Loading
Economics · Causal Analysis & Experimentation
Determine what changed because of the decision, not merely what changed afterward.
For data-mature organisations that distrust simple attribution but fear academic complexity. We design experiments and econometric analyses that separate what an intervention actually caused from seasonality, selection and coincidence, translate the result into a commercial decision, and say plainly when the evidence cannot support a causal answer.
- Cause, separated from coincidence
- Method follows the data
- 'Inconclusive' said plainly
Trusted by established businesses
The constraint
Questions that bring teams here.
Each of these looks answerable from a dashboard, and each hides a trap that routinely produces the wrong answer.
Until cause is separated from coincidence, the organisation keeps scaling interventions that never worked and killing ones that did, with each wrong lesson compounding into the next plan.
The approach
Correlation describes; causation decides
Two numbers moving together tells you they moved together, nothing more. Revenue rose after the campaign; churn fell after the feature shipped; the market, the season and three other initiatives all moved at the same time. Causal analysis asks the only question a decision actually needs answered: what would have happened without the intervention? That imagined alternative is called the counterfactual, and every credible method is a disciplined way of estimating it: from randomised tests, where chance builds the comparison for you, to econometric designs that construct one from history when a live experiment isn't possible. The method is never chosen from preference: it follows the data-generating process, meaning the way the data actually came to be: who was exposed to the change, when, and why. And when no design can credibly answer the question with the data available, we say so before you spend on the analysis, not after.
What we analyse
Eight ways to build a credible comparison
Decisions
Decisions this supports
A causal estimate is only useful once it lands on a live commercial choice.
Scale
Scale or stop the pilot
Take the scale decision on the effect the evidence supports at scale, not on the pilot's best-case result.
Spend
Keep or reallocate the spend
Fund the channels and programmes whose incremental effect is demonstrated, and recover budget from the ones riding on coincidence.
Rollout
Roll out or roll back
Commit the price change, feature or policy everywhere (or reverse it) with its measured effect and uncertainty on the table.
Targeting
Target by measured response
Direct the intervention at the segments where uplift is real, and away from those it doesn't move, or actively harms.
Planning
Plan with effect ranges
Carry the estimated effect and its uncertainty into forecasts and budgets as a range, not as a single hopeful number.
Measurement
Invest in evidence first
When feasibility shows the question can't yet be answered credibly, fund the instrumentation or design that will answer it, before the next big commitment.
Deliverables
What you receive
- ✓Causal question and estimand: the precise effect being estimated, for whom, over what period
- ✓Identification strategy: the argument, in plain language, for why the comparison isolates cause
- ✓Power assessment: whether the data can detect an effect large enough to matter
- ✓Pre-agreed analysis plan, fixed before results are seen
- ✓Data quality and bias review
- ✓Model diagnostics and robustness checks
- ✓Effect estimate with uncertainty ranges: never a bare point number
- ✓Segment analysis where the sample credibly supports it
- ✓A decision recommendation: including what the evidence cannot conclude
How we work
How an analysis runs
Seven stages, with the honest exits built in: an engagement can end at feasibility if the question cannot be answered credibly. That answer is cheaper than a misleading one.
- 01
Feasibility
We test whether the question is answerable before designing anything: a defined intervention and outcome, known exposure and timing, a credible comparison, traceable data lineage, sufficient sample for the effect size that matters, a stable environment, and ethical and privacy approval.
GateGo/no-go: can this question be answered credibly with the data available?
- 02
Design
The method is chosen to fit the data-generating process (how the intervention actually reached people), never from preference. The estimand, identification strategy and analysis plan are fixed and agreed before any result is seen.
- 03
Instrumentation and implementation
Where a live experiment runs, randomisation, tracking and metric definitions are implemented and verified (working with your engineers or ours), so the design that was agreed is the design that actually runs.
- 04
Observation
The experiment or observation window runs its planned course. Stopping rules were set in advance; peeking at results and stopping early when the numbers look good is one of the classic ways organisations manufacture false positives.
- 05
Analysis
The pre-agreed plan is executed as written. Deviations forced by data reality are documented and justified, not silently absorbed.
- 06
Diagnostics
Assumptions are tested, robustness checks run, and the estimate stress-tested against alternative specifications. An effect that only appears under one specification is reported as exactly that.
GateDoes the estimate survive its own assumptions?
- 07
Decision and learning
Results are translated into the commercial decision (including economic significance, not just statistical), and the design, data and findings are documented so the analysis is reproducible and the next question starts further ahead.
