testMGD {cSEM}

Tests for multi-group comparisons

Description

Usage

testMGD(
 .object                = NULL,
 .alpha                 = 0.05,
 .approach_p_adjust     = "none",
 .approach_mgd          = c("all", "Klesel", "Chin", "Sarstedt", 
                            "Keil", "Nitzl", "Henseler", "CI_para","CI_overlap"),
 .output_type           = c("complete", "structured"),
 .parameters_to_compare = NULL,
 .eval_plan             = c("sequential", "multicore", "multisession"),                           
 .handle_inadmissibles  = c("replace", "drop", "ignore"),
 .R_permutation         = 499,
 .R_bootstrap           = 499,
 .saturated             = FALSE,
 .seed                  = NULL,
 .type_ci               = "CI_percentile",
 .type_vcv              = c("indicator", "construct"),
 .verbose               = TRUE
 )

Arguments

.object

An R object of class cSEMResults resulting from a call to csem().

.alpha

An integer or a numeric vector of significance levels. Defaults to 0.05.

.approach_p_adjust

Character string or a vector of character strings. Approach used to adjust the p-value for multiple testing. See the methods argument of stats::p.adjust() for a list of choices and their description. Defaults to "none".

.approach_mgd

Character string or a vector of character strings. Approach used for the multi-group comparison. One of: "all", "Klesel", "Chin", "Sarstedt", "Keil, "Nitzl", "Henseler", "CI_para", or "CI_overlap". Default to "all" in which case all approaches are computed (if possible).

.output_type

Character string. The type of output to return. One of "complete" or "structured". See the Value section for details. Defaults to "complete".

.parameters_to_compare

A model in lavaan model syntax indicating which parameters (i.e, path (~), loadings (⁠=~⁠), weights (⁠<~⁠), or correlations (⁠~~⁠)) should be compared across groups. Defaults to NULL in which case all weights, loadings and path coefficients of the originally specified model are compared.

.eval_plan

Character string. The evaluation plan to use. One of "sequential", "multicore", or "multisession". In the two latter cases all available cores will be used. Defaults to "sequential".

.handle_inadmissibles

Character string. How should inadmissible results be treated? One of "drop", "ignore", or "replace". If "drop", all replications/resamples yielding an inadmissible result will be dropped (i.e. the number of results returned will potentially be less than .R). For "ignore" all results are returned even if all or some of the replications yielded inadmissible results (i.e. number of results returned is equal to .R). For "replace" resampling continues until there are exactly .R admissible solutions. Defaults to "replace" to accommodate all approaches.

.R_permutation

Integer. The number of permutations. Defaults to 499

.R_bootstrap

Integer. The number of bootstrap runs. Ignored if .object contains resamples. Defaults to 499

.saturated

Logical. Should a saturated structural model be used? Defaults to FALSE.

.seed

Integer or NULL. The random seed to use. Defaults to NULL in which case an arbitrary seed is chosen. Note that the scope of the seed is limited to the body of the function it is used in. Hence, the global seed will not be altered!

.type_ci

Character string. Which confidence interval should be calculated? For possible choices, see the .quantity argument of the infer() function. Only used if .approch_mgd is one of "CI_para" or "CI_overlap". Ignored otherwise. Defaults to "CI_percentile".

.type_vcv

Character string. Which model-implied correlation matrix should be calculated? One of "indicator" or "construct". Defaults to "indicator".

.verbose

Logical. Should information (e.g., progress bar) be printed to the console? Defaults to TRUE.

Details

This function performs various tests proposed in the context of multigroup analysis.

The following tests are implemented:

.approach_mgd = "Klesel": Approach suggested by Klesel et al. (2019)

The model-implied variance-covariance matrix (either indicator (.type_vcv = "indicator") or construct (.type_vcv = "construct")) is compared across groups. If the model-implied indicator or construct correlation matrix based on a saturated structural model should be compared, set .saturated = TRUE. To measure the distance between the model-implied variance-covariance matrices, the geodesic distance (dG) and the squared Euclidean distance (dL) are used. If more than two groups are compared, the average distance over all groups is used.

