Question: Is it "quantitatively sound" to decompose/upsample year-bins (e.g., 2002-2006) into the component years when analyzing Likert Score data that was collected as a recollection of conditions within the year-bin?
Background: A series of interviews were conducted asking to Likert Score (1-5) the condition of locations over a grid, across time within pre-defined year-bins. These data are to be the response in a Random Forest algorithm. Predictor variables include stationary habitat variables where there's only one representation for the full time period, and others that have multiple or continuous measurements.
Data structure A is un-binned, as proposed in the question. There is higher predictor variable resolution, but the temporal resolution mismatches the year-bins asked to recall upon.
A: Individual years, Likert Score repeated, Predictor variables summarized annually
Data structure B is binned according to how the Likert Question was posed to the participants. Predictor variables have comparatively lower resolution to Data Option A, though the data are on the same timescale (bin) as the Likert Response.
B: Year-bins, Likert Score not repeated, Predictor variables summarized (averaged) within year-bins
