Tetrad-Differencing CML in Directed Networks

• Tetrad-differencing CML is an estimation principle for ordered dyadic data that cancels flexible sender and receiver fixed effects through tetrad differences.
• It transforms ordinal responses into threshold-specific binary indicators and applies difference-in-differences statistics across tetrads to eliminate incidental parameters.
• The pooled tetrad logit estimator (PTLE) consistently aggregates information across outcome thresholds, offering robust performance in sparse network settings.

Searching arXiv for the specified paper and closely related tetrad-differencing CML network work. Tetrad-differencing conditional maximum likelihood (CML) is an estimation principle for ordered dyadic data in directed networks, developed for ordered logit models with sender and receiver fixed effects that can vary arbitrarily across outcome categories. In this setting, the central difficulty is an incidental parameter problem generated by the coexistence of $N(N-1)$ dyadic observations and $2NM$ category-specific fixed effects. The ordered-data extension constructs binary indicators at each outcome threshold, forms tetrad-level difference-in-differences statistics, and conditions on informative events so that the fixed effects cancel from the conditional likelihood. The resulting framework yields two estimators—the Equally-Weighted Tetrad Logit Estimator (ETLE) and the Pooled Tetrad Logit Estimator (PTLE)—with PTLE shown to be consistent under weaker identification conditions because it pools information across outcome categories (Muris et al., 22 Jul 2025).

1. Ordered logit formulation for directed dyads

The model is defined on a directed network with ordered outcomes

$Y_{ij}\in\{0,1,\dots,M\}, \qquad (i\neq j),$

and dyad-specific covariates $X_{ij}\in\mathbb R^k$. It posits a latent index

$Y^*_{ij}=X_{ij}'\beta-\epsilon_{ij},$

with threshold-crossing representation

$$Y_{ij}\ge m \Longleftrightarrow Y^*_{ij}\ge \lambda^*_{ijm},\qquad m=1,\dots,M,$$

where $\epsilon_{ij}\sim$ i.i.d. Logistic$(0,1)$.

Sender and receiver heterogeneity enter through category-specific threshold components,

$$\lambda^*_{ijm}=\lambda_{im}+\delta_{jm}, \qquad i,j\in\{1,\dots,N\},\;m=1,\dots,M.$$

Here, sender $i$ has threshold $2NM$0 at category $2NM$1, and receiver $2NM$2 has threshold $2NM$3. The full incidental-parameter vector is

$2NM$4

Under this specification, the individual-dyad probabilities are

$2NM$5

where $2NM$6.

This specification is designed to allow flexible sender and receiver fixed effects that vary arbitrarily across outcome categories. That flexibility is substantively valuable in network settings with heterogeneous propensities to send and receive ties at different ordinal intensities, but it also creates the estimation problem that motivates tetrad-differencing CML.

2. Incidental parameters and the tetrad-differencing principle

Because there are $2NM$7 fixed effects and only $2NM$8 observations, standard maximum likelihood estimation suffers from incidental parameter bias: as $2NM$9, the number of fixed effects grows at the same order as the number of observations, preventing consistent estimation of $Y_{ij}\in\{0,1,\dots,M\}, \qquad (i\neq j),$0 by joint MLE. The problem is particularly challenging under network sparsity or when some outcome categories are rare (Muris et al., 22 Jul 2025).

The ordered-data construction begins by transforming the ordinal response into threshold-specific binary indicators. For each $Y_{ij}\in\{0,1,\dots,M\}, \qquad (i\neq j),$1,

$Y_{ij}\in\{0,1,\dots,M\}, \qquad (i\neq j),$2

with conditional success probability

$Y_{ij}\in\{0,1,\dots,M\}, \qquad (i\neq j),$3

At each threshold, the ordered model is thus represented as a binary logit model with sender and receiver fixed effects indexed by $Y_{ij}\in\{0,1,\dots,M\}, \qquad (i\neq j),$4.

The tetrad construction uses four distinct nodes $Y_{ij}\in\{0,1,\dots,M\}, \qquad (i\neq j),$5. For a common cutoff $Y_{ij}\in\{0,1,\dots,M\}, \qquad (i\neq j),$6, define the $Y_{ij}\in\{0,1,\dots,M\}, \qquad (i\neq j),$7 difference-in-differences statistic

$Y_{ij}\in\{0,1,\dots,M\}, \qquad (i\neq j),$8

The associated quadruple-difference regressor is

$Y_{ij}\in\{0,1,\dots,M\}, \qquad (i\neq j),$9

The core sufficiency result states that, after conditioning on the informative event $X_{ij}\in\mathbb R^k$0, the fixed effects cancel: $X_{ij}\in\mathbb R^k$1 Theorem 1 in the source formalizes this cancellation property. The significance of the result is methodological rather than merely algebraic: the conditional likelihood becomes free of $X_{ij}\in\mathbb R^k$2, so consistent estimation remains feasible even under sparsity.

