Mr Chance of WA needs to go back to school to understand the role of chance in animal feeding experiments.

EFSA review of statistical analyses conducted for the assessment of the MON 863 90-day rat feeding study June 2007: the resolution of the Seralini et al controversy about safety of GM maize MON 863.


Previous GMO Pundit posts indicate that WA Agriculture Minister Mr Chance needs to read a new European Food Safety Authority (EFSA) report [published after peer-review November 2007, see end of this post for citation and summary] demonstrating the deceptive role played by chance in the eating habits of groups of rats.

The scientific nitty gritty from the recent EFSA report that Mr Chance needs to understand properly is presented below, with key statements in bold. Full documents relating to this study are obtainable here at the EFSA website.

But the bottom line from this substantial EFSA study, reiterated again at the bottom of this post, is quite short.

Mr Chance needs to study statistics and recognise the role of chance in scientific experiments.

He needs to understand also that the EFSA official food safety agency for the EU has now carried out such a considered assessment of statistical chances.

This considered opinion of the EU safety agency re-affirmes that in the feeding safety tests on GM maize variety MON 863 carried out for Monsanto:

• Neither final weight nor average weight were different between the GMO and their Control groups.
• The number of statistically significant test results found (in both Monsanto and Seralini et al study of the same data) is not higher than might be expected by chance alone.

Many of the nitty gritty details from the EFSA report are quoted with key statements put in bold in the remainder of this long post.

This post is long because there are many substantial disagreements between the Séralini et al. (2007) study (which has been used in the past to make strong claims about GM food's lack of safety) and the latest EFSA and related studies, including several convincing demonstrations in the EFSA reports of several significant statistical errors by Séralini et al. (2007):

Notes from the 2007 EFSA report in MON 863 feeding trials.

Evaluation of body weight
Monsanto (2002) assessed the statistical significance of differences of body weight for each of the fourteen weeks separately.
To identify significant differences in weight gain patterns over time Séralini et al. (2007) modelled the entire fourteen week observation period by fitting one three-parameter Gompertz curve to the weight data of each treatment group i.e. each combination of Genotype, Gender, and Dose. This model was fitted to fourteen data points representing the mean body weight for each week (Séralini et al., personal communication, Appendix 2).
The fitting of a Gompertz curve to the weight data thus takes into consideration the measurements conducted each week into a single analysis rather than analysing each week separately. Whereas the fitting of a Gompertz curve to the weight data may therefore represent an interesting approach, it was deemed necessary to examine whether the assumptions required to conduct a valid statistical analysis were met. Key assumptions to be able to carry out valid statistical tests are that the residuals are independent and characterized by the same distribution.
Specifically, it was investigated whether the approach chosen by Séralini et al. (2007) for the analysis of the growth curves took into account the autocorrelation present in the dataset. Autocorrelation is the correlation between data when the same variable is repeatedly measured on the same subject (e.g. same subject measured over time). When such data correlation is not addressed then the model residuals may also be correlated and the assumption of their independence is thus violated. Ignoring this increases the likelihood of finding significant differences between treatment groups when, in reality, they might not be present.

It was also verified whether the increase in the variance of the weight data, associated to the increasing weights over time, was properly considered. In addition, use of the weekly mean weight data as done by Séralini et al. (2007) removed all variability between rats from the dataset and, as a result, makes valid statistical testing of differences between treatment groups questionable.

Because of these concerns, the weight gain data have been further investigated using alternative approaches.

Evaluation of other variables
For the other variables, the Monsanto (2002) report includes a one-way analysis of variance (ANOVA), one for each Week and Gender, in which all study groups (GMO at 2 Dose levels, Control at 2 Dose levels, and 6 references) were included. Contrasts included a test for differences between the GMO and the Control for each Dose level, and a comparison of the GMO at the 33% Dose level to the average of the 6 reference lines.

