Churn Without Fragmentation: How a Party-Label Bug Reversed My Headline Finding

Between 2018 and 2022, English city councils grew to become almost twice as unstable. Median volatility rose from 12.0 to 22.5.

However the celebration system didn’t fragment.

That distinction grew to become seen solely after fixing a categorical knowledge bug.

Right here, volatility measures how a lot vote share moved between celebration households. Fragmentation measures what number of efficient events competed. A council may be extremely unstable with out turning into extra fragmented if one main celebration collapses and one other absorbs a lot of the loss.

The efficient variety of events elevated in solely 18 of 67 comparable authorities. The median change within the fragmentation index stayed barely adverse: -0.31. The vote moved sharply, but it surely principally moved inside an already-consolidating celebration system.

The primary model of this evaluation regarded dramatically completely different. It prompt fragmentation had risen in 66 of 67 councils and that median volatility had tripled. That was unsuitable. The error got here from treating poll labels reminiscent of “Labour Celebration” and “Labour and Co-operative Celebration” as separate analytical events. As soon as celebration households had been normalised earlier than computing the metrics, the headline modified fully.

What regarded like a party-label bug was actually a category-modelling failure. And its penalties propagated by way of each downstream metric.

The corrected story is much less sensational. It’s also extra helpful.

Classes are a part of the mannequin

Earlier than strolling by way of the findings, it’s value explaining what went unsuitable, as a result of that is the half that generalises most immediately past elections.

Celebration labels aren’t impartial strings. They encode messy institutional actuality: alliances, poll wording, native celebration manufacturers, nationwide celebration rebrands, and inconsistent supply coding. If these labels are grouped incorrectly, each downstream metric can look exact and nonetheless be unsuitable.

That’s precisely what occurred. Fragmentation was computed earlier than normalising celebration households. In boroughs the place “Labour Celebration” and “Labour and Co-operative Celebration” each appeared, the Laakso-Taagepera denominator handled them as separate events. That artificially inflated the efficient variety of events. The identical threat utilized to UKIP, Reform UK, and Brexit Celebration labels.

The repair was conceptually easy: compute analytical celebration households earlier than metric aggregation.

The pipeline now separates three identities:

• Metric celebration household: used for fragmentation, volatility, and swing calculations.
• Challenger celebration household: used for state of affairs and challenger identification.
• Show celebration label: used just for Tableau color and labelling.

Don’t let show labels leak into metric definitions. Don’t let uncooked strings outline analytical classes with out an specific contract.

The distinction between the unique headline (“fragmentation rose in 66 of 67 councils”) and the corrected headline (“fragmentation rose in solely 18 of 67”) will not be a rounding error. It’s a categorisation error that propagated by way of all the pipeline. Each chart and each narrative conclusion shifted as soon as the repair was utilized.

The broader precept applies properly past elections. Product classes, job titles, firm names, analysis codes, and service provider names all have the identical failure mode. If class normalisation occurs after aggregation, it’s too late. The story has already been distorted.

How the evaluation works

The undertaking follows a pattern-first method: construct the information pipeline, export the metrics, assemble the visualisation, then let the information inform you which story it really helps. The corrected fragmentation discovering, the null turnout correlation, and the geographic shift in Inexperienced features all emerged from diagnostic validation, not from the unique undertaking plan.

The pipeline ingests ward-level election outcomes from the DCLEAPIL v1.0 dataset (Leman 2025), which pulls on Andrew Teale’s LEAP archive and Democracy Membership knowledge. It normalises celebration households, aggregates vote shares to the authority stage, computes fragmentation and volatility metrics, and exports structured CSVs for an interactive Tableau dashboard.

The evaluation covers 68 English metropolitan borough, London borough, and West Yorkshire authorities throughout 5 areas. Of those, 67 have comparable fragmentation knowledge throughout the 2018-to-2022 window.

The core metrics are:

• Fragmentation Index: the Laakso-Taagepera efficient variety of events, from authority-level vote shares.
• Volatility Rating: a composite metric combining a Pedersen-style absolute swing element with the change in fragmentation.
• Turnout Delta: percentage-point change in turnout throughout the identical window.
• Celebration Swing: change in vote share by normalised celebration household.

The method generalises to any area the place you’ll want to compute derived metrics from messy categorical knowledge and current them in a validated, reproducible visualisation. The complete pipeline, calculated fields, and Tableau construct information are open-source.

