How to Use CausalLens

A step-by-step guide for non-technical users. No statistics PhD required.

5-minute quick start

1. Open the app → choose Clinical Treatment Effects
2. Click Use sample dataset
3. Set treatment = metformin, outcome = hba1c
4. Select confounders: age, bmi, baseline_severity
5. Pick 2 methods → Run causal analysis

What is causal inference?

Correlation means two things move together. Causation means one thing makes another change. CausalLens helps you estimate causal effects — but results always depend on assumptions about your data.

Type	Example	CausalLens?
Correlation	Ice cream sales and drowning both rise in summer	No
Causation	Does this drug cause lower blood sugar?	Yes
Causation	Did minimum wage policy cause employment changes?	Yes

The 5 wizard steps

Step 1 — Choose your field

Pick the domain closest to your question. This tunes AI suggestions and recommended methods.

Medicine — treatments, drugs, health outcomes
Science — experiments, lab and field data
Socioeconomic — income, employment, demographics
Marketing — campaigns, who responds to outreach
Policy — laws, subsidies, public programs

Step 2 — Load your data

You have three options:

Use sample dataset — fastest way to learn (recommended first time)
Upload file — CSV or Excel with column headers in row 1
Select existing — previously uploaded datasets

Your data needs at minimum:

A treatment column (the intervention: 0/1 or yes/no works best)
An outcome column (what you measure)
Optional confounders (factors affecting both treatment and outcome)

Step 3 — Define your question

Write your question in plain English, then map it to columns:

Example question:

“Does metformin reduce HbA1c, adjusting for age and BMI?”

Column mapping:

Treatment → metformin · Outcome → hba1c · Confounders → age, bmi, baseline_severity

Click AI suggest variables & methods for automatic suggestions. Works without an API key; richer with OpenAI configured in backend/.env.

Confounders tip: Ask “Could this factor affect who gets the treatment AND the outcome?” If yes, include it.

Step 4 — Pick methods

Select one or more analysis methods. We recommend running at least two and comparing results.

Method	Level	In plain English
Propensity Score Matching	Beginner	Match similar treated and control patients
Inverse Probability Weighting	Beginner	Reweight data to balance groups
DoWhy Backdoor	Intermediate	Formal causal model with explicit assumptions
Double / Debiased ML	Intermediate	Flexible machine learning approach
Causal Forest	Advanced	Different effects for different subgroups
T-Learner (Uplift)	Intermediate	Who benefits most from a campaign?
Difference-in-Differences	Intermediate	Before/after policy comparison
PC / LiNGAM Discovery	Advanced	Discover cause-effect structure from data

Step 5 — Read your results

Each method produces a result card with:

Effect estimate — the estimated causal effect (e.g. −0.93 on HbA1c)
95% confidence interval — plausible range for the true effect
Interpretation — what the number means in your context
Key assumptions — what must be true for a causal claim

You also get an AI summary in plain English and a causal graphshowing relationships between variables.

Important: A negative effect on HbA1c is good (lower blood sugar). Always interpret results in your domain context. CausalLens estimates effects under assumptions — it cannot prove causation from data alone.

Worked example: Diabetes study

Follow along with the built-in medicine sample:

Choose Clinical Treatment Effects
Click Use sample dataset
Question: “Does metformin reduce HbA1c?”
Treatment: metformin · Outcome: hba1c
Confounders: age, bmi, baseline_severity
Methods: Propensity Score Matching + DoWhy Backdoor
Run analysis → expect a negative effect (metformin lowers blood sugar)

Using your own CSV

Checklist

Column headers in the first row (descriptive names like age, not Column1)
Treatment column is 0/1 or yes/no
Outcome is numeric for most methods
At least 50–100 rows recommended
Confounders measured before treatment when possible

Example CSV

treatment,outcome,age,income
1,45000,35,52000
0,38000,42,48000
1,51000,28,55000

Common mistakes

Outcome before treatment — reverses causality
Missing confounders — biased estimates
Too few treated units — unreliable results
Confusing correlation with causation — wrong conclusions

Glossary

Treatment: The intervention you study (drug, policy, campaign)
Outcome: What you measure as a result
Confounder: A factor affecting both treatment and outcome
ATE: Average treatment effect for the whole population
Propensity score: Probability of receiving treatment given confounders
Difference-in-Differences: Compares change over time between treated and control groups

FAQ

How many methods should I run?

At least two. If they agree, you have more confidence. If they disagree, review assumptions and data quality.

Can the app prove causation?

No. CausalLens estimates effects under stated assumptions. Your study design and domain expertise matter.

Do I need OpenAI?

No. The app works fully without it. An API key enables richer AI explanations and hybrid causal discovery.

Where is the full written tutorial?

See TUTORIAL.md in the project folder for the complete guide with all domain examples.

Ready to start?

Try the diabetes sample — it takes about 5 minutes.

Open CausalLens