Séralini et al. (2007) compared, for each Dose level, the GMO group with its respective Control group using univariate tests. The authors point to a large number of significant results not previously reported by Monsanto, but referred for this to the Hammond et al. (2006) paper rather than the Monsanto (2002) study. However, as discussed above, the Hammond et al. (2006) provided only a subset of the results provided in the Monsanto (2002) report. Hence, it was determined which significant results in Séralini et al. (2007) were previously not already reported in Monsanto (2002) and vice versa.

In addition, any discrepancy in significant results between Monsanto (2002) and Séralini et al. (2007) was further examined in order to investigate if these could be attributed to errors in the statistical analysis.

Séralini et al. (2007) also pointed out that of the 494 comparisons carried out on these ‘other variables’ 40 differences (8%) were statistically significant at the 5% level while only 25 would be expected to be significant under the global null hypothesis of no differences between GMO and Control groups.

This statement required further examination. Firstly, it must be established how likely it is to obtain a number of significances as extreme as 40 from a sample of 494 tests at the 5% level. Secondly possible correlation within the dataset must be taken into account when examining the likelihood of finding more significant results than expected. For example, if the globulin concentration is statistically significantly increased whereas the albumin concentration is unchanged, then it would not be surprising that the albumin/globulin ratio is statistically significantly decreased as a result.

Hence, it must also be assessed how likely the finding of 40 would be from the set of tests that were conducted, given the correlation structure of the data. Thirdly, as also highlighted by Séralini et al. (2007), it is reasonable to define acceptable margins of measured differences between the tested GMO and its comparator. Hence, it is relevant to assess how likely the finding of 40 would be from a set of tests given that the null hypothesis of no effect was replaced by an assumption that a small difference between the GMO and Control groups can be expected and is therefore acceptable. Consequently, the EFSA evaluation examines the likelihood of obtaining 40 significant results by answering these three questions above.

Analyses performed in Member States: non-linear mixed model
The Commission du Génie Biomoléculaire (CGB) in France commissioned a statistical analysis of the weight data from the 90-day safety study. For this purpose Monod (2007) provided an extensive statistical analysis of the MON 863 data (Appendix 1).
In his report Monod (2007) emphasized that to be able to carry out valid statistical tests the underlying statistical assumptions must be met. Monod (2007) examined whether these assumptions were met when fitting a single Gompertz curve to the data from all the rats in a treatment group. He showed:
• That there is a large individual variability in the data between rats and this needs to be considered in the model;
• That the assumption of independence of residuals was violated. This violation of residual assumptions was present because the measurements were correlated, and such correlation invalidated the assumptions of independence; and
• That there was evidence of heteroskedasticity in the rat weight data, in the sense that, as expected, the variability of weight between rats increased with time i.e. with body weight itself.

Subsequently, Monod (2007) used the same Gompertz curve as Séralini et al. (2007), but taking into consideration the variability between rats. This was done by fitting a non-linear mixed effect model to the data, with rat as a random factor such that a separate Gompertz curve was fit to the dataset from each rat. The comparison between groups could then also be performed by comparing parameter estimates from different groups of rats. With this approach he was able to address both the heteroskedasticity and the autocorrelation of these repeated measures data.
Besides the Monod (2007) study, also the results on the weight and consumption data presented in the AFSSA (2007) report were considered.

Besides the Monod (2007) study, also the results on the weight and consumption data presented in the AFSSA (2007) report were considered.

