Network and relationship analysis

Network analysis is needed when the question is relational, that is, who connects with whom, which groups emerge or where something flows, rather than a question about independent individual attributes. It applies to phenomena such as scientific collaboration, comorbidity, patient referral, co-prescription or the spread of information, where the connection between elements is the object of study, not a detail. An analysis that treats each observation as independent ignores, by design, the structure of relationships, which is exactly the information that gives these data their value. Network analysis models those relationships explicitly: it represents the system as nodes and connections and quantifies its structure with formal metrics. It is the difference between describing the elements of a system and understanding how they are organised.

At a minimum, you need a clear definition of what a node is and what a connection is, and a list of the observed relationships between nodes (the edge list), with their direction and weight where they have them. The client provides the relational data; the service defines the unit of analysis precisely, cleans the edges, decides how to handle direction and weight, and checks structural consistency before calculating a single metric. The data do not need to arrive in network format: they are often derived from event tables, co-authorships, referral records or co-occurrences, and that transformation is part of the work. The boundary of the network also has to be defined: which nodes fall inside or outside the system, and whether the observed relationships are complete or a partial sample, because centrality, communities and density depend directly on that boundary. What does shape the whole analysis is the definition of the relationship: what exactly a connection means in the context of the study determines what can be concluded afterwards.

Centrality captures the structural relevance of a node according to its position in the network, not its volume of connections. It applies when you want to identify the elements that hold a relational system together, direct attention or resources to where the structure makes them decisive, or distinguish structural relevance from apparent presence. Different metrics are calculated depending on the question: degree centrality (number of connections), betweenness (how many paths pass through the node), closeness (distance to the rest) and eigenvector centrality (connection to nodes that are themselves well connected). A node with few connections can act as a crossing point between important parts of the network (a bottleneck or a bridge), something a simple count never captures. Each metric answers a different notion of importance: choosing the relevant one is part of the analysis, not an automatic step.

Communities are estimated with algorithms that propose groupings of nodes more densely connected to each other than to the rest of the network. This applies when you suspect that a relational system is organised into subgroups, such as collaboration teams, comorbidity clusters or referral circuits, that are not visible when the data are treated as separate observations. Modularity-based methods such as Louvain or Leiden can be used, along with analysis of cohesion, components and dense subgroups, depending on the size and density of the network. The result is not a single, definitive partition: community structure depends on the algorithm, its resolution and the nature of the network, so it is interpreted by assessing its stability, not taken as automatic truth. What you receive is a characterisation of the groupings detected and of the nodes that link them.

The analysis can go beyond description: statistical network models help explain connection patterns that a simple description of the network does not capture. This applies when describing the network, with its metrics and communities, does not fully answer the question, and you want to examine whether certain connection patterns go beyond what a simple description shows. For this, the analysis can use exponential random graph models (ERGM), which assess whether patterns such as reciprocity, transitivity or homophily are compatible with the observed network; stochastic block models (SBM), which infer latent structure; or temporal network models when the structure evolves over time. One limit applies: these models identify which mechanisms are compatible with what is observed; they do not prove causality for why each tie formed. They let you move from "this is what the network looks like" to "these mechanisms are consistent with how it came to be organised".

Network analysis is the methodological family that models any system of relationships, such as authorship, comorbidity, referral, co-prescription or diffusion, with formal metrics and models. Infodemiology fits when the question is specifically about how health information circulates in digital environments: what people search for, what spreads and how. They can be combined, since the study of digital diffusion can draw on network metrics, but they do not answer the same question: network analysis provides the general methodological toolkit for relational data of any kind; infodemiology applies that and other approaches to a specific domain, health information on social media and search engines. Here, social media networks are just one more example of a relational system, not the focus of the service.

The interpreted network indicators: centrality metrics, community structure and the global properties of density, distance and cohesion. They come with readable graph visualisations and, where appropriate, the results of the statistical modelling. This applies if you need quantitative evidence behind claims about connection, influence or cohesion that would otherwise remain intuition. The visualisation uses layout algorithms suited to the size, density and communication goal of the network, encoding metrics in node size, colour or grouping in the service of insight, not decoration; an unreadable graph does not communicate, it decorates. In large or very dense networks, communicating the result may require aggregation by community, filters or interactive views, because showing every node at once can hide more than it reveals. The tables and indicators are delivered reproducibly, ready to integrate into a report or an interactive application. What you receive is the structure of the network translated into interpretable conclusions, not an eye-catching graph with nothing to read.

Network analysis can be applied to many systems in which it is possible to define precisely what a node is, what a connection is and what that relationship means. It applies to authorship and citation networks, comorbidity, care referral, co-prescription, the spread of information and other relational systems in health and the life sciences. The same methodological logic, whether position metrics, community detection or statistical modelling, works for very different systems precisely because it operates on the structure of relationships, not on the specific content of each domain. The requirement is not the type of phenomenon but the quality of the definition: a poorly defined connection produces a network that is analysed rigorously but means nothing. That is why the first step is always to agree on what a node represents, what an edge represents and which real-world relationship it encodes, before calculating anything.