[wjnc]

Well it might be. They also include a casualty from NL at face value, while casualties are MUCH rarer here. I agree with the need for discounting of a Chinese casualty, but it's quite common practice to try to get more data points. That is good practice, not bad, and we only seem to disagree with the execution (is a China casualty 1 or .8?) In this case it wouldn't matter the base take-away much.

[AstralStorm]

You have to at least account for differences in base rate of accidents. The typical normalisation would take at least published accident total to synthesise data for the other cars. (By using sales numbers of cars.)

What he did there is statistically unacceptable.

The medical equivalent would be to compare side effect profile of current best of line treatment done in general populace to results to your new treatment's phase 1 trial results which is done in healthy people only. Then advertise that.

Regardless, can we have the raw data please? Are the rates even meaningful when the incident counts are low? Vehicle years are not the right unit as this counts cars that are not being driven. Actual range driven is the right unit.

[wjnc]

I totally agree. But keep in mind the effect "doing the right thing" has on the effect to be estimated. I venture it will not change the gist of the story. Heck, I should contact the author and build a shiny app to demonstrate. Even discounting a Chinese death 50% would put Tesla in the ballpark double the risk of luxury cars. My thought would be that Tesla is more a supersport car (McLaren, etc) in risk (for some drivers, say, male, young or midlife, etc).

That one local death in NL, the guy did double the allowed speed on a stretch where I can hardly hit 20% above the allowed speed accounting for traffic lights and local situation. The one Tesla death with the young kid, the same. Don't put maximum traction hundreds of HP in male hands. We tend to use it (sometimes).

[AstralStorm]

I suspect the correct use of Chinese data would make error bars overlap, exposing the BS.

Moreover, they should use the data where miles drive is available, say, from taxi drivers. Oops, that's 0 out of 0 total? :)

For such low incidences, even studentising will not produce right confidence intervals based on normal distribution. Poisson distribution cannot even be used, direct binomial instead, because it is likely that the error due to approximation will be important. (calculate from rule of rare events)