Actions Possible in a Relatively Short Time Frame
Presentations by representatives of various states to the committee since the November 2008 election indicated that the states’ databases generally performed well in both the primary and the general elections of that year. Nevertheless, the committee believes that a number of meaningful nontechnical actions to further improve voter registration databases could be implemented in a relatively short time frame. (All but one of these actions were discussed in the committee’s interim report of April 2008;1 for the final report, the committee chose to add one additional short-term action, presented below as Recommendation S-10 (on sharing information regarding best practices). None of the short-term actions contained in the interim report were deleted, though Recommendation S-9 (on encouraging third-party groups to turn in forms promptly) was slightly modified in the interest of clarity.)
These actions focus on two areas: (1) education and dissemination of information and (2) administrative processes and procedures. Because these actions remain relevant to future elections, they are repeated below in full (with certain editorial changes to clarify the original intent of the committee).
These short-term actions are directed primarily at election officials at the state and local/county level, and the legislatures and county commissions that make policy regarding the conduct of elections at the state and local level. In some cases, the Election Assistance Commission has a useful role to play as well in facilitating and promoting their implementation.
PUBLIC EDUCATION AND DISSEMINATION OF INFORMATION
Recommendation S-1: Raise public awareness about the legibility and the completeness of voter registration card information.
Accurate and complete data are a basic element of a high-quality VRD. But as noted in Appendix C, the quality of the data in a VRD is no better than the data that are entered into the system. For example, illegible information impairs the ability of election officials to check registrations as required by HAVA
and/or state law, possibly placing additional downstream burdens on the voter (such as having to verify information by mail or having to provide an ID when voting the first time).
Efforts to raise public awareness about the importance of legibility and fully completing voter registration forms would help to reduce the amount of illegible or missing information on these forms when they are submitted for data entry. Properly undertaken, these efforts to raise public awareness of this particular issue could be integrated with ongoing efforts to encourage people to register to vote. Jurisdictions could take some or all of the following specific steps:
Emphasize in the instructions for filling out voter registration forms the importance of legibility and completeness (for example, “Please print all responses; if your answers are illegible, your application may be mis-entered, rejected, or returned to you.”).2
Conduct media campaigns (perhaps undertaken by the Ad Council) emphasizing the importance of legibility and completeness in the information provided on voter registration forms.
Coordinate with third-party voter registration groups and public service agencies, emphasizing the need for their field volunteers to attend to legibility and completeness as they distribute and/or collect registration materials.
ADMINISTRATIVE PROCESSES AND PROCEDURES
A variety of recommended administrative processes and procedures will also help to ensure higher-quality matching and increase voter confidence in VRDs. Note, however, that large volumes of registration forms usually need to be processed as registration deadlines approach, a workload that jurisdictions commonly rely on temporary staff to handle. Unless other arrangements are made to adjust workflow (such as ensuring that actions that require human judgment are routed to permanent staff), these temporary staff will, in many cases, have to carry out these recommended processes and procedures, suggesting that training them to do so will be necessary.
Recommendation S-2: Resubmit alternate match queries if the response returned from the Social Security Administration or department of motor vehicles is a nonmatch.
An election official can use any additional information available to generate match variations for a given name. For example, a match might be sought on standard name variations (for example, Bill versus William), or transposed fields (for example, last name and first name), or compound names separated, or on a maiden name if available.
In practice, humans tend to submit only a few variations. Moreover, manual name rooting is likely to be inconsistent across different officials or even across the same official in different states of fatigue. For this reason, the committee has made a recommendation regarding the use of automated name rooting (Recommendation L-8), although implementing such functionality is difficult to do in a short time frame.
Finally, it may be possible to resolve a nonmatch result by direct contact with the voter, either by phone, in writing, or via e-mail.
Recommendation S-3: Provide human review of all computer-indicated removal decisions.
Because inaccuracies in data may lead to false matching by automated processes, the committee urges jurisdictions to provide a human review of each and every decision to remove a registered voter from a VRD. One step in human review is for a trained election official to examine every computer-indi-
cated decision to see if it makes sense, subject to the availability of such personnel. Such an examination may well reduce the need for the second step, described below.