Analyses performed by EFSA: one-way ANOVA and linear mixed model approach
A preliminary investigation of the weight data done with a factorial ANOVA performed week by
week where Gender (male or female), Dose (11% or 33% maize in diet), Genotype (GMO or Control) were the considered factors. Furthermore, a non parametric analysis, free from the assumption of normality and more robust against heteroskedasticity, was carried out according to the proposal of Scheirer el al (1976).
In addition to confirming the presence of autocorrelation in these repeated measurements data, the factorial ANOVA revealed some departures from normality and a few cases of heteroskedasticity.
It was decided to use a linear mixed model for longitudinal data in order to:
• Take into account all these factors and thus ensure a reliable assessment of the data;
• Consider all the weight data; and
• Not impose a prefixed growth curve, but consider a more generic form of autocorrelation allowing accommodating subjects’ variation in growth behaviour.
In this mixed model analysis, Rat was considered as a random factor whereas Gender (male or female), Dose (11% or 33% maize in diet), Genotype (GMO or Control) and Week (time in week from the start of experiment, 14 levels) were considered fixed factors. To estimate the model parameters the restricted maximum likelihood (REML) procedure was used.
Six different covariance models were considered in this report. A brief description of these models is provided in Appendix 3. For selecting the best covariance structure with respect to the dataset under consideration four model selection criteria were used (see Appendix 3).
These analyses were done using SPSS version 15.01.


Comparison of Monsanto and Séralini et al. results
As mentioned above, the 2004 EFSA opinions were based on evidence from the Monsanto (2002) report which was issued prior to the summary analysis of Hammond et al. (2006) which contained only a subset of the results.
To assess the agreement and discrepancies between the reported significant variables in the Monsanto (2002) report and the Séralini et al. (2007) paper, the following steps were taken:
• Clarifications were sought on the steps taken by Séralini et al. (2007) for their analysis. The authors clarified that the choice of univariate test was based on the outcome of the Shapiro test
for normality on the GMO group, the Shapiro test on the Control group, and the F tests for equal variances between the GMO and the Control group. When one of the Shapiro tests was significant (p less than or equal 0.05) the Mann-Whitney test was used. If not, the Student t test was used except for the cases where the F test was significant (p less than or equal 0.05) in which case the Student-Welch test was used. In addition, on the request of EFSA Séralini et al. provided a table indicating for each variable the p value and the statistical test that had been used (Appendix 2).
• Séralini et al. (2007) reported that they found 40 significant differences, but only identified 33 of them, as they considered 7 of those not relevant because the differences between treatment groups were considered too small (less than 5 %).
• EFSA verified whether any analyses had been reported only by Monsanto (2002) or only bySéralini et al. (2007). For those variables the statistical analyses were redone. For this purpose, the ests reported for the variable by Séralini et al. (Appendix 2) were redone (Splus 7.0). For the one-way ANOVA reported by Monsanto (2002) PROC GLM in SAS version 9.1 was used.
• All the statistically significant results from Monsanto (2002) and Séralini et al. (2007) were
then tabulated along with their p-values (Table 1,2 and 3).
• As shown in Table 4 the tested endpoints were then cross-tabulated into:
1. those which in both reports were reported as statistically not significant,
2. those which in both reports were reported as statistically significant,
3. those which indicated significance in Séralini et al. (2007) but not in Monsanto (2002), and
4. those which did not indicate significance in Séralini et al. (2007) but did so in Monsanto
(2002).

Simulation studies
To assess whether the 40 or so significant observations could be due to random variation alone it was considered worthwhile to estimate by simulation how likely it was to observe 40 significant results in the following three situations:
• Under the assumption that the endpoints are independent and that GMO and Control means are exactly the same (as assumed by Séralini et al.);
• Under the same null hypothesis but given the correlation structure as estimated from the data (derived from 119 x 4 = 476 endpoints at weeks 5 and 14); and
• Under the assumption that GMO and Control means might in fact be slightly different, given a distribution of acceptable differences. In these simulations the degree of acceptable difference was characterised by the ratio of between-group to within-group standard deviation.
Simulations were performed firstly using various pre-defined degrees of acceptable difference, and secondly using a distribution of acceptable differences as estimated by a random effect model from the data of the 6 reference groups in the MON863 study (Genstat release 9.2).
The simulation studies and the models used are described in detail in Appendix 5.

Results
...

