Psychology dissertation double machine learning combines flexible nuisance prediction, orthogonal scores, and cross-fitting to estimate a defined low-dimensional causal effect.
Usually shortened to DML, the framework addresses a difficult problem: machine-learning models can capture complex relationships between baseline covariates, exposure, and outcome, yet regularisation and overfitting can bias a treatment-effect estimate when predictions are inserted naively into a causal equation. To learn more , see our statistics and data analysis for psychology dissertations hub.
DML reduces that sensitivity through Neyman orthogonality and protected sample splitting. It does not automate causal identification. A credible psychology dissertation still needs a coherent estimand, defensible time zero, measured confounders, overlap, suitable learners, reproducible folds, appropriate inference, and honest limitations. This guide turns those requirements into a practical workflow.
Table of Contents
What double machine learning estimates
Double machine learning is a semiparametric framework for estimating a target parameter while allowing high-dimensional or complex nuisance functions. The target might be an average treatment effect, an average treatment effect among exposed participants, or a coefficient in a partially linear model. Nuisance functions describe relationships needed to estimate that target but are not themselves the main research result.
The foundational double/debiased machine learning paper identifies two central ingredients: Neyman-orthogonal scores that reduce first-order sensitivity to nuisance-estimation error, and cross-fitting that separates nuisance training from evaluation. Together they permit flexible prediction while supporting inference under stated conditions.
For a psychology example, imagine estimating the effect of voluntary therapist-guided digital support on twelve-week anxiety. Baseline severity, depression, previous therapy, medication, study load, sleep, financial strain, disability accommodations, digital access, and help-seeking attitudes may predict both programme uptake and outcome.
The target is not prediction accuracy
A learner may predict anxiety extremely well without estimating the programme effect well. DML uses prediction to remove or adjust nuisance structure, then estimates a low-dimensional parameter through an orthogonal score. The final estimate, standard error, and confidence interval answer the research question; the nuisance predictions support that calculation.
| Element | Purpose | Evidence to report |
|---|---|---|
| Target parameter | Defines the population contrast | Population, exposure, outcome, follow-up and scale |
| Outcome nuisance model | Predicts outcome from exposure and covariates | Learners, tuning, calibration and errors |
| Exposure nuisance model | Predicts exposure from baseline covariates | Probabilities, overlap and classification diagnostics |
| Orthogonal score | Reduces local sensitivity to nuisance error | Score formula and model class |
| Cross-fitting | Prevents own-observation overfitting bias | Folds, repetitions, grouping and seeds |
| Inference | Quantifies sampling uncertainty | Standard error, interval and clustering method |
When psychology dissertation double machine learning fits
DML can be useful when the dissertation has a well-defined exposure, outcome, target population, and estimand, but relationships involving measured covariates may be nonlinear, interactive, or high-dimensional. It is especially relevant when a simple adjustment model would impose a questionable functional form.
Suitable settings may include voluntary counselling, uptake of a digital intervention, school wellbeing programmes, workplace mental health support, parenting programmes, peer mentoring, and naturally selected learning formats. The method must match the exposure type, outcome type, and intended effect.
The official DoubleML model documentation distinguishes partially linear regression, interactive regression, instrumental-variable, and related models. These are not interchangeable software options. Each represents a different structural model and target parameter.
When DML is not a good default
A conventional regression may be clearer when the sample is modest, the covariate set is small, functional forms are well justified, and flexible learners would be unstable. A point-exposure DML is also inappropriate for treatment-confounder feedback, interference, poorly measured timing, severe overlap failure, or unaddressed censoring.
Do not select DML because the term sounds advanced. Choose it when its model, orthogonal score, nuisance functions, and sample requirements answer the actual research question.
Define the estimand and time zero first
Specify eligibility, target population, exposure strategies, time zero, outcome, follow-up period, and effect scale before choosing algorithms. “Does online therapy improve wellbeing?” is underspecified. A clearer question compares assignment or uptake of a defined programme with a defined comparison among an eligible population over a fixed period.
