Prepared Remarks for
The Election Reform Task Force of the NCSL

Stephen Ansolabehere, MIT
April 28, 2001

Thank you for inviting me to speak today. As a scholar I've used many of the council's reports over the years, so I realize how important forums such as this are not just for our immediate needs but for the future.

I'd like to begin by telling you a little bit about our project, and then tell you about some specific findings.

A week after the 2000 presidential election, David Baltimore, the president of Caltech, called Charles Vest, the president of MIT, with an idea. Our two institutions should collaborate to develop improved voting technologies - a new voting machine. The problems observed counting the vote in Florida and elsewhere originated with technology.

President's Vest and Baltimore assembled a team of computer scientists, mechanical engineers, and social scientists. The Carnegie Corporation and our two institutes have funded our endeavors.

I'm Steve Ansolabehere, a professor of Political Science at MIT, and co-director of the Caltech/MIT voting technology project. My counterpart at Caltech is Tom Palfrey, a professor of Economics. Our team consists of 11 faculty and many students, and our central goal is to develop new voting technology. The engineers bring expertise in electronic security, user interface design, machine design, and performance standards. The social scientists bring expertise in voter behavior, operations design, and public finance.

We are in the initial phase of our project, which I consider the learning phase. Over the last four months we have met with many voting machine manufacturers and election administrators to ascertain what the problems are and to explore ways that we can contribute to solutions. We have also conducted studies of voting machine performance and design, the public finances of election administration, and voter registration practices. A complete report of our work over the last four months is due out at the beginning of July. It will include our assessment of existing voting processes in the United States. The report will also offer specific recommendations for the industry, governments, and universities to pursue.

The second phase of our project focuses on design. We've identified a number of user interface and security features of existing equipment that can be improved upon. We have identified specific practices in voter registration and polling place administration that can be improved at minimal cost or with cost savings with the use of computer technology. We have also identified the need for a process that would involve industry, government, and universities in continual innovation in voting equipment and software.

Like most Americans, I have always taken the voting system for granted, even though I have voted with every technology. With a little introspection, it is evident that computing technology has driven changes in voting technology. Today, we are in the midst of a computing and communication revolution, and that revolution will change the voting system over the next decade.

The question before us is how can we make the transition to new voting systems a good one.

There are three big pieces to the voting system where we see technology, and in particular computer technology, changing the picture. These parts of the system are registration, casting of votes, and counting of votes.

First, registration. The registration system serves two purposes: authentication and management. It is used to authenticate the voter. That is, to make sure that those who are not allowed to vote do not and that those who are allowed to vote do so once. Registration is also used to manage the ballots. We vote on so many different offices and questions today that it has become a chore simply to keep straight who should vote on what. Registration information tells people what polling place to go to and it allows the administrators to distribute the ballots to the right polling places.

Computing advances now afford improved database management. The cost of maintaining databases electronically is now quite low. The challenge is linking the voter registration information to other county and state databases. As counties and states invest in computerized databases, problems with redundant or obsolete registrations will be reduced significantly. We are especially impressed by the efforts in Kentucky, Michigan and Oklahoma, and we see these as best practices to build on.

Internet access and cheap electronic storage now permit precincts to access the central voter files (at the county or state). In correct registration information and inability to resolve questions about registration is a major source of lines at polling places. Experiences in some counties suggests that accessing the county or state registration information using laptops at polling places can reduce about 4 in 5 registration problems at the polls. This will relieve one important bottleneck.

Second, casting votes. The act of casting a vote is fundamentally communication. Voters need a way to communicate their preferences and intentions: dropping a chit in a dish, raising their hands or voices, marking a piece of paper, pulling a lever, or, now, touching a screen.

Since 1990, there have been important developments in understanding how to make computers more user friendly to the everyday person. Improvements in user interface design - the look and feel of electronic ballots - will improve the acceptance and usefulness of electronic voting machines. But we see problems with existing designs. The challenge is how to implement better ballot designs.

Third, counting votes. With the close of polls begins an enormous computing problem. A brief history of the technology reveals the importance of computational speed.

• Hand counted paper ballots are slow. NBC, CBS, and ABC just can't wait a couple of days.
• Lever machines speed up the count by aggregating many ballots at the precinct, leaving the administrator to tally the counts on the backs of the three or four machines at the polling place.
• Punch cards improved on this by providing and fast counts of all ballots at once.
• PC connected punch cards, scanners, and DREs permit the counts to be sent in via modem - faster still.

