itemfit {mirt}

Item fit statistics

Description

Computes item-fit statistics for a variety of unidimensional and multidimensional models. Poorly fitting items should be inspected with the empirical plots/tables for unidimensional models, otherwise itemGAM can be used to diagnose where the functional form of the IRT model was misspecified, or models can be refit using more flexible semi-parametric response models (e.g., itemtype = 'spline'). If the latent trait density was approximated (e.g., Davidian curves, Empirical histograms, etc) then passing use_dentype_estimate = TRUE will use the internally saved quadrature and density components (where applicable). Currently, only S-X2 statistic supported for mixture IRT models. Finally, where applicable the root mean-square error of approximation (RMSEA) is reported to help gauge the magnitude of item misfit.

Usage

itemfit(
  x,
  fit_stats = "S_X2",
  which.items = 1:extract.mirt(x, "nitems"),
  na.rm = FALSE,
  p.adjust = "none",
  group.bins = 10,
  group.size = NA,
  group.fun = mean,
  mincell = 1,
  mincell.X2 = 2,
  return.tables = FALSE,
  pv_draws = 30,
  boot = 1000,
  boot_dfapprox = 200,
  S_X2.plot = NULL,
  S_X2.plot_raw.score = TRUE,
  ETrange = c(-2, 2),
  ETpoints = 11,
  empirical.plot = NULL,
  empirical.CI = 0.95,
  empirical.poly.collapse = FALSE,
  method = "EAP",
  Theta = NULL,
  par.strip.text = list(cex = 0.7),
  par.settings = list(strip.background = list(col = "#9ECAE1"), strip.border = list(col =
    "black")),
  auto.key = list(space = "right", points = FALSE, lines = TRUE),
  ...
)

Arguments

x

a computed model object of class SingleGroupClass, MultipleGroupClass, or DiscreteClass

fit_stats

a character vector indicating which fit statistics should be computed. Supported inputs are:

• 'S_X2' : Orlando and Thissen (2000, 2003) and Kang and Chen's (2007) signed chi-squared test (default)

• 'Zh' : Drasgow, Levine, & Williams (1985) Zh

• 'X2' : Bock's (1972) chi-squared method. The default inputs compute Yen's (1981) Q1 variant of the X2 statistic (i.e., uses a fixed group.bins = 10). However, Bock's group-size variable median-based method can be computed by passing group.fun = median and modifying the group.size input to the desired number of bins

• 'G2' : McKinley & Mills (1985) G2 statistic (similar method to Q1, but with the likelihood-ratio test).

• 'PV_Q1' : Chalmers and Ng's (2017) plausible-value variant of the Q1 statistic.

• 'PV_Q1*' : Chalmers and Ng's (2017) plausible-value variant of the Q1 statistic that uses parametric bootstrapping to obtain a suitable empirical distribution.

• 'X2*' : Stone's (2000) fit statistics that require parametric bootstrapping

• 'X2*_df' : Stone's (2000) fit statistics that require parametric bootstrapping to obtain scaled versions of the X2* and degrees of freedom

• 'infit' : Compute the infit and outfit statistics

Note that 'S_X2' and 'Zh' cannot be computed when there are missing response data (i.e., will require multiple-imputation/row-removal techniques).

which.items

an integer vector indicating which items to test for fit. Default tests all possible items

na.rm

logical; remove rows with any missing values? This is required for methods such as S-X2 because they require the "EAPsum" method from fscores

p.adjust

method to use for adjusting all p-values for each respective item fit statistic (see p.adjust for available options). Default is 'none'

group.bins

the number of bins to use for X2 and G2. For example, setting group.bins = 10 will will compute Yen's (1981) Q1 statistic when 'X2' is requested

group.size

approximate size of each group to be used in calculating the \chi^2 statistic. The default NA disables this command and instead uses the group.bins input to try and construct equally sized bins

group.fun

function used when 'X2' or 'G2' are computed. Determines the central tendency measure within each partitioned group. E.g., setting group.fun = median will obtain the median of each respective ability estimate in each subgroup (this is what was used by Bock, 1972)

mincell

the minimum expected cell size to be used in the S-X2 computations. Tables will be collapsed across items first if polytomous, and then across scores if necessary

mincell.X2

the minimum expected cell size to be used in the X2 computations. Tables will be collapsed if polytomous, however if this condition can not be met then the group block will be omitted in the computations

return.tables

logical; return tables when investigating 'X2', 'S_X2', and 'X2*'?

