goldfish Effects

Catalog of available effects

Alvaro Uzaheta and Siwei Zhang

This vignette summarizes effects that are implemented in goldfish and are thus available to be used with actor-oriented DyNAM models and tie-oriented Relational Event Models.

Effect arguments

Effect functions have the following 5 arguments:

• network(s)/attribute(s): Objects upon which the effect should be calculated. For network effects, this would be either an explanatory network or the dependent network (the default). For attribute effects, this should be the attribute name with the indication of the data frame containing the initial values of nodes attributes (data.frame$attribute), see documentation of defineNodes() for details.
• isTwoMode: Identifies whether the effect assumes the network is originated from a two-mode network or not. The default value is FALSE, it means that updates involving self-ties are ignore during the statistics update.
• ignoreRep: Identifies whether the effect recognizes actors to send additional ties beyond the first to receivers (FALSE means additional ties are taken into account). The default value is FALSE. Only available for structural effects.
• weighted: Identifies whether the effect relies on the presence (dummy value) or number of ties (TRUE means it relies also on the number of ties). The default value is FALSE.
• window: Identifies a window length within which changes should apply to events. The default value is Inf meaning that no windows are applied to the effect. Window size can be specified as a number in seconds (i.e. an hour is 3600 seconds), or as a string stating the number of time units in the format “number unit”, for example “30 seconds”, “1 minute”, “2 hours”, “3 days”, “4 weeks”, “5 months”, or “6 years”. Note: The time units are converted to seconds using their most common lengths in seconds.
• transformFun: Use this parameter to obtain transformed statistics. The function used as argument is applied to the change statistic. Note: Most of the time a transformation function is applied when an effect counts the number of ties weighted == TRUE.
• aggregateFun: Use this parameter to obtain a transformed aggregated statistics for indirect effects like tertius and tertiusDiff. The function used as argument is applied to aggregated statistic.
• type: Applies only to indegree, outdegree and nodeTrans in the case of a REM model model = "REM". In the default case of type = "alter", the effect returns change statistics for the potential receivers. In the case where type = "ego", the effect returns change statistics for the potential sender. This argument does not apply in the case of the DyNAM models: the DyNAM-choice submodel only considers change statistics for potential receivers type = "alter", while the DyNAM-rate submodel only considers change statistics for potential senders type = "ego".

The following table summaries whether the corresponding arguments can be used for the effects or not.

			isTwoMode	ignoreRep	weighted	window	transformFun	aggregateFun	type
Node/actor statistics	Structural	indeg	√	√	√	√	√	×	√
		outdeg	√	√	√	√	√	×	√
		nodeTrans	×	√	√	√	√	×	√
	Attribute	ego	√	×	×	×	×	×	×
		alter	√	×	×	×	×	×	×
	Structural + attribute	tertius	√	√	√	√	√	√	√
Dyadic statistics	Structural	tie	√	√	√	√	√	×	×
		inertia	√	√	√	√	√	×	×
		recip	√	√	√	√	√	×	×
	Attribute	same	√	×	×	×	√	×	×
		diff	√	×	×	×	√	×	×
		sim	√	×	×	×	√	×	×
		egoAlterInt	√	×	×	×	√	×	×
	Structural + attribute	tertiusDiff	√	√	√	√	√	√	×
	Closure effects	trans	×	√	×	√	√	×	×
		cycle	×	√	×	√	√	×	×
		commonSender	√	√	×	√	√	×	×
		commonReceiver	√	√	×	√	√	×	×
		four	√	√	×	√	√	×	×
		mixedTrans	×	√	×	√	√	×	×
		mixedCycle	×	√	×	√	√	×	×
		mixedCommonSender	×	√	×	√	√	×	×
		mixedCommonReceiver	×	√	×	√	√	×	×

Note that the use of some effects (combinations) are ill-advised. For example, using tie(network, ignoreRep = FALSE), where the network refers to the dependent network, will always result in a change statistic of zero, and thus cannot be used.

Notation

• $x(t)_{ij}$ denotes the value of network $x$ at time $t$ between actor $i$ and actor $j$.
• $z(t)_i$ denotes the value of an actor attribute $z$ at time $t$ for actor $i$.
• $I(y)$ denotes the indicator function. It takes the value 1 when variable $y$ is TRUE and 0 in other case.
• We refer as the weighted statistics to the outcome of the effect when weighted = TRUE, similarly we refer as the unweighted statistics when weighted = FALSE.
• We omit the difference in the computation of the statistics when the argument isTwoMode is used. In the default case when isTwoMode = FALSE self ties are excluded in the calculation of the change statistics.

