Thinking of the various tests and parameter estimates we perform with 99% confindence interval based on assumption of "normal distribution of errors" I asked myself a question what would be the 100% confidence interval for these estimated parameters ? This leads to formulation of normal distribution value Z at 0% or 100%. Is this virtually infinity or some specific number "big enough" ?

  • 3
    $\begingroup$ since the normal distribution has support on the entire real line, it would need to be infinity $\endgroup$ Aug 19, 2015 at 8:30

2 Answers 2

  1. The normal distribution covers the entire real line from $-\infty$ to $\infty$. To include 100% of the probability under the normal distribution would involve having an infinite range.

  2. A "normal distribution of errors" does not imply that the distribution you use for a parameter estimate is itself normal. For example, an interval for the mean of a $N(\mu,\sigma^2)$ (where both parameters are unknown) will be based on the $t$-distribution (also infinite in range), while an interval for $\sigma^2$ would be based on the $\chi^2$ distribution (which is semi-infinite). Other parameters might be on a finite range. For example correlations lie in $(-1,1)$.

However, in the limit as simple sizes go to infinity, in each of those three examples there will be a convergence of all the corresponding pivotal quantities to normality.

More generally, 100% intervals will typically cover the entire possible range of the parameter. The benefit of taking a smaller-than-100% interval is that you can take advantage of the way the intervals become narrower with larger sample size. A 99.9% interval for the mean of a normally distributed population will tend to shrink as $n$ becomes larger, but a 100% interval won't.


The normal distribution extends to +/- infinity, therefore to cover the 100% confidence interval you need to cover this range.

You can test this in Excel or Python using the inverse T value (i'm assuming 100 degrees of freedom. Use T.INV in Excel):

In [1]: import scipy.stats
In [2]: scipy.stats.t.ppf(1, 100)
Out[2]: inf
In [3]: scipy.stats.t.ppf(0, 100)
Out[3]: -inf

Your question about a "big enough" number will depend on your population. For example, for a distribution of the height of people, it makes no sense to have a +/- infinity range, so in practice you might want to truncate your range to say the height of the tallest man in the world.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.