Abstract
Diagnostic classification models (DCMs) are important statistical tools in cognitive diagnosis. In this paper, we consider the issue of their identifiability. In particular, we focus on one basic and popular model, the DINA model. We propose sufficient and necessary conditions under which the model parameters are identifiable from the data. The consequences, in terms of the consistency of parameter estimates, of fulfilling or failing to fulfill these conditions are illustrated via simulation. The results can be easily extended to the DINO model through the duality of the DINA and DINO models. Moreover, the proposed theoretical framework could be applied to study the identifiability issue of other DCMs.
Similar content being viewed by others
References
Bechger, T. M., Verstralen, H. H., & Verhelst, N. D. (2002). Equivalent linear logistic test models. Psychometrika, 67(1), 123–136.
Chen, Y., Liu, J., Xu, G., & Ying, Z. (2014). Statistical analysis of q-matrix based diagnostic classification models. Journal of the American Statistical Association (to appear).
Chiu, C.-Y. (2013). Statistical refinement of the Q-matrix in cognitive diagnosis. Applied Psychological Measurement, 37, 598–618.
Chiu, C.-Y., Douglas, J. A., & Li, X. (2009). Cluster analysis for cognitive diagnosis: Theory and applications. Psychometrika, 74, 633–665.
de la Torre, J. (2008). An empirically-based method of Q-matrix validation for the DINA model: Development and applications. Journal of Educational Measurement, 45, 343–362.
de la Torre, J. (2009). DINA model and parameter estimation: A didactic. Journal of Educational and Behavioral Statistics, 34, 115–130.
de la Torre, J. (2011). The generalized DINA model framework. Psychometrika, 76, 179–199.
de la Torre, J., & Douglas, J. A. (2004). Higher order latent trait models for cognitive diagnosis. Psychometrika, 69, 333–353.
DeCarlo, L. T. (2011). On the analysis of fraction subtraction data: The DINA model, classification, class sizes, and the Q-matrix. Applied Psychological Measurement, 35, 8–26.
DeCarlo, L. T. (2012). Recognizing uncertainty in the Q-matrix via a bayesian extension of the DINA model. Applied Psychological Measurement, 36(6), 447–468.
DiBello, L. V., Stout, W. F., & Roussos, L. A. (1995). Unified cognitive psychometric diagnostic assessment likelihood-based classification techniques. In P. D. Nichols, S. F. Chipman, & R. L. Brennan (Eds.), Cognitively diagnostic assessment (pp. 361–390). Hillsdale, NJ: Erlbaum Associates.
Gabrielsen, A. (1978). Consistency and identifiability. Journal of Econometrics, 8(2), 261–263.
Goodman, L. A. (1974). Exploratory latent structure analysis using both identifiable and unidentifiable models. Biometrika, 61, 215–231.
Henson, R. A., Templin, J. L., & Willse, J. T. (2009). Defining a family of cognitive diagnosis models using log-linear models with latent variables. Psychometrika, 74, 191–210.
Junker, B. W., & Sijtsma, K. (2001). Cognitive assessment models with few assumptions, and connections with nonparametric item response theory. Applied Psychological Measurement, 25, 258–272.
Koopmans, T. C. (Ed.). (1950). Statistical inference in dynamic economic models (Vol. 10). New York: Wiley.
Koopmans, T. C., & Reiersøl, O. (1950). The identification of structural characteristics. Annals of Mathematical Statistics, 21, 165–181.
Leighton, J. P., Gierl, M. J., & Hunka, S. M. (2004). The attribute hierarchy model for cognitive assessment: A variation on Tatsuoka’s rule-space approach. Journal of Educational Measurement, 41, 205–237.
Liu, J., Xu, G., & Ying, Z. (2012). Data-driven learning of Q-matrix. Applied Psychological Measurement, 36, 548–564.
Liu, J., Xu, G., & Ying, Z. (2013). Theory of self-learning Q-matrix. Bernoulli, 19(5A), 1790–1817.
Maris, G., & Bechger, T. M. (2004). Equivalent MIRID models. Psychometrika, 69(4), 627–639.
Maris, G., & Bechger, T. M. (2009). Equivalent diagnostic classification models. Measurement, 7, 41–46.
McHugh, R. B. (1956). Efficient estimatio and local identification in latent class analysis. Psychometrika, 21, 331–347.
Paulino, C. D. M., & de Bragançca Pereira, C. A. (1994). On identifiability of parametric statistical models. Journal of the Italian Statistical Society, 3(1), 125–151.
Rupp, A. A., & Templin, J. L. (2008). Effects of Q-matrix misspecification on parameter estimates and misclassification rates in the DINA model. Educational and Psychological Measurement, 68, 78–98.
Rupp, A. A., Templin, J. L., & Henson, R. A. (2010). Diagnostic measurement: Theory, methods, and applications. New York: Guilford Press.
San Martín, E., & Quintana, F. (2002). Consistency and identifiability revisited. Brazilian Journal of Probability and Statistics, 16, 99–106.
San Martín, E., Rolin, J.-M., & Castro, L. M. (2013). Identification of the 1PL model with guessing parameter: Parametric and semi-parametric results. Psychometrika, 78(2), 341–379.
