Event Data

tl;dr

Event data come in two flavors. One-mode event data (actor-on-actor: who fought who, who cited whom) becomes a unipartite network. Two-mode event data (actor-on-venue: firms in jurisdictions, attendees at meetings) becomes a bipartite network. This vignette walks through both. Along the way we cover transforming weights, single-actor highlighting, applying a ggplot theme, and how to export a figure with ggsave().

library(netify)
library(ggplot2)

one-mode event data: ucdp conflict in mexico

A common data type used to build conflict networks is event data, where actors are repeated across rows but no single column explicitly encodes an edge weight. netify() handles this case by counting interactions when no weight is provided.

We use UCDP GED data on Mexico as one example of an application to intrastate event data. The data come from https://ucdp.uu.se/downloads/ (version 23.1, subset to Mexico); the subset ships internally so the code below runs as-is.

Create: generate an aggregated, weighted network of conflict between actors in Mexico. When the user does not supply a weight value, netify() counts interactions and returns the count as the edge weight, with an informational message indicating that this has happened.

# load ucdp ged data on mexico
data(mexico)

# construct count-weighted network (number_of_events by default)
mex_network <- netify(
    mexico,
    actor1 = "side_a",
    actor2 = "side_b",
    symmetric = TRUE,
    sum_dyads = TRUE,
    diag_to_NA = TRUE,
    missing_to_zero = TRUE
)

## ! Warning: there are repeating dyads within time periods in the dataset. When `sum_dyads = TRUE` and `weight` is not supplied, edges in the outputted adjacency matrix represent a count of interactions between actors.

# summaries at the graph and actor levels
summary(mex_network)

##   net num_actors    density num_edges prop_edges_missing mean_edge_weight
## 1   1         74 0.02554609        69                  0         274.7391
##   sd_edge_weight median_edge_weight min_edge_weight max_edge_weight competition
## 1       660.3139                 58               1            3757   0.1371326
##   sd_of_actor_means transitivity
## 1          21.37264    0.1046512

actor_stats <- summary_actor(mex_network)
plot_actor_stats(actor_stats)

## Warning: Removed 55 rows containing non-finite outside the scale range
## (`stat_density()`).

## Warning: Removed 55 rows containing missing values or values outside the scale range
## (`geom_rug()`).

Event counts have a sharp interpretation gap that does not show up in, say, a 0/1 alliance network: the difference between “these two actors were both in the data and recorded zero events together” and “we have no information on whether these two actors interacted at all” is the difference between a substantive zero and a missing observation. The default missing_to_zero = TRUE collapses both into 0. For routine event-data uses this is what you want — a missing UCDP dyad really does mean “no recorded violent event between these two sides in this slice of the data,” which is closer to a substantive zero than a Bayesian “we don’t know.” But for epidemiology / contact-tracing / sparse rosters, the asymmetry matters and you’ll want missing_to_zero = FALSE.

A 5-actor toy makes the bookkeeping concrete. Suppose four contact-tracing interviews surfaced four edges among five named individuals; the remaining dyads were never asked about:

tiny <- data.frame(
    a1 = c("p1", "p1", "p2", "p3"),
    a2 = c("p2", "p3", "p3", "p4"),
    n_contacts = c(2, 1, 3, 1),
    stringsAsFactors = FALSE
)

# preserve unobserved dyads as na
net_na <- netify(
    tiny,
    actor1 = "a1", actor2 = "a2",
    weight = "n_contacts",
    symmetric = TRUE,
    missing_to_zero = FALSE,
    nodelist = c("p1", "p2", "p3", "p4", "p5")
)

# the raw matrix carries explicit nas for the never-asked dyads
get_raw(net_na)

##    p1 p2 p3 p4 p5
## p1 NA  2  1 NA NA
## p2  2 NA  3 NA NA
## p3  1  3 NA  1 NA
## p4 NA NA  1 NA NA
## p5 NA NA NA NA NA

# summary() reports the missingness fraction in its own column when
# missing_to_zero = false
summary(net_na)[, c("num_actors", "density", "num_edges",
                    "prop_edges_missing", "prop_unknown_edges")]

##   num_actors density num_edges prop_edges_missing prop_unknown_edges
## 1          5     0.4         4                0.6                0.6

prop_unknown_edges only shows up when missing_to_zero = FALSE — that is the cue that the netlet is actually carrying NA semantics downstream. Build the same netlet with the default and the NAs are replaced with 0s and the column disappears; that is fine for UCDP / GED event counts (the rest of this vignette uses the default for exactly that reason) but it is the wrong choice when “we did not measure this dyad” is a meaningful state.

