3. Occurrences of semantic fields across the entire corpus

The relative importance of the targeted semantic fields throughout the corpus can be estimated by examining the average frequency of corresponding groups of terms per article (fig. 2).

Figure 2. Average occurrence of the semantic fields per document over the complete corpus. (Semantic fields are grouped and colored by type. Surface area is proportional to the average number of occurrences of terms related to each semantic field; to give an idea of scale: suggest, in the upper right-hand corner, has an average of 5.1 occurrences per article. Numerical values are available in the “Data_for_graphs” file; see Supplementary Information section.) Figure in color online.

Disciplinary fields characterize the disciplinary orientations of the corpus (left-hand side of the graph in fig. 2, in beige). Unsurprisingly, they also match the disciplinary clusters of articles identified by Malaterre and Léonard (Reference Malaterre and Léonard2023) (table 2). A dominant share of terms concerns research about cells, genes, and molecular biology in a broad sense, as denoted by the relative significance of the semantic fields cell , gene , protein , and molecular (this result accords well with the presence of the article cluster Cell and molecular biology but also with the clusters Genetics, disease biology, others and Genomics and phylogenetics). A biomedical orientation is clearly visible through the strong presence of terms related to cancer , therapy , clinical , and pathology , as well as mortality-survival (which makes sense, given the presence of such article clusters as Cancer research; Genetics, disease biology, others; and Health and clinical research in the corpus). The semantic field demographics captures the presence of terms that relate to the description of patients, as is typically the case in clinical trials and public health studies; the fields care and behavior are also connected to health issues, notably mental health (these last three fields partly correspond to the clusters Health and clinical research and Public health). The fields animal and human denote research with animal models, as can be the case in cellular biology and cancer research (as in the clusters Cancer research and Cell and molecular biology). The three fields evolution , selection , and phylogenetics characterize research in evolutionary biology and phylogenetics (as is distinctive of the cluster Genomics and phylogenetics). A relatively less frequent disciplinary field is algorithm , which denotes research in bioinformatics (this field relates to the article cluster Bioinformatics and methods).

The four types of epistemic fields are found on the right-hand side of the graph (fig. 2). Among the epistemic fields that relate to epistemic objects (in blue), several of the most significant fields denote the centrality of modeling in the practice of science. This is the case, of course, for the field model but also for data , mathematics-statistics , parameter-variable , and estimate . By contrast, theories and generalizations are quite rarely used, as shown by the reduced size of fields such as theory and generalization . In between these extremes, there appears to be a moderately significant usage of hypothesis-related terms, as denoted by the medium-size field hypothesis . The results also show quite a strong usage of causal notions, as revealed by the relative importance of such fields as cause , effect , and response . Finally, specific targets of knowledge, such as mechanisms, structures, patterns, and processes, are also found in the corpus, as denoted by the presence of the corresponding epistemic fields mechanism , structure , pattern , and process .

The second-most-represented epistemic type concerns discovery-related epistemic fields (in orange, fig. 2). Terms that relate to analysis and observation are among the most frequent, as denoted by the significance of the corresponding epistemic fields analysis and observation . Terms that concern experiments and tests are also quite significantly present throughout the corpus, as shown by the relative importance of experiment and test . Note the relatively frequent mention of methods-related terms, as captured by the epistemic field method-approach .

When it comes to modalities of justification (in red, fig. 2), one notes the dominance of weak forms of justification that consist of suggestions, possibilities, or probabilities, as denoted by the predominance of the corresponding semantic fields possible , suggest , probable , and even alternative . The second-most-significant group of justification-related epistemic fields includes somewhat stronger forms of justification, such as confirmation or corroboration, with the fields confirm and support . On the other hand, negative forms of epistemic appraisal that would notably consist of refuting or establishing incompatibilities appear relatively rare, as shown by the relatively infrequent usage of terms that relate to inconsistency , for instance. Similarly, very strong forms of justification—for instance, in terms of truth or certainty—are also rare, as shown notably by the reduced size of the field certainty .

