation in Orange, Miami-Dade, Seminole, Duval, Collier, Martin, Palm Beach, Leon, and Clay.

Therefore, the counties in which Democratic strength at the presidential level is probably rising at the expense of Republican strength are Martin, Orange, Seminole, and Palm Beach.  The counties in which Republican strength at the presidential level is probably rising at the expense of Democratic strength are Sumter and a cluster on the eastern side of the Panhandle:  Union, Suwannee, Levy, Lafayette, Gilchrist, and Dixie. In the others, third party effects and defects inherent in the calculation processb make debatable the conclusion as to waxing or waning political strength.  

 

Figure 3.1.a -Dixie Presidential trends Normalized Presidential Performance 1992, 1996, 2000, 2004 

  Figure 3.1.b  Miami-Dade Presidential trends Normalized Presidential Performance 1992, 1996, 2000, 2004

 

Different patterns in trends are illustrated by the graphs above.  In the graph titled Dixie Presidential Trends, relative Democratic performance has declined as Republican performance has increased. In the graph titled Miami-Dade Presidential Trends, relative Republican performance has fallen sharply, while relative Democratic performance is approximately constant. 

Other counties may have voted more heavily Republican in recent years.  The analysis of this section only indicates relative performance in a single race.  As described in a footnote,a one of the virtues of using relative performance as a measure is that it compensates to some degree for the strength or weakness of a candidate in any given year, not to mention shifting national tides. Compared to the general trends, relatively few counties show the pronounced pattern of Democratic decline and Republican ascendance in presidential performance that Dixie does. 

 

3.2    The "Dixiecrat hypothesis"

 

3.2.1  -The conservative north

The trend in relative presidential performance across the state is, of course, only part of the story in understanding broader electoral trends. It has been proposed that in certain areas of Florida, voters vote Democratic for local candidates but Republican for presidential candidates, an idea termed “the Dixiecrat hypothesis.” The name comes from the 1948 Southern revolt under the flag of Strom Thurmond’s State’s Rights Party against the civil rights agenda proposed by Harry Truman. This is perhaps a misnomer, since Florida was not a major player in the Dixiecrat revolt, turning in one of the weakest performances for the State’s Rights Party of any southern state at 15% of the popular vote.   

There is, however, no question that north Florida has long been intensely conservative.  According to Dave Leip’s Atlas of presidential elections, as early as 1964 Barry Goldwater carried all of north Florida except Baker, Hamilton, Union, Bradford, Alachua, Gilchrist, Levy, and Dixie counties.

The right-wing hold on north Florida has not been absolute. Jimmy Carter, from neighboring Georgia, was able to carry much of the region in both 1976 and 1980, and Bill Clinton carried a number of the central Panhandle counties in 1992 and 1996.  Some counties in the central Panhandle (Gadsden, Jefferson, Madison) have high levels of African American participation that has helped Democratic efforts. Leon County, while largely white, is a strong regional Democratic contributor. So, north Florida is highly polarized, with most counties leaning Republican, at least at the presidential level.

One can see the shift toward voting Republican of the last 12 years manifest in the normalized presidential performance ratios of, for example, Dixie County shown in Fig. 3.1.a.  But are these trends distinguishable from a general conservative drift, as is implicitly claimed by the Dixiecrat hypothesis, or are they part of it? Proper understanding of this issue is critical to understanding whether unusual values of crossover are normal or not. 

Some statewide Democratic candidates below the presidential level have done well in north Florida. Still, it’s difficult to find evidence that there is a greater tendency of north Florida voters to vote for Democrats down the ballot than for national Democrats.c  Will Kendrick, representing Dixie, Taylor, Madison, Hamilton and parts of five other counties is perhaps the only legislative Democrat from what might be considered a Dixiecrat area. Even so, it seems likely Kendrick draws much of his support from African Americans, something that Strom Thurmond and the Dixiecrats could never have achieved.  Based on the paucity of Democrats at any level, it would seem that the swing away from the FDR coalition and toward the Republicans that began in 1948 has about run its course and that the northern rural counties might better be described as Republican, rather than “Dixiecrat.” 

 

3.2.2    Applying the crossover ratio longitudinally

The question comes down to whether the voting patterns of rural north Florida are statistically unusual or whether they simply represent the conservative drift common to much of the South.  The “Dixiecrat hypothesis” can be framed as an issue of measuring crossover between the presidential vote and races down the ballot. 

The crossover ratio provides an excellent comparison of electoral performance across different precincts or counties, but it suffers some limitations in comparing performance longitudinally through time.  Candidates change from year to year, and their relative strength in different areas will depend on many factors.

