Quantifying the dynamics of nearly 100 years of dominance hierarchy research

Abstract

Dominance hierarchies have been studied for almost 100 years. The science of science approach used here provides high-level insight into how the dynamics of dominance hierarchy research have shifted over this long timescale. To summarize these patterns, I extracted publication metadata using a Google Scholar search for the phrase ‘dominance hierarchy’, resulting in over 26 000 publications. I used text mining approaches to assess patterns in three areas: (1) general patterns in publication frequency and rate, (2) dynamics of term usage and (3) term co-occurrence in publications across the history of the field. While the overall number of publications per decade continues to rise, the percent growth rate has fallen in recent years, demonstrating that although there is sustained interest in dominance hierarchies, the field is no longer experiencing the explosive growth it showed in earlier decades. Results from title term co-occurrence networks and community structure show that the different subfields of dominance hierarchy research were most strongly separated early in the field’s history while modern research shows more evidence for cohesion and a lack of distinct term community boundaries. These methods provide a general view of the history of research on dominance hierarchies and can be applied to other fields or search terms to gain broad synthetic insight into patterns of interest, especially in fields with large bodies of literature.

This article is part of the theme issue ‘The centennial of the pecking order: current state and future prospects for the study of dominance hierarchies’.

1. Introduction

Competition is nearly ubiquitous in situations where resources are limited and contested. Because of this, conflict is inevitable in most social groups, leading to increased access to these resources for some individuals and decreased access for others. In many social species, this competition leads to the emergence of group dominance hierarchies, which can help make social life more structured and predictable and regulate overall conflict. In nearly 100 years of research on dominance, scientists have documented the presence of hierarchies that structure social conflict in a wide range of species [1,2]: groups of ants, fish, lizards, geese, parrots, elephants, hyenas, primates and many species in between form groups with detectable hierarchies, where individuals within the groups can be assigned ranks. Much research has also established that these hierarchies also matter to individuals: higher-ranked individuals often benefit from improved health or access to resources, more reproductive opportunities, more offspring, or greater longevity [3–6].

Scientific progress itself is a social process, with new research continually building on the foundations of previous work. A science of science approach can be used to synthesize the history of a large and active field like dominance research. This approach complements more typical literature reviews with a ‘big data’ perspective on publishing patterns and topics in the field. There have been too many papers published on dominance hierarchies in too many subfields for it to be feasible to read and synthesize the entire body of work. A traditional literature review may also be unintentionally biased towards certain subtopics, study species, or subfields. The quantitative approach used here allowed me to summarize a much bigger body of literature to identify general patterns and provide an overall summary of how areas of focus and study have changed over the course of nearly 100 years of research since the original publication describing ‘peck order’ in chickens [7]. However, it is important to note that what we gain from this high-level perspective is balanced by the absence of important syntheses of knowledge that are part of good literature reviews (see other papers in this special issue for this perspective, especially [8]).

Here, my goal was to quantify general patterns of interest in this long-running field of research. Specifically, I focused on three main aspects: (1) general patterns in overall publication frequency and rate, (2) dynamics of changes in term usage in titles and (3) how terms co-occur in publication titles across the history of the field. I used a combination of data scraping, text mining and network analysis to quantify these patterns. This approach provides a broad review of the history of study of dominance hierarchies to better understand where and how researchers have focused their scientific efforts.

2. Methods

(a) . Data collection and processing

I used the program ‘Publish or Perish’ [9,10] to scrape Google Scholar for publications using the search phrase [‘dominance hierarchy’]. The Google Scholar algorithm returns publications which match the search phrase anywhere in the searched documents (author, title, source, abstract, references, etc.) [9]. While the full Google Scholar search algorithm is not publicly available, the algorithm generally works by considering the full text of each document, the publication venue, the authors, and the recency and frequency with which other papers have cited it to rank all publications and return the top 1000 results. I used Google Scholar as the search engine because it has wider coverage and returns more publications than many other searches (i.e. Web of Science, [11]). It also returns a broader array of publication types including ‘grey literature’ like theses, conference proceedings, white papers and preprints [12], the inclusion of which is one way to reduce biases in the literature results [13]. While using Google Scholar is beneficial for its wider reach, there are some important limitations: it does not have strong quality control processes so data are generally not as clean as the more limited output from other sources [11], resulting in some records with missing data, corrupted text, erroneous publication years, etc. Results, including titles, are also sometimes truncated if they are too long [9].

