How to Apply Lotka’s Law to Bibliometric Data

Lotka’s Law tells us that scientific productivity is deeply unequal: a small group of authors produces a large share of the literature, while most contribute just one or two papers. But knowing the law exists is different from putting it to work. The real skill in informetrics is taking a messy list of publications and turning it into a clean dataset that you can actually fit to Lotka’s equation. This post walks through exactly that process, from organizing raw bibliographic records to estimating the parameters that define the productivity curve.

A quick refresher on the equation 🔗

Before the practical steps, it helps to fix the formula in mind. Lotka’s Law states that the number of authors producing a given number of papers is inversely proportional to a power of that number. The standard form expresses it as the number of authors publishing x papers being approximately C / xⁿ, where C is a constant and n is the productivity exponent. In Lotka’s original 1926 study of chemistry literature, he proposed n = 2, meaning the number of authors with two papers is about one-quarter of those with one paper, those with three papers about one-ninth, and so on.

Note one point of terminology that often confuses students. Some textbooks write the exponent as n and the constant as k (or C). The letters vary across sources, but the two quantities are always the same: the slope of the productivity curve and the constant that scales it. Your job when applying the law is to estimate both from real data.

Steps in data organization 🔗

You cannot estimate anything until your data is in the right shape. Lotka’s Law is a frequency distribution, so the entire process begins with counting authors by their output. This is the foundation that every later calculation rests on.

Step 1: Define your dataset and counting rules 🔗

Start by deciding what you are measuring. Most bibliometric studies draw records from a database such as Scopus, Web of Science, or a subject-specific index like LISA. Once you have the records, you must choose a counting method, and this choice genuinely affects your results. The two common approaches are normal count, where every author on a paper receives full credit, and straight count, where only the first or senior author is counted. Studies following Pao’s widely used methodology typically assign equal credit to each author who appears on a work, so a four-author paper adds one publication to each of the four authors’ tallies. State your counting rule clearly, because it determines who counts as a “productive” author.

Step 2: Build the author frequency table 🔗

Next, count how many papers each individual author has produced across your dataset. Then collapse this into a frequency distribution. You are not interested in named individuals anymore; you want to know how many authors produced exactly one paper, how many produced two, how many produced three, and so on.

The result is a simple two-column table. The first column, usually labelled x, is the number of papers. The second column, labelled y or a_x, is the number of authors who produced that many papers. A typical table looks like this in structure: x = 1 with several thousand authors, x = 2 with far fewer, x = 3 fewer still, declining steeply as you move down. This sharp drop is the signature of Lotka’s distribution.

Step 3: Transform the data into logarithms 🔗

Lotka’s equation is a power law, which plots as a curve. To fit a straight line to it, you take the logarithm of both columns. So alongside your x and y columns, you add columns for log x and log y. Taking the logarithm of both sides of the equation converts the curved power-law relationship into a linear one of the form log y = log C − n log x. This linear form is what makes the parameters estimable through standard regression. Once your table has these four columns, you are ready to estimate the parameters.

Estimation of parameters 🔗

With the data organized, the central task is finding the two parameters: the exponent n and the constant C (or k). There are three established methods, and each has a different balance of simplicity and statistical rigour. The three standard parameter estimation techniques in statistics are linear least squares, maximum likelihood estimation, and the method of moments. We will take them in turn.

Method 1: Linear least squares 🔗

This is the most common method in published bibliometric work, and the easiest to perform by hand or in a spreadsheet. It uses the logarithmic columns you already built. Because log y = log C − n log x is a straight-line equation, the slope of the best-fit line gives you n and the intercept gives you log C.

The exponent is calculated with the standard least-squares slope formula:

n = [ N Σ(XY) − ΣX ΣY ] / [ N Σ(X²) − (ΣX)² ]

Here X is log x, Y is log y, and N is the number of data points (the number of distinct productivity levels in your table). You compute the column sums for XY, X, Y, and X², plug them in, and the result is your empirical exponent. In a scientometric analysis of brain concussion research, for example, this formula produced an exponent of roughly 1.97, very close to Lotka’s theoretical value of 2.

Once you have n, you find the constant C. Many studies match the calculated exponent against a published table of values, often credited to Rousseau, to obtain the constant directly. Alternatively, the constant can be derived from the relationship that the proportions across all productivity levels must sum to one. A study of Information Science and Library Science journals used precisely this linear least squares method to find both parameters before testing the fit.

