Daily Summary for 2026-10-21

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: It's the twenty-first of October, twenty twenty-six, and this is the day's research.

Jane: 6 new papers came out today.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We'll take the day in one pass, then pull out the papers we're staying with.

Jane: Start us off with Computing the final epidemic size distributions of a multi-type Galton--Watson process.

The day's research: Tom: We are finding hidden rules in noisy systems because reality is stochastic rather than deterministic.

Jane: The main focus is recovering drift and diffusion terms from one trajectory using AKMPC.

Lu: AKMPC swaps time-indexed test functions for state-dependent spatial kernels to avoid biases.

Meng: This avoids the persistent endogeneity bias seen with fixed-width time windows in AKMPC.

Lalam: Variable selection stability is important because it addresses Lasso's weakness with correlated variables.

Tom: That finding extends beyond high-dimensional data to lower dimensions as well, according to Stability Selection via Variable Decorrelation.

Jane: This concept of improving stability by decorrelating variables relates to A Stable Lasso.

Lu: A Stable Lasso proposes integrating a weighting scheme into the penalty function based on correlation-adjusted rankings.

Tom: A Stable Lasso integrates weighting into the penalty function using correlation-adjusted ranking to stabilize variable selection when predictors are correlated.

Jane: That addresses instability during variable selection, but how does that weighting scheme specifically interact with the penalty structure?

Lu: The rationale is to maintain stability even when predictors exhibit high correlation structures within the model.

Meng: Regarding modeling, we are focusing on inferring parameters in complex epidemiological data using multi-type Galton--Watson processes.

Lalam: Computing the final epidemic size distributions of a multi-type Galton--Watson process offers a framework for understanding transmission heterogeneity.

Tom: That framework helps us understand how transmission varies across different types in real-world scenarios such as Middle East respiratory syndrome.

Lucky paper: 2606.03007: Tom: So, this paper is about using Cauchy integral contours to compute the final size distribution of multi-type Galton--Watson processes.

Jane: That sounds like a complicated way to find the expected outcome of an epidemic when you have different types of hosts.

Lu: The core idea is using these contours, specifically choosing radii r = (r one r n), to guarantee the existence and uniqueness of the integrand.

Meng: But how does that choice relate to the process itself? Are we setting those radii based on some underlying dynamics?

Lalam: The paper specifies that for subcritical or critical processes, they use a vector r = (zero point nine five, zero point nine five).

Tom: And for supercritical processes, they find a vector r where the spectral radius of the Jacobian matrix of G(z) evaluated at z equals one is less than or equal to zero point nine seven five.

Jane: That threshold of zero point nine seven five seems like a critical boundary for determining if an epidemic will go extinct or grow indefinitely.

Lu: Exactly, and they show that this framework works whether you are dealing with subcritical, critical, or supercritical processes all in one system.

Meng: Numerically, they use the discrete Fourier transform on multidimensional grids to calculate the coefficients of H(i) from a multivariate Cauchy integral representation.

Tom: The numerical procedure involves solving for h in the system Fz(h) equals h minus diag(z)G(h) equals zero using Newton’s method with backtracking, minimizing psi z (h).

Jane: So if we look at the results they got, they applied this to Middle East respiratory syndrome cluster size data.

Lalam: They estimated parameters like R equals zero point seven seven and k equals zero point three zero using Markov Chain Monte Carlo methods for that specific case study.

Tom: That likelihood surface is described as being smoothly connected across the criticality threshold where log(R) equals zero which suggests a trade-off between mean transmission and dispersion in the supercritical region.

Lu: It’s interesting because it allows parameter inference without imposing an assumption about whether the underlying Galton--Watson process is critical or not from the start.

Meng: The limitation they pointed out is that as you increase the number of host types, say to a higher dimension, computing these integrals efficiently becomes a real challenge.

Jane: So while it gives us a framework for inference, scaling this up to many different types seems like the next hurdle for this method.

Lucky paper: 2505.20864: Tom: So we are looking at Stability Selection via Variable Decorrelation, and the core idea is that decorrelating variables before applying Lasso actually improves stability across different settings and selection techniques.

Jane: That sounds like a big improvement over just running standard variable selection, but how exactly does this pre-processing step work mechanically?

Lu: The methodology involves three key steps: first you reorder the variables, then you apply Gram-Schmidt orthogonalization to project them onto a new space of orthonormal vectors, and finally you perform variable selection on that resulting matrix Q.

