--- title: "The Extended Frame of Reference model" author: "Joshua A. McGrane" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{The Extended Frame of Reference model} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r setup, include = FALSE} knitr::opts_chunk$set(collapse = TRUE, comment = "#>") options(digits = 4) source("precomputed.R") ``` ```{r library} library(rasch) ``` ## Frames and units Rasch's criterion of invariant comparison is stated within a frame of reference (Rasch 1961). The extended frame of reference model allows the unit to differ over linked combinations of item sets and person groups (Humphry 2005; Humphry and Andrich 2008). For item \(i\) in set \(s\) and person \(n\) in group \(g\), \[ P(X_{ni}=x)= \frac{\exp\{\rho_{sg}[x\theta_n-\sum_{k=1}^{x}\delta_{ik}]\}} {\sum_{y=0}^{m_i}\exp\{\rho_{sg}[y\theta_n- \sum_{k=1}^{y}\delta_{ik}]\}},\qquad \rho_{sg}=\alpha_s\phi_g . \] Following Humphry and Andrich (2008, eq. 15), the multiplicative parameter of a frame is the ratio of the common reference unit to the frame's own unit: \(\alpha_s\) carries that ratio for item set \(s\), and \(\phi_g\) for person group \(g\). Their geometric means are fixed at one; no observed set or group is the reference. A set with \(\alpha_s = 1.3\) has a smaller natural unit than the geometric-mean set unit, so its response curves are steeper on the common scale. Two observed sets can be compared directly by the ratio \(\alpha_s/\alpha_t\). The partial credit model holds within each frame in its natural unit; dichotomous items are the one-threshold case. Use this model when variation in the unit is part of the measurement account. Poor fit to an equal-unit model is not by itself evidence for a frame-dependent unit. ## Fit the model The distributed vignette uses a precomputed fit, including its bootstrap uncertainty. The code below reproduces the calculation. ```{r simulate} d <- simulate_efrm( n_per_group = 150, items_per_set = 8, set_unit_ratio = 1.30, group_unit_ratio = 1.10, seed = 25 ) truth <- attr(d, "truth") d$site <- rep(c("A", "B"), length.out = nrow(d)) d ``` The simulated data carry the units the fit below has to recover: a set-unit ratio of 1.30 and a group-unit ratio of 1.10. ```{r fit, eval = recompute} fit <- rasch_efrm( d, item_sets = truth$item_sets, groups = "group", id = "id", factors = "site", boot_reps = 50, workers = 1, seed = 25 ) fit ``` ```{r fit-precomputed, echo = FALSE} if (!recompute) { fit <- vignette_result("extended-frame-reference")$fit fit } ``` The principal model-specific tables are: ```{r tables} fit$phi_table # person-group units fit$alpha_table # item-set unit ratios (reference unit / set unit) fit$set_table # linked set locations fit$frames # complete frame units fit$linking # set-linking design ``` The frame-defining group is part of the model and is not retested as DIF. Other person factors can be examined with `dif_anova()`. Its bootstrap conditions on each person's subtotal within each observed item set, then repeats the Extended Frames fit and the complete DIF analysis. Its familywise probabilities refer to the fitted global invariant null. ```{r dif-bootstrap, eval = FALSE} dif <- dif_anova(fit) dif_bootstrap(fit, dif, B = 999, seed = 2026)$summary ``` Person-group units are estimated from common threshold patterns across groups. Set units are estimated from persons observed in more than one set. For a linked pair, the natural coordinates satisfy \(u_b=r u_a+c\). The package estimates \(r\), \(c\), and masses for the nuisance person distribution on a finite grid by semiparametric likelihood: \[ \prod_n \int P(X_{na}\mid u)P(X_{nb}\mid ru+c)\,dF_{g(n)}(u). \] The masses are estimated jointly with the link, separately for each observed person group, so neither a normal shape nor a common distribution across groups is prescribed. The link holds the conditional item thresholds and person-group units fixed while estimating only \(r\), \(c\), and the nuisance masses. The common-scale parameters are then \[ \delta_{ik}=\widetilde\delta_{ik}/\alpha_s+\mu_s, \qquad \rho_{sg}=\alpha_s\phi_g. \] The item thresholds and person-group units remain pairwise conditional estimates; the mixing distribution is used only for the relative set unit, which is not identified by that conditional stage. The discrete nonparametric margin follows Follmann (1988); its use for linked item-set units is an extension implemented here. The group-by-set frame graph and the set-linking graph must each connect to a common scale. A linking response pattern must span a score range of at least four within a set. Overlapping item sets are not