Psychology dissertation targeted maximum likelihood estimation helps researchers estimate a defined causal effect while combining flexible prediction with principled statistical inference.
Usually shortened to TMLE, the framework starts with estimates of the outcome process and exposure mechanism, then updates the outcome estimate toward the parameter the dissertation actually seeks. This targeting step distinguishes TMLE from simply fitting a machine-learning model and reporting its predictions.
TMLE is powerful, but it does not turn observational data into a randomised experiment. A defensible application still depends on a clear estimand, coherent time zero, appropriate baseline covariates, adequate positivity, reproducible nuisance modelling, valid uncertainty estimation, and honest sensitivity analysis. This guide shows how to plan those elements for a psychology dissertation.
What targeted maximum likelihood estimation does
TMLE is a substitution estimator. It produces an updated probability distribution or regression estimate and then substitutes that estimate into the formula for the target parameter. For a binary exposure, a common target is the average treatment effect, defined as the population mean outcome under exposure minus the population mean outcome under comparison.
The Targeted Learning in R handbook explains that TMLE begins with an outcome regression and a propensity-score estimate. It then updates the initial outcome regression so it is targeted to the chosen parameter. The resulting estimate can support influence-function-based standard errors and confidence intervals under the required conditions.
For a psychology study, the exposure could be enrolment in an optional wellbeing programme and the outcome could be follow-up distress. Baseline covariates might include prior distress, therapy history, medication, financial strain, study load, sleep, disability accommodations, digital access, and help-seeking attitudes.
The initial outcome model
The outcome regression predicts the outcome from exposure and baseline covariates. The researcher uses it to predict each participant’s outcome under both exposure levels, regardless of the exposure actually observed. Averaging those counterfactual predictions gives an initial plug-in estimate.
The exposure model
The exposure model estimates the probability of each participant receiving the observed exposure conditional on baseline covariates. These probabilities help form the clever covariate used in the targeting step. Scores close to zero or one can make the update unstable and reveal weak empirical positivity.
The targeted update
The update fits a low-dimensional fluctuation through the initial outcome estimate using information from the exposure mechanism. Its purpose is not to rebuild the entire predictive model. It corrects the estimate in a direction linked to the efficient influence function for the specified parameter.
| Component | Question answered | Dissertation evidence |
|---|---|---|
| Outcome regression | What outcome is expected under each exposure and covariate pattern? | Learners, calibration, residual or risk diagnostics |
| Exposure mechanism | How probable was each observed exposure? | Score distributions, overlap, influential observations |
| Target parameter | Which population contrast is being estimated? | Population, strategies, outcome, follow-up and scale |
| Targeting step | How is the initial estimate updated toward that parameter? | Loss, link, clever covariate and convergence |
| Influence function | How are uncertainty and diagnostics calculated? | Standard error, interval, clustering and mean EIF |
When psychology dissertation targeted maximum likelihood estimation fits
TMLE can fit an observational dissertation with a clearly defined point exposure, measured baseline confounders, a continuous or binary outcome, sufficient overlap, and a marginal causal estimand. It can also be used in randomised studies to improve precision and handle prespecified adjustment, although implementation and inference should reflect the trial design.
Potential psychology applications include voluntary counselling uptake, workplace mental health programmes, digital interventions, school-based support, parenting programmes, peer mentoring, treatment engagement, and naturally chosen learning formats. The point-exposure version is most manageable when exposure occurs once at a defensible baseline and the outcome follows within a fixed period.
Do not apply a simple point-treatment TMLE to repeated treatment decisions, treatment-confounder feedback, informative censoring, interference, competing events, or complex mediation without a version designed for that data structure. The CRAN tmle package documentation describes point-treatment effects for binary exposure and continuous or binary outcomes, while longitudinal settings require other estimators and software.
TMLE is estimation, not identification
Machine learning can reduce reliance on a single parametric form, but it cannot identify a causal effect from an incoherent design. A sophisticated learner cannot repair an omitted confounder, reverse temporal ordering, an ill-defined intervention, selection bias, or exposure levels unsupported by the data.
