(P)artial (in)variance specification search
Mark Lai
2024-10-02
Source:vignettes/pinsearch.rmd
pinsearch.rmdThis package implements the sequential specification search method for identifying parameters violating measurement invariance across groups. The method is discussed in, for example, Yoon and Millsap (2007), and is a popular method for identifying a partial invariance model.
Workflow
A suggested workflow for using the main function
pinSearch() is
- Define the
lavaansyntax for the configural invariance model. - Conduct a
pinSearch()with specific test statistic and stopping criteria. The default is to use modification indices as in Yoon and Millsap (2007), but the score test is also supported. The function returns a final partial invariance model and a list of noninvariant parameters in the search order. - Obtain effect size statistics for the noninvariant items.
Example
Here we use the classic data set by Holzinger & Swineford, with the goal of searching for a partial strict invariance model.
library(pinsearch)
library(lavaan)
#> This is lavaan 0.6-19
#> lavaan is FREE software! Please report any bugs.
mod <-
" visual =~ x1 + x2 + x3
textual =~ x4 + x5 + x6
speed =~ x7 + x8 + x9 "
# Output the final partial invariance model, and the noninvariant items
ps1 <-
pinSearch(mod,
data = HolzingerSwineford1939,
group = "school", type = "residuals",
inv_test = "mod" # use modification indices for search (default)
)Here are the noninvariant parameters:
ps1$`Non-Invariant Items`
#> group lhs rhs type
#> 1 1 x3 intercepts
#> 2 1 x7 interceptsAnd the final partial invariance model
ps1$`Partial Invariance Fit` |>
summary()
#> lavaan 0.6-19 ended normally after 64 iterations
#>
#> Estimator ML
#> Optimization method NLMINB
#> Number of model parameters 66
#> Number of equality constraints 25
#>
#> Number of observations per group:
#> Pasteur 156
#> Grant-White 145
#>
#> Model Test User Model:
#>
#> Test statistic 147.260
#> Degrees of freedom 67
#> P-value (Chi-square) 0.000
#> Test statistic for each group:
#> Pasteur 79.438
#> Grant-White 67.823
#>
#> Parameter Estimates:
#>
#> Standard errors Standard
#> Information Expected
#> Information saturated (h1) model Structured
#>
#>
#> Group 1 [Pasteur]:
#>
#> Latent Variables:
#> Estimate Std.Err z-value P(>|z|)
#> visual =~
#> x1 (.p1.) 0.881 0.093 9.474 0.000
#> x2 (.p2.) 0.549 0.085 6.437 0.000
#> x3 (.p3.) 0.726 0.084 8.599 0.000
#> textual =~
#> x4 (.p4.) 0.945 0.069 13.652 0.000
#> x5 (.p5.) 1.064 0.077 13.832 0.000
#> x6 (.p6.) 0.882 0.065 13.554 0.000
#> speed =~
#> x7 (.p7.) 0.564 0.071 7.981 0.000
#> x8 (.p8.) 0.674 0.073 9.189 0.000
#> x9 (.p9.) 0.603 0.070 8.666 0.000
#>
#> Covariances:
#> Estimate Std.Err z-value P(>|z|)
#> visual ~~
#> textual 0.436 0.089 4.908 0.000
#> speed 0.332 0.110 3.012 0.003
#> textual ~~
#> speed 0.315 0.097 3.240 0.001
#>
#> Intercepts:
#> Estimate Std.Err z-value P(>|z|)
#> .x1 (.25.) 4.910 0.093 52.673 0.000
#> .x2 (.26.) 6.072 0.079 76.983 0.000
#> .x3 2.487 0.090 27.772 0.000
#> .x4 (.28.) 2.784 0.086 32.195 0.000
#> .x5 (.29.) 4.029 0.096 41.811 0.000
#> .x6 (.30.) 1.927 0.081 23.746 0.000
#> .x7 4.432 0.082 53.865 0.000
#> .x8 (.32.) 5.568 0.073 75.821 0.000
#> .x9 (.33.) 5.411 0.071 76.698 0.000
#>
#> Variances:
#> Estimate Std.Err z-value P(>|z|)
#> .x1 (.10.) 0.636 0.101 6.303 0.000
#> .x2 (.11.) 1.101 0.100 10.973 0.000
#> .x3 (.12.) 0.724 0.085 8.529 0.000