permitted. Several group factors may define crossed frames: ```{r crossed, eval = FALSE} fit_crossed <- rasch_efrm( data, item_sets = item_sets, groups = c("language", "cohort"), id = "id" ) fit_crossed$phi_factorial fit_crossed$phi_factorial_tests ``` The factorial tables give the GLS decomposition of the estimated log group units. Use `p_adj`, which applies Holm's correction across the factorial terms, for decisions. The tables do not replace the frame-level estimates. ## Uncertainty and equal-unit comparisons The default hybrid covariance combines the pairwise Godambe covariance with a person bootstrap for set linking. Within-frame thresholds and group units are jointly redrawn, and the link is rebuilt from each draw. The resulting covariance is retained for the common-scale thresholds and complete frame units. Set `se_method = "bootstrap"` to refit the complete model to each person resample. The linking calculations use a compiled numerical kernel. Bootstrap replicates run on four workers by default, or fewer where the system limit is lower. A fixed `seed` gives the same samples and results for every worker count without altering the caller's random-number stream. In the Shiny application the fit runs in a separate process. The analyst can select up to four workers, follow completed batches, or cancel the process without replacing the current analysis. R scripts can supply `progress` and `cancel` callbacks for the same controls. The fit records the requested, usable and failed bootstrap counts. The requested number must exceed the free directions in the largest covariance block used by the fit. The stored bootstrap row contains several constrained blocks and is not treated as one covariance. Once estimation begins, the linking covariance needs at least 30 usable replicates and a majority of those requested; otherwise the fit stops. A numerical failure or non-convergence on any supported link invalidates the whole replicate. Pairs without enough informative common persons can be omitted only if the remaining links connect all sets. Full-bootstrap refits with unidentified group units are also discarded. The conditional calibration checks the exact likelihood curvature as well as the score. Stationary saddles and flat solutions are refused; the same check applies to bootstrap refits. Saved frame fits without a current likelihood-check record must be refitted before reopening in the app. So must an extended frame fit whose stored score curves predate the shared design enumeration, since their designs and labels are no longer the ones `test_information()` and the curve plots would give. Neither refusal alters the source file. The alpha--phi cross-covariance has the same usable-draw rule and is withheld if it cannot be estimated. A full bootstrap that falls below its minimum returns hybrid standard errors instead, retaining its attempted, usable and failed counts in `full_boot_reps_*`. ```{r comparison} fit$efrm_vs_rasch$unit_omnibus fit$efrm_vs_rasch$unit_tests ``` The omnibus Wald tests assess the set- and group-unit families. Their probabilities are adjusted together by Holm's method. Individual unit contrasts form a separate Holm-adjusted follow-up family. The raw composite-likelihood difference is descriptive and compares the group-unit stage only. Set units are identified in the linking stage and are assessed by their Wald test. `compare_fits()` withholds generic likelihood differences involving EFRM: its calibration likelihood uses within-set pairs, whereas ordinary PCM also uses cross-set pairs. The matched group-unit comparison is retained in `fit$efrm_vs_rasch`. Probabilities require at least 50 persons or effective persons in every group and a path between each pair of sets whose links each have at least 50 informative common persons. A person at the same extreme tail in both sets remains in the likelihood but does not count as link support. Sparse designs retain the unit estimates without an inferential probability. Two-set simulations under normal, bimodal and deliberately different group distributions gave set-unit bias within 0.004 log-units, empirical-to-reported SE ratios from 0.97 to 1.05, and null rejection from 4.0 to 5.0 per cent for the hybrid method. The complete person bootstrap was mildly conservative in the corresponding null design. Sets with only three dichotomous items were not identified; four-item sets were recovered without bias in adequately supported samples. A separate null study checked the complete decision families. With 200 persons per group and six items per set, 489 of 500 fits were analysed. The Holm