3 ms·

If the rate of infection in the general population was even lower, say 1 in 100,000, and they actually trialled 200,000 people, finding that the vaccine perfect

If the rate of infection in the general population was even lower, say 1 in 100,000, and they actually trialled 200,000 people, finding that the vaccine perfectly prevented infection, should they conclude this with high confidence? Or have I misunderstood you? Besides, Dr Kim is quoted elsewhere as saying "the result could be due to chance" and the vaguely phrased "31.2 percent effective" is not encouraging. Add in the likely funding source of AIDS vaccine trials, the pressure to get results and the unknown mechanism and it doesn't look so great.

timr 17y ago
The result can always be due to chance. Confidence intervals let you estimate how likely it is that your results are due to chance. In this case, the results are unlikely to be random -- but it's always possible. That's all he was saying. Maybe this study was the 1 in 10,000 that would randomly show a 30% difference in infection rates. Is that likely? No. Is it possible? Sure. Anything's possible. Confidence intervals get smaller as you include more people in an experiment, but they get smaller more slowly as more people are added. So it's fairly easy to go from a 2% to a 1% confidence interval by adding a couple thousand people, but it takes a lot more people to go from 1% to 0.5%. And of course, it's asymptotic -- you'll never get to zero. That said, in this case, the point is academic. This study was large enough that the confidence interval is going to be well below 30%.
earl 17y ago
You didn't understand the parent. See my post below. Also, of course the result could be due to chance. You could flip a fair coin 25 times in a row and come up with heads each time -- it's just pretty damn unlikely. Similarly, I'd bet what he's measured is really this: you have two equally sized groups, one treated, one untreated. In group one you get a rate of x, and in group two, you get a rate of y. How big does |y-x| and the size of the two groups have to be to conclude it was not just chance? As below, a t-test can help you answer that.
psyklic 17y ago
The key is not to find "who gets HIV" vs. "who does not get HIV." Instead, we are trying to measure "who gets exposed" vs. "who gets HIV." Although we do not know how many people are exposed in each group, we can make sure that the number of exposed people in each group is approximately the same by increasing the number of people vaccinated. If we know that the same number of people are exposed (which is less than the sample size), then the number of people who get sick is much more significant. By increasing the number of vaccinated people, the number of exposed people becomes more equal between the two groups since a few "more exposed/risky people were assigned to group A" will have a less significant impact.

timr 17y ago

The result can always be due to chance. Confidence intervals let you estimate how likely it is that your results are due to chance. In this case, the results are unlikely to be random -- but it's always possible. That's all he was saying. Maybe this study was the 1 in 10,000 that would randomly show a 30% difference in infection rates. Is that likely? No. Is it possible? Sure. Anything's possible. Confidence intervals get smaller as you include more people in an experiment, but they get smaller more slowly as more people are added. So it's fairly easy to go from a 2% to a 1% confidence interval by adding a couple thousand people, but it takes a lot more people to go from 1% to 0.5%. And of course, it's asymptotic -- you'll never get to zero. That said, in this case, the point is academic. This study was large enough that the confidence interval is going to be well below 30%.

earl 17y ago

You didn't understand the parent. See my post below. Also, of course the result could be due to chance. You could flip a fair coin 25 times in a row and come up with heads each time -- it's just pretty damn unlikely. Similarly, I'd bet what he's measured is really this: you have two equally sized groups, one treated, one untreated. In group one you get a rate of x, and in group two, you get a rate of y. How big does |y-x| and the size of the two groups have to be to conclude it was not just chance? As below, a t-test can help you answer that.

psyklic 17y ago

The key is not to find "who gets HIV" vs. "who does not get HIV." Instead, we are trying to measure "who gets exposed" vs. "who gets HIV." Although we do not know how many people are exposed in each group, we can make sure that the number of exposed people in each group is approximately the same by increasing the number of people vaccinated. If we know that the same number of people are exposed (which is less than the sample size), then the number of people who get sick is much more significant. By increasing the number of vaccinated people, the number of exposed people becomes more equal between the two groups since a few "more exposed/risky people were assigned to group A" will have a less significant impact.