A second step in human review is advance notification of cancellation—contacting the voter prior to implementing a cancellation. Such contact could be effected through the use of computer-generated letter in the mail. (Such a procedure would not conflict with the requirements of the NVRA, which call for election officials to send to the voter a “confirmation” card that requests verification of his or her voter’s address. After it is returned undeliverable, the voter stays on the rolls for two more federal elections before the registration can be cancelled. State law may or may not require a final notice of cancellation after all NVRA requirements have been met.)
For example, letters could be sent to individuals who are at risk for being removed from the voter registration list; these letters would have a “respond by date X or be deleted” notice. If a notice comes back as “undeliverable as addressed,” the name of the individual would be deleted after date X. If duplicate records are of concern (that is, if two records appear for the same individual), the incorrect record can be deleted. To determine which record is correct, election officials could check available data sources (for example, tax records, real estate records, online search engines optimized for finding people such as www.zabasearch.com, and the telephone book) and/or contact the voter.
Recommendation S-4: Improve the transparency of procedures for adding voters and for list maintenance.
There is not enough transparency in the procedures used to add and remove voters from VRDs. To improve transparency, the states and local jurisdictions should:
Make specific written procedures for the verification of new voters and the handling of removals publicly available. These procedures should address explicitly the specific field-level and record-level matching criteria used for each of these processes. These procedures both inform the public of what election officials intend to do and provide a standard by which audits, oversight, and accountability can be measured.
As an example of the need for making even small details of match algorithms public, Alvarez et al. found that using a match rule involving full first and last names, date of birth, and middle initial resulted in 2.4 times more matches than using a full middle name.3
Extrapolating from this example (not at all atypical), it is clear that “small technical details” in algorithms for matching records can have large policy consequences, such as systematic disenfranchisement of qualified voters (if such matching caused voters to be removed incorrectly)—a point that underscores the importance of transparency and openness regarding these algorithms.
Publish these procedures widely, for example, on the election office Website.
Collect relevant data, such as data on the outcomes of initial applications for registration:4 How many applications were received? Of these, how many were approved and how many rejected? Of those rejected, what were the reasons for rejection—illegibility, incompleteness, person ineligible (cite reason for ineligibility), and so on. Data on how the state handles removals from the registry are also relevant: How many removals were made? Of these, how many were due to intrastate movement, death, and so on?
R. Michael Alvarez, Jeff Jonas, William E. Winkler, and Rebecca N. Wright, “Interstate Voter Registration Database Matching: The Oregon-Washington 2008 Pilot Project,” available at www.usenix.org/events/evtwote09/tech/full_papers/alvarez.pdf.
Many jurisdictions already collect such data, and aggregations of some of these data are published in the EAC Election Day Survey. For more information on the Election Day Survey, see http://www.eac.gov/schedule/2008-election-day-survey/.
Examples of Auditing Applied to Voter Registration Database Processes
Auditing Removals from Voter Registration Rolls
Election officials need the date of receipt of registration applications, the date on which a registration-related notice was sent to the voter, the date, if any, of any response from the voter, and the date on which the corrected or completed information was received; indexes of all of these dates must be kept if correspondence and documentation are to be located. In one state, a removal letter is kept with the original application, and these are sorted by year of first receipt and then alphabetically by name. In this state’s experience, the individuals claiming that they had registered but not been found on the voter registration list had often been sent a copy of the letter, as could be demonstrated by referring to the voter’s file.
It is also possible to keep copies of voter registrations that are cancelled and removed from a voter registration database (VRD). Keeping such records (possibly in a different file) together with the reason for the cancellation would provide data that might be used to improve matching algorithms and/or cancellation procedures.
Auditing Changes in Voter Registrations Records (1)
The main text of this report suggests that voter access to a VRD should be implemented through buffered access to a synchronized copy of the VRD, not to the VRD itself. One kind of audit procedure checks expected behavior against actual behavior. For example, an audit procedure could keep a log of which records were changed in the primary source (the VRD) since the last synchronization between the VRD and the copy. This log could be used to identify the records in the copy that are supposed to be changed—changes in the copy that do not match this list would indicate a problem that election officials could and should investigate further.