Node or actor statistics

Structural Effects

effect	rate	choice	choice coordination
indeg	√	√	√
outdeg	√	√	×
nodeTrans	√	√	√

Here we refer to ego type when type = "ego" and alter type when type = "alter".

Indegree (indeg())

indeg(network,
  isTwoMode = FALSE, weighted = FALSE, window = Inf,
  ignoreRep = FALSE, type = c("alter", "ego"), transformFun = identity
)

$$\begin{align} r(i, t, x) &= \begin{cases} \sum_{j}{I(x(t)_{ji}>0)} & \text{unweighted}\\ \sum_{j}{x(t)_{ji}} & \text{weighted} \end{cases} && \text{ego type}\\ s(i,j, t, x) &= \begin{cases} \sum_{l}{I(x(t)_{lj}>0)} & \text{unweighted}\\ \sum_{l}{x(t)_{lj}} & \text{weighted} \end{cases} && \text{alter type} \end{align}$$

• DyNAM-Rate model: tendency of actor $i$ to create an event when $i$ has a high incoming degree in a covariate network (‘ego’ type).
• DyNAM-choice model: tendency to create an event $i\rightarrow j$ when $j$ has a high incoming degree in a covariate network (‘alter’ type).
• REM model: tendency to create an event $i\rightarrow j$ when either $i$ or $j$ has a high incoming degree in a covariate network. The argument type allows to choose whether to use the indegree effect for sender $i$ (type = "ego") or for receiver j (type = "alter").
• DyNAM-choice_coordination: the effects compute the total degree in a covariate network (‘alter’ type). Note: The effect statistic correspond to the total degree for an undirected network as define when using defineNetwork(net, directed = FALSE).

The degree can be transform with .

Outdegree (outdeg())

outdeg(network,
  isTwoMode = FALSE, weighted = FALSE, window = Inf,
  ignoreRep = FALSE, type = c("alter", "ego"), transformFun = identity
)

$$\begin{align} r(i, t, x) &= \begin{cases} \sum_{j}{I(x(t)_{ij}>0)} & \text{unweighted}\\ \sum_{j}{x(t)_{ij}} & \text{weighted} \end{cases} && \text{ego type}\\ s(i,j, t, x) &= \begin{cases} \sum_{l}{I(x(t)_{jl}>0)} & \text{unweighted}\\ \sum_{l}{x(t)_{jl}} & \text{weighted} \end{cases} && \text{alter type} \end{align}$$

• DyNAM-Rate model: tendency of actor $i$ to create an event when $i$ has a high outgoing degree in a covariate network (‘ego’ type).
• DyNAM-choice model: tendency to create an event $i\rightarrow j$ when $j$ has a high outgoing degree in a covariate network (‘alter’ type).
• REM model: tendency to create an event $i\rightarrow j$ when either $i$ or $j$ has a high outgoing degree in a covariate network. The argument type allows to choose whether to use the outdegree effect for sender $i$ (type = "ego") or for receiver $j$ (type = "alter").

Node embeddedness transitivity (nodeTrans())

nodeTrans(network,
  isTwoMode = FALSE, window = Inf, ignoreRep = FALSE,
  type = c("alter", "ego"), transformFun = identity
)

$$\begin{align} r(i, t, x) &= \sum_{jk}{I(x(t)_{ik}>0)I(x(t)_{kj}>0)I(x(t)_{ij}>0)} && \text{ego type}\\ s(i,j, t, x) &= \sum_{kl}{I(x(t)_{jk}>0)I(x(t)_{kl}>0)I(x(t)_{jl}>0)} && \text{alter type} \end{align}$$

Embeddedness in transitive structures as a source node.

• DyNAM-Rate model: tendency of actor $i$ to create an event when $i$ is embeddeded in more transitive structures as the source $i\rightarrow k \rightarrow j \leftarrow i$ in a covariate network (‘ego’ type).
• DyNAM-choice model: tendency to create an event $i\rightarrow j$ when $j$ is embeddeded in more transitive structures as the source $j\rightarrow k \rightarrow l \leftarrow j$ in a covariate network (‘alter’ type).
• REM model: tendency to create an event $i\rightarrow j$ when either $i$ or $j$ are embeddeded in more transitive structures as the source in a covariate network. The argument type allows to choose whether to use the node embeddedness transitivity effect for sender $i$ (type = "ego") or for receiver $j$ (type = "alter").

The statistic can be transform with , there is not weighted version for this effect.