Tatsuoka, C. (2009). Diagnostic models as partially ordered sets. Measurement, 7, 49–53.
Tatsuoka, K. K. (1983). Rule space: An approch for dealing with misconceptions based on item response theory. Journal of Educational Measurement, 20, 345–354.
Tatsuoka, K. K. (1985). A probabilistic model for diagnosing misconceptions in the pattern classification approach. Journal of Educational Statistics, 12, 55–73.
Tatsuoka, K. K. (1990). Toward an integration of item-response theory and cognitive error diagnosis. In N. Frederiksen, R. Glaser, A. Lesgold, & M. Shafto (Eds.), Diagnostic monitoring of skill and knowledge acquisition (pp. 453–488). Hillsdale, NJ: Erlbaum.
Tatsuoka, K. K. (1991). Boolean algebra applied to determination of universal set of knowledge states (Research report No. RR-91-44-ONR). Princeton, NJ: Educational Testing Service.
Tatsuoka, K. K. (2009). Cognitive assessment: An introduction to the rule space method. New York: Routledge.
Templin, J. L., & Henson, R. A. (2006). Measurement of psychological disorders using cognitive diagnosis models. Psychological Methods, 11, 287–305.
U.S. Department of Education. (2009, November). Race to the top program executive summary. Washington, D.C. Retrieved May 9, 2014 from http://www2.ed.gov/programs/racetothetop/executive-summary.pdf.
von Davier, M. (2005). A general diagnostic model applied to language testing data (Research report). Princeton, NJ: Educational Testing Service.
von Davier, M. (2008). A general diagnostic model applied to language testing data. British Journal of Mathematical and Statistical Psychology, 61, 287–307.
von Davier, M., & Yamamoto, K. (2004). A class of models for cognitive diagnosis. In 4th spearman conference. Philadelphia, PA.
Zhang, S. S., DeCarlo, L. T., & Ying, Z. (2013, March). Non-identifiability, equivalence classes, and attribute-specific classification in Q-matrix based Cognitive Diagnosis Models. ArXiv e-prints.
Author information
Authors and Affiliations
Corresponding author
Appendix: Proof of Theorems
Appendix: Proof of Theorems
We begin with two important propositions necessary to prove the main results; their own proofs are postponed to the end of this section.
Recall that identifiability and local identifiability depend on the probability density function \(f(x; \theta )\), which, when written as a function of the parameters \(\theta \) becomes the likelihood \(L(\theta )\).
Under the model specified in Section 2, given the full set of observations \(R = \{{\mathbf {R}}^i: i = 1,\ldots , N\}\) and a Q-matrix Q, the likelihood of any set of parameters \({\mathbf {c}},{\mathbf {g}},{\mathbf {p}}\) can be written as
where \(N_{{\mathbf {r}}} = |\{i \in \{1,\ldots , N\}: {\mathbf {R}}^i = {\mathbf {r}}\}|\) is the number of observations \({\mathbf {R}}^i\) equal to a particular response vector \({\mathbf {r}}\) and
is the probability of observing \({\mathbf {r}}\) given \(Q,{\mathbf {c}},{\mathbf {g}},{\mathbf {p}}\). The conditional probability \(P(R_j = r_j|Q,c_j,g_j,{\varvec{\alpha }})\) may be expressed as
Defining the likelihood leads to the first of the two propositions, which ties the T-matrix to the likelihood:
Proposition 1
For two sets of parameters \((\hat{\mathbf {c}}, \hat{\mathbf {g}}, \hat{\mathbf {p}})\) and \((\bar{\mathbf {c}}, \bar{\mathbf {g}}, \bar{\mathbf {p}})\),
for all observation matrices R if and only if the following equation holds:
The second proposition describes the linear relationship between certain pairs of T-matrices.
Proposition 2
There exists a matrix \(D({\mathbf {g}}^*)\) depending solely on \({\mathbf {g}}^*=(g_1^*,\ldots ,g_J^*)\), such that for any \({\mathbf {g}}^* \in {\mathbb {R}}^J\),
The matrix \(D({\mathbf {g}}^*)\) is always lower triangular with diagonal \(\text {diag}(D({\mathbf {g}}^*)) = {\mathbf {1}}\), and thus invertible.
The main idea of the proofs is based on the result of Proposition 1. In particular, to show the identifiability, we only need to show that for two sets of parameters \((\hat{\mathbf {c}},\hat{\mathbf {g}},\hat{\mathbf {p}})\) and \((\bar{\mathbf {c}},\bar{\mathbf {g}},\bar{\mathbf {p}})\) satisfying (10), \((\hat{\mathbf {c}},\hat{\mathbf {g}},\hat{\mathbf {p}})=(\bar{\mathbf {c}},\bar{\mathbf {g}},\bar{\mathbf {p}})\). On the other hand, by the definition, if there exist \((\hat{\mathbf {c}},\hat{\mathbf {g}},\hat{\mathbf {p}})\ne (\bar{\mathbf {c}},\bar{\mathbf {g}},\bar{\mathbf {p}})\) and (10) holds, then the model is unidentifiable. We now prove our main theorems and propositions.