3. Conditional likelihood for ordered outcomes

For each tetrad $X_{ij}\in\mathbb R^k$3 and cutoff $X_{ij}\in\mathbb R^k$4, define

$X_{ij}\in\mathbb R^k$5

The conditional log-likelihood contribution is

$X_{ij}\in\mathbb R^k$6

Summing over all tetrads $X_{ij}\in\mathbb R^k$7 gives

$X_{ij}\in\mathbb R^k$8

Applying the CML principle to ordered data yields multiple likelihood contributions corresponding to different outcome thresholds. This is the principal distinction from binary-choice tetrad CML: the ordered response generates one conditional likelihood contribution for each cutoff $X_{ij}\in\mathbb R^k$9, and estimation depends on how these contributions are aggregated.

The two estimators proposed in the source differ exactly on this point. ETLE, the Equally-Weighted Tetrad Logit Estimator, normalizes each threshold so that each contributes equally. PTLE, the Pooled Tetrad Logit Estimator, assigns equal weight to each informative tetrad-cutoff pair. The pooled estimator is defined by

$Y^*_{ij}=X_{ij}'\beta-\epsilon_{ij},$0

For ETLE, the normalization uses

$Y^*_{ij}=X_{ij}'\beta-\epsilon_{ij},$1

so that each threshold is normalized to contribute equally.

This distinction is consequential when informative tetrads are unevenly distributed across categories. A plausible implication is that equal weighting across thresholds can overemphasize rare categories relative to their information content, whereas pooling can better exploit the aggregate signal across cutoffs. The formal asymptotic results in the source support that interpretation.

4. Identification and consistency

The asymptotic analysis introduces, for each threshold $Y^*_{ij}=X_{ij}'\beta-\epsilon_{ij},$2, the informative-tetrad probability

$Y^*_{ij}=X_{ij}'\beta-\epsilon_{ij},$3

Under mild sampling and moment assumptions, the strong condition used for both ETLE and PTLE is:

$Y^*_{ij}=X_{ij}'\beta-\epsilon_{ij},$4

The weaker pooled condition used only for PTLE is: $Y^*_{ij}=X_{ij}'\beta-\epsilon_{ij},$5

$Y^*_{ij}=X_{ij}'\beta-\epsilon_{ij},$6

Theorem 2 states that under $Y^*_{ij}=X_{ij}'\beta-\epsilon_{ij},$7, both ETLE $Y^*_{ij}=X_{ij}'\beta-\epsilon_{ij},$8 and PTLE $Y^*_{ij}=X_{ij}'\beta-\epsilon_{ij},$9 are consistent, $$Y_{ij}\ge m \Longleftrightarrow Y^*_{ij}\ge \lambda^*_{ijm},\qquad m=1,\dots,M,$$0. Under $$Y_{ij}\ge m \Longleftrightarrow Y^*_{ij}\ge \lambda^*_{ijm},\qquad m=1,\dots,M,$$1, PTLE alone is consistent (Muris et al., 22 Jul 2025). The proof sketches proceed by establishing uniform convergence of the sample objective functions to strictly concave limits through U-statistic and dependency arguments for overlapping tetrads, followed by the argmax-continuous-mapping theorem.

The conceptual importance of these conditions lies in the difference between category-by-category identification and pooled identification. PTLE requires only sufficient information when pooling across categories, rather than sufficient information in each category. This is especially relevant in sparse networks and in applications with infrequent upper-tail outcome categories. A common misconception is that the ordered setting merely repeats the binary case cutoff by cutoff; the consistency result shows instead that pooling changes the identification requirements in a fundamental way.

