WEBVTT 00:00:00.000 --> 00:00:04.428 align:middle line:84% [SQUEAKING] [RUSTLING] [CLICKING] 00:00:04.428 --> 00:00:09.820 align:middle line:90% 00:00:09.820 --> 00:00:11.820 align:middle line:84% SCOTT KEMP: All right, why don't we get started? 00:00:11.820 --> 00:00:17.720 align:middle line:84% So today, we're going to talk about the linear no-threshold 00:00:17.720 --> 00:00:18.720 align:middle line:90% model. 00:00:18.720 --> 00:00:23.120 align:middle line:84% And some of you, probably most of you, have heard about it. 00:00:23.120 --> 00:00:25.720 align:middle line:84% In the last class on Wednesday, we 00:00:25.720 --> 00:00:29.360 align:middle line:84% went through this, the very basics of radiation and dose. 00:00:29.360 --> 00:00:32.400 align:middle line:84% And we came up to the point where 00:00:32.400 --> 00:00:35.900 align:middle line:84% we talked about these units of sieverts or REM, 00:00:35.900 --> 00:00:38.440 align:middle line:84% which are so-called equivalent dose. 00:00:38.440 --> 00:00:41.560 align:middle line:84% They have units of energy per unit mass, 00:00:41.560 --> 00:00:44.320 align:middle line:84% but they don't really mean that because they've 00:00:44.320 --> 00:00:48.960 align:middle line:84% been multiplied by these unitless adjustment factors. 00:00:48.960 --> 00:00:50.880 align:middle line:84% But the problem is that we don't really 00:00:50.880 --> 00:00:53.160 align:middle line:90% know what to do with this unit. 00:00:53.160 --> 00:00:57.440 align:middle line:84% It doesn't tell us anything about how much cancer that 00:00:57.440 --> 00:00:58.800 align:middle line:90% gives us. 00:00:58.800 --> 00:01:01.650 align:middle line:84% It just tells us how much energy is deposited, 00:01:01.650 --> 00:01:05.030 align:middle line:84% but the amount of energy that's actually deposited by radiation 00:01:05.030 --> 00:01:06.810 align:middle line:84% is trivially small compared to, say, 00:01:06.810 --> 00:01:10.430 align:middle line:84% taking a bath in warm water, which deposits a lot more energy 00:01:10.430 --> 00:01:12.150 align:middle line:90% into your body. 00:01:12.150 --> 00:01:14.270 align:middle line:84% And it has to do with, as we mentioned, 00:01:14.270 --> 00:01:17.870 align:middle line:84% how that interaction happens, where the energy is going. 00:01:17.870 --> 00:01:23.310 align:middle line:84% So we need some way of converting this otherwise 00:01:23.310 --> 00:01:27.590 align:middle line:84% useless unit into something that tells us about health effects. 00:01:27.590 --> 00:01:31.030 align:middle line:84% And so that is the so-called dose-response model. 00:01:31.030 --> 00:01:35.430 align:middle line:84% Now there are lots of different kinds of health effects 00:01:35.430 --> 00:01:38.870 align:middle line:84% that we could build dose-response models for. 00:01:38.870 --> 00:01:43.430 align:middle line:84% So we typically break these down into two different categories-- 00:01:43.430 --> 00:01:46.270 align:middle line:84% what are called the deterministic diseases 00:01:46.270 --> 00:01:48.390 align:middle line:90% and the stochastic diseases. 00:01:48.390 --> 00:01:52.150 align:middle line:84% So deterministic diseases have the property 00:01:52.150 --> 00:01:54.710 align:middle line:84% that the intensity of the illness 00:01:54.710 --> 00:01:59.140 align:middle line:84% is directly proportional to the amount of radiation 00:01:59.140 --> 00:02:01.020 align:middle line:90% that you get. 00:02:01.020 --> 00:02:04.900 align:middle line:84% And sometimes there's a threshold effect that 00:02:04.900 --> 00:02:07.260 align:middle line:90% might exist in those diseases. 00:02:07.260 --> 00:02:13.140 align:middle line:84% So the classic examples here are like skin burns, cataracts, 00:02:13.140 --> 00:02:16.900 align:middle line:84% hematological disorders of various kinds. 00:02:16.900 --> 00:02:20.260 align:middle line:84% And it's just basically, if you put enough energy 00:02:20.260 --> 00:02:22.980 align:middle line:84% onto the surface of your skin from radiation, 00:02:22.980 --> 00:02:26.033 align:middle line:84% more than capillary dilation is able to take away. 00:02:26.033 --> 00:02:28.200 align:middle line:84% So what happens is when you put energy on your skin, 00:02:28.200 --> 00:02:29.617 align:middle line:84% your blood vessels instantaneously 00:02:29.617 --> 00:02:32.100 align:middle line:84% expand to carry the extra heat away. 00:02:32.100 --> 00:02:35.220 align:middle line:84% If you can't do that, then you get a burn. 00:02:35.220 --> 00:02:41.040 align:middle line:84% And that's going to be worse if the energy goes up in value-- 00:02:41.040 --> 00:02:42.780 align:middle line:90% very straightforward. 00:02:42.780 --> 00:02:45.860 align:middle line:84% Stochastic diseases are different. 00:02:45.860 --> 00:02:48.020 align:middle line:84% The intensity of the disease is not 00:02:48.020 --> 00:02:51.380 align:middle line:84% proportional to the amount of radiation you get. 00:02:51.380 --> 00:02:53.540 align:middle line:84% The probability of getting the disease 00:02:53.540 --> 00:02:56.200 align:middle line:84% is what is proportional to the amount of radiation you get. 00:02:56.200 --> 00:02:59.370 align:middle line:84% And so there are also lots of different things 00:02:59.370 --> 00:03:00.622 align:middle line:90% in the stochastic category. 00:03:00.622 --> 00:03:02.330 align:middle line:84% But the big one that we always talk about 00:03:02.330 --> 00:03:05.650 align:middle line:90% is cancer, genomic mutations. 00:03:05.650 --> 00:03:07.610 align:middle line:84% And some genomic mutations do things 00:03:07.610 --> 00:03:09.650 align:middle line:90% other than causing cancer. 00:03:09.650 --> 00:03:11.930 align:middle line:84% So for example, you can have mutations 00:03:11.930 --> 00:03:14.130 align:middle line:84% that cause all kinds of classical disease 00:03:14.130 --> 00:03:19.330 align:middle line:84% by either leading to maybe a single nucleotide 00:03:19.330 --> 00:03:24.210 align:middle line:84% variant in some protein that then is miscoded and then 00:03:24.210 --> 00:03:25.910 align:middle line:84% doesn't operate correctly in your body, 00:03:25.910 --> 00:03:29.370 align:middle line:84% and you get some kind of functional disorder. 00:03:29.370 --> 00:03:35.410 align:middle line:84% But the one that kills us if we are otherwise healthy 00:03:35.410 --> 00:03:37.230 align:middle line:84% and then kills us in mid-life is cancer. 00:03:37.230 --> 00:03:39.430 align:middle line:84% And so that's the one that we tend to focus on. 00:03:39.430 --> 00:03:41.770 align:middle line:84% So we're going to just talk about cancer only. 00:03:41.770 --> 00:03:44.010 align:middle line:84% But just to remind you that there are lots of things 00:03:44.010 --> 00:03:46.090 align:middle line:90% that can happen. 00:03:46.090 --> 00:03:46.590 align:middle line:90% All right. 00:03:46.590 --> 00:03:51.890 align:middle line:84% So how do we construct this dose-response model? 00:03:51.890 --> 00:03:55.880 align:middle line:84% So there's some amount of dose, and then there's 00:03:55.880 --> 00:04:00.320 align:middle line:84% this black box, which is a blue box, which 00:04:00.320 --> 00:04:02.540 align:middle line:84% has some mysterious biological process. 00:04:02.540 --> 00:04:05.240 align:middle line:90% And out pops some cancer rate. 00:04:05.240 --> 00:04:08.040 align:middle line:84% And we need to understand what's in that blue box. 00:04:08.040 --> 00:04:11.180 align:middle line:84% So one approach is that we say, you know what? 00:04:11.180 --> 00:04:13.640 align:middle line:84% We don't bother with what's inside. 00:04:13.640 --> 00:04:17.079 align:middle line:84% We'll just build a statistical model 00:04:17.079 --> 00:04:20.640 align:middle line:84% that seems to describe what we observe 00:04:20.640 --> 00:04:22.840 align:middle line:90% happening on the outside. 00:04:22.840 --> 00:04:25.120 align:middle line:84% You put a certain amount of radiation in, 00:04:25.120 --> 00:04:28.280 align:middle line:84% and maybe we get some cancer rate out, some estimate 00:04:28.280 --> 00:04:29.480 align:middle line:90% of the cancer rate. 00:04:29.480 --> 00:04:32.080 align:middle line:84% So we'll talk a lot about this statistical model. 00:04:32.080 --> 00:04:35.560 align:middle line:84% The prevailing statistical model is called 00:04:35.560 --> 00:04:38.160 align:middle line:90% the linear no-threshold model. 00:04:38.160 --> 00:04:41.800 align:middle line:84% And we'll get into that a little bit more. 00:04:41.800 --> 00:04:43.520 align:middle line:84% The problem with statistical models 00:04:43.520 --> 00:04:46.680 align:middle line:84% is a good model requires good statistics. 00:04:46.680 --> 00:04:50.060 align:middle line:84% And it's actually rather hard to get good statistics. 00:04:50.060 --> 00:04:54.510 align:middle line:84% And I'll show you why in class a little bit later today. 00:04:54.510 --> 00:04:58.390 align:middle line:84% The other way we could try to build a model 00:04:58.390 --> 00:05:00.030 align:middle line:90% would be mechanistically. 00:05:00.030 --> 00:05:02.350 align:middle line:84% We say, OK, well, there's some Compton scattering, 00:05:02.350 --> 00:05:03.770 align:middle line:84% and there's photoelectric effect. 00:05:03.770 --> 00:05:06.103 align:middle line:84% And then this is going to do this to the genes, and then 00:05:06.103 --> 00:05:07.570 align:middle line:90% this, and et cetera, et cetera. 00:05:07.570 --> 00:05:09.710 align:middle line:84% But you can see what the problem is. 00:05:09.710 --> 00:05:13.030 align:middle line:84% We don't understand how the body works. 00:05:13.030 --> 00:05:16.010 align:middle line:90% We have some understanding. 00:05:16.010 --> 00:05:17.970 align:middle line:84% When you get down to the biochemistry of it, 00:05:17.970 --> 00:05:20.670 align:middle line:84% there's a lot we don't really understand. 00:05:20.670 --> 00:05:25.550 align:middle line:84% And so this model is not a model that we can actually build. 00:05:25.550 --> 00:05:29.930 align:middle line:84% So we are stuck with the statistical model. 00:05:29.930 --> 00:05:34.710 align:middle line:84% But we still have insights from the mechanistic world 00:05:34.710 --> 00:05:39.110 align:middle line:84% that we can use to bound how we build these statistical models. 00:05:39.110 --> 00:05:40.910 align:middle line:90% And that's what we want to do. 00:05:40.910 --> 00:05:44.750 align:middle line:84% So my goal today is to walk you through how 00:05:44.750 --> 00:05:49.630 align:middle line:84% we got to the statistical model we have and go from there. 00:05:49.630 --> 00:05:50.170 align:middle line:90% All right. 00:05:50.170 --> 00:05:52.650 align:middle line:84% So if we want to build such a model, 00:05:52.650 --> 00:05:57.710 align:middle line:84% we need first to decide what our metrics are. 00:05:57.710 --> 00:05:58.650 align:middle line:90% What is our input? 00:05:58.650 --> 00:05:59.770 align:middle line:90% What is our output? 00:05:59.770 --> 00:06:02.130 align:middle line:90% Is the input going to be dose? 00:06:02.130 --> 00:06:07.738 align:middle line:84% Is it going to be the effective dose? 00:06:07.738 --> 00:06:08.530 align:middle line:90% What is the output? 00:06:08.530 --> 00:06:11.490 align:middle line:84% Is it going to be the cancer rate, et cetera? 00:06:11.490 --> 00:06:14.950 align:middle line:90% So here's what we have chosen. 00:06:14.950 --> 00:06:18.990 align:middle line:90% We have these definitions. 00:06:18.990 --> 00:06:20.770 align:middle line:90% So the first is what is risk? 00:06:20.770 --> 00:06:23.190 align:middle line:84% It's the baseline probability that a person 00:06:23.190 --> 00:06:25.490 align:middle line:84% in a defined population has the disease. 00:06:25.490 --> 00:06:27.950 align:middle line:84% So you have a baseline risk of cancer, which 00:06:27.950 --> 00:06:30.190 align:middle line:90% turns out to be about 20%. 00:06:30.190 --> 00:06:33.510 align:middle line:84% It's about one in five of you will probably get cancer. 00:06:33.510 --> 00:06:36.190 align:middle line:84% You have a relative risk, which is 00:06:36.190 --> 00:06:38.230 align:middle line:84% to say the exposed, in this case, 00:06:38.230 --> 00:06:40.750 align:middle line:84% to radiation, the exposed population divided 00:06:40.750 --> 00:06:43.390 align:middle line:84% by the risk of the unexposed, or control group. 00:06:43.390 --> 00:06:45.110 align:middle line:90% This is a function of dose. 00:06:45.110 --> 00:06:49.950 align:middle line:84% So here is our first mechanistic choice. 00:06:49.950 --> 00:06:52.680 align:middle line:84% Just in writing this definition, it's hidden. 00:06:52.680 --> 00:06:54.940 align:middle line:84% But there is a choice about how cancer 00:06:54.940 --> 00:07:00.700 align:middle line:90% works in writing relative risk. 00:07:00.700 --> 00:07:04.100 align:middle line:84% What this means is if a person is a smoker 00:07:04.100 --> 00:07:07.340 align:middle line:84% and they have a higher baseline risk then 00:07:07.340 --> 00:07:13.660 align:middle line:84% for a given unit of radiation, their relative risk is the same, 00:07:13.660 --> 00:07:17.740 align:middle line:84% but their absolute risk is now higher. 00:07:17.740 --> 00:07:21.460 align:middle line:84% A unit of radiation for a person who has a higher baseline 00:07:21.460 --> 00:07:26.240 align:middle line:84% risk of cancer causes a higher probability of getting cancer. 00:07:26.240 --> 00:07:28.700 align:middle line:90% Does that make sense? 00:07:28.700 --> 00:07:31.680 align:middle line:84% Because it's relative to your baseline risk. 00:07:31.680 --> 00:07:33.980 align:middle line:84% And if your baseline risk is higher, 00:07:33.980 --> 00:07:37.140 align:middle line:84% then your absolute probability is higher. 00:07:37.140 --> 00:07:40.060 align:middle line:84% And so that is a choice that we are making 00:07:40.060 --> 00:07:42.660 align:middle line:90% when we build this model. 00:07:42.660 --> 00:07:45.580 align:middle line:84% And it's not one that we actually talk about a lot. 00:07:45.580 --> 00:07:48.420 align:middle line:84% Now strictly speaking, we could have various controls 00:07:48.420 --> 00:07:51.530 align:middle line:84% to interact with different kinds of baseline risks 00:07:51.530 --> 00:07:53.090 align:middle line:90% and try to control for it. 00:07:53.090 --> 00:07:55.870 align:middle line:84% We should ask the question, what is the truth of the matter? 00:07:55.870 --> 00:07:57.110 align:middle line:90% Has someone looked into this? 00:07:57.110 --> 00:07:58.027 align:middle line:90% And the answer is yes. 00:07:58.027 --> 00:07:59.170 align:middle line:90% People look into this. 00:07:59.170 --> 00:08:03.250 align:middle line:84% And as best as they can tell, this is about correct. 00:08:03.250 --> 00:08:07.650 align:middle line:84% So for example, a famous one is radon exposure, 00:08:07.650 --> 00:08:10.170 align:middle line:84% which is an inhaled form of radiation, 00:08:10.170 --> 00:08:16.690 align:middle line:84% and smoking, another inhaled kind of carcinogen. 00:08:16.690 --> 00:08:18.630 align:middle line:90% Does of relationship hold? 00:08:18.630 --> 00:08:22.010 align:middle line:84% The answer is for light smokers, it turns out 00:08:22.010 --> 00:08:25.910 align:middle line:90% to be submultiplicative. 00:08:25.910 --> 00:08:28.590 align:middle line:84% Sorry, light smokers, it's super multiplicative. 00:08:28.590 --> 00:08:32.049 align:middle line:84% And for heavy smokers, it's submultiplicative. 00:08:32.049 --> 00:08:36.490 align:middle line:90% So it's about multiplicative. 00:08:36.490 --> 00:08:41.049 align:middle line:84% Another classic one is arsenic exposure and radon in miners. 00:08:41.049 --> 00:08:44.370 align:middle line:84% And that one is also found to be basically multiplicative. 00:08:44.370 --> 00:08:49.300 align:middle line:84% So this definition is basically supported by the evidence, 00:08:49.300 --> 00:08:52.920 align:middle line:84% but there are places where the interactions are not strictly 00:08:52.920 --> 00:08:54.080 align:middle line:90% multiplicative. 