Even with this limitation, one can calculate slopes for crossover and determine whether there is a trend or a stable pattern to the crossover ratio.  With only three data points, the uncertainty in slope is high and conclusions necessarily weak. However, Santa Rosa, Okaloosa, Collier, Martin, Miami-Dade, Hendry, Indian River, Clay, Osceola, Monroe, and Seminole may show a trend by more than one standard deviation toward higher crossover ratios, while  Sumter, Suwannee, Franklin, Holmes, Broward, Sarasota, Dixie, Gadsden, and Lafayette may show a trend toward lower crossover ratios. The remaining counties have to be regarded as stable. 

 


Figure 3.2.2 Changes in normalized crossover.  This chart shows that the performance of the Democratic presidential candidate relative to the senatorial fell in Broward County, but rose in Okaloosa County over the period 1992-2004.

 

Also interesting is the result obtained by correlating the performance of each county with counties that reliably go for the Republican or the Democratic presidential candidate. Selecting Alachua, Broward, Gadsden, Jefferson, Leon, Miami, Palm Beach, St Lucie, and Volusia as Democratic counties, and selecting Escambia, Jackson, Lake, Lee, Okeechobee, and St. Johns as Republican counties, one discovers as expected that some counties correlate with the Democratic counties and against the Republican and some correlate with the Republican and against the Democratic. But some counties (Bay, Flagler, Nassau, Osceola, and Pinellas) correlate somewhat with both parties and one (Hardee) even anti-correlates to both.

 

3.2.3    The issue of bias by office level in the crossover ratio

There’s an additional point in interpretation. On average, more voters vote for president than vote for senate, independent of political party. As long as this tendency is small and distributed reasonably homogeneously, there’s no significant consequence to calculation of the crossover ratio. In the 2004 Florida election, total presidential votes exceeded senatorial by about 2%, while in 2000, it was about 1% and in 1992, about 3%. This introduces a small intrinsic upward bias into the crossover ratio but, ceterus paribus, this bias is removed by normalization.

These average vote totals, however, mask wide variations by year and location. The dispersity has sharply narrowed, with the standard deviation of total presidential/senatorial votes for the 67 counties falling from 6.4% of the mean in 1992 to 3.0% in 2000 and just 0.7% in 2004. Individual counties also show deviations from the mean. Bradford County recorded only 1% greater presidential than senatorial votes in 2004 but 4% fewer presidential than senatorial votes in 2000 (in 1992, the average was also below the mean, but only slightly).

An assumption implicit in this is that the tendency toward voting preferentially by office is distributed evenly in a geographic sense. This assumption could fail if, for example, rural voters tended were energized mostly by races other than the presidential. The narrow dispersity of 2004, however, suggests that this assumption held in that year. And, indeed, the Z-scores of only a few counties, averaged over the three elections, exceed 1 in magnitude (Bradford low; Broward, Martin, Miami-Dade, Sarasota high). So, the assumption that there is no intrinsic tendency of any county to turn out unusually high or unusually low values of the presidential/senatorial ratio is probably generally true.  

The direction in which the assumption might fail can be determined by looking at those counties recording more votes for president than for senate, on average,c by at least one standard deviation. These were Bay, Broward, Calhoun, Franklin, Lee, Martin, Palm Beach, Pasco, Pinellas, Sarasota, and Sumter.  A number of counties recorded fewer votes for president than for senate by more than one standard deviation. These were Baker, Bradford, Dixie, Gilchrist, Hernando, Jefferson, Madison, and Santa Rosa. In many cases, these excursions from the mean are the result of a single year of unusual enthusiasm for candidates at either the presidential or senatorial level and probably have little importance. However, most of the counties in the group which showed a high president/senate ratio are urban and outside the north, while most of the counties in the group with a low president/senate ratio are rural and in the north.

What this means is that if there is any bias in the crossover ratio due to a preferential interest in certain races over others, it is toward northern, rural counties tending to have slightly lower crossover ratios than other counties. This tendency - if it exists - would be due to failing to vote for either candidate at the presidential level.  At any rate, the effect of bias by office level on geographic bias in the crossover ratio would appear to be small. 

 

3.2.4    -Counties exhibiting possibly unusual crossover

The very first test to establish whether a group of counties is in some way different than others is a test of the normality of the distribution. In 1992 and 2000, deviations of the crossover ratio from normality were not statistically significant, although in 1992, it was nearly so.  In 2004, the distribution was slightly skew, with the Shapiro-Wilk probability marginally significant at 0.03.d A more detailed look shows a single low-end near outlier in 2004. If Lafayette and Holmes (which is nearly a near outlier) counties are eliminated from the calculation, the distribution becomes entirely normal.  In 1992, all but one of the outliers turn out to be at high crossover ratios in urban counties.  So, by the basic screening test used to decide whether there is anything unusual about a set of data, support for the hypothesis that the crossover ratios of some group of counties are unusual is not at all clear. 