With speed has probably come greater confidence in the process and less opportunity to tamper with ballots.

Additional gains in the speed of counting are marginal at best. But there is a need for improved security and auditability of counts. With electronic counting we sacrifice the observability or visibility of the count. In most states this is done publicly, with representatives from the parties to check the counting. That check has been lost, and that check often caught problems. One challenge is to devise a new systems of automatic checks that would highlight suspicious looking counts.

Another facet of the computerization of tabulation should be noted. We now have "internet" voting. The platforms of many of the electronic equipment are serviced over the internet. For example, many of the new DREs download the ballot formats.

A final observation is that it is evident from Kim Brace's study of voting equipment is that we are trending away from mechanical and paper technologies and toward electronic technologies (optical scanning and electronic machines). This will force more rapid upgrading of equipment. First, electronic devices do not last as long as mechanical devices. Second, to exploit the flexibility of design on electronics the software designs will have to change more quickly. Third, memory cards and other devices will become obsolete quickly. Finding "spare parts" in 10 or 15 years for a computer built today may be quite difficult.

From election administrators we've heard that the number one problem is money, especially for smaller communities. Paul Craft of the FL Elections Division put it eloquently at a meeting we had at Cal Tech. He drew an enormous dollar sign on the black board. He elaborated that the more concrete problem is that election administrators face tradeoffs. They'd like equipment that is FAST, CHEAP, and ACCURATE. "You can only have two of these."

What Mr. Craft describes is a static picture of a boundary. Technology pushes those boundaries back. With advances in computing we may be able to gain greater speed, lower prices, and make machines more accurate, and by that I mean more secure and more user friendly.

Before we design, though, we must know what is. How accurate is the technology? How much does it cost and what is our capacity to purchase voting equipment?

When we began this project one of the first questions we sought to answer was how much the US spends on elections. We found audits of several counties and projections based on those counties, but we found no estimates of nation-wide expenditures. If anyone knows of studies of election administration spending nation wide, we'd appreciate any information you have.

To fill this void, we surveyed county administrators throughout the country by sending them faxes to ask how much they budget for election administration. This gives us a ball park estimate of the nation-wide expenditure on all aspects of elections. The data are still coming in, but I can share with you our preliminary findings. In the 2000 elections, the US counties and municipalities spent (on average) approximately $8.80 per voter on all election administration. That works out to approximately $1 billion nation wide.

This figure includes all expenditures -- voter registration, salaries, office overhead, equipment purchases, equipment maintenance and storage, poll worker training and pay. Ernest Hawkins, the Sacramento County registrar, has performed an excellent cost analysis based on that county's expenditures. His total figure is slightly higher, but not much.

How much of our total expenditures are equipment acquisitions today? Industry executives estimate that total voting equipment sales range from $150 million to $200 million per year. That represents a significant part of total election administration expenditures.

However, it is a small amount of revenue for an industry. Large firms stay out, but when they enter they bring significant design innovations. IBM and Unisys are cases in point. IBM was one of the first punch card innovators, but they got out of the business in 1968 because of fears of bad publicity. IBM spun off two companies, CES and EVM, which became two of the main punch card vendors. Unisys developed the Optech scanner, the most widely used scanner in the US, but withdrew from the industry because it was not sufficiently profitable.

What implications does the $1 billion figure have for potential upgrading of equipment?

Upgrading equipment is expensive. Various firms have given us price estimates. A touch screen DRE costs about $25/voter and a precinct optical scanning system costs about $6/voter. A nation-wide upgrade to DREs would cost about $2.5 billion; a nation-wide upgrade to scanners would cost about $600 million.

It should be noted, however, that over a 15 year span (approximate life of these machines), we would spend as a nation between $2.25 billion and $3 billion on machines anyway. That is, if revenues are between $150 annually and $200 million annually, then over 15 years we expect to spend 15 times the revenues. At today's machine prices a complete upgrade to DREs would be approximately $2.5 billion, which is in the range of what the counties and municipalities would spend anyway.

One concern is the effect of spending all of that money at once. The industry may not have the capacity to fill orders. Such a large infusion of cash might increase prices. And, an immediate and complete upgrade would kill demand for the next 5 years or so, which might kill the industry.

