Skip to content

Commit

Permalink
Update README.md
Browse files Browse the repository at this point in the history
Updates to landing page linking to new Finite Sample Bias vignette.
  • Loading branch information
prodriguezsosa authored Aug 4, 2023
1 parent a5028e5 commit a629ad7
Showing 1 changed file with 2 additions and 6 deletions.
8 changes: 2 additions & 6 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -20,12 +20,8 @@ To use **conText** you will need three objects:

# Quick Start Guides

Check out this [Quick Start Guide](https://github.com/prodriguezsosa/conText/blob/master/vignettes/quickstart.md) to get going with `conText`.
Check out this [Quick Start Guide](https://github.com/prodriguezsosa/conText/blob/master/vignettes/quickstart.md) to get going with `conText` (last updated: 08/04/2023).

# Latest Updates

We'd like to thank Will Hobbs and Breanna Green for bringing to our attention implementation issues with our main estimation routine, `conText`, when samples sizes are small and/or class-ratios are highly skewed towards one group. We are actively collaborating in evaluating alternative re-sampling methods to correct for the finite sample bias associated with `conText`'s coefficient norm estimates. In the meantime, we've pushed an alternative version of the `conText` function, `conText_jackknife` that implements Jackknife debiasing. To illustrate the advantages of using Jackknife, below we showcase the results of simulations using real data where we know the true value of coefficient norm, varying both sample size and class ratios.

We observe that using OLS with bootstrapped standard errors can be problematic (i.e. resulting in a notable bias) for cases with small samples and highly skewed class ratios, especially when the true norm of the coefficient is 0. Jackknife debiasing clearly helps mitigate this bias. Nevertheless we still strongly recommend users: (1) avoid reading too much into results with highly skewed class ratios (2) benchmark their results against a random group of the same size as their group of interest and (3) always follow up any regression-based results with a qualitative comparison of nearest neighbors.

<img width="1461" alt="Screenshot 2023-07-07 at 5 39 26 PM" src="https://github.com/prodriguezsosa/conText/assets/6556873/84fe373a-d9a0-4438-be8b-d9dc84596006">
We are hugely thankful to Will Hobbs and Breanna Green for bringing to our attention clear examples where finite sample bias was larger than we had anticipated when implementing our main estimation routine, `conText`. We are actively collaborating with them to evaluate alternative fixes. In the meantime we've implemented and recommend using Jackknife debiasing. Please refer to the [Finite Sample Bias](https://github.com/prodriguezsosa/conText/blob/master/vignettes/quickstart.md) vignette for additional information on the issue and simulation results using various debiasing methods.

0 comments on commit a629ad7

Please sign in to comment.