当灵敏度为1时阴性预测值的方差是多少？是否存在类似灵敏度Score公式的NPV方差替代公式？

1. Variance of NPV When Sensitivity = 1 #

First, to directly answer your question: using the formula you provided (var(NPV) = (1 - sensitivity)), when sensitivity equals 1, the variance of NPV is indeed 0.

This makes intuitive sense too. If sensitivity is perfect (1), every actual positive case is correctly identified as positive. That means any negative test result must be a true negative—there are no false negatives. So NPV (the proportion of negative test results that are truly negative) becomes 1, a fixed value with no variability. Hence, its variance is zero.

2. Alternative Formulas for NPV Variance (Similar to Sensitivity Score) #

Yes, there are alternative formulas for calculating the variance of NPV, analogous to how we compute variance for sensitivity. Here are two practical options:

a. Observed Count-Based Formula #

If you have raw counts from your diagnostic test data:

• TN: Number of true negatives
• FN: Number of false negatives

The variance of NPV can be calculated using the standard binomial proportion variance formula (since NPV represents the proportion of true negatives among all test-negative results):

var(NPV) = (NPV * (1 - NPV)) / (TN + FN)

This mirrors the common formula for sensitivity variance (var(sensitivity) = (sensitivity * (1 - sensitivity)) / n_pos, where n_pos is the number of actual positive cases).

b. Parameter-Based Formula (Using Sensitivity, Specificity, and Prevalence) #

If you want to express variance in terms of core diagnostic parameters (sensitivity Se, specificity Sp, and disease prevalence P), you can use the delta method to derive this formula. First, recall the definition of NPV:

NPV = (Sp * (1 - P)) / (Sp*(1-P) + (1 - Se)*P)

Let D = Sp*(1-P) + (1-Se)*P (the denominator of NPV). The variance can then be written as:

var(NPV) = [ ( (1-P)^2 * var(Sp) ) + ( P^2 * var(Se) ) ] / D^2

This formula is useful when you’re working with estimated values of sensitivity and specificity (e.g., from a study with reported confidence intervals) rather than raw counts.

Key Note #

Even with these alternative formulas, when sensitivity is 1, the variance of NPV will still be 0. That’s because (1-Se) becomes 0, making NPV a fixed value of 1—there’s no room for variability in the result.

内容的提问来源于stack exchange，提问作者AziR