🌻 Analysis Filters Filter links tab
PDF

Do qualitative causal analyseis on the selected links by filtering or manipulating them.

🔍 What you can do here: Apply filters to focus your analysis on specific aspects of your data. You can trace causal pathways, group similar concepts, filter by themes or demographics, and much more. Think of this as your analysis toolkit - combine different filters to explore your data from different angles.

The Filter System: overview#

Use filters to narrow down and/or transforme the links you want to study. Filters are applied in order, from top to bottom.

Hard vs Soft recoding#

Most filters leave factor labels untouched, but these 'Transform filters' filters temporarily relabel factors:

No filters actually change your original coding.

💡Tip: If you want to permanently rename or "hard recode" your factors, there are several ways to do that:

For example, after clustering (which may give labels like C11), click a factor on the map and rename it (e.g., "Wellbeing") to save the new name permanently.

Zoom Filter #

🔍 What this does: Simplify complex factor labels by zooming to higher levels of a hierarchy. For example, turn "Health; Mental Health; Depression" into just "Health" (level 1) or "Health; Mental Health" (level 2). Perfect for getting a big-picture view of your data.

Collapse Filter #

🏷️ What this does: Merge similar factors under one common label. Type or select multiple similar terms like "money", "income", "salary" and they'll all be replaced with the first term. Great for cleaning up data where the same concept is described in different ways.

Widgets:

Remove Brackets Filter #

🧹 What this does: Clean up your factor labels by removing text in brackets. For example, "Education (primary school)" becomes just "Education". Choose between removing content in round brackets ( ) or square brackets [ ].

If you want to remove both kinds of labels, simply create another Replace brackets filter beneath this one.

Factor Label Filter #

🎯 What this does: Show links connected to factors you care about (e.g. "Education"). Choose how many steps to look upstream (causes) and downstream (effects).

Widgets:

How to use: 1) Select one or more factors.
2) Set Steps Up/Down to widen or narrow the neighbourhood.
3) (Optional) Turn on Source tracing to require paths from a single source.
4) (Optional) Turn off Highlight to hide the custom highlighting. 5) The map and tables update to show only links on those paths.

All the label and tag filters including exclude filters have three radio buttons below the selectize input called Match: Start (default), Anywhere or Exact to control how search terms match against labels/tags:

Multiple search terms are treated as OR not AND. preserve and highlight factors matching ANY of the search terms.

Focused factors show with colored borders in the map and have a star added for easy identification (when Highlight toggle is on).

Exclude Factor Label filter #

🚫 What this does: Remove unwanted factors from your analysis. Type factors like "Unclear" or "Other" to hide them from your map and tables. Useful for cleaning up your data by removing vague or irrelevant categories.

Path Tracing Filter #

🛤️ What this does: Find causal pathways between two specific points. Set a starting factor (like "Poverty") and an ending factor (like "Poor Health") to see all the causal chains that connect them. Great for understanding how problems and solutions are linked.

Exclude self-loops Filter #

You can exclude self-loops from the maps, but that is more of a visual change. This is a real filter as part of the filter pipeline. For example, if you are using a filter like Link Frequency that might be retaining link bundles which are actually self-loops, so you might get unexpected results if you use the map setting to remove the self-loops. So this filter is a better way. It simply removes all links which are self-loops from the links table.

🏷️ What this does: Filter your analysis by the tags you've added to links. Show only links tagged as "#important" or "#policy" to focus on specific themes or types of relationships you've identified.

Combine Opposites filter #

🔄 What this does: Unify opposite factor labels by matching tag numbers. If you have pairs like `Foo [99]` and `Bar [~99]` (where `~99` indicates the opposite), this filter rewrites `Bar [~99]` as `Foo [99]` to combine them under one label. The `flipped_cause` and `flipped_effect` columns track which causes and effects were flipped.

Toggle – Turn the filter on/off.

Strip tags from labels (default: on) – When enabled, removes [N] and [~N] tag patterns from labels after combining opposites. This keeps labels clean while preserving the tracking information in the flipped_cause and flipped_effect columns.

Labels can be written in pairs like:

where Bar represents the opposite of Foo. The square brackets are optional - you can use Foo 99 and Bar ~99 - but brackets make it easier to remove tags later using the Replace Brackets filter.

If there are any such pairs, with matching integers, and the filter is switched on:

rewrite any Bar [~99] filters as Foo [00] and add new columns:

to the current augmented links table, so that if the label has been flipped, the value is True and otherwise False.

Wire up the filter as part of the standard filter system with save/restore to URL etc.

