168

The answers so far have focused on the data itself, which makes sense with the site this is on, and the flaws about it. But I'm a computational/mathematical epidemiologist by inclination, so I'm also going to talk about the model itself for a little bit, because it's also relevant to the discussion. In my mind, the biggest problem with the paper is not the ...


111

My primary concern with this paper is that it focuses primarily on Google search results. It is a well-established fact that smartphone use is on the rise (Pew Internet, Brandwatch), and traditional computer sales are declining (possibly just due to old computers still functioning) (Slate, ExtremeTech), as more people use smartphones to access the internet. ...


69

It is decidedly out of the ordinary. The reason is that counts like these tend to have Poisson distributions. This implies their inherent variance equals the count. For counts near $100,$ that variance of $100$ means the standard deviations are nearly $10.$ Unless there is extreme serial correlation of the results (which is not biologically or medically ...


60

Well, this paper establishes the fact that the number of Google searches on Facebook fits a certain curve nicely. So at best it can predict that the searches on Facebook will decline by 80%. Which might be feasible, because Facebook might become so ubiquitous that nobody would need to search about it. The problem with such type of models is that they ...


37

For the US data: You are confusing two important but different concepts in epidemiology: prevalence and incidence. A Wikipedia page describes the difference. The anti-smoking warning that you show says that 9 of every 10 lung cancers that occur are caused by smoking. That's the incidence of smoking-related lung cancers among all lung cancers that occur. ...


25

Not to give a complete or authoritative answer, but just to stimulate ideas, I will report on a quick analysis I made for a lab exercise in a spatial stats course I was teaching ten years ago. The purpose was to see what effect an accurate accounting of likely travel pathways (on foot), compared to using Euclidean distances, would have on a relatively ...


24

What you're asking about is called the "Population Attributable Fraction"—the number of cases in the entire population that can be attributed to the exposure (in this case, smoking). The formula for this is: $$ PAF = \frac{P_{{\rm pop}}\times (RR-1)}{P_{{\rm pop}}\times (RR-1)+1} $$ Here, $P_{{\rm pop}}$ is the proportion of exposed subjects in the ...


24

There are several points that you can improve in the code Wrong boundary conditions Your model is fixed to I=1 for time zero. You can either changes this point to the observed value or add a parameter in the model that shifts the time accordingly. init <- c(S = N-1, I = 1, R = 0) # should be init <- c(S = N-Infected[1], I = Infected[1], R = 0) ...


24

The Krasnodar Krai case is not the only one. Below is a plot for the data from 36 regions (I selected the best examples out of 84) where we either see a similar underdispersion or at least the numbers seem to be reaching a plateau around a 'nice' number (I have drawn lines at 10, 25, 50 and 100, where several regions find their plateau) About the scale of ...


19

In [1,ยง3.2], David Freedman suggests an essentially negative answer to your question. That is, no (mere) statistical model or algorithm could solve John Snow's problem. Snow's problem was to develop a critical argument supporting his theory that cholera is a water-borne infectious disease, against the prevailing miasma theory of his day. (Chapter 3 in [1], ...


19

I will just mention one aspect that I haven't seen mentioned in the other answers. The problem with any analysis that states that this is significantly out of the ordinary is that it doesn't take into account that the data have been selected based on looking strange. At least I'd assume that the thread opener has not only seen these data but also other data ...


17

Krasnodar The data for a region is clearly not realistic in terms of its dispersion. Here's a data on Krasnodar town. The sample average is 34 in May, and the dispersion is 8.7. This is more than Poisson distribution would suggest, where the dispersion is the square root of average, i.e. 5.9. This is overdispersed but the sample size is quite small so it's ...


16

Overview quick remarks The model with three points does make a better fit. The fit with three points is only slightly better. The model with only one point is not very bad. The difference in loocv score may indicate that the model with more points is a significant/probable/likely improvement, but the effect size is only small. Even if the three points model ...


14

Google Trend in my opinion can't produce a good data set for this case of study. Google trend shows how often a term is searched with Google so there are at least two reasons for raising some doubts about the prevision: We don't know if the user searches on Google Facebook to log in or if he searches information about Facebook Facebook is not only a site ...


12

A few basic issues stand out with this paper: It assumes correlation of search engine queries about a rising social network with the membership increases. This may have correlated in the past, but may not in the future. There are very few new large social networks. You can almost count them on one hand. Friendster, Myspace, Facebook, Google+. Also, Stack ...


