Pith. sign in

REVIEW 2 major objections 4 minor 5 references

K-Means Panel Data Clustering in the Presence of Small Groups

T0 review · 2 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper shows K-means panel clustering can recover tiny groups, provided the time dimension grows fast enough; standard model-count rules fail in this setting, and new MIC rules fix them.

desk verdict Solid theoretical extension of Bonhomme–Manresa to small groups; the MIC part is a heuristic with no consistency proof and thin simulations, but the inconsistency results and rate conditions are the real contribution. read the letter →

arxiv 2508.15408 v1 pith:5SOWSZLP submitted 2025-08-21 econ.EM stat.ME

Mikihito Nishi This is my paper · ORCID
classification econ.EMstat.ME
keywords paneldatagroupstructuresmallgroupsK-meansclusteringinformationcriteriongroupedfixedeffectsasymptoticnormalitymodelselection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Relaxing the usual assumption that every latent group is proportional to N, this paper considers panels where some groups grow like N^alpha with alpha < 1. It proves the least-squares (K-means) estimators of slopes, grouped fixed effects, and memberships remain consistent and asymptotically normal only if the time dimension T grows fast enough relative to the smallest group's size. It also shows that standard information criteria for the number of groups—Bai and Ng's and the usual BIC—can be inconsistent or unreliable when small groups are present. The proposed modified information criteria (MIC) avoid both missing small groups and over-splitting. This matters because real panel applications often produce small groups, and the paper's empirical example finds a three-firm group that standard criteria miss.

What carries the argument

The central object is the group-size exponent alpha_k, with N_k = tau_k N^{alpha_k}, and the constructed partition Gamma_N(K,m) from Definition 1. For any K < K0, Gamma_N(K,m) keeps the large groups nearly intact and pools the small groups; Lemma A.4 shows its least-squares fit differs from the true K0-group fit by only Op(N^{alpha_{m+1}-1}). This vanishing gap is why standard penalties fail: they cannot detect a misspecified model whose fit is already essentially optimal. In the grouped-fixed-effects model, the same construction uses the matrix D_{T,i}, and the effective penalty becomes T h_{NT}, which drives the different proposed penalty scale.

What would settle it

Simulate a panel with N = 200, T = 10, a small group of size about N^{0.3} (roughly 5 units), well-separated slope parameters, and group sizes otherwise proportional to N. If the proposed MIC selects more than the true K0 groups, or if the K-means membership error rate fails to vanish as N grows with T fixed, the central claim that T must grow like N^{1-2 alpha} would be contradicted. More directly, run the same DGP with T = c N^{0.8} versus T = c N^{1.2} for alpha = 0.3; the theory predicts the first fails to classify the small group perfectly in large samples, while the second succeeds.

Watch

Extended reading notes

Core claim

The central claim is that small groups are not a nuisance: they change the rate conditions for inference. Writing true group sizes as N_k = tau_k N^{alpha_k}, with alpha=1 for large groups and alpha<1 for small ones, the K-means least-squares estimators of slopes, grouped fixed effects, and memberships remain consistent and asymptotically normal provided the sample length T grows faster than N^{1-2 alpha_{K0}} (or faster than a power of N when alpha_{K0} >= 1/2) for the smallest group. Smaller groups demand longer panels. For selection, if N^{1-alpha_{m+1}} h_{NT} -> infinity then any information criterion underestimates the number of groups; Bai and Ng's penalty satisfies this when N/T -> i

Load-bearing premise

The load-bearing premise is Assumption 2(g): the time-series length T must grow faster than N (or faster than N^{1-2 alpha}, where alpha is the exponent of the smallest group's size); if the panel is short relative to the number of units, the proofs of membership consistency and asymptotic normality collapse, and the paper itself flags this as a limitation.