Auditing Changes in Voter Registrations Records (2)
As a general rule, corrections or updates to an individual voter registration record should be recorded in such a way that full reconstruction of previous versions of the record is possible as well as an accounting of who made each change. Without the complete revision history available, it is impossible to determine the state of a record at any given point in time and at the same time determine who made what change when. Thus, it may be necessary either to retain the old data in their original form (changes being appended rather than overwriting old data) or to have reliable backup and restore mechanisms so that the old data can be easily retrieved.
Publish these data widely. For example, the Data.Gov Web site would appear to be an appropriate venue for publication.5 Publication of these data would enable public scrutiny of the aggregate outcomes of list maintenance operations and thus prevent politically motivated purges from being performed in secret.
Audit the processes to ensure that procedures are being followed (see Box 5.1 for examples).
Note that collecting and publishing the data suggested above can provide a basis for assessing how big a problem illegibility actually is, how many persons apply who are actually ineligible (for various
See http://www.data.gov/. The Obama Administration is strongly calling for government to publish their transparency data in this single location making it easier for consumers of this data to locate it.
reasons), and so on. The more of such data there are, the easier it will be for election officials to identify problems and to improve list maintenance procedures.
Recommendation S-5: Use printable fill-in online registration forms.
Typewritten or printed information is almost always more legible than handwritten information. Assuming they already have Web sites from which voters may obtain voter registration forms and other election-related materials, jurisdictions could encourage the use of fill-in online registration forms, such as fill-in PDF or Web forms that accept keyboard input (that can be printed, input and all); a number of states provide this service today. Although the form must still be printed, signed, and then mailed or delivered to the election officials, the information on the form will be much more legible. (Note that although the deployment of a new encoding of an old form—such as the National Mail Voter Registration Form—should be possible in a relatively short time frame (the EAC is a logical focal point for any such effort), it should not be regarded as a trivial effort that can be accomplished without some care and testing.)
Recommendation S-6: Perform empirical testing on the adequacy of processes for adding to and maintaining lists.
The only way to know how well a system is working is to test it. One way to test the adequacy of VRD adding and maintenance processes is to corrupt a copy of the most recent VRD by seeding it with artificial records with names and other identifying information from lists of felons, deaths, and mentally incompetent people and with duplicate records of individuals already in the database but with realistic types of error in them. Once corrupted in this way, the VRD can be matched against all of the usual databases (DMV, felons, and so on) to see what fractions of the corruption in each category were detected, thus providing estimates of rates of false positives and false negatives. Because “ground truth” is known in the form of the original seedings, the fractions of detected corruption are likely to be reasonable estimates of the effectiveness of the process overall.6
A corollary of such testing is that those who receive the data resulting from such testing (ultimately, the public at large) must be educated to interpret the data in context—and specifically to understand that no procedure for adding or removing voters can be perfect. At the same time, there is nothing to suggest that individual voters who are wrongly eliminated from the VRD cannot complain or seek correction of the problem through existing channels that are available for resolving such problems.
Another possible approach to testing is to audit actual acceptance, rejection, and removal decisions, not just to verify that procedures have been followed but also to estimate error rates. For example, it is often helpful to test the algorithm and matching procedures involved using a sampling approach in which some practical number of computer-indicated removals are randomly selected from the complete list. Election officials would then perform human review on those randomly selected names, and if the number of improper matches exceeded a certain threshold (which could be zero), the list-generation algorithm and parameters should be further investigated for possible flaws.
The committee emphasizes that sampling is a method for testing algorithms and procedures, not for performing list maintenance, and despite recommending the adoption of sampling for such testing, it reiterates the importance of human review of all indicated removals.
Recommendation S-7: Take steps to find and minimize errors during data entry.
A number of steps can be taken to minimize the effect of data entry errors.
Sample audits can be undertaken to assess the degree of the problem and to identify the source— some data entry personnel, for example, may be much less accurate than others. Some systems produce daily data entry reports that can be compared against the original card for errors; such systems are used in a number of jurisdictions.
