The synthetic control, augmented synthetic control and synthetic difference-in-differences methods in the preceding sections each extend basic parallel trends in a different direction. Standard synthetic control relies on a convex combination of donors that matches the treated unit’s pre-treatment path. Augmented synthetic control adds an explicit outcome model to correct residual imbalance. SDID reweights both units and time periods to focus on more comparable comparisons. Each design is robust within its own framework, but each can fail badly when its primary identification mechanism breaks down. SDID, for example, remains biased if weighted parallel trends does not hold, no matter how well we tune unit and time weights. Triply Robust Panel (TROP) estimators push this logic one step further by combining all three ingredients: unit weights, time weights and a flexible outcome model based on interactive fixed effects. Athey et al. [2025b] propose a TROP estimator that learns unit weights to balance treated and control groups, learns time weights to down-weight less informative periods, and fits a low-rank factor structure to capture heterogeneous responses to common shocks. The core theoretical claim is a conditional bias bound: under a factor-structured model for untreated outcomes, the leading bias term is bounded by a constant times the product of three errors, one for unit imbalance, one for time imbalance, and one for misspecification in the regression adjustment. This is the sense in which the estimator is “triply robust”. To connect the paper’s formalism to the monograph’s causal language, it is useful to state the target explicitly. A natural estimand in the paper is a treated-cell average over the analysis window, $\tau_{\text{TROP}} := E[Y_{it}(1) - Y_{it}(0) | D_{it} = 1]$ . This is an ATT-style object, but the expectation is taken over treated unit–period cells rather than over treated units. The formal analysis in Athey et al. [2025b] is also developed under strong restrictions that are often questionable in marketing settings. In particular, it assumes no interference (no spillovers into donors) and no dynamic effects, so that potential outcomes can be written as $Y_{it}(d)$ rather than depending on the treatment path $d_{ti}$ . We use the paper’s results as a guide to how bias behaves inside this restricted model class, not as a mechanical guarantee in settings with carryover or interference. TROP does not escape the need for structure. It replaces the additive parallel trends assumption that underlies DiD and SDID with a factor-model assumption for untreated potential outcomes. We state this assumption here for reference and refer you back to Chapter 6 for a full discussion. Assumption 19 (Factor Model for Untreated Potential Outcomes) For all units $i$ and periods $t$, untreated outcomes satisfy:

$$Y_{it}(0) = \alpha_i + \lambda_t + \sum_{r=1}^{R} \lambda_{ir} f_{tr} + \varepsilon_{it}$$$$E[\varepsilon_{it} | \{\lambda_{ir}, f_{tr}\}_{r=1}^{R}] = 0$$

where $\alpha_i$ are unit fixed effects, $\lambda_t$ are time fixed effects, $f_{tr}$ are latent factors and $\lambda_{ir}$ are unit-specific factor loadings, and $\varepsilon_{it}$ is idiosyncratic noise. The low-rank matrix with elements $L_{it} = \sum_{r=1}^{R} \lambda_{ir} f_{tr}$ captures interactive fixed effects. This factor structure is more flexible than additive parallel trends but more restrictive than a fully nonparametric model. It allows units to load differently on common shocks – something we often see in marketing panels – but still assumes that a small number of latent factors explain most of the systematic co-movement in outcomes.

Theorem 7.1 (Triple robustness bias bound; Athey et al. [2025b]) Under Assumption 19, suppose unit weights $w$ and pre-treatment time weights $v$ (both non-negative and summing to one) are fixed. Let $\boldsymbol{\lambda}_i = (\lambda_{i1}, \ldots, \lambda_{iR})^\prime$ denote the factor-loading vector for unit $i$ and let $\mathbf{f}_t = (f_{t1}, \ldots, f_{tR})^\prime$ denote the factor vector at time $t$. Define unit and time imbalance terms:

$$\Delta_u = \left\| \sum_{j \in J} w_j \boldsymbol{\lambda}_j - \boldsymbol{\lambda}_{i^*} \right\|_2$$$$\Delta_t = \left\| \sum_{s \in T_{\text{pre}}} v_s \mathbf{f}_s - \mathbf{f}_{t^*} \right\|_2$$

for a target treated unit $i^*$ and target post-treatment period $t^*$. If the regression adjustment for the low-rank component has bias controlled by an operator $B$ with operator norm $\|B\|_{\text{op}}$, then the conditional bias of the resulting counterfactual contrast is bounded by:

$$E[\hat{\tau} - \tau | L] \leq \|\Delta_u\|_2 \|\Delta_t\|_2 \|B\|_{\text{op}}$$

In the paper’s formal analysis, $B$ captures systematic shrinkage in the estimated low-rank component (for example, bias induced by nuclear-norm regularisation or rank truncation), and $\| \cdot \|_{\text{op}}$ is the operator (spectral) norm.

