# DINA - Deterministic Inputs, Noisy "And" Gate
## Model
Latent response vector from the students, given the [[q-matrix|Q-Matrix]]:
$
\eta_{ij} = \prod_{k=1}^K \alpha{jk}^{q_{jk}}
$
However, the process is inherently stochastic, due to _slip_
($s_j = P(X_{ij} = 0 \mid \eta_{ij} = 1)$) and _guessing_ parameter
($g_j = P(X_{ij} = 1 \mid \eta_{ij} = 0$) -- students with "mastery" can slip
and miss the answer, students with "non-mastery" can guess the answer. Taken
these into consideration, the probability of examinee getting item $j$ right:
$
P_j(\mathbf\alpha_i) = P(X_{ij} = 1 \mid \mathbf\alpha_i) =
g_j^{1 - \eta_{ij}} (1 - s_j)^{\eta_{ij}}
$
Note that solving a question with skills not specified in Q-vector may look like
"guessing".
- DINA model is _parsimonious_ and _interpretable_, but provides good model fit.
- DINA model deals with **multidimensional binary latent skills**, while
traditional IRMs deal with **unidimensional continuous latent traits**.
- DINA model is a conditional distribution of $X_{ij}$ given a skills vector
$\mathbf\alpha_{i}$
> [!NOTE]
>
> - $q$ - the skills needed
> - $\mathbf\alpha$ - the skills mastered
> - $\eta$ - theoretically, disregarding the noises, are you supposed to answer
> it correctly?
> - $g$ - guessed it?
> - $s$ - slipped it?
> - $P$ - the probability of answering it correctly
> - $X$ - our prediction of answering correctness (think of it as an event)
## Estimation