Phil -- No I am not saying that. I believe that Ann is probably getting positive cells above background. What I am saying is that if Ann runs another sample she might get 10 cells or 2 cells, etc. At this level of sampling, the sampling is combinatorial - which is the point we tried making in our paper on sampling statistics (and the degree of assurance you can have that the frequency you measured is likely to be accurate at a certain statistical confidence level). If you refer to Hai Qi's most recent posting to the Cytometry Mailing List you can see his estimates of frequency bouncing around. If you have enough data you can prove whether your sampling is truly within the limits of combinatorial statistics. If not then something might be changing the results, e.g. sedimentation, etc. However, what I am telling you, based on considerable experience over the past 20 years that I have performed rare-event analysis, is that you must have enough data if you want to say with any confidence what the rare cell frequency is. That is precisely why I got into development of high-speed systems. I once was approached at an ISAC conference and asked if I believed some claims on a poster. The authors claimed sensitivity of detection at one cell in a million. Their sample size was one positive cell in one sample and none in that region in the control sample. Wishing to be polite I just smiled and walked away. But underneath I was angry that I had spent at the time over 12 years of my life to address this problem by developing a large number of rare-event methods and technology, and that someone could insult everyone's intelligence at that level. Since people kept asking me how much is enough to sample, we attempted to give some answers in that Cytometry paper (Cytometry 27: 233-238, 1997). I'm not claiming that it is the definitive paper on the subject. I did hope that it would help generate a useful scientific discussion - and it has. So I am happy. I have spent a sizable fraction of my career over the past 20 years trying to open the door for others to study some pretty interesting rare cell problems that I believe will become increasingly important in the years ahead. I hope people find this work useful - and keep asking me questions! -- Jim Leary -----Original Message----- From: Barren, Phil [SMTP:BarrenP@MedImmune.com] Sent: Monday, December 06, 1999 8:46 AM To: 'Leary, James F.' Subject: RE: rare event DOes this mean that you don't believe the data on the 5 cells?/ Pb -----Original Message----- From: Leary, James F. [mailto:jleary@utmb.edu] Sent: Friday, December 03, 1999 2:34 PM To: Cytometry Mailing List Subject: RE: rare event Ann - -I am a little puzzled. The statistics of a 0.05% population in 10,000 total cells is only 5 cells. Even with the best possible background, the confidence limits of that frequency are statistically pretty poor. In the paper I gave to Hai Qi (see corrected reference on Flow Cytometry Mailing List), we explored this issue and showed that accurate frequency prediction is relatively poor below about 20 cells and only really gets stable at about 50 cells. If you see something that I don't, I would really want to know! --Jim Leary -----Original Message----- From: Ann Atzberger [mailto:Ann.Atzberger@EMBL-Heidelberg.de] Sent: Friday, December 03, 1999 7:32 AM To: cyto-inbox Subject: Re: rare event Hallo Hai Qi, if I understand correctly, you ran a total of 200 thousand cells of the same sample through the machine twice. As you say the subpopulation is not that rare at 0.15%. I have often analysed samples with rare events at 0.05% of the total population with good precision,(granted the signal was way above background) and only saving 10,000 events. If you have a lot of debris or dead cells in the sample the %s can vary; so you can either: 1.use a ficoll gradient to clean up the sample. 2.reset your threshold value to cut out as much debris as possible. 3.live gate. 4.Run the cells slowly through the machine. 5.Make sure the machine is clean, as bits of junk stuck in the sample probe can effect your values. I might be wrong but I think saving 200,000 cells has to be more than ample to detect this subpopulation. Ann At 16:49 01.12.99 -0600, you wrote: > > >It is really not too rare, a cell population about 0.04% to 0.15%, while the >background is 0.01-0.05%. The problem is the variation among different >measurements, presumably due to the systematic error. For example, I had >10e7 total cells and ran 2*10e5 out of them for two times: one time I got >0.04% and the other time I got 0.15%. In order to have smaller variation, >obviously I have to increase my sample size, but the last thing I want to do >is to run through all 10e7 cells. Can anyone give me a handy statistical >method for estimating the sample size that makes a cost-effective >compromise? Thank you very much. > >Hai Qi >Dept. of Pathology, UTMB > >
This archive was generated by hypermail 2b29 : Wed Apr 03 2002 - 11:54:18 EST