Define the causal question before the software
Write the target population, eligibility criteria, exposure strategies, time zero, outcome, follow-up, and effect measure before selecting learners. “The effect of therapy” is too vague. A better question compares two specified strategies, such as offering six sessions of guided online cognitive behavioural support versus information-only support among eligible first-year students, with anxiety measured twelve weeks after enrolment.
Choose the estimand deliberately. The average treatment effect concerns the whole eligible population. The average treatment effect among the treated concerns those who received the exposure. A risk difference, risk ratio, or mean difference may answer different practical questions. Avoid switching scales after seeing which result appears strongest.
Align time zero
Eligibility, baseline covariate measurement, and exposure assignment should align at the same conceptual start. Variables measured after enrolment, including adherence, intermediate symptoms, later medication, attendance, and dropout, are not baseline confounders. Conditioning on them in a point-treatment analysis can change the estimand or introduce bias.
State the assumptions
Consistency requires each exposure strategy to correspond to the potential outcome being discussed. Conditional exchangeability requires no unmeasured confounding after adjustment for the chosen baseline covariates. Positivity requires a non-zero probability of each exposure strategy within relevant covariate patterns. The statistical models must also estimate the necessary nuisance functions adequately.
These assumptions should appear in the methodology and limitations chapters. Balance plots and predictive diagnostics provide useful evidence, but they cannot prove exchangeability or consistency.
Select baseline covariates with causal reasoning
Use a directed acyclic graph, subject knowledge, prior research, and temporal ordering to select variables. Include measured causes of both exposure and outcome and strong baseline outcome predictors. Do not rely on automated p-value selection or choose variables only because they improve exposure prediction.
For a voluntary anxiety programme, plausible baseline covariates include initial anxiety and depression, prior counselling, medication, age, study level, workload, sleep, financial strain, referral route, digital access, disability accommodations, and help-seeking attitudes. The exact set should follow the causal question and available measurement.
Avoid post-exposure mediators and colliders. Be cautious with variables that almost determine exposure but do not predict the outcome, because they may worsen practical positivity. Document construct definitions, measurement dates, coding, transformations, and reasons for inclusion.
Plan missing-data handling
Describe missingness separately for baseline covariates, exposure, and outcome. Complete-case analysis may alter the target population and create selection bias. Multiple imputation can be appropriate, but the entire TMLE workflow should be run within each imputed dataset and combined using a justified procedure.
If outcome observation depends on measured history, a TMLE that incorporates the censoring or observation process may be needed. Do not silently treat missing outcomes as ordinary values or assume that machine learning resolves missingness.
Choose nuisance learners that match the data
TMLE can use parametric regressions or flexible learners for the outcome and exposure mechanisms. The aim is to estimate these nuisance functions well enough for the target parameter, not to win a prediction competition.
A defensible learner library might include an intercept-only benchmark, main-effects regression, penalised regression, spline-based models, and a carefully tuned tree-based learner. Include simpler learners because they can perform well in modest samples and provide stability. Preprocess variables within training folds to prevent information leakage.
The official tmle reference manual documents estimation of the outcome and exposure components and encourages data-adaptive estimation through SuperLearner. Software defaults are not a substitute for a prespecified library that suits the sample size, variable types, and exposure frequency.
Use cross-validation and cross-fitting appropriately
Cross-validation selects or combines learners according to out-of-sample risk. Cross-fitting separates nuisance-model training from prediction for the observations contributing to the estimating equation. Cross-validated TMLE can reduce overfitting bias when highly adaptive learners are used.
