Whose Votes Don't Count?: An Analysis of Spoiled Ballots in the 2000 Florida Election

Philip A. Klinkner, Associate Professor of Government
Hamilton College 198 College Hill Road Clinton, NY 13323 315-859-4344
pklinkne@hamilton.edu

Continued from Whose Votes Don't Count?

This project began in response to media reports about the findings of the U.S. Commission on Civil Rights (USCCR) that indicated higher rates of ballot spoilage in Florida counties with larger numbers of blacks. I was intrigued by this result, so I decided to run my own independent analysis of the data. I should mention that I have no official connection to the USCCR, but I have met two of its members, Abigail Thernstrom and Professor Christopher Edley. I did not, however, contact them before undertaking this analysis. In addition, I did not know Professor Alan Lichtman who conducted the USCCR analysis. Since arriving at my findings, I have spoken by telephone with Professor Lichtman and Professor Edley to inform them of my conclusions. In summary, these findings are mine and mine alone.

As a first step, I obtained data on the dependent variable-the rate of spoiled ballots in each of Florida's counties. This information came from the Governor's Select Task Force On Election Procedures, Standards and Technology, conducted by the Collins Center for Public Policy (the report is available at http://www.collinscenter.org/info-url2660/info-url.htm).

The next step was to consider the different independent variables that might explain the differential rates of ballot spoilage. Among the list of possible suspects are the following:

Different types of voting systems

The type of voting system is indicated by four variables. Op/P refers to optical scan systems in which ballots are read at the precinct where the vote is cast. Op/C refers to optical scans systems in which ballots are collected from individual precincts and read at a central location. Punchcard refers to the now infamous punchcard voting systems. Other refers to the two counties using different types of voting systems. One county uses the lever-machine system and the other uses paper/hand ballots. This information was obtained from the Governor's Select Task Force On Election Procedures, Standards and Technology, conducted by the Collins Center for Public Policy (the report is available at http://www.collinscenter.org/info-url2660/info-url.htm)

Turnout:

Data on the votes in each county is available at the Florida Elections Division website: http://election.dos.state.fl.us/elections/resultsarchive/Index.asp

Gore %:

% Hispanic

Median Income

Literacy

Education

% Black Registered Voters

Voters per Precinct

Increase in Registration

Information on registered voters in 1996 is available from the Florida Elections Division website: http://election.dos.state.fl.us/voterreg/vrArchive/1996voterreg.shtml#General

Increase in Voting

Other variables

• county crime rates,
• percent of elderly population,
• percent of population under 25,
• party of the county election supervisor,
• percent of population with less than a high school diploma,
• percent of population with some college education,
• percent of population in rural areas,
• percent of population English-only speakers,
• county population density,
• percent of blacks with less than 9th grade education,
• percent of blacks with less than a high school diploma,
• percent of blacks with some college education,
• increase in percent of registered voters who are black from 1996 to 2000

I then ran a regression model using the fourteen independent variables previously listed. The regression was run using SPSS 10.0 for the Macintosh. The results are as follows:

The results are available as a pdf Klinkner Analysis and Florida Data.

As the model shows, the following variables were not significant:

1. % Hispanic
2. 1989 Median$
3. Level 1 Literacy ¹
4. % <9th
5. 96-00% Increase in Total Reg
6 % Increase Vote 96-00

Conversely, the following variables were statistically significant at the .05 level or greater.

1. Op/C
2. Op/P
3. Punchcard
4. Other
5. Turnout
6. Gore %
7. % Black Reg Voters
8. Voters/Precincts

Interestingly, education and income appear to have no effect on the rate of spoiled ballots. Thus there is little evidence in the data for the claim that spoiled ballots in Florida resulted mostly from the individual errors of voters who lacked the education or experience to cast accurate votes. This fact can be seen in the following regression, which I call the "stupid voter" model. In it, I use only those factors that are under the control of the voter: literacy, education, poverty (arguably under the control of individuals, but I'll put it in anyway), and first time voters (measured by the increase in registration from 1996 to 2000 and the increase in votes from 1996 to 2000).

