Assessing the feasibility of linking Section G and Section GG measures of function in skilled nursing facilities

Abstract

Background and Objectives:

In skilled nursing facilities (SNFs), Section GG replaced Section G as the instrument for measuring function in 2023. To inform how previously validated scales and scores derived from Section G items might be re-produced from Section GG items, we examined key measurement assumptions in both instruments.

Research Design and Methods:

We identified 1,131,694 Minimum Data Set 3.0 assessments of unique persons with complete Sections GG and G between October 2018 and September 2023 across 2,200 SNFs in the Long-Term Care Data Cooperative. We examined: (i) assessment instructions, (ii) item-level Spearman’s correlations, (iii) measures of fit and factor structure in confirmatory and exploratory factor analyses (CFA and EFA), and (iv) graded response item response theory (IRT) parameters.

Results:

Several Section GG items were conceptually similar and highly correlated (ρ ≥ 0.7). After removing redundant items, a combined scale of both instruments was not unidimensional in CFA (Comparative Fit Index = 0.753, Tucker-Lewis Index = 0.724, root-mean-square error of approximation = 0.153) and items from each instrument generally loaded onto separate but correlated factors (inter-factor correlation 0.735) in EFA. In IRT analyses, Section GG items better captured independence while Section G items better captured dependence.

Discussion and Implications:

Sections GG and G measure related yet distinct constructs, emphasizing the need for caution when interpreting a sum of Section GG items as a reflection of underlying health in SNF patients. Scales and scores derived from Section G items cannot be presumed to be equivalent when derived from Section GG items, and further work is needed examining interchangeability.

Background

Function—the ability of a person to perform tasks independently—is a cornerstone of clinical geriatrics and aging research (Avlund, 1997; Noelker & Browdie, 2014). Originally developed to study changes in health and well-being among chronically ill and older persons in residential care settings (Katz et al., 1970), function has been used to illustrate the consequences of disease and treatment (Gure et al., 2010; Kurella Tamura et al., 2009), determine needs for services (Fries et al., 1994), measure recovery in rehabilitation (Shi et al., 2022), facilitate communication between providers, and prognosticate (Steinman et al., 2023). In biopsychosocial models of health such as the World Health Organization’s International Classification of Functioning, Disability and Health and Verbrugge and Jette’s disablement process, function is a core construct mediating how our health conditions impact our interactions with the world around us (Verbrugge & Jette, 1994; World Health Organization, 2002). Within this theoretical framing, the ability to perform a set of tasks holistically reflects underlying health (DeVellis & Thorpe, 2022).

The largest sources of data on function (referred to as “functional measures” or “measures of function” throughout this paper) in the US are the standardized assessments collected on the millions of patients cared for annually in post-acute and long-term care settings. These include the Minimum Data Set (MDS) for skilled nursing facilities (SNFs), the Outcome and Assessment Information Set (OASIS) for home health agencies, the Inpatient Rehabilitation Facility Patient Assessment Instrument (IRF-PAI), and the Long-Term Care Hospital Continuity Assessment Record Evaluation Data Set (LCDS). The data are used for clinical care, billing, quality reporting and improvement. The data have also been used for research on topics ranging from the impacts of drugs on function (Steinman et al., 2017), trajectories of dependence after the start of dialysis (Kurella Tamura et al., 2009), rehab potential in cancer patients (Singh et al., 2022), and differences in post-acute rehabilitation quality based on insurance (Kumar et al., 2018).

Historically, measures of function differed between care settings because of subtle differences in paradigms of measurement, leading to challenges in communication, care coordination, quality measurement, and research (Gage et al., 2012). In 2014, the US Congress responded by signing the Improving Medicare Post-Acute Care Transformation (IMPACT) Act to harmonize functional measures across settings, prompting development and validation of the Continuity Assessment Record and Evaluation (CARE) item set (McMullen et al., 2022). The CARE item set consists of functional measures grouped as (1) Self-Care, (2) Mobility, and (3) Instrumental Activities of Daily Living.

In SNFs, the Self-Care and Mobility CARE items were gradually incorporated into the MDS as “Section GG: Functional Abilities and Goals” starting in 2016, with the intent of eventually replacing the legacy functional measures of Section G. Section GG initially contained 12 CARE items that were completed alongside Section G. In 2018, Section GG was expanded to include 24 CARE items. In October 2023, Section GG added 2 more Mobility items and Section G was removed, leaving Section GG as the only measure of function in the MDS (eTable 1).

