test_bf {performance}

Test if models are different

Description

Testing whether models are "different" in terms of accuracy or explanatory power is a delicate and often complex procedure, with many limitations and prerequisites. Moreover, many tests exist, each coming with its own interpretation, and set of strengths and weaknesses.

The test_performance() function runs the most relevant and appropriate tests based on the type of input (for instance, whether the models are nested or not). However, it still requires the user to understand what the tests are and what they do in order to prevent their misinterpretation. See the Details section for more information regarding the different tests and their interpretation.

Usage

test_bf(...)

## Default S3 method:
test_bf(..., reference = 1, text_length = NULL)

test_likelihoodratio(..., estimator = "ML", verbose = TRUE)

test_lrt(..., estimator = "ML", verbose = TRUE)

test_performance(..., reference = 1, verbose = TRUE)

test_vuong(..., verbose = TRUE)

test_wald(..., verbose = TRUE)

Arguments

...

Multiple model objects.

reference

This only applies when models are non-nested, and determines which model should be taken as a reference, against which all the other models are tested.

text_length

Numeric, length (number of chars) of output lines. test_bf() describes models by their formulas, which can lead to overly long lines in the output. text_length fixes the length of lines to a specified limit.

estimator

Applied when comparing regression models using test_likelihoodratio(). Corresponds to the different estimators for the standard deviation of the errors. Defaults to "OLS" for linear models, "ML" for all other models (including mixed models), or "REML" for linear mixed models when these have the same fixed effects. See 'Details'.

verbose

Toggle warning and messages.

Details

Nested vs. Non-nested Models

Model's "nesting" is an important concept of models comparison. Indeed, many tests only make sense when the models are "nested", i.e., when their predictors are nested. This means that all the fixed effects predictors of a model are contained within the fixed effects predictors of a larger model (sometimes referred to as the encompassing model). For instance, model1 (y ~ x1 + x2) is "nested" within model2 (y ~ x1 + x2 + x3). Usually, people have a list of nested models, for instance m1 (y ~ 1), m2 (y ~ x1), m3 (y ~ x1 + x2), m4 (y ~ x1 + x2 + x3), and it is conventional that they are "ordered" from the smallest to largest, but it is up to the user to reverse the order from largest to smallest. The test then shows whether a more parsimonious model, or whether adding a predictor, results in a significant difference in the model's performance. In this case, models are usually compared sequentially: m2 is tested against m1, m3 against m2, m4 against m3, etc.

Two models are considered as "non-nested" if their predictors are different. For instance, model1 (y ~ x1 + x2) and model2 (y ~ x3 + x4). In the case of non-nested models, all models are usually compared against the same reference model (by default, the first of the list).

Nesting is detected via the insight::is_nested_models() function. Note that, apart from the nesting, in order for the tests to be valid, other requirements have often to be the fulfilled. For instance, outcome variables (the response) must be the same. You cannot meaningfully test whether apples are significantly different from oranges!

Estimator of the standard deviation

The estimator is relevant when comparing regression models using test_likelihoodratio(). If estimator = "OLS", then it uses the same method as anova(..., test = "LRT") implemented in base R, i.e., scaling by n-k (the unbiased OLS estimator) and using this estimator under the alternative hypothesis. If estimator = "ML", which is for instance used by lrtest(...) in package lmtest, the scaling is done by n (the biased ML estimator) and the estimator under the null hypothesis. In moderately large samples, the differences should be negligible, but it is possible that OLS would perform slightly better in small samples with Gaussian errors. For estimator = "REML", the LRT is based on the REML-fit log-likelihoods of the models. Note that not all types of estimators are available for all model classes.

REML versus ML estimator

When estimator = "ML", which is the default for linear mixed models (unless they share the same fixed effects), values from information criteria (AIC, AICc) are based on the ML-estimator, while the default behaviour of AIC() may be different (in particular for linear mixed models from lme4, which sets REML = TRUE). This default in test_likelihoodratio() intentional, because comparing information criteria based on REML fits requires the same fixed effects for all models, which is often not the case. Thus, while anova.merMod() automatically refits all models to REML when performing a LRT, test_likelihoodratio() checks if a comparison based on REML fits is indeed valid, and if so, uses REML as default (else, ML is the default). Set the estimator argument explicitely to override the default behaviour.

Tests Description

• Bayes factor for Model Comparison - test_bf(): If all models were fit from the same data, the returned BF shows the Bayes Factor (see bayestestR::bayesfactor_models()) for each model against the reference model (which depends on whether the models are nested or not). Check out this vignette for more details.

• Wald's F-Test - test_wald(): The Wald test is a rough approximation of the Likelihood Ratio Test. However, it is more applicable than the LRT: you can often run a Wald test in situations where no other test can be run. Importantly, this test only makes statistical sense if the models are nested.

  Note: this test is also available in base R through the anova() function. It returns an F-value column as a statistic and its associated p-value.

