We used data from the KNHANES, an annual survey producing representative statistics on the health status, health behaviors, and nutrition intake of the Korean population [34]. The KNHANES operates in three-year cycles [34]. For this study, we analyzed data from the second year (2020) of the 8th cycle (2019–2021). The survey follows a multistage clustered probability sampling design [35]. Primary sampling units (PSUs) were enumeration districts, stratified by region, urbanicity, and housing type [34]. Full methodology and datasets are available online (https://knhanes.kdca.go.kr/knhanes/main.do, Korean only) [34, 35]. While most survey items remain consistent across years within a cycle, some items are included only in specific years as part of a rotating module (Table 1). Using the standardized ICF linking rules [36], we previously identified 15 functioning-related items from the KNHANES 2020 questionnaire [27]. Those items were linked to their corresponding ICF categories [27], and were used in this study.

We applied three exclusion criteria: (1) participants under 19 years old; (2) more than 30% of missing responses across functioning items; (3) functioning items with a missing rate over 30%. The 30% cutoff was adopted to ensure measurement reliability [22, 37]. In the original KNHANES coding, scoring directions were inconsistent across items; in some items, higher scores indicated more difficulties in functioning, while in others, higher scores represented fewer difficulties. To ensure higher scores consistently reflect better functioning, reverse coding was applied to negatively worded items.

Missing responses were minimal (<0.7%) across all functioning items except BP17_3, which showed structural missingness because it was administered only to participants aged ≥40 years. Item-level mean replacement was applied to all functioning items with missing values to preserve the observed distribution.

A total of 7,359 individuals participated in the 2020 KNHANES. After excluding those under 19 years of age (n = 1,226) and individuals with more than 30% missing responses across the 15 functioning items (n = 720), 5,413 participants remained for analysis (Figure 1). The age range of participants was 19–80 years and over. In the KNHANES dataset, individuals aged ≥80 years were grouped together and coded as 80. Therefore, we provide the weighted median age, which was 47 years (interquartile range [IQR] = 26). Among them, 2,437 (50%, weighted) were male, and 2,279 (44.5%, weighted) resided in capital or metropolitan areas. Most participants were married and had at least a high school education. Regarding self-rated health, 4,367 (82.3%, weighted) reported their health as very good, good, or fair, while 17.7% (weighted) rated their health as poor or very poor. Descriptive statistics of the study population are shown in Table 2.

Originally, 15 functioning items were selected for analysis. However, one item—“absent from school or work due to illness or impairment (d850)”—had more than 30% missing responses and was excluded. The remaining 14 items were used to build the functioning metric, and their response distributions are shown in Table 3. These items were linked to ten distinct ICF categories: five under the Body Functions component (e.g., mental functions, sensory functions, and pain), and the rest under the Activities and Participation component (e.g., self-care and mobility) [15].

In most items, the majority of participants reported no or only minimal difficulties in functioning. However, a few specific items showed notably lower rates of no problems in functioning. The lowest rates were observed for BP1 - feeling stress in daily life (15.4%), linked to d240 Handling stress and other psychological demands, and BP_PHQ_4 - tiredness, low energy (59%), linked to b130 Energy and drive functions. Two items (BP_PHQ_3 and BP17_3), linked to b134 Sleep functions, also showed relatively low rates (65.6% and 78.6%).

The initial Rasch analysis revealed that the data did not meet the model’s assumptions. At baseline, some of the items presented LID, and threshold disordering. The scale as a whole was multidimensional. To address these issues, we followed Baghaei’s approach [39] and created two testlets combining correlated items:

Testlet 1 (Physical function & wellbeing): LQ_1EQL - exercise ability; LQ_2EQL - self-care; LQ_3EQL - daily activity; LQ_4EQL - pain/discomfort; LQ4_00 - activity limitation.

After applying testlets, the model showed no LID, supporting unidimensionality (second eigenvalue = 1.33). However, several items exhibited threshold disordering, likely due to skewed response distributions, where very few respondents selected the lowest scores (indicating severe problems in functioning) (Table 3). Accordingly, four items (BP_PHQ_1 - little interest or fun in doing work; BP_PHQ_3 - difficulty falling asleep, waking up, or sleeping too much; BP_PHQ_4 - tiredness, low energy; BP_PHQ_5 - loss of appetite or overeating) were initially recoded by collapsing adjacent response categories. For these items, original scores 0 and 1 were collapsed into 0, 2 into 1, and 3 into 2. This resolved disordering in BP_PHQ_1, BP_PHQ_3, and BP_PHQ_4. However, BP_PHQ_5 required further recoding into a dichotomous format (0–2 to 0; 3 to 1) to eliminate disordering.

In a subsequent step, item fit for Testlet 1 was above 1.2 and for Testlet 2 was 0.4, requiring adjustment. For Testlet 1, response categories were collapsed step by step until a dichotomous format was achieved: raw scores 0–8 were collapsed into 0 (“have a problem”) and 9 into 1 (“no problem”), improving the item fit to below 1.2. For Testlet 2, dichotomizing its original categories improved the item fit to 0.58. The detailed R code for these steps is provided in the Supplementary Material 1.