pv_draws

number of plausible-value draws to obtain for PV_Q1 and PV_Q1*

boot

number of parametric bootstrap samples to create for PV_Q1* and X2*

boot_dfapprox

number of parametric bootstrap samples to create for the X2*_df statistic to approximate the scaling factor for X2* as well as the scaled degrees of freedom estimates

S_X2.plot

argument input is the same as empirical.plot, however the resulting image is constructed according to the S-X2 statistic's conditional sum-score information

S_X2.plot_raw.score

logical; use the raw-score information in the plot in stead of the latent trait scale score? Default is FALSE

ETrange

range of integration nodes for Stone's X2* statistic

ETpoints

number of integration nodes to use for Stone's X2* statistic

empirical.plot

a single numeric value or character of the item name indicating which item to plot (via itemplot) and overlay with the empirical \theta groupings (see empirical.CI). Useful for plotting the expected bins based on the 'X2' or 'G2' method

empirical.CI

a numeric value indicating the width of the empirical confidence interval ranging between 0 and 1 (default of 0 plots not interval). For example, a 95 interval would be plotted when empirical.CI = .95. Only applicable to dichotomous items

empirical.poly.collapse

logical; collapse polytomous item categories to for expected scoring functions for empirical plots? Default is FALSE

method

type of factor score estimation method. See fscores for more detail

Theta

a matrix of factor scores for each person used for statistics that require empirical estimates. If supplied, arguments typically passed to fscores() will be ignored and these values will be used instead. Also required when estimating statistics with missing data via imputation

par.strip.text

plotting argument passed to lattice

par.settings

plotting argument passed to lattice

auto.key

plotting argument passed to lattice

...

additional arguments to be passed to fscores() and lattice

Author(s)

Phil Chalmers [e-mail]

References

Bock, R. D. (1972). Estimating item parameters and latent ability when responses are scored in two or more nominal categories. Psychometrika, 37, 29-51.

Chalmers, R., P. (2012). mirt: A Multidimensional Item Response Theory Package for the R Environment. Journal of Statistical Software, 48(6), 1-29. doi:10.18637/jss.v048.i06

Chalmers, R. P. & Ng, V. (2017). Plausible-Value Imputation Statistics for Detecting Item Misfit. Applied Psychological Measurement, 41, 372-387. doi:10.1177/0146621617692079

Drasgow, F., Levine, M. V., & Williams, E. A. (1985). Appropriateness measurement with polychotomous item response models and standardized indices. British Journal of Mathematical and Statistical Psychology, 38, 67-86.

Kang, T. & Chen, Troy, T. (2007). An investigation of the performance of the generalized S-X2 item-fit index for polytomous IRT models. ACT

McKinley, R., & Mills, C. (1985). A comparison of several goodness-of-fit statistics. Applied Psychological Measurement, 9, 49-57.

Orlando, M. & Thissen, D. (2000). Likelihood-based item fit indices for dichotomous item response theory models. Applied Psychological Measurement, 24, 50-64.

Reise, S. P. (1990). A comparison of item- and person-fit methods of assessing model-data fit in IRT. Applied Psychological Measurement, 14, 127-137.

Stone, C. A. (2000). Monte Carlo Based Null Distribution for an Alternative Goodness-of-Fit Test Statistics in IRT Models. Journal of Educational Measurement, 37, 58-75.

Wright B. D. & Masters, G. N. (1982). Rating scale analysis. MESA Press.

Yen, W. M. (1981). Using simulation results to choose a latent trait model. Applied Psychological Measurement, 5, 245-262.