Explore: next we use the basic plot() method to visualize the network of violent interactions in Mexico. The simplest call labels every actor:

set.seed(6886)
plot(mex_network, add_text = TRUE)

For a less cluttered view, pass a vector of actors to select_text / select_text_display:

# select 10 random names for plotting
select_names <- rownames(mex_network)
set.seed(6886)
random_indices <- sample(length(select_names), 10)
random_names <- select_names[random_indices]

plot(mex_network,
    select_text = random_names,
    select_text_display = random_names
)

The graphs above reveal the need to transform the edge weights to increase interpretability. Logging the values is a common move. The netify plot function has a built-in parameter for this — pass any function to mutate_weight (see also the general mutate_weights() function for transforming weights outside of plotting). The edge_alpha_label parameter just relabels the resulting legend:

set.seed(6886)
plot(mex_network,
    select_text = random_names,
    select_text_display = random_names,
    # log(x+1) to better see the range of connections
    mutate_weight = log1p,
    edge_alpha_label = "Log(Event Count)"
)

highlighting a single actor

Often there is one actor you really want to call out — a focal group, a state agency, a single firm. plot.netify() supports this directly via highlight=, which takes either a character vector of focal actors or a named color vector mapping actor names to colors:

focal <- "Government of Mexico"
set.seed(6886)
plot(mex_network,
    mutate_weight = log1p,
    edge_alpha_label = "Log(Event Count)",
    highlight = focal,
    highlight_color = c("Government of Mexico" = "#0A3161", "Other" = "grey70"),
    select_text = focal,
    select_text_display = focal
)

The focal actor is now visually distinct from the rest of the network — useful for storytelling and for slide decks.

applying a theme

Drop on theme_publication_netify() — a ggplot theme with a larger base font, italic legend titles, and stripped axes (the right default for force-directed layouts):

# add degree centrality from summary_actor()
actor_stats <- summary_actor(mex_network)
actor_stats$degree_group <- ifelse(
    actor_stats$degree > median(actor_stats$degree, na.rm = TRUE),
    "higher degree",
    "lower degree"
)
mex_network <- add_node_vars(
    mex_network,
    actor_stats,
    actor = "actor"
)

set.seed(6886)
plot(mex_network,
    mutate_weight = log1p,
    edge_alpha_label = "Log(Event Count)",
    node_size_by = "degree",
    node_size_label = "Degree",
    node_alpha = .5,
    check_overlap = TRUE
) +
    theme_publication_netify()

theme_publication_netify() is built for network layouts (no axes). For actor-level summaries or heatmaps — where you do want axes — use theme_publication_netify_ts() (the _ts is for “tables / time-series”, anything that isn’t a node-edge layout).

saving the figure

Once you have a plot you like, export it with ggsave(). A vector format (PDF or SVG) holds up at any size; for the web or social, a PNG at 300 dpi is fine.

# the last plot rendered is captured by ggsave() by default
ggsave("mexico_conflict.pdf", width = 7, height = 6)

# pass a named plot object
p <- plot(mex_network, mutate_weight = log1p) + theme_publication_netify()
ggsave("mexico_conflict.png", plot = p, width = 7, height = 6, dpi = 300)

a note on color choice

Colors matter for accessibility — roughly 8% of men and 0.5% of women have some form of color vision deficiency. When you encode a variable in color, prefer a colorblind-safe palette. Two easy options:

ColorBrewer qualitative palettes like Set2, Dark2, and Paired are colorblind-safe by design. Use scale_color_brewer(palette = "Set2") or just pass node_color_palette = "Set2" to plot.netify().
viridis (viridis, magma, plasma, cividis) gives a perceptually uniform, colorblind-safe gradient for continuous variables.

# example: colorblind-safe categorical fill via plot.netify()
plot(mex_network,
    node_color_by = "degree_group",
    node_color_palette = "Set2"     # colorbrewer colorblind-safe
) +
    theme_publication_netify()

# example: colorblind-safe continuous color, manual
plot(mex_network, node_color_by = "degree") +
    scale_color_viridis_c(option = "cividis") +
    theme_publication_netify()

advanced: assemble the plot from components

For most uses the calls above are enough. If you want full control over how the layers are stacked — say, to insert a custom annotation between the edge layer and the node layer — set return_components = TRUE and reassemble the pieces yourself.