Table 3.2.4.1, showing counties with unusual values of normalized crossover argues against the notion that the voters of north Florida are markedly different in their presidential preferences relative to their senatorial preferences.  The counties marked yellow are one standard deviation below the mean in crossover, while the counties marked red are two standard deviations below. The counties marked green are one standard deviation above the mean, while those marked blue are two standard deviations above. 

It is true that a number of rural north Florida counties show presidential/senatorial crossover ratios below the mean and that no urban, southern counties show crossover ratios below the mean.  However, in order to declare a value “different,” the usual requirement is that it be three standard deviations from the mean.  Only Holmes, Lafayette, and Santa Rosa counties are more than two standard deviations from the mean in any given year.  Of those, only Lafayette appears at greater than two standard deviations in more than one year.  Of counties that are more than one standard deviation from the mean, only Baker, Liberty, and Union are present in all three.  Baker, Liberty, Holmes, Lafayette, and Union show sufficiently unusual behavior to suggest they might be truly different from the rest of Florida’s counties - but only possibly so.  

Of course, there are confounding factors, especially spoilage and local preferences for a particular candidate. But until these are sorted out, one cannot say with confidence that the “Dixiecrat counties,” however they are defined, are obviously different from other conservative counties in their propensity to cross over at the senatorial level.    

 

Table 3.2.4.1 Counties with statistically unusual normalized crossover

Counties marked in blue have normalized crossover two standard deviations above the mean, green one standard deviation above the mean, yellow one standard deviation below the mean and red two standard deviations below the mean

 
 
 
 
F
A
R
T
H
E
R
 
F
R
O
M
 
M
E
A
N

1992

2000

2004

 
 

Broward

Martin

Miami-Dade

 

Gadsden

Miami-Dade

Collier

 

Sarasota

Lee

Osceola

 

Palm

Collier

Lee

 

Alachua

Broward

Charlotte

 

Sumter

Charlotte

Monroe

 

Pasco

Sarasota

Palm Beach

 

Duval

St. Lucie

Orange

 

Flagler

 

Broward

   

St. Lucie

   

Baker

 
   

Calhoun

 
   

Taylor

Wakulla

   

Bradford

Gilchrist

   

Gilchrist

Taylor

   

Dixie

Union

 

Liberty

Liberty

Hardee

 

Union

Suwannee

Baker

 

Wakulla

Union

Suwannee

 

Clay

Gulf

Liberty

 

Baker

Washington

Dixie

 

Okaloosa

Holmes

Holmes

 

Santa Rosa

Lafayette

Lafayette

 

3.2.5   -Longitudinal studies of crossover e

In principle, however, there may be a means of adapting the crossover ratio to longitudinal studies.  Using a technique called correlation, it is possible to characterize other counties as having voting patterns that are directly correlated or anticorrelated, weakly correlated, or uncorrelated to those of a reference county.  In other words, the voting patterns of the other counties could be classified as resembling the reference county, as diametrically opposed to those of the reference county, or as unrelated to the reference county. 

If the voting patterns of the reference county were very unstable, perhaps as a result of drastically changing demographics, the results might not be very meaningful. But if the voting patterns of the reference were reasonably stable across time, very interesting results could be obtained. For example, suppose the only factor in voter decisions was a candidate's political liberality or conservativism. Further suppose that the reference county were 60% conservative and 40% liberal.  Then, if the presidential candidate were liberal and the senatorial candidate conservative, the crossover ratio would be 40/60 or 0.67, while if the presidential candidate were conservative and the senatorial candidate liberal, the crossover ratio would be 60/40, or 1.5. A county that was 60% liberal and 40% conservative would show precisely opposite crossover ratios. If their choices were driven by conservative/liberal preferences, and those were stable over time, the two counties would anti-correlate, that is, move in opposite directions. 

Selecting arbitrarily nine counties in which John Kerry received over 50% (Alachua, Broward, Gadsden, Jefferson, Leon, Miami-Dade, Palm Beach, St Lucie, and Volusia) to serve as a Democratic reference, and six counties where George Bush did well in 2004 (Escambia, Jackson, Lake, Lee, Okeechobee, and St. Johns) to serve as a Republican reference, the unnormalized crossover ratio correlates to both Democratic and GOP counties, with the minimum observed value being 0.89. Since the major contributor to significance is presumably the difference between 1992, when a third party candidate strongly affected the presidential vote, this means (unsurprisingly) that the candidate roster is a key factor in the crossover ratio.

