Contents My potential involvement Swissair identification paradigm The false-positive problem Work begins WTC Kinship Planning Panel Relatives Identified American Airlines Flight 587 crashes vs attacks WTC status, May 2002 WTC status, June 2003 footnote

Links WTC DNA identification – prospectus Analysis of screening; Powerpoint presentation Tsunami victim identification considerations Forensic mathematics home page

World Trade Center Disaster Identification Diary

Hals' Willem Coymans

Back at the hotel I had switched on the TV in time to see some of the live action and numerous replays of now-familiar events, accompanied by surprisingly little voice-over aside from a short period during which a smug BBC commentator analyzed the context and future significance of the goings-on until mercifully he was given the hook.

My potential involvement

Wednesday morning we were slightly affected by the attack in the sense that we waded through security confusion when we flew out of Schiphol (Amsterdam) for London. Moreover, in considering the talk that I was scheduled to present in London, I hit on the idea to discuss the prospects for WTC identifications.^* But not until I read my email message from Dr. Howard Baum of the Forensic Biology (i.e. DNA) section of the Office of the Chief Medical Examiner of New York did the possibility of my personal involvement with the disaster occur to me. The OCME has been using the Kinship module of DNA•VIEW for several years, and it is already well established as a tool for disaster body identification. I shouldn't have needed Howard to remind me. His message was brief. He wrote, "We need help coping with the mass disaster in New York City."

Howard's immediate thought was the application of the Kinship program, but I thought I had even more to offer. There was going to be a lot of genetic data to manipulate, whose exact nature couldn't be predicted in advance. As one who has worked with computers since 1959, earned a doctorate in mathematics, and done dozens of practical or research projects involving DNA-relationship ideas and computations, I figure I am uniquely prepared to perform whatever manipulations and analysis might be necessary to wring information from the data.

Swissair identification paradigm From earlier experience in disaster identification with the Swissair 111 crash, I assumed that there would be a necessary "screening" step in making the WTC identifications based on relatives; further, I could extrapolate that due to the larger scale of the WTC problem new complexities would need to be faced (links WTC prospectus and WTC Powerpoint above). Some of the victims of the Swissair crash needed to be identified indirectly, by comparison with living (or dead) relatives. A two step paradigm emerged: Screening – a rough but quick scan (<<1 second/comparison) to select likely candidate matches between victim and reference family samples by comparing each individual reference with each victim sample (millions of screening comparisons in the WTC case); Testing – a candidate family-to-victim correspondence is evaluated by using the Kinship program to make the accurate computation (minutes/comparison) which considers the family as a group (hopefully only a few thousand WTC relationships will require testing and confirmation). Benoit LeClair, then of the RCMP DNA group, had shown how the screening could be done in that case with a mere Excel application, based on sorting the DNA profiles such that similar profiles would be sorted into neighboring positions. I considered his idea as a starting point. But for the much larger dataset that the WTC disaster might produce, I soon became convinced that more firepower and algorithmic sophistication would be necessary.

The false-positive problem

The main difficulty that I foresaw emerging as the sizes of the two lists – the victim list and the family reference list – grow, is the increasing incidence of "false positives." If both lists are small and some person C in the reference list has a brother who died, and some profile V in the victim list looks like a brother of C, it probably is. However, if the victim list has thousands of profiles, then for any given reference person C there will be dozens of victim profiles that coincidentally resemble C just as much as a typical true brother does. The proportion of false positives is proportional to the size of the victim list.

Therefore I was sure that a simple sorting program of the sort that had been adequate for the Swissair identifications, would not be very useful for the WTC disaster. Specifically, in my London talk on September 13, I hypothesized –

1. 20,000 dead
2. 5,000 recovered
3. 4,000 positively identified via e.g. toothbrush DNA (LR=10¹² per ID) dental or other ID
4. 500 with 2 close relatives (parent or child)
5. LR>10⁸ per ID. (LR=likelihood ratio)
6. 300 with 1 close relative
7. LR=1000 per ID
8. 100 with combination of sisters and uncles and cousins and aunts
9. LR inadequate unless all data taken into account

Of course the first several of my estimates have proven to be quite far off. However, #4-8 are in the ballpark. The implication of the estimate #7 is, that for every 1000 victims, there will be about one who coincidentally resembles any given reference person to the same extent as does a true child. Thus, using individual parents as references to fish victims out of the rubble would result in more false leads than true ones. On the other hand, #5 implies that if a more sophisticated trolling operation is used, wherein two reference parents are simulataneously compared with each victim to accomplish a sort of triangulation, then the number of false hits will be very small. Point #9 is correct.