Distinguish an average treatment effect for the eligible population from an effect among participants who actually selected the programme. A partially linear coefficient also carries structural assumptions about how exposure enters the outcome model. It should not be casually labelled an ATE when heterogeneity or nonlinearity changes its meaning.
Align baseline and follow-up
Eligibility, covariate measurement, and exposure assignment should share a coherent start. Adherence, attendance, later symptoms, subsequent medication, and dropout occur after exposure and should not be inserted as baseline confounders. Doing so can block part of the effect or introduce collider bias.
State the identification assumptions
For observational treatment-effect estimation, consistency links the specified exposure to potential outcomes. Conditional exchangeability requires no unmeasured confounding after adjustment for the chosen baseline covariates. Positivity requires a non-zero chance of each exposure level across relevant covariate patterns.
DML cannot test or guarantee these assumptions. Orthogonal scores reduce sensitivity to nuisance-estimation errors, not to missing confounders, measurement error, selection bias, or ill-defined interventions.
Choose the correct DML model class
The partially linear regression model represents the outcome as a linear contribution of exposure plus an unknown function of covariates. A second nuisance model predicts exposure from those covariates. The target is the coefficient on exposure after residualising both processes.
This structure may suit a continuous or meaningfully scaled exposure when a constant partially linear effect is defensible. For example, the exposure might be number of allocated coaching sessions if allocation can plausibly be interpreted under the model. It is not automatically appropriate for a binary programme where effects may vary strongly across participants.
Interactive regression for binary exposure
An interactive regression model permits the conditional outcome function to differ flexibly by exposure group. For a binary programme, it can target an ATE or ATT under unconfoundedness. It commonly estimates outcome regressions under both exposure levels and an exposure propensity. Visit Statistics and data analysis hub to learn more.
Researchers must confirm that their chosen software score targets the desired effect. A partialling-out score in a partially linear model, an ATE score in an interactive regression model, and an instrumental-variable score answer different questions.
Instrumental-variable DML
DML also has instrumental-variable formulations, but machine learning does not make an instrument valid. Relevance, exclusion, independence, and any monotonicity assumptions remain substantive. A weak or invalid instrument cannot be rescued by flexible nuisance fitting.
| Model class | Typical target | Main caution |
|---|---|---|
| Partially linear regression | Low-dimensional exposure coefficient | Partially linear structure must be meaningful |
| Interactive regression | ATE or ATT for binary exposure | Needs overlap and measured confounding control |
| Partially linear IV | Coefficient with an instrument | Instrument validity and strength remain essential |
| Interactive IV | Local average treatment effect | Targets compliers under additional assumptions |
| Heterogeneous-effect extension | Conditional or group effects | Greater complexity and multiplicity |
Select covariates using a causal model
Use a causal diagram, subject knowledge, temporal ordering, and prior evidence. Include measured common causes of exposure and outcome and strong baseline outcome predictors. Do not select variables only because a learner assigns them high importance.
For the digital-support example, baseline anxiety, depression, therapy history, medication, study load, sleep, financial strain, accessibility needs, referral pathway, digital access, and help-seeking attitudes may be plausible. Each variable needs a measurement date and causal justification.
Avoid mediators, descendants of exposure, and colliders. Variables that almost determine exposure can worsen practical positivity. High dimensionality does not excuse indiscriminate adjustment for every available field.
Handle missing data inside the workflow
Describe missingness separately for covariates, exposure, and outcome. If multiple imputation is defensible, repeat the complete DML procedure within each imputed dataset and combine estimates using a method compatible with the analysis and imputation strategy.
Preprocessing, including imputation, scaling, feature selection, and encoding, must be estimated within training folds. Performing these steps on the whole dataset before cross-fitting leaks validation information into nuisance training.
Build a defensible learner library
Machine learning in DML estimates nuisance functions, not the target effect directly. Candidate learners can include main-effects regression, penalised regression, generalised additive models, tree ensembles, boosting, or carefully chosen neural models. The library should match the sample size, outcome type, exposure type, and computational resources.
