Q: Which distribution should I use?
A: Start with the data characteristics:
- Symmetric, no bounds → Normal
- Positive, right-skewed → Lognormal, Gamma
- Bounded [0,1] → Beta
- Count data → Poisson, Binomial
- Time-to-event → Weibull, Exponential
Then test with GOF tests and compare AIC/BIC.
Q: MLE vs Moments vs Quantile - which to use?
A:
- Default: MLE (most accurate)
- MLE fails: Moments (fast, always works)
- Outliers present: Quantile (robust)
Q: How many bootstrap samples do I need?
A:
- Quick check: 1000
- Publication: 5000-10000
- Critical decisions: 10000+
Q: Fit failed with "ValueError: invalid parameters"
A: Try:
- Check data range matches distribution support
- Remove NaN/Inf values
- Try moments method instead of MLE
- Scale data to reasonable range
Q: All GOF tests reject my fit
A:
- Try different distributions
- Check for outliers (may need removal/robust method)
- Consider mixture distributions
- Data may not follow any standard distribution
Q: Bootstrap CI is very wide
A: This indicates:
- High parameter uncertainty (normal with small samples)
- Poor fit (try different distribution)
- Heavy-tailed data (expected)
Q: Bootstrap is slow
A:
- Use n_jobs=-1 for parallel processing
- Reduce n_bootstrap (1000 is usually enough)
- Use parametric instead of non-parametric
Q: Can I use GPU?
A: Not currently. Planned for v2.0.
Q: What weights should I use?
A:
- Survey: sampling weights
- Precision: 1/variance
- Frequency: counts
Q: Weighted fit gives different results than unweighted
A: This is expected! Weights change the emphasis on different observations.