The use of ‘dominance hierarchy’ as the search phrase also has limitations. First, searching for this term heavily influences search results towards English-language literature, which can limit and bias large literature reviews [14,15]. While searching for a translation of dominance hierarchy would have been feasible, matching up all the terms in titles for the analyses here would have required a large amount of translation and unfortunately was not feasible for this project. The second limitation of this search phrase is that it may miss conceptually similar papers, for example, papers focusing on rank, agonistic contests, or contest outcomes, which may not include the phrase ‘dominance hierarchy’ within the paper. This issue is likely reduced by the choice of the Google Scholar search algorithm, which considers the full text of papers, rather than just the terms in the titles or keywords, but is an important consideration.

Using Publish or Perish, I collected the titles of the top 1000 Google Scholar results of publications returned with the search phrase for each year (searches conducted from 14 April 2021 to 17 April 2021). Any records missing a title or a valid publication year were excluded from all further analyses. I also excluded any title with a publication year prior to 1920 to pin the start of analyses to the decade in which the foundational paper for the field, by Schjelderup-Ebbe, was published in 1922 [7]. Analyses were pooled by decade to facilitate identification of general patterns.

(b) . Quantifying publication trends over time

To quantify how publications trends involving the search phrase ‘dominance hierarchy’ have changed over time, I counted the number of publications in the database for each decade. I then quantified the lagged change in publication number by comparing each decade’s total publications to the previous decade. Finally, I quantified the decade over decade per cent growth in publications by dividing the lagged change in publication number by the number of publications in the previous decade and converting to a per cent.

(c) . Estimating dynamics of term usage in publication titles

All titles in the database were cleaned prior to text analyses using the R packages tm and stringr [16,17]: punctuation, numbers and whitespaces were removed, and all was text converted to lowercase. I removed all stopwords (extremely common words like a, is, the, etc.) using the snowball stopword list in the R tidytext package [18]. Prior to full analyses, all terms in the dataset were also stemmed using the tm package. Stemming reduces instances of similar words (e.g. ‘ecological’ and ‘ecology’ both reduce to the stemmed term ‘ecolog’). Limited further stemming efforts for commonly used words were conducted by hand (e.g. both ‘behaviour’ and ‘behavior’ were retained as separate terms following stemming in the tm package; I collapsed these to ‘behavior’). These cleaned and stemmed terms formed the initial corpus and are referred to here as ‘title terms’.

To estimate title term use dynamics, I filtered the terms in the title corpus to retain only the title terms used in at least 25 publications in the database. This process excluded terms that were rarely used but also helped exclude any terms that were reported by Google Scholar in a somewhat corrupted format (for example, punctuation was sometimes introduced erroneously within words in the search results, some words were split by truncation, etc.). This dataset of non-rare title terms was used in all further analyses. I then determined whether each title term was present or absent in each document’s title (this corrects for cases where a particular title term was used multiple times within a single title). I used presence data to then find the total number of publications per decade where each title term was present in the title.

To determine how title terms were used from decade to decade, I quantified the Shannon diversity of title term presence in publications in each decade, using the frequency with which title terms were present in titles compared to the total number of unique title terms. I also quantified the number of novel title terms per decade, where ‘novelty’ was defined as a unique title term which appeared in a title where it was not present in earlier decades in the corpus. Finally, using the entire history of dominance hierarchy title term corpus, I found the per cent of total title terms that were present in titles in each decade.

To measure decade to decade similarity in title term usage, I compared the presence of title terms in each decade to the presence of title terms in all other decades. I then found the number of title terms that were present in both decades, the per cent of words present in both decades, and the per cent similarity of each decade compared to the decade it was most similar to. This analysis helped determine ‘hot spots’ of title term use similarity in documents published across different decades, where the more similar term use was, the more similar those decades would be when compared.

(d) . Determining differences in title term co-occurrence

Knowing that a certain single term was present in titles in a particular decade is helpful for assessing when terms emerge and are popular. However, even more insight into how the focus of dominance hierarchy research has changed can come from quantifying how pairs of terms co-occur in documents through the history of the field. To do this, I built networks based on title term co-occurrences in publication titles for each decade. Title terms formed the nodes in these networks and were connected by edges if the two terms were used in the same title within the same decade. The strength of these edges showed how often terms co-occurred in titles at the decade scale.