Include simple benchmarks. Complex learners can fail in small exposure groups, overfit rare outcomes, or produce extreme probabilities. A diverse but compact library is usually more defensible than dozens of algorithms chosen for appearance.
The peer-reviewed Journal of Statistical Software article on DoubleML describes three core ingredients: Neyman orthogonality, high-quality machine-learning estimation, and sample splitting. It also documents partially linear and interactive models and their instrumental-variable extensions.
Tune without contaminating the target estimate
Hyperparameter tuning should occur within training data or a clearly protected nested procedure. Do not inspect the treatment-effect estimate while tuning nuisance models. Record search spaces, scoring rules, resampling, random seeds, failed learners, and fallback rules.
Understand orthogonal scores
A Neyman-orthogonal score is constructed so that small errors in nuisance functions have no first-order effect on the target estimating equation at the truth. This local insensitivity helps prevent regularisation bias from passing directly into the effect estimate.
Orthogonality is not immunity. Large nuisance errors, weak overlap, an incorrect score, model-class violations, and inadequate convergence can still bias estimates. The product of nuisance-estimation errors must generally become small enough for the desired asymptotic result.
Residualisation in a partially linear model
A common partialling-out approach predicts the outcome from covariates and separately predicts exposure from covariates. It then relates the outcome residual to the exposure residual using an orthogonal score. This removes flexible covariate structure while keeping the exposure parameter low-dimensional.
The residuals must be out-of-fold. If a flexible model predicts the same observations used for its training, the residuals can be artificially small and the effect estimate can inherit overfitting bias.
Run cross-fitting correctly
Split participants into folds. Train every nuisance model on all but one fold, then predict the held-out fold. Repeat until each participant has out-of-fold predictions. Calculate score contributions from those protected predictions and aggregate them into the target estimate.
The DoubleML guide to DML basics explains how sample splitting reduces overfitting bias and how cross-fitting reverses training and test roles so the full sample contributes to estimation.
Every training fold needs adequate exposed and comparison observations. Stratify folds when appropriate. If participants are clustered within therapists, schools, clinics, families, or sites, keep dependent units together to avoid leakage.
Repeat cross-fitting when feasible
A single random split can create noticeable Monte Carlo variation, especially in smaller samples. Repeated cross-fitting can show whether the result depends on a particular partition. Prespecify how repeated estimates and variance components will be combined.
| Diagnostic | Question | Potential repair |
|---|---|---|
| Fold exposure counts | Does every training set contain both groups? | Stratify, reduce folds, or simplify the question |
| Nuisance performance | Are out-of-fold predictions credible? | Revise justified learners or tuning |
| Propensity distribution | Are probabilities near zero or one? | Reconsider eligibility, support or trimming plan |
| Split stability | Do estimates vary across fold seeds? | Repeat cross-fitting and report variation |
| Score contributions | Do a few observations dominate? | Investigate data, overlap and influence |
| Cluster structure | Are dependent units separated across folds? | Use group-level splitting and inference |
Diagnose overlap and nuisance performance
For binary exposure, plot out-of-fold propensities by observed group and report their range and relevant percentiles. Inspect whether important covariate profiles appear in both groups. Extreme probabilities can inflate score contributions and reveal unsupported comparisons.
Assess outcome prediction with an appropriate out-of-fold loss, calibration, and subgroup checks. Assess exposure prediction with calibration and probability distributions rather than classification accuracy alone. Perfect exposure classification is usually a warning about positivity, not a success.
Do not optimise prediction metrics without considering the score and estimand. Nuisance performance matters because it supports the causal estimate, but a marginal improvement in prediction does not automatically improve inference.
Inspect influential score contributions
Summarise the distribution of observation-level scores or influence values. Investigate extreme contributions, data errors, sparse covariate patterns, and learner extrapolation. Any exclusion rule must be prespecified or clearly labelled as sensitivity analysis.
Estimate uncertainty that matches the design
DML commonly estimates a standard error from the empirical variability of orthogonal score contributions. Report the effect estimate, standard error, confidence interval, target population, outcome scale, exposure contrast, and follow-up period.