With normalized data, in which the importance of third party candidates is minimized, some clear patterns emerge, and there are a number of surprises. Figure 3.2.5.1, below, shows that a number of counties correlate to the Democratic and/or to the Republican counties in crossover. But a small number of counties anticorrelate to either or both. This suggests, unsurprisingly, that the individual candidates, not party, is the most important factor in crossover.

 



Figure 3.2.5.1 - Correlation of votes in 1992, 2000, and 2004 of individual counties to a group of eight counties that voted for the 2004 Democratic presidential candidate (Alachua, Broward, Gadsden, Jefferson, Leon, Miami-Dade, Palm Beach, St Lucie, and Volusia), shown in blue, and to six counties that voted strongly against that candidate, shown in red (Escambia, Jackson, Lake, Lee, Okeechobee, and St. Johns)

 

In the table below, Table 3.2.5.1, the counties shown are those ten at extremes above or below the mean.f  Blue indicates counties that voted more than 50% for Kerry, green indicates counties that were 45-50% for Kerry, yellow indicates 35-45% for Kerry, and red indicates less than 35% for Kerry.  Correlation with Democratic counties means that the correlation coefficient of the normalized crossover coefficient was near 1 for a given county vs. Democratic counties, while correlation against Republican counties means that the correlation coefficient was nearer -1.  

In simple terms, a correlation near 1 means that a county agreed with Democratic counties as to whether the senatorial or the presidential candidate were the stronger, while a correlation near -1 means that a county agreed with Republican counties as to which was the stronger. Table 3.2.5.1 suggests that there are at least two kinds of rural, conservative, northern counties. Bay, Bradford, Gulf, Union and Washington - even though they vote Republican - see the strength of the senatorial candidate relative to the presidential the same way that Democratic counties do. Dixie, Hardee, Holmes, Lafayette, and Suwannee see the relative strengths least like Republican counties - but not as Democratic counties do. Sarasota, a more urban and Democratic county, and Hardee, a small conservative county have the distinction of being aligned against both parties. 

 

Table 3.2.5.1 Correlation of Normalized Crossover Fractions.f Blue: counties voting more than 50% for Kerry. Green: counties voting 45-50% for Kerry. Yellow: 35-45% for Kerry. Red: less than 35% for Kerry 
CROSSOVER CORRELATION To Democratic counties Against GOP counties
  Presidential/Senatorial Presidential/Senatorial
 
L
L
 
I
I
 
K
K
 
E
E
 
 
D
R
 
E
E
 
M
P
 
O
U
 
C
C
 
R
L
 
A
I
 
T
C
 
I
A
 
C
N
 
 
C
C
 
O
O
 
U
U
 
N
N
 
T
Y
 
I
I
 
E
E
 
S
S
 
     
Pinellas
Union
Jackson
Bradford
Bay
Washington
Jefferson
Gulf
Duval
Alachua
Sarasota
Collier
Lee
Clay
Baker
Hernando
Wukulla
Martin
Hardee
Highlands
Hardee
Sarasota
Dixie
Sumter
Lafayette
Liberty
Holmes
Suwanee
Franklin
Taylor
 
Glades
Escambia
Nassau
Walton
Volusia
Hendry
Marion
Lake
Osceola
Hillsborough

 

Most surprising is how many counties align reliably neither with Democratic nor with Republican counties. These observations suggest that models of voting patterns such the “Dixiecrat hypothesis” are overly simplistic and that voter behavior is far more subtle and interesting than such models would imply.

 

3.4     Senatorial trends

A further check interpretation of voting patterns is to look at senatorial trends. Using normalized Democratic performance according to eq. 2.0.1, the slope vs. year of Union and Miami-Dade were two standard deviations low, representing a sharp swing toward Republican senatorial candidates. Baker, Dixie, Gilchrist, Hendry, Nassau, Putnam, Sumter and Walton were one standard deviation low. Brevard, Citrus, Franklin, Hillsborough, Indian River, Manatee, Martin, Orange, Pinellas, Sarasota, Seminole, and St. Lucie were one standard deviation high. No counties were two standard deviations high.  Notice that a number of the counties swinging Republican at the senatorial level are the same counties that are swinging Republican at the presidential level. Such counties should be classified as showing a general conservative trend, not as exhibiting an unusual kind of voting. 

