R: Likelihood ratio test

mxCompare {OpenMx}

R Documentation

Likelihood ratio test

Description

Compare the fit of one or more models to that of a reference (base) model or set of reference models.

Usage

mxCompare(base, comparison, ..., all = FALSE,
  boot=FALSE, replications=400, previousRun=NULL, checkHess=FALSE)
mxCompareMatrix(models,
			    diag=c('minus2LL','ep','df','AIC'),
			    stat=c('p', 'diffLL','diffdf'), ...,
  boot=FALSE, replications=400, previousRun=NULL,
 checkHess=FALSE, wholeTable=FALSE)

Arguments

`base`	A MxModel object or list of MxModel objects.
`comparison`	A MxModel object or list of MxModel objects.
`models`	A MxModel object or list of MxModel objects.
`diag`	statistic used for diagonal entries
`stat`	statistic used for off-diagonal entries
`...`	Not used.
`all`	Boolean. Whether to compare all base models with all comparison models. Defaults to FALSE.
`boot`	Whether to use the bootstrap distribution to compute the p-value.
`replications`	How many replications to use to approximate the bootstrap distribution.
`previousRun`	Results to re-use from a previous bootstrap.
`checkHess`	Whether to approximate the Hessian in each replication
`wholeTable`	Return the whole table instead of a matrix shaped summary

Details

mxCompare is used to compare the fit of one or more mxModels to one or more comparison models. mxCompareMatrix compares all the models provided against each other.

Model comparisons are made by subtracting the fit statistics for the comparison model from the fit statistics for the base model. Raw fit statistics of each ‘base’ model are also listed in the output table.

The fit statistics compared depend on the kinds of models compared. Models fit with maximum likelihood are compared based on their minus two log likelihood values. Under certain regularity conditions, the difference in minus two log likelihood values from nested models is chi-squared distributed and forms a likelihood ratio test statistic. Models fit with weighted least squares are compared based on their Satorra-Bentler (2001) scaled difference chi-squared test statistics. Under full weighted least squares, the Satorra-Bentler chi-squared value is equal to the difference in the model chi-squared values; however, for unweighted and diagonally weighted least squares, the two are no longer equal. Satorra and Bentler (2001) showed that that their test statistic behaved well under a variety of conditions, including small sample sizes. By contrast the much simpler difference in the chi-squared statistics only behaved well under large sample sizes (e.g., greater than or equal to 300 rows of data).

Specific to weighted least squares, researchers sometimes use mean-adjusted chi-squared statistics and mean-and-variance scaled chi-squared statistics. Some programs call these WLSM and WLSMV statistics. In some cases, it is fine to evaluate the total fit of a model using adjusted and scaled chi-squared statistics. However, never, ever, ever, ..., ever take differences in mean-adjusted chi-squared statistics, and use them for nested model comparisons. Similarly, never, ever, ever, ..., ever, ever take differences in mean-and-variance scaled chi-squared statistics, and use them for nested model comparisons. The differences in these adjusted and scaled chi-squared statistics are not chi-squared distributed and do not form a valid basis for model comparison. So, just don't do it.

Although not always checked by mxCompare, you should never compare models with different data sets or that use different variables from the same data set. mxCompare might not stop you from doing this, so be thoughtful when comparing models. Make sure your models are nested and use the same data. Weighted least squares models are one case of comparing different data sets that requires particular care. When comparing WLS models, make sure you are using the same exogenous covariates for all compared models. Because WLS is a multi-stage estimation approach, exogenous covariates residualize and change the data fitted in WLS. Consequently, WLS models with different exogenous covariates actually have different data. By contrast, maximum likelihood models with different exogenous covariates still use the same data and are valid to compare.

The mxCompare function makes an effort to only make valid comparisons. If a comparison is made where the comparison model has a higher minus 2 log likelihood (-2LL) than the base model, then the difference in their -2LLs will be negative. P-values for likelihood ratio tests will not be reported when either the -2LL or degrees of freedom for the comparison are negative. To ensure that the differences between models are positive and yield p-values for likelihood ratio tests, models listed in the ‘base’ argument must be more saturated (i.e., more estimated parameters and fewer degrees of freedom) than models listed in the ‘comparison’ argument. For mxCompareMatrix only the comparisons that make sense will be included.

When multiple models are included in both the ‘base’ and ‘comparison’ arguments, then comparisons are made between the two lists of models based on the value of the ‘all’ argument. If ‘all’ is set to FALSE (default), then the first model in the ‘base’ list is compared to the first model in the ‘comparison’ list, second with second, and so on. If there are an unequal number of ‘base’ and ‘comparison’ models, then the shorter list of models is repeated to match the length of the longer list. For example, comparing base models ‘B1’ and ‘B2’ with comparison models ‘C1’, ‘C2’ and ‘C3’ will yield three comparisons: ‘B1’ with ‘C1’, ‘B2’ with ‘C2’, and ‘B1’ with ‘C3’. Each of those comparisons are prefaced by a comparison between the base model and a missing comparison model to present the fit of the base model.

If ‘all’ is set to TRUE, all possible comparisons between base and comparison models are made, and one entry is made for each base model. All comparisons involving the first model in ‘base’ are made first, followed by all comparisons with the second ‘base’ model, and so on. When there are multiple models in either the ‘base’ or ‘comparison’ arguments but not both, then the ‘all’ argument does not affect the set of comparisons made.

