ケルプラインは、ジョセフ・マンガノとジャネット・シャーマン共著の研究論文 "Elevated airborne beta levels in Pacific/West Coast US States and trends in hypothyroidism among newborns after the Fukushima nuclear meltdown"「福島原発事故後の米国ハワイ州と西海岸４州での大気中のベータ線量増加と新生児における甲状腺機能低下症の傾向」について、掲載誌であるOJPedの編集者へ質問状を送った。論文内の統計的計算などに疑問を持ったからである。だが、不思議な事に、この掲載誌OJPedでは編集者への質問状は受け付けないということだった。
Elevated airborne beta levels in Pacific/West Coast US States and trends in hypothyroidism among newborns after the Fukushima nuclear meltdown
Joseph J. Mangano and Janette D. Sherman
Letter to the editor of OJPed in response to:
Joseph J. Mangano, Janette D. Sherman. Elevated airborne beta levels in Pacific/West Coast US States and trends in hypothyroidism among newborn after the Fukushima nuclear meltdown. Open Journal of Pediatrics, 2013, 3, 1-9.
Alfred Körblein, email@example.com
Mangano and Sherman’s idea to study congenital hypothyroidism (CH) case numbers among newborn babies is new and seems to be promising. Their approach is reasonable; they compare the CH case numbers after Fukushima with those before Fukushima in 5 Pacific/West Coast states (Hawaii, Alaska, Washington, Oregon, and California: the study region) where the fallout from the Fukushima plumes was higher than in other parts of the USA. And they also look at the 2010-2011 change of CH cases in 36 other US states (the control region).
Mangano and Sherman state:
“The 2010-2011 ratio representing the change in CH cases was 1.16 for the five Pacific/West Coast States, rising from 281 to 327 confirmed cases. The 1.16 ratio exceeded the 0.97 ratio (decline in cases from 1208 to 1167) for the 36 control states; the difference is significant at p < 0.03. Increases in ratios were observed in the exposed areas for the periods March 17-June 30 (1.28, significant at p < 0.04) and July 1-December 31 (1.10, not significant at p < 0.21).”
Unfortunately the authors (M&S) do not clearly say how they tested the significance of their result. In their method section they state:
“CH cases for births in the periods March 17 to December 31 (2010 and 2011) will be compared, for the Pacific/West Coast States and the remainder of the US Portions of this 290 day period will also be compared. Significance testing will be conducted using a t test, where n equals the number of Pacific/West Coast cases in 2010 and 2011, the observed change will be the change in the Pacific/West Coast, and the expected change will be the change for the remainder of the US.”
So their hypothesis to be tested is whether there is a significant difference in 2010-2011 ratios between the study region and the control region. To check whether the reported p-values of p<0.03 and p<0.04 are correct, I applied a Poisson regression of CH numbers from their Table 4 in the two years (2010 and 2011) and in the two regions (study and control) and used two dummy variables: “d11”, which is 1 for 2011 and zero for 2010, and “study”, which is 1 for the study region and zero for the control region. A third dummy variable is the interaction dint=d11*study. Dummy variable “dint” is used to estimate the difference in 2010-2011 ratios between the study and control region.
The result of the Poisson regression was a 20.5% increase in CH cases in 2011 for March 17 to December 31 (p=0.041) and a 35.6% increase for March 17 to June 30 (p=0.049). So both results are statistically significant, albeit not at their stated p<0.03 or p<0.04 levels.
But it turns out that the estimate of the parameter for “d11” is not significant; the p-value s are p=0.400 for March 17 to December 31 and p=0.451 for March 17 to June30. As a rule, parameters that have no meaningful influence on the goodness of fit should be omitted in the regression and, per convention, p>0.2 is considered not meaningful. There is also a criterion that helps choosing the best model, the so called Akaike criterion (AIC), see http://en.wikipedia.org/wiki/Akaike_information_criterion, which is a measure for the goodness of fit. This is smaller for the regression model without d11 (AIC=35.2) than for the model with d11 (AIC=36.6).
The regression without d11, yields the following result for the shorter period March 17 through June 30:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 5.96229 0.03587 166.197 <2e-16 ***
study -1.40842 0.10869 -12.958 <2e-16 ***
dint 0.25014 0.13683 1.828 0.0675 .
Now the parameter for dint (which estimates the 2010-2011 change in CH cases in the study region) is not significant (p=0.0675). A similar result is obtained for the longer period March 17 through December 31 (p=0.0624).
The use of CH case numbers instead of CH incidences, however, means that their inferences can only be made if the live birth numbers don’t change between 2010 and 2011, or if the 2010-2011 relative change in birth numbers is the same in both the study and control region.
Official numbers of live births for 2011 are not available yet, but the US Centers for Disease Control and Prevention provide preliminary annual live birth data for 2011 for individual US states on their website . The respective data for 2010 can be found on the statehealthfacts website . Therefore I was able to determine the numbers of live births for 2010 and 2011 in the two periods (March 17-December 31 and March 17-December 31) and calculate the incidence rates. I used a logistic regression of CH incidence rates for the two time windows with only one dummy variable (dint) to estimate the excess in 2011 in the study region. This yielded a 16.6% increase in March 17-December 31 2011 p=0.0087) and a 32.5% increase in March 17-June 30 2011 (p=0.0036). Both results are significant at the p<0.01 level.
The main reason for the lower p-values compared to the regression of the CH case numbers is that in 2010 the incidences agree fairly well in the study and control region (the p-value for the dummy variable study is p=0.735) so study can be omitted. There is also good agreement between the incidences in 2010 and 2011 in the study region, so the dummy variable d11 (p=0.578) can also be omitted.
Essentially, the use of incidence rates instead of case numbers increases the statistical significance of Mangano and Sherman’s findings.
Several technical gremlins appear in the published tables. The confidence intervals in Table 1 and the ratios in the last two columns of Table 3 are incorrect, and the case numbers for the 36 control states in the last two rows do not sum up to the numbers in the first row in Table 4. Also page 3 states “Results showed that for I-131, the highest depositions, in becquerels per cubic meter....“. This should read “per square meter“. This was first spotted on the ex-skf.blogspot website .
However these glitches have no influence on the authors’ main finding.