Finally, when it comes to epistemic virtues (in green, fig. 2), the dominant semantic field is predict . This shows the importance of prediction in science, which is much more significant than other virtues, such as explanation or understanding (as characterized by the fields explain and understand ). Other virtues that are notably present, although less frequent, are accuracy and, to a lesser extent, simplicity, as denoted by their respective epistemic fields.

Note that several epistemic fields are absent from the graph, their presence in the corpus being too low (average occurrence ≤ 0.25 terms/article). Among epistemic objects, law is noticeably absent, showing that the notion of law (e.g., as in “law of nature”) is clearly not mobilized in the biological and biomedical sciences of the corpus, despite the amount of conceptual work the notion has triggered in philosophy of biology, not to mention philosophy of science (for a review, see, e.g., Hamilton Reference Hamilton2007). In terms of epistemic virtues, the field unification is also hardly present at all. This notion seems to play no effective role in the practice of science, not even in connection with explanation or understanding, which in particular seems to run contrary to unification-based accounts of explanation (e.g., Friedman Reference Friedman1974; Kitcher Reference Kitcher1985). Several epistemic fields that relate to justification are also remarkably absent. This is the case for the field refute , which indicates that the notion of refutation is hardly at work in the practice of science. In fact, the results tend to show that epistemic justification mostly happens in terms of possibility and probability: negative forms of justification are rarely mobilized, if at all. In short, science is more probabilistic and Bayesian than Popperian. The field coherence is also very rare across the corpus, which shows that the notion is not much used: the fact that hypotheses or models exhibit some form of coherence with other hypotheses or models and the rest of science appears to play little or no role in practice. Similar comments apply to the field analogy . Despite extensive research by philosophers (for a review, see, e.g., Bartha Reference Bartha and Edward2019), the notion of analogy hardly seems to be mobilized, at least explicitly, in actual scientific articles. Finally, note that the field reproducibility is also barely present, if at all. Even if reproducibility and replicability should be desirable properties of scientific knowledge, the notions are not discussed in articles.

4. Correlations between semantic fields across the entire corpus

In order to analyze not just the quantitative significance of epistemic fields but also their relatedness with one another, correlation analyses were run over the whole corpus, resulting in the correlation network shown in fig. 3 (Here, semantic fields are not colored depending on their type, as was the case in fig. 2, but depending on the subnetworks they tend to form when frequently used in conjunction with one another in article paragraphs). As is immediately visible, the correlation network is dominated by three large subnetworks of semantic fields, with a small fourth one attached to the most central subnetwork.

Figure 3. Correlation network of the most significant semantic fields over the whole corpus. (Correlations calculated at the paragraph level; for analysis, only correlations of >0.1 were kept; all ps < .001. Colors indicate correlation clusters based on Louvain community detection, node size is proportional to average occurrence in the corpus, and edge thickness is proportional to correlation strength; rendering was done with ForceAtlas2 on Gephi [Bastian et al. Reference Bastian, Heymann and Jacomy2009]. To give an idea of scale: suggest, in the upper-middle part of the network, has an average of 5.1 occurrences per article. Numerical values are available in the “Data_for_graphs” file; see Supplementary Information section.) Figure in color online.