Table 4 compares psychometric properties between the initial analysis and the testlet solution (as previously described; hereafter ‘testlet solution’). With the testlet solution, Cronbach’s alpha decreased slightly from 0.79 to 0.74 but remained >0.7, indicating acceptable reliability for population surveys. In contrast, the PSI was 0.55, which may be attributed to the skewed score distribution, with most participants showing high levels of functioning. As a supplementary analysis, we used weighted sampling by 20-point intervals of functioning score (N = 120). This yielded a PSI of 0.81, suggesting the lower value in the full sample may reflect score skewness.

In the testlet solution, the mean person ability estimate decreased from 2.51 to 1.80, bringing it closer to the mean item difficulty (from −0.21 to −0.01). This indicates better alignment between person abilities and item difficulties. However, as the discrepancy still >0.5 logits, the items appeared generally too easy for the sample, implying the scale may be better suited to assessing individuals with higher disability levels than those in this sample. We conducted additional analysis to examine whether person-item alignment differed by age group. However, no significant differences in targeting were found across age groups.

The fit statistics of the testlet solution are shown in Table 5. Four items (BP_PHQ_1, BP_PHQ_4, BP_PHQ_5, and Testlet 2) had outfit values <0.7, indicating overfit. This suggests that these items/testlet predicted response patterns too well. All other items demonstrated acceptable infit and outfit MNSQ values within the commonly recommended range (0.7–1.3). Regarding item difficulty, BP1 - feeling stress in daily life had the highest value (0.97), which indicated that respondents found it hardest to endorse this item. In other words, most individuals, regardless of their level of functioning, were more likely to report stress in daily life than other items. Conversely, LQ1_sb - bedridden status in the past month had the lowest difficulty (−1.77), meaning that most respondents did not report a bedridden status in the past month. BP1 also showed the widest measurement range (−1.14–3.60) among all items. DIF analysis revealed no notable differences by sex or region, but age-related DIF was found in two items: BP17_3 - sleepiness and Testlet 1 - physical function & wellbeing. The presence of DIF suggests that these items may function differently across age groups.

The results of the testlet solution are visualized in the Person-Item Map (Figure 2), showing the distribution of person abilities and item difficulties along the latent scale (functioning score). The upper section shows person parameters, indicating the relative ability levels of respondents (i.e., their levels of functioning), with individuals at higher levels of functioning appearing further to the right. The lower section shows item locations along the same latent scale, reflecting item difficulty (i.e., how likely respondents were to report no problems on that item), with items located further to the right representing more difficult items (e.g., BP1), and those on the left indicating easier items (e.g., LQ1_sb). Item difficulties are indicated by black dots, and threshold locations between response categories by white dots. The distribution of person abilities is slightly right-skewed compared to item difficulties, suggesting a potential ceiling effect (i.e., the metric assesses levels of disability above what is found in the population).

Table 6 presents the transformation of raw scores into a user-friendly functioning scores ranging from 0 (extreme problems in functioning) to 100 (no problems in functioning). In the initial analysis, the maximum total raw score was 31, as described in Table 2. However, in the testlet solution, response categories were collapsed, resulting in a total score ranging from 0 to 14.

This study has several limitations. First, LID was detected among some functioning items, necessitating the construction of testlets. This aggregation of multiple ordinal items increased the number of response categories, which appeared to contribute to threshold disordering [24]. Accordingly, response categories were dichotomized, resulting in a certain degree of information loss [39], which may reduce the precision of measurement. Second, items LQ_1EQL - exercise ability, LQ_2EQL - self-care, and LQ_3EQL - daily activities showed extremely low response frequencies for severe problems in functioning. Such irregular distributions are common in clinical or population data, as many respondents tend to report good functioning [40]. However, infrequently used response categories may also indicate unnecessary or redundant distinctions between categories, or suboptimal item wording or content [23, 40]. Moreover, sparse endorsement in extreme categories may limit the sensitivity of these items in detecting differences among individuals with lower levels of functioning. Third, DIF was observed across age groups, particularly for item related to sleepiness and physical function & wellbeing, which is plausible given that fatigue and mobility limitations generally increase with age. This DIF may reflect true age-related variation rather than measurement bias [22, 24]. However, it still suggests that the comparability of scores across age groups should be interpreted with caution. Lastly, because the metric was developed using nationally collected data, its application may be context-specific. Nevertheless, it remains conceptually aligned with the internationally recognized ICF framework, as it was constructed using items previously linked to ICF categories. Future studies are needed to evaluate its applicability across different population contexts.

This study developed a standardized, policy-relevant functioning metric for the Korean general population using national health data. By capturing lived health, the metric can generate a functioning indicator that complements the traditional health indicators of mortality and morbidity. The study highlights the value of repurposing routine survey data for monitoring population health. With growing demand for health assessment in ageing societies, this approach offers a sustainable strategy that strengthens international comparability and supports evidence-informed policy on population functioning.