Ms. Dopp wrote to me asking about an ROI analysis for voting machines compared with manual counting of paper ballots, and I responded directly to her. I think this should also be seen here on the EM list. From Ms. Dopp's first response to this, I'll clarify that the initial figures given for speed of counting ballots are only for sorting ballots by the candidate voted for in one race, not for the complete process for that race and not for the whole ballot. So please read carefully....
At 10:56 PM 5/22/2007, Kathy Dopp wrote: The whole voting machine thing just boggles my mind. Didn't anyone >>ever do a return-on-investment analysis to compare using these >>machines with hand counting?'' > >Have you done one? Where is it? Have you done all the calculations for it? I wrote what is below, first. But to lead with the conclusion, break-even probably would be below $400 per machine. If it costs more than that, forgeddaboudit. However, this assumed 20 races on the ballot. If the number were different than that, the break-even point would vary linearly. (More races, higher break-even cost. Essentially, the machine would be used more if there are more races.) ********** Well, it seems so blatantly obvious, but ... I'm not equipped to do a formal ROI. It seems to me that I have seen some informal study of this, Steve Unger comes to mind, but I don't know. Here is a *very* rough attempt: How much use does a voting machine see? I think that it might be once per year, in some places once every two years. Let's assume once per year. If machines can be shared among jurisdictions that schedule voting on different days, that would reduce cost but introduce other problems, it only could address special, local elections. How many voters use each machine each election? I don't have figures, so I'm just going to guess wildly. 5 minutes per voter, open for 10 hours. So one machine could serve 120 voters, producing 120 ballots. Let me just stop and say, right now, that as a businessman, I wouldn't even *think* of trying to buy a machine to count the data on 120 ballots, once a year. However, let's proceed. I'm going to assume a certain counting process that should be reliable. I need to do that to estimate the labor involved in counting the ballots. I'm making up details as I go, many of these procedures, I'm sure, could be improved or might need to be altered. The ballots are sorted into stacks (batches) of a certain number. The stack counts can be quickly verified with a scale (error in this will be discovered, but little redundant checks like this, where they are easy, will decrease errors.) I'm going to assume ballot imaging, my own project, so a sticker is put on each ballot with a serial number, the time to do this is negligible. The serial number also identifies the batch number. A worker takes a batch and counts one race by physically sorting the ballots into stacks for that race. Because the worker is looking only at one specific location on the ballot, and need not do anything with the information except move his or her hands quickly in response, this can be very fast. Then the counts of candidates for that batch are made by counting the ballots in each stack. Another worker duplicates the entire process, starting with quickly shuffled ballots. This is more expensive than is really necessary, especially if there is going to be public imaging. But I have the shuffle to avoid a cause of error. If the second worker were to simply look at the counted stacks, the worker could be led into overlooking a mis-sorted ballot. I am also assuming, here, full manual counting, no machine assist, and I am not tacking in recording of individual ballot data, which, done manually, would require additional time, and, since this is not done at all now, it would be unfair to the analysis. But I will note that such counting and identification is quite simple with imaging and proper, easy-to-write software tools for doing it. It could be faster and thus cheaper than what I'm describing (and ballot imaging is essentially no-cost). The batches are thus each counted, for each race, twice, and both under observation. If the worker is picking up a ballot and putting it down on a stack representing the vote in a race, the observers can watch and detecting a misplaced ballot would actually be reasonably likely, particularly with several observers. They are looking at the exact same place on each ballot, it will end up in the same place in a stack. Then the whole process is repeated for another race. Because the batches can be handled independently, all workers can be continuously engaged. (And the double counting for a race need not be sequential; if they are going to be shuffled, the sorting on a different race partly accomplishes it, that's all.) So we can look at how long it takes to count a ballot for one race. I estimate, with the procedure, that it would be under two seconds for the sorting. I've done a lot of manual operations when I was a printer and did not, for example, have a folding machine. For small runs, maybe up to 100 sheets, a single fold in half, I could beat the machines, not in raw speed, but because it takes time to set and adjust the machine.... How many races are on the ballot? Again, this can vary widely. So I am just going to determine the cost of counting one race, initially. This can then be extrapolated to any number of races, it's linear, unless one were to reach fatigue limits, another problem, and one solved by throwing more labor at it, since the labor cost per ballot is constant. Then, to actually count the ballots would be under one second per ballot (one of the things that printers often have to do is count stacks of paper. One second is way more than enough.) 20 ballots per minute and 120 ballots per voting machine. That's 6 minutes to count one race. Double this for the redundant counting, so it is 12 minutes to count one race. There is a little extra time spent recording totals, I'm neglecting this at the moment. Let's assume a labor rate of $12 per hour, which is generous. So the cost per race is $2.40. Let's make that $3 to allow for recording results. This number should be noted, this is the cost per race per voting machine being replaced with manual counting per election. Suppose there are 20 races. So the voting machine can be replaced for $60 worth of labor per election. How inexpensive would the voting machine have to be to justify the investment, to break even? Well, suppose we think of amortizing the cost over 20 years, paying off a loan to purchase it, with, say, 6% interest. I used a mortgage calculator and came up with $400. What's the going rate for one of these little babies? Oh, and I did not consider, in comparison, the other acquisition costs, the research needed to buy the machine, though I'm sure that the voting machine people try to make that, so to speak, a pleasant experience for the government's buyer.... And I did not consider the storage and maintenance costs, which are practically zero for the hand count option. Printing costs are also negligible, on the order of, say, $2 for $120 ballots. I would not consider doing it for $400, because of the extra costs and risks. This is very informal and seat-of-the-pants. It should be independently verified before using it in some sort of official or important way. It was just done to illustrate the point. Ballot imaging with blink software (that allows comparison of ballots by blinking them, the kind of image work that is used to discover asteroids and other celestial phenomena and which is very sensitive, the eye is good at seeing variations in images that it could never detect by just staring at them, if they are blinked back and forth) could make counting faster and more accurate, plus the use of images has a host of benefits, most notably, the ballots themselves only need to be handled once to image them. Everything else is then done with images (and generic, already-available computers, etc.). Handling the ballots once and then putting them under seal would yield maximum security, and imaging allows multiply redundant counting, very quickly. Approval Voting throws a monkey wrench into the counting procedure I mentioned, but it's not difficult to vary the procedure to accomodate approval. Machine recognition would be cheaper than hand counting, if standard equipment is used, not special-purpose voting machines. Any fax can serve as an adequate scanner. The software should be public-source, and it is not difficult software, my guess is that there are already standard programs which could handle the ballot vote recognition. So it could just be a matter of feeding the ballots through a fax machine with document feeder, and faxing the images to a computer somewhere else. So all that is done at the precinct level is collect the ballots, serialize them, feed them into a fax, verify that the fax was received, and then pack them up and seal the packages. Before they are packed, though, they would be open to public inspection under the eye of multiple observers, and private imaging could take place then. I do not see this as an expensive system at all, the cost is certainly less than manual counting. And the public will manually count the ballots anyway. This is the point of serializing the ballots. If all you can do is to check totals, it's impossible to track down and determine the significance of variations. If the ballots are individually identified with a serial number, then anyone could look at the image of a ballot where there was a discrepancy in the reported vote. I'd expect that there would be a spreadsheet for each precinct, standard tab-delimited text, with a link in each record to the image. Trivial to compile.... (more labor than the manual counting described above, but many hands make short work....) (Just to be explicit, the serialization is done after the ballot boxes are opened, so that there is no way to identify the voter from the ballot if the voter has not altered the ballot in a visible way, other than by marking the positions without extraneous writing.) ---- election-methods mailing list - see http://electorama.com/em for list info