plotProfLik: Plot profile likelihood for a parameter in an NBDA
In whoppitt/NBDA: A Package For Implementing Network-Based Diffusion Analysis

Description Usage Arguments Details Value Warning Plotting profile likelihood for a difference between two parameters See Also

View source: R/profLikCI.R

Find the profile likelihood for a specific parameter, or a difference between two parameters.

1
2
3

plotProfLik(which, model, range, constraintsVect = NULL,
  resolution = 20, iterations = 150, inflation = 1, conf = 0.95,
  retainInt = NULL)

`which`	numeric giving the parameter for which the profile likelihood is to be plotted. The appropriate number can be identified from the fitted model, by entering `<modelName>@varNames` to extract the variable names from the model. Each variable name is preceded by its number.
`model`	object of class `oadaFit` or `tadaFit`.
`range`	numeric vector of length two, providing a range of parameter values within which the profile likelihood is to be plotted.
`constraintsVect`	optional numerical vector. This only needs to be used if the confidence interval for the difference between two parameters is required (see specific section below).
`resolution`	numeric giving the number of points to be plotted. The user is advised to start at a low resolution to obtain the approproate range, then increase resolution to identify ranges for endpoints.
`iterations`	optional numerical giving the maximum number of iterations to be used by the optimization alogorithms.
`inflation`	numerical to be used if the confidence intervals are to be inflated by a specified amount, as suggested by Burnham & Anderson (2000) to allow for model selection uncertainty. This simply increases the height of the dotted line above the model maximum likelihood by a factor of `inflation`.
`conf`	numerical giving the level of confidence required, defaulting to the traditional 0.95.
`retainInt`	logical, can be used to force the model to retain int_ilvs in an asocial model. This is used internally by other functions when there is an offset on the s parameters, but can be safely ignored by the user.

The profile likelihood method for finding (100-X)% confidence intervals works by finding the set of values for a parameter that would not be rejected in a likelihood ratio test at the X% level of significance. This is equivalent to finding the set of values for which the profile likelihood (-log likelihood optimized over all other parameters in the model) is within C units of the -log-likelihood for the model, where C is the critical value for rejection at the X% level of significance (1.92 for 95% confidence intervals). The plotProfLik function can be used to plot the profile likelihood for a parameter and find the approximate location of the endpoints of the confidence interval after which profLikCI can be used to locate the exact endpoints. The plotProfLik function plots the profile likelihood for the specified parameter with a dotted line at the point of rejection (C) at the X% significance level. The endpoints of the confidence interval are where the profile likelihood crosses the dotted line. If necessary the user can reduce the range and re-plot to "zoom in" on each endpoint. Note if points are plotted in red then it means the optimization algorithm did not converge when calculating the profile likelihood for that point.

A dataframe giving the plotted values, and an indicator of whether the optimization algorithms converged (0) or not (1) when finding the profile likelihood for that point.

This function does not work when trueTies are present in an OADA. Instead use plotProfLikTrueTies for the confidence intervals on a parameter, or plotProfLikDiffTrueTies for the difference between two parameters.

This can be achieved using the constraintsVect argument. e.g. if we wish to find the confidence interval for parameter 1 - parameter 2, we specify which=1 and constraintsVect=c(1,1,2,3,etc.). This constrains parameter 1 and 2 to be the same, but adds an offset to parameter 1 using the constrainedNBDAdata function. The resulting profile likelihood is for parameter 1 - parameter 2. This can only be done for parameters of the same type i.e. differences must be within the s parameters, asoc_ilv, int_ilv or multi_ilv categories. If the user wishes to find confidence intervals for the difference between two s parameters which is thought to span zero, we advise doing this as a two step process. e.g. find the upper limit for s1-s2, setting range >0, then find the upper limit for s2-s1 setting range >0. This prevents values of s1 or s2<0 being condsidered in the optimization process, which may trigger errors.

profLikCI

whoppitt/NBDA documentation built on April 25, 2021, 7:55 a.m.