Description
Test whether a predictor at one prompt foretells the outcome at the next, controlling for the outcome's carry-over and time trends, with person-varying effects and uneven intervals handled.
Concurrent associations in repeated data cannot show direction. Using the previous prompt's predictor puts the presumed cause first. Adding the lagged outcome with a random slope controls autocorrelation (emotional inertia), so the lagged predictor shows its unique effect; adding time of day and day, with random effects, guards against confounding by daily rhythms, without which autocorrelation is biased. Effective lags can be long (about 135 minutes when prompts are 90 minutes apart and events are recalled since the last beep), and intervals vary, so analysts interact the lag with interval length or use continuous-time models. Bolger and Laurenceau show a spatial-power AR(1) error structure that uses exact time gaps, and that maximum-likelihood estimates stay valid when days are missing at random (rerunning with 20% of days deleted gave the same conclusions). Overnight lags are set to missing; in the handbook example about 23% of rows were lost to lagging and missingness.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Discovery, Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Build lagged predictor and outcome variables within day and person, setting overnight lags to missing.
- Fit the outcome on the lagged predictor with a random intercept and random slope.
- Add the lagged outcome with a random slope to control autocorrelation.
- Add time of day and day with random effects, plus any group interactions.
- Where spacing varies, interact the lag with interval length or use a spatial-power (continuous-time AR(1)) error structure.
- Fit by maximum likelihood so missing occasions are handled under missing-at-random assumptions.
- Report the number of observations used and the share lost to lagging.
Sources
- The Open Handbook of Experience Sampling Methodology (Myin-Germeys & Kuppens, 2022)
- Intensive Longitudinal Methods (Bolger & Laurenceau, 2013)
Also known as
- Lagged multilevel regression
- Multilevel autoregressive model with covariates
- Autocorrelated errors for missing occasions and unequal intervals
- Spatial power covariance
- Continuous-time AR(1) errors
Categories
- Analytics & Behaviour
- Statistics & Modelling