pith. sign in

arxiv: 2606.26061 · v1 · pith:HP7BEIFBnew · submitted 2026-06-24 · 📊 stat.ME · math.ST· q-bio.QM· stat.ML· stat.TH

Deviance-style normalization for jointly overdispersed counts

Akshay Balsubramani This is my paper

Pith reviewed 2026-06-25 19:16 UTC · model grok-4.3

classification 📊 stat.ME math.STq-bio.QMstat.MLstat.TH
keywords Dirichlet-multinomialdeviance residualsoverdispersed countscompositional datasparse matricesnormalizationnegative binomial conditioning
0
0 comments X

The pith

Dirichlet-multinomial deviance residuals normalize sparse jointly overdispersed counts while preserving exact zeros and recovering the multinomial limit.

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

The paper develops a residual transform grounded in the Dirichlet-multinomial distribution to handle count matrices that are both sparse and overdispersed, the setting typical of sequencing data. The null model treats each sample's counts as a fixed-total composition whose overdispersion is governed by one scalar concentration parameter, and it arises exactly by conditioning independent negative-binomial feature counts on the observed total. The resulting deviance residuals keep every zero entry exactly zero, cost constant time per nonzero entry, match multinomial residuals on singleton counts, and shrink residuals for repeated counts in proportion to the overdispersion the null allows. The same construction extends to ordered and tree-structured features through generalized and Dirichlet-tree versions, unifying joint and feature-wise count models under one compositional logic.

Core claim

The central claim is that the Dirichlet-multinomial deviance residualization preserves exact sparsity, evaluates in constant time per nonzero entry, agrees with multinomial residuals on singleton counts, shrinks repeated-count residuals according to the overdispersion the null tolerates, and recovers the multinomial residual as the concentration parameter tends to infinity; the same fixed-dispersion comparison principle extends to ordered and tree-structured features via the generalized DM and the Dirichlet-tree multinomial, yielding a single residual family that subsumes joint and feature-wise count nulls.

What carries the argument

The Dirichlet-multinomial deviance residual, obtained by comparing each observed count to its conditional expectation under a fixed-total, single-concentration null formed by conditioning independent negative binomials.

If this is right

  • Sparse count pipelines can apply joint overdispersion correction without densifying the matrix or changing asymptotic cost.
  • The residuals apply unchanged to data whose categories are ordered or arranged in a tree via the generalized and Dirichlet-tree models.
  • Feature-wise negative-binomial models and ordinary multinomial models become limiting cases of one common residual construction.
  • Normalization remains valid on exactly sparse data because no artificial nonzeros are introduced.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The construction supplies a uniform way to produce comparable residuals across experiments that differ only in their total sequencing depth.
  • Downstream tasks such as clustering or differential testing could be rerun with these residuals to test whether joint overdispersion adjustment changes biological conclusions.
  • Analogous conditioning arguments might yield deviance residuals for other families of count distributions that admit a fixed-total interpretation.

Load-bearing premise

Each sample's count vector behaves as independent negative-binomial draws conditioned on their observed total, with one shared scalar concentration governing all overdispersion.

What would settle it

Generate data from independent negative binomials conditioned on observed totals and check whether the proposed residuals leave singleton counts unchanged while shrinking the magnitude of residuals for counts greater than one exactly in line with the chosen concentration value.

Figures

Figures reproduced from arXiv: 2606.26061 by Akshay Balsubramani.