This evolution of functional measurement produces challenges for research. With Section G’s removal in 2023, many scores and scales derived from Section G items can no longer be calculated. For example, the MDS-ADL Long Form is a highly cited 28-point scale validated as an overall measure of function based on items from the now-retired Section G of the MDS (Morris et al., 1999). It relies on self-performance of seven activities and was designed to track changes in overall health and well-being in long-term nursing home residents. Without Section G, the MDS-ADL Long Form can no longer be calculated, hindering comparisons with historical and international long-term care populations; other countries still use the items in Section G (Morris et al., 2013). Other common measures in SNF research, such as the Cognitive Function Scale (CFS) for cognition (Thomas et al., 2017), the MDS-Changes in Health, End-stage disease and Symptoms and Signs (CHESS) score for predicting readmission (Ogarek et al., 2018), the MDS 3.0 Mortality Risk Score (MRS3) (Thomas et al., 2019), and definitions of “low-care” residents rely on specific items in Section G and may similarly be affected (Mor et al., 2007).

In theory, it should be possible to construct a crosswalk to link Section GG to Section G. Section GG items assess similar activities compared with Section G items, and a study of 147 long-term SNF residents found that conceptually similar items between instruments were well-correlated, providing evidence of convergent validity (Marino et al., 2024). Another study of national MDS data of post-acute SNF patients from 2017 found Section GG scores at discharge were generally higher than scores on admission (Toth et al., 2023), supporting construct validity. The initial validations of Section GG showed good interrater reliability of items within and across care settings, with fit statistics suggesting the Self-Care and Mobility domains of Section GG represent a single construct based on exploratory factor analyses in cross-setting populations (Tobin et al., 2012).

However, no guidance exists on linking these two instruments to recreate scales and scores previously derived from Section G but using Section GG items. To develop such guidance, several assumptions about the validity of combining the items of Section GG into a scale (e.g. a reflective measure of function) and comparing it to Section G must be examined using real-world SNF data (DeVellis & Thorpe, 2022). Our study had three main aims: (i) identify which, if any, Section GG items are redundant, (ii) examine whether Section GG and Section G measure the same construct, and (iii) identify how items across both instruments align in measuring function. We hypothesized that the items of Section GG and Section G would exhibit unidimensionality, implying that they measure the same underlying construct and enabling valid development of a crosswalk.

Methods

Study design and data sources

This cross-sectional study was comprised of individuals with complete Section GG and Section G assessments, leveraging the period between 2018 and 2023 when both instruments were required (Kolen & Brennan, 2014). We obtained MDS data from the Long-Term Care (LTC) Data Cooperative, a data-sharing consortium with SNF providers with electronic medical record (EMR) and MDS data from over 2200 SNFs throughout the United States as of October 10, 2024 (Dore et al., 2022). This secondary data analysis was determined exempt by the Colorado Multiple Institutional Review Board.

Study participants

We identified MDS 3.0 v1.16 or v1.17 assessments containing complete Section GG and Section G between October 1, 2018 and September 30, 2023. These assessments were collected for patients admitted to SNF for short-term rehabilitation (i) upon admission and (ii) upon planned discharge for stays lasting 3 days or longer, excluding discharges to hospital. In October 2020, some states additionally elected to complete Section GG on (iii) routine assessments of long-term residents. We excluded MDS assessments before September 30, 2018 as they included only a limited subset of Section GG items. For individuals with multiple assessments, we randomly selected one assessment for inclusion.

Measures

Both Section GG and Section G instruments measure function, but subtle differences exist in the activities assessed, activity descriptions, the assessment period, and how performance is rated (eTable 2). These differences are common in functional measurement and can lead to differences in impairment prevalence or the underlying construct being measured (Ankuda et al., 2022; Avlund, 1997). Briefly, Section G asked trained assessors to rate a patient’s self-performance in 11 activities from 0 to 4, with 0 being independent and 4 being total dependence. Section G assessments follow the “Rule of 3” where assessors code the highest level of dependence observed at least 3 times over the last 7 days. On the other hand, Section GG asks assessors to rate a patient’s usual performance (instead of worst performance) in activities over the last 3 days from 01 to 06, with 01 being dependent and 06 being independent (the inverse of Section G). Section GG rates performance in 24 items categorized as two sub-domains: (i) Self-Care and (ii) Mobility. Some items have skip patterns. For example, persons unable to go up 1 step are not assessed for 4 steps or 12 steps.

