3.1.1 Availability

In some cases, certain network members may be unavailable to serve as KOLs, while certain other network members must serve as KOLs. Although these individuals may have restrictions on their potential KOL role, because their relationships are still potential channels for diffusing information or influencing behavior, it is important to retain them in the network when considering others’ suitability as KOLs. Therefore, interventionists may only want to consider KOL teams that exclude unavailable individuals and include required individuals, but may still want to evaluate these teams based on their members’ position in the entire network. The KOLaide package implements this using the include and exclude parameters.

3.1.2 Breadth

When the goal of a KOL team is to diffuse information, the KOL team’s members should be able to reach a large number of other network members in a relatively small number of steps so that they can share information both widely and quickly. Following Borgatti (Reference Borgatti2006), we propose using $M$ -reach to evaluate a team’s breadth of network coverage for facilitating diffusion. We adopt the definition of $M$ -reach already implemented in existing keyplayer software (An and Liu Reference An and Liu2016; Borgatti Reference Borgatti2006) – the number of unique nodes that can be reached by members of the KOL team in $M$ steps. However, because KOL team members can trivially reach themselves, we normalize this count by the number of non-KOL team members. Thus, we define $M$ -reach as the fraction of non-KOL nodes that are reachable in $M$ steps. When $M = 1$ , $M$ -reach is equivalent to group degree centrality (Everett and Borgatti Reference Everett and Borgatti1999). Larger values of $M$ yield a more liberal metric because they allow KOLs to be further away from the individuals to whom they seek to disseminate information.

When the goal of a KOL team is to encourage adoption of a behavior, multiple KOL team members should be directly reachable by other network members so that others can quickly receive support and reinforced encouragement. Following Wan et al. (Reference Wan, Riedl and Lazer2025), we propose a new metric, $M$ -reinforcement, to evaluate a team’s breadth of network coverage for facilitating adoption. We define $M$ -reinforcement as the fraction of non-KOLs who are directly connected to at least $M$ KOLs. When $M = 1$ , $M$ -reinforcement is equivalent to group degree centrality (Everett and Borgatti Reference Everett and Borgatti1999). Larger values of $M$ yield a more conservative metric because they require that non-KOLs have direct connections to more KOLs from whom they might receive reinforcing support and encouragement.

Although there are many other potential ways to measure breadth, including a vast array of group centrality measures (Everett and Borgatti Reference Everett and Borgatti1999), we follow Borgatti (Reference Borgatti2006) in using $M$ -reach to capture diffusion potential, and follow Wan et al. (Reference Wan, Riedl and Lazer2025) in using $M$ -reinforcement to capture adoption potential. In addition to choosing these metrics based on prior theoretical and empirical literature, they offer several other advantages. First, they have simple, straightforward interpretations that intervention researchers can readily explain to colleagues and community partners who are likely to have limited network expertise. In particular, they have well-defined ranges between 1 for a KOL team whose breadth covers the entire network, to 0 for a KOL team whose breadth does not cover any other network members. Second, missingness is common in field intervention data, but as local degree-based metrics, node missingness has limited impact on these metrics compared to other more global metrics such as betweenness. Finally, unlike more complex centrality metrics such as eigenvector, $M$ -reach can be computed very quickly in linear time, which is essential because it must be computed for every possible KOL team. The KOLaide package implements this using the goal and M parameters.

3.1.3 Cost

Larger KOL teams can often provide greater breadth, but also impose greater costs on both the setting and the interventionist. Therefore, interventionists may want to consider a range of potential KOL team sizes. The KOLaide package implements this using the range parameter.

3.1.4 Diversity

The suitability of a KOL team may depend primarily on the breadth of its members’ network coverage, which allows them to diffuse information to many others in the network, or to encourage many other network members to adopt a behavior. However, the KOL team’s ability to achieve this goal may depend not solely on their network breadth, but also on whether the team members are representative of the wider population. Such representativeness allows target individuals to “see themselves in” the KOLs and thereby facilitates’ buy-in by members of the population. Therefore, given a categorical attribute of population members, interventionists may consider a KOL team’s representativeness by examining the fraction of categories of the characteristic of interest that are represented on the KOL team. For example, if the characteristic of interest is smoking status, then a KOL team would achieve the maximum value of 1 if it includes at least one smoker and one non-smoker. The KOLaide package implements this using the attribute parameter.

3.1.5 Evaluation

Once potential KOL teams’ breadth, cost, and diversity have been determined, interventionists must decide how to integrate this information into an overall evaluation of each KOL team. There are an infinite number of ways that this information might be weighted and combined into an overall evaluation, but any overall evaluation should follow a couple simple rules. Other things being equal:

• • KOL teams with more breadth of network coverage are better because this is the goal of relying on KOLs for facilitating diffusion and adoption in a network.