Proof of Theorem 1
The case where \({\mathbf {s}}= {\mathbf {g}}= {\mathbf {0}}\) was shown by (Chiu, 2009). For general known \({\mathbf {s}}\) and \({\mathbf {g}}\), by Proposition 1, \({\mathbf {p}}\) is unidentifiable when \(Q,{\mathbf {c}},{\mathbf {g}}\) are known iff there exists \(\hat{\mathbf {p}},\bar{\mathbf {p}}\in {\mathbb {R}}_+^{2^K}\) such that
This occurs iff \(T(Q,{\mathbf {c}},{\mathbf {g}})\) is not a full-rank matrix.
Suppose that the Q-matrix is not complete. WLOG, we assume that the row vector corresponding to the first attribute is missing, i.e., \({\mathbf {e}}_1^\top \not \in \mathcal {R}_Q\). Then, in the T-matrix, the columns corresponding to attribute profiles \(\mathbf {0}\) and \({\mathbf {e}}_1\) are both equal to
and \(\text {rank}(T(Q,{\mathbf {c}},{\mathbf {g}}))< 2^K\).
When Q is complete, assume WLOG that \(Q_{1:K} = I_K\). The matrix \(T(Q,{\mathbf {c}},{\mathbf {g}})\) is full-rank iff \(T(Q,{\mathbf {c}}-{\mathbf {g}},{\mathbf {0}})\) is full-rank, since, by Proposition 2, \(T(Q,{\mathbf {c}}-{\mathbf {g}},{\mathbf {0}}) = D({\mathbf {g}}) T(Q,{\mathbf {c}},{\mathbf {g}})\) and \(D({\mathbf {g}})\) is invertible. Consider the rows of \(T(Q,{\mathbf {c}}-{\mathbf {g}},{\mathbf {0}})\) corresponding to combinations of the first K items, i.e., \({\mathbf {r}}\in \{0,1\}^J\) s.t. \(r_j = 0\) for all \(j > K\). This constitutes an upper-triangular submatrix of size \(2^K\times 2^K\) with diagonal entries \(\prod _{j:r_j = 1} (c_j-g_j) \ne 0\). Thus, \(T(Q,{\mathbf {c}}-{\mathbf {g}},{\mathbf {0}})\) is full-rank, and \({\mathbf {p}}\) is identifiable. \(\square \)
Proof of Theorem 2
Theorem 2 has been recently proved in Chen et al. (2014). For completeness, we include a proof under the setting of this paper. When \({\mathbf {g}}\) is known, we may combine Propositions 1 and 2 to show that two sets of parameters \((\hat{\mathbf {c}},{\mathbf {g}},\hat{\mathbf {p}})\) and \((\bar{\mathbf {c}},{\mathbf {g}},\bar{\mathbf {p}})\) produce equal likelihoods iff
Note that \(c_j \in (g_j,1] \Leftrightarrow c_j- g_j \in (0,1-g_j]\).
Sufficiency For each item \(j \in \{1,\ldots ,J\}\), condition C2 implies that there exists some set of items \(S^j\subset \{1,\ldots , J\}\) s.t. \(j\not \in S^j\) and the attributes required by item j are a subset of the attributes required by the items in \(S^j\); then the sets of attributes required by items in \(S^j\) and by items in \(S^j\cup \{j\}\) are identical. Mathematically, there exists \({\mathbf {r}}^j \in \{0,1\}^J\) s.t. \(r^j_j = 0\) and
To find \({\mathbf {r}}^j\) for each item j, first suppose that \(j\in \{1,\ldots , K\}\). Then \(Q_j = {\mathbf {e}}_j^\top \) and there is some \(j' \in \{K+1,\ldots , J\}\) s.t. \(q_{j'j}= 1\). Let \({\mathbf {r}}^j = {\mathbf {e}}_{j'}\). Otherwise, when \(j \in \{K+1,\ldots , J\}\), let \({\mathbf {r}}^j = \sum _{\{\ell :q_{j\ell }=1\}}{\mathbf {e}}_\ell \).
Then given any two sets of parameters \((\hat{\mathbf {c}},{\mathbf {0}},\hat{\mathbf {p}})\) and \((\bar{\mathbf {c}},{\mathbf {0}},\bar{\mathbf {p}})\) s.t. \(T(Q,\hat{\mathbf {c}},{\mathbf {0}})\hat{\mathbf {p}}= T(Q,\bar{\mathbf {c}},{\mathbf {0}})\bar{\mathbf {p}}\),
Thus, \(\hat{\mathbf {c}}= \bar{\mathbf {c}}\); then, by Theorem 1, \(\hat{\mathbf {p}}=\bar{\mathbf {p}}\).