Clustered observations need cluster-robust score aggregation or another procedure aligned with the sampling structure. Repeated measures, survey weights, multiple treatments, and multiple target parameters also require appropriate extensions.
Asymptotic normality does not guarantee good small-sample performance. Sparse exposure groups, unstable learners, weak overlap, many folds, and heavy-tailed scores can make intervals unreliable. A simulation based on the planned sample size and plausible data structure can be more informative than a generic minimum-sample rule.
Compare DML with nearby methods
Targeted maximum likelihood estimation also combines flexible nuisance estimates with an orthogonal or efficient influence-function structure, but TMLE updates an initial likelihood and remains a substitution estimator. DML is a broader score-based framework with common residualisation and interactive-regression implementations.
Augmented inverse probability weighting directly combines outcome predictions and inverse-probability residual corrections. An interactive-regression DML score for an ATE can be closely related, with cross-fitting playing a central role in DML implementation.
Conventional regression analysis can be preferable when functional forms are defensible and the sample is limited. DML reduces reliance on a single nuisance specification, but introduces fold, learner, tuning, and asymptotic requirements.
Sensitivity analyses and reproducibility
Compare prespecified learner libraries, fold counts, repeated seeds, and reasonable overlap rules. Compare the primary DML estimate with a transparent parametric model targeting the same estimand. Disagreement should trigger investigation, not selective reporting.
Assess missing-data assumptions, alternative defensible exposure definitions, influential observations, and unmeasured confounding. Keep a decision log containing code versions, seeds, folds, tuning, failures, diagnostics, estimates, and reasons for each analysis choice.
Do not present only the specification that produces the narrowest interval or smallest p-value. A robustness analysis is informative when each variation tests a named assumption.
A reproducible DML workflow
- Define eligibility, exposure strategies, outcome, follow-up, population and estimand.
- Draw a causal diagram and select measured baseline confounders.
- Audit timing, missingness, exposure counts, clustering and empirical overlap.
- Choose the DML model class and orthogonal score that target the question.
- Prespecify nuisance learners, preprocessing, tuning, folds and random seeds.
- Generate out-of-fold nuisance predictions with cluster-aware splitting where needed.
- Inspect prediction quality, propensities, score contributions and split stability.
- Estimate the target parameter and design-appropriate uncertainty.
- Run prespecified model, overlap, missing-data and confounding sensitivities.
- Report code, versions, diagnostics, decisions, estimates and limitations.
Frequently asked questions
Why is it called double machine learning?
The name refers to the debiasing framework, which commonly estimates multiple nuisance functions with machine learning and combines orthogonal scores with cross-fitting. It does not simply mean running two algorithms.
Is DML the same as causal machine learning?
DML is one causal machine-learning framework. Causal forests, meta-learners, targeted learning, Bayesian methods, and other approaches target different structures and questions.
Does DML remove unmeasured confounding?
No. For observational effects under unconfoundedness, all sufficient confounders must be measured and appropriately represented. Flexible learners cannot create information that was never observed.
How many cross-fitting folds should I use?
There is no universal number. The choice must leave enough observations and both exposure groups in every training set. Report sensitivity to defensible alternatives when split variation is meaningful.
Can I use random forests in DML?
Yes, when the sample and nuisance tasks support them. Use protected tuning, out-of-fold prediction, calibration checks, and simpler benchmark learners.
Is DML always better than regression?
No. Regression can be more stable and interpretable when its functional form is defensible. DML helps with complex nuisance relationships but adds implementation and sample requirements.
What belongs in the methods chapter?
Report the estimand, assumptions, causal covariates, model class, orthogonal score, learner library, preprocessing, tuning, cross-fitting, overlap diagnostics, inference, software, seeds, and sensitivity analyses.
Conclusion
Psychology dissertation double machine learning is valuable when a low-dimensional causal question sits within complex, high-dimensional nuisance relationships. Neyman orthogonality and cross-fitting reduce the first-order impact of nuisance error and overfitting, but they do not replace design, measurement, overlap, or identification.
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 support should strengthen independent understanding, never invent data or promise a preferred result.