Report the number of folds, fold construction, repetition, random seed, learner library, tuning, and failure handling. Every training fold needs enough exposed and comparison participants. When observations are clustered within therapists, schools, families, or sites, fold construction should prevent dependent units leaking across training and validation sets.
| Decision | Good practice | Warning sign |
|---|---|---|
| Learner library | Small, diverse, prespecified and sample-appropriate | Many algorithms added without justification |
| Preprocessing | Estimated inside training folds | Whole-data scaling or imputation before folds |
| Fold assignment | Preserves exposure counts and dependence structure | Empty exposure group or cluster leakage |
| Tuning | Nested or otherwise protected from outcome reuse | Manual tuning toward preferred effects |
| Reproducibility | Seeds, versions, warnings and code recorded | Only the final estimate retained |
Run the targeting step carefully
For a binary outcome, the initial outcome prediction is often updated on the logit scale so fitted values remain between zero and one. The clever covariate is constructed from exposure probabilities and the target contrast. A fluctuation coefficient is then estimated, usually with the initial prediction entered as an offset.
After the update, predict each participant’s outcome under each exposure level using the targeted outcome regression. Average these predictions across the target population and contrast the averages on the chosen scale. The updated predictions, rather than the fluctuation coefficient by itself, produce the TMLE.
Check that the targeting model converged, the update is numerically sensible, predicted outcomes remain in their valid range, and the empirical mean of the estimated efficient influence function is close to zero. Record any truncation or bounding of exposure probabilities.
Understand double robustness
Point-treatment TMLE has a double-robustness property under appropriate assumptions and regularity conditions: consistency can be retained if either the outcome regression or exposure mechanism is estimated correctly. It does not mean two misspecified models cancel each other, and it does not protect against unmeasured confounding or severe positivity violations.
This distinction is important because the previous AIPW guide also describes a doubly robust estimator. AIPW uses an estimating-equation correction directly, whereas TMLE updates the outcome distribution and remains a substitution estimator. Their first-order behaviour can be closely related, but the algorithms and finite-sample behaviour are not identical.
Diagnose positivity and influential observations
Plot estimated exposure probabilities by observed group. Report minima, maxima, relevant percentiles, and the proportions below or above prespecified thresholds. Investigate combinations of covariates for which one exposure is rare or absent.
Near-zero probabilities create large clever-covariate values and can concentrate influence in a few participants. Truncating probabilities may improve variance but introduces a bias-variance trade-off. Prespecify the primary rule, explain it, and compare defensible thresholds as sensitivity analyses rather than choosing the threshold that produces significance.
If poor overlap is structural, consider redefining eligibility or the target population. A more limited but supported causal question is preferable to an unstable estimate for a population that the data cannot represent.
Check both nuisance components
For the outcome model, assess calibration, residual patterns, plausible ranges, and performance within exposure groups. For the exposure model, assess overlap and whether weighting implied by the clever covariate creates extreme influence. Predictive accuracy alone is insufficient because a high classification score can coexist with poor overlap.
Estimate uncertainty that matches the design
The efficient influence function commonly supplies a standard error for an asymptotically linear TMLE. Report the point estimate, standard error, confidence interval, effect scale, target population, and follow-up period.
The peer-reviewed step-by-step TMLE tutorial indexed by PubMed illustrates implementation under near-positivity problems and emphasises the connection between assumptions, estimation, and practical diagnostics. Such guidance supports the workflow but does not validate a dissertation’s particular data.
Independent-observation standard errors are inappropriate when the design contains material clustering. Use cluster-aware influence-function aggregation or resampling that matches the sampling unit. Small samples may also need cautious finite-sample interpretation because asymptotic intervals can perform poorly.
| Problem | Likely consequence | Reasonable response |
|---|---|---|
| Weak overlap | Large updates and unstable intervals | Revisit population, exposure definition and truncation plan |
| Overfitted nuisance models | Biased estimates or invalid inference | Use protected prediction and suitable cross-fitting |
| Both models misspecified | Double robustness no longer protects consistency | Expand justified learners and compare specifications |
| Clustered observations | Standard errors may be too small | Use cluster-aware folds and inference |
| Unmeasured confounding | Causal interpretation may be biased | Quantitative sensitivity analysis and cautious claims |
| Outcome-driven analysis choices | Selective reporting and optimistic inference | Preregister rules and retain a decision log |
Compare TMLE with related estimators
G-computation standardises predictions from an outcome model. TMLE begins with this type of plug-in estimate and adds a parameter-specific targeting step informed by the exposure model.