Figure 1
Figure 1. Figure 1: Real-data benchmark summary across five real count regimes. All five datasets are real: PBMC 3k (10x raw UMI), GTEx [21] (a 20,000-cell subsample of the GTEx cross-tissue atlas), negative controls (PBMC 3k restricted to 20 housekeeping genes), scATAC (10x atac_pbmc_500_nextgem peak counts), and CITE-seq (10x pbmc_1k_protein_v3 antibody-capture counts). (a) Improvement in held-out per-observed-count conditi… view at source ↗
Figure 2
Figure 2. Figure 2: The parametric two-group test calibrates only under the joint-overdispersion (Dirichlet–multinomial) null, across every broad cell type of a cross-tissue atlas; the Poisson-family (multinomial, Pearson) null is anti-conservative and the per-gene negative-binomial null intermediate. Left: under a null random split, the false-positive rate of the log￾mean test at nominal α = 0.05 for each count-model normali… view at source ↗
Figure 3
Figure 3. Figure 3: Structured-data benchmarks. (a) Held-out per-sample LL: flat DM (unstructured joint baseline, highlighted) vs. three DTM strategies (independent nodewise αv,0, pooled-by-depth, single global scalar). On the synthetic tree data used here the flat DM is dramatically better: held-out likelihood for every DTM strategy is worse by more than an order of magnitude, the per-node dispersion parameters overfitting b… view at source ↗
Figure 4
Figure 4. Figure 4: Downstream analyses on GTEx. A deep generative baseline, scVI, is included wherever the comparison is commensurable. It attains the strongest embeddings here, but at far higher training cost and at the cost of differential￾expression calibration: scVI inflates the realized false-discovery proportion several-fold (Section B.2.1), whereas the DM stays calibrated. Among the linear-time normalizations the DM i… view at source ↗
Figure 5
Figure 5. Figure 5: Synthetic calibration across the four regimes. (a) Regime 1, well-specified DM: relative RMSE of αˆ0 vs. the true α0, at K = 100 and K = 1000. Recovery is accurate for α0 ≲ 104 ; at α0 = 106 the log-likelihood curvature becomes very flat and identifiability breaks down, as expected from the multinomial limit, and the larger alphabet K = 1000 tightens recovery in the well-identified regime. (b) Regime 1: mu… view at source ↗
Figure 6
Figure 6. Figure 6: The DM deviance-accounting identities hold to machine precision across overdispersed-count in￾stances. Each identity is evaluated on random count matrices sampled from a known generative law; an independent code path computes the cross-checked quantity wherever one exists. (a) Worst-case residual of each closed-form identity over the (α0, K) grid (α0 ∈ {1, 8, 66, 4700}, K ∈ {50, 200, 1000}), on a log axis.… view at source ↗
Figure 7
Figure 7. Figure 7: scIB overall integration score across the three atlases. Deviance-PCA (highlighted) is competitive with the deep methods scVI and Harmony and matches the linear pipelines, while substantially outperforming GLM-PCA. A cross (×) marks the single divergent run (scVI-LD on the lung atlas). Tabula Sapiens Human Lung Immune 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 ROC-AUC DE detection power (ROC-AUC) Ta… view at source ↗
Figure 8
Figure 8. Figure 8: Differential-expression detection power (left) and false-discovery control (right, dashed line at the nominal q = 0.05) across the three atlases. Deviance (highlighted) and the other closed-form residual methods sit at or below the nominal FDP line; scVI and scVI-LD rise far above it. 70 [PITH_FULL_IMAGE:figures/full_fig_p070_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Residual-form diagnostic (not a calibration result). A naive unit-variance two-group t-test run directly on each normalization’s residual has a false-positive rate fixed by the residual’s global variance alone: every cell-type point falls on 2 Φ(−1.96/σ). The signed-root deviance residual is over-shrunk (conservative), the Pearson-family residuals are heavy￾tailed (anti-conservative), the per-gene negative… view at source ↗
Figure 10
Figure 10. Figure 10: The concentration α0 behaves as a one-parameter regularizer, and a single scalar suffices. (a) Held-out con￾ditional Dirichlet–multinomial log-likelihood per count versus log10 α0 on the blood data and within a representative atlas stratum; the curve is unimodal with an interior optimum (ring) close to the moment estimate αˆ0 (dashed). (b) Label￾transfer accuracy of a nearest-neighbor classifier on a low-… view at source ↗
Figure 11
Figure 11. Figure 11: An exact evidence unit for the Dirichlet–multinomial contrast. The null is fitted on a peripheral-blood single￾cell split (K = 200 genes, n = 2700 cells, moment concentration αˆ0 = 524) and the calibration panels draw Nsim = 1000 synthetic null matrices at the empirical row totals. (a,b) Under a correctly specified null, the row- and column￾aggregated DM deviance against the asymptotic χ 2 K−1 and χ 2 n−1… view at source ↗
Figure 12
Figure 12. Figure 12: Fitting the flat Dirichlet–multinomial to data it cannot represent: it degrades gracefully and does no harm. Left: held-out conditional log-likelihood per count of the flat DM minus the multinomial, for each non-DM generating mechanism and severity (pooled over both scales); the gain is non-negative throughout and widens as the mechanism departs further from a multinomial (95% intervals). Middle: correlat… view at source ↗
Figure 13
Figure 13. Figure 13: The generalized Dirichlet–multinomial earns its order only when the order carries stage-varying dispersion. Left: the order-attributable held-out log-likelihood gain per observed count,GDM(true order)−GDM(permuted order), with the parameter count held fixed. On synthetic counts whose stage dispersion increases along the order, the true-order gain (filled circle, 95% bootstrap interval) sits well above the… view at source ↗
Figure 14
Figure 14. Figure 14: The Dirichlet–multinomial residual transform is linear in the number of nonzeros and preserves the sparsity pattern. Left: wall-clock time against nonzero count on log-log axes for the transform, the one-time concentration fit, and the pointwise log1p reference, across synthetic sweeps and two real matrices; the transform’s fitted exponent is one at the median (median β = 1.00; slopes annotated with boots… view at source ↗
Figure 15
Figure 15. Figure 15: Theorem-to-experiment correspondence. (a) Entry-wise residual variance over B = 2,000 DM-null replicates with n = 200, K = 50, α0 = 200. The multinomial overshoots the unit-variance target by ∼2×, exactly the overdispersion factor; the DM-impl variant (p max(c, 0)) recovers near-unit variance, while the DM-paper variant (p |c|) inflates the variance because |c| adds the under-the-null half of the cell-lev… view at source ↗
Figure 16
Figure 16. Figure 16: Microbiome application of the DM and DTM normalizations on the QIIME 2 Moving Pictures dataset (n = 34 samples, K = 770 OTUs, rooted SEPP phylogeny of depth 38). (a) Held-out per-count log-likelihood of the four candidate models; solid bars are the real Moving Pictures data, hatched bars are a synthetic DTM calibration (n = 400, K = 128, binary tree, αν,0 = 2000 · 2 −d(ν) ). The nodewise DTM-independent s… view at source ↗
Figure 17
Figure 17. Figure 17: Spatial DTM on 10x Visium mouse brain. (a) Visium spots colored by terminal quadtree-leaf id; depth-2 region rectangles overlaid in black. (b) Held-out log-likelihood per observed count; DTM (indep.) wins by 0.02 nats over the flat DM and by larger margins over every other baseline. (c) Pareto plot of depth coupling (lower is better, x-axis) versus mean Moran’s I over the top 10 PCs of each method’s spot … view at source ↗
Figure 18
Figure 18. Figure 18: DM residuals for pooled CRISPR / MPRA barcode normalization. (a) Fitted DM concentration αˆ0 across three datasets. On the MAGeCK leukemia demo counts, αˆ0 ≈ 2.9 × 104 reflects the near-multinomial behavior of a well￾represented high-library-size pooled screen; the two synthetic configurations realize overdispersion typical of dispersed screens (αˆ0 in the 100s). (b) Corollary 2.8 numerical check on MAGeC… view at source ↗
Figure 19
Figure 19. Figure 19: Dirichlet-multinomial residuals on three document-term corpora. (a) Fitted concentration αˆ0; moderate val￾ues on all three corpora confirm non-trivial word burstiness. Bars labeled “n¯” are the mean document length in tokens. (b) Held-out per-count log-likelihood under the DM (dark) and multinomial (light) models; DM has the better density on every corpus. (c) Count-stratified residual shrinkage |d Mult|… view at source ↗
Figure 20
Figure 20. Figure 20: DM residuals on MovieLens implicit feedback. (A) Fitted αˆ0 across three MovieLens datasets and three user￾count subsamples each; αˆ0 is stable under subsampling and lies in the strongly-overdispersed regime (αˆ0 ≪ K). (B) Per-count held-out log-likelihood on ML-1M’s 80/20 split; DM dominates both the multinomial and the unconditional feature-wise NB. Inset: per-user LL gain ∆DM u is positive for 100% of … view at source ↗
read the original abstract