Necessity By Theorem 1, condition C1 is necessary. Suppose condition C2 fails to hold. WLOG, it fails to hold for the first attribute and \(q_{j1} = 0\) for all \(j \ne 1\). Consider any set of parameters \((\hat{\mathbf {c}},\hat{\mathbf {p}})\) s.t. \(\hat{c}_j \in (g_j,1]\) for all \(j \in \{1,\ldots , J\}\) and \(\hat{\mathbf {p}}\in (0,1)^{2^K}\), \(\sum _{\varvec{\alpha }}p_{\varvec{\alpha }}= 1\). There exists \(\bar{c}_1\) close enough to \(\hat{c}_1\) so that \(\bar{c}_1 \in (g_1,1]\) and \(\bar{p}_{\varvec{\alpha }}\in (0,1)\) for all \({\varvec{\alpha }}\in \{0,1\}^K\), where
Then, for any \({\mathbf {r}}\in \{0,1\}^J\) s.t. \(r_1 = 0\), \(T_{\mathbf {r}}(Q,\hat{\mathbf {c}},{\mathbf {0}}) = T_{\mathbf {r}}(Q,\bar{\mathbf {c}},{\mathbf {0}})\) and
Otherwise, \(r_1 = 1\) and
Thus we have found distinct sets of parameters satisfying (10), and shown that condition C2 is necessary. \(\square \)
Proof of Theorem 3
Thanks to Theorems 1 and 2, conditions C1 and C2 are necessary for identifiability. We now show the necessity of condition C3. Suppose C3 does not hold, but C1 and C2 do. Then all attributes are required by at least two items and there exists an attribute such that it is only required by two items. WLOG, this is the first attribute.
When both items requiring the first attribute require only the first attribute, the Q-matrix can be written WLOG as
As was done for \(r_1\) in the proof of necessity for Theorem 2, consider each possible value of \((r_1,r_2) \in \{0,1\}^2\) to conclude that, for any distinct sets of parameters \((\hat{\mathbf {c}},\hat{\mathbf {g}},\hat{\mathbf {p}})\) and \((\bar{\mathbf {c}},\bar{\mathbf {g}},\bar{\mathbf {p}})\), \(T(Q,\hat{\mathbf {c}},\hat{\mathbf {g}},)\hat{\mathbf {p}}= T(Q,\bar{\mathbf {c}},\bar{\mathbf {g}})\bar{\mathbf {p}}\) if for every \({\varvec{\alpha }}\in \{0,1\}^K\) s.t. \(\alpha _1 = 0\),
Otherwise, the Q-matrix can be written WLOG as
where \({\mathbf {v}}\) is a \((K-1)\)-dimensional nonzero vector. Then \(T(Q,\hat{\mathbf {c}},\hat{\mathbf {g}},)\hat{\mathbf {p}}= T(Q,\bar{\mathbf {c}},\bar{\mathbf {g}})\bar{\mathbf {p}}\) if
Since the equations in (12) are a subset of the equations in (11), finding sets of parameters \((\hat{\mathbf {c}},\hat{\mathbf {g}},\hat{\mathbf {p}})\) and \((\bar{\mathbf {c}},\bar{\mathbf {g}},\bar{\mathbf {p}})\) fulfilling (11) completes the proof for both types of Q-matrices.
Choose a valid set of parameters \((\hat{\mathbf {c}},\hat{\mathbf {g}},\hat{\mathbf {p}})\) s.t. \(\hat{p}_{\varvec{\alpha }}/ \hat{p}_{{\varvec{\alpha }}+{\mathbf {e}}_1} = \rho \) is constant over all \({\varvec{\alpha }}\in \{0,1\}^K\) s.t. \(\alpha _1 =0\). Then, for any \(\bar{\mathbf {g}}\in {\mathbb {R}}^J\), setting
and setting
for every \({\varvec{\alpha }}\in \{0,1\}^K\) s.t. \(\alpha _1=0\) results in a solution to (11). By continuity, there is \(\bar{\mathbf {g}}\) sufficiently close to \(\hat{\mathbf {g}}\) so that \(\bar{\mathbf {c}},\bar{\mathbf {g}}\in [0,1]^J\), \({\mathbf {c}}\succ {\mathbf {g}}\), and \(\bar{\mathbf {p}}\succ {\mathbf {0}}\). Thus, the model is non-identifiable when condition C3 fails, making it a necessary condition. \(\square \)
Proof of Theorem 4
Suppose that conditions C1 and C3 hold, and let \((\hat{\mathbf {c}},\hat{\mathbf {g}},\hat{\mathbf {p}})\) and \((\bar{\mathbf {c}},\bar{\mathbf {g}},\bar{\mathbf {p}})\) be two sets of parameters solving Equation (10). According to condition C1, there is an item requiring solely the kth attribute for each \(k \in \{1,\ldots , K\}\). Moreover, by condition C3, there are also least two additional items requiring the kth attribute. We begin the proof of sufficiency by showing that for every k, there exists an item j requiring the kth attribute s.t. \(\hat{g}_j = \bar{g}_j\). The case where all these items require solely the kth attribute and the case where at least one requires multiple attributes are treated separately.