Analyses performed in Member States: non-linear mixed model
A first analysis with the Gompertz model fitted to all the data showed a number of outliers. These values, as well the subsequent measurements, were eliminated from the analysis dataset.
The main findings reported by Monod (2007)(Appendix 1) are as follows:
• Neither with males nor with females there was a significant difference in the value of the 3 model parameters when considering the four GMO and Control treatment groups in either Males or Females;
• There was no significant difference in the growth curve parameters when comparing the treatment groups the Males fed 11% GMO diet to their Controls;
• The growth curve parameters were significantly different between the Female treatment group fed the 33% GMO diet and its Control. The level of significance was only (p=0.045) though.
The difference was attributable to the slope parameter b for the fitted Gompertz curve and not to the parameter a indicating the final weight.
Those findings were also valid when the outliers were included in the analysis.

Analyses performed by EFSA: one-way ANOVA

A detailed description of the results can be found in Appendix 3. The main findings are reported below.
Three rats, were excluded from the analysis because either they were identified as outlier (B38656) or, were only present in the trial until week 5 (B38923 and B38967) (Appendix 3). Another four rats (B38612, B38642, B38690, and B38789) were characterized by unusual growth patterns (Appendix 3). In order to assess the potential influence of data from latter four animals on the statistical results, all the analyses reported in this study were carried out twice: (i) including these four rats (referred to as analyses ‘with 4 potential outliers’), and (ii) excluding them (referred to as analyses ‘without 4 potential outliers’).
The one-way ANOVA, in which weight data were analyzed separately week by week, showed, as expected, a significant main effect of Gender. There also was a significant interaction (0.01 less than p less than 0.05) of Gender and Genotype (GMO vs. Control) in (up to) 4 of the 14 study weeks. Females fed the GMO diet showed a slightly greater body weight compared to their Controls at those weeks. The opposite was the case for Males. However, for both Genders differences were small i.e. below 3% and 4%, respectively. Given that the data were affected by some heteroskedasticity problems, results of the ANOVA may show an increased type 1 error rate. With the non-parametric approach, none of the Gender by Genotype interactions were significant.

Results on linear mixed model approach
A detailed description of the results can be found in Appendix 3. The main findings are reported below.
Based on the different information criteria used, the better covariance structure was the heterogeneous Toeplitz. There was a significant interaction between Gender and Week showing that weights differed between the two genders over the course of the fourteen weeks. A significant interaction between Week and Genotype was also present, reflecting a difference in rats’ response to diet at different weeks. Finally, there was a significant three-way interaction between Gender, Week, and Genotype, indicating lack of consistency between Males and Females to Genotype at different Weeks.
The mixed model analyses conducted on the data from reference groups and the Control groups (that is, without the GMO groups) also showed the two same significant 2-way interactions of Genotype by Week and Gender by Week to be present.

Simulation Studies
Details of the simulation studies are reported in Appendix 5. The simulation studies were performed assuming multivariate normality and homoskedasticity. Whereas these assumptions may sometimes fail for the actual data, the simulation study was intended to illustrate two general points, and the results have relevance for any specific structure of data.
Details on how the correlation was computed are reported in Appendix 4. When accounting for the correlation between endpoints and time-points (weeks 5 and 14), the probability to observe 38 or more (false positive) significant results was found to be 5.5%, i.e. about 20 times greater than without accounting for this correlation. Such results were derived under the assumption that the means of simulated endpoints were exactly the same for both groups, which is conservative and is not expected to be true for all variables in equivalent groups.
Simulations were also performed accounting for acceptable difference between groups. First, when using various predefined degrees of between group variability, it turned out that the probability to observe at least 40 significant cases ranged from 3.8% to 86% when the ratio of between-group to within-group standard deviation ranged from 0.1 to 0.2. Second, the distribution of differences estimated from the data of the 6 reference groups was estimated. Among these groups the between Statistical group standard deviation was estimated as negligible in 100 of the investigated chemical and haematological endpoints, but it had a positive value up to 50 % of the within-group standard deviation in the remaining 84 endpoints. Using this empirical distribution of 184 ratios to characterise acceptable variability between-groups, at least 40 significant cases were observed out of 494 comparisons in 54% of the simulated datasets.
In reality, both correlation and background variability are present. This leads to the conclusion that it should be expected to observe 40 significant cases.