We introduce a Dirichlet--multinomial (DM) deviance residualization for sparse, jointly overdispersed count matrices, the regime that dominates sequencing-based biochemical assays. The DM null treats each sample's count vector as a fixed-total composition with a single scalar concentration $\alpha_0$ governing overdispersion, and arises exactly by conditioning independent negative-binomial feature counts on the observed sample total -- making the DM the joint conditional analogue of standard feature-wise overdispersed count models. The resulting transform preserves exact sparsity, evaluates in constant time per nonzero entry, agrees with multinomial residuals on singleton counts, shrinks repeated-count residuals according to the overdispersion the null tolerates, and recovers the multinomial residual as $\alpha_0\to\infty$. The same fixed-dispersion comparison principle extends to ordered and tree-structured features via the generalized DM and the Dirichlet-tree multinomial, giving a single residual family that subsumes joint and feature-wise count nulls under a common compositional logic and is computationally lightweight enough to drop into existing sparse pipelines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper introduces a Dirichlet-multinomial (DM) deviance residualization for sparse, jointly overdispersed count matrices common in sequencing assays. The DM null is defined as a fixed-total composition with scalar concentration α0, obtained exactly by conditioning independent negative-binomial feature counts on the observed sample total. The resulting residual is claimed to preserve exact sparsity, evaluate in O(1) time per nonzero entry, agree with multinomial residuals on singleton counts, shrink repeated-count residuals according to tolerated overdispersion, recover the multinomial residual as α0→∞, and extend via the same fixed-dispersion principle to generalized DM and Dirichlet-tree multinomial models for ordered or tree-structured features.

