Post a new comment on this article
Post Your Discussion Comment
Please follow our guidelines for comments and review our competing interests policy. Comments that do not conform to our guidelines will be promptly removed and the user account disabled. The following must be avoided:
- Remarks that could be interpreted as allegations of misconduct
- Unsupported assertions or statements
- Inflammatory or insulting language
Why should this posting be reviewed?
See also Guidelines for Comments and Corrections.
Thank you for taking the time to flag this posting; we review flagged postings on a regular basis.close
Data do not support conclusion.
Posted by RDFeinman on 04 Apr 2013 at 01:22 GMT
I think that it is good that the authors laid out all the individual (country) data in Figure 2 rather than obscuring it in group statistics like quartiles. However, accepting such data would be placing statistics ahead of common sense and the sensible and true avouch of ones own eyes. The very large scatter makes any association unreasonable. In fact there are almost as many points that support the opposite conclusion (high sugar, low diabetes incidence and vice-versa) as there are points that support the conclusion. A figure showing this is at https://dl.dropbox.com/u/...
Given how little we know about causes of diabetes, it has to be considered that what Figure 2 shows is a somewhat better chance of reducing diabetes by decreasing sugar but almost the same chance of increasing diabetes risk by lowering sugar. This might happen if, for some reason, a country felt that starch was safer than sugar -- for treatment at least, starch is clearly worse than sugar. (Such an outcome is one of the risks of taking the current paper to heart).
In addition, the outcome variable is per cent increase which does not describe increase from what. A country with a 2% incidence of diabetes that has a 1% increase would be at the same point as a country that initially had 10% diabetes and the same 2% increase although there would be many more new cases.
Incidentally, the shotgun nature of the individual outcomes makes one wonder what papers with quartiles would look like if all the data were shown.
Richard David Feinman
Professor of Cell Biology
SUNY Downstate Medical Center
RE: Data do not support conclusion.
statsguru replied to RDFeinman on 04 Apr 2013 at 03:56 GMT
The comment above confuses an added-variable plot with unadjusted correlations. The quartile comment ignores the regression results in the article; to adopt a quartile approach as displayed in the Figure in the comment would be to selectively bias the data by selecting point-in-time correlations without correcting for the overall trend in the time-series, which the authors of the paper do. Therefore the assertion of the comment's title is unsupported and somewhat insulting. The commentator appears to be advancing a personal theory about starches, and in doing so has ignored the statistical regressions in the study, presented an altered graph that could be misinterpreted by those unfamiliar with adjusted added variable plots [which should not be compartmentalized as the commentator has done], and ignored the Granger causality and Heckman selection models in the article. Indeed, the lack of understanding about how to interpret percentages in a regression result reveals a fundamental lack of understanding of statistics.
Data do not support conclusion although statistical manipulations may.
RDFeinman replied to statsguru on 04 Apr 2013 at 14:46 GMT
I meant by quartiles (or quintile, deciles or whatever the general term is) the common practice of dividing the data into groups according to the value of the independent variable (sugar availability) and then indicating the mean of the dependent variable. This practice obscures much information and I complimented you for not doing it.
The plot that I showed is not altered beyond indicating that sugar availability predicts the opposite to the conclusion almost as much as it predicts consistency with it.
Although I am quite willing to admit to many failings, I do know what an unadjusted correlation is which is what is not shown and what would undoubtedly indicate even weaker association. I don't have a personal theory about starch beyond the fact that, like sugar, it is a carbohydrate and it is carbohydrates across the board that have the greatest relation to diabetes as in the review http://www.nutritionandme.... Also, I did not personally do the experiment that Nuttal and Gannon did showing that starch is worse for treating diabetes than sugar. Wish I had.
I do admit to a fundamental lack of understanding of statistics although I do know what a percentage is and that it may be misleading in a regression analysis. For example, a country with large population that, for some reason (this is a theoretical example) had one case of diabetes but after a large increase in the availability of sugar, came up with two cases would have a 50 % increase in incidence and would be off the chart. I gave a more realistic example of how this could make the data in Figure 2 incorrect. So, it is quite possible that the actual increase in the number of cases would depend on how much sugar was available but it can't be proved from percentages. Despite my limited knowledge, I doubt that Granger causality and Heckman selection models will change that.
Again, it is good that the data were laid out so readers can draw their own conclusion. And, again, I admit to poor technical knowledge and on "advancing a personal theory" I am also the merest amateur compared to Dr. Lustig whose polemics, after all, are as ubiquitous as high fructose corn syrup.
RE: Data do not support conclusion although statistical manipulations may.
statsguru replied to RDFeinman on 05 Apr 2013 at 15:51 GMT
you clearly have no idea what you're talking about - you've ignored all the regression tables and selectively edited a graph, then you don't seem to realize that percentage change is necessary because you want to see dose-response effects, and correct for the fact that different countries start out at different levels of diabetes - hence you have to correct for the other factors involved and look for percent changes, not absolute changes. consult an introductory econometrics book.
Graph not altered, speaks for itself.
RDFeinman replied to statsguru on 06 Apr 2013 at 12:44 GMT
Again, I did not alter the graph. I highlighted parts to explain what it says.
Everything you say is probably true. I am only suggesting that you can have a statistically significant result that means nothing at all in practical terms. A country with low diabetes that reduced sugar availability would be predicted to have an equal chance of having more diabetes as having less diabetes.
The figure speaks for itself: https://dl.dropbox.com/u/...
RE: Graph not altered, speaks for itself.
statsguru replied to RDFeinman on 13 Apr 2013 at 18:42 GMT
again, you fail to understand statistics. the conclusion from the study was that decreases in sugar consumption show statistically significant correlations to decreases in diabetes prevalence. your logic is flawed: that there are variations around the regression line should always happen because of other variables. that's why they have the other variables in the regression tables. for example, tobacco and lung cancer are not perfectly correlated, and there are countries where decreases in tobacco occur but increases in lung cancer occur. that's because of other factors like genetics or time-delays,*** but doesn't negate the average regression effect of decreased tobacco on decreased lung cancer ***. that's why the authors use regression to see the effects of the average regression curve, and also other factors => the point is not that any decrease in sugar will always decrease diabetes, no matter what. the point is whether there's a statistically significant relationship between the two things, net of other factors--and their is. the authors even go farther to analyze the time trends and the net effect is in the tables, not in the diagram which shows the variations.
Answer to ad hominem...
RDFeinman replied to statsguru on 14 Apr 2013 at 00:04 GMT
...is who are you? Are you one of the authors? Before you tell me I don't understand statistics, I need to know who I am talking to. The idea that these data are as strong as the link between tobacco and lung disease is interesting but before we go on, I need to know who is insulting me.
RE: Answer to ad hominem...
PLOS_ONE_Group replied to RDFeinman on 17 Apr 2013 at 19:13 GMT
Please keep comments polite and in accordance with our guidelines: http://www.plosone.org/st...
Journal staff have raised the comments to the attention of the Academic Editor.
This was in answer to my complain about ad hominem...
RDFeinman replied to PLOS_ONE_Group on 09 Jan 2014 at 22:15 GMT
There was some confusion. Readers told me that this seemed to be directed at me whereas I was complaint about comments by statsguru about comments they made like "you clearly have no idea what you're talking about" instead of addressing the issues I raised.
RE: Data do not support conclusion.
RDFeinman replied to RDFeinman on 02 Mar 2014 at 03:13 GMT
The correct location of the figure is https://dl.dropboxusercon...