- Case 1 :
-
All items requiring the kth attribute require solely the kth attribute. WLOG, \(k=1\) and the first three rows of Q are as follows:
$$\begin{aligned} Q_{1:3}= \left( \begin{array}{c@{\quad }c} 1 &{} {\mathbf {0}}^{\top } \\ 1 &{} {\mathbf {0}}^{\top } \\ 1 &{} {\mathbf {0}}^{\top } \\ \end{array}\right) . \end{aligned}$$By Proposition 2, \(T(Q,\hat{\mathbf {c}},\hat{\mathbf {g}})\hat{\mathbf {p}}= T(Q,\bar{\mathbf {c}},\bar{\mathbf {g}})\bar{\mathbf {p}}\) iff
$$\begin{aligned} T(Q,\hat{\mathbf {c}}-\hat{\mathbf {g}},{\mathbf {0}})\hat{\mathbf {p}}= T(Q,\bar{\mathbf {c}}-\hat{\mathbf {g}},\bar{\mathbf {g}}-\hat{\mathbf {g}})\bar{\mathbf {p}}. \end{aligned}$$Then, since
$$\begin{aligned} \frac{T_{{\mathbf {e}}_1+{\mathbf {e}}_3}(\hat{\mathbf {c}}-\hat{\mathbf {g}},{\mathbf {0}})\hat{\mathbf {p}}}{T_{{\mathbf {e}}_1} (\hat{\mathbf {c}}-\hat{\mathbf {g}},{\mathbf {0}})\hat{\mathbf {p}}} = \hat{c}_3 - \hat{g}_3 =\frac{T_{{\mathbf {e}}_1+{\mathbf {e}}_2+{\mathbf {e}}_3}(\hat{\mathbf {c}}-\hat{\mathbf {g}},{\mathbf {0}})\hat{\mathbf {p}}}{T_{{\mathbf {e}}_1+{\mathbf {e}}_2}(\hat{\mathbf {c}}-\hat{\mathbf {g}},{\mathbf {0}})\hat{\mathbf {p}}} , \end{aligned}$$we may conclude that
$$\begin{aligned} \frac{T_{{\mathbf {e}}_1+{\mathbf {e}}_3}(\bar{\mathbf {c}}-\hat{\mathbf {g}},\bar{\mathbf {g}}-\hat{\mathbf {g}})\bar{\mathbf {p}}}{T_{{\mathbf {e}}_1} (\bar{\mathbf {c}}-\hat{\mathbf {g}},\bar{\mathbf {g}}-\hat{\mathbf {g}})\bar{\mathbf {p}}} =\frac{T_{{\mathbf {e}}_1+{\mathbf {e}}_2+{\mathbf {e}}_3}(\bar{\mathbf {c}}-\hat{\mathbf {g}},\bar{\mathbf {g}}-\hat{\mathbf {g}})\bar{\mathbf {p}}}{T_{{\mathbf {e}}_1+{\mathbf {e}}_2}(\bar{\mathbf {c}}-\hat{\mathbf {g}},\bar{\mathbf {g}}-\hat{\mathbf {g}})\bar{\mathbf {p}}}. \end{aligned}$$Let \(\tilde{\mathbf {c}}= \bar{\mathbf {c}}-\hat{\mathbf {g}}\) and let \(\tilde{\mathbf {g}}= \bar{\mathbf {g}}-\hat{\mathbf {g}}\). In addition, let \(\bar{p}_i = \sum _{{\varvec{\alpha }}: \alpha _1=i} p_{\varvec{\alpha }}\) for \(i = 0,1\). Then the previous equation may be written as
$$\begin{aligned} \frac{\tilde{g}_1 \tilde{g}_3 \bar{p}_0 + \tilde{c}_1 \tilde{c}_3 \bar{p}_1}{\tilde{g}_1 \bar{p}_0 + \tilde{c}_1 \bar{p}_1} =\frac{\tilde{g}_1 \tilde{g}_2 \tilde{g}_3 \bar{p}_0 + \tilde{c}_1\tilde{c}_2 \tilde{c}_3 \bar{p}_1}{\tilde{g}_1 \tilde{g}_2 \bar{p}_0 + \tilde{c}_1 \tilde{c}_2\bar{p}_1} \end{aligned}$$and
$$\begin{aligned} \tilde{g}_1 \tilde{c}_1 (\tilde{c}_2 - \tilde{g}_2) (\tilde{c}_3 - \tilde{g}_3) \bar{p}_0 \bar{p}_1 =0. \end{aligned}$$By assumption, \(\bar{\mathbf {p}}\succ {\mathbf {0}}\), \(\tilde{\mathbf {c}}\succ \tilde{\mathbf {g}}\), so \(\hat{g}_1 = \bar{g}_1 \text { or } \bar{c}_1\). By symmetry, \(\bar{g}_1 = \hat{g}_1 \text { or } \hat{c}_1\). If \(\hat{g}_1\ne \bar{g}_1\), then \(\hat{c}_1=\bar{g}_1\) and \(\bar{c}_1=\hat{g}_1\). This contradicts the assumption that \(\hat{\mathbf {c}}\succ \hat{\mathbf {g}}\) and \(\bar{\mathbf {c}}\succ \bar{\mathbf {g}}\). Thus \(\hat{g}_1= \bar{g}_1\).