Editorial extensions

If this is right

  • If the central claim is correct, then in panels with small groups, adding more cross-sections N with fixed T does not improve—and can worsen—classification accuracy and parameter estimation for the small groups; only a longer T helps.
  • Bai and Ng's information criterion can silently miss rare groups whenever N is large relative to T, and in the grouped-fixed-effects model it underestimates the number of groups even when all groups are large.
  • The BIC-type penalty overestimates the number of groups in finite samples in models without grouped fixed effects, so its use in that setting can create spurious groups.
  • The proposed MIC penalties (h_MIC1 without GFE, h_MIC2 with GFE) avoid both underestimation and overestimation in simulations, and in the empirical application they detect a 3-firm metal group that Bai and Ng misses while keeping the total group count at a parsimonious three.
  • The paper's results imply that reports of a group structure should always be accompanied by a check of whether T is large enough relative to the smallest group's size.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • One editorial extension: the rate N^{1-2 alpha} acts as an effective sample size per small-group parameter, so a practitioner could estimate alpha (e.g., from the smallest detected group) and check whether their T exceeds the required growth rate; the theory predicts a sharp threshold where classification suddenly becomes reliable.
  • A second extension: the under-penalization failure of Bai and Ng's criterion likely transfers to other clustering settings beyond panels—any partition method whose penalty decays like (ln N)/N will drop rare subpopulations when N dominates sample length; the MIC principle (penalty no faster than (ln N)/N when N<=T, and (ln N)/N scaled by 1/T with GFE) may be a portable fix.
  • A testable extension: run the Monte Carlo with T set exactly at the boundary T = c N^{1-2 alpha} for a small group with alpha = 0.3; the theory predicts that slightly below this boundary the membership error rate fails to vanish, while slightly above it succeeds—this step-like transition can be verified directly.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper studies K-means / least-squares clustering in panel data models when some latent groups are 'small,' i.e., of size N^{α_k} with α_k < 1. It makes three contributions. First, it derives sufficient conditions under which the LS slope and grouped fixed-effects estimators are consistent and asymptotically normal (Theorems 1-2), with the key rate condition that T must grow fast enough relative to N and the smallest group size. Second, it derives sufficient conditions for information criteria to under-estimate the number of groups (Propositions 1-2) and shows that the Bai-Ng penalty can satisfy these conditions, while the BIC penalty over-estimates in finite samples in models without grouped fixed effects. Third, it proposes modified information criteria (h_MIC1, h_MIC2) designed to avoid the identified under-estimation condition. The paper reports a 100-replication Monte Carlo study and an empirical application to Japanese firms' sales growth. The paper explicitly states in Remark 3 that no consistency result is proved for the proposed MIC.

Significance. The theoretical parts of the paper, if correct, fill a real gap: most clustered panel asymptotic theory assumes all group sizes are proportional to N. The paper shows the consequences of relaxing this assumption, and the proofs in the appendix are detailed and self-contained. The characterization of when Bai-Ng and BIC-type penalties fail is also of practical value. However, the proposed MIC is the least theoretically supported contribution: the design rules only ensure that a particular sufficient condition for under-estimation is avoided, and the finite-sample evidence is based on 100 replications without standard errors or sensitivity analysis. The paper is therefore more significant for its theoretical characterization of small-group asymptotics than for the specific new criterion as currently justified.

major comments (2)
  1. [§3.4, Remark 3] The paper's third contribution is the MIC, and the abstract credits it with allowing one to 'discover small groups without producing too many groups.' Yet no consistency or selection-rate theorem is proved for h_MIC1 or h_MIC2. The design rules (i)-(iv) ensure that condition (4) holds and that the sufficient under-estimation condition (5) fails, but failure of a sufficient condition is not a proof of correct selection; other channels of inconsistency remain possible. Remark 3 concedes this. This is load-bearing: if the only formal property is 'not known to under-estimate through one specific mechanism,' the abstract's promise is not backed by theory. I would ask for either a consistency result under additional conditions (e.g., a lower bound on the population gap between the true and under-fitted models), or a clear reframing of the MIC as a heuristic accompanied by a substantially stron
  2. [§5.1, Figures 1-3, footnote 8] The finite-sample evidence for the MIC is thin. The Monte Carlo uses only 100 replications and reports only means, with no standard errors, confidence bands, or full distributions. The penalty constant 0.5 in h_MIC1 is chosen in part because of the simulation results, and footnote 8 explicitly says the sensitivity of the scaling constant is not explored. The design of the simulation also motivated the penalty choice, so the evidence is partly in-sample. Please report Monte Carlo standard errors (or interquartile ranges), increase the number of replications, and provide a sensitivity analysis over the scaling constant (e.g., 0.25, 0.5, 1, 2) and over T/N ratios. Without this, the claim of 'good performance' is not robustly documented.
minor comments (4)
  1. [§5, paragraph before §5.1] The text says that for DGPs 1 and 3 a within-transformation is applied to demonstrate robustness to individual fixed effects. DGP 3 is a grouped fixed-effects model, not an individual fixed-effects model. Please clarify exactly what transformation is applied and why it is compatible with the theoretical assumptions, since within-transformation does not eliminate group-time effects.
  2. [Table 3] Group 3 in the empirical application contains only 3 firms. The reported significance stars for the slope estimates in this group should be interpreted with great caution; a sentence acknowledging the extremely small group size as a limitation would be appropriate.
  3. [§3.4, footnote 8] The sentence 'we do not explore in this direction' should be moved or expanded in the main text, because the arbitrary scaling constant is a central concern for a practitioner using the MIC.
  4. [General] The paper would benefit from a data and code availability statement, especially because the simulation study is used to support the main practical recommendation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all claimed results are derived from stated assumptions; MIC is an explicit design proposal with acknowledged lack of consistency proof.

