I would not stone them, they tried, it doesn't work well, they can fix it or scrap it. At least they tried to use their technology for something positive this time.
It's one thousand times better to fail when trying to predict the flu epidemics than successfully using technology to bomb weddings.
>It's one thousand times better to fail when trying to predict the flu epidemics than successfully using technology to bomb weddings.
That's a pointless and irrelevant comparison.
There's a cost to false positives, e.g. creating more vaccinations than necessary; fortunately I don't think the CDC or any other government organizations made any policy decisions based on Google Flu. As a person who works with a lot of data, I think applications like Google Flu have a lot of potential to do a lot of good. But it sounds like Google wasn't properly rigorous in their experiments if Google Flu is such a failure.
> When 80-90% of people visiting the doctor for “flu” don’t really have it, you can hardly expect their internet searches to be a reliable source of information.
I just attended a talk that graphed the rate of physicians' diagnoses versus the truth, and the conclusion was that about half of physicians drastically overestimate, too (and it's far more than double, closer to five times more!).
So when 80-90% of people visiting the doctor for flu don't really have it and those doctors often make type II errors, you can hardly expect their diagnoses to be a reliable source of information.
flu is like the communist, everybody is afraid of it, everybody believe they have seen one, but upon serious examination we can't find one.
I guess the flu epidemic is just a conspiration created to distract us from the fact that nobody walked on the moon or something.
On a more serious note, what's exactly the deal with misdiagnosis? If the real thing is another virus, it could also be a virus that we should treat in a similar manner? Or the doctors are loosing precious time often with a wrong first diagnosis in a significant number of cases?
Diagnosing viral illness is a pain. There are a huge number of winter circulating viruses with roughly the same pattern as the flu, and diagnostic tests for influenza are often very specific but not very sensitive. Meaning if they come up positive, you have it, but if they come up negative...shrug
Beyond that, no, it's not likely to be a virus treated in a similar manner. If by manner you mean "drugs". Oseltamivir (aka Tamiflu) only works against influenza, so if you have RSV or some other virus, it won't do you much good. The other way to handle viral diseases is to let them run their course and manage the complications, at which point a strictly accurate diagnosis isn't particularly necessary.
TL;DR: It's very hard, fairly expensive, and the payoff per-patient isn't strong.
Is it possible the CDC estimates are off by half? Two main points come to mind:
* How many Type II errors do doctors commit?
* How many people with the flu do not visit a doctor? People with good health coverage are probably more likely to visit the doctor preemptively when they are feeling unwell, potentially skewing the results.
Also, while google may be off in the exact measurement when compared to the CDC, it looks like the shape of the graph correlates with the CDC data. Overall, this looks approach promising.
The problem with "the shape of the graph" correlating is that when modeling a seasonal disease, this is trivially easy. Public health decisions are made on the magnitude of some of those measurements, not their rough shape, and so being off on the exact measurement is a pretty big problem.
This is interesting and a much needed discussion, however it too completely ignores a key difference:
CDC tracks hospitalisation based on lab reports,
GFT tracks search terms
The assumption, which admittedly Google is making, is that these two are correlated by the same amount each year. However, in periods of economic downturn without free health care, people are going to search more and seek actual medial assistance less. Given the global economy has been shot for the last six years....
> CDC tracks hospitalisation based on lab reports, GFT tracks search terms
This is the key difference, and after 5 years it is clear that this is not a valid way to predict the flu, or even monitor it. I think it was a great experiment, and thought it was pretty creative when announced, and now it should either be retired or replaced with a new hypothesis. Because as the Mythbusters would say, this myth is busted.
I disagree, but it depends what you mean by "the flu". If you mean hospitalisations, then yes, you're entirely right. If you mean lost productivity and population health outside of hospital, however, perhaps we need a better means to track this other than indirectly via either current stat.
I didn't know that, thank you! I went digging to try and find exactly what figure Google were using, and exactly how CDC came up with it, but I couldn't manage to pin them down...
One article author says the idea behind GFT has merit:
I think, actually, the core idea is a terrific one.
From [1]. [1] also has professor and doctor state that her models improved after adding GFT data. Data from search behavior does not have to replace CDC data. It can improve it. No myth.
This article argues that the system overestimated the cases (which it did) and therefore did not work. To reproduce the results of traditional systems is not the purpose of systems like GFT. If you look at the graphs you can see that GFT provides an indication of an increase in flu cases before the traditional systems. This is where it has value; it provides an early indication of an emerging problem.