00:08:54.080 --> 00:08:56.080 align:middle line:84% This is just a simplification we have 00:08:56.080 --> 00:09:01.240 align:middle line:84% to make in order to figure out how to build the model. 00:09:01.240 --> 00:09:03.360 align:middle line:84% Then we have this definition, which we will 00:09:03.360 --> 00:09:05.300 align:middle line:90% tend to use when presenting. 00:09:05.300 --> 00:09:08.740 align:middle line:84% We typically take the relative risk, and we subtract 1. 00:09:08.740 --> 00:09:11.080 align:middle line:84% Or we take one and subtract, subtract 00:09:11.080 --> 00:09:14.640 align:middle line:84% it to get what we call the excess relative risk. 00:09:14.640 --> 00:09:18.960 align:middle line:84% So if you have a 20% chance of getting cancer, 00:09:18.960 --> 00:09:23.440 align:middle line:84% and then a 21% chance of getting cancer, 00:09:23.440 --> 00:09:25.420 align:middle line:84% if you also are exposed to radiation, 00:09:25.420 --> 00:09:29.880 align:middle line:84% then the excess relative risk is just that extra 1%. 00:09:29.880 --> 00:09:34.983 align:middle line:84% So let's look at some dose responses that 00:09:34.983 --> 00:09:35.900 align:middle line:90% are in the literature. 00:09:35.900 --> 00:09:38.360 align:middle line:84% So here's the excess relative risk of a solid cancer. 00:09:38.360 --> 00:09:40.660 align:middle line:84% A solid cancer means not leukemia. 00:09:40.660 --> 00:09:43.800 align:middle line:90% 00:09:43.800 --> 00:09:48.210 align:middle line:84% By the dose to colon-- the colon is, as it turns out, 00:09:48.210 --> 00:09:49.930 align:middle line:90% a standard organ. 00:09:49.930 --> 00:09:51.970 align:middle line:84% Every organ has a slightly different response, 00:09:51.970 --> 00:09:55.230 align:middle line:84% and people tend to use the colon for whatever reasons 00:09:55.230 --> 00:09:56.530 align:middle line:90% as the standard. 00:09:56.530 --> 00:09:59.390 align:middle line:84% And so what they typically do is they adjust all the exposures 00:09:59.390 --> 00:10:02.430 align:middle line:90% to colon-weighted exposure. 00:10:02.430 --> 00:10:05.550 align:middle line:84% And this is for the Japanese nuclear bomb 00:10:05.550 --> 00:10:06.730 align:middle line:90% survivors in Hiroshima. 00:10:06.730 --> 00:10:09.150 align:middle line:84% It's called the life span survey. 00:10:09.150 --> 00:10:13.190 align:middle line:84% And you see the data here in blue. 00:10:13.190 --> 00:10:17.870 align:middle line:84% And they've binned it into different neutron-rated doses. 00:10:17.870 --> 00:10:21.710 align:middle line:90% And there's a model there. 00:10:21.710 --> 00:10:27.650 align:middle line:84% So this model is a linear no-threshold model in red. 00:10:27.650 --> 00:10:30.630 align:middle line:84% The black are the error bars of the model. 00:10:30.630 --> 00:10:33.330 align:middle line:84% And let's just look at what it means. 00:10:33.330 --> 00:10:35.510 align:middle line:90% Linear means, well, it's linear. 00:10:35.510 --> 00:10:40.590 align:middle line:84% No threshold means that the intercept is at 0, 0. 00:10:40.590 --> 00:10:45.700 align:middle line:84% There is no threshold at which radiation produces zero effect. 00:10:45.700 --> 00:10:47.580 align:middle line:90% That's what that simply means. 00:10:47.580 --> 00:10:48.840 align:middle line:90% And what do we see? 00:10:48.840 --> 00:10:52.420 align:middle line:84% It's pretty clear it doesn't fit the data. 00:10:52.420 --> 00:10:56.300 align:middle line:84% The data rolls over at high doses. 00:10:56.300 --> 00:10:59.520 align:middle line:84% So there's no actual controversy here because no, 00:10:59.520 --> 00:11:03.500 align:middle line:84% if you are, in fact, looking at a population of people 00:11:03.500 --> 00:11:07.420 align:middle line:84% who are getting doses of 3 and 4 gray, 00:11:07.420 --> 00:11:09.340 align:middle line:84% which is equivalent to basically three or four 00:11:09.340 --> 00:11:13.180 align:middle line:84% sieverts, something like that, typically. 00:11:13.180 --> 00:11:15.300 align:middle line:84% Well, this is neutron-weighted, so that 00:11:15.300 --> 00:11:16.800 align:middle line:90% wouldn't be exactly equivalent. 00:11:16.800 --> 00:11:21.420 align:middle line:84% But anyway, if you're getting people at these high doses, 00:11:21.420 --> 00:11:22.880 align:middle line:90% you don't use the linear model. 00:11:22.880 --> 00:11:25.460 align:middle line:84% You use one of these nonlinear models, which 00:11:25.460 --> 00:11:27.220 align:middle line:90% we do have in the arsenal. 00:11:27.220 --> 00:11:32.140 align:middle line:84% Where we use the linear models, only really down here below 1. 00:11:32.140 --> 00:11:35.340 align:middle line:84% And the reason why it's controversial 00:11:35.340 --> 00:11:39.740 align:middle line:84% is because when you look at nuclear accident events, 00:11:39.740 --> 00:11:41.800 align:middle line:84% this is in fact where most people lined up. 00:11:41.800 --> 00:11:45.090 align:middle line:84% Most people get almost no dose from an accident. 00:11:45.090 --> 00:11:47.150 align:middle line:84% And they're right down here in the linear regime. 00:11:47.150 --> 00:11:49.770 align:middle line:84% And the question is, what exactly is the shape down here? 00:11:49.770 --> 00:11:53.450 align:middle line:84% And it looks like it kind of is linear. 00:11:53.450 --> 00:11:56.250 align:middle line:84% But we really would like to know. 00:11:56.250 --> 00:12:00.210 align:middle line:90% So that's the story. 00:12:00.210 --> 00:12:02.930 align:middle line:90% So let's look at some more data. 00:12:02.930 --> 00:12:07.830 align:middle line:84% This is DNA mutations in thyroid tumors of Chernobyl survivors. 00:12:07.830 --> 00:12:10.610 align:middle line:90% 00:12:10.610 --> 00:12:13.690 align:middle line:84% The maximum recorded dose here is 1000 milligray, 00:12:13.690 --> 00:12:14.590 align:middle line:90% which is one gray. 00:12:14.590 --> 00:12:16.950 align:middle line:84% So we are completely in the low-dose regime. 00:12:16.950 --> 00:12:26.415 align:middle line:84% And at first, let's just look at what we have here. 00:12:26.415 --> 00:12:28.290 align:middle line:84% The first thing to notice is there's actually 00:12:28.290 --> 00:12:31.490 align:middle line:84% considerable background population at 0 dose. 00:12:31.490 --> 00:12:34.470 align:middle line:84% This is the distribution of observations. 00:12:34.470 --> 00:12:37.410 align:middle line:84% And you see that on average is something 00:12:37.410 --> 00:12:44.640 align:middle line:84% like 25 small deletions in this genome, even at no dose. 00:12:44.640 --> 00:12:48.000 align:middle line:84% We can draw a straight line, and we 00:12:48.000 --> 00:12:51.980 align:middle line:84% can call this the baseline risk of a mutation. 00:12:51.980 --> 00:12:54.340 align:middle line:84% This is like the baseline risk of cancer, if you will. 00:12:54.340 --> 00:12:56.520 align:middle line:90% It's just a different measure. 00:12:56.520 --> 00:13:02.120 align:middle line:84% And then we can look at what the effect is after that. 00:13:02.120 --> 00:13:05.160 align:middle line:84% The other thing you can observe in this data is there seems 00:13:05.160 --> 00:13:08.080 align:middle line:84% to be a lot of people clustered here at 1,000. 00:13:08.080 --> 00:13:11.320 align:middle line:84% So clearly, something happened in how they were doing the dose 00:13:11.320 --> 00:13:12.060 align:middle line:90% reconstruction. 00:13:12.060 --> 00:13:13.977 align:middle line:84% They're basically saying oh, anyone over here, 00:13:13.977 --> 00:13:15.880 align:middle line:84% we're just going to put them into this bin, 00:13:15.880 --> 00:13:18.060 align:middle line:84% which now we have to be careful of that bin 00:13:18.060 --> 00:13:23.600 align:middle line:84% and how that affects the measurements. 00:13:23.600 --> 00:13:26.640 align:middle line:84% So there are some systematic errors, which are a problem. 00:13:26.640 --> 00:13:28.660 align:middle line:84% And indeed, if you look carefully, 00:13:28.660 --> 00:13:30.520 align:middle line:84% you'll see that there appears to be 00:13:30.520 --> 00:13:33.160 align:middle line:84% a discontinuity between the zeroth bin 00:13:33.160 --> 00:13:35.880 align:middle line:84% and the first bin, at which there's actually dose 00:13:35.880 --> 00:13:37.760 align:middle line:90% in terms of the distribution. 00:13:37.760 --> 00:13:40.030 align:middle line:84% So there's something else going on there as well 00:13:40.030 --> 00:13:41.810 align:middle line:90% between the two control groups. 00:13:41.810 --> 00:13:44.550 align:middle line:90% So lots of issues. 00:13:44.550 --> 00:13:47.110 align:middle line:84% Another issue is that you might just 00:13:47.110 --> 00:13:50.470 align:middle line:84% say, oh, I'm going to throw this into a regression 00:13:50.470 --> 00:13:52.670 align:middle line:90% to fit that line. 00:13:52.670 --> 00:13:53.590 align:middle line:90% Seems reasonable. 00:13:53.590 --> 00:13:55.110 align:middle line:84% I would just throw this into Excel, 00:13:55.110 --> 00:13:58.110 align:middle line:84% and it would perform a least squares fit. 00:13:58.110 --> 00:14:00.230 align:middle line:84% Would that be a good choice here? 00:14:00.230 --> 00:14:02.110 align:middle line:90% Why? 00:14:02.110 --> 00:14:04.295 align:middle line:84% AUDIENCE: The data is really wide. 00:14:04.295 --> 00:14:05.670 align:middle line:84% SCOTT KEMP: That in and of itself 00:14:05.670 --> 00:14:07.850 align:middle line:84% is not a problem doing least squares. 00:14:07.850 --> 00:14:11.630 align:middle line:90% 00:14:11.630 --> 00:14:15.350 align:middle line:84% A least squares fit is a maximum likelihood estimate 00:14:15.350 --> 00:14:21.050 align:middle line:84% only if the underlying data have a Gaussian distribution. 00:14:21.050 --> 00:14:22.550 align:middle line:84% Gaussian distribution has tails that 00:14:22.550 --> 00:14:24.910 align:middle line:84% go off to infinity in both directions, 00:14:24.910 --> 00:14:28.310 align:middle line:84% but you can't have fewer than zero deletions. 00:14:28.310 --> 00:14:30.610 align:middle line:84% So this is not Gaussian-distributed. 00:14:30.610 --> 00:14:34.690 align:middle line:84% And if you now fit something using least squares, 00:14:34.690 --> 00:14:36.463 align:middle line:84% you're not going to get the correct fit. 00:14:36.463 --> 00:14:38.380 align:middle line:84% You're not going to get the maximum likelihood 00:14:38.380 --> 00:14:40.560 align:middle line:90% estimation of the data. 00:14:40.560 --> 00:14:45.740 align:middle line:84% So this, turns out, is more of a binomial distribution. 00:14:45.740 --> 00:14:49.780 align:middle line:84% And so you would need to use an appropriate regression model. 00:14:49.780 --> 00:14:52.940 align:middle line:84% So all this is to say that this is not a tidy problem. 00:14:52.940 --> 00:14:54.700 align:middle line:84% And there's lots of places where people 00:14:54.700 --> 00:14:59.500 align:middle line:84% can go wrong in these types of analyzing these types of things. 00:14:59.500 --> 00:15:02.020 align:middle line:84% So should we take this fit, which 00:15:02.020 --> 00:15:04.260 align:middle line:84% is the one provided to us by the authors, which 00:15:04.260 --> 00:15:05.820 align:middle line:90% is a linear model? 00:15:05.820 --> 00:15:07.700 align:middle line:90% Well, it looks all right. 00:15:07.700 --> 00:15:10.300 align:middle line:84% It does look like there's some low region here and maybe 00:15:10.300 --> 00:15:12.500 align:middle line:90% some higher region over here. 00:15:12.500 --> 00:15:15.780 align:middle line:90% Maybe we should take this. 00:15:15.780 --> 00:15:19.760 align:middle line:84% Maybe we should do put some lowness, maybe make that green. 00:15:19.760 --> 00:15:21.860 align:middle line:84% Maybe that green one is more convincing to us, 00:15:21.860 --> 00:15:25.140 align:middle line:84% looks like actually fits the data better. 00:15:25.140 --> 00:15:28.280 align:middle line:90% Or maybe we should do that. 00:15:28.280 --> 00:15:30.700 align:middle line:84% maybe that actually fits the data better. 00:15:30.700 --> 00:15:31.380 align:middle line:90% I don't know. 00:15:31.380 --> 00:15:35.940 align:middle line:84% I mean, the data do definitely seem to go up, 00:15:35.940 --> 00:15:37.670 align:middle line:84% and then they definitely seem to go down. 00:15:37.670 --> 00:15:40.230 align:middle line:84% And then there's this big outlier point over here. 00:15:40.230 --> 00:15:42.410 align:middle line:90% And what do we do? 00:15:42.410 --> 00:15:43.710 align:middle line:90% How do we choose these models? 00:15:43.710 --> 00:15:50.590 align:middle line:84% So here's where we're going to have to make a choice. 00:15:50.590 --> 00:15:57.890 align:middle line:84% And this is the practice of model selection. 00:15:57.890 --> 00:16:00.770 align:middle line:84% And I don't think people really-- 00:16:00.770 --> 00:16:03.845 align:middle line:84% people know about overfitting data as a problem. 00:16:03.845 --> 00:16:05.970 align:middle line:84% But I don't think people are really given the tools 00:16:05.970 --> 00:16:08.210 align:middle line:90% to talk about model selection. 00:16:08.210 --> 00:16:17.130 align:middle line:84% So as a general rule, we want to choose simple models. 00:16:17.130 --> 00:16:19.610 align:middle line:84% But there's a good reason behind why we 00:16:19.610 --> 00:16:21.850 align:middle line:90% want to choose simple models. 00:16:21.850 --> 00:16:23.750 align:middle line:84% We don't need to choose the simplest model. 00:16:23.750 --> 00:16:25.810 align:middle line:84% We still want to describe the data well. 00:16:25.810 --> 00:16:27.910 align:middle line:84% But there's a preference for simple models. 00:16:27.910 --> 00:16:30.930 align:middle line:84% And that comes from Occam's razor, which is to say, 00:16:30.930 --> 00:16:37.290 align:middle line:84% basically, simple things make the most sense. 00:16:37.290 --> 00:16:39.070 align:middle line:90% Let's say I have a picture. 00:16:39.070 --> 00:16:41.850 align:middle line:90% 00:16:41.850 --> 00:16:46.930 align:middle line:84% And there's a picture of the NW14 building. 00:16:46.930 --> 00:16:50.490 align:middle line:84% And remember, there's a road here in front of it, 00:16:50.490 --> 00:16:51.710 align:middle line:90% and then there's a sidewalk. 00:16:51.710 --> 00:16:58.090 align:middle line:84% And then there are these trees planted on the sidewalk 00:16:58.090 --> 00:17:00.570 align:middle line:90% along the front of NW14. 00:17:00.570 --> 00:17:03.050 align:middle line:90% There's a little door here. 00:17:03.050 --> 00:17:12.430 align:middle line:84% And then there are some windows in the NW14 building. 00:17:12.430 --> 00:17:15.690 align:middle line:90% 00:17:15.690 --> 00:17:16.670 align:middle line:90% We're behind the trees? 00:17:16.670 --> 00:17:22.329 align:middle line:90% Maybe I should make the windows. 00:17:22.329 --> 00:17:24.270 align:middle line:84% I don't really have any different colors. 00:17:24.270 --> 00:17:27.569 align:middle line:90% 00:17:27.569 --> 00:17:28.270 align:middle line:90% There we go. 00:17:28.270 --> 00:17:30.970 align:middle line:90% 00:17:30.970 --> 00:17:31.470 align:middle line:90% All right. 00:17:31.470 --> 00:17:35.400 align:middle line:84% How many windows are in this picture? 00:17:35.400 --> 00:17:36.640 align:middle line:90% Four. 00:17:36.640 --> 00:17:37.140 align:middle line:90% All right. 00:17:37.140 --> 00:17:39.515 align:middle line:84% But how do you know there's not like a little tiny window 00:17:39.515 --> 00:17:40.060 align:middle line:90% right here? 00:17:40.060 --> 00:17:43.240 align:middle line:90% 00:17:43.240 --> 00:17:48.800 align:middle line:84% OK, maybe, but it's strange credulity. 