As the regressions for the "stupid voter" model indicates, these individually controlled variables exert little explanatory power. The adjusted r² is only .452, meaning that the model explains less than half the variance in the pattern of spoiled ballots across Florida counties. Furthermore, none of the variables, save for literacy is statistically significant. This model and the previous one show quite clear that the pattern of spoiled ballots in Florida was much more influenced by systemic factors rather than individual ones.

I then re-ran the model using only the statistically significant variables. The results are as follows:

The previous model seems to do a good job of explaining the rate of ballot spoilage in Florida counties. The adjusted r² is .85, indicating that 85 percent of the variation between counties in the rate of spoiled ballots can be explained by these variables.

The model shows that the type of voting system used in a county has a clear impact on the rate of ballot spoilage. In addition, so does the level of turnout, Gore's percent of the vote, the percent of registered voters who are black, and the number of voters per precincts. The fact that higher levels of turnout and more voters per precinct are negatively associated with ballot spoilage seems a bit counterintuitive. On the other hand, higher levels of turnout might also reflect greater political interest and knowledge, and thus probably mean less chance of spoiled or mistaken ballots. With the number of voters per precinct, the result is probably due to the fact that where counties have reduced the number of precincts, they will likely have more poll workers and election officials per precinct, thus making it easier for voters to obtain help in filling out their ballot.

To more precisely the impact of different voting systems on the other variables I set up several interactive variables by multiplying the following variable with one another:

Op/P * Turnout
Op/P * Gore%
Op/P * % Black Reg Voters
Op/P * Voters/Precincts
Op/C * Turnout
Op/C * Gore%
Op/C * % Black Reg Voters
Op/C * Voters/Precincts
Punch * Turnout
Punch * Gore%
Punch * % Black Reg Voters
Punch * Voters/Precincts
Other * Turnout
Other * Gore%
Other * % Black Reg Voters
Other * Voters/Precincts

I then re-ran the model using the existing variables and these new interactive variables. After dropping out the non-significant variables, I came up with the following results:

Thus, the statistically significant variables that exercise and independent effect are:

1. Punchcard
2. Turnout
3. % of Black Reg Voters
4. Voters/Precincts
5. Op/P * Gore%
6. Op/P * Voters/Precincts
7. Punch * % Gore

In overall terms, the following factors led to increased levels of spoiled ballots:

1. Counties with punchcard ballots
2. Counties with higher percentages of black registered voters
3. Counties with Op/P votings sytems with higher numbers of voters per precinct

In addition, the following factors led to lower levels of spoiled ballots:

1. Counties with higher levels of turnout
2. Counties with more voters per precinct
3. Counties with Op/P voting systems with higher percentages for Gore.
4. Counties with punchcard voting systems with higher percentages for Gore.

One aspect of these regressions seems incongruous. According to the initial model, the percent for Gore negatively correlates with the percent of spoiled ballots. In other words, if a county has a higher percent for Gore, it is likely to have fewer spoiled ballots. On the other hand, the percent of registered voters who are black is positively correlated with spoiled ballots-counties with a greater percentage of black registered voters were likely to have more spoiled ballots. This finding seems odd, since the percent of black voters and the percent for Gore are moderately correlated (.415). Another way to look at it is that the model seems to suggest that ballot spoilage is likely to be highest in counties with high percentages of blacks and low votes for Gore.

This is an interesting finding, since if we were to suspect racial disenfranchisement in Florida, we would expect to find it in certain types of counties. Racial disenfranchisement would be least likely to take place in strongly Democratic counties. Election officials would have no incentive to disenfranchise some of their most loyal voters. Conversely, we would also be unlikely to find racial disenfranchisement in heavily Republican areas with very few black voters. Such tactics would yield little benefit and most likely they would be difficult to carry, since the black vote would be less concentrated and identifiable.

On the other hand, racial disenfranchisement would be most likely in counties where there is a significant or at least non-trivial percentage of black voters, but at the same time the county is strongly Republican. In such cases there are enough black votes to create political incentives for racial disenfranchisement and the black vote would be more concentrated and identifiable.