- Case 2 :
-
At least one item requiring the kth attribute requires multiple attributes. WLOG, \(k = 1\) and
$$\begin{aligned} Q_{1:3}= \left( \begin{array}{c@{\quad }c@{\quad }c} 1 &{} 0 &{}{\mathbf {0}}^{\top } \\ 1 &{} 1 &{} {\mathbf {v}}^{\top } \\ 0 &{} 1 &{} {\mathbf {0}}^{\top } \\ \end{array}\right) , \end{aligned}$$for some vector \({\mathbf {v}}\in \{0,1\}^{K-2}\). We will show that \(\hat{g}_2 = \bar{g}_2\). Since
$$\begin{aligned} \frac{T_{{\mathbf {e}}_1+{\mathbf {e}}_2}(\hat{\mathbf {c}}-\hat{\mathbf {g}},{\mathbf {0}})\bar{\mathbf {p}}}{T_{{\mathbf {e}}_2}(\hat{\mathbf {c}}-\hat{\mathbf {g}},{\mathbf {0}})\bar{\mathbf {p}}}= \hat{c}_1 - \hat{g}_1 = \frac{T_{{\mathbf {e}}_1+{\mathbf {e}}_2+{\mathbf {e}}_3}(\hat{\mathbf {c}}-\hat{\mathbf {g}},{\mathbf {0}})\bar{\mathbf {p}}}{T_{{\mathbf {e}}_2+{\mathbf {e}}_3}(\hat{\mathbf {c}}-\hat{\mathbf {g}},{\mathbf {0}})\bar{\mathbf {p}}}, \end{aligned}$$we know that
$$\begin{aligned} \frac{T_{{\mathbf {e}}_1+{\mathbf {e}}_2}(\bar{\mathbf {c}}-\hat{\mathbf {g}},\bar{\mathbf {g}}-\hat{\mathbf {g}})\bar{\mathbf {p}}}{T_{{\mathbf {e}}_2} (\bar{\mathbf {c}}-\hat{\mathbf {g}},\bar{\mathbf {g}}-\hat{\mathbf {g}})\bar{\mathbf {p}}} = \frac{T_{{\mathbf {e}}_1+{\mathbf {e}}_2+{\mathbf {e}}_3}(\bar{\mathbf {c}}-\hat{\mathbf {g}},\bar{\mathbf {g}}-\hat{\mathbf {g}})\bar{\mathbf {p}}}{T_{{\mathbf {e}}_2+{\mathbf {e}}_3}(\bar{\mathbf {c}}-\hat{\mathbf {g}},\bar{\mathbf {g}}-\hat{\mathbf {g}})\bar{\mathbf {p}}}. \end{aligned}$$Thus,
$$\begin{aligned}&\frac{\tilde{g}_1\tilde{g}_2 \bar{p}_{0,0} +\tilde{c}_1\tilde{g}_2 \bar{p}_{1,0} +\tilde{g}_1\tilde{g}_2 \bar{p}_{0,1} +\tilde{c}_1\tilde{c}_2 \bar{p}_{1,1}}{\tilde{g}_2 \bar{p}_{0,0} +\tilde{g}_2 \bar{p}_{1,0} +\tilde{g}_2 \bar{p}_{0,1} +\tilde{c}_2 \bar{p}_{1,1}}\\&\quad = \frac{\tilde{g}_1\tilde{g}_3\tilde{g}_2 \bar{p}_{0,0} +\tilde{c}_1\tilde{g}_3\tilde{g}_2 \bar{p}_{1,0} +\tilde{g}_1\tilde{c}_3\tilde{g}_2 \bar{p}_{0,1} +\tilde{c}_1\tilde{c}_3\tilde{c}_2 \bar{p}_{1,1}}{\tilde{g}_3\tilde{g}_2 \bar{p}_{0,0} +\tilde{g}_3\tilde{g}_2 \bar{p}_{1,0} +\tilde{c}_3\tilde{g}_2 \bar{p}_{0,1} +\tilde{c}_3\tilde{c}_2 \bar{p}_{1,1}}, \end{aligned}$$where \(\bar{p}_{i,j}=\sum _{{\varvec{\alpha }}: (\alpha _1,\alpha _2) = (i,j)} \bar{p}_{\varvec{\alpha }}\) for \((i,j)\in \{0,1\}^2\), \(\tilde{g}_j=\bar{g}_j-\hat{g}_j\) for \(j=1,2,3\), \(\tilde{c}_j=\bar{c}_j-\hat{c}_j\) for \(j=1,3\), and
$$\begin{aligned} \tilde{c}_2=\frac{(\bar{c}_2-\hat{c}_2)\sum _{{\varvec{\alpha }}: {\varvec{\alpha }}\succeq Q_2}p_{\varvec{\alpha }}+(\bar{g}_2-\hat{g}_2)\sum _{{\varvec{\alpha }}:\alpha _1=\alpha _2=1,{\varvec{\alpha }}\not \succeq Q_2} \bar{p}_{\varvec{\alpha }}}{\bar{p}_{1,1}}. \end{aligned}$$Cross-multiply and cancel to obtain that