5. Monte Carlo evidence

The simulation design uses sample sizes $$Y_{ij}\ge m \Longleftrightarrow Y^*_{ij}\ge \lambda^*_{ijm},\qquad m=1,\dots,M,$$2, a single covariate $$Y_{ij}\ge m \Longleftrightarrow Y^*_{ij}\ge \lambda^*_{ijm},\qquad m=1,\dots,M,$$3 with $$Y_{ij}\ge m \Longleftrightarrow Y^*_{ij}\ge \lambda^*_{ijm},\qquad m=1,\dots,M,$$4 Bernoulli$$Y_{ij}\ge m \Longleftrightarrow Y^*_{ij}\ge \lambda^*_{ijm},\qquad m=1,\dots,M,$$5, and two thresholds $$Y_{ij}\ge m \Longleftrightarrow Y^*_{ij}\ge \lambda^*_{ijm},\qquad m=1,\dots,M,$$6, corresponding to outcomes $$Y_{ij}\ge m \Longleftrightarrow Y^*_{ij}\ge \lambda^*_{ijm},\qquad m=1,\dots,M,$$7. Node-specific threshold heterogeneity is controlled by

$$Y_{ij}\ge m \Longleftrightarrow Y^*_{ij}\ge \lambda^*_{ijm},\qquad m=1,\dots,M,$$8

Errors are logistic, and thresholds are set as

$$Y_{ij}\ge m \Longleftrightarrow Y^*_{ij}\ge \lambda^*_{ijm},\qquad m=1,\dots,M,$$9

under two common-threshold scenarios, $\epsilon_{ij}\sim$0 and $\epsilon_{ij}\sim$1.

The comparison criteria are Bias, StdDev, and IQR of $\epsilon_{ij}\sim$2, with ETLE and PTLE evaluated against single-cutoff binary-CMLE at $\epsilon_{ij}\sim$3 or $\epsilon_{ij}\sim$4. The simulation results confirm the theoretical preference for PTLE. In dense networks, corresponding to small $\epsilon_{ij}\sim$5, ETLE and PTLE are approximately equal. As sparsity increases, ETLE becomes unstable, with large bias and variance, because some cutoffs yield few tetrads. PTLE remains stable. Single-cutoff methods are extremely sensitive to cutoff choice and ignore information from other thresholds (Muris et al., 22 Jul 2025).

These findings clarify the practical meaning of the asymptotic theory. The issue is not only whether a given category is informative, but whether identification should be imposed separately at each threshold or on the pooled ordered structure. The simulations indicate that the latter can be materially more robust when ordered outcomes are imbalanced.

6. Empirical application and practical implications

The empirical application studies friendship networks among Dutch university students. The data consist of $\epsilon_{ij}\sim$6 students, with friendship rated on a six-point scale $\epsilon_{ij}\sim$7–$\epsilon_{ij}\sim$8, and covariates “common gender,” “both smokers,” and “common program.” The missing category “unknown person” is treated as missing. Informative tetrads are constructed at cutoffs $\epsilon_{ij}\sim$9.

For the time-5 snapshot, the comparison across estimators is substantively revealing. A naïve ordered logit without fixed effects produces weak homophily signs. Ordered logit with two-way fixed effects yields large positive estimates for all three variables, but is susceptible to incidental parameter bias. ETLE gives positive but imprecise estimates, especially for the “best friends” category. PTLE yields all three homophily effects as highly significant: $(0,1)$0 Single-cutoff binary CML estimates vary dramatically by chosen cutoff, sometimes even with wrong sign (Muris et al., 22 Jul 2025).

The substantive conclusion is strong, positive homophily on gender, smoking, and program, together with a methodological warning: standard methods without flexible fixed effects can produce misleading signs. Over seven waves, PTLE yields smooth, interpretable time-series of homophily coefficients, whereas ETLE and single-cutoff approaches often spike wildly when categories are rare.

The practical recommendations follow directly from these results. The source recommends always pooling information across thresholds through PTLE in ordered dyadic CML, checking the total number of informative tetrads through the pooled quantity $(0,1)$1, and implementing the estimator via a single logistic regression on the “flattened” dataset of $(0,1)$2 for all $(0,1)$3. It also recommends robust sandwich standard errors obtained by grouping score contributions at the dyad level to account for network dependence.

7. Limitations and open directions

The framework is developed for ordered logit models with logistic errors and threshold decomposition $(0,1)$4. Within that scope, the source identifies several limitations and open questions. Finite-sample bias may remain when $(0,1)$5 is very small, and this may require bias-correction or penalization. Extensions to dynamic panel-network models or to non-logistic errors are left open. Inference under more complex dependence structures, such as triadic closure, may require higher-order U-statistic theory.

These limitations indicate that tetrad-differencing CML is not a generic cure for all fixed-effects problems in network analysis. Its contribution is more specific: it provides the first estimation method for ordered directed-network outcomes with flexible sender and receiver fixed effects varying across categories, while preserving consistent estimation of $(0,1)$6 through conditional elimination of the incidental parameters. Within that domain, PTLE occupies the central role because its consistency relies on pooled information across thresholds rather than separate identification within every category.