The largest subnetwork (18 semantic fields, center position in fig. 3, in blue) includes several disciplinary fields that relate to molecular biology, genetics, and cellular biology (in decreasing order of occurrence: gene , molecular , cell , and protein but also human and animal ). The epistemic field role-function is one of the fields that are most connected to these disciplinary fields, meaning that the identification of functions and roles is one major epistemic target for research on molecular processes or entities, notably proteins and genes (see text excerpts A10–18 in table S1, available in the Supplementary Information). Response and pattern are also frequently used in connection with molecular (A20–21). Role-function is correlated with process but also, and more significantly, with suggest , which is in turn correlated with evidence and observation ; this tends to indicate that the role or function of molecular objects such as proteins is often, at most, suggested by observation and evidence (A30–34). This hypothetical or suggested epistemic status is also reinforced by the relatively frequent usage of possible (A35–38). Another significant epistemic object in this subnetwork—although only about half as significant as role-function —is mechanism . This field is strongly connected to molecular but also to gene and cell , thereby corroborating the significant role that mechanisms play in molecular and cellular biology. Similarly to role-function , mechanism is also significantly correlated with suggest , indicating an epistemic status that is often taken to be hypothetical (A40–41). Mechanism is also frequently found in conjunction with underlie , as well as with clarify and understand . Mechanisms are thereby addressed by scientists as things that underlie a phenomenon of interest (e.g., an effect, a disease) and that are to be uncovered and clarified (A42–44). The correlation of mechanism with understand is ambivalent: it could mean that mechanisms provide some form of understanding, for instance, of molecular phenomena (as is defended by the new mechanicist philosophy). But it could be taken to mean that mechanisms are things that need to be understood. Most text excerpts by far corroborate the latter (A45–48). This is not to say that mechanisms never provide understanding but that they are usually rather thought of as objects that need to be understood. This finding attributes a quite different epistemic role to the concept of mechanism than the one identified by the mechanicist philosophy (Machamer et al. Reference Machamer, Darden and Craver2000): mechanisms are rarely thought of as epistemic objects that have a form of explanatory epistemic virtue (understanding) but are more often conceived as research objectives, so to speak, in a context of discovery.

The second-largest subnetwork (in green in fig. 3, left-hand side) includes disciplinary fields that relate to data analysis and the elaboration of models: data , analysis , mathematics-statistics , and model . The field data plays a very central role (in conjunction with information ): it is strongly correlated with interpretation and analysis , the latter being, in turn, correlated with mathematics-statistics (and with the disciplinary field gene ). This suggests that data, in practice, are often submitted to interpretation and analysis that are mediated by mathematical and statistical approaches (see excerpts B10–19). The epistemic field mathematics-statistics is also strongly correlated with the fields parameter-variable and model , the latter also being correlated with fit . This shows the importance of model fitting (B20–22), together with the strong role of statistics as the basis for both model building and model testing (as shown by the significant correlation of test with mathematics-statistics ) (B23–26). Interestingly, the notion of test is much more connected to the notion of model (through the use of statistical tests) than to the notions of hypothesis, theory, and experiment. But note the presence of hypothesis and its correlation with model and mathematics-statistics : this is explained in particular by the practice of null-hypothesis testing in modeling (B27) but also by the testing of more general hypotheses, as would be expected in science (B28–29). The epistemic function of models is clearly one of prediction, as shown by the strong correlation between model and predict (B30–33). In this respect, accuracy is the most sought-after epistemic virtue (strong correlation between predict and accuracy ) (B34–36). In the disciplinary context of the present corpus, prediction notably concerns the structure of molecular entities such as proteins and others (correlation of predict with structure and protein ) (B37–39). On the contrary, note the absence of a significant correlation between model and explain or understand . This means that neither explanation nor understanding is found to be a desirable characteristic of models, a finding that runs somewhat contrary to theses defending the explanatory role of models in science (e.g., Strevens Reference Strevens2008; Potochnik Reference Potochnik2017).

The third largest subnetwork (in red in fig. 3, right-hand side) has to do with the health sciences as broadly construed. Note the central position of demographics in this cluster, which relates to the frequent usage of terms describing populations of patients. This field is most strongly correlated with pathology , therapy , and clinical , which denotes research on public health and epidemiology, as well as clinical studies (C10–15). It is also correlated with care and support , denoting research on health support and health care (C16–19) (note that support is not used here in the sense of epistemic justification). Demographics is correlated with mortality-survival and cancer , too, denoting the presence of cancer research in the corpus (C20–23). This latter field sits at the border between the health sciences subnetwork and the molecular biology subnetwork, in particular because of its strong correlation with cell and, to a lesser extent, with molecular . From an epistemic viewpoint, causal notions appear to be relatively frequent in this health sciences subnetwork, and possibly one of its epistemic characteristics, as shown by the size of cause and effect , notably in conjunction with the disciplinary fields pathology and therapy (C30–33). Interestingly, the field experiment is frequently used in this disciplinary context (correlation with clinical ) and not in the molecular biology context (as characterized by the central subnetwork, in blue in fig. 3). Text excerpts indicate that terminological usage often relates to experiments on animal models and comparison with clinical data (C40–41).

