Report

An Empirical Assessment of Taxic Paleobiology

See allHide authors and affiliations

Science  07 Jul 2000:
Vol. 289, Issue 5476, pp. 110-112
DOI: 10.1126/science.289.5476.110

Abstract

The analysis of major changes in faunal diversity through time is a central theme of analytical paleobiology. The most important sources of data are literature-based compilations of stratigraphic ranges of fossil taxa. The levels of error in these compilations and the possible effects of such error have often been discussed but never directly assessed. We compared our comprehensive database of trilobites to the equivalent portion of J. J. Sepkoski Jr.'s widely used global genus database. More than 70% of entries in the global database are inaccurate; however, as predicted, the error is randomly distributed and does not introduce bias.

The publication of J. J. Sepkoski Jr.'s (1) factor-analytical description of the marine fossil record was an epochal event in modern evolutionary paleobiology. The compilation of marine families on which it was based (2, 3) has served as the raw material for many influential papers, including studies of extinction (4–7), the periodicity of mass extinction (8–10), and evolutionary rates (11–15). Recognizing the need for a more detailed level of analysis, Sepkoski (16) began a more ambitious compilation of fossil genera (17), which now serves as the foundation for the majority of current work in the field.

There are critics of taxic paleobiology (18). Some have pointed out that taxa of a particular Linnean rank have no natural equivalence (19), and others (20, 21) that traditional taxonomy contains a large number of polyphyletic or paraphyletic groups, which hamper the estimation of large-scale pattern (22). The most widespread complaint (23,24), however, has been that the basic accuracy of global databases is suspect because they are compiled by workers who are not systematic specialists. Because nonspecialists have no means to evaluate or correct taxonomic data, they might compile considerable error, and this error might significantly bias the result. This charge has been answered on conceptual grounds by proponents of taxic paleobiology, who claim that statistical treatment of the data in itself presumes that error is present and accounts for it (18), and that even large amounts of error, if randomly distributed, are unlikely to bias the underlying pattern (25). Others have pointed to the fact that major revisions of large databases have had little impact on the results produced by the use of those databases (26). A seminal study by Sepkoski himself (27) demonstrated that significant additions and corrections to his family compilation had minimal effect on diversity patterns. To date, however, there has never been an empirical assessment of either the amount of error present in large databases or of its effect.

We tested the accuracy of the Sepkoski genus database by comparing a portion of it to our own independent database of trilobite genera (28). Our database has been assembled over a period of years by systematic specialists and represents a taxonomically standardized (29), critically evaluated compilation that we regard as essentially complete and exhaustive for the Ordovician and Silurian periods. Although it might still contain errors and omissions, it is considered “correct” data against which to compare the equivalent portion of the Sepkoski database. Taking the study interval from the beginning of the Ordovician through the end of the Lower Silurian (Wenlock), the Sepkoski compilation yields a sample size of 941 trilobite genera. Trilobites are among the most common fossils of the Lower Paleozoic (28) and are likely to be a good proxy for other components of the global genus database.

Our sampling intervals for the Lower Silurian, the five global standard stages, are identical to Sepkoski's (2,3). We used 9 intervals for the Ordovician, in contrast to Sepkoski's 13. Problems of global correlation, regional biostratigraphic resolution, and reported resolution make his more detailed scheme impossible to apply to records from many parts of the world, and our scheme is based on widely correlated biohorizons (30). This creates no difficulty, because Sepkoski's scheme is easily and directly transferable to our own.

Of Sepkoski's trilobite data, the stratigraphic ranges of about one-third of the records are not resolved to his sampling scheme (31). Sepkoski (16, 32) had explicit strategies to distribute the known error introduced by this. We followed Sepkoski in distributing the resolution error by assigning the bin-occurrence of poorly resolved taxa in proportion to that of the fully resolved taxa, and we calculated range error with known errors both included and excluded (that is, using only the fully resolved data).

We compared the databases by checking each genus in the Sepkoski compilation and assessing its validity and stratigraphic range relative to our estimate. We also tracked its stratigraphic interval occurrence, both as presence and absence, and scaled to include the distribution of uncertainty in stratigraphic resolution. If a genus was for some reason not accepted (for example, if it was a junior synonym of a previously established taxon), all of its assigned occurrence was treated as error. If a genus was accepted, then the accuracy of its stratigraphic first appearance (FA) and last appearance (LA) was recorded. If either was in error, the direction and magnitude of that error was tabulated.