To test this possibility, I developed an additional interactive variable, the percent of registered voters who are black multiplied by the winning vote margin for George W. Bush.  This variable provides a good proxy for the types of counties where racial disenfranchisement is most likely to occur.  Counties that rank highest have a sizeable black vote and a larger Republican vote.  The lowest ranking counties have both a large black vote and go strongly Democratic.

I then re-ran the regressions including this variable.

As the tables indicate, the adjusted r2 is now quite high-the model explains over 92 percent of the variance in spoiled ballots. In addition, the percent of black voters remains significant. Finally, in areas where the combined result of multiplying the percent of voters who are black by the voter margin for Bush is positive, there is a positive correlation with spoiled ballots. To put it another way, not only does being black matter in the model, it also matters where you are black. Strongly Republican areas that also had a sizeable proportion of blacks had a greater incidence of spoiled ballots. While this finding is only suggestive, it is exactly what one would expect to find in a situation where racial disenfranchisement is likely to occur--black voters are a sizeable part of the electorate, but lacked the political power to ensure that their ballots are counted accurately and fairly.

In conclusion, this analysis offers two important findings:

1. There is no evidence that higher rates of spoiled ballots resulted from such individual factors as education and literacy. Instead, the factors influencing spoiled ballots were systemic. Thus, rather than speaking of individuals who spoiled their ballots, we should speak of individuals who were placed in situations in which it was more likely that their ballots would be spoiled. Furthermore, this finding indicates that any effort to reduce the rate of spoiled ballots must focus on systemic solutions--improved technology, more and better election workers, and stronger efforts to investigate and prosecute any instances of corruption and/or racial disenfranchisement.

2. Even after controlling for other factors, rates of ballot spoilage remain higher in predominantly black areas than in other areas of Florida. As the last model indicates, with all else being equal, for every 1-point increase in the percentage of registered voters who are black, there was a .07 percentage point increase in spoiled ballots.

In addition, these rates were even higher where substantial numbers of blacks were found in counties with large margins for George W. Bush. All of this corresponds to and further reinforces the findings of the USCCR that there is evidence of racial disenfranchisement in the 2000 election in Florida. Consequently, it is important that federal authorities should investigate this matter more thoroughly.

Addendum 1: Evaluating John Lott's Analysis

In response to the findings of the majority of USCCR, the minority members of the Commission submitted an alternative analysis vote spoilage in Florida by Professor John Lott (his report is available at http://www.manhattan-institute.org). According to Professor Lott's analysis, the percentage of black voters in a county has no statistical relationship to the percent of spoiled ballots in that county. Rather, the percent of people in poverty is a more important factor for explaining the rate of ballot spoilage. In this section I will evaluate Professor Lott's analysis.

Lott begins his analysis by suggesting that Lichtman's cross-sectional analysis is insufficient, arguing that if African Americans were more likely to spoil their ballots, then changes in spoiled ballots across time should closely correlate with changes in the percentage of African Americans across the same period of time. In a series of scatterplots, Lott shows that there is little if any relationship between the change in percent of spoiled ballots between 1996 and the change in percent of voters who are black between the same years. But Lott makes a critical error by assuming that all other factors that might influence ballot spoilage remained equal between 1996 and 2000. This is extremely doubtful. Consequently, even if increased percentages of black voters led to increased percentages of spoiled ballots for a particular county, this finding might not be apparent if, for example, that county moved to a more accurate voting system. Evidence of this can be seen in Table 1.

In each year, blacks spoiled their ballots at twice the rate of whites. In Year 1, blacks made up 50 percent of the county population and the rate of spoiled ballots was 7.5 percent. Year 2a represents what Lott expects to find. The percent of voters who are black rises from 50 to 55 percent, leading to a corresponding increase in the percent of spoiled ballots from 7.5 percent to 7.7 percent. But in Year 2a, nothing changes but the percent of voters who are black.

