Index: The Book of Statistical ProofsStatistical ModelsUnivariate normal dataUnivariate Gaussian ▷ Power analysis for one-sample t-test

Theorem: Let there be $n$ i.i.d. observations $Y_1, \dots, Y_n \sim N(\mu, \sigma^2)$ with both $\mu and \sigma^2$ unknown. Further, let $z_x = \Phi^{-1}(1 - x)$, where $\Phi$ is the cumulative distribution function of a standard normal distribution. Equivalently, with $Z \sim \mathcal{N}(0,1)$, let

\[\label{eq:zx-definition} z_x = \Phi^{-1}(1 - x) \quad \Leftrightarrow \quad \Phi(z_x) = 1 - x \quad \Leftrightarrow \quad \mathbb{P}(Z \leq z_x) = 1 - x \Leftrightarrow \mathbb{P}(z_x \leq Z) = x \; .\]

1) One-sided one-sample t-test:

We perform a one-sided one-sample t-test of the null hypothesis $H_0: \mu \leq \mu_0$ vs. $H_1: \mu > \mu_0$. The test has significance level $\alpha$. If $H_1$ is true and $\mu = \mu_0 + \delta$, then the power of the test is

\[\label{eq:power-one-sided} \Phi\left(\frac{\delta\sqrt{n}}{\sigma} - z_\alpha\right) \; .\]

If we keep $\alpha$, $n$, $\sigma^2$ the same as before and require that the power is at least $1 - \beta$, then the minimum detectable effect is

\[\label{eq:mde-one-sided} \frac{(z_\alpha + z_\beta)\sigma}{\sqrt{n}} \; .\]

If we keep $\alpha$, $\sigma^2$, $\delta$, $\beta$ the same as before, then the minimum required sample size is

\[\label{eq:mrss-one-sided} \frac{\sigma^2(z_\alpha + z_\beta)^2}{\delta^2} \; .\]

2) Two-sided one-sample t-test:

We perform a two-sided one-sample t-test of hypothesis $H_0: \mu = \mu_0$ vs. $H_1: \mu \neq \mu_0$. The test has size $\alpha$. If $H_1$ is true and $\mu = \mu_0 + \delta$ (assuming $\delta > 0$ without loss of generality), then the power of the test is

\[\label{eq:power-two-sided} \Phi\left(\frac{\delta\sqrt{n}}{\sigma} - z_{\alpha/2}\right) + \Phi\left(-z_{\alpha/2} - \frac{\delta\sqrt{n}}{\sigma}\right) \approx \Phi\left(\frac{\delta\sqrt{n}}{\sigma} - z_{\alpha/2}\right) \; .\]

If we keep $\alpha$, $n$, $\sigma^2$ the same as before, and require that the power is at least $1 - \beta$, then the minimum detectable effect is

\[\label{eq:mde-two-sided} \frac{(z_{\alpha/2} + z_\beta) \sigma}{\sqrt{n}} \; .\]

If we keep $\alpha$, $\sigma^2$, $\delta$, $\beta$ the same as before then the minimum required sample size is

\[\label{eq:sample-size-two-sided} \frac{\sigma^2(z_{\alpha/2} + z_\beta)^2}{\delta^2} \; .\]

Proof: The $t$-statistic is

\[\label{eq:t-statistic} T = \frac{\bar{Y} - \mu_0}{s/\sqrt{n}}.\]

which has a t-distribution with $n-1$ degrees of freedom. Unless $n$ is very small, the $t(n-1)$ distribution is very close to a $\mathcal{N}(0,1)$ distribution and for this proof, we will assume that this is a good approximation.

1) One-sided one-sample t-test:

The test has size $\alpha$, we reject the null, if

\[\label{eq:one-sided-rejection} T > z_\alpha \; .\]

So the power of the test is

\[\label{eq:one-sided-power-prob} \mathbb{P}\left(\frac{\bar{Y} - \mu_0}{s/\sqrt{n}} > z_\alpha\right) \; .\]

If the true mean is $\mu = \mu_0 + \delta$, where $\delta > 0$, then the power is

\[\label{eq:one-sided-power-shift} \mathbb{P}\left(\frac{\bar{Y} - \mu_0 - \delta}{s/\sqrt{n}} + \frac{\delta\sqrt{n}}{s} > z_\alpha\right) \; .\]

Now that we’re subtracting the true mean, the first term in the sum has a t-distribution, which we approximate with a $\mathcal{N}(0,1)$ distribution, denoted as $Z$. We also approximate $s$ with $\sigma$, since it should be a good approximation, and the error here doesn’t make much difference to the calculation. The power is therefore approximately

\[\label{eq:one-sided-power-final} \mathbb{P}\left(Z > z_\alpha - \frac{\delta\sqrt{n}}{\sigma}\right) = \mathbb{P}\left(Z < \frac{\delta\sqrt{n}}{\sigma} - z_\alpha\right) = \Phi\left(\frac{\delta\sqrt{n}}{\sigma} - z_\alpha\right) \; .\]