Significance. If the derivation and claimed properties hold, the method would supply a computationally lightweight, sparsity-preserving normalization that unifies joint compositional and feature-wise overdispersed count models under a common conditional logic, with direct applicability to high-throughput biochemical count data.

major comments (1)
  1. Abstract: The manuscript states conceptual properties and the derivation route (conditioning negative binomials on totals) but supplies no equations, explicit residual formula, proofs, simulations, or data examples; therefore the mathematical support for the central claims (sparsity preservation, O(1) evaluation, shrinkage behavior, and limit recovery) cannot be verified from the available information.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the review and recommendation. We address the major comment below.

read point-by-point responses

  1. Referee: [—] Abstract: The manuscript states conceptual properties and the derivation route (conditioning negative binomials on totals) but supplies no equations, explicit residual formula, proofs, simulations, or data examples; therefore the mathematical support for the central claims (sparsity preservation, O(1) evaluation, shrinkage behavior, and limit recovery) cannot be verified from the available information.

    Authors: We agree that the abstract is intentionally high-level and narrative, without explicit equations or proofs, due to typical length constraints. The full manuscript supplies the explicit DM deviance residual formula (derived exactly by conditioning independent negative-binomial counts on the observed sample total), the proofs of sparsity preservation and O(1) evaluation per nonzero entry, the recovery of the multinomial residual as α₀→∞, the shrinkage behavior under finite α₀, and the supporting simulation studies and data examples. To address the concern about verifiability from the abstract, we will revise it to include a concise statement of the key residual formula. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The manuscript presents the DM null and its deviance residual as following directly from the standard statistical construction of the Dirichlet-multinomial distribution (conditioning independent negative-binomial counts on their sum). All listed properties—sparsity preservation, O(1) evaluation, agreement with multinomial residuals at count=1, shrinkage under overdispersion, and the α0→∞ limit—are algebraic consequences of that construction plus the definition of deviance residuals under a fixed-α0 null. No load-bearing step reduces to a fitted parameter renamed as a prediction, a self-citation chain, or an ansatz smuggled via prior work by the same authors. The extension to generalized DM and Dirichlet-tree multinomial applies the identical fixed-dispersion comparison principle to the corresponding joint distributions. The derivation is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the Dirichlet-multinomial as the appropriate joint null obtained by conditioning negative binomials, plus the scalar concentration parameter that controls tolerated overdispersion.

free parameters (1)
  • α0
    Scalar concentration parameter that governs the level of overdispersion tolerated by the DM null model.
axioms (1)
  • domain assumption The DM distribution arises exactly by conditioning independent negative-binomial feature counts on the observed sample total.
    This conditioning step is invoked to establish DM as the joint conditional analogue of standard feature-wise overdispersed count models.

pith-pipeline@v0.9.1-grok · 5713 in / 1381 out tokens · 56898 ms · 2026-06-25T19:16:47.269685+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

71 extracted references · 7 canonical work pages

  1. [1]

    Comparison of transformations for single-cell RNA-seq data

    Constantin Ahlmann-Eltze and Wolfgang Huber. Comparison of transformations for single-cell RNA-seq data. Na- ture Methods, 20(5):665–672, 2023

    2023

  2. [2]

    Comparison of transformations for single-cell RNA-seq data

    Constantin Ahlmann-Eltze and Wolfgang Huber. Comparison of transformations for single-cell RNA-seq data. Na- ture Methods, 20(5):665–672, 2023. Alias of AhlmannEltzeHuber2023

    2023

  3. [3]

    The statistical analysis of compositional data

    John Aitchison. The statistical analysis of compositional data. Journal of the Royal Statistical Society: Series B , 44(2): 139–160, 1982

    1982

  4. [4]

    The Statistical Analysis of Compositional Data

    John Aitchison. The Statistical Analysis of Compositional Data. Chapman & Hall, 1986

    1986

  5. [5]

    John Aitchison and C. H. Ho. The multivariate Poisson-log normal distribution. Biometrika, 76(4):643–653, 1989

    1989

  6. [6]

    John Aitchison and S. M. Shen. Logistic-normal distributions: Some properties and uses. Biometrika, 67(2):261–272, 1980

    1980

  7. [7]

    Differential expression analysis for sequence count data

    Simon Anders and Wolfgang Huber. Differential expression analysis for sequence count data. Genome Biology , 11(10):R106, 2010. doi: 10.1186/gb-2010-11-10-r106. url-verified 2026-05-26: https://link.springer.com/article/10.1186/gb-2010-11-10-r106

    work page doi:10.1186/gb-2010-11-10-r106 2010

  8. [8]

    Latent Dirichlet allocation

    David M Blei, Andrew Y Ng, and Michael I Jordan. Latent Dirichlet allocation. Journal of Machine Learning Research, 3:993–1022, 2003

    2003

  9. [9]

    Reproducible, interactive, scalable and extensible microbiome data science using QIIME 2

    Evan Bolyen, Jai Ram Rideout, Matthew R Dillon, Nicholas A Bokulich, Christian C Abnet, Gabriel A Al-Ghalith, et al. Reproducible, interactive, scalable and extensible microbiome data science using QIIME 2. Nature Biotechnology, 37 (8):852–857, 2019

    2019

  10. [10]

    Mueller, Fabian J

    Maren Buettner, Johannes Ostner, Christian L. Mueller, Fabian J. Theis, and Benjamin Schubert. scCODA is a Bayesian model for compositional single-cell data analysis. Nature Communications, 12(1):6876, 2021

    2021

  11. [11]

    Global patterns of 16S rRNA diversity at a depth of millions of sequences per sample

    J Gregory Caporaso, Christian L Lauber, William A Walters, Donna Berg-Lyons, Catherine A Lozupone, Peter J Turn- baugh, Noah Fierer, and Rob Knight. Global patterns of 16S rRNA diversity at a depth of millions of sequences per sample. Proceedings of the National Academy of Sciences , 108(Supplement 1):4516–4522, 2011

    2011

  12. [12]

    Variable selection for sparse Dirichlet-multinomial regression with an application to microbiome data analysis

    Jun Chen and Hongzhe Li. Variable selection for sparse Dirichlet-multinomial regression with an application to microbiome data analysis. The Annals of Applied Statistics, 7(1):418–442, 2013

    2013

  13. [13]

    Comparison and evaluation of statistical error models for scRNA-seq

    Saket Choudhary and Rahul Satija. Comparison and evaluation of statistical error models for scRNA-seq. Genome Biology, 23(1):27, 2022. 63

    2022

  14. [14]

    Connor and James E

    Robert J. Connor and James E. Mosimann. Concepts of independence for proportions with a generalization of the dirichlet distribution. Journal of the American Statistical Association , 64(325):194–206, 1969

    1969

  15. [15]

    Deep neural networks for YouTube recommendations

    Paul Covington, Jay Adams, and Emre Sargin. Deep neural networks for YouTube recommendations. In Proceedings of the 10th ACM Conference on Recommender Systems (RecSys) , pages 191–198, 2016

    2016

  16. [16]

    Performance of recommender algorithms on top- n recom- mendation tasks

    Paolo Cremonesi, Yehuda Koren, and Roberto Turrin. Performance of recommender algorithms on top- n recom- mendation tasks. In Proceedings of the 4th ACM Conference on Recommender Systems (RecSys) , pages 39–46, 2010

    2010

  17. [17]

    Sur les lois de probabilité à estimation exhaustive

    Georges Darmois. Sur les lois de probabilité à estimation exhaustive. C. R. Acad. Sci. Paris, 260:1265–1268, 1935

    1935

  18. [18]

    Samuel Y. Dennis. On the hyper-dirichlet type 1 and hyper-liouville distributions. Communications in Statistics— Theory and Methods, 20(12):4069–4081, 1991

    1991

  19. [19]

    Exponential Families in Theory and Practice

    Bradley Efron. Exponential Families in Theory and Practice. Institute of Mathematical Statistics T extbooks. Cambridge University Press, 2022. doi: 10.1017/9781108773157

    work page doi:10.1017/9781108773157 2022

  20. [20]

    Clustering documents with an exponential-family approximation of the dirichlet compound multi- nomial distribution

    Charles Elkan. Clustering documents with an exponential-family approximation of the dirichlet compound multi- nomial distribution. In Proceedings of the 23rd International Conference on Machine Learning , pages 289–296, 2006

    2006

  21. [21]

    Rouhana, Julia Waldman, et al

    Gökcen Eraslan, Eugene Drokhlyansky, Shankara Anand, Evgenij Fiskin, Ayshwarya Subramanian, Michal Slyper, Jiali Wang, Nicholas Van Wittenberghe, John M. Rouhana, Julia Waldman, et al. Single-nucleus cross-tissue molec- ular reference maps toward understanding disease gene function. Science, 376(6594):eabl4290, 2022. Alias of EraslanDrokhlyanskyAnandEtAl2022

    2022

  22. [22]

    Symmetric Multivariate and Related Distributions

    Kai-Tai Fang. Symmetric Multivariate and Related Distributions. Chapman and Hall/CRC, 2018

    2018

  23. [23]

    Farewell and Vernon T

    Daniel M. Farewell and Vernon T. Farewell. Dirichlet negative multinomial regression for overdispersed correlated count data. Biostatistics, 14(2):395–404, 2013

    2013

  24. [24]

    Unifying the analysis of high-throughput sequencing datasets: characterizing RNA-seq, 16S rRNA gene se- quencing and selective growth experiments by compositional data analysis

    Andrew D Fernandes, Jennifer NS Reid, Jean M Macklaim, Thomas A McMurrough, David R Edgell, and Gregory B Gloor. Unifying the analysis of high-throughput sequencing datasets: characterizing RNA-seq, 16S rRNA gene se- quencing and selective growth experiments by compositional data analysis. Microbiome, 2:15, 2014

    2014

  25. [25]

    On the mathematical foundations of theoretical statistics

    Ronald A Fisher. On the mathematical foundations of theoretical statistics. Philosophical transactions of the Royal Society of London. Series A, containing papers of a mathematical or physical character , 222(594-604):309–368, 1922

    1922

  26. [26]

    Stochastic Finance: An Introduction in Discrete Time

    Hans Föllmer and Alexander Schied. Stochastic Finance: An Introduction in Discrete Time . Walter de Gruyter, 3rd edition, 2011

    2011

  27. [27]

    Gloor, Jean M

    Gregory B. Gloor, Jean M. Macklaim, Vera Pawlowsky-Glahn, and Juan J. Egozcue. Microbiome datasets are com- positional: and this is not optional. Frontiers in Microbiology, 8:2224, 2017

    2017

  28. [28]

    Lindrooth

    Paulo Guimaraes and Richard C. Lindrooth. Controlling for overdispersion in grouped conditional logit models: a computationally simple application of dirichlet-multinomial regression. The Econometrics Journal, 10(2):439–452, 2007

    2007

  29. [29]

    Gupta and Donald St P

    Rameshwar D. Gupta and Donald St P. Richards. Multivariate liouville distributions. Journal of Multivariate Analysis, 23(2):233–256, 1987

    1987

  30. [30]

    Normalization and variance stabilization of single-cell rna-seq data using regularized negative binomial regression

    Christoph Hafemeister and Rahul Satija. Normalization and variance stabilization of single-cell rna-seq data using regularized negative binomial regression. Genome Biology, 20(1):296, 2019

    2019

  31. [31]

    The MovieLens datasets: history and context

    F Maxwell Harper and Joseph A Konstan. The MovieLens datasets: history and context. ACM Transactions on Inter- active Intelligent Systems, 5(4):1–19, 2015

    2015

  32. [32]

    Harrison, W

    Joshua G. Harrison, W. John Calder, Vivaswat Shastry, and C. Alex Buerkle. Dirichlet-multinomial modelling out- performs alternatives for analysis of microbiome and other ecological count data. Molecular Ecology Resources, 20(2): 481–497, 2020. 64

    2020

  33. [33]

    BAGEL: a computational framework for identifying essential genes from pooled library screens

    Traver Hart and Jason Moffat. BAGEL: a computational framework for identifying essential genes from pooled library screens. BMC Bioinformatics, 17:164, 2016

    2016

  34. [34]

    Evaluation and design of genome-wide CRISPR/SpCas9 knockout screens

    Traver Hart, Amy Hin Yan T ong, Katie Chan, Jolanda Van Leeuwen, Ashwin Seetharaman, Michael Aregger, Megha Chandrashekhar, Nicole Hustedt, Sahil Seth, Avery Noonan, et al. Evaluation and design of genome-wide CRISPR/SpCas9 knockout screens. G3: Genes, Genomes, Genetics , 7(8):2719–2727, 2017

    2017

  35. [35]

    A closer look at the deviance

    Trevor Hastie. A closer look at the deviance. The American Statistician, 41(1):16–20, 1987

    1987

  36. [36]

    Hall, and Zvi Griliches

    Jerry Hausman, Bronwyn H. Hall, and Zvi Griliches. Econometric models for count data with an application to the patents-r&d relationship. Econometrica, 52(4):909–938, 1984

    1984

  37. [37]

    Howard, Aaditya Ramdas, Jon McAuliffe, and Jasjeet Sekhon

    Steven R. Howard, Aaditya Ramdas, Jon McAuliffe, and Jasjeet Sekhon. Time-uniform, nonparametric, nonasymp- totic confidence sequences. The Annals of Statistics, 49(2):1055–1080, 2021. doi: 10.1214/20-AOS2002

    work page doi:10.1214/20-aos2002 2021

  38. [38]

    Collaborative filtering for implicit feedback datasets

    Yifan Hu, Yehuda Koren, and Chris Volinsky. Collaborative filtering for implicit feedback datasets. In Proceedings of the 8th IEEE International Conference on Data Mining (ICDM) , pages 263–272, 2008

    2008

  39. [39]

    On distributions admitting a sufficient statistic

    Bernard Osgood Koopman. On distributions admitting a sufficient statistic. Transactions of the American Mathemat- ical society, 39(3):399–409, 1936

    1936

  40. [40]

    Koslovsky

    Matthew D. Koslovsky. A bayesian zero-inflated dirichlet-multinomial regression model for multivariate composi- tional count data. Biometrics, 79(4):3239–3251, 2023

    2023

  41. [41]

    Analytic Pearson residuals for normalization of single-cell RNA- seq UMI data

    Jan Lause, Philipp Berens, and Dmitry Kobak. Analytic Pearson residuals for normalization of single-cell RNA- seq UMI data. Genome Biology , 22:258, 2021. doi: 10.1186/s13059-021-02451-7. url-verified 2026-05-26: https://link.springer.com/article/10.1186/s13059-021-02451-7

    work page doi:10.1186/s13059-021-02451-7 2021

  42. [42]

    MAGeCK enables robust identification of essential genes from genome-scale CRISPR/Cas9 knockout screens

    Wei Li, Han Xu, T engfei Xiao, Le Cong, Michael I Love, Feng Zhang, Rafael A Irizarry, Jun S Liu, Myles Brown, and X Shirley Liu. MAGeCK enables robust identification of essential genes from genome-scale CRISPR/Cas9 knockout screens. Genome Biology, 15:554, 2014

    2014

  43. [43]

    Love, Wolfgang Huber, and Simon Anders

    Michael I. Love, Wolfgang Huber, and Simon Anders. Moderated estimation of fold change and dispersion for rna- seq data with deseq2. Genome Biology, 15(12):550, 2014

    2014

  44. [44]

    Macosko, Anindita Basu, Rahul Satija, James Nemesh, Karthik Shekhar, Michael Goldman, Itay Tirosh, Al- lison R

    Evan Z. Macosko, Anindita Basu, Rahul Satija, James Nemesh, Karthik Shekhar, Michael Goldman, Itay Tirosh, Al- lison R. Bialas, Nolan Kamitaki, Emily M. Martersteck, et al. Highly parallel genome-wide expression profiling of individual cells using nanoliter droplets. Cell, 161(5):1202–1214, 2015

    2015

  45. [45]

    Marshall, Ingram Olkin, and Barry C

    Albert W. Marshall, Ingram Olkin, and Barry C. Arnold. Inequalities: Theory of Majorization and Its Applications . Academic Press, 1979

    1979

  46. [46]

    McCullagh and J

    P. McCullagh and J. A. Nelder. Generalized Linear Models. Chapman and Hall, 2nd edition, 1989

    1989

  47. [47]

    Thomas P. Minka. Estimating a dirichlet distribution. T echnical report, MIT, 2000

    2000

  48. [48]

    Thomas P. Minka. The dirichlet-tree distribution. T echnical report, 2004. T echnical note, 1999; revised 2004

    2004

  49. [49]

    Carl N. Morris. Natural exponential families with quadratic variance functions. The Annals of Statistics, 10(1):65–80, 1982

    1982

  50. [50]

    Mosimann

    James E. Mosimann. On the compound multinomial distribution, the multivariate β-distribution, and correlations among proportions. Biometrika, 49(1/2):65–82, 1962

    1962

  51. [51]

    Mosimann

    James E. Mosimann. On the compound negative multinomial distribution and correlations among inversely sampled pollen counts. Biometrika, 50(1/2):47–54, 1963

    1963

  52. [52]

    Dirichlet and Related Distributions: Theory, Methods and Applica- tions

    Kai-Wang Ng, Guo-Liang Tian, and Man-Lai Tang. Dirichlet and Related Distributions: Theory, Methods and Applica- tions. Wiley, 2011

    2011

  53. [53]

    Sufficient statistics and intrinsic accuracy

    Edwin James George Pitman. Sufficient statistics and intrinsic accuracy. InMathematical Proceedings of the cambridge Philosophical society, volume 32, pages 567–579. Cambridge University Press, 1936. 65

    1936

  54. [54]

    Game-theoretic statistics and safe anytime-valid inference

    Aaditya Ramdas, Peter Grünwald, Vladimir Vovk, and Glenn Shafer. Game-theoretic statistics and safe anytime-valid inference. Statistical Science, 38(4):576–601, 2023. doi: 10.1214/23-STS894

    work page doi:10.1214/23-sts894 2023

  55. [55]

    Using tf-idf to determine word relevance in document queries

    Juan Ramos et al. Using tf-idf to determine word relevance in document queries. In Proceedings of the First Instruc- tional Conference on Machine Learning, volume 242, pages 29–48, 2003

    2003

  56. [56]

    Robinson, Davis J

    Mark D. Robinson, Davis J. McCarthy, and Gordon K. Smyth. edger: a bioconductor package for differential expres- sion analysis of digital gene expression data. Bioinformatics, 26(1):139–140, 2010

    2010

  57. [57]

    T erm-weighting approaches in automatic text retrieval

    Gerard Salton and Christopher Buckley. T erm-weighting approaches in automatic text retrieval. Information Pro- cessing & Management, 24(5):513–523, 1988

    1988

  58. [58]

    Farrell, David Gennert, Alexander F

    Rahul Satija, Jeffrey A. Farrell, David Gennert, Alexander F. Schier, and Aviv Regev. Spatial reconstruction of single- cell gene expression data. Nature Biotechnology, 33(5):495–502, 2015

    2015

  59. [59]

    Fast mle computation for the dirichlet multinomial

    Max Sklar. Fast mle computation for the dirichlet multinomial. arXiv preprint arXiv:1405.0099, 2014

    arXiv 2014

  60. [60]

    Mauck III, Yuhan Hao, Marlon Stoeckius, Peter Smibert, and Rahul Satija

    Tim Stuart, Andrew Butler, Paul Hoffman, Christoph Hafemeister, Efthymia Papalexi, William M. Mauck III, Yuhan Hao, Marlon Stoeckius, Peter Smibert, and Rahul Satija. Comprehensive integration of single- cell data. Cell, 177(7):1888–1902.e21, 2019. doi: 10.1016/j.cell.2019.05.031. url-verified 2026-05-26: https://www.cell.com/cell/fulltext/S0092-8674(19)30559-8

    work page doi:10.1016/j.cell.2019.05.031 1902

  61. [61]

    Droplet scrna-seq is not zero-inflated

    Valentine Svensson. Droplet scrna-seq is not zero-inflated. Nature Biotechnology, 38(2):147–150, 2020

    2020

  62. [62]

    Thorson, Kelli F

    James T. Thorson, Kelli F. Johnson, Richard D. Methot, and Ian G. Taylor. Model-based estimates of effective sample size in stock assessment models using the dirichlet-multinomial distribution. Fisheries Research, 192:84–93, 2017

    2017

  63. [63]

    Tiao and Irwin Cuttman

    George C. Tiao and Irwin Cuttman. The inverted dirichlet distribution with applications. Journal of the American Statistical Association, 60(311):793–805, 1965

    1965

  64. [64]

    William T ownes, Stephanie C

    F. William T ownes, Stephanie C. Hicks, Martin J. Aryee, and Rafael A. Irizarry. Feature selection and dimension reduction for single-cell RNA-Seq based on a multinomial model. Genome Biology , 20:295, 2019. doi: 10.1186/ s13059-019-1861-6. url-verified 2026-05-26: https://link.springer.com/article/10.1186/s13059-019-1861-6

    work page doi:10.1186/s13059-019-1861-6 2019

  65. [65]

    Williams, Geo Pertea, Ali Mortazavi, Grace Kwan, Marijke J

    Cole Trapnell, Brian A. Williams, Geo Pertea, Ali Mortazavi, Grace Kwan, Marijke J. van Baren, Steven L. Salzberg, Barbara J. Wold, and Lior Pachter. Transcript assembly and quantification by rna-seq reveals unannotated transcripts and isoform switching during cell differentiation. Nature Biotechnology, 28(5):511–515, 2010

    2010

  66. [66]

    Duncan Wadsworth, Raffaele Argiento, Michele Guindani, Jessica Galloway-Peña, Samuel A

    W. Duncan Wadsworth, Raffaele Argiento, Michele Guindani, Jessica Galloway-Peña, Samuel A. Shelburne, and Ma- rina Vannucci. An integrative bayesian dirichlet-multinomial regression model for the analysis of taxonomic abun- dances in microbiome data. BMC Bioinformatics, 18:94, 2017

    2017

  67. [67]

    A dirichlet-tree multinomial regression model for associating dietary nutrients with gut microorganisms

    Tao Wang and Hongyu Zhao. A dirichlet-tree multinomial regression model for associating dietary nutrients with gut microorganisms. Biometrics, 73(3):792–801, 2017

    2017

  68. [68]

    Universal inference

    Larry Wasserman, Aaditya Ramdas, and Sivaraman Balakrishnan. Universal inference. Proceedings of the National Academy of Sciences, 117(29):16880–16890, 2020

    2020

  69. [69]

    A logistic normal multinomial regression model for microbiome compositional data analysis

    Fengzhu Xia, Jun Chen, Wing Kam Fung, and Hongzhe Li. A logistic normal multinomial regression model for microbiome compositional data analysis. Biometrics, 69(4):1053–1063, 2013

    2013

  70. [70]

    Mm algorithms for some discrete multivariate distributions.Journal of Computational and Graphical Statistics, 19(3):645–665, 2010

    Hua Zhou and Kenneth Lange. Mm algorithms for some discrete multivariate distributions.Journal of Computational and Graphical Statistics, 19(3):645–665, 2010

    2010

  71. [71]

    Nonparametric bayesian negative binomial factor analysis

    Mingyuan Zhou. Nonparametric bayesian negative binomial factor analysis. Bayesian Analysis, 13(4):1065–1093, 2018. 66 Figure 4: Downstream analyses on GTEx. A deep generative baseline, scVI, is included wherever the comparison is commensurable. It attains the strongest embeddings here, but at far higher training cost and at the cost of differential- expre...

    2018