Attribute effects

effect	rate	choice	choice coordination
ego	√	×	×
alter	×	√	×

Ego attribute (ego())

ego(attribute, isTwoMode = FALSE)

$$\begin{align} r(i, t, z) &= z(t)_i \end{align}$$

• DyNAM-Rate model: tendency of actors to be more active when they score high on an attribute.
• REM model: tendency to create an event $i\rightarrow j$ when $i$ has a high score on an attribute.

Alter attribute (alter())

alter(attribute, isTwoMode = FALSE)

$$\begin{align} s(i, j, t, z) &= z(t)_j \end{align}$$

• DyNAM-choice/DyNAM-choice_coordination model: tendency to create an event $i\rightarrow j$ when $j$ score high on an attribute.
• REM model: tendency to create an event $i\rightarrow j$ when $j$ has a high score on an attribute.

Structural $+$ Attribute effect

effect	rate	choice	choice coordination
tertius	√	√	√

Tertius (tertius())

tertius(network, attribute,
  isTwoMode = FALSE, window = Inf,
  ignoreRep = FALSE, type = c("alter", "ego"), transformFun = identity,
  aggregateFun = function(x) mean(x, na.rm = TRUE)
)

$$\begin{align} r(i, t, x, z) &= \frac{\sum_{j:~x(t)_{ji} > 0}{z(t)_j}}{\sum_{j}{I(x(t)_{ji} > 0)}} && \text{ego type}\\ s(i,j, t, x, z) &= \frac{\sum_{k:~x(t)_{kj} > 0}{z(t)_k}}{\sum_{k}{I(x(t)_{kj} > 0)}} && \text{alter type} \end{align}$$

• DyNAM-Rate submodel: tendency of actor $i$ to create an event when $i$ has a high aggregate (aggregateFun) value of its in-neighbors ($\forall j:~ x[j, i] > 0$) in a covariate network (‘ego’ type).
• DyNAM-choice/choice_coordination submodels: tendency to create an event $i\rightarrow j$ when $j$ has a high aggregate (aggregateFun) value of its in-neighbors ($\forall k:~ x[k, j] > 0$) in a covariate network (‘alter’ type).
• REM model: tendency to create an event $i\rightarrow j$ when either $i$ or $j$ has a high aggregate (aggregateFun) value of its in-neighbors in a covariate network. The argument type allows to choose whether to use the tertius effect for sender $i$ (type = "ego") or for receiver $j$ (type = "alter").

Note: When a node does not have in-neighbors, the tertius effect is impute as the average of the aggregate values of nodes with in-neighbors.

Dyadic statistics

Structural effects

effect	rate	choice	choice coordination
tie	×	√	√
inertia	×	√	√
recip	×	√	×

Tie (tie())

tie(network,
  weighted = FALSE, window = Inf, ignoreRep = FALSE,
  transformFun = identity
)

$$\begin{equation} s(i,j, t, x) = \begin{cases} I(x(t)_{ij}>0) & \text{unweighted}\\ x(t)_{ij} & \text{weighted} \end{cases} \end{equation}$$

Tendency to create an event $i\rightarrow j$ if the tie $i\rightarrow j$ exists in a covariate network. Parameter weighted can be set to TRUE if the value in the covariate network for the dyad $i\rightarrow j$ is to be taken as a statistic. It can be transformed by using transformFun (This might make sense when weighted = TRUE).

Inertia (inertia())

inertia(network, weighted = FALSE, window = Inf, transformFun = identity)

$$\begin{equation} s(i,j, t, x) = \begin{cases} I(x(t)_{ij}>0) & \text{unweighted}\\ x(t)_{ij} & \text{weighted} \end{cases} \end{equation}$$

Usually used as the “intercept” for the choice submodel, inertia is the tendency to create an event $i\rightarrow j$ if the event $i\rightarrow j$ happened before. It can be interpreted as the differential tendency to update existing ties rather than creating new ones. Thus, inertia is similar to tie, but defined on the network to which the dependent events relate. Parameter weighted can be set to TRUE if the count of past events $i\rightarrow j$ is to be taken as a statistic. It can be transformed by using transformFun (this might make sense when weighted = TRUE). Note: inertia can never be used in combination with a ignoreRep = TRUE parameter as this would replace all positive statistics with zeros.

Tie reciprocation (recip())

recip(network,
  weighted = FALSE, window = Inf, ignoreRep = FALSE,
  transformFun = identity
)

$$\begin{equation} s(i,j, t, x) = \begin{cases} I(x(t)_{ji}>0) & \text{unweighted}\\ x(t)_{ji} & \text{weighted} \end{cases} \end{equation}$$

Effect of a tie j->i on event $i\rightarrow j$. Recip cannot be used with binary = T in undirected dependent network

Tendency to create an event $i\rightarrow j$ if one or several $j\rightarrow i$ happened before. Parameter weighted can be set to TRUE if the count/weight of $j\rightarrow i$ events/ties is to be taken as a statistic. It can be transformed by using transformFun (this might make sense when weighted = TRUE). This effect cannot be used for two-mode networks and for DyNAM-choice_coordination submodel.

