Skip to contents

silentema: dynamic missingness graphs and sensitivity analysis for EMA data

Source: R/silentema-package.R
silentema-package.Rd

Participants in ecological momentary assessment (EMA) and other experience-sampling studies skip prompts, and the reasons are rarely unrelated to the momentary states being measured. silentema provides a graphical language for stating which nonresponse mechanism is assumed, a checker that says which quantities of a two-level vector autoregressive model of order one (VAR(1)) remain estimable under that mechanism and by which estimator, two tests that ask the data whether skipped prompts were informative, estimators for the recoverable case, a sensitivity analysis for self-censoring (the mechanism that estimation alone cannot repair), three calibration designs for the sensitivity parameter, worst-case bounds, a reporting template, and a simulator.

The workflow

The recommended order of analysis is

  1. Declare the assumed mechanism as a dynamic missingness graph with dm_graph, built from the motifs M0 (completely random), M1 (lagged-state dependence), M2 (self-censoring), M3 (burden or fatigue), M4 (person propensity), M5 (context confounding) and M6 (reactivity), with optional sensor, probe and context nodes; inspect it with plot and query it with dsep.

  2. Check recoverability with recoverability: which estimands (transition kernel, person means, between-person law) are structurally recoverable, and which estimator recovers them.

  3. Test whether silence was informative with silence_test (the state after a skipped prompt) and, when an always-observed channel exists, sensor_gap_test; describe response persistence with fatigue_check.

  4. Estimate the within-person VAR(1) from answered adjacent prompts with fit_pairs (person intercepts, half-panel jackknife, cluster-robust inference through summary.pairs_fit), with covariate adjustment or inverse-probability weighting for an observed context (fit_ipw), or by full-information maximum likelihood under missing at random (fit_fiml).

  5. Profile the estimates over a self-censoring sensitivity value with fit_tilt and tilt_profile, and summarize the profile with break_even (sign changes, significance changes, identified sets and bands).

  6. Calibrate the sensitivity value from a passive sensor, randomized probes or the post-skip contrast with calibrate_delta; compare with the worst-case bounds of bounds_support.

  7. Report with missingness_declaration, which writes a Markdown declaration for preregistrations and papers.

Simulation and design

simulate_ema generates two-level VAR(1) data under any combination of motifs with a target response rate, optional sensor, probe, context and day structure; simulate_from_fit simulates from a fitted (tilted) model; default_params, stationary_cov and partial_cors give the population parameters used in the accompanying article and the derived quantities.

Data format

All estimators take a long data frame with one row per scheduled prompt, answered or not: a person identifier (id), an integer prompt index that increases by one from one scheduled prompt to the next (time), a response indicator (R, 1 = answered), the state columns (NA at skipped prompts), and optionally a day column (pairs are then formed within days only), a sensor column, a probe column and context columns. See make_pairs.

References

Yu, H.-T. (2026). What skipped prompts hide: Detecting, diagnosing, and correcting informative nonresponse in ecological momentary assessment. Manuscript under review. Materials: https://osf.io/x6d2t/.

Mohan, K., & Pearl, J. (2021). Graphical models for processing missing data. Journal of the American Statistical Association, 116, 1023-1037. doi:10.1080/01621459.2021.1874961

Dhaene, G., & Jochmans, K. (2015). Split-panel jackknife estimation of fixed-effect models. The Review of Economic Studies, 82, 991-1030. doi:10.1093/restud/rdv007

Shumway, R. H., & Stoffer, D. S. (1982). An approach to time series smoothing and forecasting using the EM algorithm. Journal of Time Series Analysis, 3, 253-264. doi:10.1111/j.1467-9892.1982.tb00349.x

See also

The vignettes: vignette("silentema-workflow") for a start-to-finish analysis, vignette("dm-graphs") for the motif taxonomy and recoverability, vignette("testing-informativeness") for the tests, vignette("sensitivity-analysis") for tilting and calibration, and vignette("simulation-and-design") for the simulator.

Author

Maintainer: Hsiu-Ting Yu hsiutingyu@gmail.com (ORCID) [copyright holder]