00:17:48.800 --> 00:17:52.400 align:middle line:90% 00:17:52.400 --> 00:17:55.680 align:middle line:84% Maybe there's actually a secret little window right here. 00:17:55.680 --> 00:17:59.720 align:middle line:84% OK, It gets obviously the more complex your model space 00:17:59.720 --> 00:18:02.700 align:middle line:84% where you're having to consider all kinds of possibilities, 00:18:02.700 --> 00:18:05.100 align:middle line:90% the less likely it is. 00:18:05.100 --> 00:18:07.780 align:middle line:84% So that's the kind of intuitive explanation. 00:18:07.780 --> 00:18:11.120 align:middle line:84% But we can actually show this mathematically. 00:18:11.120 --> 00:18:13.980 align:middle line:84% So we can go back and derive Bayes' rule. 00:18:13.980 --> 00:18:18.000 align:middle line:84% So let's say we have the probability of A and B-- 00:18:18.000 --> 00:18:23.120 align:middle line:90% A and B. Well, what is that? 00:18:23.120 --> 00:18:30.760 align:middle line:84% Well, it's the probability of A times the probability of B 00:18:30.760 --> 00:18:34.250 align:middle line:84% if A happens given A. This is straightforward. 00:18:34.250 --> 00:18:37.090 align:middle line:84% And then we can just flip the variables and write the reverse. 00:18:37.090 --> 00:18:45.910 align:middle line:90% 00:18:45.910 --> 00:18:47.910 align:middle line:90% Everyone is happy with that? 00:18:47.910 --> 00:18:50.430 align:middle line:84% And does everyone agree that the probability of A and B 00:18:50.430 --> 00:18:53.230 align:middle line:84% is the same thing as the probability of B and A? 00:18:53.230 --> 00:18:56.910 align:middle line:84% Yes, so then we can just create these. 00:18:56.910 --> 00:19:01.270 align:middle line:84% Probability of A times the probability of B given A 00:19:01.270 --> 00:19:07.670 align:middle line:84% is equal to the probability of B times the probability of A given 00:19:07.670 --> 00:19:11.850 align:middle line:84% B. And then I can just divide both sides by this. 00:19:11.850 --> 00:19:15.470 align:middle line:90% 00:19:15.470 --> 00:19:17.750 align:middle line:90% And that's Bayes' law. 00:19:17.750 --> 00:19:20.890 align:middle line:84% So that's a very useful way to derive Bayes' equation 00:19:20.890 --> 00:19:23.090 align:middle line:90% if you don't really remember. 00:19:23.090 --> 00:19:26.550 align:middle line:90% 00:19:26.550 --> 00:19:28.890 align:middle line:84% This actually embodies Occam's razor. 00:19:28.890 --> 00:19:30.710 align:middle line:90% And let me just show you why. 00:19:30.710 --> 00:19:32.560 align:middle line:90% So let's make a substitution. 00:19:32.560 --> 00:19:36.500 align:middle line:90% 00:19:36.500 --> 00:19:44.080 align:middle line:84% Let's call B here a hypothesis, and let's call A the data. 00:19:44.080 --> 00:19:47.020 align:middle line:90% 00:19:47.020 --> 00:19:48.660 align:middle line:84% If I write over here, people are going 00:19:48.660 --> 00:19:50.520 align:middle line:84% to have to turn their heads too much? 00:19:50.520 --> 00:19:51.020 align:middle line:90% No? 00:19:51.020 --> 00:19:51.560 align:middle line:90% OK. 00:19:51.560 --> 00:19:52.380 align:middle line:90% All right. 00:19:52.380 --> 00:19:54.280 align:middle line:84% So we can write something like this. 00:19:54.280 --> 00:19:58.460 align:middle line:84% The probability of hypothesis one, given some data, 00:19:58.460 --> 00:20:03.520 align:middle line:84% is equal to the probability of the data given hypothesis one. 00:20:03.520 --> 00:20:05.680 align:middle line:84% Or let's actually call this model one. 00:20:05.680 --> 00:20:08.540 align:middle line:90% It's a better model one. 00:20:08.540 --> 00:20:15.300 align:middle line:84% Model one, given the data, times the probability of the model 00:20:15.300 --> 00:20:19.860 align:middle line:84% divided by the probability of the data. 00:20:19.860 --> 00:20:20.360 align:middle line:90% All right. 00:20:20.360 --> 00:20:23.660 align:middle line:84% So this is the probability that the model is true 00:20:23.660 --> 00:20:28.940 align:middle line:84% given the observed data, which is what we would like to know. 00:20:28.940 --> 00:20:31.295 align:middle line:84% And so we could also have a second model. 00:20:31.295 --> 00:20:33.170 align:middle line:84% We say, OK, well, we can have the probability 00:20:33.170 --> 00:20:37.110 align:middle line:84% of model 2 given the observed data, which are unchanging. 00:20:37.110 --> 00:20:47.450 align:middle line:90% 00:20:47.450 --> 00:20:48.770 align:middle line:90% All right. 00:20:48.770 --> 00:20:51.610 align:middle line:84% And we can just divide these two equations. 00:20:51.610 --> 00:20:55.650 align:middle line:84% So we get the probability of this minus this. 00:20:55.650 --> 00:20:58.170 align:middle line:90% These terms cancel. 00:20:58.170 --> 00:21:00.990 align:middle line:90% And so we have this times this. 00:21:00.990 --> 00:21:04.010 align:middle line:84% Let me just rewrite this here, make it a little bit cleaner. 00:21:04.010 --> 00:21:08.370 align:middle line:84% This is equal to the probability of the data given model 00:21:08.370 --> 00:21:13.770 align:middle line:84% 1 times the probability of model 1 divided 00:21:13.770 --> 00:21:18.570 align:middle line:84% by the probability of the data model 2 times 00:21:18.570 --> 00:21:21.650 align:middle line:90% probability of model 2. 00:21:21.650 --> 00:21:23.970 align:middle line:90% All right. 00:21:23.970 --> 00:21:30.880 align:middle line:84% So maybe it will make more sense if I just do that. 00:21:30.880 --> 00:21:34.000 align:middle line:90% So what is this term here? 00:21:34.000 --> 00:21:35.277 align:middle line:90% So this is your prior. 00:21:35.277 --> 00:21:37.360 align:middle line:84% This is the probability that you believe the model 00:21:37.360 --> 00:21:39.000 align:middle line:90% without the data. 00:21:39.000 --> 00:21:41.720 align:middle line:84% So is my prior belief in the probability of model 00:21:41.720 --> 00:21:44.300 align:middle line:84% 1 versus my prior belief in model 2. 00:21:44.300 --> 00:21:47.320 align:middle line:84% And if I'm agnostic between models, 00:21:47.320 --> 00:21:49.540 align:middle line:90% then this is basically unity. 00:21:49.540 --> 00:21:50.500 align:middle line:90% This goes to 1. 00:21:50.500 --> 00:21:53.360 align:middle line:90% 00:21:53.360 --> 00:21:54.400 align:middle line:90% All right. 00:21:54.400 --> 00:21:55.940 align:middle line:90% So now we simply have-- 00:21:55.940 --> 00:21:59.280 align:middle line:90% 00:21:59.280 --> 00:22:02.360 align:middle line:90% this looks a bit sloppy. 00:22:02.360 --> 00:22:03.980 align:middle line:90% Let me just rewrite over here. 00:22:03.980 --> 00:22:11.360 align:middle line:84% Probability of model 1 given the data probability of model 2. 00:22:11.360 --> 00:22:12.960 align:middle line:84% So now we have this, which is what 00:22:12.960 --> 00:22:16.240 align:middle line:84% is the ratio of that we believe in view of the data, 00:22:16.240 --> 00:22:18.120 align:middle line:90% model 1 over model 2. 00:22:18.120 --> 00:22:20.720 align:middle line:84% And it's just the ratio of the likelihoods. 00:22:20.720 --> 00:22:23.240 align:middle line:84% This is what's called the likelihood-- the probability 00:22:23.240 --> 00:22:26.710 align:middle line:84% that I see the data given the model. 00:22:26.710 --> 00:22:28.950 align:middle line:90% So here's the insight. 00:22:28.950 --> 00:22:33.030 align:middle line:84% If model one is the simple model, 00:22:33.030 --> 00:22:36.390 align:middle line:84% it has a small number of possible states. 00:22:36.390 --> 00:22:38.310 align:middle line:84% If model 2 is a complex model, it 00:22:38.310 --> 00:22:41.230 align:middle line:90% has many more possible states. 00:22:41.230 --> 00:22:46.270 align:middle line:84% So therefore, the probability of seeing the given data 00:22:46.270 --> 00:22:50.550 align:middle line:84% for a complex model is always going to be smaller. 00:22:50.550 --> 00:23:03.150 align:middle line:84% So for model 2 complex relative to 1, 00:23:03.150 --> 00:23:09.070 align:middle line:84% this ratio here is going to be greater than 1 00:23:09.070 --> 00:23:12.590 align:middle line:84% because this will be less than this. 00:23:12.590 --> 00:23:13.910 align:middle line:90% Does that make sense? 00:23:13.910 --> 00:23:17.510 align:middle line:84% And so that means we will prefer by some amount model 00:23:17.510 --> 00:23:18.510 align:middle line:90% 1 over model 2. 00:23:18.510 --> 00:23:19.790 align:middle line:90% I can just take model 2. 00:23:19.790 --> 00:23:22.870 align:middle line:84% I can multiply it by both sides, put it over here, 00:23:22.870 --> 00:23:25.580 align:middle line:84% and it has 1 times model 2, which means something 00:23:25.580 --> 00:23:29.420 align:middle line:84% greater than 1 times model 2 is the probability of model 1. 00:23:29.420 --> 00:23:33.420 align:middle line:84% So that just tells us that we have a preference for models 00:23:33.420 --> 00:23:38.100 align:middle line:84% which have more simple descriptions, which 00:23:38.100 --> 00:23:43.020 align:middle line:84% is essentially this, just mathematically. 00:23:43.020 --> 00:23:55.420 align:middle line:84% So it takes a lot of data or some kind of prior reason 00:23:55.420 --> 00:23:59.620 align:middle line:84% to support choosing a different model. 00:23:59.620 --> 00:24:03.160 align:middle line:84% And so let's look at how we decide. 00:24:03.160 --> 00:24:07.620 align:middle line:84% So one of the things that comes up often in radiation 00:24:07.620 --> 00:24:10.140 align:middle line:90% is should there be a threshold. 00:24:10.140 --> 00:24:15.940 align:middle line:84% And so one question might be, is this model actually 00:24:15.940 --> 00:24:18.240 align:middle line:84% really more complicated than this model? 00:24:18.240 --> 00:24:21.260 align:middle line:90% 00:24:21.260 --> 00:24:22.600 align:middle line:90% And the answer is yes. 00:24:22.600 --> 00:24:27.290 align:middle line:84% So in the no-threshold model, our model is something like y 00:24:27.290 --> 00:24:29.930 align:middle line:90% equals alpha x. 00:24:29.930 --> 00:24:34.450 align:middle line:84% And in the threshold model, our model is piecewise, 00:24:34.450 --> 00:24:39.170 align:middle line:84% and it looks like y equals 0 if x less than 00:24:39.170 --> 00:24:41.510 align:middle line:90% or equal to some threshold. 00:24:41.510 --> 00:24:44.690 align:middle line:84% Otherwise, x greater than threshold 00:24:44.690 --> 00:24:49.970 align:middle line:90% is beta x minus some intercept. 00:24:49.970 --> 00:24:56.810 align:middle line:84% So we have now extra parameters in the threshold model. 00:24:56.810 --> 00:25:00.210 align:middle line:84% And the threshold model thus requires more data 00:25:00.210 --> 00:25:03.250 align:middle line:84% to defend it than the no-threshold model. 00:25:03.250 --> 00:25:05.290 align:middle line:90% All right. 00:25:05.290 --> 00:25:08.850 align:middle line:90% So how is model selection done? 00:25:08.850 --> 00:25:11.170 align:middle line:84% So the problem is that when we look 00:25:11.170 --> 00:25:13.290 align:middle line:84% at this going back to the Bayes rule 00:25:13.290 --> 00:25:17.510 align:middle line:84% here, when we look at the probability of the data which. 00:25:17.510 --> 00:25:20.130 align:middle line:90% 00:25:20.130 --> 00:25:22.840 align:middle line:90% Well I deleted. 00:25:22.840 --> 00:25:26.660 align:middle line:84% And this, I was able to delete it here and not worry about it. 00:25:26.660 --> 00:25:29.440 align:middle line:84% But when you want to calculate it directly, 00:25:29.440 --> 00:25:32.800 align:middle line:84% you need to do this bit of math here. 00:25:32.800 --> 00:25:34.760 align:middle line:84% And the problem is that this bit of math 00:25:34.760 --> 00:25:39.580 align:middle line:84% is often impossible to do or very difficult to do. 00:25:39.580 --> 00:25:41.720 align:middle line:84% The way you traditionally do it numerically 00:25:41.720 --> 00:25:45.360 align:middle line:84% is through a Markov chain Monte Carlo. 00:25:45.360 --> 00:25:50.080 align:middle line:84% But analytically, it's a really difficult thing to do. 00:25:50.080 --> 00:25:53.160 align:middle line:84% So instead, what we tend to do is come up 00:25:53.160 --> 00:25:56.280 align:middle line:84% with some kind of approximation method. 00:25:56.280 --> 00:26:03.240 align:middle line:84% So the best way to do this if you have a large data set 00:26:03.240 --> 00:26:05.840 align:middle line:90% is called cross-validation. 00:26:05.840 --> 00:26:13.760 align:middle line:84% So what you do is you have a lot of samples in your data. 00:26:13.760 --> 00:26:17.000 align:middle line:84% Here's a whole bunch of people in my data set. 00:26:17.000 --> 00:26:20.990 align:middle line:84% And I train my model on this population. 00:26:20.990 --> 00:26:25.870 align:middle line:84% And then I use these people who didn't go into the trained model 00:26:25.870 --> 00:26:29.250 align:middle line:84% to test how well the model works to see, OK, 00:26:29.250 --> 00:26:33.590 align:middle line:84% what is the prediction error for each of these observations? 00:26:33.590 --> 00:26:36.490 align:middle line:84% And then I do that again, but I choose a different subset. 00:26:36.490 --> 00:26:38.910 align:middle line:84% And I go through all the possible subsets. 00:26:38.910 --> 00:26:44.750 align:middle line:84% And I basically measure how badly the model 00:26:44.750 --> 00:26:48.470 align:middle line:90% predicts the outliers. 00:26:48.470 --> 00:26:51.830 align:middle line:84% That only works if you have a lot of data. 00:26:51.830 --> 00:26:54.770 align:middle line:84% And you can actually train the model on a small subset. 00:26:54.770 --> 00:26:59.430 align:middle line:84% If you don't have a lot of data, then you're kind of in trouble. 00:26:59.430 --> 00:27:03.310 align:middle line:84% The other way to do this is that you can use a scoring function. 00:27:03.310 --> 00:27:06.690 align:middle line:84% And there are two good scoring functions. 00:27:06.690 --> 00:27:09.310 align:middle line:84% One is called the Bayesian information Criterion, 00:27:09.310 --> 00:27:14.350 align:middle line:84% and the other one is the Akaike Information Criterion. 00:27:14.350 --> 00:27:21.270 align:middle line:84% If you have reason to believe that the true model that 00:27:21.270 --> 00:27:24.430 align:middle line:84% actually contains the real physics 00:27:24.430 --> 00:27:26.710 align:middle line:84% is somehow in the set of models that you're 00:27:26.710 --> 00:27:30.870 align:middle line:84% evaluating, which is almost never the case, 00:27:30.870 --> 00:27:36.270 align:middle line:84% then BIC can be shown to be the correct way to approach it. 00:27:36.270 --> 00:27:39.910 align:middle line:84% Otherwise, the Akaike Information Criterion 00:27:39.910 --> 00:27:41.870 align:middle line:84% is usually a better way to do it. 00:27:41.870 --> 00:27:43.310 align:middle line:84% Here it is-- the number of samples 00:27:43.310 --> 00:27:46.030 align:middle line:84% times the log of the residual sums of squares, 00:27:46.030 --> 00:27:49.310 align:middle line:84% plus the number of free parameters. 00:27:49.310 --> 00:27:54.790 align:middle line:84% Basically, if you run if you run any data through NumPy 00:27:54.790 --> 00:27:59.510 align:middle line:84% or something, it has a way of just calling for this value. 00:27:59.510 --> 00:28:01.110 align:middle line:90% And you can get out the value. 00:28:01.110 --> 00:28:05.230 align:middle line:84% And basically, what you do is you get this number out 00:28:05.230 --> 00:28:07.390 align:middle line:84% for every model, and you just choose the model 00:28:07.390 --> 00:28:09.210 align:middle line:90% that has the lowest AIC. 00:28:09.210 --> 00:28:12.790 align:middle line:84% That is the model that contains the most possible information. 00:28:12.790 --> 00:28:14.310 align:middle line:84% This is a very simple result which 00:28:14.310 --> 00:28:18.140 align:middle line:84% comes from a long information-theoretic 00:28:18.140 --> 00:28:21.440 align:middle line:84% derivation, and you can go look it up if you're interested. 00:28:21.440 --> 00:28:24.020 align:middle line:84% But it's basically simply a way to penalize 00:28:24.020 --> 00:28:28.020 align:middle line:84% overfitting-- that is, at least has a foundation in information 00:28:28.020 --> 00:28:28.740 align:middle line:90% theory. 