The headline: volatility rose, fragmentation didn’t

The primary dashboard panel maps volatility by authority. Circle dimension represents the volatility rating. Color represents the change in fragmentation: teal the place it rose, amber the place it fell.

The map reveals two issues without delay. First, volatility genuinely elevated: about 1.9 instances larger than the prior window. Second, fragmentation didn’t rise in most locations. Solely 18 of 67 comparable authorities had the next efficient variety of events in 2022 than in 2018.

The very best-volatility authorities had been Solihull (67.6), Kingston upon Thames (60.3), Sutton (48.7), South Tyneside (47.4), and Havering (45.2). 5 of the highest eight are London boroughs, however the highest total is Solihull. This isn’t merely a capital-city story.

Knowledge science takeaway: when two associated metrics (volatility and fragmentation) transfer in reverse instructions, the analytical story modifications fully. All the time examine whether or not your headline metric and your supporting metrics agree earlier than publishing. The hole between the 2 is the place the precise discovering lives.

Brexit consolidated the vote. 2022 didn’t undo it.

The second view plots the efficient variety of events throughout three factors: every council’s final pre-2018 election, 2018, and 2022.

The previous model described this chart as a V-shape: consolidation into 2018, then fragmentation after 2022. The corrected knowledge doesn’t assist that. The higher studying is consolidation, then partial stabilisation.

Tier medians present the sample: London declined from 2.87 to 2.16. Metropolitan boroughs declined from 3.22 to 2.65 (with a slight uptick from the 2018 low of two.62). West Yorkshire declined sharply from 4.13 to 2.01.

The 2022 cycle was disruptive, but it surely was not a generalised splintering of the celebration system.

The mechanism: Conservative collapse, uneven absorption

The party-swing chart explains how volatility can rise whereas fragmentation falls.

Throughout 67 councils, the median party-family swing between 2018 and 2022 was: Labour +8.5 proportion factors, Conservative -8.3, Liberal Democrats -2.3. Each different celebration moved lower than 0.3 factors in both course.

These swings are calculated on normalised celebration households. Labour and Labour Co-operative are grouped collectively, as are UKIP, Reform UK, and Brexit Celebration labels. With out this normalisation, the uncooked knowledge would present deceptive Labour Co-operative features alongside Labour losses in the identical borough. The normalisation logic is documented within the data source metadata.

On the median, it is a Conservative-loss and Labour-gain story, not a third-party surge. However medians flatten geography. Labour absorbed the standard Conservative loss, whereas Liberal Democrats and Greens surged in particular councils.

Utilizing an insurgency filter of a minimum of a 5-point acquire from a 2018 base of a minimum of 2%: Liberal Democrats surged in 9 councils, Greens in 7, and the Yorkshire Celebration in 1. Independents and Reform/UKIP didn’t clear the brink on this window.

Knowledge science takeaway: threshold choice in categorical filters deserves the identical rigour as hyperparameter tuning. The preliminary insurgency filter (5pp swing, no baseline flooring) produced 12 Inexperienced “surge” councils. Diagnostic inspection revealed 5 had been low-base artifacts: events going from 0.5% to five.5%. Including a 2% baseline flooring lowered the depend to 7 and adjusted the geographic composition totally. The analytical discovering (Northern metros, not interior London) solely emerged after the filter was corrected. Any threshold utilized earlier than a headline discovering needs to be stress-tested by inspecting the sting circumstances it admits.

That’s the mechanism: uneven absorption. The place Labour absorbed Conservative losses cleanly, volatility rose however fragmentation usually fell. The place a 3rd celebration absorbed a part of the loss, native competitors grew to become extra advanced.

The Inexperienced story is geographic, not nationwide

The Inexperienced median swing was +0.1 proportion factors. That quantity is correct and deceptive.

It’s correct as a result of the standard council didn’t see a big Inexperienced advance. It’s deceptive as a result of Inexperienced assist moved geographically.

In a number of interior London boroughs, Greens fell sharply:

Desk 1: Interior London Inexperienced retreat, 2018 to 2022. Three boroughs the place Greens held double-digit vote share in 2018 noticed sharp declines by 2022.

