2 ms·

I don't want to discuss on the level of "pro" or "anti" but I would like to analyze some numbers from this paper. Take a look at this picture: https://www.nejm

by pc123 8y ago

I don't want to discuss on the level of "pro" or "anti" but I would like to analyze some numbers from this paper. Take a look at this picture: https://www.nejm.org/na101/home/literatum/publisher/mms/journals/content/nejm/2002/nejm_2002.347.issue-19/nejmoa021134/production/images/img_medium/nejmoa021134_t1.jpeg https://www.nejm.org/na101/home/literatum/publisher/mms/jour... There were 440,655 vaccinated children and 96,648 unvaccinated. Among the vaccinated 621 were autistic (of any kind), 117 among unvaccinated. Which gives us 0.14% and 0.12% respectively. Why doesn't this prove that vaccination has some impact on autism? Please refrain from attacking me, I just want to know the science explanation.

gerbilly 8y ago
I haven't read the paper, but it could be that the 0.02% difference is not statistically significant.[1][2] Say you pick two groups of people from a population, two groups of 10 000 people at random. Imagine then that you count the proportion of autistic people in each group. You might get say, 0.15% in group A and 0.14% in group B. If by chance the people in groups A and B had been partitioned differently between the two groups, i.e.: if enough autistic people in group A had instead ended up in group B, you can easily imagine that the numbers would change. Or even that they could flip. Some differences between the two were simply due to random chance. Statistical significance is a concept used to try to tease apart how much of an effect is due to a suspected cause (vaccination in this case) vs chance. Now in the case of this study, you might say, well there was no other way to partition the data, since all the vaccinated kids had to go in group A. While this is true, if autism has a different cause or causes than the partitions you chose effectively, your partitioning into groups would still be random with respect to the true cause of autism. So to 'prove'[3] whether a vaccines cause autism, we would prefer to see large effect sizes, which would strongly suggest that the effect is due to the cause we suspected. If instead, we had seen numbers like this: group A 15% and group B 7% then you could argue that the effect size is large enough to go beyond the effect caused by the sampling. Mathematical techniques in statistics and probability theory exist which allow you to choose a threshold (p value) in order to consider an effect to be statistically significant. [1] https://en.wikipedia.org/wiki/Statistical_significance https://en.wikipedia.org/wiki/Statistical_significance [2] https://towardsdatascience.com/statistical-significance-hypothesis-testing-the-normal-curve-and-p-values-93274fa32687 https://towardsdatascience.com/statistical-significance-hypo... [3] 'Prove' in quotes, because you can't really prove cause and effect using a study like this. To get closer to proof, you'd need an experiment where the conditions are specifically assigned to each group.

by pc123 8y ago

gerbilly 8y ago

I haven't read the paper, but it could be that the 0.02% difference is not statistically significant.[1][2] Say you pick two groups of people from a population, two groups of 10 000 people at random. Imagine then that you count the proportion of autistic people in each group. You might get say, 0.15% in group A and 0.14% in group B. If by chance the people in groups A and B had been partitioned differently between the two groups, i.e.: if enough autistic people in group A had instead ended up in group B, you can easily imagine that the numbers would change. Or even that they could flip. Some differences between the two were simply due to random chance. Statistical significance is a concept used to try to tease apart how much of an effect is due to a suspected cause (vaccination in this case) vs chance. Now in the case of this study, you might say, well there was no other way to partition the data, since all the vaccinated kids had to go in group A. While this is true, if autism has a different cause or causes than the partitions you chose effectively, your partitioning into groups would still be random with respect to the true cause of autism. So to 'prove'[3] whether a vaccines cause autism, we would prefer to see large effect sizes, which would strongly suggest that the effect is due to the cause we suspected. If instead, we had seen numbers like this: group A 15% and group B 7% then you could argue that the effect size is large enough to go beyond the effect caused by the sampling. Mathematical techniques in statistics and probability theory exist which allow you to choose a threshold (p value) in order to consider an effect to be statistically significant. [1] https://en.wikipedia.org/wiki/Statistical_significance https://en.wikipedia.org/wiki/Statistical_significance [2] https://towardsdatascience.com/statistical-significance-hypothesis-testing-the-normal-curve-and-p-values-93274fa32687 https://towardsdatascience.com/statistical-significance-hypo... [3] 'Prove' in quotes, because you can't really prove cause and effect using a study like this. To get closer to proof, you'd need an experiment where the conditions are specifically assigned to each group.