9.1 Attribute Profile Estimation
Attribute Profile Estimation in Cognitive Diagnostic Modeling (CDM) refers to the scientific process of estimating an individual’s skills/attributes across a set of predefined attributes or skills based on their performance on assessment items. The attribute estimation process aims to provide mastery of skills/attributes in specific cognitive areas, which can inform tailored instructional strategies and interventions.
Recall: CDM is often mentioned as assessment for learning
To estimate attributes for each individual test-taker, we can employ different estimation approaches.
Kreitchmann et al. (2023) summarized the attribute profile estimation in several steps:
For Maximum Likelihood Estimation:
We first obtain the marginal maximum likelihood of the model parameters i.e., \(\hat{\delta}\) and \(\hat{P(\alpha_l)}\).
We obtain the MLE of \(\hat{\alpha}\) for respondent \(i\) will be of class \(l\) for which likilihood for \((Y_i|\alpha_l, \delta)\) is maximum, assumming that MMLE estimates of \(\delta\) is correct.
For Bayesian Estimation:
In Bayesian approach, the estimated classification of attributes for each test-taker is estimated based on the expected or maximum posterior probability of each attribute profile known as expected-a-posterior(EAP) or maximum-a-posterior (MAP).
Expected-A-Posterior (EAP) represents the expected value of the posterior distribution for an attribute profile.
Maximum-A-Posterior (MAP) refers to the profile with the highest posterior probability for each examinee.
The posterior probability of attribute profile αl for examinee i is numerically approximated as
\[P(\alpha_l \mid \mathbf{Y}_i, \delta) =\frac{\operatorname{lik}\left(\mathbf{Y}_i \mid \alpha_l, \delta\right) P\left(\alpha_l\right)}{\sum_{l=1}^L \operatorname{lik}\left(\mathbf{Y}_i \mid \alpha_l, \delta\right) P\left(\alpha_l\right)}\] - the MAP estimator classifies each examinee to its most probable attribute profile.
- The EAP estimator calculates the marginal probability for each attribute individually, then assigns each attribute to either mastery or non-mastery by comparing the probability against a set threshold (e.g., 0.5).