Attribute effects

effect	rate	choice	choice coordination
same	×	√	√
diff	×	√	√
sim	×	√	√
egoAlterInt	×	√	√

Same value in attribute (same())

same(attribute)

$$\begin{equation} s(i,j, t, z) = I(z(t)_i = z(t)_j) \end{equation}$$

Homophily (same value). The tendency of an event $i\rightarrow j$ to happen if actors i and j have the same attribute value. This effect cannot be used for two-mode networks and for the DyNAM-rate submodel.

Differences in attribute (diff())

diff(attribute, transformFun = abs)

$$\begin{equation} s(i,j, t, z) = |z(t)_i - z(t)_j| \end{equation}$$

Heterophily. The tendency of an event $i\rightarrow j$ to happen if actors i and j have different attribute values (high absolute differences regarding attribute if transformFun = abs). This effect cannot be used for two-mode networks.

Similarity in attribute (sim())

sim(attribute, transformFun = abs)

$$\begin{equation} s(i,j, t, z) = -|z(t)_i - z(t)_j| \end{equation}$$

Homophily (similar value). The tendency of an event $i\rightarrow j$ to happen if actors i and j have similar attribute values (low absolute differences regarding attribute if transformFun = abs). This effect cannot be used for two-mode networks.

Ego Alter interaction (egoAlterInt())

egoAlterInt(attribute = list(attribute1, attribute2))

$$\begin{equation} s(i,j, t, z^{(1)}, z^{(2)}) = z(t)_i^{(1)} * z(t)_j^{(2)} \end{equation}$$

In a model that includes an alter effect using attribute2, the ego alter interaction helps to study the tendency to create an event $i \rightarrow j$ when $j$ score high on attribute2 moderated by the score of ego on attribute1.

Structural $+$ Attribute effect

effect	rate	choice	choice coordination
tertiusDiff	×	√	√

Tertius difference (tertiusDiff())

tertiusDiff(network, attribute,
  isTwoMode = FALSE, weighted = FALSE,
  window = Inf, ignoreRep = FALSE, transformFun = abs,
  aggregateFun = function(x) mean(x, na.rm = TRUE)
)

$$\begin{equation} s(i,j, t, x, z) = \left|\frac{z(t)_i - \sum_{k:~x(t)_{kj} > 0}{z(t)_k}}{ \sum_{k}{I(x(t)_{kj} > 0)}}\right| \end{equation}$$

The tendency to create an event $i\rightarrow j$ when $i$ has a similar value as $j$ aggregate (aggregateFun) value of its in-neighbors ($\forall j:~ x[j, i] > 0$). Note: When the node $j$ does not have in-neighbors, its value is imputed by the average of the similarities computed for the pairs $i$, $k$ for all $k$ that has at least one in-neighbor.

Closure effect statistics (Dyadic effects)

effect	rate	choice	choice coordination
trans	×	√	√
cycle	×	√	×
commonSender	×	√	×
commonReceiver	×	√	×
four	×	√	√
mixedTrans	×	√	√
mixedCycle	×	√	×
mixedCommonSender	×	√	×
mixedCommonReceiver	×	√	×

Transitivity (trans())

trans(network, window = Inf, ignoreRep = FALSE, transformFun = identity)

$$\begin{equation} s(i,j, t, x) = \sum_{k}{I(x(t)_{ik}>0)I(x(t)_{kj}>0)} \end{equation}$$

It is the tendency to create an event $i\rightarrow j$ when it closes more two-paths ($i\rightarrow k\rightarrow j$) observed in the past events in a covariate network. It can be transformed by using transformFun. This effect cannot be used for two-mode networks.

Cycle (cycle())

cycle(network, window = Inf, ignoreRep = FALSE, transformFun = identity)

$$\begin{equation} s(i,j, t, x) = \sum_{k}{I(x(t)_{jk}>0)I(x(t)_{ki}>0)} \end{equation}$$

It is the tendency to create an event $i\rightarrow j$ when it closes more two-paths ($j\rightarrow k\rightarrow i$) observed in the past events in a covariate network. It can be transformed by using transformFun. This effect cannot be used for two-mode networks and DyNAM-choice_coordination.

