Grid Search Box (Data)

This page describes how the Grid Search box behaves when it is connected to a Data box.
In this mode, Grid Search performs systematic comparison of causal discovery methods on a fixed dataset.

No simulations are run in this mode. All results are deterministic and fully reproducible.



When This Mode Is Used

Grid Search automatically enters data mode when:

  • A Data box is connected to the Grid Search box, and

  • No simulation is selected.

In this case:

  • Simulation controls are hidden or disabled

  • Each algorithm–parameter combination is evaluated once

  • Results reflect only variation across modeling choices, not random variation


Basic Setup

To use Grid Search with data:

  1. Load your dataset into a Data box

  2. Draw an edge from the Data box to the Grid Search box

  3. Open the Grid Search editor

The editor will configure itself for data-based comparison.


Algorithms Tab

In the Algorithms tab, you select:

  • One or more causal discovery algorithms

  • Required independence tests and/or scores

  • Parameter ranges for algorithms, tests, and scores

Parameter values may be entered as comma-separated lists.
Grid Search will evaluate all combinations of the selected parameters.

Only tests and scores compatible with the data type (continuous, discrete, mixed) are shown.


Table Columns Tab

In the Table Columns tab, you choose which quantities appear in the comparison table.

Available columns include:

  • Algorithm and parameter values

  • Model complexity measures (e.g., number of edges)

  • Diagnostic statistics (e.g., Markov check results)

When working from data, statistics that require knowledge of the true graph are not shown, since no ground truth is available.

Columns may be added, reordered, or removed using the Add and Manage buttons.


Comparison Tab

The Comparison tab controls how results are evaluated and displayed.

Key options include:

  • Comparison graph type (e.g., CPDAG or PAG)

  • Markov Checker test

  • Utility settings for ranking models

When you click Run Comparison, Grid Search:

  1. Runs each selected algorithm for every parameter combination

  2. Evaluates resulting graphs using selected diagnostics

  3. Populates the comparison table with results

Progress and detailed output are shown in the Verbose Output tab.


Interpreting Results

Each row in the comparison table corresponds to a distinct model.

Typical analysis focuses on:

  • Whether the model passes Markov checking

  • Model complexity (e.g., number of edges)

  • Stability of features across nearby parameter settings

Rather than selecting the single highest-utility model, it is usually more informative to identify minimal models that pass diagnostics.


View Graphs Tab

After a comparison is complete, the View Graphs tab allows you to inspect individual output graphs.

  • Selecting a row in the table highlights the corresponding graph

  • Graph selections are remembered when the editor is reopened

This makes it easy to compare candidate models visually.


Notes and Best Practices

  • Sweep only a small number of meaningful parameters at a time

  • Prefer systematic comparison over isolated runs

  • Use diagnostics early to detect mismatched assumptions

  • Treat fragile edges and orientations with caution


Summary

When connected to data, the Grid Search box provides a structured, reproducible way to:

  • Explore algorithm and parameter sensitivity

  • Evaluate candidate causal models

  • Identify parsimonious models consistent with the data

This mode forms the backbone of Tetrad’s recommended causal discovery workflow.