Finally, the fourth subnetwork (in pink in fig. 3) denotes research in evolutionary biology and phylogenetics. This subnetwork comprises only three disciplinary fields ( evolution , selection , and phylogenetics ) and no epistemic field at this level of analysis. Although conjunctions of terms pertaining to evolution and selection usually tend to characterize research on Darwinian evolution (D10–11), conjunctions of terms related to evolution and phylogenetics appear to specifically pick out taxonomical works (D12–13). In any case, most of these studies are gene based (hence the connection of this subnetwork to the field gene of the molecular biology subnetwork).

Note that the overall network comprises 48 of the 77 targeted semantic fields, meaning that 29 fields were left out. This can be explained by their relatively low level of occurrence in corpus articles and by insufficient correlation with other fields. In particular, missing fields of the type object include property , generalization , law , and theory . This indicates that corresponding epistemic objects are actually rarely mobilized in scientific practice, despite being frequently discussed in philosophy of science (e.g., Nagel Reference Nagel1961; Carnap Reference Carnap1966; Suppe Reference Suppe1989; Savage Reference Savage1990; Hamilton Reference Hamilton2007). Noticeable justification fields that are absent are analogy and coherence , as well as certainty and confirm (as noted in sec. 3). Although it could be expected that certainty and truth would not be much present (being too strong for the biological and biomedical sciences), the results show that confirmation-related concepts are also very little used in the practice of science, which can appear somewhat counterintuitive. Finally, noticeable missing fields related to epistemic virtues include explain , unification , integration , and simplicity . Explanation therefore does not appear to be a much-discussed topic during the course of scientific investigation (nor does understanding, as discussed earlier). Similarly, unification, integration, and simplicity do not play a significant role. This contrasts quite starkly with the attention these concepts have attracted from numerous philosophers, notably since the 1950s (e.g., Hempel Reference Hempel1965; Friedman Reference Friedman1974; Kitcher Reference Kitcher1985; Mitchell Reference Mitchell2003; Salmon Reference Salmon1989; Sober Reference Sober2015).

5. Effect of disciplinary context on semantic fields

To investigate whether the usage of epistemic concepts depended on disciplinary context, we made use of the relative diversity of the articles found in the corpus. Although all articles stem from the biological and biomedical sciences, broadly speaking, there are notable differences in terms of subdisciplines, ranging from molecular biology to phylogenetics and clinical research. To capture these differences, we used the article clustering conducted by Malaterre and Léonard (Reference Malaterre and Léonard2023) to build semantic field correlation networks for all seven disciplinary clusters (see fig. 4). As could be expected, the relative significance of disciplinary fields reflects the disciplinary orientations of the article clusters (as depicted by their topical profiles; see table 2; for instance, as shown in fig. 4b, cancer is a dominant field in the correlation network stemming from the cluster Cancer research). In what follows, we mostly leave aside disciplinary fields and focus on epistemic fields. The networks partly display features that were already apparent at the level of the entire corpus but also—and this is what matters here—specificities that depend on clusters’ disciplinary orientations and their topical profiles.

Figure 4. Correlation network of the most significant semantic fields for the seven disciplinary clusters. (Correlations calculated at the paragraph level; for analysis, only correlations > 0.1 were kept; all ps < .001. Node size is proportional to average occurrence in the cluster; edge thickness is proportional to correlation strength. Subnetwork identification is based on Louvain community detection; subnetwork colors depend on the size of the network, from the largest subnetwork in blue to the smallest subnetwork in brown. Rendering was done with ForceAtlas on Gephi. To give an idea of scale: model has an average of 25.6 occurrences per article in network (a) and 6.1 in network (g). Numerical values are available in the “Data_for_graphs” file; see Supplementary Information section.) Figure in color online.

