I'm an experimental physicist who mainly needs statistics for the calculation of uncertainties/confidence intervals. Since my results are usually normally distributed, I simply take $N$ measurements of the same signal and then calculate the sample standard deviation, i.e. $$ \sigma = \sqrt{\frac{\sum_{i=1}^N (x_i - \bar{x})^2}{N-1}} \,. $$ I usually then assume that my measurement uncertainty is $\Delta y = \sigma$.

However, there are people who say that a correction is necessary here. And I know of Student's t-distribution, which is a correction of the normal distribution for small sample sizes. Some people suggest that this should be used if you have only a small amount of measurements, which would be true in my case because I usually have $N=10$.

In practice, I have never seen anybody use this correction, I do not frequently read about it in papers, and in my own statistics courses I have never been told to do it. I would not know exactly how to use it, except that the Wikipedia entry has a bit about confidence intervals, including this:

Therefore, the interval whose endpoints are $$\bar{X_n} \pm A\frac{S_n}{\sqrt{n}}$$ is a 90% confidence interval for μ.

This is difficult to parse for me. As far as I understand it, A is (in this example) the tabulated correction factor that gives the two-sided 90% confidence interval, which is $A=1.833$ for $N=10$ because I have 9 degrees of freedom. $S_n$ seems to be the standard deviation. So if I say that my uncertainty is half the confidence interval, then the total correction factor is $1.833/\sqrt{10} \approx 0.58$, and $\Delta y = 0.58 \sigma$.

So is this something I should do with my data?

I am unsure because there is so little information on it, and nobody else seems to be doing it. Actually, a different correction method I have heard frequently about is to divide my standard deviation by a factor of $\sqrt{N}$, i.e. $$ \Delta y = \frac{1}{\sqrt{N}} \sigma \,, $$ which would be $\Delta y = 0.32 \sigma$ for $N=10$. The usual explanation for this is that the likelihood of the mean value being the correct value should increase if $N$ increases. That makes some sense, but I have never seen a statistical explanation for this correction factor. And the difference to the correction factor from Student's t-distribution is significant. So which one is correct?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.