From these title term co-occurrence networks, I could then determine if particular title terms in these networks could be assigned to different network communities, depending on how each term was connected to others. Inspired by methods used to reconstruct the cultural evolution of a music genre [19], I used the R package igraph and fastgreedy community detection [20,21], to identify communities of co-occurring title terms by decade, where title terms often found together in publication titles in the same decade were more likely to be assigned to the same title term co-occurrence community. This approach differs from common bibliometric network analyses, which often focus on communities driven by shared co-authorship patterns; here, the focus is solely on the content of titles rather than the identity of the authors. I calculated the modularity of title term co-occurrence networks in each decade, which measures how well the community-detection algorithm partitions a network into communities [20]. Finally, for each decade, I quantified how title term co-occurrence communities were interconnected or separated from each other [22]. To do this, I calculated a cohesion index [19] to represent the ratio of within-community edges compared to connections between title terms assigned to different title term co-occurrence communities, using unweighted binary edges. A cohesion index near 0 indicates that few connections between terms assigned to different communities exist; a value near 1 indicates that most connections occur between terms assigned to different communities.

Finally, because community detection was done independently by decade, I needed a way to track similarity in use of title terms across decades. Title term co-occurrence communities detected in each decade were assigned numerical codes, but ‘Community 1’ in one decade is not necessarily comprised of a similar assortment of title terms in the next decade—in other words, there was no consistent naming continuity in communities across decades. To determine cross-decade community similarity, I identified which title term co-occurrence communities in one decade were the most similar in term composition to a community detected in the next decade. I used Jaccard similarity, which measures overlap in membership between sets as the size of the intersection divided by the size of the union of two sets. In my case, I used Jaccard similarity to find the proportion of title terms found in a title term co-occurrence community in one decade compared to the title term compositions of title term co-occurrence communities in the next decade. The similarity measure ranges from 0 (no overlap in terms) to 1 (exactly the same terms present in both communities). This analysis allowed me to look at how consistently terms were assigned together as a group in one decade compared to the next decade.

3. Results and discussion

(a) . Quantifying publication trends over time

Analysing publication trends can provide insight into overall interest in a field of study. Research on dominance hierarchies has resulted in an impressive number of publications in nearly 100 years of research, with over 26 000 publications in the scraped dataset which were returned from a keyword search for ‘dominance hierarchy’. The Google Scholar algorithm returns a maximum of 1000 results per query, so queries were split by time periods as small as a single year to maximize results (2011 was the only year where 1000 results were returned; all other years were below this thresholding limit and represent complete search results). After data cleaning (which excluded any publication missing a publication year and/or a title, or with a publication date prior to 1922), 25 219 publications were retained for the analyses.

Figure 1 shows three views of publication trends by decade. While publication numbers have risen from decade to decade during the entire history of the field (figure 1a), it is also important to account for the overall increase in modern publication rates. Quantifying how the number of publications in one decade compares to the number of publications in the last decade helps somewhat normalize for this general increase in overall publications in modern science and helps better visualize changes in interest on a decade-by-decade scale. For example, the greatest increase in numbers of papers in a decade compared to the last decade occurred during the 2000s. 2010 was the first decade in which this explosive growth rate decreased: fewer additional papers were published in the 2010s compared to the increased number of publications when comparing the 2000s to the 1990s. The decade-over-decade per cent growth rate provides different insight into publication trends, with the highest per cent growth seen in the 1930s compared to the 1920s. This high growth rate indicates just how quickly the initial number of publications grew when compared to the ‘founding’ of dominance hierarchy research in the 1920s. We see another peak in per cent growth rates in the 1950s to the 1970s with decade over decade growth rates around 300%. These publication trends provide strong evidence for sustained interest in dominance hierarchy research despite nearly 100 years of study, but do indicate that in the most recent decades, the earlier explosive growth in numbers of publications has tapered off.

Figure 1.

Publications by decade showing (a) total publications for each decade, (b) the number of publications in each decade compared to the previous decade, and (c) the per cent growth in number of publications compared to the previous decade.

(b) . Estimating dynamics of term usage

Of 22 406 unique title terms identified via stemming words used in titles in the entire database, only 1295 were used in at least 25 separate publications. I used these more commonly used terms to estimate dynamics of term usage. This 25-publication threshold filtered out a large proportion of potential title terms, but allowed me to focus on ones that were more used and was an easy way to exclude the many corrupted terms found in the Google Scholar search.