$$\begin{aligned} \bar{p}_{0,1} \bar{p}_{1,0} (\tilde{c}_1- \tilde{g}_1) \tilde{g}_2^2(\tilde{c}_3- \tilde{g}_3)= \bar{p}_{0,0} \bar{p}_{1,1}(\tilde{c}_1 - \tilde{g}_1) \tilde{c}_2 \tilde{g}_2 (\tilde{c}_3 - \tilde{g}_3) \end{aligned}$$Now suppose that \(\hat{g}_2 \ne \bar{g}_2\). Since \(\tilde{c}_j > \tilde{g}_j\) for \(j = 1,2,3\),
$$\begin{aligned} \bar{p}_{1,0}\bar{p}_{0,1}(\bar{g}_2 - \hat{g}_2)= \bar{p}_{0,0}\bar{p}_{1,1}(\bar{c}_2-\hat{g}_2). \end{aligned}$$(13)In addition, by symmetry,
$$\begin{aligned} \hat{p}_{1,0}\hat{p}_{0,1} (\hat{g}_2 - \bar{g}_2)= \hat{p}_{0,0}\hat{p}_{1,1}(\hat{c}_2 - \bar{g}_2), \end{aligned}$$(14)where \(\hat{p}_{i,j}=\sum _{{\varvec{\alpha }}: (\alpha _1,\alpha _2) = (i,j)} \hat{p}_{\varvec{\alpha }}\) for \((i,j)\in \{0,1\}^2\). Taken together, (13) and (14) imply that either \(\hat{c}_3 > \hat{g}_3> \bar{c}_3>\bar{g}_3\) or \(\bar{c}_3 > \bar{g}_3> \hat{c}_3>\hat{g}_3\). However, since \(T_{{\mathbf {e}}_2}(\hat{\mathbf {c}},\hat{\mathbf {g}})\hat{\mathbf {p}}= T_{{\mathbf {e}}_2}(\bar{\mathbf {c}},\bar{\mathbf {g}})\bar{\mathbf {p}}\),
$$\begin{aligned} \hat{g}_2(\hat{p}_{0,0}+\hat{p}_{1,0}+\hat{p}_{0,1})+\hat{c}_2 p_{1,1} = \bar{g}_2(\bar{p}_{0,0}+\bar{p}_{1,0}+\bar{p}_{0,1})+\bar{c}_2 p_{1,1}. \end{aligned}$$This is a contradiction; thus \(\hat{g}_2 = \bar{g}_2\).
WLOG, the Q-matrix can be written as
We have shown that for each \(k \in \{1,\ldots , K\}\), there exists some item \(j_k >K\) requiring the kth attribute s.t. \(\hat{g}_{j_k} = \bar{g}_{j_k}\). For each item \(j >K\), let \({\mathbf {r}}^j = \begin{pmatrix}Q_j^\top&{\mathbf {0}}\end{pmatrix}\) be the response vector selecting those among the first K items requiring attributes required by the jth item. Then \({\mathbf {r}}^j\) and \({\mathbf {r}}^j+{\mathbf {e}}_j\) denote distinct sets of items with identical attribute requirements and
Thus, \(\hat{c}_j = \bar{c}_j\) if \(\hat{g}_j = \bar{g}_j\); by the proof of Case 2, this includes all items j requiring multiple attributes. Otherwise, \(Q_j = {\mathbf {e}}_k\) for some \(k \in \{1,\ldots ,K\}\), and the response vectors \({\mathbf {e}}_j + {\mathbf {e}}_{j_k}\) and \({\mathbf {e}}_{j_k}\) represent distinct combinations of items with identical attribute requirements, so that
Thus, \(\hat{c}_j = \bar{c}_j\) for every \(j\in \{1,\ldots , J\}\), i.e., \(\hat{\mathbf {c}}= \bar{\mathbf {c}}\).
We now consider the identifiability of the remaining \(g_j\). For each \(j >K\) s.t. \(Q_j = {\mathbf {e}}_k\) for some \(k \in \{1,\ldots ,K\}\), let \({\mathbf {c}}^* = \hat{c}_k {\mathbf {e}}_k + \hat{c}_j {\mathbf {e}}_j\). Then
and \(\hat{g}_k = \bar{g}_k\). Thus \(g_j\) is identifiable for all \(j >K\).