Consider the correlation network for the cluster Bioinformatics and methods (fig. 4a): the fields model , fit , parameter , and mathematics-statistics are all strongly correlated, showing the importance of statistical model fitting and testing (CA10–11). Refute , which was absent from the corpus-level network, is here seen in conjunction with hypothesis : refutation is therefore sometimes at work in science, although not very frequently, and in relation to hypotheses, notably null hypotheses, but not theories (CA12–13). Note also the usage of causal notions such as effect and cause in connection with hypothesis , typically denoting that hypotheses are sometimes about observed effects and possible causes (CA14–15). Note the correlation threads between model , predict , structure , and protein : as noted earlier, models are often used to predict, and such prediction often targets the structure of molecular entities such as proteins (CA20–21). A key virtue of prediction clearly is accuracy, as denoted by the strong correlation between predict and accuracy (CA22–23), with such accuracy resulting from the use of methods (correlation with method-approach ) that are specifically developed for improving predictive accuracy (CA24–25). Confirmation tends to apply to prediction (correlation of confirm and predict ) (CA26–27, although sometimes confirmation concerns prediction methods, e.g., CA28). The relatively strong correlations between data , method-approach , and analysis indicate the importance of data analysis, and additional correlations with selection reveal the use of feature- and variable-selection methods and data-reduction techniques (CA30–32). Note the epistemic field integration and its correlation with data , which captures the practical issue of integrating data from multiple sources (CA33–34) rather than theoretical integration (as is the case in, e.g., Mitchell [Reference Mitchell2003]). The field mechanism exhibits some correlation with understand and molecular , but its role turns out to be modest (low frequency and connectivity). As noted earlier, these correlations usually do not denote any explanatory role of mechanisms but rather point to mechanisms as objects in need of understanding (CA40, but see CA41). Both role-function and property appear correlated to protein (in addition to structure , as mentioned earlier), meaning that proteins are the focus of much research on their characteristics (CA50–52).

The correlation network of the cluster Cancer research (fig. 4b) exhibits strong correlations between model , mathematics-statistics , and parameter , as in the network of Bioinformatics and methods, denoting the importance of statistical model building (CB10–11). One notable specificity is the correlation of model with animal , which indicates the frequent use of animal models in this domain of research (CB12–13), even when statistics are involved (CB14–15). As also noted earlier, the main epistemic virtue of models is prediction, with accuracy if possible (as shown by the correlation between model , predict , and accuracy ). Here, prediction is directed toward a specific objective: predicting survival or mortality (correlation with mortality-survival ) (CB16–17). The fields data , method-approach , analysis , mathematics-statistics , and interpretation are again all connected, showing that data are often statistically analyzed and interpreted (CA18–19). Defining a subnetwork of its own, mechanism is found to be mostly correlated to understand , underlie , and clarify ; these correlations actually denote that mechanisms are things to be understood, not things that provide understanding (CB20–22). In addition, discussions often concern how-possibly mechanisms, as shown by correlations with possible and suggest (CB23), and the elucidation of related roles or functions (correlation with role-function ) (CB24). Note the correlations between support , evidence , and hypothesis , highlighting the role of evidence in providing epistemic support, notably to hypotheses (CB30). Interestingly, explain is found to be correlated with possible : often, explanations of specific phenomena are not known but offered as possible (CB31–35). This indicates the key role of how-possibly explanations, over and above how-actual explanations. There is also a correlation between confirm and gene : confirmation is here predominantly tied to the presence or role of specific genes (CB36–38), as opposed to epistemic objects such as hypotheses, models, or theories (as could have been expected).