7.5 Triply Robust Panel (TROP) Estimators

The theorem clarifies what “triply robust” does and does not mean. It does not say that arbitrary modelling errors cancel out. It says that, under a shared factor structure, the leading bias term is small when at least one of the three components is small: unit imbalance, time imbalance, or regression-adjustment misspecification. This is the sense in which TROP is “triply robust.” It does not mean that you can be cavalier about all three ingredients. It means that you now have three levers which, if any one of them is correctly tuned relative to the underlying factor structure, can salvage identification. In practice, all three will be estimated with error and all three may be misspecified to some degree. The value of TROP is that small mistakes in several places multiply rather than add, so that moderate imbalance and moderate model error can still produce modest bias.

Why Factor Structure Matters in Marketing Panels The factor model is not an abstract technicality. It is designed to capture a pattern we routinely see in marketing data: outcomes co-move because of common shocks, but different units respond with different intensity. Consider again a national recession in a campaign evaluation. Under additive parallel trends, every market experiences the same time effect $\lambda_t$, so recession quarters shift all units by the same amount. That is implausible. Affluent DMAs with high discretionary spending might see a sharp fall in premium-category sales. Industrial regions hit by plant closures see a different pattern again. Budget markets where consumers trade down to cheaper brands may see smaller or even opposite shifts in category volume. These heterogeneous responses are exactly what cause additive DiD and SDID designs to struggle. In a factor framework, the recession is represented by a common factor $f_t$ that spikes during downturn quarters. Affluent, industrial and budget DMAs carry different loadings $\lambda_i$ on that factor. A treated DMA’s untreated path during the recession therefore depends on both the factor and its own loading. TROP’s lowrank component aims to learn these heterogeneous loadings from the donor panel, while its unit and time weights try to align the treated DMA with a weighted combination of donors that shares a similar factor profile. When this works, the method can deliver credible counterfactuals in settings where additive two-way fixed effects are clearly inadequate. The same story applies to seasonality, category trends and platform shocks in digital advertising. Common shocks exist, but units react differently. TROP’s model structure is built to capture exactly that pattern.

Estimator Sketch and Relation to Existing Methods At a high level, TROP works by fitting a two-way fixed-effects plus factor model to the donor panel, while reweighting both units and time periods to focus on observations that are most informative for a given treated unit and treatment period. In the most general formulation in Athey et al. [2025b], the weights can depend on the target treated cell (i, t). For exposition, we suppress this dependence and describe the three components in a simpler global-weight form. Let $D_{it}$ denote treatment, $Y_{it}$ the outcome and assume we observe a panel with many untreated (i, t) pairs. The estimator proceeds in three steps. First, it defines unit weights $w_j$ that give more weight to donors whose pre-treatment paths lie close to the treated unit’s path. Conceptually, these play the same rôle as the unit weights in SC and SDID. Distance can be measured by root mean squared differences in pre-treatment outcomes, much as in synthetic control. A tuning parameter controls how sharply the weights decay with distance; when the parameter is zero the weights become uniform and the method behaves like a global factor model, whereas large values make the estimator focus on close neighbours. Second, it defines time weights $v_s$ that give more weight to pre-treatment periods close to the treatment date. This reflects the idea that recent history is often more informative about current behaviour than distant history. A separate tuning parameter controls how quickly the weights decay as you move away from the treatment date. When this parameter is zero all pre-treatment periods are weighted equally. When it is large the estimator behaves more like a local time-differencing design. Third, given these weights, it fits a low-rank factor model to the untreated panel by solving a nuclear-norm-regularised least-squares problem over unit and time fixed effects and the interactive component. The nuclear norm penalty controls the effective rank of the factor structure and is tuned by cross-validation. The fitted model yields predictions of untreated outcomes for any unit and time, including treated units in post-treatment periods. The TROP treatment effect estimate for unit i in period t is the difference between the observed outcome and this predicted counterfactual. This framework connects to several familiar designs. If you set all weights equal and choose a large rank, you obtain something close to matrix completion on the donor panel. If you turn off the factor component and keep only unit and time fixed effects with data-driven weights, you obtain an estimator similar in spirit to SDID. If you remove time weighting and focus on unit weighting and a simple outcome model, you move back towards ASCM. TROP is best thought of as a unifying language and a flexible class of estimators rather than a single closed-form formula.