• Likelihood Ratio Test (LRT) - test_likelihoodratio(): The LRT tests which model is a better (more likely) explanation of the data. Likelihood-Ratio-Test (LRT) gives usually somewhat close results (if not equivalent) to the Wald test and, similarly, only makes sense for nested models. However, maximum likelihood tests make stronger assumptions than method of moments tests like the F-test, and in turn are more efficient. Agresti (1990) suggests that you should use the LRT instead of the Wald test for small sample sizes (under or about 30) or if the parameters are large.

  Note: for regression models, this is similar to anova(..., test="LRT") (on models) or lmtest::lrtest(...), depending on the estimator argument. For lavaan models (SEM, CFA), the function calls lavaan::lavTestLRT().

  For models with transformed response variables (like log(x) or sqrt(x)), logLik() returns a wrong log-likelihood. However, test_likelihoodratio() calls insight::get_loglikelihood() with check_response=TRUE, which returns a corrected log-likelihood value for models with transformed response variables. Furthermore, since the LRT only accepts nested models (i.e. models that differ in their fixed effects), the computed log-likelihood is always based on the ML estimator, not on the REML fits.

• Vuong's Test - test_vuong(): Vuong's (1989) test can be used both for nested and non-nested models, and actually consists of two tests.

  • The Test of Distinguishability (the Omega2 column and its associated p-value) indicates whether or not the models can possibly be distinguished on the basis of the observed data. If its p-value is significant, it means the models are distinguishable.

  • The Robust Likelihood Test (the LR column and its associated p-value) indicates whether each model fits better than the reference model. If the models are nested, then the test works as a robust LRT. The code for this function is adapted from the nonnest2 package, and all credit go to their authors.

Value

A data frame containing the relevant indices.

References

• Vuong, Q. H. (1989). Likelihood ratio tests for model selection and non-nested hypotheses. Econometrica, 57, 307-333.

• Merkle, E. C., You, D., & Preacher, K. (2016). Testing non-nested structural equation models. Psychological Methods, 21, 151-163.

See Also

compare_performance() to compare the performance indices of many different models.

Examples

# Nested Models
# -------------
m1 <- lm(Sepal.Length ~ Petal.Width, data = iris)
m2 <- lm(Sepal.Length ~ Petal.Width + Species, data = iris)
m3 <- lm(Sepal.Length ~ Petal.Width * Species, data = iris)

test_performance(m1, m2, m3)

test_bf(m1, m2, m3)
test_wald(m1, m2, m3) # Equivalent to anova(m1, m2, m3)

# Equivalent to lmtest::lrtest(m1, m2, m3)
test_likelihoodratio(m1, m2, m3, estimator = "ML")

# Equivalent to anova(m1, m2, m3, test='LRT')
test_likelihoodratio(m1, m2, m3, estimator = "OLS")

if (require("CompQuadForm")) {
  test_vuong(m1, m2, m3) # nonnest2::vuongtest(m1, m2, nested=TRUE)

  # Non-nested Models
  # -----------------
  m1 <- lm(Sepal.Length ~ Petal.Width, data = iris)
  m2 <- lm(Sepal.Length ~ Petal.Length, data = iris)
  m3 <- lm(Sepal.Length ~ Species, data = iris)

  test_performance(m1, m2, m3)
  test_bf(m1, m2, m3)
  test_vuong(m1, m2, m3) # nonnest2::vuongtest(m1, m2)
}

# Tweak the output
# ----------------
test_performance(m1, m2, m3, include_formula = TRUE)


# SEM / CFA (lavaan objects)
# --------------------------
# Lavaan Models
if (require("lavaan")) {
  structure <- " visual  =~ x1 + x2 + x3
                 textual =~ x4 + x5 + x6
                 speed   =~ x7 + x8 + x9

                  visual ~~ textual + speed "
  m1 <- lavaan::cfa(structure, data = HolzingerSwineford1939)

  structure <- " visual  =~ x1 + x2 + x3
                 textual =~ x4 + x5 + x6
                 speed   =~ x7 + x8 + x9

                  visual ~~ 0 * textual + speed "
  m2 <- lavaan::cfa(structure, data = HolzingerSwineford1939)

  structure <- " visual  =~ x1 + x2 + x3
                 textual =~ x4 + x5 + x6
                 speed   =~ x7 + x8 + x9

                  visual ~~ 0 * textual + 0 * speed "
  m3 <- lavaan::cfa(structure, data = HolzingerSwineford1939)

  test_likelihoodratio(m1, m2, m3)

  # Different Model Types
  # ---------------------
  if (require("lme4") && require("mgcv")) {
    m1 <- lm(Sepal.Length ~ Petal.Length + Species, data = iris)
    m2 <- lmer(Sepal.Length ~ Petal.Length + (1 | Species), data = iris)
    m3 <- gam(Sepal.Length ~ s(Petal.Length, by = Species) + Species, data = iris)

    test_performance(m1, m2, m3)
  }
}