While general patterns in the number of publications per decade provide evidence of sustained interest in dominance hierarchy research, how terms are used each decade provides insight into areas of focus for research efforts. I found that the overall diversity of title term use per decade increased sharply up to 1950, then continued to increase at a slower rate until 1990 (figure 2a). The diversity of title term use has been relatively stable from 1990 to 2020 and reflects the highest diversity period in the history of dominance hierarchy research. This high diversity period coincides with high overall numbers of publications during these decades, as seen in figure 1. Interestingly, despite very few years of publications so far in the 2020s, publications from January 2020 to April 2021 already share the high diversity of title term use seen across the whole 2010 decade. Overall patterns of term diversity and network measures (see below) are unlikely to result from decade-to-decade changes in publications norms such as title lengths; title length was relatively consistent across the dataset with the median number of title terms per title per decade ranging from 3 to 6 and an average median number of title terms per publication per decade of 4.77.

Figure 2.

Title term usage in titles by decade showing (a) Shannon diversity in title term use in each decade (with diversity calculated on total number of publications using each title term per decade), (b) the number of novel title terms introduced in each decade and (c) the per cent of title terms in the entire corpus that were used in titles in each decade.

The number of novel title terms introduced in publication titles peaked sharply in 1970 (figure 2b). On a decade by decade basis, figure 2c shows how terms present across the entire history of dominance hierarchy research are used in publication titles: by 1970, nearly 75% of all title terms in the corpus were in use.

Each decade can also be compared to other decades in the dataset to determine the levels of similarity in title term use in titles over time (figure 3). The number of title terms present in both decades peaked when comparing 1990, 2000 and 2010. Breaking these patterns down by the per cent of terms present in both decades (compared to the number of terms present in either decade) shows an even larger hotspot of similarity that highlights how similar term use in 1980–2020 has been to other decades within that same time span. Comparing term use in each decade to the decade with which term use is most similar shows an even wider hotspot coinciding with all decades compared to 1990–2010; this hotspot also coincides with data in figure 2c, showing that nearly 100% of title terms used in the entire historical corpus were used in 1990–2010.

Figure 3.

Title term usage in publications showing ‘hotspots’ of similarity across decades: (a) similarity in the raw number of title terms present in both decades, (b) scaled similarity showing the per cent of title terms present in decades on the y-axis compared with the title terms present in decades on the x-axis and (c) relative similarity showing each decade’s similarity scaled by maximum similarity to decades on the y-axis, with maximum similarity shown in red.

(c) . Determining patterns in title term co-occurrences

In addition to quantifying overall publication trends and the use of single terms in titles, the co-occurrence of terms in titles can provide insight into areas of focus for dominance hierarchy research and how the use of pairs of terms in titles has changed over time. Using network measures (eigenvector centrality and community detection) I found each term’s community membership in each decade, with community membership determined by how each term co-occurred in titles with other terms. These term co-occurrences and community assignments across decades are depicted as wordclouds in figure 4, visualized with the R package ‘wordcloud’ [23].

Figure 4.

Wordclouds showing title term community assignment across decades. Text size indicates each term’s eigenvector centrality in the title term co-occurrence network for that decade, with higher centrality terms printed in larger text. Text colour indicates each term’s assignment to a community in each decade; terms in the same colour within the same decade were assigned to the same title term co-occurrence community. To improve legibility, wordclouds show (at most) the top 10 most central terms per community per decade.

Figure 5 shows how terms can shift in how they co-occur in titles over time. In the figure, title term use and title term co-occurrence community membership are shown for the top three most-used terms (‘behavior’, ‘social’ and ‘domin’). In 1930, all three title terms were assigned to separate communities, while in 1940, the title terms ‘behavior’ and ‘social’ were grouped together in Community 1 and ‘domin’ was assigned to Community 2. In 1950 this pattern changed again as ‘behavior’ was assigned to Community 1 and ‘social’ and ‘domin’ were both assigned to Community 2. This pattern illustrates that title term co-occurrence community membership composition changed from year to year, with even the top most-used terms changing in how they were grouped. Interestingly, ‘social’ and ‘behavior’ both appeared in a higher proportion of titles in almost every decade than ‘domin’, which could be an indication of a long-standing trend in dominance hierarchy research for placing dominance within a social and behavioural context.

Figure 5.

Title term use and co-occurrence community membership over time for the top three most-used title terms: ‘behavior’, ‘social’ and ‘domin’. Points show the proportion of titles containing each title term per decade; point colour indicates title term co-occurrence community membership.