00:28:28.740 --> 00:28:30.995 align:middle line:84% So this is how you would do model selection. 00:28:30.995 --> 00:28:32.620 align:middle line:84% And you should know this because you're 00:28:32.620 --> 00:28:34.980 align:middle line:84% going to have to do model selection if you're 00:28:34.980 --> 00:28:37.780 align:middle line:84% a researcher, especially if you're a grad student. 00:28:37.780 --> 00:28:42.140 align:middle line:84% And people frequently say, oh, here's some experimental data, 00:28:42.140 --> 00:28:43.280 align:middle line:90% and it looks like this. 00:28:43.280 --> 00:28:44.780 align:middle line:84% And you fit something, and you maybe 00:28:44.780 --> 00:28:47.180 align:middle line:90% don't give it a second thought. 00:28:47.180 --> 00:28:51.420 align:middle line:84% And really, you should be trying multiple models that possibly 00:28:51.420 --> 00:28:54.220 align:middle line:84% make sense and scoring them either using 00:28:54.220 --> 00:28:59.780 align:middle line:84% k-fold cross-validation or using AIC before just choosing one 00:28:59.780 --> 00:29:01.420 align:middle line:90% for your thesis. 00:29:01.420 --> 00:29:02.860 align:middle line:90% Don't go on R-squared. 00:29:02.860 --> 00:29:07.840 align:middle line:84% R-squared only tells you how well it matches the data points. 00:29:07.840 --> 00:29:10.680 align:middle line:84% It doesn't tell you how much information is in the model. 00:29:10.680 --> 00:29:13.540 align:middle line:90% 00:29:13.540 --> 00:29:18.310 align:middle line:84% This is elementary, but if I have these data points like this 00:29:18.310 --> 00:29:20.650 align:middle line:84% and I have a model that does this, 00:29:20.650 --> 00:29:22.750 align:middle line:84% it fits the fits the data perfectly, 00:29:22.750 --> 00:29:27.210 align:middle line:84% but probably has almost no predictive value. 00:29:27.210 --> 00:29:27.970 align:middle line:90% OK. 00:29:27.970 --> 00:29:32.810 align:middle line:84% So this is where linear no threshold comes from. 00:29:32.810 --> 00:29:36.410 align:middle line:84% If you actually go through this process rigorously and you start 00:29:36.410 --> 00:29:42.090 align:middle line:84% testing a bunch of models, as I have been doing lately, 00:29:42.090 --> 00:29:45.470 align:middle line:84% you will find that the LNT model, 00:29:45.470 --> 00:29:47.690 align:middle line:84% this is a k-fold cross-validation result 00:29:47.690 --> 00:29:50.490 align:middle line:90% for a 20-fold cross-validation. 00:29:50.490 --> 00:29:52.715 align:middle line:84% This is for the lifespan survey data. 00:29:52.715 --> 00:29:54.090 align:middle line:84% These are the people in Hiroshima 00:29:54.090 --> 00:29:56.530 align:middle line:84% who were exposed to radiation from the atom bomb, which 00:29:56.530 --> 00:29:59.470 align:middle line:84% is the largest single data set for low-dose exposure. 00:29:59.470 --> 00:30:02.770 align:middle line:84% I think I have some information about how many people 00:30:02.770 --> 00:30:05.770 align:middle line:90% are in a subsequent slide. 00:30:05.770 --> 00:30:09.650 align:middle line:84% Only below 1 sievert-- so we are in the potentially linear 00:30:09.650 --> 00:30:10.490 align:middle line:90% regime. 00:30:10.490 --> 00:30:12.470 align:middle line:84% And these are different models-- linear model, 00:30:12.470 --> 00:30:14.370 align:middle line:84% linear-quadratic model, a threshold model 00:30:14.370 --> 00:30:16.680 align:middle line:84% where we allow the threshold to exist 00:30:16.680 --> 00:30:19.840 align:middle line:84% at the optimal possible location that give the best 00:30:19.840 --> 00:30:21.840 align:middle line:90% possible fit to the data. 00:30:21.840 --> 00:30:23.460 align:middle line:90% Quadratic, this is a Gaussian. 00:30:23.460 --> 00:30:25.700 align:middle line:84% These are non-parametric models, Gaussian process, 00:30:25.700 --> 00:30:27.840 align:middle line:90% Linear Gaussian process other. 00:30:27.840 --> 00:30:31.960 align:middle line:84% And you see that basically the linear model just 00:30:31.960 --> 00:30:33.440 align:middle line:90% constantly wins. 00:30:33.440 --> 00:30:38.960 align:middle line:84% And this is frustrating because I would like to-- 00:30:38.960 --> 00:30:43.400 align:middle line:84% I would like a model that explains why this happens. 00:30:43.400 --> 00:30:47.320 align:middle line:84% But that shape, you'll see it over and over. 00:30:47.320 --> 00:30:52.280 align:middle line:84% But statistically, the only thing we can really defend 00:30:52.280 --> 00:30:53.640 align:middle line:90% is the linear model. 00:30:53.640 --> 00:30:54.140 align:middle line:90% All right. 00:30:54.140 --> 00:30:55.807 align:middle line:84% Now I have to go through all this again. 00:30:55.807 --> 00:30:58.360 align:middle line:90% 00:30:58.360 --> 00:31:05.240 align:middle line:84% So there are some other pitfalls that you 00:31:05.240 --> 00:31:09.040 align:middle line:84% should be aware of when looking at radiation dose-response 00:31:09.040 --> 00:31:12.840 align:middle line:84% models that sometimes people fall into. 00:31:12.840 --> 00:31:15.830 align:middle line:84% And here's an example of one of them. 00:31:15.830 --> 00:31:19.910 align:middle line:84% So this is the excess relative risk 00:31:19.910 --> 00:31:29.310 align:middle line:84% for cancer for exposures per sievert of radiation 00:31:29.310 --> 00:31:31.730 align:middle line:90% to different organs. 00:31:31.730 --> 00:31:34.970 align:middle line:84% And so remember we talked about colon-weighted dose. 00:31:34.970 --> 00:31:39.510 align:middle line:84% You can see here that it's not uniform depending on the tissue. 00:31:39.510 --> 00:31:40.230 align:middle line:90% All right. 00:31:40.230 --> 00:31:40.970 align:middle line:90% So let's see. 00:31:40.970 --> 00:31:45.990 align:middle line:84% So 0 excess risk means that that's your baseline risk 00:31:45.990 --> 00:31:47.550 align:middle line:90% for no radiation. 00:31:47.550 --> 00:31:51.750 align:middle line:84% And here, we see that for uterus exposure, 00:31:51.750 --> 00:31:55.750 align:middle line:84% the probability of cancer is actually lower. 00:31:55.750 --> 00:32:00.170 align:middle line:84% So should we go around irradiating women's uteruses? 00:32:00.170 --> 00:32:04.670 align:middle line:90% 00:32:04.670 --> 00:32:05.750 align:middle line:90% AUDIENCE: Just for fun? 00:32:05.750 --> 00:32:09.430 align:middle line:84% SCOTT KEMP: No, for their benefit. 00:32:09.430 --> 00:32:12.460 align:middle line:84% Because we will reduce the risk of uterine cancer. 00:32:12.460 --> 00:32:13.880 align:middle line:90% Should we go and do that? 00:32:13.880 --> 00:32:17.260 align:middle line:90% 00:32:17.260 --> 00:32:18.720 align:middle line:90% This data is correct. 00:32:18.720 --> 00:32:20.540 align:middle line:90% I mean, the real data. 00:32:20.540 --> 00:32:22.020 align:middle line:90% AUDIENCE: How much data is it? 00:32:22.020 --> 00:32:24.040 align:middle line:84% SCOTT KEMP: Well, that's a good question. 00:32:24.040 --> 00:32:25.100 align:middle line:90% There's the error bar. 00:32:25.100 --> 00:32:31.403 align:middle line:90% 00:32:31.403 --> 00:32:33.320 align:middle line:84% AUDIENCE: The gist is it doesn't cause cancer, 00:32:33.320 --> 00:32:35.230 align:middle line:84% but then maybe there's other effects 00:32:35.230 --> 00:32:36.480 align:middle line:90% that we're not accounting for. 00:32:36.480 --> 00:32:39.263 align:middle line:90% 00:32:39.263 --> 00:32:40.680 align:middle line:84% SCOTT KEMP: There's other effects. 00:32:40.680 --> 00:32:41.960 align:middle line:90% So you're being precautionary. 00:32:41.960 --> 00:32:43.780 align:middle line:84% But on the basis of cancer, you don't 00:32:43.780 --> 00:32:45.627 align:middle line:90% have an argument against it. 00:32:45.627 --> 00:32:46.960 align:middle line:90% AUDIENCE: Cancer for the mother? 00:32:46.960 --> 00:32:48.380 align:middle line:90% SCOTT KEMP: For the uterus. 00:32:48.380 --> 00:32:50.300 align:middle line:90% AUDIENCE: Yeah. 00:32:50.300 --> 00:32:52.420 align:middle line:84% AUDIENCE: You just have to look at how 00:32:52.420 --> 00:32:54.200 align:middle line:84% cancer affects the quality of life, 00:32:54.200 --> 00:32:56.617 align:middle line:90% not just the incidence of some. 00:32:56.617 --> 00:32:58.200 align:middle line:84% SCOTT KEMP: Well, if you don't get it, 00:32:58.200 --> 00:32:59.325 align:middle line:90% that's better than nothing. 00:32:59.325 --> 00:33:02.940 align:middle line:90% 00:33:02.940 --> 00:33:05.380 align:middle line:84% This is called the look-again effect. 00:33:05.380 --> 00:33:07.580 align:middle line:84% Have you heard of the look-again effect. 00:33:07.580 --> 00:33:08.160 align:middle line:90% All right. 00:33:08.160 --> 00:33:13.690 align:middle line:84% So if these error bars are 95% confidence intervals, which 00:33:13.690 --> 00:33:16.690 align:middle line:84% means 95% of the observations are estimated 00:33:16.690 --> 00:33:20.490 align:middle line:84% to fall within the error bar, which means one out of every 20 00:33:20.490 --> 00:33:22.450 align:middle line:84% is going to fall out of the error. 00:33:22.450 --> 00:33:25.230 align:middle line:84% The estimation will be wrong by the amount of the error bar. 00:33:25.230 --> 00:33:29.150 align:middle line:84% So there's about 20 things up here. 00:33:29.150 --> 00:33:31.590 align:middle line:90% And one of them is wrong. 00:33:31.590 --> 00:33:33.090 align:middle line:90% And that's what it is. 00:33:33.090 --> 00:33:37.250 align:middle line:84% But if you look at all of these things 00:33:37.250 --> 00:33:42.930 align:middle line:84% and then just picked this result out and wrote a paper, 00:33:42.930 --> 00:33:45.270 align:middle line:90% then you would have a problem. 00:33:45.270 --> 00:33:49.490 align:middle line:84% So in fact, this is easily shown if we also plot the mortality 00:33:49.490 --> 00:33:50.670 align:middle line:90% from radiation risk. 00:33:50.670 --> 00:33:53.870 align:middle line:84% And you see, it was just a statistical fluke. 00:33:53.870 --> 00:33:57.290 align:middle line:84% The mortality shows that, in fact, radiation exposure 00:33:57.290 --> 00:34:00.170 align:middle line:84% does increase your mortality of uterine cancer, 00:34:00.170 --> 00:34:02.170 align:middle line:84% even though it appears to reduce the probability 00:34:02.170 --> 00:34:03.370 align:middle line:90% of uterine cancer. 00:34:03.370 --> 00:34:06.810 align:middle line:84% And here you see here the effect has shown up here 00:34:06.810 --> 00:34:10.260 align:middle line:84% for oral and pancreas in the opposite direction. 00:34:10.260 --> 00:34:11.780 align:middle line:90% It's just this effect. 00:34:11.780 --> 00:34:15.480 align:middle line:84% So there was a famous XKCD cartoon about this. 00:34:15.480 --> 00:34:17.139 align:middle line:90% Jelly beans cause acne. 00:34:17.139 --> 00:34:18.400 align:middle line:90% Scientists investigate. 00:34:18.400 --> 00:34:21.360 align:middle line:84% And there's this claim that they found this link. 00:34:21.360 --> 00:34:24.739 align:middle line:90% And then here is this thing. 00:34:24.739 --> 00:34:27.120 align:middle line:84% It appears to be a certain color. 00:34:27.120 --> 00:34:29.199 align:middle line:84% And so they go through all the colors. 00:34:29.199 --> 00:34:34.120 align:middle line:84% Found no link between purple and brown and pink 00:34:34.120 --> 00:34:37.280 align:middle line:84% and blue and teal and salmon color, 00:34:37.280 --> 00:34:40.400 align:middle line:84% red color, turquoise, magenta, yellow. 00:34:40.400 --> 00:34:44.840 align:middle line:84% The p-value is always too large to defend any link. 00:34:44.840 --> 00:34:47.699 align:middle line:84% And then, lo and behold, they get to green. 00:34:47.699 --> 00:34:54.880 align:middle line:84% And the p-value appears to justify that there is acne 00:34:54.880 --> 00:34:55.900 align:middle line:90% caused by jelly beans. 00:34:55.900 --> 00:34:57.700 align:middle line:84% And then you have this publication. 00:34:57.700 --> 00:35:00.680 align:middle line:90% 00:35:00.680 --> 00:35:05.260 align:middle line:84% If you look at the data and you start ignoring things, 00:35:05.260 --> 00:35:10.510 align:middle line:84% you actually have to adjust how you calculate the p-value, 00:35:10.510 --> 00:35:12.730 align:middle line:84% because every time you pull a sample, 00:35:12.730 --> 00:35:16.630 align:middle line:84% every time you run a test, you are actually adjusting. 00:35:16.630 --> 00:35:19.990 align:middle line:84% You are making choices and filtering the data, 00:35:19.990 --> 00:35:22.550 align:middle line:84% and therefore you have to actually change 00:35:22.550 --> 00:35:24.090 align:middle line:90% how you do that calculation. 00:35:24.090 --> 00:35:27.070 align:middle line:84% This is frequently forgotten about. 00:35:27.070 --> 00:35:30.190 align:middle line:84% And people do this, and they run their science experiment 00:35:30.190 --> 00:35:32.610 align:middle line:84% on their instrument until they get the result they want. 00:35:32.610 --> 00:35:35.230 align:middle line:84% Now finally, the instrument worked right. 00:35:35.230 --> 00:35:36.830 align:middle line:90% And then they publish. 00:35:36.830 --> 00:35:39.710 align:middle line:84% It's like, well, actually, you have 00:35:39.710 --> 00:35:42.790 align:middle line:84% to take into account all of those unsuccessful runs, 00:35:42.790 --> 00:35:45.470 align:middle line:90% and maybe it's just noise. 00:35:45.470 --> 00:35:46.850 align:middle line:90% So this happens all the time. 00:35:46.850 --> 00:35:48.270 align:middle line:84% And one of the ways this manifests 00:35:48.270 --> 00:35:51.910 align:middle line:84% is that people get these results out of small little laboratory 00:35:51.910 --> 00:35:53.670 align:middle line:84% experiments about radiation, and they 00:35:53.670 --> 00:35:56.010 align:middle line:84% show that some radiation is good for you. 00:35:56.010 --> 00:35:58.590 align:middle line:84% And then they say, Lo, and we have discovered hormesis, 00:35:58.590 --> 00:36:03.310 align:middle line:84% this idea that a little bit of radiation is good for you, 00:36:03.310 --> 00:36:05.380 align:middle line:90% and it's not real. 00:36:05.380 --> 00:36:11.940 align:middle line:90% So you've been warned. 00:36:11.940 --> 00:36:13.180 align:middle line:90% All right. 00:36:13.180 --> 00:36:19.140 align:middle line:84% This is kind of like the grand plot from the lifespan survey. 00:36:19.140 --> 00:36:21.512 align:middle line:84% This is from ICRP Publication 99, which 00:36:21.512 --> 00:36:22.720 align:middle line:90% is actually assigned reading. 00:36:22.720 --> 00:36:24.620 align:middle line:90% I hope you looked at it. 00:36:24.620 --> 00:36:28.103 align:middle line:84% And here is up to two sieverts of dose. 00:36:28.103 --> 00:36:29.520 align:middle line:84% And you can see the linear regime. 00:36:29.520 --> 00:36:31.620 align:middle line:84% And then you see in the low-dose regime, 00:36:31.620 --> 00:36:35.360 align:middle line:84% there does appear to be this like super linear region. 00:36:35.360 --> 00:36:40.820 align:middle line:90% 00:36:40.820 --> 00:36:42.940 align:middle line:84% You could say, well, maybe that's just it does 00:36:42.940 --> 00:36:45.200 align:middle line:84% appear consistently between data points. 00:36:45.200 --> 00:36:47.580 align:middle line:90% So that could be. 00:36:47.580 --> 00:36:49.900 align:middle line:84% It looks like it's not just noise. 00:36:49.900 --> 00:36:53.260 align:middle line:84% But I have tried as I could to find a non-linear model 00:36:53.260 --> 00:36:55.300 align:middle line:90% to describe this. 00:36:55.300 --> 00:36:56.580 align:middle line:90% You can't do it. 00:36:56.580 --> 00:36:57.860 align:middle line:90% It does not. 00:36:57.860 --> 00:37:01.500 align:middle line:84% Nothing is statistically justifiable. 00:37:01.500 --> 00:37:05.730 align:middle line:84% So I mentioned this data set earlier. 