To show the identifiability of \(g_1,\ldots , g_K\), for each \(k \le K\) let
represent the set of items in \(Q'\) not requiring the kth attribute. When condition C4 holds, there is some item \(\ell >K\) requiring the kth attribute and no other attributes not required by the set of items denoted by \({\mathbf {r}}^k\). Let \({\mathbf {g}}^* = (\hat{c}_1,\ldots , \hat{c}_k, \hat{g}_{K+1},\ldots , \hat{g}_J)^\top \). Then, for any set of parameters \(({\mathbf {c}},{\mathbf {g}},{\mathbf {p}})\) s.t. \(g_j = \hat{g}_j\) for all \(j>K\),
for all response vectors \({\mathbf {r}}\) s.t. \(r_j = 0\) for all \(j\le K\). Since \(\hat{\mathbf {c}}=\bar{\mathbf {c}}\) and \(\bar{g}_j = \hat{g}_j\) for all \(j>K\), this implies that
for all such \({\mathbf {r}}\). Consider the row of \(T(Q,{\mathbf {c}}-{\mathbf {g}}^*,{\mathbf {g}}-{\mathbf {g}}^*)\) corresponding to the combination of the kth item with all the items denoted by \({\mathbf {r}}^k\). The entries of this row-vector are non-zero only for attribute profiles denoting mastery of the skills required by \({\mathbf {r}}^k\) and non-mastery of the kth attribute. Thus,
When condition C4 holds, there is some \({\mathbf {r}}\) s.t. \(r_j = 0\) for all \(j \le K\) and \(T_{{\mathbf {e}}_k + {\mathbf {r}}^k}(Q) = T_{{\mathbf {r}}+ {\mathbf {r}}^k}(Q)\). Then, by (15)
Since \(T_{{\mathbf {e}}_k + {\mathbf {r}}^k}(Q,\hat{\mathbf {c}}-{\mathbf {g}}^*,\hat{\mathbf {g}}-{\mathbf {g}}^*)\hat{\mathbf {p}}=T_{{\mathbf {e}}_k + {\mathbf {r}}^k}(Q,\bar{\mathbf {c}}-{\mathbf {g}}^*,\bar{\mathbf {g}}-{\mathbf {g}}^*)\bar{\mathbf {p}}\), it must be true that \(\hat{g}_k = \bar{g}_k\). Thus, \({\mathbf {g}}\) is fully identifiable and by Theorem 1 so is \({\mathbf {p}}\). \(\square \)
Proof of Proposition 1
The observations follow a multinomial distribution over the set of possible responses \({\mathbf {r}}\in \{0,1\}^J\), with probabilities \(\pi _{\mathbf {r}}\) as defined in (9). For a particular \((\hat{\mathbf {c}},\hat{\mathbf {g}},\hat{\mathbf {p}})\) and \((\bar{\mathbf {c}},\bar{\mathbf {g}},\bar{\mathbf {p}})\),
for all observation matrices R iff \(\pi _{\mathbf {r}}(Q,\hat{\mathbf {c}},\hat{\mathbf {g}},\hat{\mathbf {p}})= \pi _{\mathbf {r}}(Q,\bar{\mathbf {c}},\bar{\mathbf {g}},\bar{\mathbf {p}})\) for all \({\mathbf {r}}\in \{0,1\}^J\). Consider a P-matrix \(P(Q,{\mathbf {c}},{\mathbf {g}})\) indexed like the T-matrix by item subsets \({\mathbf {r}}\in \{0,1\}^J\) and attribute profiles \({\varvec{\alpha }}\in \{0,1\}^K\). The entries of \(P(Q,{\mathbf {c}},{\mathbf {g}})\) are denoted by the quantities
Then the statement \(\pi _{\mathbf {r}}(Q,\hat{\mathbf {c}},\hat{\mathbf {g}},\hat{\mathbf {p}})= \pi _{\mathbf {r}}(Q,\bar{\mathbf {c}},\bar{\mathbf {g}},\bar{\mathbf {p}})\) for all \({\mathbf {r}}\in \{0,1\}^J\) can be written in matrix notation as \(P(Q,\hat{\mathbf {c}},\hat{\mathbf {g}})\hat{\mathbf {p}}= P(Q,\bar{\mathbf {c}},\bar{\mathbf {g}})\bar{\mathbf {p}}\). Since
there is a one-to-one linear transformation between \(P(Q,{\mathbf {c}},{\mathbf {g}})\) and \(T(Q,{\mathbf {c}},{\mathbf {g}})\) that is not dependent on \(Q,{\mathbf {c}},{\mathbf {g}}\), and
\(\square \)
Proof of the Proposition 2
In what follows, we construct such a D matrix satisfying the conditions in the proposition, i.e., \(D({\mathbf {g}}^*)\) is a matrix only depending on \({\mathbf {g}}^*\) such that \(D_{{\mathbf {g}}^*} T(Q,{\mathbf {c}},{\mathbf {g}}) =T_{{\mathbf {c}}-{\mathbf {g}}^*,{\mathbf {g}}-{\mathbf {g}}^*}(Q)\) for any \(Q,{\mathbf {c}},{\mathbf {g}}\). Recall that for any \(Q,{\mathbf {c}},{\mathbf {g}}\),
We may extend this definition to include \({\mathbf {c}},{\mathbf {g}}\not \in [0,1]^M\), though in such cases the \(t_{{\mathbf {r}},{\varvec{\alpha }}}\) will no longer correspond to probabilities. Then for any \({\mathbf {g}}^* \in {\mathbb {R}}\),
where \(b_j = c_j^{\xi _j(Q,{\varvec{\alpha }})}g_j^{1-\xi _j(Q,{\varvec{\alpha }})} = t_{{\mathbf {e}}_j,{\varvec{\alpha }}}(Q,{\mathbf {c}},{\mathbf {g}})\). By polynomial expansion,
Define the entries \(d_{{\mathbf {r}},{\mathbf {r}}'}({\mathbf {g}}^*)\) of \(D({\mathbf {g}}^*)\) as
Then
where \(D({\mathbf {g}}^*)\) is a lower triangular matrix depending solely on \({\mathbf {g}}^*\) with eigenvalues equal to its diagonal. Since \(\text {diag}(D({\mathbf {g}}^*)) = {\mathbf {1}}\), \(D({\mathbf {g}}^*)\) is invertible. \(\square \)
Rights and permissions
About this article
Cite this article
Xu, G., Zhang, S. Identifiability of Diagnostic Classification Models. Psychometrika 81, 625–649 (2016). https://doi.org/10.1007/s11336-015-9471-z
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11336-015-9471-z