#> .x4 (.13.) 0.383 0.047 8.096 0.000
#> .x5 (.14.) 0.435 0.057 7.613 0.000
#> .x6 (.15.) 0.353 0.042 8.336 0.000
#> .x7 (.16.) 0.738 0.076 9.696 0.000
#> .x8 (.17.) 0.479 0.073 6.592 0.000
#> .x9 (.18.) 0.580 0.069 8.386 0.000
#> visual 1.000
#> textual 1.000
#> speed 1.000
#>
#>
#> Group 2 [Grant-White]:
#>
#> Latent Variables:
#> Estimate Std.Err z-value P(>|z|)
#> visual =~
#> x1 (.p1.) 0.881 0.093 9.474 0.000
#> x2 (.p2.) 0.549 0.085 6.437 0.000
#> x3 (.p3.) 0.726 0.084 8.599 0.000
#> textual =~
#> x4 (.p4.) 0.945 0.069 13.652 0.000
#> x5 (.p5.) 1.064 0.077 13.832 0.000
#> x6 (.p6.) 0.882 0.065 13.554 0.000
#> speed =~
#> x7 (.p7.) 0.564 0.071 7.981 0.000
#> x8 (.p8.) 0.674 0.073 9.189 0.000
#> x9 (.p9.) 0.603 0.070 8.666 0.000
#>
#> Covariances:
#> Estimate Std.Err z-value P(>|z|)
#> visual ~~
#> textual 0.506 0.123 4.125 0.000
#> speed 0.634 0.161 3.929 0.000
#> textual ~~
#> speed 0.418 0.134 3.115 0.002
#>
#> Intercepts:
#> Estimate Std.Err z-value P(>|z|)
#> .x1 (.25.) 4.910 0.093 52.673 0.000
#> .x2 (.26.) 6.072 0.079 76.983 0.000
#> .x3 (.27.) 1.951 0.114 17.044 0.000
#> .x4 (.28.) 2.784 0.086 32.195 0.000
#> .x5 (.29.) 4.029 0.096 41.811 0.000
#> .x6 (.30.) 1.927 0.081 23.746 0.000
#> .x7 (.31.) 3.992 0.099 40.135 0.000
#> .x8 (.32.) 5.568 0.073 75.821 0.000
#> .x9 (.33.) 5.411 0.071 76.698 0.000
#> visual 0.062 0.146 0.423 0.672
#> textual 0.608 0.129 4.722 0.000
#> speed -0.127 0.157 -0.807 0.420
#>
#> Variances:
#> Estimate Std.Err z-value P(>|z|)
#> .x1 (.10.) 0.636 0.101 6.303 0.000
#> .x2 (.11.) 1.101 0.100 10.973 0.000
#> .x3 (.12.) 0.724 0.085 8.529 0.000
#> .x4 (.13.) 0.383 0.047 8.096 0.000
#> .x5 (.14.) 0.435 0.057 7.613 0.000
#> .x6 (.15.) 0.353 0.042 8.336 0.000
#> .x7 (.16.) 0.738 0.076 9.696 0.000
#> .x8 (.17.) 0.479 0.073 6.592 0.000
#> .x9 (.18.) 0.580 0.069 8.386 0.000
#> visual 0.856 0.208 4.114 0.000
#> textual 0.980 0.183 5.369 0.000
#> speed 1.401 0.326 4.294 0.000Effect Size
The package supports the effect size statistics by Nye and Drasgow (2011) for noninvariant items, and the generalization to an statistic for more than two groups.
These can be obtained by setting the effect_size = TRUE
argument:
pinSearch(mod,
data = HolzingerSwineford1939,
group = "school", type = "residuals",
inv_test = "mod",
effect_size = TRUE
)
#> $`Partial Invariance Fit`
#> lavaan 0.6-19 ended normally after 64 iterations
#>
#> Estimator ML
#> Optimization method NLMINB
#> Number of model parameters 66
#> Number of equality constraints 25
#>
#> Number of observations per group:
#> Pasteur 156
#> Grant-White 145
#>
#> Model Test User Model:
#>
#> Test statistic 147.260
#> Degrees of freedom 67
#> P-value (Chi-square) 0.000
#> Test statistic for each group:
#> Pasteur 79.438
#> Grant-White 67.823
#>
#> $`Non-Invariant Items`
#> group lhs rhs type
#> 1 1 x3 intercepts
#> 2 1 x7 intercepts
#>
#> $effect_size
#> x3-visual x7-speed
#> dmacs 0.4865937 0.4161147Note that in the above output, the effect sizes are printed three
times for x3 and x7, which have non-invariant
intercepts. This is because there are three latent variables in the
model, and the function allows for the possibility of cross-loadings.
When each item is only loaded on one latent variable, user can ignore
the duplicated values (we will try to fix this in future releases).
Another possibility is to use the pin_effsize()
function, which also works for partial invariance model not obtained
from pinSearch().