To reproduce the results of traditional systems is exactly the purpose of Google Flu trends. This is evident when you go back and read the original paper.
The question has always been "Can we match CDC estimates faster and with different information?"
Mean absolute error (MAE) during the out-of-sample period is 0.486 for GFT, 0.311 for lagged CDC, and 0.232 for combined GFT and CDC.
A combination of GFT search behavior and actual CDC numbers produced the best model for the authors of this paper. The (media) spin is "GFT as a stand-alone flu tracker fails".
Perhaps the article would be of a different tone if the researchers had full access to Google's data. I don't think the mentions of irony, failure, misleading and hubris are deserved nor offer a fair portrayal.
One issue is that the number of historical data points used in the model is really very very small. The author's point out that the model initially was fit to 1,152 CDC data points.
However, these 1,152 data points are highly temporally correlated (the value at time t is a strong predictor of the value at time t+1). As such, they aren't "worth" very much when building a model. Therefore, the effectively independent number of points (what you generally need to build a good model) is in reality much smaller.
Looking at the graphs, it appears that for each flu season there is a pretty regular spike of flu incidences. This spike could probably be summarized effectively by four parameters:
- Mean
- Standard Deviation
- Skewness
- Kurtosis
Looking at it this way, the number of "independent" data points of a flu season is not 365 (one for every day the year), 52 (one for every week), or even 12 (one for every month). Instead it is only 4.
Thus the total amount of data they had to build their model was 4*(# years). When you have such small data sets, you can't expect to obtain good results.
TL;DR. Time series are very hard to use in predictive models. It is often impossible to generate good predictive models based off them.
The most vexing thing about the research I did on influenza was that my yearly estimates meant despite having a staggering amount of data, my functional N for many experiments was ~ 35, and that was because I reached way into the past.
This seems like a case of circular reasoning. "Google's results are not as good as the CDC's because Google's results are different than the CDC's." But why do we believe that the CDC estimates are better?
I have never gone to the doctor when I had a flu. If it had lasted more than one or two days, or if I had a temperature higher than 102, then I would have gone. But as it was, nobody ever knew that I had the flu except myself and my family.
Even if I had wanted to go to the doctor for every time I got a flu or a cold, it takes at least a day or two to schedule a doctor's appointment. By the time the appointment had rolled around I would probably have cancelled, since the flu doesn't usually last very long.
This being the US, there are also a lot of uninsured people who can't go to the doctor even if they wanted to. For those people, Google is the only option. The CDC can collect as much data as it wants from the laboratories, but you can't collect what isn't there.
When you add in the uninsured people and the people who get better before visiting the doctor, it's not at all surprising that the lab estimates are 2x lower than reality. In fact, I would kind of expect them to be even lower than that.
Sure, it's great to compare the Google model with the CDC's. But if there are persistent differences, does it make sense to immediately assume that the Google model is worse at predicting reality than the CDC? Saying that CDC is right "by definition" is not very helpful.
[edit: For example, the author of the paper could have compared Google and CDC versus a third data source, like sales of cold medicine. That would have been an interesting comparison, and might have shed some light on which estimates were a better reflection of reality.]
The CDC's definition is pretty much a gold standard.
Things like "sales of cold medicine" and the like fall under the heading of "syndromic surveillance", and believe me, they've been tried for the better part of a decade without yielding much that's impressive. Which is frustrating, because theoretically they should be amazing, but they just aren't, especially when asked to predict rather than back fit.
Things like the sales of cold medicine have been used before, and generally speaking they're pretty poor at detecting anything. They can occasionally work well if you know something happened already, and are looking to see when it started, but as any sort of surveillance system, they haven't panned out despite some very intensive research.
The CDC numbers (which are part of ILINet) are pretty decent, and one of the things they do, which hasn't been mentioned here, is they're not trying to find all cases of influenza. They have a network of physicians in all 50 states reporting "Influenza-like Illness". These people then also send samples in for laboratory confirmation of virus type, drug resistance, etc. The numbers are also backwards revised as more data comes in, meaning they improve over the course of the year.
It's a very good system for getting an idea of how "bad" the flu season is - the main problem is that it's slow. The ideal, and what people were hoping from Google Flu Trends (and other systems like it, it's not alone) is "ILINet, but in real time".
The CDC's model "fails pretty spectacularly" when it comes to capturing any of the times I've had the flu. I have had the flu several times and the CDC has never been aware of it. As I explained earlier, I never told my doctor that I had the flu or went to a lab. My local health department was never aware of these cases either. It might, however, have appeared on Google flu trends, since since I might have searched for what the impact of a certain level of fever was. It's been a few years so I don't remember if I did or not. And of course, Google has never published which search keywords they use.