Scope
What this is, and what it isn't
What's included
- Feasibility assessment before any analysis is commissioned
- Experiment and quasi-experiment design matched to how your data is actually generated
- Pre-agreed metrics, analysis plans and stopping rules: fixed before results exist
- Experiment governance: change control during the window, predefined stopping rules, reproducible analysis
- Plain-language translation of every method, assumption and result
- Explicit statements of what the evidence cannot conclude
- Design and documentation that builds toward an internal experiment portfolio
What this is not
- Attribution dashboards or last-click reporting: that is descriptive analytics, not causal evidence
- Guaranteed 'positive' findings: the result is whatever the evidence supports, including inconclusive
- Academic research for publication: every analysis targets a commercial decision
- Data pipelines and warehousing: Data Engineering builds the foundations where they're missing
- Analyses whose feasibility check fails: we decline work the data cannot credibly support
- Data-mature organisations with a material intervention worth measuring properly
- A pilot, campaign, price change or feature whose true impact would change a real decision
- Leaders willing to fix success metrics before seeing results
- Teams that can trace where their outcome data comes from
- Executives prepared to hear 'inconclusive' when that is the honest answer
- Organisations wanting experimentation as a repeatable capability, not a one-off
Probably not the right fit
- −Organisations without reliable outcome data yet: data foundations come first, and we'll say so
- −Decisions that can't wait for an observation window: a credible answer takes the time the design takes
- −Teams seeking confirmation of a conclusion already reached: the analysis may disagree
- −Questions where the honest feasibility answer is no: we decline rather than dress up weak evidence
- −Anyone wanting a dashboard rather than an answer to a causal question
Where this fits
The evidence layer under every other economic decision
Pricing, product, forecasting and investment work all generate causal questions; this practice answers them, and engineering builds the data and instrumentation the answers depend on.
Econometrics & Causal Analysisthe evidence layer
Owns: experiment design · quasi-experiments · causal estimation · interpretation and governance.
Econometrics & Causal AnalysisPricing & Monetisation
When it comes into play
A price change needs its true effect measured before it rolls out everywhere.
What it owns
Price architecture and monetisation decisions.
Relationship to Econometrics & Causal Analysis
Pricing frames the change; causal analysis measures what it actually did.
Proof
Selected work.
Payment Processing Cost Reduction. An ecommerce retailer was losing a significant percentage of revenue to payment processing and invoice platform fees. Web Lifter redesigned the entire sales and payment workflow, replacing Stripe and Paycove with a direct Westpac PayWay integration and a custom-built invoicing platform. The new architecture reduced transaction costs, streamlined operations, and delivered immediate profit improvements without requiring any increase in sales volume.
Read the caseWhat clients say
“We can't recommend Web Lifter highly enough … a digital partner who could understand our operations, connect the dots between marketing and backend systems, and deliver real results.”
Questions
Common questions
Fit
Analytics tools describe what happened and attribution tools allocate credit by rule: last click, position, algorithmic weighting. Neither estimates what would have happened without the intervention, which is the question a spend, rollout or scale decision actually turns on. Causal analysis is designed around that missing comparison.
The rigour is academic; the output is not. Every method, assumption and result is translated into plain language, every analysis targets a named commercial decision, and 'statistically fascinating but commercially irrelevant' is treated as a failure. If a simpler design answers the decision, we use the simpler design.
We say so, plainly, and tell you what it would take to get a conclusive answer: more time, more sample, better instrumentation or a different design. An honest 'the evidence can't support that claim' routinely saves more money than a flattering estimate, because it stops a false lesson from scaling.
Feasibility & data
A defined intervention and outcome; knowledge of who was exposed and when; a credible comparison group or a way to construct one; data whose lineage you can trace; a sample large enough to detect an effect worth acting on; a reasonably stable environment during the window; and ethical and privacy approval where people's data is involved. The feasibility stage tests all of this before design starts, and the engagement can honestly end there.
Often not. Quasi-experimental and econometric designs (difference-in-differences, matching, synthetic control, discontinuity) exist precisely for interventions that were never run as experiments. Whether one applies depends on how the intervention rolled out and what comparison the history contains, which is exactly what feasibility establishes.
Never through the website: we do not accept raw datasets publicly. Data is exchanged only through secure channels agreed at engagement start, under your privacy and governance requirements, and any published write-up is redacted so no confidential data is exposed.
Interpretation
Statistical significance says an effect is probably not zero: that it's unlikely to be noise. Economic significance asks whether it's big enough to act on. With enough data, a trivial effect becomes statistically significant while remaining commercially worthless; with too little, a genuinely valuable effect can fail to reach significance. Every result we report addresses both, which is why estimates come as ranges tied to the decision, never as a lone asterisk.
Always: it's a named deliverable. Every write-up states the assumptions the estimate rests on, the populations and periods it generalises to, and the claims it does not license. The fastest way to waste a good analysis is to stretch it one claim further than the design supports.
Governance
The discipline that keeps results trustworthy: metrics and stopping rules defined before launch, change control so nothing else shifts mid-window, and analysis code documented for reproducibility. Without it, organisations unconsciously manufacture positive results: stopping tests when the numbers look good, redefining success after the fact, or quietly excluding awkward segments.
Yes, that's the natural trajectory. Designs, instrumentation and governance from early engagements are documented to be reusable, building toward an experiment portfolio your own teams operate: a standing pipeline of tests with shared standards, where each analysis makes the next one cheaper and faster.
Evidence before commitment