• • When diversity on a given attribute is relevant in a given setting, KOL teams with more diversity are better.

• • A KOL team’s breadth of network coverage matters more than its diversity, otherwise a network-based approach to KOL selection would not be necessary.

• • KOL teams with lower cost (i.e., smaller KOL teams) are better.

The KOLaide package returns each KOL team’s score on breadth, diversity, and cost, which interventionists can examine separately, or combine in ways that are most appropriate given the community context. However, to facilitate the evaluation of KOL teams, the package also returns one convenient overall evaluation metric that obeys these rules:

\begin{equation*}\frac {B^\alpha D^{1-\alpha }}{C^\beta }\end{equation*}

where B is breadth, C is cost, D is diversity, and $\alpha$ and $\beta$ are weighting parameters. Alternatively, in contexts where attribute data is not available or diversity is not relevant,

\begin{equation*}\frac {B}{C^\beta }\end{equation*}

may be appropriate. The KOLaide package implements these evaluation metrics using the alpha and beta parameters.

The $\alpha$ parameter controls the relative weight of breadth versus diversity, and can range $0.5 \leq \alpha \leq 1$ . Intuitively, $\alpha$ can be viewed as the proportion of weight that is assigned to breadth, and $1-\alpha$ can be viewed as the proportion of weight that is assigned to diversity. We recommend a default of $\alpha = 0.9$ , which prioritizes a KOL team’s ability to cover the network, but still considers the diversity of its members. This mirrors the approach used by Cappella et al. (Reference Cappella, DeShazer, Park, Neal, Exner-Cortens and Owens2025), who used the diversity of potential KOL teams as a “tie breaker” for teams that offered similar breadth of network coverage.

The $\beta$ parameter controls how larger KOL teams are penalized to reflect their great cost, and can range $0 \leq \beta \leq 2$ . The value of $\beta$ can capture scenarios when additional KOLs have a diminishing marginal cost ( $\beta \lt 1$ ), a linear cost ( $\beta = 1$ ), or an increasing marginal cost ( $\beta \gt 1$ ). Given the economies of scale that are likely involved in recruiting and training additional KOL team members, we recommend a default of $\beta = 0.9$ , which implies a small diminishing marginal cost.

3.2.1 A small example

Table 2 shows some example code for using the KOLaide package to identify and plot KOL teams. The code to reproduce these examples is also available at https://osf.io/e2m8g.

Table 2. Example code to identify and plot KOLs using the KOLaide package

The KOLaide package can be installed from CRAN using install.packages(“KOLaide”), and can be loaded in R using library(KOLaide). Selecting KOLs from a network is performed using the pick_kols() function, which takes an adjacency matrix or igraph network as its input. In this example, we use a randomly generated small-world network of 26 nodes, where each node has been assigned a categorical attribute called “role” with three levels. This network might represent advice seeking among teachers in a school, where each teacher is located in one of three buildings, or teaches at one of three grade levels.

The function allows several optional arguments that control how it addresses the practical challenges that arise in KOL selection. To address the challenge of availability, exclude = “C” indicates that person C is not available to serve as a KOL team and therefore should be excluded from all possible KOL teams, while include = “H” indicates that person H is not only available to serve on KOL teams, but must be included on all possible KOL teams. To address the challenge of breadth, m = 2 and goal = “diffusion” indicate that the goal of the KOL team is to diffuse information, which causes the function to assess each team’s breadth using $M$ -step where $M = 2$ . To address the challenge of cost, range = c(2,4) requests that the function identify KOL teams of 2–4 members, which will vary in the costs they impose on both the setting and interventionist. To address the challenge of diversity, attribute = “role” requests that the function consider each team’s diversity with respect to members’ role. Finally, to address the challenge of evaluation, alpha = 0.9 and beta = 0.9 uses the default values for the $\alpha$ and $\beta$ weights in the summary evaluation score; these values would also be used if these lines were omitted. For convenience, file = “diffuse” requests that a list of all possible KOL teams, with their characteristics, should be saved to the file diffuse.csv for use after exiting R.

In less than 1 second (using an Apple M1 Max processor from 2021), the function returns a list of all 2,324 valid KOL teams in a KOL object called teams and in a CSV file called diffuse.csv. Table 3A shows the top five teams identified by the function. The "best” team consists of H and T, who can reach 88% of the rest of the network in two or fewer steps, at the cost associated with recruiting 2 KOLs, who represent 67% of the roles in the setting. Combining these characteristics using the default weights, this team is assigned an overall evaluation score of 0.46. The other teams vary in their breadth and diversity, but have similarly high overall evaluations. Notably, as requested, these teams always include person H, and never include person C. Based on this output, an interventionist may choose to recruit H and T as a KOL team, but in the face of other considerations, might view teams such as H-S and H-R as also promising.