There are three datasets here: the CDC's, Google's, and reality. Some folks here seem unable to differentiate between the CDC data and reality. But in fact, I really did have the flu, even though CDC didn't think I did. Reality is what matters.
Maybe there is reason to believe that CDC's data is closer to reality than Google's. If so, let's hear it.
The key is, like all population studies, the CDC doesn't have to know about you. It has to know about some people like you, and honestly, if you were well enough to not need to go to the doctor, it's also not really the influenza public health authorities are worried about.
The CDC is unapologetic about their data being estimates - but they're very solid estimates, and match more intensive but smaller scale studies pretty well. But we don't need to capture every single flu case - no surveillance system will ever do that, nor need to.
I just feel like either I am missing something, or there is a lack of rigor here. Clearly, sampling introduces errors and under-reporting introduces errors. What I am looking for is a case (based on data, not hand-waving) that Google's errors are worse than the CDC's, and so far I'm just not seeing it. Maybe I am missing something which is obvious to people in the field.
Edit: I see how the CDC's numbers could be a lower bound, but not upper.
I have never gone to the doctor when I had a flu. If it had lasted more than one or two days, or if I had a temperature higher than 102, then I would have gone. But as it was, nobody ever knew that I had the flu except myself and my family.
Then perhaps you didn't have strictly the flu, but some other viral upper respiratory tract infection instead.
It is hard to judge from symptoms alone (which is why lab confirmations are required for diagnosis), but as a general rule of thumb influenza A infections in adults tend to be much more severe and long lasting than what you describe [1], with distinctively higher malaise, fever and incapacitation than your regular winter cold. It is so debilitating that you probably would have gone to the doctor in any case, I would guess.
35 comments
[ 64.6 ms ] story [ 317 ms ] threadGist of it is, Google Trends overestimated flu cases by about double for 100 out of 108 weeks, according to the paper.
Edit: There is actually an interesting podcast available on it without a paywall: http://podcasts.aaas.org/science_podcast/SciencePodcast_1403...
Well, there's a simple solution to that... Just halve all of the results from the current algorithm!
That's a pointless and irrelevant comparison.
There's a cost to false positives, e.g. creating more vaccinations than necessary; fortunately I don't think the CDC or any other government organizations made any policy decisions based on Google Flu. As a person who works with a lot of data, I think applications like Google Flu have a lot of potential to do a lot of good. But it sounds like Google wasn't properly rigorous in their experiments if Google Flu is such a failure.
/newest can be so random.
I just attended a talk that graphed the rate of physicians' diagnoses versus the truth, and the conclusion was that about half of physicians drastically overestimate, too (and it's far more than double, closer to five times more!).
So when 80-90% of people visiting the doctor for flu don't really have it and those doctors often make type II errors, you can hardly expect their diagnoses to be a reliable source of information.
I guess the flu epidemic is just a conspiration created to distract us from the fact that nobody walked on the moon or something.
On a more serious note, what's exactly the deal with misdiagnosis? If the real thing is another virus, it could also be a virus that we should treat in a similar manner? Or the doctors are loosing precious time often with a wrong first diagnosis in a significant number of cases?
Beyond that, no, it's not likely to be a virus treated in a similar manner. If by manner you mean "drugs". Oseltamivir (aka Tamiflu) only works against influenza, so if you have RSV or some other virus, it won't do you much good. The other way to handle viral diseases is to let them run their course and manage the complications, at which point a strictly accurate diagnosis isn't particularly necessary.
TL;DR: It's very hard, fairly expensive, and the payoff per-patient isn't strong.
Is it possible the CDC estimates are off by half? Two main points come to mind:
* How many Type II errors do doctors commit?
* How many people with the flu do not visit a doctor? People with good health coverage are probably more likely to visit the doctor preemptively when they are feeling unwell, potentially skewing the results.
Also, while google may be off in the exact measurement when compared to the CDC, it looks like the shape of the graph correlates with the CDC data. Overall, this looks approach promising.
CDC tracks hospitalisation based on lab reports, GFT tracks search terms
The assumption, which admittedly Google is making, is that these two are correlated by the same amount each year. However, in periods of economic downturn without free health care, people are going to search more and seek actual medial assistance less. Given the global economy has been shot for the last six years....
I think, actually, the core idea is a terrific one.