Discussion
Model choice
Séralini et al. (2007) stated that a multivariate analysis had been conducted on the food consumption and the body weight data. Séralini et al. (personal communication) confirmed that they performed a multivariate analysis on the food consumption data, which was however not described in the paper.
Their statistical approach for the analysis of the body weight data consisting of the fitting of a nonlinear growth curve (Gompertz curve) to a series of weekly treatment group does not represent what is known as a multivariate analysis. Nevertheless, the fitting of a Gompertz curve to the weight data represents an interesting approach. However, the fact that this curve was fitted on the mean weights of each group essentially removed all variability between rats from the data. This greatly inflated the probability to find statistically significant results because it ignored the variability of the rat weights within each group. In addition, this approach did not allow for testing of independency of residuals with acceptable power, as the number of residuals was very low (n=14).
Indeed, other statistical methods such as linear or non linear mixed-effects modelling proposed in this report are preferable because these approaches take into account the correlation structure of the data.
Mixed models were considered very suitable to analyze the weight data of the MON863 90-day rat feeding study for the following reasons:
• The between-rat variability was accounted for;
• The models allowed to handle for the presence of correlated data and non-constant variability as well as some missing observations. Mixed models offer also the possibility to use different covariance structures within subjects over time; and
• They are robust to deviations from the normality assumption, unless the datasets are strongly unbalanced (which was not the case for the weight data)
The possibility of testing different covariance structure models in the linear mixed model represented an advantage in this specific case because it allowed overcoming the problem of lack of fit of non linear fitting in the weeks following the blood collection that took place during the 5th week.
Interestingly, although heterogeneous Toeplitz could be clearly identified as the most appropriate covariance structure model, there was consistency of results among all the six covariance structure models tested in this report, suggesting the obtained results are robust.

Potential effects of Genotype and Dose
Several questions can be raised with regard to the identification of potential differences in body weight that might be attributable to differences between the GMO groups and their corresponding Control groups. These include:
• Were there differences in the final weights?
• Were there differences in the average weight or the weight gain during the 14 week study period?
• Were there differences in the weight gain patterns during the study period?
The Monod (2007) and the EFSA results confirm the Monsanto (2002) report that there were no statistically significant differences between the GMO treatment groups and their respective Controls in the average body weight during the study nor in the final weight at the end of the study, respectively.
Monod (2007) showed that, when the Gompertz model was used with appropriate model assumptions, the parameters of the growth curve models were not significantly different between the GMO and the Control treatment groups, neither for males nor for females, with the exception of one treatment group (33% Females) which showed a slight (p=0.045) statistically significant difference. This is in contrast with Séralini et al. (2007) who reported that the growth curves for two of the four GMO treatment groups showed a strong statistical difference (pless than 0.001), when compared to their respective Controls.


The significant interaction of Gender and Genotype in (up to) five of the fourteen weeks in the oneway ANOVA as performed by EFSA and the significant two-way interaction between Genotype and Week and the three-way interaction between Genotype, Gender, and Week in the linear mixed model analysis reflect:
• Differences between treatment groups in different weeks; and
• Lack of consistency between Males and Females for those differences.
This could indicate an effect of GMO on the growth rate of rats. However, when the analysis was carried out excluding the GMO from the dataset (i.e. analysis only on the commercial lines and the Controls) the two-way interactions of Week with Gender and Week with Genotype were also present.
Therefore, it can not be concluded that the latter interaction is due to the presence of MON863 in the diet.
The 3-way interaction might suggest a pattern of higher increase in body size of the Male GMO treatment groups when compared to their Control groups and a pattern of slightly higher body weight gain of Female GMO treatment groups, when compared to their Control groups. However, for both Male and Female genders the weight differences between GMO treatment groups and Controls very small i.e. below 3% and 4%, respectively.