 

3.5    Races further down the ballot

It is more difficult to study races further down the ballot. Races very far down the ballot are, of course, run only in one or at most a few counties. These cannot be compared statewide. Also, since there are two separate US senatorial races but only one race for state level offices, the senatorial races occur with greater frequency than other offices. The gubernatorial races are off-cycle to the presidential.  Therefore, studies of crossover between president and other offices are more anecdotal than systematic.

However, some interesting observations can be made. For example, in 2000, there were races for Treasurer and Education Commissioner in addition to the presidential and senatorial races, from which additional crossover ratios can be formed. Rank-ordering of the deviations from the means of unnormalized crossover ratios shows Holmes and Lafayette two standard deviations below the mean in two cases and one standard deviation low in one case. Dixie is two standard deviations below the mean in one case and one standard deviation below in two. Counties that are one standard deviation below the mean in all three cases are Baker, Gilchrist, and Suwannee. Calhoun, Gulf, Hamilton, Liberty, Taylor, and Union show up one standard deviation below the mean in two cases.  Thus, Gilchrist and Suwannee, which are some of the most deviant cases, do not even show up on the Caltech/MIT list. Franklin and Washington, which show up on the Caltech/MIT list, cannot be considered unusual in crossover.

In more formal terms, analysis of variance (ANOVA) of these data, unnormalized, tells us that there is a significant difference between counties, but a far more significant difference between races. To express it one way, the mean squares of the difference of political office is 32 times greater than that of difference of the counties. This ratio stays almost the same if one looks at the 1998 race, using the governorship as the reference race.

Or, since we know that Lafayette tends to be an outlier in crossover, one could ask the (post hoc) question, “Are there any counties whose crossover is not distinctly greater than Lafayette’s?” Expressed as a one-tailed paired t-test, seven other counties (Baker, Dixie, Gilchrist, Hamilton, Holmes, Madison, and Suwannee) meet this very lenient test. Four of the six are not on the Caltech/MIT list at all.  When the same procedure is applied to 1998, using the crossover to Governor, a different list of counties that are not distinctly different from Lafayette can be discerned. These are Baker, Bay, Bradford, Clay, Collier, Dixie, Nassau, Okaloosa, Santa Rosa, St. Johns, and Walton.

The point here is that any reasonably unbiased test is likely to come to different conclusions as to which counties exhibit unusual crossover than one might expect from the Dixiecrat hypothesis. This is because the primary issue in voting is - as the ANOVA reported above indicated - the candidate.  Not the party, although political party is a useful shorthand for a set of positions comprising a candidate’s worldview. Not the county, though counties may differ in the degree to which polarization inhibits crossover. And certainly not racial issues per se, although doubtless those enter into the assessment of the party and the candidate. 

 

 

4.0  Summary of findings. Certainly race plays a role in Florida politics. It is a factor in all American politics. For example, African Americans hold only one Senate seat and Hispanics hold just two.  However, for the term to have any meaning, “Dixiecrat” voting patterns must be quantifiable. One would expect that Dixiecrat voting would, as in 1948, be expressed as a clear and consistent difference in party preference between the presidential race and races down lower on the ballot. This means that in a longitudinal study, one must disentangle preferences for individual personalities and larger trends in party preference from a shift in the crossover tendency. 

Martin, Orange, Seminole, and Palm Beach are unambiguously trending Democratic at the presidential level.  Sumter County and a cluster of counties on the eastern side of the Panhandle (Union, Suwannee, Levy, Lafayette, Gilchrist, and Dixie) are unambiguously trending Republican at the presidential level.  Some of those (Dixie, Gilchrist, Sumter, and Union) are probably also trending Republican at the senatorial level, reducing or perhaps reversing effects on crossover.

However, only a few counties can be said to show weaker Democratic presidential performance than senatorial performance. The Shapiro-Wilk test can be regarded as definitive in this regard. Based on distance from the mean in section 3.2.4, Baker, Holmes, Lafayette Liberty, and Union might be regarded as different, but the evidence is painfully weak. One might as easily conclude that the northern rural counties are entirely normal and that it is the urban counties that have unusually high crossover ratios. The crossover behavior of a number of conservative northern counties correlated well with Democratic counties. Other such counties anticorrelated to Republican counties. No rural northern county strongly anti-correlated to Democratic counties.

Examining races down the ballot in a (necessarily) anecdotal rather than systematic way, Dixie, Baker, Gilchrist, Holmes, Lafayette, and Suwannee might be regarded as different from other Florida counties based on the 2000 race.  Applying an extremely lenient (one-tailed) post hoc test, Baker, Dixie, Gilchrist, Hamilton, Holmes, Madison, and Suwannee can be grouped with Lafayette, the one county most likely to be a genuine outlier. None of these tests matched very well with the counties claimed to be Dixiecrat counties by Caltech/MIT. 