See Also

personfit, itemGAM

Examples

## Not run: 

P <- function(Theta){exp(Theta^2 * 1.2 - 1) / (1 + exp(Theta^2 * 1.2 - 1))}

#make some data
set.seed(1234)
a <- matrix(rlnorm(20, meanlog=0, sdlog = .1),ncol=1)
d <- matrix(rnorm(20),ncol=1)
Theta <- matrix(rnorm(2000))
items <- rep('2PL', 20)
ps <- P(Theta)
baditem <- numeric(2000)
for(i in 1:2000)
   baditem[i] <- sample(c(0,1), 1, prob = c(1-ps[i], ps[i]))
data <- cbind(simdata(a,d, 2000, items, Theta=Theta), baditem=baditem)

x <- mirt(data, 1)
raschfit <- mirt(data, 1, itemtype='Rasch')
fit <- itemfit(x)
fit

# p-value adjustment
itemfit(x, p.adjust='fdr')

# two different fit stats (with/without p-value adjustment)
itemfit(x, c('S_X2' ,'X2'), p.adjust='fdr')
itemfit(x, c('S_X2' ,'X2'))

# Conditional sum-score plot from S-X2 information
itemfit(x, S_X2.plot = 1) # good fit
itemfit(x, S_X2.plot = 2) # good fit
itemfit(x, S_X2.plot = 21) # bad fit

itemfit(x, 'X2') # just X2
itemfit(x, 'X2', method = 'ML') # X2 with maximum-likelihood estimates for traits
itemfit(x, group.bins=15, empirical.plot = 1, method = 'ML') #empirical item plot with 15 points
itemfit(x, group.bins=15, empirical.plot = 21, method = 'ML')

# PV and X2* statistics (parametric bootstrap stats not run to save time)
itemfit(x, 'PV_Q1')

if(interactive()) mirtCluster() # improve speed of bootstrap samples by running in parallel
# itemfit(x, 'PV_Q1*')
# itemfit(x, 'X2*') # Stone's 1993 statistic
# itemfit(x, 'X2*_df') # Stone's 2000 scaled statistic with df estimate

# empirical tables for X2 statistic
tabs <- itemfit(x, 'X2', return.tables=TRUE, which.items = 1)
tabs

#infit/outfit statistics. method='ML' agrees better with eRm package
itemfit(raschfit, 'infit', method = 'ML') #infit and outfit stats

#same as above, but inputting ML estimates instead (saves time for re-use)
Theta <- fscores(raschfit, method = 'ML')
itemfit(raschfit, 'infit', Theta=Theta)
itemfit(raschfit, empirical.plot=1, Theta=Theta)
itemfit(raschfit, 'X2', return.tables=TRUE, Theta=Theta, which.items=1)

# fit a new more flexible model for the mis-fitting item
itemtype <- c(rep('2PL', 20), 'spline')
x2 <- mirt(data, 1, itemtype=itemtype)
itemfit(x2)
itemplot(x2, 21)
anova(x, x2)

#------------------------------------------------------------

#similar example to Kang and Chen 2007
a <- matrix(c(.8,.4,.7, .8, .4, .7, 1, 1, 1, 1))
d <- matrix(rep(c(2.0,0.0,-1,-1.5),10), ncol=4, byrow=TRUE)
dat <- simdata(a,d,2000, itemtype = rep('graded', 10))
head(dat)

mod <- mirt(dat, 1)
itemfit(mod)
itemfit(mod, 'X2') # less useful given inflated Type I error rates
itemfit(mod, empirical.plot = 1)
itemfit(mod, empirical.plot = 1, empirical.poly.collapse=TRUE)

# collapsed tables (see mincell.X2) for X2 and G2
itemfit(mod, 'X2', return.tables = TRUE, which.items = 1)

mod2 <- mirt(dat, 1, 'Rasch')
itemfit(mod2, 'infit', method = 'ML')

# massive list of tables for S-X2
tables <- itemfit(mod, return.tables = TRUE)

#observed and expected total score patterns for item 1 (post collapsing)
tables$O[[1]]
tables$E[[1]]

# can also select specific items
# itemfit(mod, return.tables = TRUE, which.items=1)

# fit stats with missing data (run in parallel using all cores)
dat[sample(1:prod(dim(dat)), 100)] <- NA
raschfit <- mirt(dat, 1, itemtype='Rasch')

# use only valid data by removing rows with missing terms
itemfit(raschfit, c('S_X2', 'infit'), na.rm = TRUE)

# note that X2, G2, PV-Q1, and X2* do not require complete datasets
thetas <- fscores(raschfit, method = 'ML') # save for faster computations
itemfit(raschfit, c('X2', 'G2'), Theta=thetas)
itemfit(raschfit, empirical.plot=1, Theta=thetas)
itemfit(raschfit, 'X2', return.tables=TRUE, which.items=1, Theta=thetas)


## End(Not run)