# investigate each component of the plot
set.seed(6886)
comp <- plot(
    mex_network,
    remove_isolates = TRUE,
    select_text = random_names,
    select_text_display = random_names,
    mutate_weight = log1p,
    return_components = TRUE
)

comp

##

## ── Netify plot components ──

##

## • Base plot: ggplot object

## • Edges: geom_segment/geom_curve layer

## • Points: geom_point layer

## • Text Repel: geom_text_repel layer

## • Theme: theme_netify

## ℹ Use `assemble_netify_plot()` to build or construct manually with
##   `netify_edge()`, `netify_node()`, etc.

# modify component
comp$base +
    netify_edge(comp) +
    labs(alpha = "Log(Event Count)") +
    reset_scales() +
    netify_node(comp) +
    netify_text_repel(comp) +
    comp$theme

a note on missingness

When you build a network from event data, netify() by default fills unobserved dyads with 0 (missing_to_zero = TRUE). For some applications — contact tracing, animal interactions, partially-observed survey rosters — the difference between observed-no-interaction (0) and not-observed-at-all (NA) matters. Set missing_to_zero = FALSE to keep NAs as NAs. Similarly, actor_time_uniform = FALSE lets actor composition vary across periods (useful for open-cohort longitudinal data); the netlet then stores an actor_pds data.frame recording each actor’s entry / exit window. See ?netify for details.

two-mode (bipartite) event data: offshore jurisdictions

Event data are not limited to one-mode actor-on-actor interactions. People attending venues, students enrolling in classes, firms registered in jurisdictions — anywhere two distinct sets of entities are linked by observed events — calls for a bipartite network.

A concrete data-journalism example: leaks like the Panama Papers or Paradise Papers connect entities (companies, trusts) to jurisdictions (tax havens). Each entity-jurisdiction pairing is an event (“firm X was incorporated in haven Y”). The result is a two-mode network with firms on one side and jurisdictions on the other.

# toy offshore-leak data: firms incorporated in jurisdictions
set.seed(6886)
firms <- paste0("firm_", LETTERS[1:8])
havens <- c("BVI", "Cayman", "Bermuda", "Jersey", "Panama", "Mauritius")

offshore <- data.frame(
    firm = sample(firms, 40, replace = TRUE),
    haven = sample(havens, 40, replace = TRUE,
        prob = c(0.30, 0.25, 0.15, 0.12, 0.10, 0.08))
)
head(offshore)

##     firm     haven
## 1 firm_C    Cayman
## 2 firm_F   Bermuda
## 3 firm_D   Bermuda
## 4 firm_B Mauritius
## 5 firm_F    Cayman
## 6 firm_G       BVI

To build the bipartite netlet, set mode = "bipartite" and point actor1 / actor2 at the two entity columns. sum_dyads = TRUE aggregates repeated firm-haven incorporations into an edge weight.

bp <- netify(
    offshore,
    actor1 = "firm", actor2 = "haven",
    mode = "bipartite",
    sum_dyads = TRUE
)

## ℹ `missing_to_zero` is set to "TRUE" (the default).
## ! Missing dyads will be filled with zeros. For latent space or other
##   statistical network models, structural zeros and missing data have different
##   meanings. Set `missing_to_zero = FALSE` to preserve NAs if this distinction
##   matters for your analysis.
## ! Warning: there are repeating dyads within time periods in the dataset. When `sum_dyads = TRUE` and `weight` is not supplied, edges in the outputted adjacency matrix represent a count of interactions between actors.
## 
## This message is displayed once per session.

bp

## ✔ Network data created.
## • Bipartite
## • Sum of Binary Weights
## • Cross-Sectional
## • # Unique Row Actors: 8
## • # Unique Column Actors: 6
## • # Unique Actors: 8
## Network Summary Statistics:
##             dens miss  mean
## weight_var 0.542    0 1.538
## • Nodal Features: None
## • Dyad Features: None

Because the row actors (firms) and column actors (havens) are different sets, the resulting adjacency is rectangular and asymmetric. A heatmap is a natural display here – one cell per firm-haven pair, color encoding the number of incorporations.

Use plot(bp, style = "heatmap") for a quick plot:

plot(bp, style = "heatmap")

The code below extracts the same adjacency with get_raw() and draws the tile plot directly with ggplot2::geom_tile():

# turn a netify bipartite adjacency into a tidy long data.frame
heatmap_long <- function(net) {
    m <- get_raw(net)
    data.frame(
        row = factor(rep(rownames(m), times = ncol(m)),
            levels = rownames(m)),
        col = factor(rep(colnames(m), each = nrow(m)),
            levels = colnames(m)),
        value = as.numeric(m)
    )
}

bp_long <- heatmap_long(bp)

ggplot(bp_long, aes(x = col, y = row, fill = value)) +
    geom_tile(color = "white", linewidth = 0.3) +
    scale_fill_viridis_c(option = "cividis", name = "Incorporations") +
    labs(x = "Jurisdiction", y = "Firm",
         title = "Offshore footprint by jurisdiction") +
    theme_publication_netify_ts(base_size = 11) +
    theme(axis.text.x = element_text(angle = 45, hjust = 1))

Note the use of theme_publication_netify_ts() here instead of theme_publication_netify() — heatmaps need their axes, so the _ts variant (which keeps axes on) is the right choice. The cividis viridis palette is colorblind-safe and prints well in grayscale.

