The recent news about the FDA stopping the google backed 23andMe service selling any more kits got me thinking. I understand the company may have been selling it as a medical tool prematurely, but what is it about SNP analysis that makes it so poor as at predicting disposition? Even other SNP companies have similar problems with accuracy and can't agree on disease disposition as shown in this paper. Why might they differ from 23andMe's predictions?
The long and short of it is that genetic variation is actually not very predictive in comparison to "environmental" effects such as lifestyle.
Only a quarter of the variation in lifespan between twins is attributable to inherited factors (including genetics and epigenetics)  - the rest is environmental, from lifestyle to air quality.
Most genetic variants (such as those measured on the chip used by 23andMe) have a very small impact on outcomes such as "Risk of disease X, Y or Z". Many risk loci have been identified, but without a full understand of what this risk means, the results of the 23andMe analysis can be misleading.
Some few rare genetic variants have a larger impact on diseases (such as early onset diabetes), but this comes under the "monogenic" category of genetic variation, rather than the common variants typically observed in large populations.
Basically, genetic risk scores are, at the moment, only indicative of an individuals risk, and to get the real picture a lot more information is required - for instance smoking is a better predictor of early mortality than any "common" (>=5% frequency in a population) genetic variants, likewise a high fat diet is more indicative of cardiovascular disease.