A second concern with an immediate upgrade concerns the public financing. Because equipment is mainly sold rather than leased, county budgets would have to absorb sizable capital costs. A separate capital request is required, which is often more difficult than a request for additional operating funds.

Upgrading equipment is particularly difficult for smaller counties. There are strong economies of scale in election administration expenditures. The cost per voter falls off quickly the larger the number of voters in a jurisdiction. In the lingo of economics, election administration involves a big fixed cost and a low variable cost. Any financial solutions will have to take this into account.

In the end, it may make more sense for a gradual upgrade, over say 5 to 7 years. The questions are where to begin and how to finance the transition in a way that is equitable to all counties.

A second study we have undertaken concerns the accuracy of existing equipment. This study has been pretty widely circulated. I have brought copies of the current version and it is posted on our website (www.vote.caltech.edu).

We undertook this study to establish some benchmarks.
How many votes are unmarked, spoiled, or uncounted, and thus problematic in the event of a recount?

• Does the incidence of such ballots depend on the equipment used?

If the incidence of such ballots ( which we call residual votes) is unrelated to machines then it may be unlikely that design improvements could help.

We collected data on election results and equipment used in each of the counties in the US for the 1988 to 2000 election. We began with EDS's data and proceeded to fill in data for states not covered in that data base. We also augmented that data with data from 2000, and we carefully checked the data for errors.

There are some odd observations in the data (very few) that we were unable to resolve. We have omitted these cases.

The metric we use is the percent of total ballots cast for which no presidential vote was registered. This does not capture all errors, but it does capture those votes that would be problematic in the event of a recount or an audit of the election.

The average county in the United States has a residual vote rate of approximately 2.5 percent. The percent of all ballots cast that had no presidential vote recorded equals 2.1 percent. The figures differ because larger counties have lower residual vote rates. We suspect that this is because they have more resources to administer elections.

We then performed several statistical analyses to assess the extent to which the residual vote rate depends on what voting equipment in use. It does. And some of the results surprised us.

First, we looked at simple averages. For each type of voting equipment, what is the average residual vote rate.

Counties using punch cards, either Votomatic or DataVote, had the highest residual vote rates - 3 percent of total ballots cast.

Counties using direct recording electronic equipment also averaged residual vote rates above the national mean.

The average residual vote rate among counties using optical scanning, lever machines, or hand counted paper ballots were below the national mean.

Results that we will include in subsequent versions of the report look at Senate offices. Here again optical scanning and hand counted paper are well below the national average. Punch cards are again above the mean. And DREs are double the residual vote rate of counties with scanning or paper. Lever machines also have higher than average residual vote rates for US Senate races.

Many factors may affect the residual vote rate. So we tried to control for these other factors statistically. Doing so did not change the results.

Holding constant county-level factors, such as racial composition, literacy rates, income, and age, we find the same pattern. Hand counted paper ballots, optically scanned ballots are significantly better than Direct Recording Electronic equipment and punch cards.

This is not to say that those other factors do not matter. County income, racial composition, percent over 65, participation rates, and other factors all strongly affect the incidence of residual votes. Rather, holding those factors constant we still find the same pattern of effects.

Still concerned that we had missed something we performed a third sort of analysis, tracking changes in the residual vote rate as counties changed their equipment.

We took lever machines as a baseline because that was the modal machine in the 1980s. Counties that had lever machines in 1988 but switched to optical scanning saw their residual vote rates drop by one-half of one percent (about 25 percent drop). Counties that had lever machines in 1988 but switched to DREs saw their residual vote rates increase by slightly more than one-half of one percent. This implies a difference between scanners and DREs of approximately one percent of total ballots cast, and it is consistent with our other estimates.

Our immediate reactions to these results were two-fold.

First, there is a good case to be made against punch cards. They are an established technology, and they are, on average, performing poorly.

Second, these results sparked a heated debate within our group between adherents of paper (optically scanned or hand counted) and adherents of electronics. We have subsequently study much of the equipment on the market, and we feel that design improvements for DREs are possible. This is the challenge facing our engineers. The results clearly set paper - hand counted or optically scanned - as the benchmark, the thing to beat.

Doing so will require less attention to designs that speed up the count and more attention to designs that are easy-to-use, that start with the many different types of voters in mind.

Tim Storey or Jennie Drage (303) 364-7700
or Susan Frederick (202) 624-3566