So, the Dixiecrat effect, if it exists, must manifest itself even further down the ballot than at the senatorial level. Measuring trends over different counties, not to mention over different election cycles, however, is difficult for offices other than statewide offices. Having eliminated the senatorial rung, the Dixiecrat hypothesis appears to be either false or effectively untestable. As was pointed out, the proximity between the rural northern counties and Tallahassee with the implications on media and candidate visits is as good, if better, an explanation for unusual crossover than the Dixiecrat hypothesis. Most important, the Dixiecrat hypothesis cannot explain unusual crossover within counties, especially since unusual crossover has been found in counties that are not regarded as Dixiecrat, such as Hillsborough, Duval, and Escambia.

Other evidence runs against the Dixiecrat hypothesis as a viable explanation of north Florida voting patterns. As previously mentioned, there are very few Democrats holding state or federal office in north Florida, and of those, probably only Will Kendrick could be considered a classic Dixiecrat. Even he probably relies heavily on African American support. Therefore, a better characterization of rural north Florida may be that it has been trending Republican at all levels ever since at least 1992. This is unsurprising, since it is historically a conservative area. Perhaps it might be more correct to characterize the Clinton years as a temporary defection of rural north Florida toward the Democrats.

 

5.0   Post Hoc Analyses. Post hoc analysis - gathering data, coming to a conclusion, and then formulating a statistical hypothesis for testing - is what is referred to by the saying (often attributed to Disraeli or to Mark Twain but probably actually from Leonard Henry Courtney) that there are “lies, d-ned lies, and statistics.”[5] However, intellectual integrity demands that one make the best case possible for the other side. In the case of the Dixiecrat hypothesis, the only way to do that is to go to post hoc analysis.

One can indeed construct post hoc analyses to show that the ten counties named by Caltech/MIT as Dixiecrat have a mean (unnormalized) crossover different than other counties.  The significance (two-tailed, assumed heteroscedastic) seems to be high, at least in 2000 and 2004 where it is in the realm of 10-5 to 10-6; the 1992 value is marginally significant at 0.02.  So, intellectual integrity also requires one to note that there are over 1011 ways to group 67 counties in two groups, one of 10 and one of 57. In other words, there are so many ways to partition the counties that one could arrive at this result by chance - even three times running.  

To illustrate just how susceptible to manipulation arbitrary partitioning can be, suppose that an alternate hypothesis is brought forward: that in certain counties, people cross over much more strongly toward the Democratic presidential candidate than in the rest of Florida.  It is noticed that crossover ratios tend to be high in Broward, Charlotte, Collier, Flagler, Martin, Miami-Dade, Palm Beach, St. Lucie, Sarasota, and Volusia counties. A catchy slogan to describe this voting behavior, say, The New Progressive Movement, is coined.  The probabilities for the comparable t-tests say that this hypothesis is as likely to be correct as the Dixiecrat hypothesis in 1992, but 20 times more likely in 2000 and 120 times more likely in 2004!  So, it’s in some ways a better hypothesis than the Dixiecrat hypothesis, but ultimately no more sound.  

Finally, for completeness, suppose we examine the counties that were deemed by the ad hoc procedures of section 3.2.4 to be possibly different than the rest of Florida: Baker, Liberty, Holmes, Lafayette, and Union.  These show up with probabilities on the order of 1 in 10-3 to 10-4.  With almost ten million ways to partition 67 counties into a group of 62 and a group of 5, even this best effort falls far short of persuading.

 

6.0   Conclusion. There is a rule of thumb in statistics that if one needs statistics to prove a hypothesis, one should get a better hypothesis. The Dixiecrat hypothesis - at least as it applies to presidential/senatorial crossover - is, at a minimum, not clearly evident from the data. Nor, again in reference to presidential/senatorial crossover, can it be said to constitute a historical trend or pattern.  Rather, the voters of Florida, in all their wonderful variety of outlooks, ideals, and interests, vote for candidates as individuals.  There may be areas of the state in which crossover voting is more common but, if so, that tendency is buried among many others.

As for the deviations between registration and voting, one correspondent made the sensible suggestion that Democratic primaries, especially in counties where Republicans or Republican-leaners are an overwhelming majority, are the only interesting local races.[6] Given what we know, this is as good an explanation as any.


