The question behind the paper

Count data appear everywhere. An economist counts firm failures, strikes, defaults, patents, accidents, or entries into a market. A political scientist counts protests, cabinet changes, violent events, legislative vetoes, or international disputes. A biologist counts surviving organisms, mutations, infections, or offspring. The elementary model for many such problems is the Poisson distribution. But real counts often fluctuate more than a Poisson model allows. Their variance is larger than their mean: the data are overdispersed.

The Negative Binomial II—usually abbreviated NB2—is one of the central statistical instruments for precisely that situation. Yet Gómez Julián’s paper is not content to present the NB2 probability mass function, list its moments, and move on. Its organizing question is broader: what kind of object is a probability distribution, where does it come from, what other distributions does it contain or presuppose, and what is gained when we understand its construction rather than merely its final formula?

The paper therefore has two explicit axes. The first, and more important one, studies NB2 as the outcome of a wider theoretical structure involving hierarchical models and probability mixtures. The second studies NB2 as an individual distribution—its form, interpretation, moments, and practical use, including hand calculations and R. The intended unity between these axes is philosophical as well as mathematical: probability distributions are treated as instruments for measuring natural and social phenomena under uncertainty, and the paper argues that their history, formal structure, scientific interpretation, and application should not be torn apart.

The distribution is easier to understand when we see not only the finished formula, but also the process that produces it.

Why begin with the geometric series?

The paper begins surprisingly far away from count-data regression: with the geometric series, its historical roots, and its relation to the binomial theorem. This is deliberate. Gómez Julián wants the reader to see the Negative Binomial II as part of a mathematical genealogy rather than as a formula that appeared fully formed.

The route is roughly this: the geometric series provides a simple infinite expansion; differentiation exposes a recurring combinatorial pattern; that pattern is used to motivate the binomial expansion; and replacing the ordinary exponent with a negative one leads to the negative binomial series. The point is not merely algebraic. The paper repeatedly emphasizes the movement from simpler structures to more general ones and from one family of mathematical objects to another.

For a nontechnical reader, the important idea is simple: a probability distribution can be understood through the transformations and relationships that generate it. The paper later reinforces this visually with a large network diagram of probability distributions, reproduced from ProbOnto, in which distributions are connected by transformations, limiting relations, and special cases. NB2 is therefore presented as one node in a densely connected mathematical ecology, not as an isolated technique.

Probability as a measurement problem

Before building the NB2, the paper stops to ask what “probability” means. This is not a decorative philosophical detour. Gómez Julián’s position is that formal probability calculus and philosophical interpretation cannot be completely divorced, because statistical conclusions depend on what we think probabilities are measuring and on how the scientific problem is conceptualized.

The paper adopts an explicitly objective and dialectical-materialist orientation. Randomness is treated primarily as an epistemological condition: events appear random because their causes are unknown, too numerous, or too complicated to represent completely. At the same time, the Kolmogorov axioms provide the formal mathematical framework that makes probability calculations coherent. The paper also draws on objective Bayesianism to argue that degrees of belief should be constrained by evidence, scientific theory, and the probability calculus rather than reduced to arbitrary personal opinion.

A family, not an isolated formula

The paper next introduces the distributions needed for its construction: Bernoulli, Binomial, Poisson, and Exponential. Each plays a distinct role.

This sequence already contains the paper’s methodological intuition. A complex phenomenon can be decomposed into simpler probabilistic stages, each corresponding to a different part of the scientific story. Instead of forcing all uncertainty into a single flat formula, a hierarchical model lets uncertainty enter at more than one level.

Hierarchies, mixtures, and latent variables

A hierarchical model specifies variables and parameters in stages. A mixture model appears when a parameter in one probability distribution is itself treated as a random quantity governed by another distribution. The parameter that seemed fixed at the lower level becomes variable at the higher level.

