LogLikelihood: Log-likelihood functions and Likelihood Ratio Tests for moderated networks

Description

Computes log-likelihood, AIC, and BIC for a whole network, or for each node in the network. Also compares two or more networks using a likelihood ratio test (LRT).

Usage

modLL(
  net0,
  net1 = NULL,
  nodes = FALSE,
  lrt = NULL,
  all = FALSE,
  d = 4,
  alpha = 0.05,
  orderBy = NULL,
  decreasing = TRUE
)
SURll(
  net0,
  net1 = NULL,
  nodes = FALSE,
  lrt = NULL,
  all = FALSE,
  d = 4,
  alpha = 0.05,
  s = "res",
  orderBy = NULL,
  decreasing = TRUE,
  sysfits = FALSE
)
modTable(
  net0,
  nodes = FALSE,
  orderBy = TRUE,
  d = 4,
  alpha = 0.05,
  decreasing = TRUE,
  names = NULL,
  rmsea = FALSE
)
SURtable(
  net0,
  nodes = FALSE,
  orderBy = TRUE,
  d = 4,
  alpha = 0.05,
  decreasing = TRUE,
  names = NULL,
  rmsea = FALSE,
  s = "res"
)

Arguments

net0

Output from one of the main modnets functions. Alternatively, a list of network models can be provided. This list should be named for the easiest interpretation of results.

net1

For modLL and SURll, can be used to supply a second network to compare to net0 via an LRT. Or if lrt = FALSE, then relevant statistics will be returned for both net0 and net1. Importantly, if one network is provided for net0, and another is provided for net1, then the names in the output will reflect these argument names. This can be somewhat confusing at times, so ultimately it is not recommended to use this argument. Instead, try supplying both networks (or more) as a named list to the net0 argument for the most customization.

nodes

Logical. Determines whether to compute omnibus or nodewise statistics and tests. If TRUE, then LL values for nodewise models will be returned, and any LRTs requested will reflect nodewise tests.

lrt

Logical. Determines whether to conduct an LRT or not. If FALSE, then only LL-related statistics will be returned for all models supplied. Only relevant for modLL and SURll

all

Logical. If TRUE, then omnibus LL statistics as well as nodewise statistics are returned for either LL function.

Number of decimal places to round outputted statistics to.

alpha

Alpha level for LRTs. Defaults to .05.

orderBy

Can be one of "LL", "df", "AIC", "BIC" to indicate which statistic to order the table by. If using modTable or SURtable, then a value of TRUE will organize the output by the LRT column, which indicates the number of times that a particular model performed better than another based on the pairwise LRTs. Higher values indicate that the model was selected more often in comparison with other models in the list.

decreasing

Logical. Determines whether to organize output from highest to lowest, or vice versa, in accordance with the value of orderBy.

Character string indicating which type of residual covariance matrix to compute for SUR models. Options include "res", "dfres", "sigma". "sigma" uses the residual covariance matrix as computed by the systemfit::systemfit function. "res" and "dfres" compute the matrix based directly on the residual values. "dfres" is the sample estimator that uses N - 1 in the denominator, while "res" just uses N.

sysfits

Logical, only relevant to SURll when multiple networks are included in a list. Does not currently work when there are two networks in the list, but does work with 3 or more. Returns the omnibus model statistics based on functions available to output from the systemfit::systemfit function. This allows for some additional statistics such as SSR, detSigma, OLS.R2, McElroy.R2.

names

Character vector containing the names of the models being compared. Only relevant to the modTable and SURtable. Alternatively, models can be named by supplying a named list to the net0 argument.

rmsea

Logical. Relevant to modTable and SURtable. Determines whether to return RMSEA values, as well as tests comparing RMSEA across each pair of models.

Value

A table or list of results depending on which function is used.

Details

Fits LRT to a list of network models to compare them all against each other. Obtain all possible LRTs comparing a list of SUR models. Can include tests comparing RMSEA values. The nodes argument determines whether to perform these computations in an omnibus or nodewise fashion.

One key thing to note is that when using modTable or SURtable, the LRT column indicates the number of times that each network was selected over others with respect to the pairwise LRTs.

Examples

Run this code

# NOT RUN {
data <- na.omit(psychTools::msq[, c('hostile', 'lonely', 'nervous', 'sleepy', 'depressed')])

##### Use modLL() for GGMs
ggm1 <- fitNetwork(data[, -5])
ggm2 <- fitNetwork(data, covariates = 5)
ggm3 <- fitNetwork(data, moderators = 5)

modLL(ggm1)
modLL(ggm2)

modLL(ggm1, ggm2)
modLL(ggm1, ggm2, nodes = TRUE)

modLL(list(ggm1 = ggm1, ggm2 = ggm2))
modLL(list(GGM1 = ggm1, GGM2 = ggm2), nodes = TRUE)

ggms <- list(ggm1, ggm2, ggm3)

modLL(ggms)
modTable(ggms)
modTable(ggms, names = c("GGM1", "GGM2", "GGM3"))

names(ggms) <- c("GGM1", "GGM2", "GGM3")
modTable(ggms)
modLL(ggms)

##### Use SURll() for SUR networks
sur1 <- fitNetwork(data[, -5], lags = TRUE)
sur2 <- fitNetwork(data, covariates = 5, lags = TRUE)
sur3 <- fitNetwork(data, moderators = 5, lags = TRUE)

SURll(sur1)
SURll(sur2)

SURll(sur1, sur2)
SURll(sur1, sur2, nodes = TRUE)
SURll(list(SUR1 = sur1, SUR2 = sur2), nodes = TRUE)

surs <- list(sur1, sur2, sur3)

SURll(surs)
SURtable(surs, names = c('SUR1', "SUR2", "SUR3"))

names(surs) <- c("SUR1", "SUR2", "SUR3")
SURll(surs)
SURtable(surs)
# }