7.0  Note added in proof. g It was pointed out [7] that an (arbitrary) criterion had been offered to define Dixiecrat counties, namely the disparity between a party’s share of the registration and its share of the presidential vote. There is certainly no disagreement that in certain counties, the Democratic Party significantly underperforms at the presidential relative to its share of the registration. A logical measure to use in examining this is:

(Presidential vote share – Registration Share)/Registration share eq. 7.0.1
   

This is positive if a party captures a greater fraction of the vote than its share of registration and negative if it captures less of the vote than its share of the registration. It is arbitrarily designated here as the capture statistic. There is no disagreement that a number of small and midsize northern counties seem to show a negative capture statistic.  In 1996, 2000, and 2004, Baker, Calhoun, Dixie, Gulf, Hardee, Holmes, Lafayette, Liberty, Suwannee, Taylor, Washington, and Union showed a negative capture statistic of more than one standard deviation less than the mean in all three years.    

The problem arises in the interpretation. Unlike the crossover ratio, the capture statistic is not normally distributed.  As shown in Figure 7.0.1, it is better described as bimodal.  There are approximately 34 counties in the leftmost peak (most negative capture statistic) of the Figure. These include a number that are not rural, not small, not northern, or in a couple of cases, about half African-American. The use of nonparametric methods is strongly indicated, and parametric methods are probably invalid, a point that seems to have been missed entirely by all parties to the debate.  Similar results are obtained if one uses the simple difference between vote and registration.



 

Figure 7.0.1 Distribution of the capture statistic for the 1996 election. The blue line is what a normal distribution would show. The distribution was non-normal (p< 0.0006), nominally due to skewness (p<0.0001).  The 2000 and 2004 distributions were similar, though not as pronounced.  No data points were outliers in any year.

So, which group is really unusual for Florida?

There is, indeed, a strong correlation between the crossover ratio and the capture statistic. But due to the bimodality of the capture statistic, this amounts to drawing a line between two data points, and therefore means very little.  

As was correctly raised in criticism [8] of papers by authors from US Count Votes [9] (and relevant to other analyses),[10] Florida counties are demographically clustered, so there is distributional inhomogeneity. While researchers have tended to focus on variables, such as population, income and education to describe the clustering, sometimes looking at the facts on the ground is more helpful.  In this case a desire to have some competitive races in an area of one-party domination may be a better explanation for what is observed in terms of capture. At any rate, it’s not really surprising to find that one can partition some counties into this group or that and discover that they are different from one another.

Of course they are.

But what’s the connection to racism? 
  




All data and calculations are available on request by contacting the author.
OliverDawshed[at]aol.com
or

 

Acknowledgements.  Data were received December 6, 2004 from Vincent Lipsio, attributed to Jon M. Ausman.  Helpful comments were received from Kathy Dopp, Dr. Gus H. Miller, Paul Lukasiak and Josh Mittledorf. Thanks, of course, to M. E. Cowan for hosting the webspace in which this appears and to Elizabeth Jordan for design. 

 

Footnotes

a In the 1992 election, Perot took between 9.9 (Miami-Dade) and 29.0 (St. Lucie) percent of the vote. The dispersity was very low (4.0 percentage points on a mean of 22.3 percent). In 49 counties, the Perot vote was within a standard deviation of the mean. Few of the counties where Perot support was a standard deviation greater than or less than the mean were in the rural north, those being Gadsden, Gilchrist, Taylor, and Wakulla. The debate on how much of that vote was drawn from Republican-leaning voters and how much from Democratic-leaning voters has been a subject of spirited debate. For this reason, one can certainly debate what the impact on the normalized presidential vote would be. However, on the assumption that equal percentage points came from both candidates, the net effect on the normalized presidential value would be approximately zero. Indeed, a Gallup survey evidently showed that, had Perot been removed as a candidate, roughly a third of Perot voters would have voted for Clinton, a third for Bush, and a third would have spoiled their vote.[11]

b One example should suffice. Suppose that Democratic performance ranges from 80% to 20% in years 1, 2, 3 and 4. In one county, Democratic performance drops 10% each year. In a second county, it drops 30% between year 1 and year 2, and then stays constant. The first slope is -0.17. The second is -0.15. So, equal declines in performance lead to different results depending on the details of when decline occurs. But also note a key virtue of this method of calculation. Suppose that in each of several years, Democrats nominate candidates that are progressively less well known, such that in each county, support drops equally. Then the relative rankings of the counties will remain the same. This correctly reflects the fact that there is not really a weakening of Democratic strength but a temporary effect that can be reversed by nominating a candidate with good name recognition.

c To perform this calculation, take the difference between a value in a given year and the mean for that year. The three values are then averaged, and the result compared to the standard deviation for the results of all counties. In the paragraph above, by contrast, the difference between the value and mean is taken and divided by the standard deviation before averaging.