00:37:05.730 --> 00:37:09.310 align:middle line:84% It's the single largest data set. 00:37:09.310 --> 00:37:13.970 align:middle line:84% It tracked 80,000 individuals who were exposed to radiation 00:37:13.970 --> 00:37:16.850 align:middle line:90% because of the atom bomb. 00:37:16.850 --> 00:37:21.010 align:middle line:84% A lot of people think, oh, well, that's a lot of radiation. 00:37:21.010 --> 00:37:25.890 align:middle line:84% Well, it turns out that 64,000 of those 80,000 were exposed 00:37:25.890 --> 00:37:28.690 align:middle line:90% to less than 100 millisieverts. 00:37:28.690 --> 00:37:33.670 align:middle line:84% So most of these people are really in the low-dose regime. 00:37:33.670 --> 00:37:36.290 align:middle line:84% To put that into perspective, 100 million millisieverts is 00:37:36.290 --> 00:37:39.730 align:middle line:84% about the same amount of dose a radiation worker in the United 00:37:39.730 --> 00:37:43.050 align:middle line:84% States would be allowed to receive over a 20-year career. 00:37:43.050 --> 00:37:47.810 align:middle line:90% So it's not that much radiation. 00:37:47.810 --> 00:37:53.730 align:middle line:84% So we do have 64,000 observations in here. 00:37:53.730 --> 00:37:56.930 align:middle line:84% And it still doesn't support anything other 00:37:56.930 --> 00:37:58.550 align:middle line:90% than the linear no threshold. 00:37:58.550 --> 00:38:01.290 align:middle line:90% 00:38:01.290 --> 00:38:03.870 align:middle line:84% One of the things I want to talk about, 00:38:03.870 --> 00:38:07.822 align:middle line:84% though, is just to look at the slopes here. 00:38:07.822 --> 00:38:11.010 align:middle line:84% Right now, the ICRP, the International Committee 00:38:11.010 --> 00:38:14.250 align:middle line:84% on Radiation Protection, which is one of several bodies 00:38:14.250 --> 00:38:17.850 align:middle line:84% that recommend safety standards for radiation, 00:38:17.850 --> 00:38:22.170 align:middle line:84% they need to publish a number for the slope of what this dose 00:38:22.170 --> 00:38:23.370 align:middle line:90% response is-- 00:38:23.370 --> 00:38:27.210 align:middle line:84% how much cancer you're going to get for a certain dose. 00:38:27.210 --> 00:38:31.930 align:middle line:90% So the ICRP recommends 0.53. 00:38:31.930 --> 00:38:39.450 align:middle line:84% But if you read the slope of that chart, it's actually 0.64. 00:38:39.450 --> 00:38:43.310 align:middle line:84% And if you really limit yourself to the low-dose regime, 00:38:43.310 --> 00:38:47.290 align:middle line:84% you would actually read one, about one per sievert, 00:38:47.290 --> 00:38:50.610 align:middle line:84% excess relative risk of one per sievert. 00:38:50.610 --> 00:38:55.430 align:middle line:84% So should we conclude that the ICRP is being conservative, 00:38:55.430 --> 00:38:58.270 align:middle line:84% and that the linear no-threshold model is conservative, 00:38:58.270 --> 00:38:59.570 align:middle line:90% as a lot of people say? 00:38:59.570 --> 00:39:01.840 align:middle line:90% No, in fact, the opposite-- 00:39:01.840 --> 00:39:04.960 align:middle line:84% the data show that it is not conservative. 00:39:04.960 --> 00:39:08.720 align:middle line:84% They are being generous in allowing more radiation out 00:39:08.720 --> 00:39:13.360 align:middle line:84% there than is defendable and as opposed to the idea 00:39:13.360 --> 00:39:16.400 align:middle line:84% that LNT is somehow a conservative model. 00:39:16.400 --> 00:39:19.180 align:middle line:84% It is also not supported because of the statistics. 00:39:19.180 --> 00:39:24.960 align:middle line:84% The threshold, the threshold model is not really supported. 00:39:24.960 --> 00:39:32.000 align:middle line:84% So let's say, we really liked the idea of a threshold 00:39:32.000 --> 00:39:35.080 align:middle line:84% because most people in the nuclear community 00:39:35.080 --> 00:39:39.160 align:middle line:90% would love that to be true. 00:39:39.160 --> 00:39:43.600 align:middle line:84% Someone here was doing a thesis on this. 00:39:43.600 --> 00:39:49.000 align:middle line:84% And so we could ask, OK, what would it 00:39:49.000 --> 00:39:54.120 align:middle line:84% take to actually establish that a threshold does 00:39:54.120 --> 00:39:56.800 align:middle line:90% exist at some point? 00:39:56.800 --> 00:39:58.920 align:middle line:84% And so we can do that calculation. 00:39:58.920 --> 00:40:02.210 align:middle line:84% So let's just do it really quickly here. 00:40:02.210 --> 00:40:04.950 align:middle line:90% 00:40:04.950 --> 00:40:07.950 align:middle line:84% We set up a little thought experiment. 00:40:07.950 --> 00:40:11.010 align:middle line:84% And this is by the way, borrowed from ICRP Publication 99-- 00:40:11.010 --> 00:40:15.590 align:middle line:84% section 2.42 if you want to go read their version of it. 00:40:15.590 --> 00:40:17.630 align:middle line:84% And what we're going to do is ask 00:40:17.630 --> 00:40:19.950 align:middle line:84% how many observations or basically, what does it 00:40:19.950 --> 00:40:24.630 align:middle line:84% take to observe a threshold at one millisievert? 00:40:24.630 --> 00:40:30.230 align:middle line:84% Now, one millisievert is not all that low, 00:40:30.230 --> 00:40:35.110 align:middle line:84% but I think it's like just 50% of natural background. 00:40:35.110 --> 00:40:37.890 align:middle line:84% So that's not really super low dose. 00:40:37.890 --> 00:40:39.670 align:middle line:84% People who talk about threshold models 00:40:39.670 --> 00:40:42.950 align:middle line:84% often talk about even lower thresholds. 00:40:42.950 --> 00:40:46.435 align:middle line:84% But it seems like a reasonable place to put your test. 00:40:46.435 --> 00:40:47.810 align:middle line:84% If you put it really, really low, 00:40:47.810 --> 00:40:50.150 align:middle line:84% it gets really, really hard to see because you 00:40:50.150 --> 00:40:51.977 align:middle line:90% have very, very small effect. 00:40:51.977 --> 00:40:54.310 align:middle line:84% So we're going to be a little generous, and we'll put it 00:40:54.310 --> 00:40:57.150 align:middle line:84% we'll put it at one millisievert-- 00:40:57.150 --> 00:40:59.140 align:middle line:90% half of background. 00:40:59.140 --> 00:40:59.640 align:middle line:90% All right. 00:40:59.640 --> 00:41:01.300 align:middle line:84% Now we need to make some assumptions 00:41:01.300 --> 00:41:05.520 align:middle line:84% about how the world works in order to run this model. 00:41:05.520 --> 00:41:07.020 align:middle line:84% That first assumption is we're going 00:41:07.020 --> 00:41:11.540 align:middle line:84% to assume that the baseline lifetime lethal cancer risk 00:41:11.540 --> 00:41:13.620 align:middle line:90% is 10%. 00:41:13.620 --> 00:41:17.420 align:middle line:84% Now, earlier, I said it was 20%, which is true. 00:41:17.420 --> 00:41:19.420 align:middle line:84% But what we're going to do is assume 00:41:19.420 --> 00:41:23.260 align:middle line:84% it's 10% in this little thought experiment. 00:41:23.260 --> 00:41:24.540 align:middle line:90% Why? 00:41:24.540 --> 00:41:27.920 align:middle line:84% Well, if the background rate is lower, 00:41:27.920 --> 00:41:32.260 align:middle line:84% it's easier to see the effect of the radiation. 00:41:32.260 --> 00:41:37.060 align:middle line:84% So it's going to bias the result of our model 00:41:37.060 --> 00:41:40.060 align:middle line:84% to make it appear to be easy to see the threshold, 00:41:40.060 --> 00:41:42.060 align:middle line:90% easier than reality. 00:41:42.060 --> 00:41:44.900 align:middle line:90% Does that make sense? 00:41:44.900 --> 00:41:47.180 align:middle line:90% All right. 00:41:47.180 --> 00:41:48.320 align:middle line:90% We need another assumption. 00:41:48.320 --> 00:41:53.180 align:middle line:84% We need an assumption about what of excess risk of cancer 00:41:53.180 --> 00:41:55.660 align:middle line:90% is per unit radiation exposure. 00:41:55.660 --> 00:41:59.850 align:middle line:84% So let's adopt the excess relative risk of one 00:41:59.850 --> 00:42:02.250 align:middle line:90% per sievert. 00:42:02.250 --> 00:42:07.610 align:middle line:84% Now remember that the ICRP recommended 0.53. 00:42:07.610 --> 00:42:10.290 align:middle line:84% But we are going to take one, which 00:42:10.290 --> 00:42:14.290 align:middle line:84% is to assume that the effective radiation is stronger 00:42:14.290 --> 00:42:17.730 align:middle line:84% than the consensus view, which makes it easier 00:42:17.730 --> 00:42:20.190 align:middle line:84% to see the effects of a little bit of dose of radiation 00:42:20.190 --> 00:42:22.810 align:middle line:84% and makes it artificially easy to estimate 00:42:22.810 --> 00:42:25.110 align:middle line:90% the existence of a threshold. 00:42:25.110 --> 00:42:28.110 align:middle line:84% So both of these are biased in the same direction. 00:42:28.110 --> 00:42:31.690 align:middle line:84% It's making the experiment easier than it really is. 00:42:31.690 --> 00:42:35.410 align:middle line:84% And three, we're going to assume perfect experimental conditions. 00:42:35.410 --> 00:42:37.930 align:middle line:84% You say we can have any population. 00:42:37.930 --> 00:42:40.630 align:middle line:84% We'll know the dose that everyone gets perfectly. 00:42:40.630 --> 00:42:43.850 align:middle line:84% There's not going to be any uncertainties there. 00:42:43.850 --> 00:42:46.730 align:middle line:84% And we'll know all the control parameters exactly. 00:42:46.730 --> 00:42:51.410 align:middle line:84% We don't need a fancy control group. 00:42:51.410 --> 00:42:54.490 align:middle line:84% We don't have confounders from people smoking and other stuff 00:42:54.490 --> 00:42:55.860 align:middle line:90% noising up the data. 00:42:55.860 --> 00:42:59.783 align:middle line:84% We have exactly the right number of identical individuals, 00:42:59.783 --> 00:43:01.200 align:middle line:84% and we're going to all expose them 00:43:01.200 --> 00:43:04.480 align:middle line:84% to the exact unambiguous amount of dose, 00:43:04.480 --> 00:43:08.000 align:middle line:84% which, again, is making the experiment artificially easy. 00:43:08.000 --> 00:43:11.880 align:middle line:84% So we have made all of these generous assumptions that 00:43:11.880 --> 00:43:16.000 align:middle line:84% will cause us to underestimate the difficulty of actually 00:43:16.000 --> 00:43:18.040 align:middle line:90% looking for a threshold. 00:43:18.040 --> 00:43:18.540 align:middle line:90% All right. 00:43:18.540 --> 00:43:21.200 align:middle line:90% So now we have some hypotheses. 00:43:21.200 --> 00:43:25.780 align:middle line:84% And we're going to set the null hypothesis to the threshold. 00:43:25.780 --> 00:43:28.800 align:middle line:84% We'll say the null hypothesis is that at one millisievert, 00:43:28.800 --> 00:43:30.040 align:middle line:90% the dose is zero. 00:43:30.040 --> 00:43:31.600 align:middle line:90% The effect is zero. 00:43:31.600 --> 00:43:33.620 align:middle line:84% At 1 millisievert dose, the effect is zero. 00:43:33.620 --> 00:43:34.840 align:middle line:90% That's our null. 00:43:34.840 --> 00:43:36.562 align:middle line:90% That's the threshold model. 00:43:36.562 --> 00:43:38.520 align:middle line:84% And then we'll have this alternative hypothesis 00:43:38.520 --> 00:43:41.960 align:middle line:84% that says that the relative risk at one millisievert 00:43:41.960 --> 00:43:47.040 align:middle line:84% will be as predicted by the linear model, one per sievert. 00:43:47.040 --> 00:43:49.600 align:middle line:90% So those are our two hypotheses. 00:43:49.600 --> 00:43:50.280 align:middle line:90% OK. 00:43:50.280 --> 00:43:53.400 align:middle line:84% So let's see what we would expect 00:43:53.400 --> 00:43:58.010 align:middle line:84% will assume to make the math easy incorrectly, 00:43:58.010 --> 00:44:03.790 align:middle line:84% that these are Gaussians which enlarge n is 00:44:03.790 --> 00:44:06.630 align:middle line:90% a reasonable approximation. 00:44:06.630 --> 00:44:11.730 align:middle line:84% And so here we have the null hypothesis, 00:44:11.730 --> 00:44:13.910 align:middle line:90% which is the threshold model. 00:44:13.910 --> 00:44:17.470 align:middle line:84% We expect that the population sample 00:44:17.470 --> 00:44:23.950 align:middle line:84% will have a mean at the baseline cancer risk of 10%. 00:44:23.950 --> 00:44:25.630 align:middle line:84% Because at one millisievert, there's 00:44:25.630 --> 00:44:26.930 align:middle line:90% no effect from radiation. 00:44:26.930 --> 00:44:29.190 align:middle line:84% So you just get the baseline risk. 00:44:29.190 --> 00:44:32.950 align:middle line:84% And then there will be some statistical variation 00:44:32.950 --> 00:44:36.030 align:middle line:90% around that expected mean. 00:44:36.030 --> 00:44:40.230 align:middle line:84% If hypothesis one is correct, then we expect the background 00:44:40.230 --> 00:44:45.070 align:middle line:84% rate plus the excess risk, which is the background rate-- 00:44:45.070 --> 00:44:46.830 align:middle line:90% remember because it's relative-- 00:44:46.830 --> 00:44:51.030 align:middle line:84% times 1 per sievert times 1 millisievert. 00:44:51.030 --> 00:44:54.420 align:middle line:90% So it's going to be 0.1001. 00:44:54.420 --> 00:44:57.360 align:middle line:84% So that's where we expect the mean of the distribution to be. 00:44:57.360 --> 00:45:01.060 align:middle line:84% And again, we'll have some variance around that. 00:45:01.060 --> 00:45:03.700 align:middle line:84% So the question is, how big of a sample 00:45:03.700 --> 00:45:08.580 align:middle line:84% do we need so that the uncertainty is squished 00:45:08.580 --> 00:45:11.740 align:middle line:84% sufficiently small that we can tell 00:45:11.740 --> 00:45:14.620 align:middle line:84% the difference between these two distributions. 00:45:14.620 --> 00:45:17.700 align:middle line:90% That's the test of what we need. 00:45:17.700 --> 00:45:20.140 align:middle line:90% So we need to choose two things. 00:45:20.140 --> 00:45:23.860 align:middle line:84% We need to choose this critical value, which will be our cutoff. 00:45:23.860 --> 00:45:28.820 align:middle line:84% And we need to choose the sample size, which will determine 00:45:28.820 --> 00:45:31.380 align:middle line:90% the width of this distribution. 00:45:31.380 --> 00:45:36.140 align:middle line:84% So the cutoff is basically where we decide to accept or reject 00:45:36.140 --> 00:45:39.780 align:middle line:90% the null hypothesis. 00:45:39.780 --> 00:45:43.660 align:middle line:84% If we see a cancer rate above this level, 00:45:43.660 --> 00:45:45.100 align:middle line:90% we'll reject the null. 00:45:45.100 --> 00:45:49.260 align:middle line:84% Otherwise, we're going to assume that the null hypothesis, which 00:45:49.260 --> 00:45:52.810 align:middle line:84% is in this case is the threshold model is correct. 00:45:52.810 --> 00:45:54.930 align:middle line:84% So we are being very generous because what 00:45:54.930 --> 00:46:00.090 align:middle line:84% we've done is we've used standard 95% confidence here 00:46:00.090 --> 00:46:05.690 align:middle line:84% and 80% power, which is to say that we will accidentally reject 00:46:05.690 --> 00:46:11.010 align:middle line:90% the null only 5% of the time. 00:46:11.010 --> 00:46:16.250 align:middle line:84% And by setting 80% power for the other test, 00:46:16.250 --> 00:46:23.210 align:middle line:84% it's OK if we reject the LNT theory at least 20% of the time 00:46:23.210 --> 00:46:25.250 align:middle line:90% that it was correct. 00:46:25.250 --> 00:46:28.170 align:middle line:84% We don't really care about getting the LNT part wrong, 00:46:28.170 --> 00:46:32.690 align:middle line:84% but we want to be very careful not to accidentally reject 00:46:32.690 --> 00:46:34.703 align:middle line:84% the threshold model as being wrong. 00:46:34.703 --> 00:46:36.870 align:middle line:84% So we're being very generous to the threshold model. 00:46:36.870 --> 00:46:38.790 align:middle line:84% Normally, you would do this in the reverse. 