Meng: So Q becomes your surrogate for X in the variable selection results because it retains the component of Xj contributing uniquely to its span when variables are correlated.

Tom: That's where I get excited; using Q as a surrogate for X when there's correlation is a clever way to handle that interdependence.

Jane: The paper claims this pre-processing satisfies the irrepresentable condition for Lasso under two specific assumptions, which they show in Corollary one.

Lu: Corollary one shows that if the columns of the design matrix X aren't linearly dependent and if your variable ordering places relevant variables earlier than correlated irrelevant ones, then Q satisfies that condition.

Tom: So it’s a conditional guarantee based on how you order things relative to each other in the design matrix.

Meng: And when you have two highly correlated and equally relevant variables, Corollary two suggests applying stability selection with Lasso to Q results in higher selection stability compared to applying it directly to X.

Jane: That's a concrete result for handling those tricky situations where variables are very similar in their predictive power.

Lalam: From my perspective, the numerical results are interesting because Figure 1a shows the pre-processing significantly enhances stability in simulation scenarios.

Tom: And they also show that even when covariances have reversed signs, like in Figure 3a and 3b, both stability and accuracy improvements are maintained.

Jane: That suggests this method is robust to sign flips in the correlation structure, which is something we often worry about when dealing with real data.

Lu: Furthermore, they found that this method outperforms other candidates like Elastic Net, SCAD, MCP, and Sparse Group Lasso in terms of selection stability.

Meng: It also works well even in low-dimensional settings because Corollary two demonstrates that the improvement in stability holds there too.

Tom: That’s good news for models where we don't have massive amounts of data to worry about high dimensionality issues.

Jane: The paper does mention future work suggesting further exploration into non-linear settings, which is a big area for research.

Lu: They hypothesize that this approach could also be effective for non-regularization-based methods like filter or wrapper methods in those complex situations.

Meng: We need to think about the computational cost of the Gram-Schmidt orthogonalization and projection steps before we can deploy this at scale in production.

Tom: That’s a fair point; we have to check the runtime of Algorithm two before we can use it on our actual traffic data.

Jane: The authors did mention that the Riboflavin dataset is available via the hdi package in R, which is helpful for checking out their simulations.

Lalam: The availability of the source code through GitHub allows others to check if they can reproduce these results using the DVS R package as well.

Lucky paper: 2503.00326: Tom: We're looking at a paper called A Bayesian Additive Regression Tree Model for Learning Conditional Average Treatment Effects in Regression Discontinuity Designs.

Jane: This is about causal inference, specifically how to estimate conditional average treatment effects or CATE when you have a regression discontinuity design.

Lu: The authors are using a variant of the BART model, which is an additive regression tree ensemble, with linear leaf-level regressions on the running variable and a treatment dummy.

Meng: So they're adapting BART to partition covariate space into regions where the slope on the running variable changes significantly?

Tom: Exactly. They are using this structure because standard methods often assume a fixed basis expansion, which is too restrictive for real data.

Jane: And how do they actually calculate that conditional average treatment effect at the cutoff point?

Lu: The model defines the prediction function for tree j as gj(x, w, z) equals psi(x, z) gamma bj (x w). This means the CATE at x equals zero is a sum of the delta b j zero w elements across all trees.

Tom: That summation structure is key because it regularizes the treatment effect through priors on those delta coefficients.

Jane: So they are using those priors to stabilize that estimation process, which is what I was wondering about with Lasso stability.

Lalam: The paper also describes a simulation protocol where the evaluation criteria is the average root-mean-squared estimation error expressed as a fraction of a default ATE estimator.

Meng: Lower RMSE than other estimators on average is one thing, but how does this compare to what we see when deploying models in production?

Tom: The results show that BARDDT produces lower RMSE than the other estimators on average and is most often better than the homogeneous estimator.

Jane: It sounds like it's more robust to the data structure than previous methods, which is a big deal for complex real-world datasets.

Lu: They also found that course load, which we call credits attempted in the academic term leading to probation, is a strong moderator for this CATE estimation.

Meng: That means we might need to include that feature when building our next deployment pipeline if we want good results.

Lalam: The paper also demonstrates the ability of BARDDT to handle more features because the correlation between their CATE predictions and each of the null features is effectively zero.

Lucky paper: 2411.09097: Tom: So we’re looking at On the Selection Stability of Stability Selection and Its Applications, which is paper number four of five this day.

Jane: It seems like the core idea here is shifting from just checking individual variables to evaluating the overall stability of all those variable selection results.