Bipartite networks use a different family of summaries than unipartite ones: there is no transitivity or mutuality (a triangle would have to span the bipartition), and degree / centralization are reported separately for row and column actors. Run summary(bp) for the full set.

summary(bp)

##   net num_row_actors num_col_actors   density num_edges prop_edges_missing
## 1   1              8              6 0.5416667        26                  0
##   mean_edge_weight sd_edge_weight median_edge_weight min_edge_weight
## 1         1.538462      0.9046886                  1               1
##   max_edge_weight competition_row competition_col sd_of_row_means
## 1               4         0.13875         0.21125       0.2954684
##   sd_of_col_means covar_of_row_col_means
## 1       0.4721405                     NA

To save the heatmap for a story or a slide deck:

ggsave("offshore_heatmap.pdf", width = 7, height = 5)

cookbook: twitter `@mentions`

Event data is not just for conflict. Anything that arrives as rows of (source, target, time) – a tweet that mentions another handle, an email between coworkers, a code review that requests a reviewer – can be netified the exact same way. Here is a tiny self-contained Twitter @mentions example so you can copy/paste and adapt.

# tiny synthetic edge list: who mentioned whom, on which day
mentions <- data.frame(
    from = c("@alice", "@alice", "@bob",  "@bob",  "@carol",
             "@dave",  "@alice", "@eve",  "@eve",  "@frank"),
    to   = c("@bob",   "@carol", "@alice","@dave", "@dave",
             "@carol", "@eve",   "@bob",  "@alice","@alice"),
    day  = c(1, 1, 1, 2, 2, 2, 3, 3, 3, 3)
)

# directed (a mention has a sender and receiver); sum repeat mentions as
# the edge weight by letting netify count them for us
mention_net <- netify(
    mentions,
    actor1 = "from", actor2 = "to",
    time = "day",
    symmetric = FALSE,
    sum_dyads = TRUE
)

mention_net

## ✔ Network data created.
## • Unipartite
## • Asymmetric
## • Sum of Binary Weights
## • Longitudinal: 3 Periods
## • # Unique Actors: 6
## Network Summary Statistics (averaged across time):
##             dens miss recip trans
## weight_var 0.111    0 0.561     0
## • Nodal Features: None
## • Dyad Features: None

A quick plot() already gives you the picture:

plot(mention_net)

And the usual summary* helpers tell you who is most active and who is most mentioned:

summary(mention_net)

##   net num_actors   density num_edges prop_edges_missing competition_row
## 1   1          6 0.1000000         3                  0       0.5555556
## 2   2          6 0.1000000         3                  0       0.3333333
## 3   3          6 0.1333333         4                  0       0.3750000
##   competition_col sd_of_row_means sd_of_col_means covar_of_row_col_means
## 1       0.3333333       0.1673320       0.1095445              0.6546537
## 2       0.5555556       0.1095445       0.1673320              0.6546537
## 3       0.3750000       0.1632993       0.1632993              0.4000000
##   reciprocity    mutual transitivity
## 1   0.6296296 0.5000000            0
## 2   0.6296296 0.5000000            0
## 3   0.4230769 0.3333333            0

# directed networks split degree into in / out / total
mention_actor <- summary_actor(mention_net)
head(mention_actor[, c("actor", "time", "degree_in", "degree_out", "degree_total")])

##    actor time degree_in degree_out degree_total
## 1 @alice    1         1          2            3
## 2   @bob    1         1          1            2
## 3 @carol    1         1          0            1
## 4  @dave    1         0          0            0
## 5   @eve    1         0          0            0
## 6 @frank    1         0          0            0

Same recipe, different domain – the only thing that changes is the column names you hand to netify().

references

Sundberg, Ralph and Erik Melander (2013) Introducing the UCDP Georeferenced Event Dataset. Journal of Peace Research 50(4).
Boschee, E., Lautenschlager, J., O’Brien, S., Shellman, S., Starz, J., & Ward, M. (2015). ICEWS Coded Event Data. Harvard Dataverse. DOI 10.7910/DVN/28075.

Cassy Dorff and Shahryar Minhas

2026-06-05

tl;dr

one-mode event data: ucdp conflict in mexico

tl;dr sidebar: na versus zero in event-count weights

highlighting a single actor

applying a theme

saving the figure

a note on color choice

advanced: assemble the plot from components

a note on missingness

two-mode (bipartite) event data: offshore jurisdictions

cookbook: twitter `@mentions`

references

Event Data

Cassy Dorff and Shahryar Minhas

2026-06-05

tl;dr

one-mode event data: ucdp conflict in mexico

tl;dr sidebar: na versus zero in event-count weights

highlighting a single actor

applying a theme

saving the figure

a note on color choice

advanced: assemble the plot from components

a note on missingness

two-mode (bipartite) event data: offshore jurisdictions

cookbook: twitter @mentions

references

cookbook: twitter `@mentions`