LogitMap Pairs in Categorical Analysis

• LogitMap Pairs are explicit correspondences between parameterizations in log-linear and logistic models, enabling systematic transfer of estimates and inferential properties.
• They employ algebraic mappings, such as incidence matrices with corner-point constraints, to relate log-linear parameters with logistic coefficients and ensure equivalent MLEs and deviance measures.
• Extended applications include latent-feature models for dyadic prediction and RC association models, offering robust frameworks for model selection and hypothesis testing in categorical data analysis.

LogitMap Pairs refer to mathematical correspondences between parameterizations and inferential structures in different classes of log-linear and logit (logistic) models, as well as the explicit mappings in more recent latent-feature log-linear approaches and generalized row–column (RC) association models for categorical data. These pairings originate from foundational results in contingency table analysis and have been systematically formalized to enable the transfer of model parameters, interval estimates, deviances, and inferences between distinct but structurally related models for categorical and dyadic data.

1. Classical LogitMap Pairs: Log-linear and Logistic Regression Correspondence

The archetypal LogitMap Pair arises from the correspondence between Poisson log-linear models for multidimensional contingency tables containing at least one binary factor $Y$, and conditional binary logistic regression models where $Y$ is the binary response. If $\mathcal{P}$ denotes the set of factors, $Y \in \mathcal{P}$ the binary response, and $X = \mathcal{P} \setminus \{Y\}$, counts are arranged as $n_{j_1,\dots,j_p}$, indexing over all factor combinations.

Fitting a saturated Poisson GLM with main effects and all interactions among $X$ and all such interactions with $Y$ produces parameter estimates $\lambda$ (log-linear parameters) fulfilling

$\log E[n_{j_1,\dots,j_p}] = X_{LL} \lambda$

with the requirement that the model incorporates all joint structure among $X$ and their interactions with $Y$. The implied conditional odds for $Y=1|X$ then possess the exact structure of a logistic regression:

$\logit\ \Pr(Y=1|X=j_2\dots j_p) = X_{\text{logit}} \beta$

where $\beta$ is a vector of logistic regression parameters. No reduced log-linear model yields that identical logistic form unless all $X$ interactions are included (Jing et al., 2017).

2. Formal Parameter Mapping and Structural Incidence Relations

Critical to the LogitMap Pair is the algebraic mapping from the log-linear parameters $\lambda$ to the logistic regression coefficients $\beta$. With corner-point constraints and suitable indexing:

• $\lambda_0$ represents the overall intercept,
• $\lambda_u$ indexes the effect for marginal subset $u \subseteq \mathcal{P}$, including main and interaction effects.

The mapping is realized as:

$\beta = T \lambda$

where $T$ is an incidence matrix effecting the difference between log-linear parameters that involve $Y$ and those that do not:

$$T_{(u),(v)} = \begin{cases} +1 & \text{if } v = u \cup \{Y\} \ -1 & \text{if } v = u \ 0 & \text{otherwise} \end{cases}$$

For a concrete $X=x$,

$\logit\ \Pr(Y=1|X=x) = \sum_{u \subseteq X} [\lambda_{Y \cup u} - \lambda_u] = \sum_{u \subseteq X} \beta_u 1_{(x\, \text{has}\, u)}$

This algebraic structure ensures that effect estimates for $Y$ in the logistic layer are linear combinations of the relevant log-linear effects, enabling exact parameter-pair correspondences (Jing et al., 2017).

3. MLE, Inferential Equivalence, and Deviance Correspondence

The maximum likelihood estimates (MLEs) of $\lambda$ and $\beta$ under the conditions described above coincide for the mapped components: $$\widehat{\beta} = T \widehat{\lambda}_{LL,\,\text{relevant}}$$. This extends to standard errors, as the relevant Fisher information block in the log-linear model exactly matches the information in the logistic model for those effects involving $Y$. Thus, asymptotically, the Wald confidence intervals for $\beta$ coincide with those for the corresponding $\lambda$ coefficients.

Deviance equivalence between the log-linear (Poisson) and logistic (product-binomial) models holds precisely when no cell merging occurs in the logistic data: $D_{LL} = D_{\text{logit}}$. Each log-likelihood contribution matches term by term if the logistic regression is carried out without collapsing cells in the contingency structure (Jing et al., 2017).