00:46:38.790 --> 00:46:41.770 align:middle line:84% You would have a higher standard for the threshold model 00:46:41.770 --> 00:46:44.130 align:middle line:84% because it's a more complicated model. 00:46:44.130 --> 00:46:46.290 align:middle line:84% But here we're going to be generous and assume 00:46:46.290 --> 00:46:50.080 align:middle line:84% the threshold model is the notional model. 00:46:50.080 --> 00:46:54.040 align:middle line:84% So now we need to basically write 00:46:54.040 --> 00:46:57.280 align:middle line:84% what I've drawn here in some equations 00:46:57.280 --> 00:47:01.360 align:middle line:84% with these type I and type II error. 00:47:01.360 --> 00:47:06.000 align:middle line:84% And so this is just the z-test, as you recall. 00:47:06.000 --> 00:47:08.640 align:middle line:90% Here's the definition. 00:47:08.640 --> 00:47:11.400 align:middle line:84% Z is the number of standard deviations away from the mean. 00:47:11.400 --> 00:47:16.080 align:middle line:90% X-bar is your sample mean. 00:47:16.080 --> 00:47:17.960 align:middle line:90% Mu is what you expect. 00:47:17.960 --> 00:47:22.520 align:middle line:84% Sigma over root n is your standard deviation 00:47:22.520 --> 00:47:25.880 align:middle line:90% of your sample. 00:47:25.880 --> 00:47:28.240 align:middle line:84% Just some criteria-- this technically 00:47:28.240 --> 00:47:31.220 align:middle line:90% requires a normal distribution. 00:47:31.220 --> 00:47:32.720 align:middle line:84% We're using the central limit theory 00:47:32.720 --> 00:47:37.280 align:middle line:84% to basically say that as long as the sample size is large, 00:47:37.280 --> 00:47:40.120 align:middle line:84% the number of people that we're going to test is large. 00:47:40.120 --> 00:47:41.952 align:middle line:90% Over 30-ish, this will hold. 00:47:41.952 --> 00:47:43.160 align:middle line:90% We'll go back and check this. 00:47:43.160 --> 00:47:46.338 align:middle line:84% When we do the math, if we get a number less than 30 individuals, 00:47:46.338 --> 00:47:47.880 align:middle line:84% then we'll say this was a bad choice. 00:47:47.880 --> 00:47:51.720 align:middle line:90% And we'll see what it is. 00:47:51.720 --> 00:47:54.600 align:middle line:90% So let's just put in the math. 00:47:54.600 --> 00:47:57.120 align:middle line:84% For type I error, you can look this up. 00:47:57.120 --> 00:48:01.280 align:middle line:84% If you want a 95% confidence of accepting the model, 00:48:01.280 --> 00:48:06.020 align:middle line:84% or only a 5% chance of wrongly rejecting it is 1.5 sigma out. 00:48:06.020 --> 00:48:07.380 align:middle line:90% So there's your equation. 00:48:07.380 --> 00:48:10.120 align:middle line:84% There's the equation for type II error 00:48:10.120 --> 00:48:15.400 align:middle line:84% on these other distribution -0.84. 00:48:15.400 --> 00:48:18.620 align:middle line:84% And we have two equations and two unknowns. 00:48:18.620 --> 00:48:20.500 align:middle line:84% And so we do the math, and we solve it. 00:48:20.500 --> 00:48:24.200 align:middle line:84% And we see the threshold is 1.0007. 00:48:24.200 --> 00:48:31.400 align:middle line:84% Anything above this, we'll reject the linear model 00:48:31.400 --> 00:48:34.320 align:middle line:90% at 56 million people. 00:48:34.320 --> 00:48:35.900 align:middle line:84% And actually, to be more precise, 00:48:35.900 --> 00:48:39.460 align:middle line:84% 56 million people is the minimum size. 00:48:39.460 --> 00:48:45.540 align:middle line:84% If your observed incidence of cancer rate deviates from this, 00:48:45.540 --> 00:48:50.110 align:middle line:84% you will need actually potentially a larger population. 00:48:50.110 --> 00:48:55.710 align:middle line:90% So here's the issue. 00:48:55.710 --> 00:49:01.110 align:middle line:84% Everything we've done favors, underestimates the difficulty 00:49:01.110 --> 00:49:04.870 align:middle line:90% of seeing the threshold. 00:49:04.870 --> 00:49:06.990 align:middle line:84% And the question is, can you do an experiment 00:49:06.990 --> 00:49:13.030 align:middle line:84% with 56 identically humans and give them 00:49:13.030 --> 00:49:16.320 align:middle line:84% all one millisievert of radiation 00:49:16.320 --> 00:49:17.570 align:middle line:90% in order to see the threshold? 00:49:17.570 --> 00:49:21.950 align:middle line:84% That is an experiment you probably cannot practically do. 00:49:21.950 --> 00:49:23.550 align:middle line:90% That's the point. 00:49:23.550 --> 00:49:25.910 align:middle line:84% We'll never have this experiment. 00:49:25.910 --> 00:49:28.290 align:middle line:90% And this is really optimistic. 00:49:28.290 --> 00:49:31.910 align:middle line:90% 00:49:31.910 --> 00:49:36.190 align:middle line:84% If the threshold does exist, it is extremely difficult 00:49:36.190 --> 00:49:39.070 align:middle line:90% to show that it exists. 00:49:39.070 --> 00:49:43.910 align:middle line:84% So since we can't prove or disprove 00:49:43.910 --> 00:49:47.540 align:middle line:84% the existence of a threshold with any reasonable experiment, 00:49:47.540 --> 00:49:48.440 align:middle line:90% what happens? 00:49:48.440 --> 00:49:51.300 align:middle line:90% 00:49:51.300 --> 00:49:55.140 align:middle line:84% People just choose to believe what they want to believe, 00:49:55.140 --> 00:49:56.740 align:middle line:90% but they're not supposed to. 00:49:56.740 --> 00:50:00.220 align:middle line:84% They're supposed to follow Occam's razor 00:50:00.220 --> 00:50:04.420 align:middle line:84% and choose the model that has the best predictive power given 00:50:04.420 --> 00:50:05.860 align:middle line:90% what we know. 00:50:05.860 --> 00:50:09.820 align:middle line:84% So that's the story between threshold models 00:50:09.820 --> 00:50:14.100 align:middle line:84% and no-threshold models, is that the no-threshold is not 00:50:14.100 --> 00:50:19.420 align:middle line:84% defensible because it is a higher complexity model, 00:50:19.420 --> 00:50:23.128 align:middle line:84% but because you can't prove the threshold people wrong, 00:50:23.128 --> 00:50:25.420 align:middle line:84% there will be people who will insist on using threshold 00:50:25.420 --> 00:50:26.240 align:middle line:90% no matter what. 00:50:26.240 --> 00:50:28.940 align:middle line:90% 00:50:28.940 --> 00:50:29.600 align:middle line:90% All right. 00:50:29.600 --> 00:50:33.780 align:middle line:84% Now, could we actually do better than this? 00:50:33.780 --> 00:50:35.280 align:middle line:90% And the answer is yes. 00:50:35.280 --> 00:50:38.460 align:middle line:84% If we take a mechanistic perspective, 00:50:38.460 --> 00:50:42.460 align:middle line:84% we can learn something about how radiation biology works 00:50:42.460 --> 00:50:47.090 align:middle line:84% and inform us as to whether we think there might be a threshold 00:50:47.090 --> 00:50:47.890 align:middle line:90% or not. 00:50:47.890 --> 00:50:51.370 align:middle line:84% So this is basically saying, now we're 00:50:51.370 --> 00:50:58.130 align:middle line:84% going to allow this term here to be something other than unity. 00:50:58.130 --> 00:51:03.210 align:middle line:84% We're going to inject a prior into the model based 00:51:03.210 --> 00:51:06.370 align:middle line:90% on some additional information. 00:51:06.370 --> 00:51:11.010 align:middle line:84% So ready to learn how cancer happens? 00:51:11.010 --> 00:51:11.590 align:middle line:90% All right. 00:51:11.590 --> 00:51:15.390 align:middle line:84% So you have some amount of radiation coming in. 00:51:15.390 --> 00:51:18.530 align:middle line:84% And one of the things it does is it hits water 00:51:18.530 --> 00:51:23.890 align:middle line:84% because you are mostly water, as usually what happens. 00:51:23.890 --> 00:51:24.990 align:middle line:90% And so what does it do? 00:51:24.990 --> 00:51:27.890 align:middle line:84% It ionizes the water, kicks off an electron, 00:51:27.890 --> 00:51:31.490 align:middle line:90% and it produces hydronium. 00:51:31.490 --> 00:51:37.530 align:middle line:84% And eventually, you get OH hydronium and OH dot. 00:51:37.530 --> 00:51:39.570 align:middle line:90% And this is a free radical. 00:51:39.570 --> 00:51:46.680 align:middle line:84% And it goes and attacks one of the sugars on your DNA 00:51:46.680 --> 00:51:49.000 align:middle line:90% and causes a problem. 00:51:49.000 --> 00:51:51.140 align:middle line:90% So this is usually what happens. 00:51:51.140 --> 00:51:54.440 align:middle line:84% And this is happening all the time in your body, 00:51:54.440 --> 00:51:55.620 align:middle line:90% not just from radiation. 00:51:55.620 --> 00:51:58.760 align:middle line:84% Also, chemicals cause free radicals and do this, 00:51:58.760 --> 00:52:00.780 align:middle line:84% and your body has to deal with it. 00:52:00.780 --> 00:52:04.000 align:middle line:90% 00:52:04.000 --> 00:52:05.640 align:middle line:84% Another thing that you can have is 00:52:05.640 --> 00:52:09.200 align:middle line:84% that you could have the radiation, either the photon 00:52:09.200 --> 00:52:12.640 align:middle line:84% or an electron, directly interact 00:52:12.640 --> 00:52:20.280 align:middle line:84% with an atom in the chain in your DNA, the non-water atom, 00:52:20.280 --> 00:52:24.240 align:middle line:84% and cause that molecule to break apart. 00:52:24.240 --> 00:52:26.720 align:middle line:84% And this is where things get interesting. 00:52:26.720 --> 00:52:31.500 align:middle line:84% So recall that it's not really like the total energy deposit. 00:52:31.500 --> 00:52:33.560 align:middle line:84% It's how and where it's deposited. 00:52:33.560 --> 00:52:36.900 align:middle line:84% So I showed you this on Wednesday. 00:52:36.900 --> 00:52:39.600 align:middle line:84% And what I've done now is I've superimposed in approximately 00:52:39.600 --> 00:52:42.670 align:middle line:90% to scale the size of a cell. 00:52:42.670 --> 00:52:47.310 align:middle line:84% And you can see here that for low LET radiation, 00:52:47.310 --> 00:52:49.870 align:middle line:84% you can have lots of interactions 00:52:49.870 --> 00:52:52.270 align:middle line:90% on the scale of the nucleus. 00:52:52.270 --> 00:52:57.790 align:middle line:84% And so it is possible to have radiation. 00:52:57.790 --> 00:53:01.470 align:middle line:84% A single radioactive particle, a single electron or single photon 00:53:01.470 --> 00:53:04.950 align:middle line:84% enter the body and undergo photoelectric effect and Compton 00:53:04.950 --> 00:53:08.230 align:middle line:84% scattering and cause all of these subsequent changes 00:53:08.230 --> 00:53:10.830 align:middle line:90% multiple times within a cell. 00:53:10.830 --> 00:53:14.290 align:middle line:84% And this may generate multiple free radicals from water, 00:53:14.290 --> 00:53:18.470 align:middle line:84% these OH dot ions, or it may directly attack the DNA. 00:53:18.470 --> 00:53:25.830 align:middle line:84% And so the probability of getting multiple breaks 00:53:25.830 --> 00:53:31.650 align:middle line:84% to the genome goes up because of the nature of radiation, 00:53:31.650 --> 00:53:34.690 align:middle line:84% because of this low LET radiation. 00:53:34.690 --> 00:53:37.230 align:middle line:84% So this is looking at different kinds 00:53:37.230 --> 00:53:38.410 align:middle line:90% of breaks that might happen. 00:53:38.410 --> 00:53:42.460 align:middle line:84% The top one is what we get when we have that OH dot showing 00:53:42.460 --> 00:53:44.060 align:middle line:84% up just at some random place, and it 00:53:44.060 --> 00:53:45.400 align:middle line:90% breaks one side of the DNA. 00:53:45.400 --> 00:53:46.900 align:middle line:84% And then we have this process called 00:53:46.900 --> 00:53:48.580 align:middle line:90% non-homologous end joining. 00:53:48.580 --> 00:53:50.180 align:middle line:84% And it's like a little thing that 00:53:50.180 --> 00:53:53.460 align:middle line:84% comes along and detects the break and glues 00:53:53.460 --> 00:53:55.460 align:middle line:90% it back together. 00:53:55.460 --> 00:53:59.020 align:middle line:84% And it turns out we have one of these 00:53:59.020 --> 00:54:05.500 align:middle line:84% happen every half-second per cell in our body. 00:54:05.500 --> 00:54:06.875 align:middle line:90% It's happening all the time. 00:54:06.875 --> 00:54:09.000 align:middle line:84% All the cells breaking, breaking, fixing, breaking, 00:54:09.000 --> 00:54:11.020 align:middle line:84% fixing, breaking, fixing all the time. 00:54:11.020 --> 00:54:13.980 align:middle line:90% It's amazing. 00:54:13.980 --> 00:54:20.700 align:middle line:84% But this kind of break where we have a double-strand break 00:54:20.700 --> 00:54:24.380 align:middle line:90% is pretty rare. 00:54:24.380 --> 00:54:27.980 align:middle line:84% That just two happen to happen in close proximity 00:54:27.980 --> 00:54:29.060 align:middle line:90% to each other. 00:54:29.060 --> 00:54:32.500 align:middle line:84% It's something like once every-- turns out 00:54:32.500 --> 00:54:35.460 align:middle line:90% this is like every two hours. 00:54:35.460 --> 00:54:38.130 align:middle line:90% So it still happens. 00:54:38.130 --> 00:54:39.890 align:middle line:84% And you still have non-homologous end 00:54:39.890 --> 00:54:42.050 align:middle line:84% joining coming and gluing this together. 00:54:42.050 --> 00:54:46.530 align:middle line:84% But sometimes you get something like this. 00:54:46.530 --> 00:54:47.830 align:middle line:90% And what is the end? 00:54:47.830 --> 00:54:52.107 align:middle line:84% This thing floats away, and the enzyme comes along and goes, 00:54:52.107 --> 00:54:53.690 align:middle line:84% uhh. and then it cuts these things off 00:54:53.690 --> 00:54:58.250 align:middle line:84% and it glues it together and just takes out a few base pairs. 00:54:58.250 --> 00:55:00.610 align:middle line:84% And this happens all the time in your genome. 00:55:00.610 --> 00:55:06.850 align:middle line:84% And sometimes it's benign, and sometimes it's not. 00:55:06.850 --> 00:55:11.610 align:middle line:84% And when it's not, you get cancer. 00:55:11.610 --> 00:55:15.130 align:middle line:90% So here's the key takeaway. 00:55:15.130 --> 00:55:19.450 align:middle line:84% The repair process is what's giving you the cancer. 00:55:19.450 --> 00:55:23.145 align:middle line:84% If you simply break the genome and you don't repair it, 00:55:23.145 --> 00:55:24.770 align:middle line:84% then generally what happens is the cell 00:55:24.770 --> 00:55:28.450 align:middle line:84% becomes senescent and doesn't continue to operate 00:55:28.450 --> 00:55:31.770 align:middle line:90% and just dies generally. 00:55:31.770 --> 00:55:34.890 align:middle line:84% But if you go and you introduce these genomic errors 00:55:34.890 --> 00:55:36.880 align:middle line:90% from this repair process. 00:55:36.880 --> 00:55:38.680 align:middle line:90% That's when you get cancer. 00:55:38.680 --> 00:55:45.080 align:middle line:84% So sometimes you hear people who have seen the idea of hormesis, 00:55:45.080 --> 00:55:47.320 align:middle line:84% that a little bit of radiation is good for you, 00:55:47.320 --> 00:55:49.860 align:middle line:84% and they are looking for ways to justify that. 00:55:49.860 --> 00:55:54.720 align:middle line:84% Remember, most likely that is actually the look again effect 00:55:54.720 --> 00:55:56.220 align:middle line:84% that they've picked up on initially. 00:55:56.220 --> 00:55:58.060 align:middle line:84% And they say, well what would describe this? 00:55:58.060 --> 00:56:00.160 align:middle line:84% And they'll say, oh, it turns out 00:56:00.160 --> 00:56:01.940 align:middle line:84% that if you give a cell radiation, 00:56:01.940 --> 00:56:05.352 align:middle line:84% we can observe that the repair process is upregulated. 00:56:05.352 --> 00:56:06.060 align:middle line:90% Of course, it is. 00:56:06.060 --> 00:56:07.435 align:middle line:84% You've done damage to the genome. 00:56:07.435 --> 00:56:09.960 align:middle line:84% It's going to have more repair to do. 00:56:09.960 --> 00:56:12.240 align:middle line:84% But then they think oh, but maybe that repair process 00:56:12.240 --> 00:56:14.740 align:middle line:84% is preventing background cancers. 