Common sender closure (commonSender())

commonSender(network, window = Inf, ignoreRep = FALSE, transformFun = identity)

$$\begin{equation} s(i,j, t, x) = \sum_{k}{I(x(t)_{ki}>0)I(x(t)_{kj}>0)} \end{equation}$$

It is the tendency to create an event $i\rightarrow j$ when it closes more two-paths ($i\leftarrow k\rightarrow j$) observed in the past events in a covariate network. It can be transformed by using transformFun. This effect cannot be used for two-mode networks and DyNAM-choice_coordination.

Common receiver closure (commonReceiver())

commonReceiver(network,
  window = Inf, ignoreRep = FALSE,
  transformFun = identity
)

$$\begin{equation} s(i,j, t, x) = \sum_{k}{I(x(t)_{ik}>0)I(x(t)_{jk}>0)} \end{equation}$$

It is the tendency to create an event $i\rightarrow j$ when it closes more two-paths ($i\rightarrow k\leftarrow j$) observed in the past events in a covariate network. It can be transformed by using transformFun. This effect cannot be used for two-mode networks and DyNAM-choice_coordination.

Four cycle (four())

four(network,
  isTwoMode = FALSE, window = Inf, ignoreRep = FALSE,
  transformFun = identity
)

$$\begin{equation} s(i,j, t, x) = \sum_{kl}{I(x(t)_{ik}>0)I(x(t)_{lk}>0)I(x(t)_{lj}>0)} \end{equation}$$

Closure of three-paths. It is the tendency to create an event $i\rightarrow j$ when it closes more three-paths ($i\rightarrow k\leftarrow l\rightarrow j$) observed in the past events in a covariate network. It can be transformed by using transformFun. This effect cannot be used for two-mode networks and DyNAM-choice_coordination.

Mixed Transitivity (mixedTrans())

mixedTrans(
  network = list(network1, network2), window = Inf,
  ignoreRep = FALSE, transformFun = identity
)

$$\begin{equation} s(i,j, t, x^{(1)}, x^{(2)}) = \sum_{k}{I(x(t)_{ik}^{(1)}>0)I(x(t)_{kj}^{(2)}>0)} \end{equation}$$

Transitivity within 2 networks. It is the tendency to create an event $i\rightarrow j$ when it closes more two-paths with events ($i\rightarrow k$) in network1 and ($k\rightarrow j$) in network2 observed in the past events in the covariate networks. It can be transformed by using transformFun. This effect cannot be used for two-mode networks.

Mixed Cycle (mixedCycle())

mixedCycle(
  network = list(network1, network2), window = Inf,
  ignoreRep = FALSE, transformFun = identity
)

$$\begin{equation} s(i,j, t, x^{(1)}, x^{(2)}) = \sum_{k}{I(x(t)_{ki}^{(1)}>0)I(x(t)_{jk}^{(2)}>0)} \end{equation}$$

Cycle within 2 networks. It is the tendency to create an event $i\rightarrow j$ when it closes more two-paths with events ($k\rightarrow i$) in network1 and ($j\rightarrow k$) in network2 observed in the past events in the covariate networks. It can be transformed by using transformFun. This effect cannot be used for two-mode networks and DyNAM-choice_coordination.

Mixed common sender closure (mixedCommonSender())

mixedCommonSender(
  network = list(network1, network2), window = Inf,
  ignoreRep = FALSE, transformFun = identity
)

$$\begin{equation} s(i,j, t, x^{(1)}, x^{(2)}) = \sum_{k}{I(x(t)_{ki}^{(1)}>0)I(x(t)_{kj}^{(2)}>0)} \end{equation}$$

Closure common sender within 2 networks. It is the tendency to create an event $i\rightarrow j$ when it closes more two-paths with events ($k\rightarrow i$) in network1 and ($k\rightarrow j$) in network2 observed in the past events in the covariate networks. It can be transformed by using transformFun. This effect cannot be used for two-mode networks and DyNAM-choice_coordination.

Mixed common receiver closure (mixedCommonReceiver())

mixedCommonReceiver(
  network = list(network1, network2), window = Inf,
  ignoreRep = FALSE, transformFun = identity
)

$$\begin{equation} s(i,j, t, x^{(1)}, x^{(2)}) = \sum_{k}{I(x(t)_{ik}^{(1)}>0)I(x(t)_{jk}^{(2)}>0)} \end{equation}$$

Closure receiver within 2 networks. It is the tendency to create an event $i\rightarrow j$ when it closes more two-paths with events ($i\rightarrow k$) in network1 and ($j\rightarrow k$) in network2 observed in the past events in the covariate networks. It can be transformed by using transformFun. This effect cannot be used for two-mode networks and DyNAM-choice_coordination.