That is the key conceptual move in the paper. It is also why Gómez Julián brings in the language of latent variables: some of the forces producing observed variation may not be directly observed, but their consequences can still be represented probabilistically. In applied work, this is familiar. Two factories, municipalities, firms, hospitals, or individuals may face different underlying event rates even if we initially write one common Poisson equation for all of them.

The paper links this mathematical construction to the Hegelian distinction between Being-in-itself and Being-for-itself. Stripped of the philosophical vocabulary, its statistical meaning is fairly intuitive. Studying NB2 “in itself” means studying the wider process and network of relationships from which it emerges. Studying NB2 “for itself” means taking the marginalized distribution as a distinct object and examining its own formula, parameters, moments, and applications.

The first mixture: Binomial inside Poisson

The first major construction uses a biological example. Imagine an insect that lays many eggs. Conditional on a mother having laid Y eggs, each egg survives independently with probability p. The number of surviving eggs, X, is therefore Binomial. But instead of fixing the number of eggs Y, the paper lets it vary according to a Poisson distribution with mean λ.

Marginalizing means asking for the distribution of X after summing out the intermediate variable Y. Algebraically, we combine all the possible values of Y, weighted by how probable each one is. The result is elegant:

In modern probability language, this is a version of Poisson thinning. If the total number of opportunities is Poisson and each opportunity independently survives with probability p, then the surviving count is also Poisson, with its mean reduced from λ to λp. The paper also derives the same conclusion through iterated expectations:

For the paper, this is more than a computational trick. It shows how a hierarchical process that appears to contain two random layers can be “compressed” into a simpler marginal law without erasing the scientific interpretation that motivated the hierarchy.

When the Poisson rate itself varies

The next step introduces another level of heterogeneity. Suppose there are many insect mothers, and the Poisson mean is not the same for every mother. The paper now treats the rate itself as random:

The statistical intuition is important. Variation does not occur only in the observed count; it can also occur in the underlying rate that generates the count. Once the rate differs across observational units, the final distribution becomes more dispersed than a single-rate Poisson model. This is precisely the kind of latent heterogeneity that makes the negative binomial family useful in economics, epidemiology, demography, political science, and many other count-data settings.

The paper shows that the mean of the full hierarchy can be obtained by repeatedly conditioning and averaging, arriving at E[X] = pβ under its parameterization. It then integrates out the random Poisson rate and identifies the resulting expression with the negative binomial form.

This qualification actually sharpens the paper’s broader message. The geometric distribution, the exponential distribution, the gamma family, Poisson processes, and the negative binomial are not unrelated objects. They sit inside a network of special cases and mixture relationships. The more general Gamma mixing distribution preserves the same hierarchical logic while extending the model beyond the r = 1 case.

The NB2 “for itself”: what the finished distribution tells us

Once the hierarchy has been “compressed,” the paper turns to the Negative Binomial II as an object in its own right. One common parameterization writes the probability of observing x failures before the r-th success as:

Here, p is the success probability and r is the target number of successes. The distribution answers a reversed version of the familiar Binomial question. The Binomial asks: how many successes occur in a fixed number of trials? The negative binomial asks: how many failures occur before a fixed number of successes is reached?

The paper’s most important statistical property appears in its first two moments. With q = 1 − p:

That final equality is the bridge to modern count-data econometrics. The Poisson distribution imposes Var(X) = μ. NB2 allows:

In other words, variance can grow faster than the mean. The smaller r is—or, equivalently, the larger the heterogeneity parameter α is—the more dispersion the model permits. As heterogeneity vanishes, NB2 approaches the Poisson benchmark. This is why the paper presents NB2 as a more flexible instrument for count data when the Poisson equality between mean and variance is too restrictive.

For an econometrician, this variance function is often the fastest route to understanding NB2. For a broader reader, an intuitive translation is enough: NB2 expects the world to be more uneven than a simple Poisson process. Some units have persistently higher event rates than others; unobserved conditions vary; clusters form; the same mean can coexist with much wider dispersion.

