Citation Distribution Modeling (Lognormal/Tsallis)

Because raw counts conflate a paper's merit with its field's citing culture and its age. Biomedicine cites far more, and far faster, than mathematics, so a raw comparison systematically favors high-citation fields and older papers. Radicchi, Fortunato, and Castellano showed that dividing each paper's citations by its field-and-year average removes this bias: the rescaled distributions from all fields collapse onto one universal curve. The resulting relative indicator means the same thing everywhere, so a value of two represents twice the local average whether the paper is in physics or sociology, making cross-field comparison meaningful.

Radicchi and colleagues found that the rescaled universal distribution is well described by a lognormal, which is consistent with citations accumulating through a multiplicative, preferential-attachment-like process. Some researchers argue the extreme tail is closer to a power law, and others fit Tsallis q-exponential or discounted-cumulative forms. Distinguishing these is statistically demanding and depends on how the tail is sampled. For most normalization purposes the lognormal fit is adequate and well behaved; the debate matters mainly when the precise shape of the high-impact tail is the object of study.

It provides its theoretical justification. Many evaluation systems normalize citations by field and year so that institutions or researchers are not penalized or rewarded for their disciplinary mix. The universality result shows that such normalization is not arbitrary: once you rescale by the cohort average, distributions genuinely become comparable. Indicators that compute the ratio of observed to expected citations, or that count papers in field-normalized top percentiles, are operational descendants of this modeling. The main caveat is that the chosen field-classification scheme strongly affects the cohorts and therefore the resulting normalized scores.