Work begins

The upshot was, on October 2, the first of several trips to NY. Dr. Shaler organized a one-day meeting, a "summit of genetics experts" (Wall Street Journal) to discuss various problems and possible approaches for sorting through the inevitable masses of data. Five laboratories – the city Office of the Chief Medical Examiner, NY State Police, Bode Technologies, Myriad, and Celera – who were expected to do parts of the DNA analysis were represented. Coincidentally Myriad now included my old colleague Benoit formerly of the RCMP. The FBI and myself were also present to discuss software, as well as Howard Cash and others from a company called GeneCodes that is contracted with the OCME to provide software. Finally, there were a few people from the NIJ (National Institutes of Justice, which is another arm of the DOJ). Following introductory explanations by Bob Shaler in the morning several of us presented our ideas about making the necessary victim-to-reference identifications. The afternoon was mostly discussions, and of course mostly rather general and oriented toward planning. From time to time, though, people inevitably succumb to the temptation to discuss details even when it is obvious that there is neither time nor yet sufficient information to make a detailed discussion productive. When this happened, Chief Inspector Dale of the State Police patiently suggested, each time as if it were the first, that it would be appropriate to make general plans. I enjoyed that.

At some juncture, concerned that the plans might be steering toward an unnecessary and ponderous software project, I made a comment to the same effect as I have indicated above, that once I am able to get my hands on the data, I will quite quickly be able to produce the tentative identifications by myself. At this Howard Cash piped in, "Surely, Charles, even your work can stand a second opinion." I told him he had a fair point.

WTC Kinship and Data Analysis Planning Panel (KADAP)

The three-day meeting ranged over a variety of topics. The one topic originally mentioned to me was the same that Bob Shaler had already asked of me: choose which screening program to use. To that end I put together a Powerpoint presentation to explain the difficulties and pitfalls as I foresaw – and, by now, had computed.

In assessing the candidate screening programs, I had in mind several design requirements:

• "batch" operation
  Most of the available programs stored a list of victim profiles, and expected the user to perform, for each family to search against the victim profile list, some manual operatons such as typing in the family DNA profiles. That would be very awkward.

  • It implies spending at least several moments per family just for screening, times a thousand or so families
  • and again each time there is an updated victim list.
  • It doesn't cater to "cherry picking" – concentrating first on those families for which clear victim candidates pop up, rather than wasting time on difficult or impossible cases.
• likelihood ratio evaluation
  The quick and dirty way to estimate if two profiles come from related people is to count common alleles. This may be too dirty.
  My simulations convinced me that more sophisticated (LR) computations, taking into account for example that sharing rare alleles is more significant than sharing common ones, would greatly improve the accuracy of the screening.

  triangulation

  look for potential victim-family associations that are indicated by at least two members of the same family The already available programs scored poorly by the above criteria. However, my preferred idea, to produce something from scratch, might reasonably be considered deficient with respect to such standards as:

  • proven support; multiple support personel
  • validated performance
  • track record of reliability

  Nonetheless, I naturally decided to go ahead with my own screening program.

  Relatives identified

  In any case, there was certainly no harm in trying. A lot of the most difficult work for my intended screening module was embodied in programs that I had already written in order run simulations to decide what kind of screening program was likely to be effective, and what kind of data likely to be sufficient. The first version of the program took only a few days to write. Writing programs in a sterile environment is easy. The acid test of a program is whether it works with real data and – almost as important – real users. I was anxious to get my hands on the lists of DNA profiles, which by now were flowing out of Bode and Myriad labs and accumulating at the NY State Police.

  November 8, 2001 – I went to the OCME lab in NYC for a scheduled two-day visit to deliver Mark I of my new software, collect some data, and try it out. Predictably, the first look at the data showed some small surprises: the format of the data was not exactly what I expected. I decided to stay around a few extra days to sort everything out. Thus, I was in New York for the astonishing incident of the following Monday.

