Running Searches and Grid Search Tips

Once you have explored your data and chosen a starting set of assumptions and methods (see Algorithm Selection and Assumptions), the next step is to run causal searches systematically.

Rather than treating causal discovery as a one-shot operation, Tetrad is designed to support exploring how results change across reasonable choices of algorithms, tests or scores, and tuning parameters. The Grid Search tool provides a structured way to do this.






Interpreting Grid Search Results

A Grid Search produces a table where each row corresponds to a specific algorithm and parameter combination.

Two aspects are especially informative:

1. Markov Consistency

  • Does the graph’s implied conditional independence structure agree with the data?

  • Diagnostics such as the Markov Checker are designed to assess this.

Graphs that consistently fail Markov diagnostics typically warrant closer scrutiny or revised assumptions.


2. Model Complexity

  • Number of edges

  • Degrees of freedom (when available)

Among models that pass diagnostics, simpler graphs are often preferred unless there is a clear reason to accept additional complexity.


A Practical Starter Pattern

A commonly effective approach is:

  1. Choose one algorithm family (e.g., PC or FCI).

  2. Sweep one key parameter (Ξ± or penalty).

  3. Evaluate results using:

    • Markov Checker statistics

    • Visual inspection of graphs

  4. Identify minimal models that pass diagnostics.

  5. Optionally repeat with a second algorithm family.

This balances systematic exploration with interpretability.


Reading Grid Search Output

When examining the results table:

  • Each row corresponds to a distinct model.

  • Selecting a row allows you to inspect the associated graph.

  • Pay attention to:

    • Adjacencies that appear across many settings

    • Orientations that remain stable

    • Edges that appear or disappear easily (these are typically less robust)

The aim is not to identify a single β€œbest” graph, but to understand which features are consistently supported.


Common Pitfalls to Avoid

Sweeping Too Many Parameters at Once

Large grids can become difficult to interpret. Starting with a small, focused sweep is usually more productive.


Changing Background Knowledge Too Early

It can be helpful to first see what the data suggest before adding strong constraints.


Delaying Diagnostics

If many models fail Markov diagnostics, it may be worth revisiting assumptions, tests, or parameter ranges early.


Not Recording What Was Tried

Keeping brief notes on parameter choices and outcomes can greatly simplify interpretation and reporting.


Where Grid Search Fits in the Workflow

Grid Search sits at the center of the causal analysis workflow:

  • After choosing assumptions and methods

  • Before final interpretation and reporting

It turns causal discovery from a single run into a structured, evidence-based exploration.


🧭 Next Step

Once you have identified promising candidate models, continue to Model Evaluation and Markov Checking to assess consistency, robustness, and plausibility in greater detail.