Getting Started with Social Network Analysis Through John Scott's Framework
Most people trying to learn network analysis hit a wall somewhere between understanding nodes and edges and actually producing something useful. John Scott's approach in his 1991 handbook and subsequent works cuts through a lot of the noise by grounding the methodology in clear structural concepts rather than abstract mathematical formalism. That distinction matters more than it might seem at first. Scott's core contribution centers on how relationships between actors produce structural positions that constrain or enable action. He didn't invent network analysis, but he systematized a way of thinking about it that works when you actually need to analyze data instead of just visualize it. The key terms to get straight early are structural equivalence, density, centrality measures, and blockmodeling. Get those down before you touch any software.
Network Analysis John Scott: The Practical Path
Here is how I would approach this if you are starting from zero and need to produce actual results, not just pass an assignment. First, define your boundary condition. This is the step everyone skips and then regrets. You need to decide what constitutes your network. Is it a single organization? A community? A set of interlocking boards? Scott emphasizes that the boundary determines everything that follows, and picking the wrong one will make your analysis meaningless regardless of how sophisticated your methods are. I spent two weeks on a project where I hadn't properly bounded the network. The data came back showing apparent cliques that turned out to be artifacts of my boundary being too tight. Once I expanded it by two degrees of separation, the real structure appeared. That cost me about ten business days I'll never get back. Next, collect your data. You have three main options: survey-based ego networks, complete relational data from records or archives, and observed interaction data. Scott tends to favor complete relational datasets because they let you see the actual structure rather than individual perceptions of it. But complete data is expensive to collect. For a small organization of maybe two hundred people, you can feasibly ask everyone to report their ties. For anything larger, you usually need to compromise.
The third step is encoding your data into a matrix format. Rows and columns represent the same set of actors. An entry of 1 means a tie exists, 0 means it does not, or you can use weighted values if the strength of the relationship matters. Scott pays particular attention to valued networks because most real relationships are not simply present or absent. They vary in intensity, frequency, or quality. Converting that into a numeric matrix is straightforward in principle. It gets messy fast when you have missing data or ambiguous responses. Once your matrix is ready, run the basic descriptive measures. Calculate density to understand how connected the overall network is. Compute degree centrality to identify who has the most ties. Look at betweenness centrality to find people who bridge otherwise disconnected groups. These three measures alone will tell you roughly forty percent of what matters about any social network. Scott's handbook walks through the calculations, but honestly you should just use software for the heavy lifting. UCINET, Pajek, or even Python's NetworkX library will handle the matrix algebra without making you derive it by hand. I stopped doing hand calculations around 2008. There is no reason to. Blockmodeling is where Scott's framework becomes particularly useful. The idea is to collapse actors who share similar tie patterns into structural positions, then examine the relationships between those positions rather than between individuals. This makes large networks interpretable. A two hundred person network is impossible to read directly. A blockmodel reduction to eight or ten positions is manageable. The trick is choosing the right equivalence criterion. Structural equivalence means two actors have identical ties to every other actor in the network. Relational equivalence means they have similar patterns of ties even if not identical. Scott leans toward structural equivalence for cleaner results, but real data rarely cooperates that nicely. You will often need to relax the criterion and accept some noise.
Get the Full Details
A common pitfall I see repeatedly is treating centrality scores as if they measure influence. They do not. A person with high betweenness centrality is structurally positioned to control information flow, which may or may not translate into actual influence. I analyzed a nonprofit network where the highest-betweenness individual was technically a gatekeeper but had almost no substantive power. Their position was more accidental than strategic. People who assumed she was influential based purely on the metric made flawed decisions about where to direct resources. Always triangulate centrality measures with qualitative information about actual power dynamics. Another thing beginners miss is the difference between one-mode and two-mode networks. One-mode networks have a single type of actor. Two-mode networks link two different types, like people and organizations or authors and journals. Scott discusses this distinction but many tutorials gloss over it. If you have two-mode data, you need to convert it appropriately before running standard analyses. The conversion process itself can introduce artifacts, so check your assumptions carefully. When you interpret your results, resist the temptation to overstate causal claims. Network analysis shows structural patterns, not mechanisms. A correlation between network position and outcomes like job performance or innovation rate suggests a relationship but does not prove directionality. Does being central cause better performance, or do high-performing people naturally accumulate more ties? The data alone cannot answer that. Longitudinal network data helps, but it is rare and expensive to obtain. Be honest about what your analysis can and cannot tell you.
There is also a practical issue with software that few people mention upfront. Most network analysis tools assume your data is complete and your network is static. Real networks are neither. People leave organizations. Ties weaken and strengthen. If you are working with historical data or survey snapshots, acknowledge that limitation explicitly in any report you produce. I learned this the hard way when reviewing someone else's work that presented a cross-sectional analysis as if it captured an entire organizational dynamic. The conclusions were defensible only under very narrow conditions that the authors had not stated. If you want to go further after getting comfortable with Scott's basic framework, look into stochastic actor-oriented models for longitudinal data, exponential random graph models for testing hypotheses about tie formation, and latent space models for visualizing multidimensional structural positions. These are more advanced but build directly on the foundation Scott helped establish. The honest assessment is that network analysis has real limitations. It requires data that is often difficult to obtain ethically and practically. Small networks produce measures that are highly sensitive to individual cases. Large networks become computationally intensive and hard to interpret. The field has produced enough published work showing inflated claims about what network metrics can predict that a certain healthy skepticism is warranted. Scott himself was never guilty of overclaiming, which is part of why his handbook remains relevant decades after publication.
For practical purposes, start with Scott's 1991 book as your reference text alongside whichever software you choose. Work through a small complete dataset by hand first, even if it is just a class of thirty students mapping friendships. The matrix algebra will make sense in a way that abstract explanation never achieves. Then move to software and larger projects. The gap between knowing the concepts and actually applying them is where most people stall, and closing that gap requires doing the work, not just reading about it.