We extracted additional baseline characteristics from the MDS and EMR data. These included age (<65, 65–79, 80–89, 90+; the LTC Data Cooperative suppresses birth year for persons 90 and older), sex (male / female), self-reported race and ethnicity (Black non-Hispanic, Hispanic White, White non-Hispanic, Other, which included American Indian or Alaska Native, Asian, Native Hawaiian or Other Pacific Islander, Unknown, or missing), assessments per facility, first year each facility joined the LTC Data Cooperative, U.S. Census region based on facility state, year of MDS being completed, reason for MDS completion, presence of dementia, and presence of Parkinson’s disease. Years were defined from October onwards as MDS version changes occur in October. Presence of dementia was defined using MDS or diagnosis codes documented in the EMR problem list (McCarthy et al., 2022). Presence of Parkinson’s disease was defined using MDS and EMR diagnosis codes and orders for Parkinson’s disease-specific medications (eTable 3) (Noyes et al., 2007).

Analysis

We examined similarities and differences in how function is measured in Section GG and Section G and reviewed annual MDS manuals to record changes in coding instructions and items over time. We generated a heatmap of pairwise Spearman’s correlations to visualize the relationship between all Section GG and Section G items and inform analyses.

We used classical test theory (CTT) and item response theory (IRT) to examine the relationships between the Sections. CTT assumes that the observed scores are a combination of latent true scores and measurement errors and that all items have equal weight in their summary of the underlying trait. IRT assumes scores are a function of latent individual ability and item parameters (Krabbe, 2017), modeling items as having varying levels of difficulty based on response patterns across respondents. IRT methods are used to develop standardized tests of aptitude such as the SAT (Scholastic Aptitude Test) and are the basis for assigning higher credit for questions that are more difficult (e.g., fewer students answer correctly). IRT methods are increasingly being used to model function based on the intuition that activities like eating are easier to complete than walking 150 feet (Marfeo et al., 2019).

Analyses were grouped by two types of linkages we envisioned other researchers might want to conduct:

1. To link scales, we needed to examine the extent to which items in Section GG could be combined to serve as an overall measure of function like the MDS-ADL Long Form. A prerequisite for scaling and linking scales is unidimensionality (Dorans & Holland, 2000). We used CTT-based confirmatory factor analyses (CFA) to test the unidimensionality of (i) all Section GG and Section G items together as a combined scale, (ii) Section GG items together and as sub-domains of Self-Care and Mobility, and (iii) Section G items together and as MDS-ADL items alone. We assessed fit using the Comparative Fit Index (CFI; >.90 = adequate fit, >.95 = very good fit), Tucker-Lewis Index (TLI; >0.90 = adequate fit, >.95 = very good fit) and root-mean-square error of approximation (RMSEA; <.10 = adequate fit, <.05 = very good fit). Several Section GG items were conceptually similar and highly correlated based on the Spearman’s correlation. We therefore repeated CFA using a parsimonious set of Section GG items to reduce the influence of redundant items distorting measures of fit (DeVellis & Thorpe, 2022). The parsimonious set included only one Section GG item for each Section G item in the MDS-ADL – specifically, the Section GG item with the highest correlation to a Section G item. CFA was conducted using the default maximum likelihood estimation in the ‘lavaan’ package in R.

   How items group together based on covariance is known as a factor structure. While CFA tests the fit of a hypothesized factor structure, exploratory factor analyses (EFA) identify possible factor structures without imposing a priori constraints. We therefore used EFA to see if alternative factor structures existed outside the theoretical constructs of unidimensional function, Self-Care, and Mobility specified in our CFA. Starting with the subset of MDS-ADL items and the parsimonious set of Section GG items, we used Bartlett’s test of sphericity (significant, p < .001) and the Kaiser-Meyer-Olkin criterion (suitable, overall 0.953) to determine suitability for EFA and determined the number of factors using parallel analysis and a scree plot. Parallel analysis recommended 9 factors, but large samples are prone to over-retain factors. Based on the scree plot (eFigure 1), we implemented 2- and 3-factor models with an oblique Promax rotation. EFA was conducted with the ‘EFAtools’ package.

