Cronbach's alpha与pair-wise correlation的关系及模拟验证咨询

Great question—let's break this down using your simulation setup, which perfectly illustrates the clean relationship between these two metrics when the tau-equivalence assumption (from classical test theory, Lord & Novick 1969) holds.

First, let's recap your scenario: you generated 10 responses with a multivariate normal distribution (mean=5, SD=2), fixed all pairwise correlations at 0.5, and found Cronbach's alpha landed around 0.91 for 250 simulated participants. This setup is exactly where the relationship between alpha and pairwise correlations is most straightforward.

The Exact Mathematical Relationship (Tau-Equivalence Only) #

When all items in your measure satisfy tau-equivalence—meaning every item has the same true score mean, and all pairwise correlations between items are identical (like your fixed 0.5)—Cronbach's alpha simplifies to this formula:

α = (k * r) / (1 + (k - 1) * r)

Where:

• (k) = number of items (10 in your simulation)
• (r) = the uniform pairwise correlation between items (0.5 here)

Let's plug in your numbers to confirm:

α = (10 * 0.5) / (1 + (10 - 1) * 0.5) = 5 / (1 + 4.5) = 5 / 5.5 ≈ 0.909

That's exactly the ~0.91 you observed—perfect alignment!

What This Means Intuitively #

• Higher pairwise correlations ((r)) directly boost alpha: the more your items measure the same underlying construct, the higher your internal consistency.
• More items ((k)) also increase alpha, even if the pairwise correlation stays the same. This is why longer scales often have higher reliability (assuming you're adding valid, tau-equivalent items).

Beyond Tau-Equivalence #

If your items don't meet tau-equivalence (e.g., some pairs have higher correlations than others, or true score means differ), the relationship gets messier. Cronbach's alpha becomes a lower bound for the average split-half reliability of your scale, and it will be less than the value you'd calculate using the average pairwise correlation in the formula above. But your simulation intentionally controlled for this, so we get that clean, predictable result.

内容的提问来源于stack exchange，提问作者Denis Cousineau