Table 3. Selected KOL teams for (A) diffusion and (B) adoption identified by pick_kols(); B = breadth, C = cost, D = diversity, E = evaluation

Figure 1. Example of KOL teams selected using pick_kols() and plotted using plot_kols() in a 26-node network with the goal of (A) diffusing information or (B) encouraging adoption of a new behavior. Node colors represent a node attribute of interest (e.g., role). Red-bordered nodes are KOLs, while green-bordered nodes are network members whom the KOLs can influence.

To facilitate evaluation of the identified KOL teams, the plot_kols() function plots the network and the KOL team’s coverage. The function takes a KOL object generated by pick_kols() as its input, and allows any additional arguments that are available in igraph’s plot() function. Figure 1A shows the plot generated from the above example. By default, the nodes are colored according to their attribute, with KOLs depicted using a red border and network members covered by KOLs depicted using a green border. This plot highlights that the team of H and T is diverse (H is orange, T is aqua) and cover the majority of the network (most nodes have green borders).

Using all the same settings, but changing to goal = “adoption” generates a list of valid KOL teams where breadth is measured using $M$ -reinforcement, which also requires less than 1 s. Table 3B shows the top five teams identified by the function, while Figure 1B plots the best team. Several differences are apparent when identifying KOL teams with the goal of adoption rather than diffusion. First, the best teams for adoption are larger, despite the fact that larger teams are more costly. Second, even the best KOL teams for adoption have a smaller breadth, covering fewer of the other network members. These differences highlight that encouraging adoption is often more challenging than merely facilitating diffusion, but they also highlight the importance that users (know how to) select the measure of breadth that corresponds to their goal.

3.2.2 A large example

Identification of a KOL team using a keyplayer strategy like that implemented in keyplayer software (An and Liu Reference An and Liu2016; Borgatti Reference Borgatti2006) and KOLaide requires knowledge of the whole network. Because whole networks are challenging to collect in community settings, this requirement imposes some limits on the size of settings where keyplayer strategies are typically used. For example, a network of 26 teachers may be feasible to collect, such as in the small example, and thus offers a common application for KOL team selection using a keyplayer strategy. In contrast, a network of all students in a large high school may be less feasible to collect, and thus may be a less common application for keyplayer strategies.

Although it may be less common to apply keyplayer strategies to selecting KOL teams in large networks, it is still possible using the KOLaide package. In the small example, the pick_kols() function returns a complete list of all valid KOL teams, and therefore guarantees finding the optimal (as well as each sub-optimal) team. However, because the number of possible teams is defined by $\binom {N}{k}$ , where $N$ is the network size and $k$ is the KOL team size, this is not practical in a large network where the number of valid teams may be very large. For example, there are over one billion teams of 4 in a network of 400.

To address the challenge of selecting KOLs in large networks, the top = X argument restricts the scope of individuals considered eligible to the $X$ individuals with the highest degree, or highest closeness, or highest betweenness. Because an individual is deemed eligible if they have the highest value on any one of these three metrics, the pool of eligible individuals can range from $X$ if the three centralities are perfectly correlated, to $3X$ if the three centralities are perfectly uncorrelated. This approach assumes that high-centrality individuals are more likely to be useful KOL team members than low-centrality individuals. However, because this assumption is not necessarily correct (Borgatti Reference Borgatti2006), this approach is not guaranteed to find optimal teams, and only offers a heuristic.

To illustrate, we use a randomly generated small-world network of 400 nodes. We can still use pick_kols() to identify 2–4 member KOL teams for diffusion, but include top = 30 to restrict the scope of individuals considered as possible KOLs. This takes about 50 s, and returns a list of nearly 200,000 possible teams. The best team identified using this approach is a four-member KOL team that can reach 27% of other network members in up to 2 steps, and has an overall evaluation score of 0.078. Figure 2 is a plot generated by plot_kols() that illustrates this KOL team, with the KOL team’s members shown in red and reachable non-KOLs shown in green. The visualization suggests that this team’s breadth of coverage includes parts of the core, but misses most of the periphery, and may suggest considering still larger KOL teams or a different dissemination strategy.

Figure 2. Example of KOL teams selected using pick_kols() and plotted using plot_kols() in a 400-node network with the goal of diffusing information. Red nodes are KOLs, while green nodes are network members whom the KOLs can influence.