In the Cell and molecular biology network (fig. 4c), the fields model and animal are strongly correlated (as in the cluster Cancer research), denoting the prevalence of models in the sense of animal models (CC10–13). Note the absence of correlation with predict in this disciplinary context; this corroborates the presence of a different type of model concept in this article cluster compared to Bioinformatics and methods, where models are used for prediction. This also makes sense given the correlation of model with pathology , which notably captures research on animal models of specific diseases (CC14–15). As in other clusters, data , analysis , mathematics-statistics , and interpretation are all correlated. A specificity is the correlation with confirm . Data are often said to confirm other epistemic objects, for instance, other studies or other results (CC20–21), although the reverse can also be the case: sometimes results or findings are said to confirm data (CC22). As seen in other clusters, mechanism turns out to be correlated with underlie , understand , and clarify . It is also correlated with role-function , which may denote different use contexts: for instance, a system can have a role that is made possible by an underlying mechanism (CC30); elsewhere, some entity may have a role within a mechanism (CC31). In any case, much uncertainty is found in connection with mechanisms and associated roles and functions, as shown by correlations with such fields as possible , probable , uncertainty , and suggest (CC32–35). Support can be directed toward hypotheses (correlation of support with hypothesis ; CC40) or toward mechanisms (correlation with mechanism ; CC41). Note again the correlation between explain and possible ; this denotes much of the uncertainty that surrounds explanations, the latter often being formulated as how-possibly explanations only (CC51–52).

As shown in the network for Bioinformatics and methods, the fields model , fit , parameter , and mathematics-statistics exhibit significant correlation in the network for Genetics, disease biology, others (fig. 4d). This shows the importance of statistical model fitting in this disciplinary context (CD10–12). The correlation of model with hypothesis maps with the confrontation of hypotheses (notably, null hypotheses) in the context of model building (CD13–14). Further correlation of mathematics-statistics with test reveals the ubiquity of statistical tests, for instance, to identify significant variables in datasets (CD15–16), sometimes associated with specific statistical models (CD17). Prediction accuracy is significant in this context, as shown earlier in the network for Bioinformatics and methods (correlation of predict and accuracy ) (CD18–19). Note that accuracy is also often discussed in association with specific methods, with some methods being known to provide more or less accuracy than others (correlation of accuracy and method-approach ) (CD20–21). The fields method-approach , data , analysis , and mathematics-statistics exhibit tight correlations, notably showing the importance of data analysis by means of different statistical approaches (CD22–23). The correlation of data with integration reveals the practical significance of multiple-source data integration, notably to generate and test novel hypotheses on the basis of varied sets of data (CD24–27). As shown in the network for Bioinformatics and methods, this specific epistemic role for integration concerns data and not models, as might have been expected. Note the presence of much uncertainty in research, with multiple correlations of the fields suggest , possible , and probable with such fields as mechanism or role-function , indicating numerous discussions about how-possibly mechanisms and how-possibly roles or functions (CD30–33). As seen in other disciplinary networks, mechanism is also correlated with both underlie and understand , typically in the sense that mechanisms that underlie phenomena of interest need to be (better) understood (CD34–35). Note also the use of causal notions such as cause and effect, mostly in connection with research on specific diseases (correlation of cause and pathology ) or modeling (correlation of effect and model ): diseases are often said to be caused by specific agents, pathogens, viruses, and so forth (CD40–41), whereas models can be used to study the effects of certain variables (CD42–43).

