Exploratory only: not recommended for reporting measures.
TARGET=Y down-weights off-target observations. This lessens the effect of guessing on the measure estimates, but can reduce reliabilities and increase reported misfit. A big discrepancy between the measures produced by TARGET=N and TARGET=Y indicates much anomalous behavior disturbing the measurement process.
Unwanted behavior (e.g. guessing, carelessness) can cause unexpected responses to off-target items. The effect of responses on off-target items is lessened by specifying TARGET=Y. This weights each response by its statistical information during estimation. Fit statistics are calculated as though the estimates were made in the usual manner. Reported displacements show how much difference targeting has made in the estimates.
Example: Some low achievers have guessed wildly on a MCQ test. You want to reduce the effect of their lucky guesses on their measures and on item calibrations.
TARGET=Y
How Targeting works:
a) for each observation:
calculate probability of each category (0,1 for dichotomies)
calculate expected score (= probability of 1 for dichotomy)
calculate variance = information
= probability of 1 * probability of 0 for dichotomies,
so maximum value is 0.25 when person ability measure = item difficulty measure
b) for targeting:
weighted observation = variance * observation
weighted expected score = variance * expected score
c) sum these across persons and items (and structures)
d) required "targeted" estimates are obtained when, for each person, item, structure sum (weighted observations) = sum (weighted expected scores)
e) for calculation of fit statistics and displacement, weights of 1 are used but with the targeted parameter estimates. Displacement size and excessive misfit indicate how much "off-target" aberrant behavior exists in the data.
For targeting, there are many patterns of responses that can cause infinite measures, e.g. all items correct except for the easiest one. The convergence criteria limit how extreme the reported measures will be.
Polytomous items: look at the Item Information function from the Graphs menu to see the weighting. With TARGET=Y, responses are weighted by their information.