  November 10, 2001 – The first effective run of the screening program poured out a sorted list of victim-family potential associations in about 5 minutes. The families on the top of the list figured to be genuine identifications; families with no strong resemblance to any victim sample would sort further down. Indeed, the first victim-family pair on the list had an extremely high score. Checking the actual profiles showed that this case concerned identical twins, as I expected since I'd been previously told that among the DNA identifications already made through "direct" references was a person who was the identical twin of a survivor.

  Next on the list was another identical twin case. This one – or at any rate at least one of the two – was a new identification, the first I had found. The third candidate identifications on the list was a case where the mother, the daughter, and a brother had presented themselves as references. The screening report only told me that two of these people bore a resemblance to the same victim. To confirm the identity of that victim, I needed to make a family-specific computation with the Kinship program to check that the entire assemblage is genetically consistent and numerically convincing.

  It was. According to the kinship computation, it was either the right victim, or it was a one-in-twenty-billion coincidence. That's good enough to call it a confirmed identification.

  And so it went, easily, for the first thirty or so cases, that Saturday morning. Of the next twenty, some looked ok but the data was often insufficient to be confident according to the criteria we had established. A few looked to be spurious. After that, the list started to run dry. The data simply didn't exist to say who went with whom. More data was necessary, both as to obtaining more reference family members, and recovering further victims. Nor was I in a position to know which of the identities I had found might already be known from direct DNA references or from some other method.

  American Airlines Flight 587

  9am Monday, November 12, 2001 – An American Airlines flight crashed in Queens, almost within sight of our hotel window. By the time I showed up at the OCME offices with my list of relationship identifications, the WTC problem was temporarily demoted in priority; plans were already being made to identify the new tranche of victims. Obviously, this would be another test of my new software.

  We discussed the codings schemes that would be used for the airplane victims, and I considered what minor modifications would be needed. Plane crashes, unlike office environments, tend to include related people. It's important to know about them and to consider them in the analysis. Shaler and his colleagues were already thinking about how to improve on the WTC experience in collecting samples from relatives. It may seem a ghoulish observation, but an experienced disaster identification team was swinging into action.

  We left New York on Wednesday, figuring that I had a bit of work to do.

  Airplane crashes versus WTC attacks

  Thursday, November 22, 2001 – It took only a few days to customize the program to accept the airline crash data, and it also took the OCME only a week to begin producing DNA typing data from the crash.

  I believe that every disaster is unique. Contrary to the hope of a few of the KADAP group, I don't believe it is practical or realistic to expect a "disaster identification" program to be a result of the current identification effort, or efforts. Useful tools, yes. Worthwhile experience, also.

  1. For example, a large part of the difficulty of the WTC identifications is gathering, collating, checking, and correcting data from a large variety of sources. There is so much DNA typing to be done that it was of necessity farmed out to several vendors, with inevitable consequent complications of disparate sample coding schemes, and complexities in receiving the data and in following up with questions about it. By contrast, the airplane crash DNA analyses would all be done at the OCME lab.
  2. Also, I have already mentioned the unique screening complexity of the WTC disaster because of the sheer magnitude of combinations, compared to those in a smaller disaster such as an airplane crash.
  3. On the other hand, an airplane crash brings complexities of its own: Typically there are numerous families who flew and perished together. The fact that several different remains are related to one another, and to the same sets of living references, is both a difficulty and a hint that must be utilitized in establishing identities (Kinship analysis when there are many possibilities). The Trade Center victims include a few relatives, and if we forgot to consider the possibility we would surely announce some incorrect identifications, but they are the exception.
  4. In an airplane crash there is a reasonable chance to identify every victim, and this "closed system" quality of the problem, like a completed jig-saw puzzle, is helpful.
  5. A war zone like Bosnia has other characteristics. For example, there are many mass graves, but within each grave it is to be expected that the victims originated from nearby. Therefore geographical and other non-DNA or, as we have come to call it, "meta-data", plays an important role.

  In the case of AA587, the screening program sorted together the families and victims rather easily as the data flowed in. There were a few family groups, as point #3 above, that were tricky to unravel. In order to help myself sort them out I added a new feature to the Kinship program so that it instantly shows a helpful diagram of all likely pair-wise relationship among a group of people.

  Most of the AA587 victims were identified within a few weeks, which is sensational. Announcement of the final identification took about three months, as inevitably a few cases were delayed by special problems.

