# Stack question groups

> Reshape a multi-item grid so each item becomes its own row. Stacking lets you analyse the items as a single variable instead of comparing N parallel columns.

*Source:* https://recense.ai/docs/stacking

## What stacking is

Many surveys ask the same thing about several items. A grid like "How likely are you to recommend each of these brands?" produces N parallel variables — one per brand. The data is wide: one row per respondent, N columns for the brands.

Stacking reshapes the data so each item becomes its own row. Instead of N columns, you get one likelihood column plus a new "brand" variable that says which item the row is about. The data is long: N rows per respondent.

> **ℹ Wide vs long, in one line**  
> Wide: one row per respondent. Long (stacked): one row per (respondent × item).

## When to stack

- You want to compare the items as a single variable — for example "average likelihood across all brands, broken out by age group".
- You want to filter or weight on the item — for example "only show responses about brands the respondent had heard of".
- You want to model item-level effects (regression, clustering, etc.) where each item is an observation.

## When to leave as-is

- You're studying combinations or cross-patterns — for example "who picked brand A and brand B together?". Stacking flattens that relationship away.
- The items aren't really parallel. If three columns measure different things and only happen to share a prefix, stacking would smush incompatible data.
- You only need the per-item summary tables Recense already builds for the wide form.

## How to stack in Recense

1. **Look for stacking suggestions on the Datasets tab.** — After import, Smart prep scans your question groups and surfaces "Prep suggestion" cards for groups that look stackable.
2. **Review the proposed reshape.** — Each card shows what the new long form will look like — predicted row count, the new item variable, and any helper variables.
3. **Apply or dismiss.** — Apply turns the suggestion into a derived stacked dataset. Dismiss hides the suggestion if the items are not really parallel. You can always re-detect later.

> **ℹ Stacking doesn't lose data**  
> The original wide form stays available. Stacking produces a new derived view — switch back any time.

## Next steps

- **[Build tables and analysis](/docs/tables-and-analysis)** — Tabulate your stacked dataset like any other.
- **[Methodology](/docs/methodology)** — How Recense treats long-form data in tests and weighting.
- **[Explore datasets](/docs/explore-datasets)** — Confirm group structure before stacking decisions.