full rationale

The paper's derivations are self-contained. Theorems 1-2 prove consistency and asymptotic normality from Assumptions 1-3 / 1'-3' by standard least-squares arguments; the key rate condition Assumption 2(g) is a sufficient condition used in the proofs, not a restatement of the conclusions. Proposition 1 derives sufficient conditions (4) and (5) for no-overestimation and underestimation directly from the SSR decomposition in (6); the behavior of the Bai-Ng and BIC penalties under (5)/(11) is then a calculation, not a fit. The MIC in Section 3.4 is explicitly constructed to satisfy (i)-(iv), which are consequences of (4) and (5); Remark 3 candidly states 'we do not prove the consistency of K-hat(h_MIC1_NT),' so the paper does not present a consistency theorem for MIC as a derived prediction. Its finite-sample and empirical support are external to the theoretical derivation. The simulation's 100 replications, absence of standard errors, and the unexplored 0.5 scaling constant (footnote 8) are robustness/evidence concerns, not circularity. No load-bearing step reduces an output to an input by construction, and there are no self-citations used as evidence.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the small-group size model and on the rate and separation assumptions. No invented entities are introduced, and the only hand-chosen numbers are the irrelevant asymptotic scaling constants in the proposed penalties.

free parameters (1)
  • MIC penalty scaling constants (0.5 in h_MIC1, 2 in h_MIC2) = 0.5 and 2
    Chosen by hand to preserve continuity and to match the relative behavior of h_BN and h_BIC in different N/T regimes. Asymptotically irrelevant but may affect finite-sample performance.
assumptions (4)
  • domain assumption Assumption 1: group sizes follow N_k^0 = tau_k N^{alpha_k} with at least one group proportional to N (alpha = 1), and for GFE all groups diverge (alpha > 0).
    Defines the small-group setting used in all theorems and propositions.
  • domain assumption Assumption 2(g): T must grow faster than N or faster than N^{1-2alpha_K0} depending on the smallest group exponent.
    Necessary for the op(1) misclassification probabilities and for the asymptotic normality claims. The paper itself identifies this as a limitation in short panels.
  • domain assumption Assumption 2(h): minimal group separation in slope parameters, and for GFE also in group fixed effects.
    Ensures groups are identifiable and is standard in the clustering literature.
  • domain assumption Assumption 3: high-level CLT and variance conditions for oracle estimators.
    Standard sufficient conditions used to establish asymptotic normality of the oracle LS estimators.

how reviews work

0 comments
Cite this review

Pith. "Pith review of K-Means Panel Data Clustering in the Presence of Small Groups." pith.science (2026). https://pith.science/paper/5SOWSZLP

@misc{pith2026250815408,
  author       = {Pith},
  title        = {Pith review of: K-Means Panel Data Clustering in the Presence of Small Groups},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5SOWSZLP}},
  note         = {Machine review of arXiv:2508.15408}
}
read the original abstract

We consider panel data models with group structure. We study the asymptotic behavior of least-squares estimators and information criterion for the number of groups, allowing for the presence of small groups that have an asymptotically negligible relative size. Our contributions are threefold. First, we derive sufficient conditions under which the least-squares estimators are consistent and asymptotically normal. One of the conditions implies that a longer sample period is required as there are smaller groups. Second, we show that information criteria for the number of groups proposed in earlier works can be inconsistent or perform poorly in the presence of small groups. Third, we propose modified information criteria (MIC) designed to perform well in the presence of small groups. A Monte Carlo simulation confirms their good performance in finite samples. An empirical application illustrates that K-means clustering paired with the proposed MIC allows one to discover small groups without producing too many groups. This enables characterizing small groups and differentiating them from the other large groups in a parsimonious group structure.