00:56:14.740 --> 00:56:17.120 align:middle line:90% It's like no. 00:56:17.120 --> 00:56:20.600 align:middle line:84% That repair process is causing the cancers. 00:56:20.600 --> 00:56:25.880 align:middle line:84% And so this is a key insight here. 00:56:25.880 --> 00:56:28.200 align:middle line:84% The other insight I want you to take away this 00:56:28.200 --> 00:56:32.360 align:middle line:84% is possible from a single photon-- 00:56:32.360 --> 00:56:33.760 align:middle line:90% comes in, it comes through. 00:56:33.760 --> 00:56:35.090 align:middle line:90% You get Compton scatters. 00:56:35.090 --> 00:56:36.010 align:middle line:90% You get an electron. 00:56:36.010 --> 00:56:38.090 align:middle line:84% You get another photon, another Compton. 00:56:38.090 --> 00:56:39.130 align:middle line:90% Two electrons. 00:56:39.130 --> 00:56:40.190 align:middle line:90% Boom, boom, boom. 00:56:40.190 --> 00:56:44.390 align:middle line:90% This is technically possible. 00:56:44.390 --> 00:56:51.190 align:middle line:84% Therefore, with some very, very, very small probability, 00:56:51.190 --> 00:56:55.150 align:middle line:84% a single photon could give you cancer-- 00:56:55.150 --> 00:56:57.172 align:middle line:84% very very, very small, not normal. 00:56:57.172 --> 00:56:58.630 align:middle line:84% Normally, it requires multiple hits 00:56:58.630 --> 00:57:01.310 align:middle line:84% to the genome and multiple places, et cetera. 00:57:01.310 --> 00:57:05.750 align:middle line:84% But if there's a very small non-zero probability, 00:57:05.750 --> 00:57:07.410 align:middle line:90% a threshold cannot exist. 00:57:07.410 --> 00:57:10.710 align:middle line:90% 00:57:10.710 --> 00:57:14.590 align:middle line:84% So this is the mechanistic argument 00:57:14.590 --> 00:57:17.070 align:middle line:84% for why we don't have a threshold model. 00:57:17.070 --> 00:57:19.200 align:middle line:90% Yes. 00:57:19.200 --> 00:57:20.950 align:middle line:84% AUDIENCE: What energy photons are required 00:57:20.950 --> 00:57:22.050 align:middle line:90% to do a damage like that? 00:57:22.050 --> 00:57:24.790 align:middle line:90% Is there a threshold? 00:57:24.790 --> 00:57:26.710 align:middle line:84% SCOTT KEMP: You need to deposit at least 00:57:26.710 --> 00:57:32.580 align:middle line:84% basically an EV in the interaction to ionize something. 00:57:32.580 --> 00:57:36.580 align:middle line:84% And typically, these photons have hundreds of hundreds 00:57:36.580 --> 00:57:38.460 align:middle line:90% of keV of energy-- 00:57:38.460 --> 00:57:39.720 align:middle line:90% so plenty of energy. 00:57:39.720 --> 00:57:40.220 align:middle line:90% Yeah. 00:57:40.220 --> 00:57:44.220 align:middle line:90% 00:57:44.220 --> 00:57:45.180 align:middle line:90% OK. 00:57:45.180 --> 00:57:49.920 align:middle line:84% So we do live in a radioactive environment. 00:57:49.920 --> 00:57:53.220 align:middle line:84% Our bodies do have these repair mechanisms. 00:57:53.220 --> 00:57:56.020 align:middle line:84% There are error-proof repair mechanisms. 00:57:56.020 --> 00:58:03.580 align:middle line:84% There's something called, sorry, homologous recombination, which 00:58:03.580 --> 00:58:06.580 align:middle line:84% is a way of error-checking the genome. 00:58:06.580 --> 00:58:12.620 align:middle line:84% But in human cells, it only exists during cell division. 00:58:12.620 --> 00:58:15.900 align:middle line:84% In the normal growth phase of the cell, 00:58:15.900 --> 00:58:17.640 align:middle line:84% we only have L1 copy of the genome, 00:58:17.640 --> 00:58:20.460 align:middle line:84% so we have nothing to fact-check it against. 00:58:20.460 --> 00:58:23.660 align:middle line:84% Certain bacteria, like Deinococcus radiodurans, 00:58:23.660 --> 00:58:25.400 align:middle line:84% has multiple copies of its genome. 00:58:25.400 --> 00:58:27.340 align:middle line:84% And every time it fixes it, it checks repair 00:58:27.340 --> 00:58:28.720 align:middle line:90% against the other copies. 00:58:28.720 --> 00:58:32.220 align:middle line:84% And so it makes it extremely robust against large doses 00:58:32.220 --> 00:58:33.280 align:middle line:90% of radiation. 00:58:33.280 --> 00:58:35.380 align:middle line:90% But human cells, only once-- 00:58:35.380 --> 00:58:37.740 align:middle line:84% only copy, except during cell division. 00:58:37.740 --> 00:58:41.320 align:middle line:84% And so we have no way of enduring 00:58:41.320 --> 00:58:43.080 align:middle line:90% those radioactive assaults. 00:58:43.080 --> 00:58:48.360 align:middle line:84% So yes, we are evolved to live in the radioactive environment. 00:58:48.360 --> 00:58:52.520 align:middle line:84% We have repair processes that keep the cells going, 00:58:52.520 --> 00:58:53.700 align:middle line:90% keep them alive. 00:58:53.700 --> 00:58:57.460 align:middle line:84% But it doesn't guarantee us that the cells don't eventually 00:58:57.460 --> 00:58:58.500 align:middle line:90% get cancer. 00:58:58.500 --> 00:59:03.980 align:middle line:84% It only guarantees that on average, most people will 00:59:03.980 --> 00:59:06.900 align:middle line:90% live long enough to procreate. 00:59:06.900 --> 00:59:10.300 align:middle line:84% And that's all evolution has given us. 00:59:10.300 --> 00:59:16.180 align:middle line:84% So little bits of radiation do matter very likely. 00:59:16.180 --> 00:59:19.340 align:middle line:84% And we shouldn't be seduced by any 00:59:19.340 --> 00:59:22.780 align:middle line:84% of the ideas that radioactive environment, 00:59:22.780 --> 00:59:25.440 align:middle line:84% and we're accustomed to this, cetera, cetera. 00:59:25.440 --> 00:59:28.860 align:middle line:90% 00:59:28.860 --> 00:59:30.410 align:middle line:90% That's it. 00:59:30.410 --> 00:59:33.330 align:middle line:84% That's why LNT is right, or very likely. 00:59:33.330 --> 00:59:34.680 align:middle line:90% Yep. 00:59:34.680 --> 00:59:36.230 align:middle line:90% AUDIENCE: Here's a question. 00:59:36.230 --> 00:59:40.730 align:middle line:84% So if the cluster double-strand breaks, 00:59:40.730 --> 00:59:44.930 align:middle line:84% if the repair mechanism happens to repair 00:59:44.930 --> 00:59:50.170 align:middle line:84% in a way that's malignant, is there 00:59:50.170 --> 00:59:55.170 align:middle line:84% a threshold for a certain amount of those dangerous repairs 00:59:55.170 --> 00:59:55.830 align:middle line:90% have to happen? 00:59:55.830 --> 00:59:58.390 align:middle line:84% Because I'm assuming if probabilistically, 00:59:58.390 --> 01:00:01.130 align:middle line:90% there's been cancerous repairs. 01:00:01.130 --> 01:00:02.850 align:middle line:90% SCOTT KEMP: Yep. 01:00:02.850 --> 01:00:03.350 align:middle line:90% Yeah. 01:00:03.350 --> 01:00:07.610 align:middle line:84% So they call them promoters or potentiators. 01:00:07.610 --> 01:00:10.070 align:middle line:84% So I believe-- and this is going from memory, 01:00:10.070 --> 01:00:11.570 align:middle line:84% so you should double-check this-- 01:00:11.570 --> 01:00:13.370 align:middle line:84% but I believe the general consensus 01:00:13.370 --> 01:00:22.170 align:middle line:84% is that you need on average about six or so of these things 01:00:22.170 --> 01:00:24.730 align:middle line:84% to get the cell really going into the cancerous mode. 01:00:24.730 --> 01:00:27.505 align:middle line:90% AUDIENCE: In a specific cell? 01:00:27.505 --> 01:00:29.140 align:middle line:84% SCOTT KEMP: In a specific genome. 01:00:29.140 --> 01:00:29.680 align:middle line:90% Yeah. 01:00:29.680 --> 01:00:32.500 align:middle line:84% That doesn't mean you actually need six. 01:00:32.500 --> 01:00:37.200 align:middle line:84% If the damage happens in the right region of the genome, 01:00:37.200 --> 01:00:40.840 align:middle line:84% you in fact can cause problems right away. 01:00:40.840 --> 01:00:42.963 align:middle line:84% You can also, if the damage happens 01:00:42.963 --> 01:00:44.380 align:middle line:84% in the right region of the genome, 01:00:44.380 --> 01:00:46.320 align:middle line:84% you can also just instantly kill the cell, 01:00:46.320 --> 01:00:47.420 align:middle line:90% which is not a problem. 01:00:47.420 --> 01:00:49.160 align:middle line:90% Cell die is fine. 01:00:49.160 --> 01:00:53.960 align:middle line:84% But on average, it does need to happen more than once. 01:00:53.960 --> 01:00:56.840 align:middle line:84% But remember, it is happening all the time in your body. 01:00:56.840 --> 01:01:01.520 align:middle line:84% So you are accumulating these potentiated cells 01:01:01.520 --> 01:01:03.560 align:middle line:84% over your lifetime, which is why you typically 01:01:03.560 --> 01:01:07.320 align:middle line:90% get cancer when you're old. 01:01:07.320 --> 01:01:10.940 align:middle line:84% And this is also why the relative risk model makes sense. 01:01:10.940 --> 01:01:14.560 align:middle line:84% The idea that a smoker has a higher probability 01:01:14.560 --> 01:01:18.460 align:middle line:84% of getting cancer from radiation makes sense, 01:01:18.460 --> 01:01:21.240 align:middle line:84% because they've already trashed their genome 01:01:21.240 --> 01:01:22.800 align:middle line:90% from cigarette smoking. 01:01:22.800 --> 01:01:26.600 align:middle line:84% So they've accumulated these potentiated genomes. 01:01:26.600 --> 01:01:29.030 align:middle line:84% So an additional damage from radiation 01:01:29.030 --> 01:01:32.550 align:middle line:84% is more likely to actually cause the cancer. 01:01:32.550 --> 01:01:37.730 align:middle line:84% So yeah, that is what's likely going on in most cells. 01:01:37.730 --> 01:01:38.230 align:middle line:90% Yeah. 01:01:38.230 --> 01:01:39.330 align:middle line:90% AUDIENCE: That's a good point. 01:01:39.330 --> 01:01:40.240 align:middle line:90% It's like you have these-- 01:01:40.240 --> 01:01:41.907 align:middle line:84% SCOTT KEMP: Just depends on where it is. 01:01:41.907 --> 01:01:44.250 align:middle line:90% It's a stochastic process. 01:01:44.250 --> 01:01:46.190 align:middle line:90% It's not a threshold phenomenon. 01:01:46.190 --> 01:01:47.870 align:middle line:90% It's stochastic. 01:01:47.870 --> 01:01:48.730 align:middle line:90% Yeah. 01:01:48.730 --> 01:01:52.150 align:middle line:84% AUDIENCE: What is the effect of how long you get the radiation 01:01:52.150 --> 01:01:53.870 align:middle line:90% dose over [INAUDIBLE]? 01:01:53.870 --> 01:01:55.630 align:middle line:90% SCOTT KEMP: The dose rate. 01:01:55.630 --> 01:02:01.932 align:middle line:84% Yeah, so there is evidence of dose-rate effects. 01:02:01.932 --> 01:02:03.390 align:middle line:84% If you want to read about that, you 01:02:03.390 --> 01:02:06.710 align:middle line:84% can go to our ICRP, which is in the reading. 01:02:06.710 --> 01:02:10.470 align:middle line:84% Generally, what the dose rate effects show is like 01:02:10.470 --> 01:02:15.670 align:middle line:84% if you increase the dose rate, then it tips over, 01:02:15.670 --> 01:02:19.190 align:middle line:84% and then it's to say that the additional units of dose 01:02:19.190 --> 01:02:22.202 align:middle line:84% at a high rate don't have as big of an effect. 01:02:22.202 --> 01:02:23.410 align:middle line:90% And that kind of makes sense. 01:02:23.410 --> 01:02:29.660 align:middle line:84% You have a lot of damages occurring, 01:02:29.660 --> 01:02:34.020 align:middle line:84% and they're all happening before the repair process happens 01:02:34.020 --> 01:02:35.220 align:middle line:90% kind of a thing. 01:02:35.220 --> 01:02:40.280 align:middle line:84% So that is the extent of my understanding of it. 01:02:40.280 --> 01:02:41.960 align:middle line:90% And you can go and look more. 01:02:41.960 --> 01:02:42.460 align:middle line:90% Yeah. 01:02:42.460 --> 01:02:47.700 align:middle line:90% 01:02:47.700 --> 01:02:48.380 align:middle line:90% Yep. 01:02:48.380 --> 01:02:50.500 align:middle line:84% AUDIENCE: Any of the nonlinear models 01:02:50.500 --> 01:02:54.580 align:middle line:84% end up kind of converging on setting those nonlinear 01:02:54.580 --> 01:02:55.900 align:middle line:90% weights to 0? 01:02:55.900 --> 01:02:58.700 align:middle line:84% I guess, do you see those models converging on a linear model 01:02:58.700 --> 01:02:59.420 align:middle line:90% still? 01:02:59.420 --> 01:03:03.460 align:middle line:84% SCOTT KEMP: The nonlinear models are 01:03:03.460 --> 01:03:10.100 align:middle line:84% defensible once you allow the high-dose regime like this. 01:03:10.100 --> 01:03:14.340 align:middle line:84% Then they do converge for very high doses. 01:03:14.340 --> 01:03:16.500 align:middle line:90% Yeah. 01:03:16.500 --> 01:03:19.880 align:middle line:84% For what we would still permit in the lower-dose regime. 01:03:19.880 --> 01:03:22.500 align:middle line:84% I think we now have enough data on leukemia 01:03:22.500 --> 01:03:28.410 align:middle line:84% to support a nonlinear model on leukemia, 01:03:28.410 --> 01:03:32.170 align:middle line:84% but they're basically all linear in the extreme low-- and so 01:03:32.170 --> 01:03:34.610 align:middle line:90% in the limit of ultra-low doses. 01:03:34.610 --> 01:03:36.030 align:middle line:90% They all kind go like that. 01:03:36.030 --> 01:03:39.810 align:middle line:90% 01:03:39.810 --> 01:03:42.450 align:middle line:90% Yep. 01:03:42.450 --> 01:03:44.430 align:middle line:84% AUDIENCE: [INAUDIBLE] why would it be? 01:03:44.430 --> 01:03:47.650 align:middle line:84% You said earlier that people in industry 01:03:47.650 --> 01:03:49.090 align:middle line:90% would want this to be true. 01:03:49.090 --> 01:03:50.870 align:middle line:84% Why is there such a small effect? 01:03:50.870 --> 01:03:51.785 align:middle line:90% Why would it be-- 01:03:51.785 --> 01:03:52.410 align:middle line:90% SCOTT KEMP: Oh. 01:03:52.410 --> 01:03:55.890 align:middle line:84% Well, maybe we can have an answer from the audience. 01:03:55.890 --> 01:03:57.810 align:middle line:84% AUDIENCE: I mean, I'm not really trying to-- 01:03:57.810 --> 01:04:00.510 align:middle line:84% I haven't really looked into threshold models or not. 01:04:00.510 --> 01:04:03.090 align:middle line:84% Just what would the cost difference be? 01:04:03.090 --> 01:04:10.330 align:middle line:90% And I'm going to tell you. 01:04:10.330 --> 01:04:11.550 align:middle line:90% It's going to be small. 01:04:11.550 --> 01:04:14.070 align:middle line:90% 01:04:14.070 --> 01:04:18.710 align:middle line:84% The purpose of the research is to say that for nuclear power, 01:04:18.710 --> 01:04:20.130 align:middle line:90% it does not matter. 01:04:20.130 --> 01:04:23.840 align:middle line:84% Just focus on why the containment is so damn expensive 01:04:23.840 --> 01:04:25.760 align:middle line:90% and all these things. 01:04:25.760 --> 01:04:28.480 align:middle line:90% That's essentially what I do. 01:04:28.480 --> 01:04:32.020 align:middle line:84% I'm not looking into why threshold. 01:04:32.020 --> 01:04:33.060 align:middle line:90% Why not threshold? 01:04:33.060 --> 01:04:35.480 align:middle line:84% Obviously, I've done my [INAUDIBLE] review. 01:04:35.480 --> 01:04:38.480 align:middle line:90% But it's [INAUDIBLE] 01:04:38.480 --> 01:04:42.140 align:middle line:84% SCOTT KEMP: But what's motivating your study? 01:04:42.140 --> 01:04:45.180 align:middle line:84% So you believe that it's likely to be or have a small impact? 01:04:45.180 --> 01:04:47.200 align:middle line:84% Bongiorno believes likely to have a small impact 01:04:47.200 --> 01:04:48.900 align:middle line:90% on the total cost of reactors. 01:04:48.900 --> 01:04:50.783 align:middle line:90% I agree with that. 01:04:50.783 --> 01:04:52.200 align:middle line:84% But there are people out there who 01:04:52.200 --> 01:04:54.920 align:middle line:84% don't believe that, which is why you're doing 01:04:54.920 --> 01:04:56.420 align:middle line:90% the study in the first place. 01:04:56.420 --> 01:04:57.420 align:middle line:90% AUDIENCE: Yeah, exactly. 