2. Other researchers may want to link items, as some scores and definitions rely on individual Section G activities (e.g., eating) alone (Morris et al., 1994; Thomas et al., 2017). This requires item-level comparisons to examine if individual Section G items are equivalent to Section GG items. Using CTT, we examined the correlation between overall independence (as a count of Section GG and Section G activities completed independently) with independence in individual activities. Using IRT, we estimated item difficulty threshold parameters based on the graded response model (GRM) for ordered categorical responses, wherein responses reflect increasing levels of ability (Hays et al., 2000). We first estimated item parameters for MDS-ADL items because of their unidimensionality and familiarity in research. We then used fixed parameter calibration to produce parameters for Section GG items (Kim, 2006). IRT analyses used the ‘mirt’ package.

In post-hoc analyses, we hypothesized dimensionality and factor structure could vary by subgroups. We explored this variance by repeating unidimensional CFA and 2-factor EFA on the parsimonious set of items, stratified by routine vs. non-routine assessment, presence of dementia, duration of facility participation in the LTC Data Cooperative, and sex.

Results

We identified 1,131,694 unique individuals with complete Section GG and Section G data (Table 1; eFigure 2). 167,754 (15%) individuals were younger than 65 and 248,278 (22%) individuals were 90 years or older. 652,464 (58%) were female, 148,510 (13%) were Black non-Hispanic, 22,092 (2%) were Hispanic, 199,506 (18%) were Other, and 781,586 (67%) were White non-Hispanic. 274,351 (24%) had dementia, and 46,326 (4%) had Parkinson’s disease or secondary parkinsonism. Most assessments (76%) were completed for a rehabilitation stay paid for by Part A Medicare. Of these, most were admission assessments or discharge assessments. Fewer assessments (35%) were OBRA assessments (completed to satisfy federal requirements to assess all persons cared for by SNFs), of which most were admission, routine, or discharge assessments.

Table 1.

Baseline characteristics of unique SNF short-stay patients and residents with complete Sections GG and Section G in the Long-Term Care Data Cooperative.

	n	%
Overall	1,131,694	100
Age
<65	167,754	15
65–74	285,214	25
75–89	430,448	38
90+	248,278	22
Sex
Female	652,464	58
Male	478,900	42
Unknown	330	<1
Race / ethnicity
Black, non-Hispanic	148,510	13
Hispanic	22,092	2
Other race / ethnicitya	199,506	18
White, non-Hispanic	761,586	67
Dementiab	274,351	24
Parkinson’s disease or secondary parkinsonismc	46,326	4
US Census region
Midwest	412,657	36
Northeast	151,877	13
South	494,268	44
West	72,892	6
Year of MDS assessmentd
2018 – 2019	165,331	15
2019 – 2020	169,296	15
2020 – 2021	237,811	21
2021 – 2022	255,719	23
2022 – 2023	303,537	27
Number of assessments per facility, mean (SD)	558.0	503.2
Rehabilitation assessmentse	866,026	76
Admission	710,376	62
Dischargef	151,579	13
Routine	3,878	<1
Change in status	193	<1
OBRA assessmentse
Admission	393,094	35
Discharge	190,105	17
Routine (annual or quarterly)	126,222	11
Change in status	17,784	2
Correction	47	<1

^aOther race / ethnicity included American Indian or Alaska Native, Asian, Native Hawaiian or Other Pacific Islander, Unknown, or missing.

^bDetermined by ICD-10 codes in electronic health record data and MDS codes.

^cDetermined by ICD-10 codes in electronic health record data, MDS codes, and Parkinson’s disease medication orders.

^dYears of assessment start in October, the month when MDS changes are implemented.

^eRehabilitation and OBRA assessment types were determined based on A0310, which indicates the reason why the MDS assessment was completed. Rehabilitation and OBRA assessments are not mutually exclusive.

^fThe number of rehabilitation admission assessments can exceed the number of discharge assessments as discharge assessments were not required for death, discharge to hospital, unplanned discharges, and if the initial assessment occurred within the past 3 days.