familywise rejection rates were 3.5 per cent for the omnibus tests and 1.6 per cent for the individual unit contrasts. A boundary design with 100 persons per group and four items per set refused 493 of 1,000 fits and did not converge in another 24. Among the 483 analysed fits, the corresponding rates were 5.2 and 2.5 per cent. A design should not be used merely because it is formally identifiable. ## Item invariance across frames The fitted EFRM gives an item one location across its frames, scaled by the frame unit. `frame_invariance()` examines that restriction by calibrating each frame separately. Each compared set and frame needs at least 50 distinct persons contributing an informative item pair: not both zero or both at their item maxima after any item removal or category recoding in that frame's separate calibration. Items with weak separate-frame standard errors are excluded from the comparison. For frame \(f\), the common-scale item location is \(\hat\delta_{if}^{*}=\hat\delta_{if}/\hat\rho_f\). The separate calibrations have independent origins, so differences are centred over the common thresholds before testing. Let \(w_i=m_i/\sum_jm_j\), where \(m_i\) is the number of thresholds for item \(i\). The default covariance conditions on the fitted frame units: \[ C\{V_1/\hat\rho_1^2+V_2/\hat\rho_2^2\}C^{\mathsf T}, \qquad C=I-\mathbf 1\mathbf w^{\mathsf T} . \] ```{r invariance, eval = FALSE} inv <- frame_invariance(fit, se_method = "conditional") inv$summary inv$locations inv$discrimination ``` A whole-person bootstrap within person group retains each sampled row's item-set response pattern and includes uncertainty in the fitted units. A replicate is used only if it retains the same item-comparison family, and therefore the same centred frame origin, as the observed analysis: ```{r invariance-bootstrap, eval = FALSE} inv_boot <- frame_invariance( fit, se_method = "bootstrap", boot_reps = 300, seed = 1 ) ``` The conditional method returns raw and Holm-adjusted probabilities for the location comparisons. Its standardised-infit and slope-ratio columns are descriptive. The bootstrap returns discrimination probabilities as well and applies Holm adjustment to the combined location and discrimination family. In 300 null simulations with 500 persons per frame, the bootstrap SE ratios were 1.00 for locations and 1.03 for log discrimination ratios; 95 per cent coverage was 0.948 and 0.954, and combined Holm familywise error was 3.0 per cent. A separate 2,000-replicate check found that probabilities based only on the conditional standardised-infit comparison gave 7.1 per cent combined familywise error. The conditional method therefore leaves those probabilities undefined. At the same sample size, bootstrap power was 96.3 per cent for two items shifted by one logit, but 9.6 per cent for two items made 1.5 times as steep. Discrimination departures require appreciably more information. The location contrasts are relative to the mean difference of the common items. If a small number of items is shifted, centring also gives the remaining items non-zero relative contrasts. The table identifies the pattern of departures; item content or external anchors are needed to decide which items provide the defensible reference. ## Refit after a departure `resolve_frames()` gives an item a separate location in each frame. The item continues to contribute to person measurement but no longer links those frames. `drop_items()` removes it. Both functions refit the model and update the frame, item and person estimates. Each set must retain common items linking its groups' origins; a unit link through another set is not sufficient. ```{r repair, eval = FALSE} resolved <- resolve_frames(fit, "S1I02") removed <- drop_items(fit, "S1I02") ``` The remaining common items and linked sets must still identify the frame units. The functions preserve the original crossed factors, uncertainty method and fitting controls, and return a result only when the revised model converges. ## References Follmann, D. (1988). Consistent estimation in the Rasch model based on nonparametric margins. *Psychometrika*, 53, 553--562. https://doi.org/10.1007/BF02294407 Humphry, S. M. (2005). *Maintaining a Common Arbitrary Unit in Social Measurement*. PhD thesis, Murdoch University. Humphry, S. M., and Andrich, D. (2008). Understanding the unit in the Rasch model. *Journal of Applied Measurement*, 9(3), 249--264. Rasch, G. (1961). On general laws and the meaning of measurement in psychology. In *Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability* (Vol. 4, pp. 321--333). University of California Press.