The registrant can be provided with a copy of the data that were actually entered (for example, when a voter receives his or her registration card, which should in most cases reflect all of the data entered on behalf of the voter), reminded to check the data, and given information on how to contact the election jurisdiction if there are errors on the card. (To maintain security of personal information, the card should be mailed in a sealed envelope.)
During the input process, the entered values can be tested against domains (for example, common names,7 valid addresses including street name and postal code, valid phone numbers, valid dates of birth).
Data can be entered twice by different people and compared for discrepancies (an expensive way to check, but effective in most instances).
Discrepancies can be found when matching new inputs to previously known values (an ideal way to detect transposition keying errors in dates of birth, for example).
When errors or inconsistencies in the entered data are found, they should be immediately corrected. In some cases, an examination of the records themselves will indicate how corrections should be made; in other cases, it may be necessary to consult additional data sources or even the voter to make the necessary corrections. For example, election officials might provide a special telephone number for voters to call to make corrections.
The first two of these steps (sample audits and the voter being provided a copy) can be taken in a relatively short time frame. The other three require a nontrivial amount of new technology deployment.
Lastly, voter registration forms should ask the applicant to provide (voluntarily) e-mail addresses and/or SMS (text message)-enabled cell numbers in order to facilitate contact with the applicant—such contact may well be necessary in order to clarify other information that is unclear on the registration form.
Recommendation S-8: Allow selected individuals to suppress address information on public disclosures of voter registration status.
Although voter registration information is nominally public in most states, certain individuals (e.g., domestic violence victims, undercover police officers, witness protection program participants, and so on) have legitimate and understandable reasons for wanting to make address information inaccessible to the public, and an administrative process should be available to protect such information on request. Indeed, some states (such as New Mexico, California, Oregon, Missouri, and Kansas) already provide for such suppression under certain circumstances.
Defining who can and cannot suppress address information involves a balancing of privacy and other interests. Advocates of more open records might argue that the public interest is best served by a relatively narrow scope of individuals who should be granted such privileges, whereas some privacy advocates might argue for the broadest possible scope. The committee is silent on this particular point.
The committee notes that enacting this recommendation may require legislation in many jurisdictions.
Recommendation S-9: Encourage entities sponsoring voter registration drives to submit voter registration forms in a timely manner to reduce massive influxes at the registration deadline.
Election offices can be overwhelmed by the mechanics of data entry if large numbers of voter registration applications must be processed in a very short time. Such a volume reduces the time for error checking or multiple attempts to verify voter information, and often forces election officials to hire inexperienced temporary workers for data entry. These conditions in turn are likely to increase the error rate of data entry and may invalidate more registration applications that would be the case if more time were available to handle the applications.
In addition, forms that are not processed (not entered into the relevant VRD) in a timely manner can cause confusion for the applicant. Election officials often have to deal with a large number of inquiries from applicants who filled out an application and gave it to someone weeks or months before. Seeking information about whether their applications have been received and processed, these applicants are often told that their applications are not in the system, and are encouraged to submit another application because the election officials do not know if the party collecting applications will actually turn in the form. This chain of events is frustrating for the applicant, and it causes more work and confusion (more duplicates) if and when the original form is eventually submitted.
Some states (e.g., Oregon) require third-party groups to submit applications to election officials within a certain number of days of collecting them. Nevertheless, imposing such requirements more broadly may entail politically controversial state legislation.
Recommendation S-10: Improve information sharing regarding best practices and lessons learned regarding VRD acquisition, operation, and maintenance.
Election officials in various states operate a variety of voter registration databases and face many different problems in acquiring, operating, and maintaining their databases. Election officials, and the technologists who support them, would be likely to find value in reports of best practices and lessons learned in the ongoing database enterprise. The content of such venues would logically include both process and substance knowledge. In the former category might be lists of state database administrators and their contact information, provided so that others might call to inquire about how various things are done within their jurisdictions. In the latter category might be data related to database performance or published procedures for list maintenance.
The committee thus recommends the establishment of continuing venues that could act as repositories of such wisdom. Such venues should be regular and continuing, either through face-to-face meetings and/or a Web site. In the committee’s view, the most logical focal point of action would be the National Association of State Election Directors.