Statistical Considerations for Drawing Conclusions About Recovery

Background. Numerous studies have found associations when change scores are regressed onto initial impairments in people with stroke (slopes ~0.7). However, there are important statistical considerations that limit the conclusions we can draw about recovery from these studies. Objective. To provide an accessible "check-list" of conceptual and analytical issues on longitudinal measures of stroke recovery. Proportional recovery is an illustrative example, but these considerations apply broadly to studies of change over time. Methods. Using a pooled dataset of N = 373 Fugl-Meyer Assessment (FMA) upper extremity scores, we ran simulations to illustrate three considerations: (1) how change scores can be problematic in this context; (2) how "nil" and non-zero null-hypothesis significance tests can be used; and (3) how scale boundaries can create the illusion of proportionality, while other analytical procedures (e.g., post-hoc classifications) can augment this problem. Results. Our simulations highlight several limitations of common methods for analyzing recovery over time. Critically, we find that uniform recovery (in the population) leads to similar group-level statistics (regression slopes) and individual-level classifications (into fitters and non-fitters) that have been claimed as evidence for the proportional recovery rule. Conclusions. Our results highlight that one cannot identify whether proportional recovery is true or not based on commonly used methods. We illustrate how these techniques (regressing change scores onto baseline values), measurement tools (bounded scales), and post-hoc classifications (e.g., "non-fitters") can create spurious results. Going forward the field needs to carefully consider the influence of these factors on how we measure, analyze, and conceptualize recovery.