Which statistic is more informative depends on what null hypothesis you are concerned about:
Null hypothesis: "the data fit the model (perfectly)" - use the t-test significance = Winsteps Zstd
Null hypothesis: "the data fit the model (usefully)" - use the chi-square divided by degrees of freedom = mean-square.
"The first of the distributions characteristic of modern tests of significance, though originating with F.R. Helmert [1875], was rediscovered by Karl Pearson in 1900, for the measure of discrepancy between observation and hypothesis, known as χ2 [chi-square].. ... It supplies an exact and objective measure of the joint discrepancy from their expectations of a number of normally distributed ... variates" (R. A. Fisher, Statistical Methods for Research Workers.)
It is the χ2 distribution which underlies many Rasch-model fit statistics. Even those based on the likelihood of the data capitalize on the fact that -2 log ( likelihood ) is asymptotically χ2.
A χ2 statistic with k degrees of freedom, d.f., is the sum of the squares of k random unit-normal deviates. Therefore its expected value is k, and its model variance is 2k. This provides the convenient feature that the expected value of a mean-square statistic, i.e., a χ2 statistic divided by its d.f. is 1.0. But the model variance of a mean-square statistic is 2/k. Thus, as the number of degrees of freedom, i.e., the sample size, increases, the power to detect small divergences increases, and ever smaller departures of the mean-square from 1.0 become statistically "significant", i.e., surprising, if the data are indeed as modeled.
For terminology, etc., see www.rasch.org/rmt/rmt162f.htm
The relationship between the size and significance of mean-square statistics is shown in the Figure. The statistical significance is expressed as the value of the corresponding value on a unit normal distribution. For 2-sided t-tests, 1.96 corresponds to p=.05. For dichotomous responses, d.f. is a little less than sample size (for an item) or test length (for a person). For polytomous responses, d.f. is somewhat less than (sample size or test length)*(polytomous categories - 1).
Test of Perfect Fit
The null hypothesis for a significance test of "perfect" fit of these data would be "Mean-square=1.0". Since the Rasch model is a mathematical ideal, like a Pythagorean triangle, we never expect to encounter empirical data that match it exactly. So this is an instance in which we know, a priori, that the null hypothesis cannot be accepted.
A mean-square of 1.2 means 1 unit of modeled information and .2 of unmodeled noise. The plot indicates that items with as little misfit as this would be flagged as significantly misfitting if observed in samples of over 200 persons. On the other hand, grossly noisy items, with more unmodeled noise than modeled information, i.e., with mean-squares of 2.0 or more, are not flagged in samples of less than 10. Overall, useful sample sizes for standardized fit statistics appear to be in the range 50-250 data points for the "perfect fit" null hypothesis.
Indication of Useful Fit
An indicator of "useful" fit could be "mean-square = 1.5 or less" (e.g., RMT 14:2, p. 743). Then, as the sample size (d.f.) increases, especially beyond 30, there is increasing certainty as to whether these data are productive for measurement (mean-square ≤ 1.5) or unproductive (mean-square > 1.5). This could be formulated as a one-sided t-test of the hypothesis that the mean-square is ≤1.5, with only values >1.5 being of concern.
John Michael Linacre
|
|
Size vs. Significance: Standardized Chi-Square Fit Statistic. J.M. Linacre … Rasch Measurement Transactions, 2003, 17:1, 918.
The URL of this page is www.rasch.org/rmt/rmt171n.htm
Website: www.rasch.org/rmt/contents.htm