To calculate the minimum detectable effect, we want the smallest $\delta$, such that the power is at least $1-\beta$. If the power is at least $1-\beta$, then

\[\label{eq:one-sided-power-threshold} 1 - \beta \leq \Phi\left(\frac{\delta\sqrt{n}}{\sigma} - z_\alpha\right) \; .\]

Using $z_\beta = \Phi^{-1}(1 - \beta)$ and rearranging, we get

\[\label{eq:one-sided-delta-ineq} z_\beta + z_\alpha \leq \frac{\delta \sqrt{n}}{\sigma} \; .\]

Therefore,

\[\label{eq:one-sided-delta} \frac{(z_\beta + z_\alpha)\sigma}{\sqrt{n}} \leq \delta \; ,\]

and taking the minimum value for $\delta$ gives the minimum detectable effect.

To calculate the minimum required sample size, we want the smallest $n$, such that the power is at least $1 - \beta$. Rearranging \eqref{eq:one-sided-power-threshold}, we have

\[\label{eq:one-sided-n} \frac{\sigma^2(z_\alpha + z_\beta)^2}{\delta^2} \leq n \; .\]

2) Two-sided one-sample t-test:

The test has size $\alpha$, we reject the null, if

\[\label{eq:two-sided-rejection} |T| > z_{\alpha/2} \; .\]

So the power of the test is

\[\label{eq:two-sided-power-prob} \mathbb{P}\left(\left(\frac{\bar{Y} - \mu_0}{s/\sqrt{n}} > z_{\alpha/2}\right) \wedge \left(\frac{\bar{Y} - \mu_0}{s/\sqrt{n}} < -z_{\alpha/2}\right)\right) \; .\]

If the true mean is $\mu = \mu_0 + \delta$ (where $\delta > 0$ without loss of generality), then the power is

\[\label{eq:two-sided-power-shift} \mathbb{P}\left(\left(\frac{\bar{Y} - \mu_0 - \delta}{s/\sqrt{n}} + \frac{\delta\sqrt{n}}{\sigma} > z_{\alpha/2}\right) \wedge \left(\frac{\bar{Y} - \mu_0 - \delta}{s/\sqrt{n}} + \frac{\delta\sqrt{n}}{\sigma} < -z_{\alpha/2}\right)\right) \; .\]

As before, we use a $\mathcal{N}(0,1)$ distribution and we approximate $s$ with $\sigma$. The power is therefore approximately

\[\label{eq:two-sided-power-almost-final} \Phi\left(\frac{\delta\sqrt{n}}{\sigma} - z_{\alpha/2}\right) + \Phi\left(-z_{\alpha/2} - \frac{\delta\sqrt{n}}{\sigma}\right) \; .\]

If we have a large enough $\delta$, then the second term, which is the probability of rejecting the null hypothesis, but with a sign error on $\delta$, will be small. For the later calculations, we will set it to $0$ and use the approximation

\[\label{eq:two-sided-power-final} \Phi\left(\frac{\delta\sqrt{n}}{\sigma} - z_{\alpha/2}\right)\]

for the power. To calculate the minimum detectable effect, we want the smallest $\delta$, such that the power is at least $1-\beta$. If the power is at least $1-\beta$, then

\[\label{eq:two-sided-power-threshold} 1 - \beta \leq \Phi\left(\frac{\delta\sqrt{n}}{\sigma} - z_{\alpha/2}\right) \; .\]

Using $z_\beta = \Phi^{-1}(1 - \beta)$ and rearranging, we get

\[\label{eq:two-sided-delta-ineq} z_\beta + z_{\alpha/2} \leq \frac{\delta \sqrt{n}}{\sigma} \; .\]

Therefore,

\[\label{eq:two-sided-delta} \frac{(z_\beta + z_{\alpha/2})\sigma}{\sqrt{n}} \leq \delta \; ,\]

and taking the minimum value for $\delta$ gives the minimum detectable effect.

To calculate the minimum required sample size, we want the smallest $n$, such that the power is at least $1 - \beta$. Rearranging \eqref{eq:two-sided-power-threshold}, we have

\[\label{eq:two-sided-n} \frac{\sigma^2(z_{\alpha/2} + z_\beta)^2}{\delta^2} \leq n \; .\]
Sources:

Metadata: ID: P511 | shortcut: ug-ttest1power | author: alexanderdbolton | date: 2025-07-23, 00:00.