4. LogitMap in Dyadic Prediction: Latent-Feature Log-linear Models

Expanding the LogitMap concept, latent-feature log-linear (LFL) models for dyadic prediction define a LogitMap from a dyad $(i,j)$, possibly endowed with side-information $s_{ij}$, to the space of conditional label probabilities. For dyads indexed by $(i,j)$ with label $y_{ij} \in \mathcal{Y}$ (of cardinality $L$), the LFL approach assigns each label $k$ a row-latent matrix $\Alpha^k$, column-latent matrix $\Beta^k$, per-label bias $\gamma_k$, and side-weight $w^k_s$. The natural parameter is

$$\eta_{ij}^{(k)} = (\alpha^k_{i:})^\top \beta^k_{j:} + \gamma_k + (w^k_s)^\top s_{ij}$$

with multinomial logit probabilities

$$P(y_{ij}=k\,|\,i,j) = \frac{\exp(\eta_{ij}^{(k)})}{\sum_{k'=1}^L \exp(\eta_{ij}^{(k')})}$$

This framework generalizes LogitMap Pairs to dyadic data, allowing both identifier-only and side-information-rich regimes. The approach enables learning well-calibrated, low-rank log-linear predictors for nominal and ordinal outcomes with discriminative objectives, maintaining scalability and resistance to sample-selection bias (Menon et al., 2010).

5. Extended LogitMap Pairs in RC Association Models

A further generalization is realized in the extended class of row-column (RC) association models for two-way tables. Here, arbitrary pairs of logit functions—local ($L$), global/cumulative ($G$), continuation ($C$), and reverse-continuation ($R$)—can be paired to define odds-ratio-type interactions on different scales via $\phi$-divergence functions:

$$\gamma_{ij}(F; l_1, l_2) = F[\varrho_{ij}(1,1)] - F[\varrho_{ij}(1,0)] - F[\varrho_{ij}(0,1)] + F[\varrho_{ij}(0,0)]$$

with $\varrho_{ij}(u,v)$ the generalized odds-ratio at cutpoints $(i, j)$ and $F$ derived from a convex $\phi$. Reconstruction theorems guarantee that, given all marginal logits and the interaction matrix $\gamma_{ij}$ (subject to a rank-$K$ constraint), the joint probability table is uniquely identified. This explicitly defines LogitMap Pairs between sets of marginal logit effects and pairs of association parameters, covering diverse logit types and scaling choices (Forcina et al., 2019).

6. Uniqueness, Structural Constraints, and Positive Association

Within the extended RC association framework, the LogitMap is a bijective correspondence: the pair of marginal logits and $\phi$-scaled logit-pairwise interactions determine the full joint distribution uniquely. This is guaranteed by properties of the function $G = F^{-1}$, the additive Lagrange structure in the representation theorems, and supporting monotonicity arguments.

Rank constraints on the interaction matrix $\gamma_{ij}$, implemented via specific algebraic operators, enable fitting parsimonious association structures (e.g., rank-1 for strong monotone dependence). Nonnegative $\gamma_{ij}$ induce positive association properties, with tractable stochastic ordering implications depending on the logit type pair—e.g., quadrant orderings for $(G,G)$ and strong dependence for $(L,L)$ (Forcina et al., 2019).

7. Applications and Modeling Implications

The LogitMap Pair correspondence is leveraged in classical contingency table analysis, dyadic prediction scenarios (collaborative filtering, link prediction), and ordinal association studies (e.g., social mobility). In each case, the LogitMap yields a structured mapping between different parameterizations and inferential frameworks, enabling:

• Transfer of MLEs and their asymptotic properties,
• Direct comparison of deviance and fit statistics,
• Efficient model selection and hypothesis testing,
• Robustness in the presence of sample selection or structural zeros.

Notable applications include the coronary-heart-disease cohort re-analysis, where the mapped estimates and deviance were shown to coincide exactly, and the British social mobility study, where extended RC models with selected logit-pairings and rank constraints provided robust, interpretable inferences (Jing et al., 2017, Forcina et al., 2019).

In summary, LogitMap Pairs formalize a robust, algebraically explicit, and practically valuable connection between diverse models for categorical data, enabling precise inferential translation across log-linear, logistic, latent-feature, and generalized association frameworks.