The results of the linear and the non-linear mixed model analyses therefore confirm the results of Monsanto (2002) in that there are differences for the 33% Female and 11% Male treatment groups in some weeks.
The weekly data on food consumption indicate that some of these transient differences in weight and weight gain are consistent with the observed patterns in the significant differences in food intake between these GMO treatment groups and their Control counterparts (Monsanto, 2002):
• There was a significantly lower food consumption in weeks 3-6 with 11% Males; and
• A significantly higher food intake in week 8 in the 33% Male.
The AFSSA report (2007) illustrates that the food consumption in the 11% Dose GMO Male group was almost always lower than in the equivalent Control group throughout the study. Thus it is not surprising that the weight gain in that group tended to be numerically lower. The AFSSA (2007) report also illustrates that from week 7 the food intake in the Male 33% GMO group was higher than in its Control group. This is consistent with a pattern of higher weight gain in that period

Part 2: Other variables
For the other variables, Séralini et al. (2007) highlight a number of differences to be significantly different from zero. Many (n=25) of the significant results on variables tested in both studies were already previously reported in the Monsanto (2002) report examined by the EFSA GMO panel. Thus, in effect, the Séralini et al. (2007) study provides independent confirmation for many of the previous findings. Nevertheless, among the 40 reported variables some were found to be significant in the Séralini et al. (2007) paper and not in the Monsanto (2002) report and vice versa. There are two main reasons for these differences:
• Differences in statistical significance appeared when a t-test was used in both studies.In the Monsanto (2002) study the comparison between the GMO and Controls was based on an ANOVA approach in which all ten study groups were included. In contrast, Séralini et al. (2007) compared the mean of the GMO11% versus Control 11% group and the GMO 33% versus Control 33% using simple univariate tests. Hence, where they occur these differences are due to the use of different variance estimates in both approaches. Determining which of these variance estimates is the more appropriate one to use for the t test is a matter of debate and depends on how much the variance of the reference lines is considered to provide information that is of relevance for the hypothesis being tested.
• In addition, instead of using at t-test, Séralini et al. (2007) performed a Student-Welch test when the equal variance assumption between the two groups was not met and a non parametric test (Mann-Whitney) in case of non-normality. While the use of these tests may be quite appropriate it is not the case that they led to the identification of more statistically significant results. In other words, sometimes results were not identified as significant by Séralini et al. (2007) using these other tests whereas the t-test used in the Monsanto (2002) indicated the result to be significant.
More relevant is the concordance of both approaches in the identification of 25 statistically variables. These represent, from a statistical perspective, the more robust results.
There is no evidence that the test methods used by Séralini et al. (2007) resulted in the identification of more significant differences than in the approach used by Monsanto (2002). The total number of significant results was in fact very similar for both (Table 4). Séralini et al. (2007) wrote that on average 25 ‘false positive’ significant results were expected, while 40 were found. However in both the Séralini and Monsanto analysis the statistical tests were performed assuming that no differences exist between the treatment groups in the averages of the baseline values of their test parameters. Due to the correlation between the variables (which was clearly present in this study) and the fact that
variable means from different treatment groups may sometimes actually be slightly different, the probability that 40 or more positive results in a set of 494 tests are obtained by chance alone is substantial.
As the aim of safety studies is to identify any potential adverse effect, they should be organised and focussed on finding differences with primary attention for restricting the false negative error rate (finding no difference where a biologically relevant difference does in fact exist). This may often come at the cost of finding statistically significant results which are not biologically relevant and thus need careful review and interpretation.