While the composition of title term co-occurrence communities has changed over time, the distinctness of borders between these communities has also changed. Figure 6 shows how title term co-occurrence networks were much more modular early in dominance hierarchy research, but have decreased in modularity in recent years, indicating that terms have become less-strictly co-occurring with specific other terms in titles. The number of detected communities has also shifted from a small number of communities to a peak of 12 detected communities in 1950, followed by a gradual decline to fewer communities in modern dominance hierarchy research (it is uncertain if the rise in community number in the 2020s will persist as more papers are published in this decade so those results should be treated with caution).

Figure 6.

Title term co-occurrence network summaries by decade: (a) modularity over time, (b) number of communities detected in each decade and (c) per cent of edges connecting a title term node in one community to a title term node in a different community.

Within decades, the per cent of edges between title terms assigned to the same community compared to title terms assigned to different communities has changed over time. Connections between terms assigned to different communities have increased over the history of dominance hierarchy research to plateau in modern times at about 50%, indicating a balance between within-community and outside-community edges. This connectivity pattern contributed to the decrease in overall modularity as communities became more interconnected as well as the reduction in the total number of detectable communities. Whether this decrease in modularity comes from more integrative studies or potentially from a increase in cross-disciplinary work that makes more connections to research across subfields remains to be seen. Future work focused on citation networks, analysing how authors cite each other’s work, could provide important insight into this process in the field of dominance hierarchy research (e.g. [24]).

Across decades, title term community continuity was highest from 1990 to 2010, with an average maximal Jaccard similarity of over 25% shared title terms for a community in one decade compared to the set of title terms in the most-similar community in the next decade. This period of more stable title term assignments to similar communities co-coincided with a period of fewer overall identified communities. This result also shows that even in decades with few overall communities, the majority of title terms in any one community were not assigned together in the same community in the next decade, indicating a high level of remixing of title term use in publication titles.

4. Conclusion

The science of science perspective used here provided insight into general publication trends, title term use and term co-occurrence in titles returned from a search for ‘dominance hierarchy’ in publications across the nearly 100 years of dominance hierarchy research. From this, we can infer how investment in publishing dominance hierarchy research has changed, but also how connectivity between different subfields and topics has shifted.

An important limitation of this approach is that the analysis only considers terms used in titles of publications. This analysis obviously cannot capture the complexity of how topics are treated in the full text of these publications, so cannot provide a detailed account of exactly how research trends or concepts have shifted over the history of dominance hierarchy research. However, this summary demonstrates that dominance research has had sustained interest over its long history and the evidence, especially the high connectivity between what could have been isolated subfields, provides a high-level perspective on historical and modern trends in the science of hierarchies.

An open question among researchers working on dominance hierarchies is whether we have ‘solved’ dominance. Informally, and depending on who you talk to at conferences, the question of dominance has been ‘solved’ in the 1960s, the 1970s, the 1980s or the 2000s. However, the sustained interest and investment in new publications demonstrated here, as well as greater cohesion and cross-community connections, suggests that there is still much interest in all that we still have to learn about dominance hierarchies.

Recent dominance studies may be moving to a new stage of research focus, particularly if we have solved some of the more basic hierarchy questions. In particular, new genetic methods (e.g. [25]), computational approaches [26], and a focus on the information contained in both networks of aggression and rank within social groups [26,27] provide many new avenues for novel insight into animal sociality. Both theoretical work (e.g. [28,29]) and empirical work [26] have also recently suggested that rank acquisition can be remarkably sensitive to stochastic events. This new work has the potential to enrich our understanding of how rank forms and is maintained in different groups [29]. Other empirical work has shown that ‘rule-breaking’ via coalition formation can cause disruptions to expected rank inheritance patterns [30], and that these dynastic changes can gain momentum and persist despite the lack of underlying characteristics or quality to differentiate these individuals from less-successful ones in the group. Finally, comparative analyses across species have strong potential to advance our understanding of dominance hierarchy structures. New compilations of dominance data across species, such as in the R package DomArchive (see Strauss et al. [30]), provide easier access to historical datasets which researchers can use to test hypotheses about dominance and how the social and cognitive features required for dominance to emerge may have evolved. These datasets can also form the basis for new analyses to understand how rank influences individual health and how competition can influence the outcome of fitness-related traits more broadly.

The integration of new tools as well as new, more complex ways of studying the decisions animals make about who, when, and how they fight each other, and the consequences of different conflict management styles, provide a strong foundation for the next 100 years of dominance hierarchy research.

Data accessibility

All code and data are available at https://doi.org/10.5281/zenodo.5736801.

Competing interests

I declare I have no competing interests.

Funding

No funding has been received for this article.