# Fit the partial strict model by hand
pstrict_fit <- cfa(mod, data = HolzingerSwineford1939,
group = "school",
group.equal = c("loadings", "intercepts",
"residuals"),
group.partial = c("x3~1", "x7~1"))
# Effect sizes
pin_effsize(pstrict_fit)
#> x3-visual x7-speed
#> dmacs 0.4865945 0.4161159Cross-Loadings
The function also supports models with cross-loadings. For example,
mod2 <- "
visual =~ x1 + x2 + x3 + x9
textual =~ x4 + x5 + x6
speed =~ x7 + x8 + x9
"
ps2 <-
pinSearch(mod2,
data = HolzingerSwineford1939,
group = "school", type = "residuals",
inv_test = "mod",
effect_size = TRUE
)
ps2
#> $`Partial Invariance Fit`
#> lavaan 0.6-19 ended normally after 55 iterations
#>
#> Estimator ML
#> Optimization method NLMINB
#> Number of model parameters 68
#> Number of equality constraints 26
#>
#> Number of observations per group:
#> Pasteur 156
#> Grant-White 145
#>
#> Model Test User Model:
#>
#> Test statistic 113.828
#> Degrees of freedom 66
#> P-value (Chi-square) 0.000
#> Test statistic for each group:
#> Pasteur 67.913
#> Grant-White 45.915
#>
#> $`Non-Invariant Items`
#> group lhs rhs type
#> 1 1 x3 intercepts
#> 2 1 x7 intercepts
#>
#> $effect_size
#> x3-visual x7-speed
#> dmacs 0.4721376 0.4069837It just happened in this examplle that the final results are not affected.
Score Test
One can also use the score test instead of the modification indices,
as starting
in version 0.6-17, lavaan recommends using that for
equality constraints. This can be achieved with
inv_test = "score".
pinSearch(mod2,
data = HolzingerSwineford1939,
group = "school", type = "residuals",
inv_test = "score",
effect_size = TRUE
)
#> $`Partial Invariance Fit`
#> lavaan 0.6-19 ended normally after 59 iterations
#>
#> Estimator ML
#> Optimization method NLMINB
#> Number of model parameters 68
#> Number of equality constraints 25
#>
#> Number of observations per group:
#> Pasteur 156
#> Grant-White 145
#>
#> Model Test User Model:
#>
#> Test statistic 109.770
#> Degrees of freedom 65
#> P-value (Chi-square) 0.000
#> Test statistic for each group:
#> Pasteur 66.273
#> Grant-White 43.497
#>
#> $`Non-Invariant Items`
#> group lhs rhs type
#> 1 2 x3 intercepts
#> 2 2 x7 intercepts
#> 3 2 x7 x7 residuals
#>
#> $effect_size
#> x3-visual x7-speed
#> dmacs 0.4717618 0.4071113The test is slightly more sensitive and identifies an additional
non-invariant uniqueness for item x7.
Likelihood Ratio Test (LRT)
LRT is generally more accurate, but more computationally expensive.
pinSearch(mod2,
data = HolzingerSwineford1939,
group = "school", type = "residuals",
inv_test = "lrt",
effect_size = TRUE
)
#> $`Partial Invariance Fit`
#> lavaan 0.6-19 ended normally after 59 iterations
#>
#> Estimator ML
#> Optimization method NLMINB
#> Number of model parameters 68
#> Number of equality constraints 25
#>
#> Number of observations per group:
#> Pasteur 156
#> Grant-White 145
#>
#> Model Test User Model:
#>
#> Test statistic 109.770
#> Degrees of freedom 65
#> P-value (Chi-square) 0.000
#> Test statistic for each group:
#> Pasteur 66.273
#> Grant-White 43.497
#>
#> $`Non-Invariant Items`
#> group lhs rhs type
#> 1 1 x3 intercepts
#> 2 1 x7 intercepts
#> 3 1 x7 x7 residuals
#>
#> $effect_size
#> x3-visual x7-speed
#> dmacs 0.4717618 0.4071113In this example, the results using LRT are the same as the score test.
Controlling for False Discovery Rate
The function also supports adjusting for multiplicity by using a more
stringent threshold for flagging a parameter as noninvariant, using the
multistage false discovery rate procedure by Benjamini
& Gavrilov (2009) for forward selection. It is up for debate
whether controlling for false discovery rate is as important in
measurement/factorial invariance analyses as in other context, and it
probably depends on whether one is interested in identifying
non-invariant items or in inferences for the latent parameters. If this
is of interest, one can use the control_fdr = TRUE
option.
pinSearch(mod2,
data = HolzingerSwineford1939,
group = "school", type = "residuals",
inv_test = "score",
effect_size = TRUE,
control_fdr = TRUE
)
#> $`Partial Invariance Fit`
#> lavaan 0.6-19 ended normally after 55 iterations
#>
#> Estimator ML
#> Optimization method NLMINB
#> Number of model parameters 68
#> Number of equality constraints 26
#>
#> Number of observations per group:
#> Pasteur 156
#> Grant-White 145
#>
#> Model Test User Model:
#>
#> Test statistic 113.828
#> Degrees of freedom 66
#> P-value (Chi-square) 0.000
#> Test statistic for each group:
#> Pasteur 67.913
#> Grant-White 45.915
#>
#> $`Non-Invariant Items`
#> group lhs rhs type
#> 1 2 x3 intercepts
#> 2 2 x7 intercepts
#>
#> $effect_size
#> x3-visual x7-speed
#> dmacs 0.4721376 0.4069837The uniqueness for x7 is no longer flagged.