Several Section G items had Section GG counterparts; these were grouped into conceptually similar sets (eTable 2). For example, Section G “dressing” (G0110G) asked assessors to rate how individuals put on all items of clothing, but Section GG assessed “upper body dressing” (GG0130F), “lower body dressing” (GG0130G), and “putting on/taking off footwear” (GG0130H).

The heatmap of correlations showed conceptually similar Section GG items were highly correlated (all Spearman’s correlations ≥ 0.7; Figure 1). Some – such as “sit to lying” (GG0170B) and “lying to sitting on bed” (GG0170C) – were very highly correlated (ρ = 0.97). This suggested redundancy within Section GG, motivating identification of a parsimonious set of Section GG items with only one Section GG item per counterpart Section G item (shown bolded on the left side of Figure 1) as described in our Methods. Wheelchair mobility items (GG0170R and GG0170S) were poorly correlated with most other items (ρ = 0.19 – 0.51). For the parsimonious set, we therefore collapsed wheelchair mobility items (GG0170R and GG0170S) with walking items (GG0170J and GG0170K) to recreate locomotion items (GG0170JR and GG0170KS), coded as the most independent of wheelchair mobility and walking. These composite items were more closely correlated with the Section G locomotion items than walking items alone (e.g., ρ = 0.53 vs. 0.47 with locomotion on unit, ρ 0.44 vs. 0.37 with locomotion off unit).

Figure 1.

Heatmap of pairwise correlations between Section G and Section GG measures of function.

Colored regions represent correlations within the pre-specified constructs of Section G (red), Section GG Self-Care (blue), and Section GG Mobility (green). Uncolored regions represent correlations between constructs. Cells with white text show high correlations: bolded text indicates very high correlation (≥0.8) and non-bolded text indicates high correlation (≥0.6). In the left column, bolded items are in the MDS-ADL (for section G) and the subset of parsimonious items (for Section GG). The parsimonious set of Section GG items kept only one item for each conceptually similar Section G item based on highest correlation, along with picking up objects and car transfers.

Scaling analyses

The combined set of all Section G and Section GG items was not unidimensional (CFI = 0.663, TLI = 0.641, RMSEA = 0.155; Table 2). This was driven by the lack of unidimensionality in the Mobility sub-domain of Section GG (CFI = 0.692, TLI 0.641, RMSEA = 0.242). The MDS-ADL items were unidimensional with good to very good fit (CFI = 0.975, TLI = 0.962, RMSEA = 0.094). Unidimensionality of Section GG improved after removing redundant items (CFI = 0.840, TLI = 0.808, RMSEA = 0.157 vs. CFI = 0.702, TLI = 0.671, RMSEA = 0.191), but the combined set of parsimonious Section GG items with MDS-ADL items was not unidimensional (CFI = 0.694, TLI = 0.660, RMSEA = 0.171).

Table 2.

Confirmatory factor analyses of Section G and Section GG.

Model being tested	Model χ2	CFI	TLI	RMSEA
All items
Section G and Section GG combined	13401419.267	0.663	0.641	0.155
Section GG (all items)	8615692.751	0.702	0.671	0.191
Self-Care subscale only	830381.356	0.895	0.842	0.229
Mobility subscale only	5961970.389	0.692	0.641	0.242
Section G (all items)	1188868.023	0.868	0.835	0.155
MDS-ADL subscale only	139893.867	0.975*	0.962*	0.094*
Parsimonious set
MDS-ADL and Section GG combined	18349471.749	0.753	0.724	0.153
Section GG	1810932.649	0.840	0.808	0.157
Self-Care subscale only	532698.413	0.879	0.757	0.307
Mobility subscale only	732048.765	0.865	0.811	0.180

CFI = Comparative Fit Index; TLI = Tucker-Lewis Index; RMSEA = root-mean-square error of approximation.

All models were unidimensional unless otherwise noted. Values marked with an asterisk (*) denote adequate measures of fit, per the following criteria: (1) CFI; >.90 = adequate fit, >.95 = very good fit, (2) TLI; >0.90 = adequate fit, >.95 = very good fit and (3) RMSEA; <.10 = adequate fit, <.05 = very good fit.