.approach_mgd = "Sarstedt": Approach suggested by Sarstedt et al. (2011)

Groups are compared in terms of parameter differences across groups. Sarstedt et al. (2011) tests if parameter k is equal across all groups. If several parameters are tested simultaneously it is recommended to adjust the significance level or the p-values (in cSEM correction is done by p-value). By default no multiple testing correction is done, however, several common adjustments are available via .approach_p_adjust. See stats::p.adjust() for details. Note: the test has some severe shortcomings. Use with caution.

.approach_mgd = "Chin": Approach suggested by Chin and Dibbern (2010)

Groups are compared in terms of parameter differences across groups. Chin and Dibbern (2010) tests if parameter k is equal between two groups. If more than two groups are tested for equality, parameter k is compared between all pairs of groups. In this case, it is recommended to adjust the significance level or the p-values (in cSEM correction is done by p-value) since this is essentially a multiple testing setup. If several parameters are tested simultaneously, correction is by group and number of parameters. By default no multiple testing correction is done, however, several common adjustments are available via .approach_p_adjust. See stats::p.adjust() for details.

.approach_mgd = "Keil": Approach suggested by Keil et al. (2000)

Groups are compared in terms of parameter differences across groups. Keil et al. (2000) tests if parameter k is equal between two groups. It is assumed, that the standard errors of the coefficients are equal across groups. The calculation of the standard error of the parameter difference is adjusted as proposed by Henseler et al. (2009). If more than two groups are tested for equality, parameter k is compared between all pairs of groups. In this case, it is recommended to adjust the significance level or the p-values (in cSEM correction is done by p-value) since this is essentially a multiple testing setup. If several parameters are tested simultaneously, correction is by group and number of parameters. By default no multiple testing correction is done, however, several common adjustments are available via .approach_p_adjust. See stats::p.adjust() for details.

.approach_mgd = "Nitzl": Approach suggested by Nitzl (2010)

Groups are compared in terms of parameter differences across groups. Similarly to Keil et al. (2000), a single parameter k is tested for equality between two groups. In contrast to Keil et al. (2000), it is assumed, that the standard errors of the coefficients are unequal across groups (Sarstedt et al. 2011). If more than two groups are tested for equality, parameter k is compared between all pairs of groups. In this case, it is recommended to adjust the significance level or the p-values (in cSEM correction is done by p-value) since this is essentially a multiple testing setup. If several parameters are tested simultaneously, correction is by group and number of parameters. By default no multiple testing correction is done, however, several common adjustments are available via .approach_p_adjust. See stats::p.adjust() for details.

.approach_mgd = "Henseler": Approach suggested by Henseler (2007)

Groups are compared in terms of parameter differences across groups. In doing so, the bootstrap estimates of one parameter are compared across groups. In the literature, this approach is also known as PLS-MGA. Originally, this test was proposed as an one-sided test. In this function we perform a left-sided and a right-sided test to investigate whether a parameter differs across two groups. In doing so, the significance level is divided by 2 and compared to p-value of the left and right-sided test. Moreover, .approach_p_adjust is ignored and no overall decision is returned. For a more detailed description, see also Henseler et al. (2009).

.approach_mgd = "CI_param": Approach mentioned in Sarstedt et al. (2011)

This approach is based on the confidence intervals constructed around the parameter estimates of the two groups. If the parameter of one group falls within the confidence interval of the other group and/or vice versa, it can be concluded that there is no group difference. Since it is based on the confidence intervals .approach_p_adjust is ignored.

.approach_mgd = "CI_overlap": Approach mentioned in Sarstedt et al. (2011)

This approach is based on the confidence intervals (CIs) constructed around the parameter estimates of the two groups. If the two CIs overlap, it can be concluded that there is no group difference. Since it is based on the confidence intervals .approach_p_adjust is ignored.

