# The training group changed more—in this simulation

The estimated increase was **7.53 points in training** and **0.42 points in control**. The planned difference in change was therefore **7.11 points**, with a **95% confidence interval of [4.25, 9.97]**. These numbers describe a deliberately simulated study, not evidence that a real training intervention works.

## Keeping repeated observations together

The dataset contained 100 people, 50 per group, each observed twice. The 200 measurements were analysed with a participant random-intercept model, `outcome ~ time * group + (1 | id)`. Estimation used REML with Satterthwaite degrees of freedom and 95% t-based intervals. All observations were retained. See the [NLSS mixed-model reference](https://github.com/docmh/nlss/blob/v2.0.1/references/subskills/mixed-models.md), [lmerTest inference](https://search.r-project.org/CRAN/refmans/lmerTest/html/anova.lmerModLmerTest.html) and [emmeans model support](https://rvlenth.github.io/emmeans/articles/models.html#group-l--lmermod-models).

| Group | Occasion | Estimated mean | 95% CI |
| --- | --- | ---: | --- |
| Control | Pre | 52.09 | [49.18, 54.99] |
| Control | Post | 52.51 | [49.60, 55.41] |
| Training | Pre | 48.82 | [45.91, 51.72] |
| Training | Post | 56.35 | [53.44, 59.25] |

These are model-estimated means; in this balanced, complete design their point estimates equal the observed group-by-time means. Their intervals concern individual means and must not be subtracted to construct the interval for change.

## Ask the intended question

We specified exactly one contrast: **(training post − training pre) − (control post − control pre)**. It gave *t*(98.00) = 4.93, *p* < .001. No multiplicity adjustment was applied to this single planned contrast, and no additional pairwise contrasts were requested.

The model's reference occasion is **post**, and its reference group is **control**. Consequently, the `timepre:grouptraining` coefficient is −7.11, the opposite sign to the planned change contrast. Named contrast weights preserve the intended direction. This is why an agent should inspect coding before translating a coefficient into a substantive answer.

![Group means from pre to post: control rises slightly, while training starts lower and rises more.](figures/trajectory.png)

*Figure 1.* Four observed group means, rendered from NLSS's saved graphical data with explicit occasion labels, stronger contrast and distinct point/line shapes. Lines connect occasions, not individual trajectories. No confidence bands are shown; uncertainty for the change comparison is reported above. The original NLSS figure is retained with the evidence.

## What still deserves attention

The fit converged and was not singular. Estimated participant-intercept variance was 81.61 and residual variance was 25.99. A conditional-residual normality screen gave *p* = .040, so the report does not declare that all assumptions passed. This screen does not assess random-effect normality or independence. The current model also assumes a common residual variance; robustness to a different variance structure has not been demonstrated.

The synthetic training group started lower, despite balanced random assignment. In an actual study, the design, outcome validity, missing follow-up and planned estimand would need scrutiny. A difference in change is the pre-specified target here; it is not automatically interchangeable with a baseline-adjusted post-treatment comparison. No threshold for a practically important change was supplied.

NLSS 2.0.1 · Material 1.0 · Mike Hammes · [CC BY 4.0](LICENSE.md). See [reproduction and evidence](reproduce.md).