Our database contains 1383 historically proposed trilobite genera with occurrence in the study interval, of which 389 are junior synonyms, leaving 994 valid genera. The Sepkoski database lists 941 genera for the same interval, of which 681 contain error of some kind (Fig. 1): 202 are invalid records for various reasons and 479 are valid genera with errors in FA, LA, or both. Only 260 of 941 records are valid genera with accurate ranges (33). Figure 2 shows the distribution of error in reported FA and LA among fully resolved records.

Figure 1

Accuracy of genus records in the Sepkoski trilobite database (N = 941 total genera). 1, valid genera with accurate stratigraphic ranges; 2, valid genera with incorrect stratigraphic FA but correct LA; 3, valid genera with incorrect LA but correct FA; 4, valid genera with incorrect FA and LA; 5, junior synonyms; 6, valid genera with LA before or FA after the study interval; 7, other errors: uninterpretable genera restricted to types, misspelled repetitions, and nomina nuda.

Figure 2

(A) Proportion of all range data among fully resolved records that is incorrect, by interval. (B) Proportion of total reported FA and LA among fully resolved records that is incorrect, by interval.

It is evident that most of the data in the Sepkoski trilobite data set are erroneous. The database captures only 74% of the valid genera for the interval and includes an additional 20.3% of taxonomic “noise,” while 55.2% of raw reported FA (46.5% of fully resolved FA) and 53.6% of reported LA (45.0% of fully resolved LA) are erroneous. The magnitude of raw individual sampling bin error (the reported occurrence of a genus in a sampling interval that is incorrect, or failure to report a valid genus in an interval in which it occurs) is 52.9% that of the total taxon-bin occurrence. And finally, only 27.6% of the records in the database comprise valid genera with correct stratigraphic ranges.

Does this error result in bias? Figure 3 plots the cumulative genus diversity by sampling interval as indicated by both the raw Sepkoski data and our own data, as well as the respective percentage changes in diversity from interval to interval and the distribution of taxonomic “noise” by interval in the Sepkoski data set. The raw Sepkoski data describe a curve nearly identical to, and almost coincident with, our own. AG test of goodness of fit between the two curves demonstrates that they are not significantly different (P > 0.70). The Sepkoski data track both the direction and magnitude of diversity changes almost exactly. Cumulative diversity curves differ most at the upper and, especially, lower ends of the Ordovician-Wenlock study interval, and it is clear that this is due to higher levels of taxonomic noise (Fig. 3B). Noise levels are elevated at each end of the distribution by the inclusion of genera whose stratigraphic durations actually fall in entirely older or younger strata. Such errors are less likely to be a factor toward the midpoint of the distribution, where noise levels are reduced. The bias toward more noise at the lower end of the distribution (that is, the earliest Ordovician) reflects historical uncertainty in the position and correlation of the Cambrian-Ordovician boundary.

Figure 3

(A) Cumulative genus diversity by interval, as indicated by the Sepkoski data and by our data. (B) Distribution of taxonomic “noise” in the Sepkoski data set. (C) Direction and magnitude of the percent change in diversity from interval to interval as indicated by either data set.

Stratigraphic range error appears to be a less important factor. With the exception of a few sample bins, the frequency of resolved range error is essentially constant (Fig. 2A) and plots in a narrow band around a mean value of 47%. Errors in raw first or last appearances are symmetrically distributed (Fig. 4). Stratigraphic ranges are commonly overestimated, but in similar amounts in either direction, so that no pervasive bias is introduced. These results indicate that studies in analytical paleobiology that are based on large compilations of data are likely to be highly resilient to error.

Figure 4

Frequency distribution of magnitude and direction of range error among valid Sepkoski genera.

We conclude that Sepkoski's global genus database, although rife with error and of little value for low-level systematic studies, is adequate for its intended application. As far as can be determined, it accurately estimates the large-scale patterns of Phanerozoic biodiversity, and its widespread use in current studies of analytical paleobiology is justified.

  • * To whom correspondence should be addressed. E-mail: jonathan-adrain{at}uiowa.edu

REFERENCES AND NOTES

View Abstract

Navigate This Article