intermediate 9 min read
Social sciences · Topic
Rasch Model
probability theory · optimization · information theory
The Rasch model is a one-parameter logistic model in which the log-odds of a correct response depends solely on the difference between person ability and item difficulty. Its elegant mathematical structure yields specific objectivity: person and item parameter estimates are separable, making comparisons independent of the particular sample or item set used.

The Rasch Equation

For person $n$ with ability $\theta_n$ on item $i$ with difficulty $b_i$, the log-odds (logit) of success is:

\[\log\frac{P_{ni}}{1 - P_{ni}} = \theta_n - b_i\]

Solving for the probability:

\[P_{ni} = \frac{e^{\theta_n - b_i}}{1 + e^{\theta_n - b_i}}\]

When $\theta_n = b_i$, $P_{ni} = 0.5$. Both $\theta$ and $b$ are measured in logits on an unbounded scale, typically centred by setting $\bar{b} = 0$.

Parameter Separation

A key property of the Rasch model is that the raw score $r_n = \sum_i u_{ni}$ is a sufficient statistic for $\theta_n$, and the item total $s_i = \sum_n u_{ni}$ is sufficient for $b_i$. This allows conditional maximum likelihood (CML) estimation of item difficulties free of person parameters.

Fit Statistics

Statistic Formula Expectation
Outfit MSQ $\sum_n z_{ni}^2 / N$ weighted by extreme persons $\approx 1.0$
Infit MSQ Variance-weighted residuals $\approx 1.0$

Values between 0.7 and 1.3 are generally acceptable. Outfit is sensitive to unexpected responses far from the item difficulty; infit is more sensitive to unexpected responses near the item difficulty.