Use .approach_mgd to choose the approach. By default all approaches are computed (.approach_mgd = "all").

For convenience, two types of output are available. See the "Value" section below.

By default, approaches based on parameter differences across groups compare all parameters (.parameters_to_compare = NULL). To compare only a subset of parameters provide the parameters in lavaan model syntax just like the model to estimate. Take the simple model:

model_to_estimate <- "
Structural model
eta2 ~ eta1
eta3 ~ eta1 + eta2

# Each concept os measured by 3 indicators, i.e., modeled as latent variable
eta1 =~ y11 + y12 + y13
eta2 =~ y21 + y22 + y23
eta3 =~ y31 + y32 + y33
"

If only the path from eta1 to eta3 and the loadings of eta1 are to be compared across groups, write:

to_compare <- "
Structural parameters to compare
eta3 ~ eta1

# Loadings to compare
eta1 =~ y11 + y12 + y13
"

Note that the "model" provided to .parameters_to_compare does not need to be an estimable model!

Note also that compared to all other functions in cSEM using the argument, .handle_inadmissibles defaults to "replace" to accommodate the Sarstedt et al. (2011) approach.

Argument .R_permuation is ignored for the "Nitzl" and the "Keil" approach. .R_bootstrap is ignored if .object already contains resamples, i.e. has class cSEMResults_resampled and if only the "Klesel" or the "Chin" approach are used.

The argument .saturated is used by "Klesel" only. If .saturated = TRUE the original structural model is ignored and replaced by a saturated model, i.e. a model in which all constructs are allowed to correlate freely. This is useful to test differences in the measurement models between groups in isolation.

Value

If .output_type = "complete" a list of class cSEMTestMGD. Technically, cSEMTestMGD is a named list containing the following list elements:

⁠$Information⁠

Additional information.

⁠$Klesel⁠

A list with elements, Test_statistic, P_value, and Decision

⁠$Chin⁠

A list with elements, Test_statistic, P_value, Decision, and Decision_overall

⁠$Sarstedt⁠

A list with elements, Test_statistic, P_value, Decision, and Decision_overall

⁠$Keil⁠

A list with elements, Test_statistic, P_value, Decision, and Decision_overall

⁠$Nitzl⁠

A list with elements, Test_statistic, P_value, Decision, and Decision_overall

⁠$Henseler⁠

A list with elements, Test_statistic, P_value, Decision, and Decision_overall

⁠$CI_para⁠

A list with elements, Decision, and Decision_overall

⁠$CI_overlap⁠

A list with elements, Decision, and Decision_overall

If .output_type = "structured" a tibble (data frame) with the following columns is returned.

Test

The name of the test.

Comparision

The parameter that was compared across groups. If "overall" the overall fit of the model was compared.

⁠alpha%⁠

The test decision for a given "alpha" level. If TRUE the null hypotheses was rejected; if FALSE it was not rejected.

p-value_correction

The p-value correction.

CI_type

Only for the "CI_para" and the "CI_overlap" test. Which confidence interval was used.

Distance_metric

Only for Test = "Klesel". Which distance metric was used.

References

Chin WW, Dibbern J (2010). “An Introduction to a Permutation Based Procedure for Multi-Group PLS Analysis: Results of Tests of Differences on Simulated Data and a Cross Cultural Analysis of the Sourcing of Information System Services Between Germany and the USA.” In Handbook of Partial Least Squares, 171–193. Springer Berlin Heidelberg. doi:10.1007/978-3-540-32827-8_8.

Henseler J (2007). “A new and simple approach to multi-group analysis in partial least squares path modeling.” In Martens H, Næ s T (eds.), Proceedings of PLS'07 - The 5th International Symposium on PLS and Related Methods, 104–107. PLS, Norway: Matforsk, As.