Differences in the constructs measured by Section GG and Section G were further reinforced in 2- and 3-factor EFA with oblique rotation (eTable 4). Section GG and Section G items generally loaded onto separate albeit correlated factors (inter-factor correlation ≥ 0.7). In 2-factor EFA, one factor included all Section G items and Section GG “eating” and “oral hygiene”, while the second factor included the remaining Section GG items. In 3-factor EFA, eating and oral hygiene items from both instruments loaded onto their own factor.

Item-level analyses

CTT and IRT analyses illustrated a hierarchy of item difficulty for Section GG activities relative to MDS-ADL activities (Table 3). For example, eating and oral hygiene were “easy” activities – the ones most likely to be preserved at low levels of ability. No Section GG item had a difficulty threshold comparable to total dependence in Section G “eating” (G0110H; b = −3.352 for extensive assistance, Figure 2, eTable 5). Conversely, Section GG included activities that were more “difficult” (e.g. 12 steps; b = 2.947 for independence) than any MDS-ADL items. This implies no items in Section GG precisely measured low levels of ability like the MDS-ADL, and no items in the MDS-ADL precisely measured high levels of ability like Section GG.

Table 3.

Relationship between the count of activities performed independently and independence in specific activities.

		Count of Section G activities (MDS-ADL only) performed independently
		7	6	5	4	3	2	1	0
Overall, n (1000s)		10.7	7.5	7.6	9.4	13.9	35.6	256.4	790.6
Indep. in… (col. %)	Dressing	100	79	51	34	23	7	0	0
	Personal hygiene	100	83	56	44	42	30	1	0
	Locomotion on unit	100	77	70	63	54	43	4	0
	Toilet use	100	90	73	42	18	4	0	0
	Transfer	100	94	81	62	33	8	0	0
	Bed mobility	100	94	85	72	51	21	1	0
	Eating	100	84	83	83	79	86	92	0

		Count of Section GG activities (parsimonious set only) performed independently
		13	12	11	10	9	8	7	6	5	4	3	2	1	0
Overall, n (1000s)		5.2	5.9	7.4	10.1	12.6	12.9	12.9	14.6	17.7	22.3	37.6	92.1	188.4	692
Indep. in… (col. %)	12 steps	100	42	20	9	4	4	2	1	0	0	0	0	0	0
	Car transfer	100	83	69	44	23	18	12	8	5	2	1	0	0	0
	1 step (curb)	100	99	80	38	17	13	8	5	2	1	1	0	0	0
	Pick up object	100	96	84	70	51	35	27	20	16	8	5	2	1	0
	Shower/bathe self	100	83	69	62	44	25	23	20	14	3	1	0	0	0
	Walk / wheel 50 feet	100	100	99	98	95	85	71	59	51	50	35	19	12	0
	Lower body dressing	100	99	96	93	89	78	62	51	37	18	5	1	0	0
	Walking 10 feet	100	100	99	98	95	83	64	48	37	25	8	2	1	0
	Toileting hygiene	100	99	98	96	94	88	78	66	55	35	18	3	1	0
	Chair / bed transfer	100	100	100	99	99	97	89	70	54	38	14	4	0	0
	Oral hygiene	100	100	99	98	96	91	84	77	68	66	66	55	2	0
	Lying to sitting on bed	100	100	100	100	99	98	97	94	82	79	62	30	16	0
	Eating	100	98	97	95	94	87	83	81	77	75	83	83	66	0

Columns group individuals by the total number of activities they can perform independently. Columns to the right indicate greater dependence. Cells indicate the likelihood of independence in each activity (by row) given the overall level of independence. For example, 92% of people independent in 1 MDS-ADL activity are independent in eating.

Figure 2.

Item map of difficulty of Section G and Section GG items.

Each dot is an item difficulty threshold showing how difficult it is to move from one level of dependence to a higher level of independence for a given activity. Thresholds to the right are more difficult to achieve and reflect higher levels of ability. Section G items are in red and have 4 markers, representing the difficulty of moving between total dependence (4), extensive assistance (3), limited assistance (2), supervision (1) and independence (0). Section GG Self-Care items are in blue and Mobility items are in green. Each item has 5 markers, representing the difficulty of moving between dependence (01), substantial / maximal assistance (02), partial / moderate assistance (03), supervision or touching assistance (04) setup or clean-up assistance (05) and independent (06). Items are grouped by domain and ordered by difficulty. An example of how to read this Figure: an individual with low functional ability (eg, θ = −3) is likely to be totally dependent in all activities except Section G Eating (G0110H), for which they would require extensive assistance (3) but not be totally dependent (4).