Lu: Exactly, they introduce a 'stable stability selection' methodology to identify variables that have highly consistent results across different resamples.

Meng: Can you tell us more about that stability estimator (M(lambda)) ? What exactly are we looking at there?

Tom: It's defined as one minus one over p times the sum over all j, squared, of the ratio of the unbiased sample variance of binary selection statuses for the j-th variable to one minus the average number selected under a regularization parameter lambda.

Jane: That sounds really dense. So they are using that specific estimator to define a stability measure for their results.

Lu: Yes, they bound this stability estimator between negative one over B and one, where B is some constant related to the setup.

Meng: And then they propose an optimal regularization value lambda stable as the smallest regularization value at which this stability measure surpasses zero point seven five.

Tom: That gives us a concrete threshold for tuning Lasso or any of these methods, which is important for practitioners who want high stability.

Jane: If that lambda stable doesn't exist, they use a strategy called lambda one se, defined using the maximum value in the set of that stability measure minus its standard deviation.

Lu: That lambda one se strategy uses the maximum lambda in multiplied by the stability measure minus its standard deviation, which is a technique introduced earlier with Hastie et al. in two thousand nine.

Meng: The paper evaluates this on synthetic data, and they show that transitioning from that lambda one se to lambda stable improves stability significantly without much loss in accuracy across all three correlation scenarios.

Tom: That's a big deal for model deployment; it suggests we can find a better tuning point.

Jane: But they also tested this on real datasets like the Riboflavin data and Affymetrix Rat Genome two hundred thirty two point zero Array microarray data where lambda stable didn't actually exist.

Lu: In those real cases, they used lambda stable-one sd, and they found selection frequencies above zero point five for certain genes in some instances.

Meng: That finding on the real data is interesting because it shows that even when the perfect stability point isn't there, we can still get decent results using this adapted method.

Tom: It also points out that as data complexity increases, these stability values show a significant decline, which is a useful warning sign for us.

Jane: So what does this mean for someone just listening to the show who isn't deep into variable selection? It means when you have messy data and you use Lasso, you need to look beyond just one metric.

Lu: It means that focusing on the overall stability of those selections tells us more than traditional methods when the data complexity is high.

Meng: For production, it suggests that this approach could help us calibrate our regularization parameters more reliably before we commit to a final model run.

Tom: It gives us a defined path for tuning hyper-parameters based on stability, which is something we've been missing in some of our setups.

Jane: So the conclusion is that lambda stable represents the point where the model hits high stability with minimal loss in prediction accuracy, offering a better trade-off.

Lucky paper: 2511.02306: Tom: So we’re talking about A Stable Lasso, which is trying to fix the stability issues Lasso has when predictors are correlated.

Jane: The paper focuses on how integrating a weighting scheme into the penalty function can improve that selection stability.

Lu: They define these weights as an increasing function of a correlation-adjusted ranking based on Air-HOLP.

Meng: That sounds like it's trying to give the Lasso more prior information about which variables are actually useful within correlated groups.

Lalam: The methodology involves ranking predictors first, then defining the weight function as wj r j equals one minus one over r j for every predictor j from one to p.

Tom: So it's basically prioritizing variables based on how predictive they are relative to their neighbors in the correlation structure.

Jane: And they embed this within the Stability Selection framework to test stability across random sub-samples and tune the hyperparameter lambda.

Lu: A key finding is that applying a fixed, non-uniform weighting scheme across all sub-samples yields greater stability than the standard equally weighted Lasso.

Meng: That’s interesting because it suggests that a one-size-fits-all weighting might actually be better than a uniform approach in this context.

Tom: In their empirical evaluations, they looked at synthetic data with n equals one hundred and p equals one thousand variables with strong correlations across five groups.

Jane: And in the real world test on a mouse PCR dataset with n equals sixty and p equals twenty-two thousand five hundred seventy-five genes, Stable Lasso got a maximum selection stability of zero point eight five compared to the standard Lasso's maximum stability of zero point zero five four.

Lu: That jump from zero point zero five four to zero point eight five is significant, showing how much the weighting scheme helps when correlations are up to zero point eight in that real dataset structure.

Meng: It also showed that Stable Lasso achieved the highest average F1-score across all decision thresholds when tested under less extreme correlation structures, up to zero point eight.

Tom: So, it’s not just about stability; they also found better prediction accuracy when correlations are moderate.

Jane: The conclusion is that the Stable Lasso is a more stable variant of the Lasso and a generalized form of the Adaptive Lasso because it shows enhanced stability and accuracy in complex settings.