The maintenance example: counting failures before the fourth alarm

The paper gives a concrete industrial interpretation. Imagine fixed capital—a machine—producing parts. A part is either satisfactory or defective. A monitoring system treats a defective part as the event of interest because it signals the need for maintenance. Assume independent Bernoulli trials and a constant defect probability.

Gómez Julián sets r = 4, p = 0.005, and asks for the probability of observing exactly x = 100 failures before the fourth success, using the negative binomial mass function. Substitution gives:

The number is tiny—about 6.7 chances in 100,000. The statistical interpretation is not that “100 failures happen and then four successes happen” as two separate blocks. Rather, among an ongoing sequence of independent trials, exactly 100 non-events occur before the fourth event of interest is reached.

The paper checks the result manually, with a Texas Instruments calculator, and with R. It then plots the corresponding distribution, illustrating how a small success probability pushes substantial probability mass toward relatively large counts before the required number of successes is accumulated.

What the R simulation is trying to show

The final applied part of the main text shifts from evaluating a probability to generating pseudo-random data. The paper constructs a custom card-drawing experiment: repeatedly sample from a 52-card deck until a specified rank appears, record how many draws were required, and repeat the experiment many times. It then compares the histogram produced by that “from first principles” counting procedure with a histogram generated by R’s built-in rnbinom function.

The figures for 50, 100, 150, and 200 repetitions show the same qualitative pattern: a strongly right-skewed count distribution with many small values and a long tail. The pedagogical purpose is clear. Software is not magic. A built-in random generator is implementing a probabilistic structure that can also be approximated through an explicit sequence of elementary trials.

The broader lesson is useful far beyond R. Simulation can reveal what a distribution means operationally: not merely how its formula looks, but what repeated mechanism would generate data with that shape.

NB1, Bayes, and the annexes: why the paper keeps widening the frame

The annexes extend the same relational approach. The paper distinguishes NB2 from NB1, emphasizing that different negative-binomial parameterizations answer slightly different counting questions. In the NB1 presentation used there, the random variable is the total number of Bernoulli trials needed to reach r events of the chosen type; in the NB2 presentation, the random variable is restricted to the number of failures before those r successes.

The paper also returns to conditional probability, total probability, inverse probability, and objective Bayesianism. This may seem far removed from overdispersed count data, but it serves the same philosophical thesis: statistical formulas must be understood through the relationships they encode. Conditional probability is not merely a ratio; it represents a structured dependence between events. Bayesian updating is not merely algebra; it connects prior knowledge, evidence, and posterior assessment.

Finally, the paper gives a more general NB2 expression in terms of the Gamma function and reports estimators based on the first two sample moments. This again links the abstract distribution to empirical work: the population parameters acquire meaning only because sample information provides a route to estimation.

What should we take away?

Gómez Julián’s paper is best read as an extended argument against treating the Negative Binomial II as a black-box formula. Its distinctive contribution is not a new NB2 estimator or a new regression algorithm. It is an attempt to reconstruct the distribution through several layers at once: historical, algebraic, probabilistic, computational, applied, and philosophical.

For the nontechnical reader, the central statistical lesson can be stated in one sentence: when counts vary more than a single-rate Poisson model permits, the extra variability can often be understood as heterogeneity in the underlying event rate, and the negative binomial family provides a natural way to represent it.

For the econometrician, the key signature is the NB2 variance function, Var(Y) = μ + αμ². For the mathematician, the paper is an invitation to follow the transformations linking geometric series, binomial expansions, gamma functions, conditional distributions, and marginalization. For the philosopher of science, its central claim is methodological: the formal instrument, the scientific object, the history of the instrument, and the interpretation of uncertainty should be studied in relation rather than isolation.

And for the applied researcher, perhaps the most useful question is the simplest one: what process would have to be operating for this distribution to be a sensible measurement instrument? Once that question is asked, NB2 stops being just a convenient correction for overdispersion. It becomes a hypothesis about how heterogeneity enters the data-generating process.

A distribution is most informative when its probability law and its generating story tell the same scientific story.