The following columns appear in the output for maximum likelihood comparisons:

base: Name of the base model.
comparison: Name of the comparison model. Is <NA> for the first
ep: Estimated parameters of the comparison model.
minus2LL: Minus 2*log-likelihood of the comparison model. If the comparison model is <NA>, then the minus 2*log-likelihood of the base model is given.
df: Degrees in freedom of the comparison model. If the comparison model is <NA>, then the degrees of freedom of the base model is given.
AIC: Akaike's Information Criterion for the comparison model. If the comparison model is <NA>, then the AIC of the base model is given.
diffLL: Difference in minus 2*log-likelihoods of the base and comparison models. Will be positive when base model -2LL is higher than comparison model -2LL.
diffdf: Difference in degrees of freedoms of the base and comparison models. Will be positive when base model DF is lower than comparison model DF (base model estimated parameters is higher than comparison model estimated parameters)
p: P-value for likelihood ratio test based on diffLL and diffdf values.

Weighted least squares reports a similar set of columns with four substitutions:

chisq: Replaces the minus2LL column. This is the comparison model's chi-squared statistic from Browne (1984, Equation 2.20a), accounting for some misspecification of the weight matrix.
AIC: Although this has the same name as that in maximum likelihood, it is really a pseudo-AIC using the comparison model chi-squared and the number of estimated parameters. It is the chi-squared value plus two times the number of free parameters.
SBchisq: Replaces the diffLL column. This is the Satorra-Bentler (2001, p. 511) scaled difference chi-squared statisic between the base model and the comparison model. If your models use full weighted least squares, then this will be the same as the difference between the individual model chi-squared statistics. However, for unweighted and diagonally weighted least square, the SB chisq will not be equal to the difference between the component model chi-squared statistics.
p: p-value for the Satorra-Bentler chi-squared statistic.

In addition to the particular columns for maximum likelihood and weighted least squares, there are three general columns that are not printed but are accessible via the $ and [ extractors.

fit: The individual model fit value: m2ll for maximum likelihood models, chisq for WLS models.
fitUnits: The units of the fit function: "-2LL" for ML models, "r'Wr" for WLS models.
diffFit: The difference in fit values between the base and comparison models: diffLL for ML models, SBchisq for WLS models.

mxCompare will give a p-value for any comparison in which both ‘diffLL’ and ‘diffdf’ are non-negative. However, this p-value is based on the assumptions of the likelihood ratio test, specifically that the two models being compared are nested. The likelihood ratio test and associated p-values are not valid when the comparison model is not nested in the referenced base model. For a more accurate p-value, the empirical bootstrap distribution can be computed (‘boot=TRUE’). However, ‘replications’ must be set high enough for an accurate approximation. The Monte Carlo SE of a proportion for B replications is \sqrt(p*(1-p)/B), but this will be zero if p is zero, which is nonsense. Note that a parametric-bootstrap p-value of zero must be interpreted as p < 1/B, which, depending on B and the desired Type I error rate, may not be "statistically significant."

When ‘boot=TRUE’, the model has a default compute plan, and ‘checkHess’ is kept at FALSE then the Hessian will not be approximated or checked. On the other hand, ‘checkHess’ is TRUE then the Hessian will be approximated by finite differences. This procedure is of some value because it can be informative to check whether the Hessian is positive definite (see mxComputeHessianQuality). However, approximating the Hessian is often costly in terms of CPU time. For bootstrapping, the parameter estimates derived from the resampled data are typically of primary interest.

note: The mxCompare function does not directly accept a digits argument, and depends on the value of the 'digits' option. To set the minimum number of significant digits printed, use options('digits' = N) (see example).

Value

Returns a new MxCompare object. If you want something more like a table of results, use as.data.frame() on the returned MxCompare object.

Examples


data(demoOneFactor)
manifests <- names(demoOneFactor)
latents <- "G1"
model1 <- mxModel(model="One Factor", type="RAM",
      manifestVars = manifests,
      latentVars = latents,
      mxPath(from = latents, to=manifests),
      mxPath(from = manifests, arrows = 2),
      mxPath(from = latents, arrows = 2, free = FALSE, values = 1.0),
      mxData(cov(demoOneFactor), type = "cov", numObs = 500)
)

fit1 <- mxRun(model1)

latents <- c("G1", "G2")
model2 <- mxModel(model="One factor Rasch equated", type="RAM",
      manifestVars = manifests,
      latentVars = latents,
      mxPath(from = latents[1], to=manifests[1:5], labels='raschEquated'),

      mxPath(from = manifests, arrows = 2),
      mxPath(from = latents, arrows = 2, free = FALSE, values = 1.0),
      mxData(cov(demoOneFactor), type = "cov", numObs=500)
)
fit2 <- mxRun(model2)

mxCompare(fit1, fit2) # Rasch equated is significantly worse

# Vary precision (rounding) of the table 
oldPrecision = as.numeric(options('digits')) 
options('digits' = 1)
mxCompare(fit1, fit2)
options('digits' = oldPrecision)

[Package OpenMx version 2.21.11 Index]