技术问询：能否计算逻辑回归中伪R²的置信区间（含已发表文献中的伪R²）
Alright, let's tackle your two questions about pseudo-R² confidence intervals in logistic regression—this is a common point of confusion, so great questions!
1. Can we calculate confidence intervals for pseudo-R² in logistic regression?
Short answer: Yes, though there’s no universal analytical formula (unlike linear regression’s R²), we can use resampling methods (most commonly bootstrapping) to estimate confidence intervals (CIs) for nearly any type of pseudo-R² (e.g., McFadden’s, Nagelkerke’s, Cox-Snell’s).
Here’s a step-by-step breakdown of the bootstrapping approach:
- Start with your original dataset (n observations, binary outcome, predictor variables).
- Resample with replacement: Create a new dataset of size n by randomly selecting observations from the original data (allowing repeats).
- Fit your logistic regression model to this resampled dataset, then calculate the pseudo-R² of interest.
- Repeat this process many times (typically 1,000–10,000 iterations) to build a distribution of pseudo-R² values.
- The confidence interval is then the percentiles of this distribution—for a 95% CI, you’d take the 2.5th and 97.5th percentiles.
This works because bootstrapping captures the variability in the pseudo-R² that comes from sampling variation in your data. Most statistical software (R, Python, Stata) has built-in functions or packages to automate this (e.g., the boot package in R).
2. Can we compute confidence intervals for pseudo-R² reported in published literature?
It depends on how much supplementary information the paper provides:
If the paper includes key details
Yes, you can estimate CIs if you have access to:
- Sample size (n)
- Log-likelihood values for the full model (
LL_full) and the null model (LL_null) (critical for calculating McFadden’s R² = 1 - (LL_full/LL_null) or Nagelkerke’s adjusted version) - Number of predictors in the model (k)
Here’s how to proceed:
- Parametric bootstrapping: Simulate new binary outcome data using the published model’s coefficients (if provided) or by deriving them from the reported log-likelihoods. For each simulated dataset, fit the logistic model, compute the pseudo-R², and build a distribution to extract the CI.
- Approximate analytical methods: For some pseudo-R² measures (like McFadden’s), there are rough variance approximations available in statistical literature. Note that these are less reliable than bootstrapping, but can serve as a quick estimate if resampling isn’t feasible.
If the paper only reports the pseudo-R² value (no other details)
Unfortunately, no—you can’t estimate a meaningful CI without knowing the sample size, model complexity, or underlying likelihood values. Pseudo-R² alone doesn’t capture the sampling uncertainty tied to the original data and model structure.
Pro tip
If the paper lacks necessary data, consider reaching out to the authors to request supplementary information (like log-likelihoods or raw data). Many researchers are willing to share to support reproducibility and further analysis.
内容的提问来源于stack exchange，提问作者pomodoro
