What Is the TOST Equivalence Test (Two One-Sided Tests) and How Do You Use It?

When you need to prove that two processes, products, or measurement methods are practically the same, a standard null-hypothesis test (like a t-test) won't help. A t-test can only show a difference; it cannot demonstrate equivalence. The TOST (Two One-Sided Tests) procedure, originally proposed by Schuirmann, is the classic method to statistically confirm that a difference is small enough to be considered equivalent within a pre-defined bound.

What It Is

TOST is a frequentist hypothesis-testing framework for equivalence testing. Instead of testing whether two means are equal (which is impossible to prove), TOST tests whether the true difference lies within an acceptable interval, \([-Δ, +Δ]\), where Δ (delta) is the equivalence margin you define based on practical or regulatory requirements.

The name comes from the procedure: you run two separate one-sided tests at a significance level of α (typically 0.05). If both tests are rejected, you conclude equivalence.

How It Works / Steps

1. Define the equivalence margin (Δ). This is the largest difference you are willing to accept as "not practically important." It must be set before collecting data, based on subject-matter knowledge, regulatory standards, or clinical relevance.
2. Collect your data and compute the sample means and the standard error (SE) of the difference.
3. Perform the two one-sided tests:

- Test 1 (lower bound): \(H_{01}: \mu_1 - \mu_2 \le -Δ\) vs. \(H_{a1}: \mu_1 - \mu_2 > -Δ\)
- Test 2 (upper bound): \(H_{02}: \mu_1 - \mu_2 \ge +Δ\) vs. \(H_{a2}: \mu_1 - \mu_2 < +Δ\)

1. Calculate the test statistics (assuming equal or unequal variances, similar to a t-test):

- \(t_1 = \frac{(\bar{x}_1 - \bar{x}_2) + Δ}{SE}\)
- \(t_2 = \frac{Δ - (\bar{x}_1 - \bar{x}_2)}{SE}\)

1. Reject both null hypotheses if \(t_1\) and \(t_2\) exceed the critical t-value (or if the p-values are both < α). Equivalently, you can check whether the 100×(1–2α)% confidence interval for the mean difference lies entirely within \([-Δ, +Δ]\). For α = 0.05, this is a 90% confidence interval.

A Worked Illustrative Example

Example data (illustrative only): Suppose you want to show that a new manufacturing process is equivalent to the current one for product strength. You set Δ = 2 units. From samples, you find:

• Mean difference (new – current) = 0.8 units
• Standard error (SE) = 0.5 units
• α = 0.05, so the critical t-value ≈ 1.96 (large sample)

Step 1 – Lower one-sided test:
\(t_1 = (0.8 + 2) / 0.5 = 5.6\) → p < 0.05, reject \(H_{01}\).

Step 2 – Upper one-sided test:
\(t_2 = (2 – 0.8) / 0.5 = 2.4\) → p < 0.05, reject \(H_{02}\).

Step 3 – Confidence interval check: The 90% CI for the difference is \(0.8 ± 1.96 × 0.5 = [-0.18, 1.78]\), which lies entirely within \([-2, +2]\).

Conclusion: Both one-sided tests are rejected, and the CI is within the bound. You can claim the new process is statistically equivalent to the current one.

Common Pitfalls

• Choosing Δ too large makes the test trivial; always justify the margin before seeing data.
• Using a 95% CI instead of 90% for the equivalence check — this is a frequent error. TOST at α = 0.05 corresponds to a 90% CI, not 95%.
• Confusing "no significant difference" with equivalence. A non-significant t-test does not prove equivalence; it only means you failed to detect a difference.
• Ignoring the assumption of independence and normality — check your data before applying TOST.

---

To avoid manual calculation errors and get instant results with proper confidence intervals, try the free, interactive Equivalence Test (TOST) tool at https://www.6sq.com/tools/equivalence/. It handles the math so you can focus on setting a meaningful Δ and interpreting the outcome correctly.