📊 statistics

Infer-and-widen, or not?

This paper demonstrates that in various selective inference scenarios, the widely used "infer-and-widen" framework produces confidence intervals that are unnecessarily wider than alternative approaches, even when comparing against the theoretically narrowest possible "infer-and-widen" interval.

Original authors: Ronan Perry, Zichun Xu, Olivia McGough, Daniela Witten

Published 2026-02-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ronan Perry, Zichun Xu, Olivia McGough, Daniela Witten

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery in a city with 1,000 suspects. You have a hunch, but you don't know who the criminal is. So, you look at all the evidence and pick the one suspect who looks the most guilty. Let's call him "Mr. Winner."

Now, you want to build a case against Mr. Winner. You want to say, "I am 95% sure Mr. Winner is the criminal, and here is the range of evidence that proves it."

This is the core problem of Selective Inference. The tricky part is that you didn't pick Mr. Winner randomly; you picked him because the data made him look guilty. If you use standard statistical rules (which assume you picked a suspect before looking at the evidence), you will likely be wrong. This is known as the "Winner's Curse." You picked the winner because they were lucky, not necessarily because they are the best.

The Old Way: "Infer-and-Widen"

For a long time, statisticians have used a strategy the authors call "Infer-and-Widen."

Think of it like this: You know that picking the "most guilty" suspect introduces a huge bias (you might be picking an innocent person just because they got lucky). So, the old strategy says:

"Okay, we know our guess is biased. Let's just make our 'range of evidence' (confidence interval) massive to cover our mistakes. We'll widen the net so wide that it definitely catches the truth, even if our guess is off."

The authors compare this to a fisherman who knows his net is full of holes. Instead of fixing the net, he just makes the net so huge that he's sure to catch a fish eventually.

The Problem: This "Infer-and-Widen" approach creates intervals that are painfully wide. They are so wide they aren't very useful. It's like saying, "The criminal is somewhere between the North Pole and the South Pole." Technically, you might be right, but you haven't learned anything useful.

The New Idea: "Fix the Center, Don't Just Widen"

The authors of this paper argue: "Why are we just making the net bigger? Why don't we fix the net itself?"

They propose alternative methods (like Data Fission or Randomized Conditional Inference) that don't just widen the interval. Instead, they:

Acknowledge the bias: They realize the "center" of their guess is skewed.
Shift the center: They adjust the math so the guess is actually centered on the truth.
Keep the net tight: Because the guess is now accurate, they don't need a massive net. They can use a small, precise net.

The Three "Vignettes" (Case Studies)

The authors tested their idea in three different scenarios to see if "Infer-and-Widen" is really the best we can do.

1. The Winner's Curse (The Lottery)

Scenario: You have 100 lottery tickets. You pick the one with the highest number. Is that number a good estimate of the "true" average? No, it's inflated.
The Test: They compared the "Infer-and-Widen" method (making the interval huge) against a new method that splits the data to correct the bias.
Result: The new method produced a much narrower, more useful interval. Even a theoretical "perfect" Infer-and-Widen interval (one that knows the answer but still uses the old method) was wider than the new method.

2. Maximal Contrasts (The Best Athlete)

Scenario: You have 100 athletes. You pick the one who ran the fastest on that specific day. You want to estimate their true speed.
The Test: Again, the "Infer-and-Widen" method made the interval too wide. The new method, which accounted for the fact that you picked the fastest runner after seeing the race, gave a much tighter, more accurate range.

3. The Lasso (The Filter)

Scenario: You have a huge list of potential causes for a disease (1,000 variables). You use a computer algorithm (the Lasso) to filter them down to the top 5. You then want to know how strong those 5 are.
The Test: This is a very common problem in modern data science. The "Infer-and-Widen" methods (like the famous Bonferroni correction) are very conservative and produce huge intervals. The new methods (like Hybrid or Conditional inference) produced intervals that were much more precise.

The Big Conclusion

The paper's main message is a wake-up call for statisticians and data scientists:

"Stop just widening the net!"

The "Infer-and-Widen" framework is like using a sledgehammer to crack a nut. It works (you get the right answer), but it's clumsy and inefficient. The authors show that by using smarter, more nuanced methods that correct for how the data was selected, we can get narrower, more precise, and more useful answers without sacrificing accuracy.

In short: Don't just make your safety net bigger; fix the hole in the middle so you don't need such a big net at all.

1. Problem Statement

The paper addresses the challenge of selective inference (or post-selection inference), where the target of inference (e.g., a parameter $\theta_i$ ) is not fixed a priori but is selected from the data using a selection rule $S(Y)$ .

The Core Issue: Classical confidence intervals (CIs) assume a fixed target. When the target is data-driven, standard CIs fail to achieve nominal coverage because the selection process introduces bias (e.g., the "winner's curse").
The Dominant Paradigm: The authors critique the prevailing "infer-and-widen" (I&W) framework. In this approach:
1. A target is selected from the data.
2. "Classical" inference is performed on that target, but the critical values are symmetrically widened (adjusted) to account for the selection bias.
3. Limitation: Because the midpoint of the CI remains the biased point estimate, the interval must be made excessively wide to ensure the true parameter is captured. The authors argue this is an inefficient use of the error budget.

2. Methodology and Framework

The paper evaluates the I&W framework against alternative selective inference methods across three specific vignettes.