Henseler J, Ringle CM, Sinkovics RR (2009). “The use of partial least squares path modeling in international marketing.” Advances in International Marketing, 20, 277–320. doi:10.1108/S1474-7979(2009)0000020014.

Keil M, Tan BC, Wei K, Saarinen T, Tuunainen V, Wassenaar A (2000). “A cross-cultural study on escalation of commitment behavior in software projects.” MIS Quarterly, 24(2), 299–325.

Klesel M, Schuberth F, Henseler J, Niehaves B (2019). “A Test for Multigroup Comparison Using Partial Least Squares Path Modeling.” Internet Research, 29(3), 464–477. doi:10.1108/intr-11-2017-0418.

Nitzl C (2010). “Eine anwenderorientierte Einfuehrung in die Partial Least Square (PLS)-Methode.” In Arbeitspapier, number 21. Universitaet Hamburg, Institut fuer Industrielles Management, Hamburg.

Sarstedt M, Henseler J, Ringle CM (2011). “Multigroup Analysis in Partial Least Squares (PLS) Path Modeling: Alternative Methods and Empirical Results.” In Advances in International Marketing, 195–218. Emerald Group Publishing Limited. doi:10.1108/s1474-7979(2011)0000022012.

See Also

csem(), cSEMResults, testMICOM(), testOMF()

Examples

## Not run: 
# ===========================================================================
# Basic usage
# ===========================================================================
model <- "
# Structural model
QUAL ~ EXPE
EXPE ~ IMAG
SAT  ~ IMAG + EXPE + QUAL + VAL
LOY  ~ IMAG + SAT
VAL  ~ EXPE + QUAL

# Measurement model

EXPE <~ expe1 + expe2 + expe3 + expe4 + expe5
IMAG <~ imag1 + imag2 + imag3 + imag4 + imag5
LOY  =~ loy1  + loy2  + loy3  + loy4
QUAL =~ qual1 + qual2 + qual3 + qual4 + qual5
SAT  <~ sat1  + sat2  + sat3  + sat4
VAL  <~ val1  + val2  + val3  + val4
"

## Create list of virtually identical data sets
dat <- list(satisfaction[-3,], satisfaction[-5, ], satisfaction[-10, ])
out <- csem(dat, model, .resample_method = "bootstrap", .R = 40) 

## Test 
testMGD(out, .R_permutation = 40,.verbose = FALSE)

# Notes: 
#  1. .R_permutation (and .R in the call to csem) is small to make examples run quicker; 
#     should be higher in real applications.
#  2. Test will not reject their respective H0s since the groups are virtually
#     identical.
#  3. Only exception is the approach suggested by Sarstedt et al. (2011), a
#     sign that the test is unreliable.
#  4. As opposed to other functions involving the argument, 
#     '.handle_inadmissibles' the default is "replace" as this is
#     required by Sarstedt et al. (2011)'s approach.

# ===========================================================================
# Extended usage
# ===========================================================================
### Test only a subset ------------------------------------------------------
# By default all parameters are compared. Select a subset by providing a 
# model in lavaan model syntax:

to_compare <- "
# Path coefficients
QUAL ~ EXPE

# Loadings
EXPE <~ expe1 + expe2 + expe3 + expe4 + expe5
"

## Test 
testMGD(out, .parameters_to_compare = to_compare, .R_permutation = 20, 
        .R_bootstrap = 20, .verbose = FALSE)

### Different p_adjustments --------------------------------------------------
# To adjust p-values to accommodate multiple testing use .approach_p_adjust. 
# The number of tests to use for adjusting depends on the approach chosen. For
# the Chin approach for example it is the number of parameters to test times the
# number of possible group comparisons. To compare the results for different
# adjustments, a vector of p-adjustments may be chosen.

## Test 
testMGD(out, .parameters_to_compare = to_compare, 
        .approach_p_adjust = c("none", "bonferroni"),
        .R_permutation = 20, .R_bootstrap = 20, .verbose = FALSE)

## End(Not run)