Design Choices and Tuning In applied work you face three main design choices with TROP: how sharply to concentrate unit weights, how local to make time weights, and how complex to make the factor structure. Athey et al. [2025b] propose choosing these tuning parameters by cross-validation on untreated cells, with an objective designed so that pseudo treatment effects on untreated cells are close to zero. This tuning logic is appealing because it directly targets the imputation problem that drives TROP. Good predictive performance on held-out untreated outcomes is necessary for credible identification but not sufficient. Cross-validation cannot, on its own, guarantee that the factor structure and weighting scheme satisfy the causal assumptions needed for post-treatment counterfactuals. In practice, you would define a grid of candidate tuning parameters for unit weights, time weights and the nuclear norm penalty. For each combination you fit the model on a training subset of untreated observations

7.5 Triply Robust Panel (TROP) Estimators and measure prediction error on a validation subset. You then select the parameters that minimise this error. If the data exhibit strong interactive fixed effects, cross-validation will favour a non-trivial factor structure. If additive two-way fixed effects suffice, it will push the nuclear norm penalty high and effectively shut down the factor component. If all history is informative, it will choose slowly decaying time weights. If only recent periods matter, it will pick fast decay. This tuning mechanism is conceptually appealing because it lets the data decide which components matter most, rather than forcing the analyst to commit ex ante to SC, SDID or matrix completion. It is also computationally intensive. Each tuning combination requires solving a large convex optimisation problem. For small-to-medium marketing panels (say a few dozen units over a few dozen periods) this is feasible with modern optimisation toolkits. For very large panels, staged or random-search strategies are needed to keep computation reasonable.

When to Use TROP TROP is an ambitious method. It makes sense in applications where you have good reason to suspect interactive fixed effects, where the sample is large enough to estimate a low-rank structure and where simpler methods struggle to achieve credible pre-treatment fit. Marketing panels with many units and moderate time depth are a natural candidate. Think of 40 DMAs followed for 20 quarters, with rich common shocks and clear heterogeneity in how markets respond. In such settings, standard SC may fail because the treated unit sits outside the convex hull of donors, ASCM may struggle because a simple regression cannot capture the full pattern of co-movement and SDID may fail because additive time effects are too crude. If, in that context, a carefully tuned TROP model achieves substantially better pre-treatment fit than SC, ASCM or SDID while using plausible weights and a modest factor rank, it deserves serious consideration. By contrast, when the donor pool is small or the time dimension is short, the extra complexity of TROP is unlikely to pay off. Estimating a factor structure with only eight pre-treatment months or with only five control markets is more art than science. In those cases, you are better off with the simpler designs developed earlier in the chapter, which are easier to estimate and explain. At the time of writing, TROP is a frontier method. The underlying research is at preprint stage and stable, well-documented software implementations are not yet widely available. Implementing it today requires custom optimisation code, typically built on general-purpose tools such as cvxpy in Python or convex optimisation libraries in R. That makes it more suitable for methodological work or for teams with strong in-house econometrics capability than for routine marketing analytics.

Positioning TROP in the Hybrid Methods Hierarchy From a strategic point of view, the main value of TROP for this book is conceptual. It shows how weighting, differencing and factor modelling can be combined in a single framework and clarifies what “triple robustness” means in panels. It also gives you a way to think about why SC, ASCM, SDID and matrix completion succeed or fail in particular designs. For example, SC failures typically reflect convex-hull and unit-imbalance problems, ASCM failures reflect regression misspecification on top of those, SDID failures reflect violations of (weighted) parallel trends, and matrix-completion failures reflect poor low-rank approximation. In empirical marketing work, the right approach is to treat TROP as one more estimator in a small ensemble. You start with methods that are transparent and well understood – SC, ASCM, regularised SC and SDID – and only move to TROP when those methods either fail to fit the pre-treatment data or produce estimates that are clearly at odds with economic common sense. When you do estimate TROP, you should present it alongside these simpler benchmarks, explain what its tuning parameters are doing and show how its pre-treatment diagnostics compare. Agreement across methods strengthens confidence in your conclusions. Disagreement is a signal to dig deeper, not a licence to cherry-pick the method that tells the most appealing story. In that sense, TROP completes the conceptual hierarchy rather than replacing the tools you already have. It is a powerful idea – and potentially a powerful estimator – but its practical role in marketing analytics will depend on how the method and its software ecosystem evolve beyond the time this book is written.

7.6 Identification and Assumptions

References

  • Shaw, C. (2025). Causal Inference in Marketing: Panel Data and Machine Learning Methods (Community Review Edition), Section 7.5.