reciprocityStat: Calculate reciprocity statistics
In brandenberger/rem: Relational Event Models (REM)

Description Usage Arguments Details Author(s) See Also Examples

View source: R/rem.R

Calculate the endogenous network statistic reciprocity for relational event models. reciprocity measures the tendency for senders to reciprocate prior events where they were targeted by other senders. One-mode network statistic only.

reciprocityStat(data, time, sender, target, halflife, 
    weight = NULL, 
    eventtypevar = NULL,
    eventtypevalue = "valuematch", 
    eventfiltervar = NULL,
    eventfiltervalue = NULL, 
    eventvar = NULL,
    variablename = "recip", 
    returnData = FALSE,
    showprogressbar = FALSE, 
    inParallel = FALSE, cluster = NULL)

`data`	A data frame containing all the variables.
`time`	Numeric variable that represents the event sequence. The variable has to be sorted in ascending order.
`sender`	A string (or factor or numeric) variable that represents the sender of the event.
`target`	A string (or factor or numeric) variable that represents the target of the event.
`halflife`	A numeric value that is used in the decay function. The vector of past events is weighted by an exponential decay function using the specified halflife. The halflife parameter determines after how long a period the event weight should be halved. E.g. if `halflife = 5`, the weight of an event that occurred 5 units in the past is halved. Smaller halflife values give more importance to more recent events, while larger halflife values should be used if time does not affect the time between events that much.
`weight`	An optional numeric variable that represents the weight of each event. If `weight = NULL` each event is given an event weight of `1`.
`eventtypevar`	An optional variable that represents the type of the event. Use `eventtypevalue` to specify how the `eventtypevar` should be used to filter past events.
`eventtypevalue`	An optional value (or set of values) used to specify how paste events should be filtered depending on their type. `eventtypevalue = "valuematch"` indicates that only past events that have the same type as the current event should be used to calculate the reciprocity statistic. `eventtypevalue = "valuemix"` indicates that past and present events of specific types should be used for the reciprocity statistic. All the possible combinations of the eventtypevar-values will be used. E.g. if `eventtypevar` contains three unique values "a" and "b", 4 reciprocity statistics will be calculated. The first variable calculates the reciprocity effect where the present event is of type "a" and all the past events are of type "b". The next variable calculates reciprocity for present events of type "b" and past events of type "a". Additionally, a variable is calculated, where present events as well as past events are of type "a" and a fourth variable calculates reciprocity for events with type "b" (i.e. valuematch on value "b"). `eventtypevalue = c(.., ..)`, similar to the "`nodmix`"-option, all different combinations of the values specified in `eventtypevalue` are used to create reciprocity statistics.
`eventfiltervar`	An optional numeric/character/or factor variable for each event. If `eventfiltervar` is specified, `eventfiltervalue` has to be provided as well.
`eventfiltervalue`	An optional character string that represents the value for which past events should be filtered. To filter the current events, use `eventtypevar`.
`eventvar`	An optional dummy variable with 0 values for null-events and 1 values for true events. If the `data` is in the form of counting process data, use the `eventvar`-option to specify which variable contains the 0/1-dummy for event occurrence. If this variable is not specified, all events in the past will be considered for the calulation of the reciprocity statistic, regardless if they occurred or not (= are null-events).
`variablename`	An optional value (or values) with the name the reciprocity statistic variable should be given. To be used if `returnData = TRUE` or multiple reciprocity statistics are calculated.
`returnData`	`TRUE/FALSE`. Set to `FALSE` by default. The new variable(s) are bound directly to the `data.frame` provided and the data frame is returned in full.
`showprogressbar`	`TRUE/FALSE`. Can only be set to TRUE if the function is not run in parallel.
`inParallel`	`TRUE/FALSE`. An optional boolean to specify if the loop should be run in parallel.
`cluster`	An optional numeric or character value that defines the cluster. By specifying a single number, the cluster option uses the provided number of nodes to parallellize. By specifying a cluster using the `makeCluster`-command in the `doParallel`-package, the loop can be run on multiple nodes/cores. E.g., `cluster = makeCluster(12, type="FORK")`.

The reciprocityStat()-function calculates an endogenous statistic that measures whether senders have a tendency to reciprocate events.

The effect is calculated as follows:

G_t = G_t(E) = (A, B, w_t),

G_t represents the network of past events and includes all events E. These events consist each of a sender a in A and a target b in B and a weight function w_t:

w_t(i, j) = ∑_{e:a = i, b = j} | w_e | * exp^{-(t-t_e)* (ln(2)/T_{1/2})} * (ln(2)/T_{1/2}),

where w_e is the event weight (usually a constant set to 1 for each event), t is the current event time, t_e is the past event time and T_{1/2} is a halflife parameter.