For conceptually similar items, the range of item difficulty thresholds for Section GG items overlapped the range of counterpart MDS-ADL items and were narrower. For example, the item difficulty threshold for independence in Section GG “lower body dressing” was 1.450 compared to 2.141 for Section G “dressing”. For total dependence in “upper body dressing”, the difficulty threshold was −1.247 compared to −1.628 for “dressing” (eTable 5). This implies that, taken together, sets of Section GG items (e.g., all three dressing measures) provide more precise measurement within a narrow range of abilities compared to counterpart MDS-ADL items, but the MDS-ADL items are better at differentiating between individuals across a broader range of ability.

Post-hoc analyses

In post-hoc analyses examining measurement invariance across subgroups, unidimensional fit was higher in the subgroup of routine assessments compared to the full SNF population (CFI = 0.793, TLI = 0.769, RMSEA = 0.152 vs. CFI 0.753, TLI = 0.724, RMSEA = 0.153; eTable 6). Better fit was not observed when stratifying by dementia, duration of facility participation in LTC Data Cooperative, or sex. In 2-factor EFA of routine assessments, MDS-ADL and Section GG items loaded onto a single factor, while difficult Mobility items such as “car transfer”, “1 step (curb)”, “12 steps” and “pick up object” loaded onto a second factor with lower correlation compared to the full population (0.643 vs. 0.735; eTable 7).

Discussion

Sections GG and G seem to both measure function. However, contrary to our hypothesis, Sections GG and G failed to demonstrate unidimensionality when combined in CFA and each Section’s items generally loaded onto distinct factors in EFA even after we removed redundant items. At the item level, no Section GG item captured the lowest levels of ability indicated by total dependence in eating in Section G, and several Section GG items captured higher levels of ability not measured by Section G. Furthermore, key measurement assumptions for re-creating established Section G scales and scores using Section GG items—such as unidimensionality and overlap in measured ability levels—were not met. This implies that even though others observed correlations between these instruments and with other external variables (Marino et al., 2024; Toth et al., 2023), Section G items should not be assumed to be directly interchangeable with Section GG items, and scales and scores will require empirical assessments of comparability before use in research.

The distinction between instruments appears to reflect underlying conceptual differences between them. Section G items, and more specifically the MDS-ADL, were designed and scaled to measure overall health and well-being in a long-term care population (Katz, 1963; Morris et al., 1999; Saliba & Buchanan, 2012). In contrast, Section GG was validated in an older adult population spanning care settings and is based on a rehabilitation paradigm, which emphasizes documenting improvements attributable to therapy (Dodds et al., 1993; Tobin et al., 2012). This paradigm entails assessing activities with greater granularity (e.g. “upper body dressing” vs. “dressing”) and assessing average performance over shorter time windows. In theory, this granularity helps specify which activities improve during rehabilitation. But in practice—either because MDS assessors are not recognizing these distinctions or for physiologic reasons (e.g. the body functions required to dress the upper body and lower body are similar)—the increased granularity introduced redundancy and distorted unidimensional fit, which must be accounted through the removal of some items when producing a summed scale of Section GG.

A rehabilitation paradigm must account for patients improving day by day and reflect not only long-term health and well-being but also responsiveness to rehabilitation. Per this paradigm, Section GG instructed raters to assess performance as an average over the past three days, making it more likely to produce better ratings than Section G in patients who improve quickly. Unsurprisingly, we observed worse unidimensional fit when combining items of both instruments in rehabilitation assessments, reflecting differences in the underlying constructs evaluated by each Sections because of the change in time period being assessed. Nevertheless, we observed poor unidimensional fit even in routine assessments of long-stay SNF residents in whom function is unlikely to change. This further suggests that the constructs underlying Section G and Section GG differ in ways beyond just the time period being assessed. The differences could lie in how clinicians approached Section GG assessments or the changes in how activities are defined and rated (Ankuda et al., 2022). For example, “bathing” in Section G includes transfers in and out of the tub, whereas “shower/bathe self” in Section GG excludes transfers.

