Network TMLE: Structural Models, Dimension Reduction, and Inference
Suppose an intervention changes treatment across a network. To estimate its average outcome, we must account for how each person responds to the treatment of others and for the dependence between observations. Network TMLE in (Van Der Laan 2014) approaches the first problem through shared summaries of the network and a targeted correction to an outcome regression. A separate dependence argument is needed for inference.
1. The target and identification
Let \(W\) be baseline covariates, \(A\) the treatment vector, and \(Y_i\) an outcome. Suppose known summaries \(C_i^Y=c_i(A,W)\) suffice for the conditional outcome mean, shared across units as \(\overline Q(C_i^Y)\). For an intervention distribution \(g^*(a\mid w)\), the target is
Consistency, the stipulated causal/no-unmeasured-confounding conditions, and positivity identify this intervention mean. In a longitudinal model, the corresponding formula iterates over time-specific covariate and treatment histories.
Sequential exchangeability and conditional independence of different individuals’ treatment errors are distinct restrictions. Neither should be used as a synonym for the other.
2. The influence-function result
Proposition 1. Under the paper’s independent-baseline model and shared conditional mechanisms, the efficient influence function contains a baseline component and a weighted outcome-residual component. For the latter,
where \(\overline h\) and \(\overline h^*\) are the average distributions of the outcome summary under observation and intervention. The full gradient also accounts for how each baseline covariate affects the intervention mean.
The weight must change the distribution of every treatment feature entering the outcome regression. When that regression includes neighbors’ treatments, an individual treatment propensity alone does not describe the required change of measure.
3. Why the residual correction works
The residual correction has two useful properties. With a correct outcome regression, the residual has conditional mean zero, and multiplying it by a function of the summary preserves that zero expectation. With a correct summary density, the weight instead transfers the residual average to the intervention distribution, where it cancels the regression error. These are the two bias cancellations behind double robustness, under the maintained baseline-distribution and causal assumptions.
TMLE implements that correction by fluctuating the initial regression along a submodel whose score includes the weighted residual. Fitting the fluctuation approximately solves the empirical influence equation. Evaluating the intervention formula at the updated regression yields a plug-in estimate compatible with the outcome range.
4. What remains for inference
For inference, write the estimation error as an average of influence contributions plus a remainder. Nuisance rates and empirical-process bounds must make that remainder negligible. A network CLT then applies to the average, whose variance includes covariances between units with overlapping neighborhoods.
The single-network asymptotic result determines the sampling scale after targeting. It explains how growing network degree can slow convergence and when the usual \(\sqrt N\) normalization is available.
5. Why the weight is a pooled summary-density ratio
Introduce an auxiliary index \(I\) drawn uniformly from \(\{1,\ldots,N\}\), independently of the network data, and set \(C=C_I^Y\). Its observed density is \(\overline h=N^{-1}\sum_i h_i\), and its intervention density is \(\overline h^*=N^{-1}\sum_i h_i^*\). For an integrable function \(v\),
Multiplying by the ratio of intervention density to observed density cancels the observed density in this integral:
This calculation uses a common regression: the same function of \(c\) must describe each unit’s conditional mean. If an omitted unit characteristic changes that mean, pooling densities leaves the regression misspecified. Thus both the shared regression model and positivity are needed for this change-of-measure argument.
An individual propensity weight changes only the distribution of own treatment. If \(C_i^Y\) includes the number of treated neighbors, the intervention also changes that count’s distribution. The pooled ratio accounts for the whole summary entering the outcome regression.
6. The exact bias cancellation in a simplified subproblem
Hold the baseline distribution and intervention summary density fixed for this calculation. Let \(Q\) be the true shared regression, \(\widetilde Q\) a candidate, and \(\widetilde h\) a candidate observed summary density. The population augmented functional is
Write \(e(c)=\widetilde Q(c)-Q(c)\). Conditioning the residual on \(C\) and subtracting the target yields
If either \(e=0\) or \(\widetilde h=\overline h\), the bias vanishes. Otherwise it is a product of nuisance errors. With a denominator bounded away from zero and bounded change-of-measure weights, Cauchy–Schwarz bounds its magnitude by a constant times the product of the corresponding \(L^2\) errors.