The Infer-and-Widen (I&W) Framework

Formalized in Section 2, I&W constructs a CI $CI_{S(Y)}(Y)$ centered at the biased estimator $\hat{\theta}_{S(Y)}$ but with a widened critical value $H(\alpha, S, Y)$ .

Mechanism: It adjusts the significance level $\alpha$ to $\alpha' = H(\alpha, S, Y)$ to guarantee unconditional coverage: $P(\theta_{S(Y)} \notin CI) \leq \alpha$ .
Examples: Bonferroni, Holm-Bonferroni, Scheffé, and modern algorithmic stability approaches (e.g., Zrnic & Jordan, 2023).
Drawback: The midpoint remains biased, forcing the interval to be wide to compensate.

Alternative Approaches

The authors compare I&W against methods that do not rely on symmetric widening:

Data Fission / Data Thinning: Splitting data (or a single observation) into independent training and test sets. The training set selects the target; the test set performs inference on a fixed target, avoiding selection bias in the inference step.
Randomized Conditional Inference: Conditioning on the selection event and auxiliary randomization (e.g., noise added to the selection rule) to derive exact conditional distributions (often truncated normals) for the parameter.
Hybrid Methods: Combining conditional inference (for strong signals) with unconditional methods.

The "Oracle" Benchmark

To determine if the I&W framework is theoretically capable of producing narrow intervals, the authors define an "Oracle I&W" interval. This is the narrowest possible interval centered at the biased estimator that achieves the required coverage. If even the Oracle I&W is wider than alternative methods, the I&W framework is fundamentally limited.

3. Key Contributions and Vignettes

The authors conduct empirical and theoretical evaluations in three vignettes:

Vignette 1: The Winner's Curse

Setup: Selecting the index of the maximum of $n$ Gaussian variables ( $S(Y) = \arg\max Y_i$ ).
Comparison: I&W (based on algorithmic stability with Laplace noise) vs. Data Fission (using the same selection event).
Result: The Data Fission interval is significantly narrower than the I&W interval.
Oracle Finding: Even the Oracle I&W (narrowest possible centered interval) is wider than the Data Fission interval. This suggests that in this setting, no I&W method can outperform the alternative.

Vignette 2: Maximal Contrasts

Setup: Selecting the feature $j$ that maximizes $|X_j^\top Y|$ (maximal contrast).
Comparison: I&W (algorithmic stability) vs. Randomized Conditional Selective Inference (RCSI).
Result: RCSI intervals are generally narrower than I&W intervals.
Oracle Finding: Here, the Oracle I&W is theoretically narrower than RCSI. However, the authors note that no practical method currently exists to construct this Oracle interval. Thus, while I&W could theoretically win, current practical I&W methods lose.

Vignette 3: Inference After the Lasso

Setup: Inference on coefficients selected by the Lasso in a linear model.
Comparison: Various existing methods (Locally Simultaneous Inference, Hybrid, Conditional) vs. Oracle I&W.
Result: The Hybrid method (which uses conditional inference for strong signals) outperforms existing practical I&W methods (like LSI).
Oracle Finding: The Oracle I&W is the narrowest possible interval. However, practical I&W methods are much wider. The Hybrid method outperforms all practical I&W methods.

4. Theoretical Analysis of Bias

In Section 6, the authors provide a theoretical bound on the bias of the I&W midpoint.

Proposition 5: For a selection rule using Gaussian noise, the bias of the I&W midpoint ( $E[\hat{Y}_{S(Y)} - \mu_{S(Y)}]$ ) grows on the order of $\sqrt{\log n}$ .
Implication: As the sample size increases, the bias of the center of the I&W interval increases, necessitating wider intervals to maintain coverage. In contrast, alternative methods (like data thinning) can produce unbiased midpoints.

5. Key Results and Findings

Inefficiency of I&W: In all three vignettes, practical I&W methods produce confidence intervals that are wider than non-I&W alternatives (Data Fission, RCSI, Hybrid).
Theoretical Limits:
- In the "Winner's Curse" setting, the I&W framework is fundamentally flawed; even the theoretical best I&W interval is wider than the alternative.
- In "Maximal Contrasts" and "Lasso," the I&W framework could theoretically be superior (via the Oracle), but no computable method currently achieves this. Practical I&W methods remain inferior.
Bias is the Culprit: The primary reason I&W intervals are wide is that they attempt to correct for selection bias by widening the interval around a biased center, rather than correcting the center itself.

6. Significance and Conclusion

Paradigm Shift: The paper challenges the assumption that "infer-and-widen" is the optimal or only path for selective inference. It argues that methods which avoid the bias in the point estimate (via data splitting, thinning, or conditioning) are often superior.
Practical Guidance: Researchers should prefer methods that yield narrower intervals (like Data Fission or Hybrid approaches) over standard I&W approaches, even if the latter are conceptually simpler.
Future Directions: The paper suggests that further development of I&W methods is unlikely to yield significant gains in the "Winner's Curse" setting and that the focus should shift toward methods that correct the point estimate bias rather than just widening the interval.

Conclusion: The authors conclude that while the I&W framework is conceptually attractive for its simplicity, it often leads to suboptimal (overly wide) confidence intervals. In many settings, alternative strategies that decouple selection from inference or correct the bias directly are statistically more efficient.