Measures of unidimensional fit approached adequacy once we removed redundant items. Based on our findings, future work scaling Section GG should build off the parsimonious set of Section GG items and use composite locomotion variables. Future work must also explicitly clarify the goals for scaling (Dorans, 2004). For example, recreating the MDS-ADL from Section GG would benefit from excluding items in Section GG like “car transfer” and “12 steps” which extend beyond the construct measured by the MDS-ADL. This would reduce error and bias in crosswalks, but at the cost of greater ceiling effects, which are a limitation Section GG was designed to address (Tobin et al., 2012). By definition, crosswalk methods impose restrictions to allow bidirectional conversions of scales (e.g. Section GG scores to Section G, and Section G to Section GG). However, it is not clear if we need this bidirectionality because the MDS-ADL Long Form is no longer in use; a Section GG scale that keeps the difficult items and uses one-way prediction to compare to the MDS-ADL may suffice for historic comparisons, without compromising advantages of the new instrument (Dorans et al., 2007).

Our description of items (eTable 2) and item map (Figure 2) can serve as a guide to reproduce scores that depend on specific Section G items. Researchers should identify items and response levels measuring a conceptually similar activity and with a similar item difficulty parameter. Even so, specific Section GG items may not be able to differentiate between patients as well as their counterpart Section G items when assessing populations with a broad range of abilities. Importantly, we note that several established scores derived from Section G rely on total dependence in eating (Morris et al., 1994; Ogarek et al., 2018; Thomas et al., 2017, 2019), which we empirically showed is not captured well by any item in Section GG. These differences could impact the prognostic significance of derived scores, and scores should be re-evaluated when substituting Section GG items for Section G items to confirm that they still have the same interpretation.

Clinically, our findings emphasize the need for caution when interpreting Section GG as a reflection of underlying health alone, especially in the subset of SNF patients undergoing rehabilitation after hospitalization. In a chronically ill long-term care SNF resident, an inability to eat usually reflects advanced decline. However, in a SNF patient receiving post-acute care, an inability to eat could also reflect features of a recent acute hospitalization such as treatment for a head and neck cancer. Furthermore, our work has implications for policy. Starting in fiscal year 2027, eleven Section GG items will be summed as a Discharge Function Score and included as a quality metric in CMS’ pay-for-performance program (SNF Value Based Purchasing or VBP) (RTI International, 2025). The VBP program assumes that skilled services provided in SNFs improve function. A fair incentive structure requires that functional measures appropriately differentiate between SNF patients and within patients over time. Further work should evaluate if differences in how the new Section GG assesses function leads to differences in how SNFs are reimbursed.

Our study has a few limitations. The cohort represents participants in the LTC Data Cooperative, which is not a random sample of all SNF patients and residents in the US. We nevertheless confirmed results were similar between facilities with long-standing and more recent membership. We coded all non-assessments (including activity not performed due to environmental limitations, a new rating option in Section GG) as dependence. This follows the MDS-ADL convention of coding dependence whenever an activity is not performed for any reason, but may overestimate dependence in resource-limited settings where assessments are skipped (Technical Expert Panel (TEP) for Cross-Setting Function Measure Development, 2022). Finally, we conducted IRT despite imperfect unidimensional fit. We nevertheless confirmed the general ordering of item difficulties matched original validation studies in the CARE item set in cross-setting populations (Tobin et al., 2012).

Conclusion

By testing key assumptions underlying linking methods using real-world SNF data, we identified major concerns about redundancy, dimensionality, and the range of difficulty thresholds measured by Section GG items compared to Section G. We identified subsets of items that can be used to reproduce previously validated scales and scores, but scales and scores derived from Section G items cannot be presumed to be equivalent when derived from Section GG items. Given the role of function in policy and research, future work must examine the measurement properties of Section GG in the SNF population and recognize distinctions in the overall functional trajectory of SNF residents receiving long-term care versus and SNF patients receiving post-acute care.

Data Availability:

Data from the Long-Term Care Data Cooperative is available by application at https://www.ltcdatacooperative.org/pages/default.aspx. Code for cleaning functional measures from Section GG across different types of Minimum Data Set 3.0 assessments will be made available on the Brown Digital Repository at https://repository.library.brown.edu/studio/.