bootstrap

[Tommaso Salvitti]

Good afternoon to everybody If I use a bootstrap model with 5000 replications, what percentage of replication can be considered valid? does it have to be above 80%? thanks in advance

[Felix Bittmann]

There is no general rule of thumb, but higher is always better. Do you have an idea why some bootstrap results cannot be estimated?

[Tommaso Salvitti]

Given the small sample size (51 observations) and sparse data (only 4 events for var1= 1; 1 meaning exposure), both standard logistic regression and bootstrap faced significant limitations.The bootstrap method, while improving confidence interval robustness, achieved only 60% valid replicates in the unadjusted analysis and 25% in the adjusted analysi.. with 5000 replication also i see these article "Carvalho, C., & Simoes, J. A. (2016). Improved bootstrap confidence intervals in small samples. Computational Statistics & Data Analysis, 100, 200-212."

[Tommaso Salvitti]

Bootstrap methods: A guide for practitioners and researchers

[Felix Bittmann]

Given these limitations, there is not much you can do. As an alternative, you can try the jackknife. In any case, make sure to report these limitations transparently in your report. Simply state the number of converged bootstrap estimates, this is much more useful than relying on a (potentially arbitrary) cutoff value.

[Joseph Coveney]

Given the small sample size (51 observations) and sparse data (only 4 events for var1= 1; 1 meaning exposure) . . .

I'm curious to see the dataset that's giving you so much trouble expressed here in this thread and in its sister thread. It's only 51 observations long—would you mind posting it using -dataex- (or some other method)?

Among the unaddressed questions members of the list might have in their efforts to assist: you haven't said explicitly what you're actually interested in; with your emphasis on bootstrap, can we infer that it's solely in the confidence interval for the log odds ratio for events if var1=1 (exposure) versus var1 = 0? Does the model that you're trying to fit have other predictors that would complicate things?

. . . also i see these article "Carvalho, C., & Simoes, J. A. (2016). Improved bootstrap confidence intervals in small samples. Computational Statistics & Data Analysis, 100, 200-212."

I cannot find such an article.