d The astute reader knows that if tests of significance are done on more than one group of data, it becomes increasingly likely that one or more of those samples will fail a test of significance by random chance. This is the case here. When one case is mildly significant because it has low outliers, a second case is almost significant because of high outliers, and a third case is not statistically significant, it's more likely that case-to-case variations are more important than variations within a single dataset.

e Because data from only three elections were used in this analysis, conclusions in this section are necessarily tentative.

f Because the distribution is very far from normal, the standard deviation is very large and no counties are more than one standard deviation above the mean for the GOP correlation. The choice of the ten most extreme counties was purely arbitrary. Also, as is evident from Figure 3.2.5.1, there were few counties that actually anticorrelated to GOP counties and very few counties that actually anticorrelated with Democratic counties. Perhaps half the counties listed as the ten most extreme in anticorrelation would be better described as not correlating.

g Data for calculations of the capture statistic are drawn from Mebane.[1]


References


1.

Mebane, W. R., Jr. Letter to CommonDreams.org. November 8, 2004.

Mebane, W. R., Jr. Letter to K. Dopp. , November 12, 2004.

Mebane, W. R., Jr. personal communication. December 27, 2004. 

2.

Jordan, E. and O. Dawshed. "Bush’s Fifth Ace: A Crooked Panhandle." Online Journal, July 2001.

3.

Dawshed, O. "A Model for Interpreting Voting Patterns with Application to Florida" Rev. 1.0, November 21, 2004.

4.

Caltech/MIT Voting Technology Project. "On the Discrepancy Between Party Registration and Presidential Vote in Florida." November 10, 2004.

5.

Courtney, L. H. “To My Fellow-Disciples at Saratoga Springs.” The National Review [London] 26 (1895) 21-26, cited at <http://www.york.ac.uk/depts/maths/histstat/lies.htm>.

6.

Lipsio, V. personal communication. December 6, 2004.

7.

Mebane, W. R., Jr. personal communication. February 17, 2005.

8.

Sekhon, J. “The 2004 Florida Optical Voting Machine Controversy: A Causal Analysis Using Matching.” <http://jsekhon.fas.harvard.edu/papers/
SekhonOpticalMatch.pdf
>,

November 13, 2004. updated November 13, 2004.

9.

Liddle, E. untitled. <http://uscountvotes.org/index.php?option=com_
content&task=view&id=32&Itemid=43
>,
November 13, 2004.

Mittedorf, J. untitled. <http://uscountvotes.org/index.php?option=com_
content&task=view&id=31&Itemid=43
>
, November 13, 2004.

10.

Trexel, D. “2004 Election Anomalies in Florida - Was There a Major Panhandle Democrat Defection?<http://www.socsci.umn.edu/~trex0003/FL2004.html>,
downloaded February 20, 2005.

11. See, for example, Cranor, L. “Case Study:  The 1992 US Presidential Election.” <http://lorrie.cranor.org/pubs/diss/node19.html>,
downloaded February 9, 2005.


Appendix: Description of the data.  Data for the capture statistic were drawn from Mebane.[1] Voting data for 1992-2004 were obtained from Vincent Lipsio.[6]  In most cases, percentages of the vote were used to compute ratios. To examine simple time series in presidential and senatorial races, however, raw vote totals were used. The races that were used in calculations follow.

1992:   Clinton/Bush/Perot (Presidential) and Graham/Grant (US Senate)

1994:   Rodham/Mack (Senate), Chiles/Bush (Governor), Saunders/Mortham (Secretary of State), Butterworth/Ferro (Attorney General), Lewis/Milligan (Comptroller), and Nelson/Ireland (Treasurer)

1996:   Clinton/Dole/Perot (President)

1998:   Graham/Crist (US Senate), MacKay/Bush (Governor) Gievers/Harris (Secretary of State), Butterworth/Bludworth (Attorney General), Daughtrey/Milligan (Comptroller), Nelson/Ireland (Treasurer), Wallace/Gallagher (Education Commissioner), and Crawford/Faircloth (Agriculture Commissioner)

2000:   Gore/Bush (Presidential), Nelson/McCollum (Senatorial) and Cosgrove/Gallagher (Treasurer)

2002:   McBride/Bush (Governor), Crist/Dyer (Attorney General), Bronson/Nelson (Agriculture)

2004:   Kerry/Bush (Presidential) and Castor/Martinez (Senatorial)



Copyright, © 2005 Oliver T. Dawshed. All Rights Reserved.
Permissions for reproduction or extensive quotation may be obtained from
.