From [1]. [1] also has professor and doctor state that her models improved after adding GFT data. Data from search behavior does not have to replace CDC data. It can improve it. No myth.
[1] http://www.npr.org/blogs/health/2014/03/13/289802934/googles...
The question has always been "Can we match CDC estimates faster and with different information?"
The emerging answer is: "No".
A combination of GFT search behavior and actual CDC numbers produced the best model for the authors of this paper. The (media) spin is "GFT as a stand-alone flu tracker fails".
Perhaps the article would be of a different tone if the researchers had full access to Google's data. I don't think the mentions of irony, failure, misleading and hubris are deserved nor offer a fair portrayal.
However, these 1,152 data points are highly temporally correlated (the value at time t is a strong predictor of the value at time t+1). As such, they aren't "worth" very much when building a model. Therefore, the effectively independent number of points (what you generally need to build a good model) is in reality much smaller.
Looking at the graphs, it appears that for each flu season there is a pretty regular spike of flu incidences. This spike could probably be summarized effectively by four parameters:
- Mean
- Standard Deviation
- Skewness
- Kurtosis
Looking at it this way, the number of "independent" data points of a flu season is not 365 (one for every day the year), 52 (one for every week), or even 12 (one for every month). Instead it is only 4.
Thus the total amount of data they had to build their model was 4*(# years). When you have such small data sets, you can't expect to obtain good results.
TL;DR. Time series are very hard to use in predictive models. It is often impossible to generate good predictive models based off them.
I have never gone to the doctor when I had a flu. If it had lasted more than one or two days, or if I had a temperature higher than 102, then I would have gone. But as it was, nobody ever knew that I had the flu except myself and my family.
Even if I had wanted to go to the doctor for every time I got a flu or a cold, it takes at least a day or two to schedule a doctor's appointment. By the time the appointment had rolled around I would probably have cancelled, since the flu doesn't usually last very long.
This being the US, there are also a lot of uninsured people who can't go to the doctor even if they wanted to. For those people, Google is the only option. The CDC can collect as much data as it wants from the laboratories, but you can't collect what isn't there.
When you add in the uninsured people and the people who get better before visiting the doctor, it's not at all surprising that the lab estimates are 2x lower than reality. In fact, I would kind of expect them to be even lower than that.
Error in the Google model is by definition the difference from the CDC levels.
[edit: For example, the author of the paper could have compared Google and CDC versus a third data source, like sales of cold medicine. That would have been an interesting comparison, and might have shed some light on which estimates were a better reflection of reality.]
Things like "sales of cold medicine" and the like fall under the heading of "syndromic surveillance", and believe me, they've been tried for the better part of a decade without yielding much that's impressive. Which is frustrating, because theoretically they should be amazing, but they just aren't, especially when asked to predict rather than back fit.
The CDC numbers (which are part of ILINet) are pretty decent, and one of the things they do, which hasn't been mentioned here, is they're not trying to find all cases of influenza. They have a network of physicians in all 50 states reporting "Influenza-like Illness". These people then also send samples in for laboratory confirmation of virus type, drug resistance, etc. The numbers are also backwards revised as more data comes in, meaning they improve over the course of the year.
It's a very good system for getting an idea of how "bad" the flu season is - the main problem is that it's slow. The ideal, and what people were hoping from Google Flu Trends (and other systems like it, it's not alone) is "ILINet, but in real time".
It also fails fairly spectacularly on a local level when compared to health department data from some top flight state agencies, like NY.
There are three datasets here: the CDC's, Google's, and reality. Some folks here seem unable to differentiate between the CDC data and reality. But in fact, I really did have the flu, even though CDC didn't think I did. Reality is what matters.
Maybe there is reason to believe that CDC's data is closer to reality than Google's. If so, let's hear it.
The CDC is unapologetic about their data being estimates - but they're very solid estimates, and match more intensive but smaller scale studies pretty well. But we don't need to capture every single flu case - no surveillance system will ever do that, nor need to.
Edit: I see how the CDC's numbers could be a lower bound, but not upper.
Then perhaps you didn't have strictly the flu, but some other viral upper respiratory tract infection instead.
It is hard to judge from symptoms alone (which is why lab confirmations are required for diagnosis), but as a general rule of thumb influenza A infections in adults tend to be much more severe and long lasting than what you describe [1], with distinctively higher malaise, fever and incapacitation than your regular winter cold. It is so debilitating that you probably would have gone to the doctor in any case, I would guess.
[1] http://www.influenzareport.com/ir/cp.htm