Also, when calculating new links table, create new text columns:

The embellished counts always show all variants with custom SVG icons (no total prefix). Four circle icons represent the flipped status:

So if a bundle has 12 citations where 5 are unflipped/unflipped, 2 have flipped cause and flipped effect, and 1 has flipped cause but non-flipped effect, the text is: light-blue-circle5, dark-red-circle2, mixed-circle1. If nothing were flipped, the label would just be light-blue-circle12. Do the same with source counts too, counting the unique sources in each variant.

When the filter is on, and source count or citation count is selected, the graphviz and graphviz maps change to use source_count_with_opposites or citation_count_with_opposites just for the labels. The edge width calculation remains driven by source count or citation count, as selected.

🚫 What this does: Remove specific types of links from your analysis. Exclude links tagged as "#uncertain" or "#duplicate" to focus on higher-quality data. Helpful for filtering out questionable or irrelevant causal claims.

Exclude self-loops Filter #

You can exclude self-loops from the maps, but that is more of a visual change. This is a real filter as part of the filter pipeline. For example, if you are using a filter like Link Frequency that might be retaining link bundles which are actually self-loops, so you might get unexpected results if you use the map setting to remove the self-loops. So this filter is a better way. It simply removes all links which are self-loops from the links table.

📊 What this does: Focus on the most important causal relationships by filtering out rare ones. Choose "Top 10" to see only the most frequently mentioned connections, or set a minimum threshold like "at least 3 sources" to ensure reliability.

Examples:

By default, setting the slider to 6 means we are selecting only links with at least 6 citations.

If you switch to “Sources”, we are selecting only links with at least 6 sources.

If you switch to “Top” we are selecting only the top 6 links by citation count, etc. The selection respects ties, so that if there are several links with the same count, either all of them or none of them will be selected.

Factor Frequency Filter #

📈 What this does: Similar to Link Frequency, but focuses on the most important factors (causes and effects). Show only the most frequently mentioned themes or concepts to identify the key issues in your data.

Same controls as Link Frequency but applies to factors instead of links.

Source Groups filter #

👥 What this does: Filter your analysis by participant demographics or document characteristics. For example, show only responses from "women aged 25-35" or interviews from "urban areas." Perfect for comparing how different groups see causal relationships. This is very similar to the [Source Groups widget](../source-groups-sub-panel/) in the Sources sub-panel, but having it here too means you can add multiple source filters to the pipeline.

Everything Filter #

👥 What this does: Filter your analysis by any characteristic of your links and their sources. Useful for anything not covered by the other filters, for example, - show only links with negative sentiment - Show only links from one source Also shows source separators and their values, often used for common sections within multiple sources texts.

Soft Relabel Filter #

👥 What this does: Temporarily relabel factors.

Effect: all factors exactly matching any of the labels in the Old list are relabelled with the corresponding labels from the New list. factors not listed are not relabelled but preserved.

Many use cases:

Keyboard shortcuts (Win/Linux ⇄ macOS):

Potentially, one NEW label might have multiple OLD labels.

Soft Recode Plus filter #

Requires AI subscription

🧲 What this does: Group messy factor labels under clearer names you choose (called "magnets"). Example magnets: `Improved health`, `Education programs`, `Income changes`. The filter finds the closest magnet for each label and replaces it.

Controls:#

Create Suggestions for Magnets#

(collapsed by default): Optional. Ask AI to propose clear names from your current labels. Insert adds them to your magnets box to review/edit.

Main panel#

Recoded columns#

When you use Soft Recode Plus, the Links and Factors tables show special columns that track which labels have been recoded:

These columns track recoding from any filter that transforms labels: Soft Recode Plus, Zoom, Collapse, Remove Brackets, Soft Relabel, Cluster, Hierarchical Cluster, and Combine Opposites.

Process only unmatched NEW#

the point of this is: what if I apply some (maybe standard) magnetisation and matches plenty of factors but there might be some important material left unmatched, not just noise. so i can use a PAIR of these filters. in the first one, I leave OFF its Discard Unmatched toggle and in the second filter switch ON its Only Unmatched filter. (if there is no preceding SRP filter with Discard Unmatched=OFF, this second SRP filter does nothing).

So now,

Meaning Space (2‑D embeddings)#

Go to the map formatting and select Layout → Meaning Space to see a 2‑D scatter of your factors in “meaning space”.

Motivation for Remove Hierarchy#

"Remove hierarchy", default off. if on, strip any text before a final semi-colon, if no semi-colon, do not change the text. 

something; another thing

is treated same as

another thing

.... but it continues to be treated as "something; another thing" in the rest of the filter pipeline.

Quick workflow:

1) (Optional) Open Create Suggestions for Magnets panel → set Number of clusters and use Insert to get AI suggestions.
2) Use these suggestions and/or edit them, paste or type your own magnets (one per line).
3) Click Save.