For the full network target, baseline distributions and the induced intervention law are also estimated. Their errors contribute additional remainder terms. The calculation above isolates one cancellation that the full argument must retain.
7. How the targeting update solves the score equation
For an outcome in \([0,1]\), take an initial regression \(Q_0(c)\) in \((0,1)\) and a fitted clever covariate \(H(c)=\widehat{\overline h}^*(c)/\widehat{\overline h}(c)\). A one-dimensional fluctuation is
For the Bernoulli log-likelihood criterion, the empirical score and curvature are
An interior maximizer therefore sets the weighted residual sum to zero. Dependence among observations does not alter this differentiation; it changes the statistical argument for the fitted score and its variance. For bounded nonbinary outcomes, the Bernoulli loss is a working loss for the mean, rather than a claim that the outcome distribution is Bernoulli.
The updated regression stays in the outcome range. Evaluating the intervention mean using it gives a substitution estimator with the same range constraint. Solving an influence equation and respecting the parameter space are compatible here because the fluctuation is chosen to target the relevant residual direction.
The score equation also shows exactly how targeting differs from an arbitrary regression refit. At \(\epsilon=0\), write \(S_0=\sum_iH_i(Y_i-Q_{0i})\) and \(I_0=\sum_iH_i^2Q_{0i}(1-Q_{0i})\). If \(I_0>0\), one Newton step is
The update therefore responds to the intervention-weighted residual. Its size is divided by the curvature, which measures how quickly the score changes. An observation with a large intervention weight may have little influence on unweighted prediction error but a large influence on this update. Extremely large \(H_i\) can make a few observations dominate both quantities.
For the simplified fixed-baseline functional, the preceding bias identity gives a more precise orthogonality calculation. Along \(\widetilde Q=Q+tu\) and \(\widetilde h=\overline h+tv\), with denominators positive,
The absence of a term proportional to \(t\) is the useful feature of this identity. The two nuisance errors enter as a product, which may be small enough for inference even when neither nuisance estimate has the final parameter’s convergence rate. By comparison, the uncorrected regression average has the first-order error \(\int(\widetilde Q-Q)\overline h^*\).
There are therefore two arguments to complete in a network application. The pooled density identity explains the bias correction under a shared summary model. The dependence argument establishes how accurately its empirical average estimates its expectation.
8. Baseline uncertainty is not an ordinary empirical mean by default
Let \(G(W)\) denote the intervention average for a fixed baseline vector. Changing one baseline variable can change several units’ intervention outcomes through the network. In an illustrative model with separately varying independent baseline marginals, perturbing the \(j\)th marginal along a mean-zero score \(s_j\) changes the target by
This explains the conditional-expectation terms in baseline gradient calculations. Shared marginal models impose further pooling restrictions. Simply inserting \(G(W_j)-\psi\) would ignore the fact that \(G\) is a function of the entire baseline vector.
Finally, the nuisance-product remainder must be negligible at the actual network normalization. If the estimator fluctuates at rate \(C_N^{-1/2}\), a schematic requirement is \(r_Qr_h=o(C_N^{-1/2})\), alongside the other remainder and empirical-process conditions. Targeting removes a leading bias term; the network dependence argument still determines the sampling scale and covariance calculation.
Making the product-error condition precise
For the simplified functional in Section 6, a useful sufficient condition is \(\overline h^*/\widetilde h\le M\) on the common support. Write \(r=(\widetilde h-\overline h)/\overline h\) where \(\overline h>0\). The exact remainder becomes
Thus the density error is measured here as a relative error under the observed summary law. This formulation makes the measure in each norm explicit and avoids assuming that a Lebesgue density is bounded below over an unbounded support. If the two norms are \(O_p(r_Q)\) and \(O_p(r_h)\), multiplication by the network normalization gives \(\sqrt{C_N}R=O_p(\sqrt{C_N}r_Qr_h)\). The condition \(\sqrt{C_N}r_Qr_h\to0\) removes this term. The baseline-law and empirical-average terms described above remain separate obligations in the full estimator.