Figures

Figures reproduced from arXiv: 2508.15408 by the authors.

Figure 1
Figure 1. Mean of K b (hNT ) over 100 replications, DGP 1 : N = 60, : N = 90, : N = 120, : T = 10, : T = 20 28 [PITH_FULL_IMAGE:figures/full_fig_p029_1.png] view at source ↗
Figure 2
Figure 2. Mean of K b (hNT ) over 100 replications, DGP 2 : N = 60, : N = 90, : N = 120, : T = 10, : T = 20 29 [PITH_FULL_IMAGE:figures/full_fig_p030_2.png] view at source ↗
Figure 3
Figure 3. Mean of Kb(hNT ) over 100 replications, DGP 3 : N = 60, : N = 90, : N = 120, : T = 10, : T = 20 30 [PITH_FULL_IMAGE:figures/full_fig_p031_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: PPC over 100 replications : N = 60, : N = 90, : N = 120, : T = 10, : T = 20 32 [PITH_FULL_IMAGE:figures/full_fig_p033_5.png]
Figure 6
Figure 6. Figure 6: Plots of the group averages of variables [PITH_FULL_IMAGE:figures/full_fig_p034_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages

  1. [1]

    Ando, T. and J. Bai (2017) Clustering Huge Number of Financia l Time Series: A Panel Data Approach With High-Dimensional Predictors and Factor Structures. Journal of the American Statistical Association , 112 (519), 1182–1198. Aparicio-P´ erez, D. and J. Ripoll´ es (2025) Disentanglingthe Heterogeneous Effect of Natural Resources on Economic Growth. Econom...

    work page 2017

  2. [4]

    The proof closely follows that of Lemma S.4 of Bonhomme and Ma nresa (2015)

    Proof. The proof closely follows that of Lemma S.4 of Bonhomme and Ma nresa (2015). 34 Using Assumption 2(e), we have ˜Q (θ, γ N ) − ˜Q ( θ 0, γ 0 N ) = 1 N N∑ i=1 ( θ0 k0 i − θki )′ 1 T T∑ t=1 xitx′ it ( θ0 k0 i − θki ) = K 0 ∑ k=1 N 0 k N K 0 ∑ ˜k=1 ( θ0 k − θ˜ k )′ { 1 N 0 k N∑ i=1 1{k0 i = k}1{ki =˜ k} 1 T T∑ t=1 xitx′ it } ( θ0 k − θ˜ k ) ≥ K 0 ∑ k=1...

    work page 2015

  3. [5]

    The proof closely follows that of Lemma S.4 of Bonhomme and Ma nresa (2015)

    Proof. The proof closely follows that of Lemma S.4 of Bonhomme and Ma nresa (2015). Using Assumption 2’(e’), we obtain ˜Q (θ, µ , γ N ) − ˜Q ( θ 0, µ 0, γ 0 N ) = K 0 ∑ k=1 K 0 ∑ ˜ k=1 1N N∑ i=1 1{k0 i = k}1{ki =˜ k} 1 T T∑ t=1 {( θ0 k − θ˜ k )′ xitx′ it ( θ0 k − θ˜ k ) + 2 ( θ0 k − θ˜ k )′ xit(µkt − µ˜ kt) + (µkt − µ˜ kt)2 } ≥ K 0 ∑ k=1 N 0 k N K 0 ∑ ˜k=...

    work page 2015

  4. [188]

    Shi, and P

    Su, L., Z. Shi, and P. C. B. Phillips (2016) Identifying Laten t Structures in Panel Data. Econometrica, 84 (6), 2215–2264. Sun, Y., L. Tan, W. Zhang, and Z. Zhu (2025) Homogeneity Pursu it in Clustered Data Analysis When Cluster Sizes Are Small*. Journal of Business & Economic Statistics , 0 (ja), 1–24. Wang, W. and Z. Zhu (2024) Homogeneity and Sparsity...

    work page 2016

  5. [1279]

    Bai, J. and S. Ng (2002) Determining the Number of Factors in A pproximate Factor Models. Econometrica, 70 (1), 191–221. Bonhomme, S. and E. Manresa (2015) Grouped Patterns of Heter ogeneity in Panel Data. Econometrica, 83 (3), 1147–1184. Chetverikov, D. and E. Manresa (2022) Spectral and Post-Spe ctral Estimators for Grouped Panel Data Models. https://ar...

    work page Pith review arXiv 2002

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.