On the identical time, Greens surged in Northern and Midlands authorities plus Westminster:

Desk 2: Inexperienced surge councils, 2018 to 2022. Seven authorities the place Greens gained 5+ proportion factors from a base of a minimum of 2%. Six are Northern and Midlands metropolitan boroughs. Westminster is the only real London borough on the listing.

The interior London Inexperienced surge seems to have occurred earlier than 2018. Between 2018 and 2022, a few of that vote moved again towards Labour. In the meantime, Greens gained from decrease bases in post-industrial metros.

The dataset can not show voter motivation. But it surely reveals {that a} nationwide Inexperienced median is the unsuitable stage of study. A flat mixture median can disguise giant offsetting actions throughout subgroups. The true sample is redistribution throughout locations, and also you want the authority-level view to seek out it.

Regional volatility: group-level summaries aren’t explanations

Median volatility by area: North East 27.8, Yorkshire 25.7, London 22.0, North West 16.0, West Midlands 15.6.

The West Midlands has probably the most unstable council within the dataset (Solihull at 67.6) however the lowest regional median. Aggregating by area helped orient the evaluation, but it surely additionally confirmed why group-level summaries aren’t explanations. Council-level components dominate regional geography.

Turnout and volatility moved independently

I anticipated unstable councils to have falling turnout.

Throughout 67 authorities, the Pearson correlation between turnout change and volatility is -0.12 (p = 0.35). Limiting to 64 election-active authorities: r = -0.15, p = 0.25. Each statistically null.

Knowledge science takeaway: publishing null findings prevents unhealthy narratives from turning into defaults. The unique wireframes assumed a adverse turnout-volatility correlation. When the computation returned r = -0.12 (p = 0.35), the headline was rewritten reasonably than the information re-scoped. Each scopes are reported transparently. Null findings are undervalued in knowledge evaluation. Letting knowledge override assumptions is straightforward to explain and genuinely laborious to practise.

What the corrected story says

English councils skilled a lot larger voter churn between 2018 and 2022. Median volatility rose from 12.0 to 22.5. However the efficient variety of events didn’t rise in most councils. Fragmentation elevated in solely 18 of 67 comparable authorities, and the median change remained barely adverse.

Native electoral churn may be intense with out producing a extra fragmented celebration system. Voters moved, however in lots of locations they moved from one dominant pole to a different. The place smaller events superior, they did so domestically and erratically, not as a uniform nationwide wave.

The true lesson is upstream: classes are a part of the mannequin. Get them unsuitable, and each chart tells a convincing however incorrect story.

Knowledge sources and licensing

The underlying election outcomes come from the DCLEAPIL v1.0 dataset (Leman, Jason, 2025), launched below CC BY-SA 4.0. Supplementary knowledge from the House of Commons Library is used below Open Parliament Licence v3.0. Derived datasets and pipeline code are launched below MIT licence. Knowledge provenance is documented in DATA_SOURCE_METADATA.md.

Methodology notes

68 in-scope authorities. 67 with comparable fragmentation values for the 2018-to-2022 window (Rotherham excluded from FI comparisons). Fragmentation makes use of the Laakso-Taagepera index. Volatility is a composite of Pedersen-style swing and fragmentation change. Celebration swings use normalised analytical households. The insurgency filter excludes Labour and Conservatives and requires a 2% 2018 baseline flooring. Causal language is interpretive; the information captures outcomes, not motivations.

What comes subsequent

A companion evaluation will discover 2026 eventualities: baseline continuity, Reform local-surge assumptions, and major-party reconsolidation. These are eventualities below algebraic assumptions, not forecasts.

The central query: if Conservative losses proceed, does Labour take up them once more, or do the geographically concentrated LibDem and Inexperienced surges unfold to new councils? And does that absorption sample lastly push fragmentation upward, or does the celebration system proceed to consolidate at the same time as particular person councils churn?

That distinction between churn and fragmentation is what the undertaking is designed to measure.

The interactive dashboard is printed on Tableau Public and the complete knowledge pipeline is on the market at github.com/Wisabi-Analytics/civic-lens.

Obinna Iheanachor is a Senior AI/Knowledge Engineer and founding father of Wisabi Analytics, a UK-based knowledge engineering and AI consultancy. He creates content material round manufacturing AI methods, knowledge pipelines, and utilized analytics at @DataSenseiObi on X and Wisabi Analytics on YouTube. Civic Lens is an open-source political knowledge undertaking at github.com/Wisabi-Analytics/civic-lens.