01:04:57.420 --> 01:05:00.820 align:middle line:84% SCOTT KEMP: Yeah, and so, just to drive the point home, 01:05:00.820 --> 01:05:04.600 align:middle line:84% in May of this year, the White House issued Executive Order 01:05:04.600 --> 01:05:09.040 align:middle line:84% 14300, which instructed the Nuclear Regulatory Commission 01:05:09.040 --> 01:05:14.120 align:middle line:84% to quote, "specifically consider adopting a radiation limit," 01:05:14.120 --> 01:05:18.400 align:middle line:84% which is, again, the politicization of our regulation 01:05:18.400 --> 01:05:20.780 align:middle line:90% with nonscientific views. 01:05:20.780 --> 01:05:25.670 align:middle line:84% There are people out there pushing hard for this 01:05:25.670 --> 01:05:30.150 align:middle line:84% because they believe, without evidence, that this 01:05:30.150 --> 01:05:33.750 align:middle line:84% is going to make it cheap to build nuclear power. 01:05:33.750 --> 01:05:37.510 align:middle line:90% So yeah. 01:05:37.510 --> 01:05:39.210 align:middle line:84% AUDIENCE: A more general question. 01:05:39.210 --> 01:05:45.110 align:middle line:84% So when you're not familiar with the subject-- so for example, 01:05:45.110 --> 01:05:47.690 align:middle line:84% this is going back to the look-again effect. 01:05:47.690 --> 01:05:54.190 align:middle line:84% Yeah, so using the example of the jelly beans or whatever. 01:05:54.190 --> 01:05:57.390 align:middle line:90% Say I was familiar. 01:05:57.390 --> 01:06:01.050 align:middle line:84% How do you tune your radar to be aware of those types of things, 01:06:01.050 --> 01:06:03.190 align:middle line:84% besides knowing the subject and being like, 01:06:03.190 --> 01:06:05.970 align:middle line:84% that's very contrary to all the existing evidence? 01:06:05.970 --> 01:06:10.705 align:middle line:90% 01:06:10.705 --> 01:06:12.330 align:middle line:84% SCOTT KEMP: How do you tune your radar? 01:06:12.330 --> 01:06:16.190 align:middle line:90% 01:06:16.190 --> 01:06:17.690 align:middle line:90% I'm not sure. 01:06:17.690 --> 01:06:20.787 align:middle line:90% 01:06:20.787 --> 01:06:22.620 align:middle line:84% Are you asking the question, if you had only 01:06:22.620 --> 01:06:25.300 align:middle line:84% looked at the green jelly bean scenario? 01:06:25.300 --> 01:06:28.215 align:middle line:84% AUDIENCE: I think I mean, as a reader who 01:06:28.215 --> 01:06:29.340 align:middle line:90% looks at various articles-- 01:06:29.340 --> 01:06:31.882 align:middle line:84% SCOTT KEMP: Oh, how do they're not doing the green jelly bean 01:06:31.882 --> 01:06:32.580 align:middle line:90% scenario? 01:06:32.580 --> 01:06:33.260 align:middle line:90% Yeah. 01:06:33.260 --> 01:06:34.380 align:middle line:90% Yeah. 01:06:34.380 --> 01:06:37.300 align:middle line:90% You don't often. 01:06:37.300 --> 01:06:38.060 align:middle line:90% Yeah. 01:06:38.060 --> 01:06:41.540 align:middle line:84% And if you go read the medical literature, 01:06:41.540 --> 01:06:43.900 align:middle line:84% you will find all kinds of articles where you're like? 01:06:43.900 --> 01:06:45.200 align:middle line:90% Why did you look at this? 01:06:45.200 --> 01:06:46.940 align:middle line:90% Why did you study this? 01:06:46.940 --> 01:06:50.860 align:middle line:84% And I think a lot of that is oh, look. 01:06:50.860 --> 01:06:54.020 align:middle line:84% And then they write a paper about it. 01:06:54.020 --> 01:06:56.940 align:middle line:84% So that's why you, you want to avoid 01:06:56.940 --> 01:07:01.420 align:middle line:84% what I call single-study syndrome, where you just 01:07:01.420 --> 01:07:04.120 align:middle line:84% find one study that supports something. 01:07:04.120 --> 01:07:08.360 align:middle line:84% If that study has not been replicated, you shouldn't do it. 01:07:08.360 --> 01:07:10.860 align:middle line:84% And so we try to teach people not to do this. 01:07:10.860 --> 01:07:13.700 align:middle line:84% One of the things is something called 01:07:13.700 --> 01:07:18.715 align:middle line:84% hypothesizing after the fact, which is to say, 01:07:18.715 --> 01:07:19.590 align:middle line:90% oh, I saw this thing. 01:07:19.590 --> 01:07:20.350 align:middle line:90% I wonder why it happened. 01:07:20.350 --> 01:07:21.350 align:middle line:90% Oh, maybe this is the reason. 01:07:21.350 --> 01:07:22.310 align:middle line:90% This is my hypothesis. 01:07:22.310 --> 01:07:24.050 align:middle line:90% Here's the data, publish. 01:07:24.050 --> 01:07:26.690 align:middle line:90% You're not supposed to do that. 01:07:26.690 --> 01:07:28.530 align:middle line:84% What you're doing is you're exacerbating 01:07:28.530 --> 01:07:29.750 align:middle line:90% the look-again effect. 01:07:29.750 --> 01:07:31.490 align:middle line:84% You're supposed to have the hypothesis do 01:07:31.490 --> 01:07:33.490 align:middle line:90% the test of your hypothesis. 01:07:33.490 --> 01:07:36.490 align:middle line:90% And then if it fails, it fails. 01:07:36.490 --> 01:07:40.090 align:middle line:84% But unfortunately, this is a problem 01:07:40.090 --> 01:07:44.150 align:middle line:90% with how we promote people. 01:07:44.150 --> 01:07:46.650 align:middle line:84% It's a social problem because we say 01:07:46.650 --> 01:07:51.690 align:middle line:84% you need publications to show your impact as a scientist. 01:07:51.690 --> 01:07:54.090 align:middle line:84% And therefore people are incentivized to have 01:07:54.090 --> 01:07:57.690 align:middle line:84% bad behavior and to publish everything they can 01:07:57.690 --> 01:07:59.610 align:middle line:90% when it ought not be published. 01:07:59.610 --> 01:08:02.050 align:middle line:84% And then journals also don't want 01:08:02.050 --> 01:08:05.130 align:middle line:84% to publish null results because they're not exciting. 01:08:05.130 --> 01:08:06.970 align:middle line:84% And so there are lots of experiments 01:08:06.970 --> 01:08:11.450 align:middle line:84% that are done that are never reported in the literature. 01:08:11.450 --> 01:08:14.570 align:middle line:84% And all kinds of people are coming up 01:08:14.570 --> 01:08:16.840 align:middle line:84% with the same hypothesis, doing the same experiment, 01:08:16.840 --> 01:08:18.720 align:middle line:84% getting the same null result, and not knowing 01:08:18.720 --> 01:08:21.479 align:middle line:84% that the repeating people's experiments because we 01:08:21.479 --> 01:08:23.359 align:middle line:90% don't publish null results. 01:08:23.359 --> 01:08:25.220 align:middle line:90% It's a real big problem. 01:08:25.220 --> 01:08:25.720 align:middle line:90% Yeah. 01:08:25.720 --> 01:08:26.220 align:middle line:90% Yeah. 01:08:26.220 --> 01:08:27.840 align:middle line:84% AUDIENCE: Follow-up question of mine-- 01:08:27.840 --> 01:08:31.600 align:middle line:84% since an important part of the scientific process 01:08:31.600 --> 01:08:35.240 align:middle line:90% is repeatability. 01:08:35.240 --> 01:08:38.200 align:middle line:84% This is also coming from a lack of knowledge. 01:08:38.200 --> 01:08:40.160 align:middle line:84% Is the system set up so that it is 01:08:40.160 --> 01:08:43.240 align:middle line:84% hard to get funding to just prove or disprove someone else's 01:08:43.240 --> 01:08:45.279 align:middle line:90% stuff, because someone would? 01:08:45.279 --> 01:08:46.772 align:middle line:84% We want you to make breakthroughs. 01:08:46.772 --> 01:08:47.939 align:middle line:90% SCOTT KEMP: Something novel. 01:08:47.939 --> 01:08:49.200 align:middle line:90% Yeah, sure. 01:08:49.200 --> 01:08:50.899 align:middle line:90% Absolutely, it's really hard. 01:08:50.899 --> 01:08:53.760 align:middle line:84% It would be really hard to justify a replication study 01:08:53.760 --> 01:08:56.779 align:middle line:84% unless the initial result was, for some reason, 01:08:56.779 --> 01:09:00.220 align:middle line:84% very controversial or very important. 01:09:00.220 --> 01:09:03.399 align:middle line:84% Someone say, oh, it's already been done. 01:09:03.399 --> 01:09:07.000 align:middle line:84% AUDIENCE: So does that make it such that single studies set 01:09:07.000 --> 01:09:12.319 align:middle line:84% the tone of the body of literature, when in reality, it 01:09:12.319 --> 01:09:14.140 align:middle line:84% should be, a collection of studies? 01:09:14.140 --> 01:09:15.740 align:middle line:84% SCOTT KEMP: Yes, they absolutely do. 01:09:15.740 --> 01:09:20.080 align:middle line:84% And I mean, in the best of all worlds, 01:09:20.080 --> 01:09:24.140 align:middle line:84% the first study changes the direction of research, 01:09:24.140 --> 01:09:25.720 align:middle line:90% but research goes and checks. 01:09:25.720 --> 01:09:28.800 align:middle line:90% 01:09:28.800 --> 01:09:32.800 align:middle line:84% It is the case that sometimes we go down the wrong path, 01:09:32.800 --> 01:09:35.040 align:middle line:90% sometimes for a very long time. 01:09:35.040 --> 01:09:38.220 align:middle line:84% So I mean, maybe the classic example, 01:09:38.220 --> 01:09:40.840 align:middle line:84% which is not really relevant to how modern research is done 01:09:40.840 --> 01:09:45.160 align:middle line:84% and published, but would be Ptolemaic epicycles 01:09:45.160 --> 01:09:52.279 align:middle line:84% to explain how the sun and the planets orbit the Earth. 01:09:52.279 --> 01:09:58.720 align:middle line:84% And it took hundreds of years before that view was refuted 01:09:58.720 --> 01:10:02.220 align:middle line:90% because of the firmness. 01:10:02.220 --> 01:10:05.000 align:middle line:84% There are other phenomena which I 01:10:05.000 --> 01:10:07.320 align:middle line:84% can't bring to mind, which have existed 01:10:07.320 --> 01:10:11.460 align:middle line:90% for various periods of time. 01:10:11.460 --> 01:10:14.710 align:middle line:90% Does anyone remember the-- 01:10:14.710 --> 01:10:20.190 align:middle line:84% during the COVID pandemic, there was a big push for-- 01:10:20.190 --> 01:10:23.202 align:middle line:84% what is the name of the anti-malarial drug? 01:10:23.202 --> 01:10:24.410 align:middle line:90% AUDIENCE: Hydroxychloroquine? 01:10:24.410 --> 01:10:25.702 align:middle line:90% SCOTT KEMP: Hydroxychloroquine. 01:10:25.702 --> 01:10:28.190 align:middle line:84% Do you remember hydroxychloroquine? 01:10:28.190 --> 01:10:34.990 align:middle line:84% So that didn't just come from some wacky person. 01:10:34.990 --> 01:10:38.070 align:middle line:84% I was about to Google it, but you got it. 01:10:38.070 --> 01:10:41.950 align:middle line:84% That actually came from a scientific study published 01:10:41.950 --> 01:10:44.950 align:middle line:90% in Nature, top-quality journal. 01:10:44.950 --> 01:10:47.315 align:middle line:84% And what they did is they found-- 01:10:47.315 --> 01:10:49.190 align:middle line:84% people who want to go and don't want to learn 01:10:49.190 --> 01:10:53.830 align:middle line:90% about rate biology, you can go. 01:10:53.830 --> 01:10:59.910 align:middle line:84% So what they did is they used Vero cells, which 01:10:59.910 --> 01:11:01.470 align:middle line:90% are derived from liver. 01:11:01.470 --> 01:11:05.870 align:middle line:84% And it's a standard test cell to see if different drugs would 01:11:05.870 --> 01:11:07.550 align:middle line:90% prevent COVID infection. 01:11:07.550 --> 01:11:13.700 align:middle line:84% So they put the cell inside a solution of hydroxychloroquine. 01:11:13.700 --> 01:11:19.620 align:middle line:84% They infect the supernatant with little virus of COVID. 01:11:19.620 --> 01:11:25.420 align:middle line:84% And lo and behold, it reduced the probability of infection. 01:11:25.420 --> 01:11:28.940 align:middle line:84% So everyone assumed that this would work. 01:11:28.940 --> 01:11:33.700 align:middle line:84% The thing is that it turns out that viruses can infect cells 01:11:33.700 --> 01:11:35.380 align:middle line:90% in a couple of different ways. 01:11:35.380 --> 01:11:37.380 align:middle line:84% So one of the classic ways is the virus lands 01:11:37.380 --> 01:11:41.820 align:middle line:84% on the top of the cell, like here. 01:11:41.820 --> 01:11:45.460 align:middle line:84% And the cell detects something is wrong, 01:11:45.460 --> 01:11:48.900 align:middle line:84% and it endocytosis, creates a vesicle, 01:11:48.900 --> 01:11:51.460 align:middle line:90% brings the virus inside. 01:11:51.460 --> 01:11:55.235 align:middle line:84% And then what hydroxychloroquine does is it acidifies. 01:11:55.235 --> 01:11:56.360 align:middle line:90% This is called an endosome. 01:11:56.360 --> 01:12:00.460 align:middle line:84% And it acidifies this endosome and destroys the virus. 01:12:00.460 --> 01:12:04.300 align:middle line:84% And so it prevents the viral, which is being endocytosed 01:12:04.300 --> 01:12:08.540 align:middle line:84% is the term, from infecting the cell. 01:12:08.540 --> 01:12:12.490 align:middle line:84% So it does indeed prevent infection via this mechanism. 01:12:12.490 --> 01:12:15.490 align:middle line:84% But this mechanism is the dominant mechanism 01:12:15.490 --> 01:12:17.490 align:middle line:84% in liver cells, but not the dominant mechanism 01:12:17.490 --> 01:12:20.410 align:middle line:84% in your respiratory tract, which happens to express 01:12:20.410 --> 01:12:23.650 align:middle line:90% a protein here called ACE2. 01:12:23.650 --> 01:12:29.130 align:middle line:84% And most of the viruses land on ACE2 and get in this way. 01:12:29.130 --> 01:12:31.050 align:middle line:84% And so hydroxychloroquine does nothing 01:12:31.050 --> 01:12:34.010 align:middle line:90% to prevent infection via ACE2. 01:12:34.010 --> 01:12:37.290 align:middle line:84% So it was good science, but they forgot 01:12:37.290 --> 01:12:38.970 align:middle line:84% that this cell type did not express 01:12:38.970 --> 01:12:43.810 align:middle line:84% this protein, which is expressed in your airway epithelial. 01:12:43.810 --> 01:12:46.150 align:middle line:90% And so the result was wrong. 01:12:46.150 --> 01:12:48.570 align:middle line:84% But you will not believe how hard 01:12:48.570 --> 01:12:52.850 align:middle line:84% it was to get hydroxychloroquine for a large number of months 01:12:52.850 --> 01:12:56.230 align:middle line:84% because of this one paper, which had never been validated. 01:12:56.230 --> 01:12:58.510 align:middle line:84% And that wasn't even the look-again effect. 01:12:58.510 --> 01:13:01.210 align:middle line:84% It was just that they used the wrong cell type. 01:13:01.210 --> 01:13:05.330 align:middle line:84% So it was good science, but it wasn't complete. 01:13:05.330 --> 01:13:06.410 align:middle line:90% So yeah. 01:13:06.410 --> 01:13:08.890 align:middle line:84% AUDIENCE: Continuing unless someone else has a question. 01:13:08.890 --> 01:13:10.307 align:middle line:84% AUDIENCE: I was just going to say, 01:13:10.307 --> 01:13:15.160 align:middle line:84% I feel like there'd be a need for a scientific magazine called 01:13:15.160 --> 01:13:19.060 align:middle line:84% Null, Not Void, which is just null results. 01:13:19.060 --> 01:13:23.000 align:middle line:84% SCOTT KEMP: I think there is a journal of null results. 01:13:23.000 --> 01:13:25.880 align:middle line:84% But I don't think most people want to bother taking the time 01:13:25.880 --> 01:13:28.440 align:middle line:90% to write up their work. 01:13:28.440 --> 01:13:31.560 align:middle line:90% So yeah. 01:13:31.560 --> 01:13:34.060 align:middle line:90% Anyway, this is a problem. 01:13:34.060 --> 01:13:36.400 align:middle line:84% We'll talk more about this in the final class. 01:13:36.400 --> 01:13:38.920 align:middle line:84% No other questions about radiation biology? 01:13:38.920 --> 01:13:41.680 align:middle line:90% All right, good. 01:13:41.680 --> 01:13:47.560 align:middle line:84% So on Wednesday, we will calculate how many people die 01:13:47.560 --> 01:13:49.640 align:middle line:90% from nuclear reactor accidents. 01:13:49.640 --> 01:13:53.120 align:middle line:84% And that will tell us something about whether nuclear power is 01:13:53.120 --> 01:13:53.980 align:middle line:90% safe or not. 01:13:53.980 --> 01:13:55.980 align:middle line:84% And we'll use the linear no-threshold model, 01:13:55.980 --> 01:14:00.430 align:middle line:84% which I hope you will believe now that I've defended it. 01:14:00.430 --> 01:14:09.000 align:middle line:90%