In Year 2b, however, two things change. In addition to the increase in the percent of voters who are black from 50 to 55 percent (as in Year 2a), the county also institutes a new voting systems, cutting the spoilage rate in half, from 10 to 5 percent for blacks and from 5 to 2.5 percent for whites. This change means that in Year 2b, the percent of spoiled ballots declines to 4 percent from 7.5 percent in Year 1, despite the fact that blacks now make up a larger share of the county's voters. Consequently, it is impossible to know the true relationship between change in the percent of voters who are black and change in the percent of spoiled ballots without controlling for other factors, especially voting systems.

Lott then goes on to develop a series of models that, he argues, better explain the rate of spoiled ballots in Florida counties. According to the results of his models, the percentage of black voters is not a significant factor in explaining the percentage of spoiled ballots in a county. His models also suggest a county's poverty rate has more to do with spoiled ballots than the racial and ethnic makeup of the county.

The ultimate test of any regression model is the amount of variance in the dependent variable that is explained by the independent variables. In this case, how much of the variation in the percent of spoiled ballots across counties is explained. On this score Lott's models are inferior to those that I have developed. Of the eight models listed in Lott's Table 2, the highest r2 is .7859, meaning that his independent variables explain almost 79 percent of the variation in spoiled ballots across Florida counties (I'll give Professor Lott the benefit of the doubt by assuming that these are adjusted r2's. If not, then his models have even less explanatory power). Furthermore, a model using only variables for the different types of voting systems for each county yields an adjusted r2 of .594. This indicates that the bulk of the explanatory power in Lott's models comes from these variables.

In comparison, even the simplest model that I developed has an adjusted r2 of .837 and my final model had an r2 of .927. Put in other terms, my final model is 17 percent more powerful in explaining the pattern of spoiled ballots than Lott's. Furthermore, I also entered several of Lott's variables, namely party of the county election supervisor, percent Hispanic population, and county median income into my earlier model and all came up as statistically insignificant.

Nonetheless, perhaps Lott is on to something with his variables and they might turn up as significant in my final model. Consequently, I reran my model and added in the variables for median income, percent of the population in poverty, and variables for the party of the county election supervisor.

As the tables indicate, adding in Lott's variables to my final model increases the unadjusted r2 only from .945 to .946, a trivial increase. Indeed, I added a random variable (the number of letters in the name of the county) and the r2 also increased from .945 to .946, indicating that the addition of Lott's variables adds no more explanatory power than the addition of a random variable. (In fact, the significance of this random variable was .283, making it more statistically significant that any of Lott's variables!)

Since the addition of any variable (even a random variable like the number of letters in the name of the county) will increase the r2 of a model, the more important statistic is the adjusted r2, which controls for the number of variables in the model. Adding Lott's variables to my model drops the adjusted r2 from .927 to .922. In other words, adding in the variables that Lott claims have the most explanatory power actually makes my model less, not more, powerful. Furthermore, none of Lott's variables is statistically significant, and their addition to the model causes only two of the original variables, Op/P*BushMargin and Other*BushMargin, to fall out of significance. The percent of registered voters who are black remains statistically significant and the correlation coefficient remains largely the same as in my earlier model.

In conclusion, Lott's findings do not hold up under scrutiny. Not only do they under-explain the variance in the rate of spoiled ballots, but when his variables are added to a more sophisticated model, they lack statistical significance. As a result, nothing in Lott's analysis detracts from the finding of the USCCR majority report or the analysis that I've offered here.

Addendum 2: First Time Voters

In the earlier models I used two proxies for first time voters: increase in registration from 1996 to 2000 and increase in the number of votes cast from 1996 to 2000. Neither variable was significant. Fortunately, I was just able to come up with the number of first time voters from the Florida Secretary of State Voter File (available on CD-ROM for free by calling the Florida Elections Division). These files contain data on every current voter in Florida, including their voting history since 1994. Using this data, I determined how many voters in each county voted only in the 2000 general election. While this would also included some non-first time voters (voters who voted prior to 1994 or voters who voted in other states by voted in Florida for the first time in 2000), it is a better estimate than the previous ones.

I reran my simple model and included a new variable for the % of first time voters. The results are as follows: