When data arrive only as coarse summaries—ranges, binned frequency tables, or grouped counts—most analysts either throw away the aggregates and hunt for raw records, or treat each summary as if it were a single approximate observation. A new study argues that neither shortcut is necessary, and that a carefully constructed likelihood built directly on the summaries can recover nearly all of the statistical information that would have been available from the full, unaggregated data. The work, published in Discover Data, develops the method for discrete count data and demonstrates it on a season of National Basketball Association free-throw statistics.
The framework belongs to Symbolic Data Analysis, a branch of statistics in which the observational units are not individual numbers but symbols: intervals, multi-valued categories, or frequency histograms produced by aggregating underlying micro-data. Such symbols arise constantly in practice, because privacy rules, storage limits, and reporting policies often prevent agencies from releasing individual-level records in public health, climatology, finance, or sports analytics. The challenge is that a symbol like “between 25 and 29 free-throw attempts” is not itself a measurement from any simple distribution, so standard estimation tools do not apply to it directly.
The paper’s central move is to treat each symbol as the deterministic output of a known aggregation map applied to an unobserved sample of latent counts. If the latent counts follow a parametric discrete distribution, the probability of observing a given symbol is obtained by summing the latent model over every configuration of counts that maps to that symbol. This induced probability mass function on the symbol space becomes the object of inference, and maximizing its likelihood yields what the authors call the symbolic maximum likelihood estimator. The approach explicitly respects the discrete, countable support of the data and avoids ad hoc assumptions about how values are distributed inside each interval or bin.
A key theoretical result is a score identity: the derivative of the symbolic log-likelihood equals the conditional expectation of the complete-data score given the observed symbol. For discrete exponential-family models such as the Poisson, this identity translates into estimating equations in which the usual sufficient statistics are replaced by their conditional expectations given the aggregates. That structure naturally suggests EM-type algorithms, where the expectation step averages sufficient statistics over all latent configurations compatible with each symbol and the maximization step mirrors the full-data update. For Poisson counts summarized in fixed bins, the authors derive a particularly simple fixed-point iteration involving bin-wise conditional means, making the computation essentially one-dimensional.
The method is not automatically valid for every aggregation scheme. The paper formalizes symbolic identifiability, requiring that different parameter values produce distinguishable distributions on the symbol space, and shows with a small example how a poorly chosen partition can destroy identifiability entirely: binning a Binomial(2, q) variable into the sets {0, 2} and {1} makes the bin probabilities symmetric under q and 1 minus q, so the parameter cannot be recovered. Under suitable regularity conditions, however, the symbolic maximum likelihood estimator inherits the standard large-sample guarantees of consistency and asymptotic normality.
To quantify how much information survives aggregation, the authors ran a simulation study with Poisson latent counts of mean 3, grouped into blocks of 20 observations and summarized in different ways. Fixed-bin frequency summaries performed strikingly well: the symbolic estimator’s root mean squared error stayed very close to that of the full-data estimator across sample sizes from 10 to 500, indicating only a modest efficiency loss from binning. Min-max intervals, which compress each block to just its smallest and largest values, remained approximately unbiased but lost noticeably more precision—at a sample size of 100, their root mean squared error was about 0.067 versus 0.039 for the full-data estimator.
The simulations also exposed the cost of heuristic shortcuts. A midpoint rule that treats the average of each interval’s endpoints as a proxy for the latent mean carried a persistent positive bias of roughly 0.5 that did not shrink as the sample grew, so its error was dominated by bias rather than variance. In a second experiment holding storage fixed at five integers per block, storing a random subsample of five raw counts produced an essentially unbiased estimator whose error decreased at the expected rate, while treating Tukey five-number summaries as pseudo-observations introduced a stable positive bias of about 0.13. The lesson is that the statistical value of a summary depends critically on whether the summary is analyzed through a likelihood compatible with how it was generated.
The real-data application uses box-score records from the 2022-2023 NBA regular season, accessed through the R package ISAR, focusing on the Dallas Mavericks’ 82 games. Game-level free-throw attempts ranged from 8 to 42, with a median of 25 and a mean of 25.09; the sample variance of 45.39 signaled overdispersion relative to a Poisson model. The authors compressed the season into a single five-bin frequency symbol over volume categories from 0-14 up to 30-or-more attempts, yielding observed bin counts of 4, 17, 16, 27, and 18 games. Fitting the symbolic Poisson likelihood to this one symbol gave an estimated mean of 24.54, remarkably close to the full-data estimate of 25.09 despite the drastic compression.
Model assessment was carried out entirely in the symbol space using a Pearson-type goodness-of-fit statistic calibrated by parametric bootstrap, with 3,000 synthetic seasons generated under each fitted model. The Poisson model was strongly rejected, with a statistic of 20.30 and a bootstrap p-value of about 0.002: it underestimated very low-volume games, overestimated the central 20-24 bin, and underestimated high-volume games of 30 or more attempts. A two-parameter Negative Binomial model, which allows the variance to exceed the mean, fit far better, with a statistic of 4.68 and a p-value near 0.32. Its fitted variance of 45.67 closely matched the empirical 45.39. Substantively, the Mavericks’ game-to-game free-throw volume was more variable than an equidispersed count process would predict, with both unusually quiet and unusually high-volume nights occurring more often than a Poisson model allows.
The authors note limitations: the framework currently covers univariate discrete latent variables and simple symbol types, and extending it to multivariate or temporally dependent aggregates, hierarchical settings with multiple teams or seasons, and Bayesian formulations remains open work, as does a systematic study of how to choose the aggregation map to optimize inference under a fixed storage budget. Even so, the combination of theory, simulation, and the basketball case study makes a clear practical point: when only coarse count summaries are available, fitting an explicit likelihood on the summaries themselves can preserve nearly all of the information in the underlying data, while naive pseudo-data approximations can quietly introduce bias that no amount of additional data will fix.
Subject of Research: Likelihood-based statistical inference from aggregated discrete count data, illustrated with NBA free-throw attempts
Article Title: Symbolic likelihood inference for discrete aggregated count data: an NBA application
Article References: Sadeghkhani, A. (2026). Symbolic likelihood inference for discrete aggregated count data: an NBA application. Discover Data, 4(1), Article 15. https://doi.org/10.1007/s44248-026-00114-1
Image Credits: AI Generated
DOI: 10.1007/s44248-026-00114-1
Keywords: symbolic data analysis, aggregated count data, likelihood inference, EM algorithm, Poisson model, negative binomial, overdispersion, goodness-of-fit, parametric bootstrap, NBA free throws, symbolic maximum likelihood, fixed-bin frequency
News Source: Denise Maddox. (October 28, 2026). Research report: Symbolic likelihood inference for discrete aggregated count data: an NBA application. Scienmag.