When a multiplicity of statistical tests is performed, one on each of a large number of variables, it is to Statistical report – main document, page 19 be expected that some will attain statistical significance by chance alone. There are several ways one can make allowance for this:
• Classically this is done by adjusting the level at which significance is assessed, e.g. Bonferroni adjustments (Bonferroni, 1935), or by using protected pairwise tests which are conducted only after a significant omnibus F-test. Recently proposed approaches to address massively multiple comparisons are based on controlling false discovery rates (FDR) as introduced by Benjamini and Hochberg (1995).
• Monsanto (2002), although they reported every significant contrast, also used a protected approach. This was done by only considering for further evaluation those variables for which the overall test results of the one-way ANOVA on 10 diets was significant. If this was the case, then they tested whether the 33% GMO diet was significantly different from the means of the reference lines. Finally, if these were significant, then the means of the 33% GMO diets were compared to the range of the commercial diets to assess whether they were biologically meaningful. They reported that there were 19 findings that satisfied all the above criteria.
• Séralini et al. (2007) used unprotected tests and reported the significant results of each of the 494 variables tested except for those with a difference of less than 5%. They then assessed whether differences could be due to differences in the diet by comparing the control groups to the mean of the reference groups and finally compared the remaining significant parameters to the mean of the six reference groups. They reported that seven results satisfied all the above criteria.
• Of the seven variables proposed as key findings by Séralini et al. (2007) (highlighted in Tables 1) three were also significant in the Monsanto (2002) analysis. The remaining four concern urine parameters in the 33% males (Table 1) (Séralini et al., 2007): urine phosphorus (weeks 5 and 14), and urine sodium (week 14), and urine sodium excretion (week 14).
Statistical analysis is a useful tool to point out possible safety problems. However, it does not in itself constitute a sufficient condition to demonstrate presence of an ill effect. For that purpose it is necessary to interpret these results and consider whether the differences are of a biologically relevant magnitude. The allowed discrimination between genders and between weeks, but no visual difference between treatment groups could be found.
This report contains the following information that provides a biological context for any differences found between treatment groups and can thus be used to aid the interpretation of statistical significance tests (Appendix 4):
• The results on variability of the data and the presence of outliers within each treatment group (Box plots);
• The correlation matrices between variables; and
• The principal components analysis which helped identified the main sources of variation in the dataset (Gender and Week) and thus helps to put into perspective any findings attributed to Genotype.


Conclusions
Overall, the various studies that were conducted on body weight show a consistent picture in that:
• Neither final weight nor average weight were different between the GMO and their Control groups;
• During the study the differences in the weight data between the GMO and their Control
groups sometimes varied by Week and Gender, but differences were always below 4 %. These
differences are probably attributable to fluctuations in food intake.
The results of the statistical analyses on haematology, serum and urine chemistry, and organ weight data reported by Séralini et al. (2007) are largely consistent with the findings previously reported by Monsanto (Monsanto, 2002). The differences resided in the interpretation of these results i.e. which statistically significant results were considered biologically meaningful. The number of statistically significant test results found (in both studies) is not higher than might be expected by chance alone.


References to the EFSA 2007 Statistical Report
(Available at link that starts this post)

AFSSA, 2007. Agence Française de la Sécurité Sanitaire des Aliments. Dossier n 2007-SA-0109
http://www.afssa.fr/Ftp/Afssa/40728-40729.pdf

EFSA, 2004a. Opinion of the Scientific Panel on Genetically Modified Organisms on a request from the Commission related to the Notification (Reference C/DE/02/9) for the placing on the market of insect-protected genetically modified maize MON 863 and MON 863 x MON 810, for import and processing, under Part C of Directive 2001/18/EC from Monsanto. The EFSA Journal, 49, 1-25.
http://www.efsa.europa.eu/etc/medialib/efsa/science/gmo/gmo_opinions/381.Par.0001.File.dat/opinion_gmo_06_en1.pdf

EFSA, 2004b. Opinion of the Scientific Panel on Genetically Modified Organisms on a request from the Commission related to the safety of foods and food ingredients derived from insect-protected genetically modified maize MON 863 and MON 863 x MON 810, for which a request for placing on the market was submitted under Article 4 of the Novel Food Regulation (EC) No 258/97 by Monsanto. The EFSA Journal, 50, 1-25.
http://www.efsa.europa.eu/etc/medialib/efsa/science/gmo/gmo_opinions/383.Par.0001.File.dat/opinion_gmo_07_en1.pdf