See tips on using the history to reuse both your labelling prompt and magnet sets.

Motivation for "recycle weakest magnets": suppose you create 20 magnets, and then apply more filters like say a link frequency filter so that you end up with say only 5 factors. If you then remove those factors from the magnets list which are not included in the final output, you will usually increase the coverage of your map (re-assigning raw labels which fit best with one of the "lost" labels but still fit well with one of the "surviving" labels). This is what the Recycle slider does: it recycles the specified number of smaller magnets and reassigns them to the larger magnets. So in the example, if you start off with 20 magnets but your final map only shows 5, try recycling say 10 or even 15 of the missing factors.

Note that Recycle Weakest Magnets is applied BEFORE Drop Unmatched.

Clustering filter #

Requires AI subscription

🎯 What this does: Automatically discover groups of similar factors in your data using machine learning. The system finds natural clusters of related concepts and labels them with cluster numbers. Great for exploratory analysis when you're not sure what causal themes exist.

Auto Recode filter #

🪜 What this does: Quickly turns your current set of labels (after any previous filters like Zoom) into a simple tree you can "roll up" or "open out". Pick a small number of clusters for a big‑picture view, then nudge the Balance and Similarity to tidy results. Designed for fast, practical exploration on real projects.

Motivation#

Making sense of hundreds or thousands of factor labels is hard.

You might use something like soft Recode Plus, but often you'll ask for 20 clusters to cover a wide range of meanings. Then after filtering out insignificant data, you end up with only 7 clusters — losing coverage. Ideally you'd go back and recreate just 7 clusters, but that gives different results. Frustrating!

The point of this Auto Recode filter: have your cake and eat it. Ask for an foldable/unfoldable hierarchical solution. When you move the slider to 15, you get the best solution for 15 clusters. Slide it to 3, you instantly get the best solution for 3 clusters.

Controls:#

NEW: AI labelling prompt with history controls. Use this to suggest clearer names for each cluster:

NEW: Seed labels (optional) with history and strength:

How to use (quick):

Notes:

A good prompt looks something like this:

This is a list of many raw labels grouped into two different clusters, with their cluster IDs, together with a reference list of other labels. Return a list of two new labels, one for each cluster ID. Each label should capture the meaning of the whole cluster, using similar language to the original raw labels, but in such a way that the labels you create are distinct from one another in meaning. Try not to be too generic, try to be as concrete as you can. Do NOT provide labels which include causal ideas, like "X through Y" or "X leading to Y" or "X results in Y" or "X improves Y" etc. Equally, don't include conjunctions in the title like "X and Y". The meaning of the labels you give me should ideally not overlap in meaning with one another or with the labels in the reference list.

Optimized Cluster filter #

⚠️ DEPRECATED Requires AI subscription

🎯 What this does: Automatically finds the most optimal factor labels to use as centroids through genuine optimization. Unlike regular clustering that just groups similar items, this finds the best possible n≤N labels that maximize coverage with similarity ≥S. Perfect for discovering the most representative concepts in your data.

Controls:

How it works:

  1. Extracts all unique labels from your current data (1K-30K labels supported)
  2. Runs iterative optimization with multiple strategies (random, frequency-based, diverse selection)
  3. Uses hill-climbing optimization to find the best possible centroids
  4. Shows coverage percentage and timing information
  5. Returns recoded links table with optimal centroid labels

Optimization Strategies:

Performance Features:

Technical Implementation:

Soft Recode Integration:

This filter implements the optimization challenge described in the technical documentation: finding optimal centroids that maximize label coverage within similarity constraints.

Tribes filter #

Requires AI subscription

🏛️ What this does: Group your sources (participants/documents) by how similarly they describe causal relationships. This reveals different "tribes" or perspectives in your data - for example, optimists vs. pessimists, or urban vs. rural viewpoints.

Controls:

.

It returns:

We can then show maps for each tribe and/or for the most typical source in each tribe. we could also then create a typical story centred around the current factors, i.e. told in terms of our concepts.

--->

🏷️ What this does: Configure how link labels appear on your map based on any field in your data. Choose what information to display (like tribe memberships or custom attributes) and how to show it (counts, percentages, or statistical significance).

Controls:

To use:

  1. Add the Custom Links Label filter to your pipeline
  2. Select a field (e.g., custom_tribeId after running the Tribes filter)
  3. Choose a display mode
  4. In Map Formatting, set Link Labels to "Custom Links label"

Example use cases:

<!---

Technical details:

This is a non-filtering filter - it doesn't change which links appear, only configures how they're labeled on the map.

Display modes: All modes use the Counts toggle:

The filter populates its field dropdown from currentFilteredLinks (the output of the filter pipeline), so it sees all fields added by previous filters.