For the reciprocity effect, the past events G_t are filtered to include only events where the senders are the present targets and the targets are the present senders:

reciprocity(G_t , a , b) = w_t(b, a)

An exponential decay function is used to model the effect of time on the endogenous statistics. Each past event that involves the sender as target and the target as sender, and fulfills additional filtering options specified via event type or event attributes, is weighted with an exponential decay. The further apart the past event is from the present event, the less weight is given to this event. The halflife parameter in the reciprocityStat()-function determines at which rate the weights of past events should be reduced.

The eventtypevar- and eventattributevar-options help filter the past events more specifically. How they are filtered depends on the eventtypevalue- and eventattributevalue-option.

Laurence Brandenberger laurence.brandenberger@eawag.ch

rem-package

# create some data with 'sender', 'target' and a 'time'-variable
# (Note: Data used here are random events from the Correlates of War Project)
sender <- c('TUN', 'NIR', 'NIR', 'TUR', 'TUR', 'USA', 'URU', 
            'IRQ', 'MOR', 'BEL', 'EEC', 'USA', 'IRN', 'IRN', 
            'USA', 'AFG', 'ETH', 'USA', 'SAU', 'IRN', 'IRN',
            'ROM', 'USA', 'USA', 'PAN', 'USA', 'USA', 'YEM', 
            'SYR', 'AFG', 'NAT', 'NAT', 'USA')
target <- c('BNG', 'ZAM', 'JAM', 'SAU', 'MOM', 'CHN', 'IRQ', 
            'AFG', 'AFG', 'EEC', 'BEL', 'ITA', 'RUS', 'UNK',
            'IRN', 'RUS', 'AFG', 'ISR', 'ARB', 'USA', 'USA',
            'USA', 'AFG', 'IRN', 'IRN', 'IRN', 'AFG', 'PAL',
            'ARB', 'USA', 'EEC', 'BEL', 'PAK')
time <- c('800107', '800107', '800107', '800109', '800109', 
          '800109', '800111', '800111', '800111', '800113',
          '800113', '800113', '800114', '800114', '800114', 
          '800116', '800116', '800116', '800119', '800119',
          '800119', '800122', '800122', '800122', '800124', 
          '800125', '800125', '800127', '800127', '800127', 
          '800204', '800204', '800204')
type <- sample(c('cooperation', 'conflict'), 33,
               replace = TRUE)
important <- sample(c('important', 'not important'), 33,
                    replace = TRUE)

# combine them into a data.frame
dt <- data.frame(sender, target, time, type, important)

# create event sequence and order the data
dt <- eventSequence(datevar = dt$time, dateformat = "%y%m%d", 
                    data = dt, type = "continuous", 
                    byTime = "daily", returnData = TRUE,
                    sortData = TRUE)

# create counting process data set (with null-events) - conditional logit setting
dts <- createRemDataset(dt, dt$sender, dt$target, dt$event.seq.cont, 
                          eventAttribute = dt$type, 
                          atEventTimesOnly = TRUE, untilEventOccurrs = TRUE, 
						  returnInputData = TRUE)
## divide up the results: counting process data = 1, original data = 2
dtrem <- dts[[1]]
dt <- dts[[2]]
## merge all necessary event attribute variables back in
dtrem$type <- dt$type[match(dtrem$eventID, dt$eventID)]
dtrem$important <- dt$important[match(dtrem$eventID, dt$eventID)]
# manually sort the data set
dtrem <- dtrem[order(dtrem$eventTime), ]

# calculate reciprocity statistic
dtrem$recip <- reciprocityStat(data = dtrem,
                            time = dtrem$eventTime, 
                            sender = dtrem$sender, 
                            target = dtrem$target,
                            eventvar = dtrem$eventDummy,
                            halflife = 2)

# plot sender-outdegree over time
library("ggplot2")
ggplot(dtrem, aes(eventTime, recip, 
               group = factor(eventDummy), color = factor(eventDummy)) ) +
  geom_point()+ geom_smooth() 

# calculate reciprocity statistic with typematch
# if a cooperated with b in the past, does
# b cooperate with a now?
dtrem$recip.typematch <- reciprocityStat(data = dtrem,
                               time = dtrem$eventTime, 
                               sender = dtrem$sender, 
                               target = dtrem$target,
                               eventvar = dtrem$eventDummy,
                               eventtypevar = dtrem$type,
                               eventtypevalue = 'valuematch',
                               halflife = 2)

# calculate reciprocity with valuemix on type
dtrem <- reciprocityStat(data = dtrem,
                         time = dtrem$eventTime, 
                         sender = dtrem$sender, 
                         target = dtrem$target,
                         eventvar = dtrem$eventDummy,
                         eventtypevar = dtrem$type,
                         eventtypevalue = 'valuemix',
                         halflife = 2, 
                         returnData = TRUE)

# calculate reciprocity and count important events only
dtrem$recip.filtered <- reciprocityStat(data = dtrem,
                               time = dtrem$eventTime, 
                               sender = dtrem$sender, 
                               target = dtrem$target,
                               eventvar = dtrem$eventDummy,
                               eventfiltervar = dtrem$important, 
                               eventfiltervalue = 'important',
                               halflife = 2)