EFSA, 2004c. Statement of the Scientific Panel on Genetically Modified Organisms on the evaluation of the 13-week rat feeding study on MON 863 maize, submitted by the German authorities to the European Commission.
http://www.efsa.europa.eu/etc/medialib/efsa/science/gmo/statements/666.Par.0001.File.dat/sr_gmo01_statement_study_MON 863_en1.pdf

Benjamini, Y. and Hochberg, Y., 1995. Controlling the false discovery rate: A practical and powerful approach to multiple testing. J. Roy. Statist. Soc. B 57, 289–300.

Bonferroni, C. E., 1935. Il calcolo delle assicurazioni su gruppi di teste. In: Studi in Onore del
Professore Salvatore Ortu Carboni. Rome: Italy, 13-60.

Hammond, B.G., Dudek, R., Lemen, J.K., Nemeth, M.A., 2006. Results of a 90-day safety assurance study with rats fed grain from corn borer-protected corn. Food Chem. Toxicol., 44, 1092-1099.

Monod H., 2007. Expérience sur rats menée par Monsanto en 2001-2002 avec certains régimes
comportant du maïs génétiquement modifié: analyse statistique des courbes d’évolution du
poids. Study conducted upon request of the CGB (Commission du Génie Moléculaire, France).
http://www.ogm.gouv.fr/experimentations/evaluation_scientifique/cgb/autres_avis/Avis_CGB_MON 863_15juin2007.pdf

Monsanto, 2002. 13-Week Dietary Subchronic Comparison Study with MON 863 Corn in Rats
Preceded by a 1-Week Baseline Food Consumption Determination with PMI Certified Rodent
Diet #5002
http://www.monsanto.com/monsanto/content/products/technicalandsafety/fullratstudy.pdf

Scheirer J.C., Ray W.S., Hare N., 1976. The Analysis of Ranked Data Derived from Completely
Randomized Factorial Designs Biometrics, 32, 429-434.

Séralini, G.E., Cellier, D., de Vendomois, J.,S., 2007. New analysis of a rat feeding study with a
genetically modified maize reveals signs of hepatorenal toxicity. Arch. Environ. Contam.
Toxicol., 52, 596-602.



Updates

See later post Deja vu all over again for a full explication of the French mess associated with Seralini et coll. including documentation of their funding by Carrefour, a transnational supermarket chain.

19 December 2007
EFSA response now published as

Doull, J; Gaylor, D; Greim, H; Lovell, DP; Lynch, B; Munro, IC (2007) Report of an expert panel on the reanalysis by Seralini et al. (2007) of a 90-day study conducted by Monsanto in support of the safety of a genetically modified corn variety (MON 863) FOOD AND CHEMICAL TOXICOLOGY, 45 (11): 2073-2085 NOV 2007

MON 863, a genetically engineered corn variety that contains the gene for modified Bacillus thuringiensis Cry3Bb1 protein to protect against corn rootworm, was tested in a 90-day toxicity study as part of the process to gain regulatory approval. This study was reanalyzed by Seralini et al. who contended that the study showed possible hepatorenal effects of MON 863. An Expert Panel was convened to assess the original study results as analyzed by the Monsanto Company and the reanalysis conducted by Seralini et al. The Expert Panel concludes that the Seralini et al. reanalysis provided no evidence to indicate that MON 863 was associated with adverse effects in the 90-day rat study. In each case, statistical findings reported by both Monsanto and Seralini et al. were considered to be unrelated to treatment or of no biological or clinical importance because they failed to demonstrate a dose-response relationship, reproducibility over time, association with other relevant changes (e.g., histopathology), occurrence in both sexes, difference outside the normal range of variation, or biological plausibility with respect to cause-and-effect. The Seralini et al. reanalysis does not advance any new scientific data to indicate that MON 863 caused adverse effects in the 90-day rat study. (C) 2007 Elsevier Ltd. All rights reserved.