  WTC Status

  Status May 30, 2002

  Trade Center excavation officially ended today with a ceremony at the site. Since I was in the city and in six previous visits since September 11 I never happened to visit ground zero, I decided to take this last chance to have a look. It was a look a the back of a lot of heads and some fences, and most excitingly of some neighboring buildings that were draped with giant shrouds. Not much to see, but it interested me.

  No new victims will be found. It will take a few weeks for the DNA profiles of the most recently excavated victim pieces to be reported. Once those have been checked against the reference materials – direct or indirect – already in hand, I expect a pause in the identifications. Further progress will depend mainly on success in the new DNA techologies that are being attempted, namely SNP's and mtDNA.

  Status June 24, 2002

  • 1200 victims have been identified.
  • Over 600 DNA identifications have been made; 500 by DNA only.
  • 1/3 of these were by kinship, and 1/4 exclusively by kinship.
  • The remainder were by comparison with direct references
    • Most often helpful are hairbrushes, toothbrushes, hair, and razors.
    • For some victims blood samples were available, e.g. from blood donors.
    • Clothing, towels were rarely productive.

  80% of the victims were men.

  Status June 24, 2003

  The 1000th victim identification was announced on April 29, 2002, seven months after the attack. The 1500th identification came out on June 18, 2003, after 21 months. The rate now seems to be two or three per week. It is to be expected that progress has slowed, but not much can be extrapolated. Identifications from traditional forensic STR loci and primers ground pretty much to a halt several months ago. Special, shorter versions of them – credit to John Butler and Bode Technologies – then revived progress as I understand from Bob Shaler, and similarly assistance from the other new technologies may have either incremental or dramatic benefits.

  Last July Bruce Weir suggested that I submit an article on World Trade identifications to an edition of Theoretical Population Genetics – as the name implies a rather high-brow scientific journal – that he was editing. He thought "it would have high interest," and so it proved. The article, completed with Bruce's considerable help and co-authorship, attracted what is by my standards a lot of interest. I gave an interview to a German radio (in English) and a newspaper, to the children's science magazine Odyssey, and to Nature.com.

  One point I made in the article was an estimate of the maximum number of bodies for which any DNA has been found. I did this probabalistically, trying to account for those victim fragments that produced any DNA at all. If we then make the optimistic assumption that all such fragments can eventually, through ultimately sophisticated DNA typing methods, be identified, as many as 2100 victims might eventually be identified. Perhaps a better way to put it is that no more than 2100 will be found. Dr. Hirsch, the NY Chief Medical Examiner, has expressed his commitment to continue with the work possibly for a long time, with a goal in mind of 2000 identifications. I agree that even that number seems very difficult.

  SNP progress

  Identification through SNP's seems promising. The theory is that SNP's may succeed where STR's fail because the necessary extant DNA fragment size is smaller. Successful STR typing depends on finding intact some particular 120 to 300 (depending on the locus and the allele) base pair stretch of DNA. Of this stretch, perhaps 20 to 120 bases is the variable part that represents information (the information is in fact the length of the variable part) and the remainder, the flanking region, is partly a handle or label that selects the fragment of interest uniquely from the entire genome, and partly a spacer included in order to enable multiplexing – the assay of multiple loci in the same test tube.

  Orchid Biosciences has developed a high-speed, high-throughput, largely automated method for SNP typing, which is being used for WTC samples.

  For SNP's, the region of interest is a single nucleotide, which considerably reduces the fragment size. Moreover, the method by which multiple assays are acheived is entirely different, so there is no need for a spacer in the flanking region. Consequently SNP's may succeed where STR's fail when the DNA has degraded to the point where the typical fragment size is around 100 bp.

  Over 5000 samples, some victims, some from living relatives, have been typed. The attempted panel of typing is 70 loci. Given that many of the samples are degraded, as we already know from the mixed success of STR typing, there is mixed success in SNP typing. Often only a partial profile is obtained. However, for sure there are times that SNP is quite successful even though STR was not. For these cases the SNP technology is quite likely to provide new identifications. Additionally, there will be cases where both technologies are only moderately successful but the combination is just good enough.

  mtDNA progress

  ---

  *In subsequent months of course my (my) idea, to link anything and everything to the Trade Tower attack, has been widely copied. At the time, it seemed quite bright and original.

  Go to the top of this page