The appeal of least squares is its transparency: every step is visible and reproducible in Excel. Its drawback is that fitting a line on a log-log scale can be statistically biased, because the logarithmic transformation distorts the weighting of data points.

Method 2: Maximum likelihood estimation 🔗

Maximum likelihood estimation, or MLE, takes a different philosophy. Instead of treating your frequency table as points to draw a line through, it treats your full set of individual author counts as a sample drawn from an underlying probability distribution. It then asks: what value of n makes the observed data most probable?

The distinction matters. Curve-fitting approaches use the binned frequency data, while the maximum likelihood approach works directly from the individual data points without grouping them into a table. This avoids the loss of information that binning causes. For Lotka’s Law expressed as a power law, the MLE involves the Riemann zeta function as a normalizing constant, which is why it is usually computed with software rather than by hand.

The strong argument for MLE is accuracy. Research on fitting power-law distributions has shown that graphical methods based on a linear fit on the log-log scale are biased and inaccurate, and that maximum likelihood estimation is far more robust. For students working on a serious dissertation rather than a quick classroom exercise, MLE is the methodologically safer choice when tools are available.

Method 3: Method of moments 🔗

The method of moments estimates parameters by matching the theoretical moments of the distribution, such as its mean, to the moments actually observed in your data. You calculate the sample mean number of publications per author, set it equal to the theoretical expression for the mean implied by Lotka’s distribution, and solve for the exponent.

This method is conceptually simple and computationally light, which is its main attraction. However, it is used less often for Lotka’s Law than the other two. Power-law distributions like Lotka’s have heavy tails, and for certain exponent values the theoretical moments are very large or even undefined, which makes the estimates unstable. It is best treated as a useful cross-check rather than your primary estimator.

Testing the goodness of fit 🔗

Estimating the parameters is not the end. You still have to ask whether Lotka’s Law actually describes your data, or whether you have merely forced a curve onto numbers that do not follow it. This is where the goodness-of-fit test comes in, and the standard tool in bibliometrics is the Kolmogorov-Smirnov (K-S) test.

The K-S test compares your observed cumulative distribution against the theoretical one predicted by Lotka’s Law. The test is based on the empirical distribution function and works by finding the maximum absolute difference between the observed and expected cumulative frequencies. This largest gap is the test statistic, often written as D-max.

You then compare D-max against a critical value calculated for your chosen significance level and sample size. If D-max is smaller than the critical value, the difference between your data and the theoretical distribution is not statistically significant, and you conclude that the data fits Lotka’s Law. In a study of artificial intelligence research output, the computed D-max of around 0.007 fell well below the critical value, so the authors concluded the productivity pattern conformed to the law. The widely used bibliometrix R package automates this entire process, estimating the exponent by regression and running the K-S test against the theoretical distribution with the exponent set to 2.

Putting it all together 🔗

The full application of Lotka’s Law follows a clear sequence. You define your dataset and counting rule, build a frequency table of authors by number of papers, transform it into logarithms, estimate the exponent and constant using one of the three methods, and finally test the fit with the Kolmogorov-Smirnov statistic. Linear least squares remains the workhorse for student projects because of its transparency, but maximum likelihood is the more defensible choice when you have the software and the stakes are higher. The method of moments serves best as a secondary check.

What this whole exercise reveals is that bibliometric laws are not magic constants. They are statistical models fitted to real, imperfect data, and the conclusions you draw depend heavily on the choices you make along the way, from how you count authors to which estimator you trust.

What do you think? If two researchers analyse the same dataset but one uses linear least squares while the other uses maximum likelihood, and they reach different conclusions about whether the data fits Lotka’s Law, whose result should we trust? And does the freedom to choose a counting method weaken the claim that Lotka’s Law is a universal regularity of science?

References

1. https://rdrr.io/cran/bibliometrix/man/lotka.html
2. https://revistas.marilia.unesp.br/index.php/bjis/article/download/44/62/188
3. https://www.math.wm.edu/~leemis/2006jscs.pdf
4. https://arxiv.org/pdf/2102.11983
5. https://www.researchgate.net/publication/325218378_Empirical_Examination_of_Lotka's_Law_for_Information_Science_and_Library_Science
6. https://arxiv.org/pdf/1512.08754
7. https://link.springer.com/article/10.1140/epjb/e2004-00316-5
8. https://www.itl.nist.gov/div898/handbook/eda/section3/eda35g.htm
9. https://arxiv.org/pdf/2102.09182