Inverse probability weighting relies mainly on the exposure mechanism and can be highly sensitive to extreme weights. TMLE uses exposure probabilities in the update but also incorporates an outcome regression.
AIPW and TMLE are both doubly robust for common point-treatment parameters under suitable conditions. TMLE’s substitution structure can respect natural parameter bounds, while AIPW’s direct correction is often simpler to explain. Method choice should be planned, not based on which estimate is most favourable.
Sensitivity analyses that add information
Compare a prespecified primary learner library with simpler parametric nuisance models. Vary reasonable truncation thresholds, fold assignments, and learner sets. Present outcome regression, weighting, AIPW, or g-computation estimates when they answer the same estimand and help diagnose dependence on modelling choices.
Assess missing-data assumptions, alternative defensible exposure definitions, influential observations, and measurement decisions. Add a quantitative sensitivity analysis for unmeasured confounding that matches the effect measure. Agreement across estimates is reassuring but never proves that causal assumptions hold.
Keep a structured log of every specification, diagnostic, warning, software version, random seed, and reason for retaining or rejecting a model. Do not run many variants and display only those that support a preferred conclusion.
A reproducible dissertation workflow
- Define eligibility, target population, exposure strategies, time zero, outcome, follow-up and estimand.
- Draw the causal diagram and justify every baseline adjustment variable.
- Audit missingness, measurement, exposure counts, clustering and overlap.
- Prespecify the outcome and exposure learner libraries, folds and truncation rule.
- Generate protected nuisance predictions using the planned validation or cross-fitting scheme.
- Inspect calibration, overlap, clever-covariate values and influential observations.
- Run the parameter-specific targeting step and confirm numerical convergence.
- Calculate targeted potential-outcome means, the contrast and valid uncertainty.
- Run prespecified robustness, missing-data and unmeasured-confounding analyses.
- Report code, versions, seeds, diagnostics, estimates, assumptions and limitations.
Frequently asked questions
Is TMLE a machine-learning algorithm?
No. TMLE is a general estimation framework that can incorporate machine-learning algorithms for nuisance functions. The targeting step connects those predictions to a defined statistical parameter and its influence function.
What does “targeted” mean in TMLE?
It means the initial outcome estimate is updated in a direction designed for the specific parameter, such as an average treatment effect, rather than optimised only for overall prediction.
Does TMLE remove unmeasured confounding?
No. Observational causal interpretation still requires conditional exchangeability based on measured covariates. Sensitivity analysis can explore plausible departures but cannot create missing information.
Do I need Super Learner?
Not always. Parametric nuisance models can be used, but a prespecified ensemble can reduce reliance on a single functional form when sample size and computing resources support it.
Is TMLE always better than AIPW?
No. Both can be efficient and doubly robust for common parameters. Performance depends on the estimand, nuisance estimates, positivity, implementation, sample size and inference procedure.
Can TMLE analyse longitudinal exposures?
Yes, but not with a simple point-treatment implementation. Longitudinal TMLE must represent treatment, time-varying confounding, censoring and the intervention history correctly.
What should I report in the methods chapter?
Report the causal question, estimand, timeline, assumptions, covariates, missing-data method, nuisance learners, preprocessing, folds, targeting specification, positivity diagnostics, truncation, influence-function inference, software and sensitivity analyses.
Conclusion
Psychology dissertation targeted maximum likelihood estimation is most useful when a clear causal question can support flexible nuisance modelling and a parameter-specific update. Its strengths include substitution estimation, double robustness, machine-learning compatibility, and influence-function inference. Those strengths depend on careful design, overlap, protected prediction, diagnostics, and transparent reporting.
If you need support, seek ethical methodological guidance that helps you define the estimand, review code, interpret diagnostics, and document decisions while you retain data ownership and authorship. Responsible assistance should strengthen your independent understanding, never promise significance or invent results.