12

So I think these are the data: month day new delta tens ones 4 29 63 NA 6 3 4 30 66 3 6 6 5 1 65 -1 6 5 5 2 79 14 7 9 5 3 82 3 8 2 5 4 96 14 9 6 5 5 97 1 9 7 5 6 97 0 9 7 5 7 99 2 9 9 5 8 ...


11

First, observational studies can have control. Like prospective cohort studies (people choosing to smoke versus people choosing not to) or case-control studies (people with outcome versus people without outcome.) A more proper contrast for observational studies is probably "intervention studies" or "experimental studies", in which researchers get to assign ...


11

Survival bias and competing risks. Also, elderly having a high value of a risk factor who have not been affected by that risk factor have demonstrated a robustness to that particular factor in general. This is why age $\times$ risk factor interactions can be important to pre-specify in a model.


10

I too speculate at the prevalence of logistic models in the literature when a relative risk model would be more appropriate. We as statisticians are all too familiar with adherence to convention or sticking to "drop-down-menu" analyses. These create far more problems than they solve. Logistic regression is taught as a "standard off the shelf tool" for ...


10

"If all you have is a hammer, everything looks like a nail." The dataset you have is small, possibly underrepresented, and of unknown quality, since it is argued that many cases could have not been diagnosed. You observe an exponential growth, a common phenomena in many natural and artificial processes. The curve fits well, but I'd bet that other similar ...


9

Two thoughts in addition to the other answers: prevalence is a fraction (ratio), not a rate. a rate is a fraction where the units in enumerator and denominator differ. The difference is usually a time (duration) in the denominator. Examples: incidence rate, growth rate, decay rate. e.g. incidence rate: number of newly diagnosed disease X cases per (...


9

First, it's worth recognizing that you cannot typically change sensitivity independently of specificity, and vice versa. This is the point of a ROC curve. Given the nature of the data generating process, and your specific data and model, you will always be stuck with some tradeoff between sensitivity and specificity. You would of course prefer to have 100%...


8

The growth of infected cases $y$ is more or less exponential but the growth rate $c$ is not constant. $$ \frac{\partial y}{\partial t} \approx c y$$ For instance, note in the graph how the change in cases from day to day depends on the number of cases in a particular day and the increase in cases is larger when the current cases are large. But, instead of a ...


8

This is actually a handbook example of determining the sample size needed for estimating binomial proportion (e.g. Jones et al, 2004, Naing, 2003 for other references and examples). First of all, to make it more precise, we are talking about finding such sample size, that with probability $\alpha$, the difference between the true probability of being ...


7

"All", "always", etc. are dangerous words. Most epidemiology studies are observational - as a field epidemiology tends to concern itself with study questions that are not amenable to randomization and controlled trials. The dominant form of studies one would encounter while doing graduate work in epidemiology or working in a public health department, or ...


7

I think the best example of this may likely be the controversy around hormone replacement therapy and cardiovascular risk - large cohort epidemiological studies seem to suggest a protective effect and health policy and physician recommendations were made on this information. Follow-up RCTs then seem to show that there's actually an increased risk of ...


7

The issue that the linear predictor can take the parameter outside its admissable range is real, but not limited to this case (common examples are seen when using the identity link with Poisson or Gamma GLMs). If the data stay away from the problem area that shouldn't necessarily pose any real difficulty. However, the two link functions don't correspond ...


7

I would be surprised if this result held up. Consider that the overall effect size is very small - the point estimate risk of a specific type of cancer (breast cancer) increased by 5% (alternatively: by a factor of 1.05) and was barely significant at the 95% level of confidence. Consider that data dredging indicated that the effect only held for a) pre-...


7

I am going to confine my comments to the SEIR model - the issues for the SIR model are similar and it can be treated as a special limiting case of the SEIR model anyway (for large $\delta$). What you've done so far I've had a look at your MATLAB code, which seems absolutely fine to me. For a given set of model parameters, your code solves the SEIR ...


7

As already said by others, sensitivity and specificity don't depend on prevalence. Sensitivity is the proportion of true positives among all positives and specificity is proportion of true negatives among all negatives. So if sensitivity is 90%, then the test will be correct for 90% of the cases that are positive. Obviously 90% of something smaller and 90% ...


Only top voted, non community-wiki answers of a minimum length are eligible