In Genomics and phylogenetics (fig. 4e), the fields model , parameter-variable , mathematics-statistics , and fit are correlated with one another, as seen also in Bioinformatics and methods and Genetics, disease biology, others, highlighting here, too, the significance of (statistical) model fitting (CE10). Hypothesis and test are also correlated to model , whereas refute is connected with hypothesis : both models and hypotheses can be tested (CE11–12); sometimes hypotheses are tested with models (CE13), and at other times, hypotheses are refuted (CE14–15; note here a relatively rare occurrence of refutation). A peculiarity is the lack of a significant correlation of model with predict (as encountered in Cell and molecular biology). The major epistemic role of models here is to be fitted to (phylogenetic) data, not to make predictions. Prediction is more frequently used in the context of molecular structures (correlation between prediction , protein , structure , and gene ) (CE20–21). Prediction accuracy is still a major concern and is affected by the use of specific methods (correlation between predict and accuracy , and with method-approach ) (CE22–23). As can be expected, predictions can get confirmed (CE24), and experiments may play a confirmatory role (CE25) (correlations of predict with confirm and correlations of confirm with experiment ). Note here again that much uncertainty surrounds mechanisms and functions (CE30–31) (correlations among suggest , possible , role-function , and mechanism ) and that mechanisms are epistemic objects in need of being understood (CE32–33) (correlations of mechanism with underlie and understand ). Interestingly, mechanism is here correlated with evolution , which denotes discussions about specific mechanisms of evolution (CE34–36) (not about whether evolution itself should be considered a mechanism, as in, e.g., Barros [Reference Barros2008]). As for the correlation between explain and hypothesis , it tends to capture instances of how-possibly explanations (CE37–39). Data is again correlated with analysis and method-approach , as well as with phylogenetics (disciplinary field). In this specific context, correlation with simplicity denotes the significant use of the notion of simplicity/parsimony as an epistemic value to sort explanations, interpretations, and hypotheses (CE40–41), but also the use of specific methods that make use of parsimony, such as statistical parsimony analyses and networks (CE42–43). Also note the correlation of phylogenetics with support and inconsistency , for which this disciplinary context is a rare case of occurrence. This reveals the complementary use of supporting and contradicting judgments (CE44–46). In other words, justification is not just a matter of corroboration but also of noncontradiction of accepted hypotheses (or of contradiction of rival hypotheses). Finally, causal notions appear to be connected to disease research (which is still somehow present in this cluster, likely through genomics studies), as denoted by correlations between pathology , effect , and cause .

The fields model , parameter-variable , mathematics-statistics , and fit are correlated in the network for Health and clinical research, as seen elsewhere (fig. 4f). Modeling appears to play a significant prediction-related role (CF10–11) (correlation with predict ). Here, too, analysis , data , and interpretation are strongly correlated, showing the importance of data analysis (CF20). Uncertainty is also much present, as indicated by correlations between evidence and suggest , as well as explain and possible : evidence often only has a suggestive status (as opposed to being corroborative) (CF30–32), whereas explanation is mentioned in the role of how-possibly explanation (CF33–34). Causal notions are again seen very much in connection with disease and pathology (correlation of cause with pathology and mortality-survival , correlation of effect with therapy ) (CF40–43). Mechanism plays a minor role in this disciplinary context (as opposed to the central role it had in Cancer research; Cell and molecular biology; or Genetics, disease biology, others) and is only connected to underlie . Experiments appear to be most often conducted in a clinical setting (as denoted by the correlation between experiment and clinical ) (CF50).

The network for Public health (fig. 4g) bears some similarities with the network for Health and clinical research. The fields model , parameter-variable , mathematics-statistics , and fit are all correlated, with correlation also to predict , highlighting the predictive role of statistical models (CG10). In this cluster, behavior is one particular target of prediction, as well as what triggers the search for causes, notably causes that may influence specific behavior patterns (CG20–22) (correlations between prediction , cause , and behavior ). Yet, and most interestingly in comparison to other clusters, social psychology and behavior research appear to rely on the formulation of theories, which in turn play an explanatory role (correlation of behavior , theory , and explain ). Mentioned theories include the theory of reasoned action, the theory of planned behavior, and other diverse social cognitive theories (CG23–25), which are all frequently endowed with explanatory power (CG26–28). Data , analysis , method-approach , and interpretation are all correlated as well, as seen elsewhere, denoting the importance of data analysis and the role of specific methods in this respect (CG30–31). Notions of support, integration, and role or function tend, in this cluster, to be used in a nonepistemic sense in connection with the notion of care—for instance, to discuss whether specific organizational support may favor the integration of different types of health-care services (CG40–43) (correlation of care with support , integration , and role-function ). Finally, note the correlation between process and understand , which sometimes denotes research on the process of understanding (CG50) but more frequently on biological processes as objects that need to be understood (CG51), quite similar to the way mechanisms are also taken as targets for understanding in other clusters, such as Cell and molecular biology; Cancer research; and Genetics, disease biology, others.