Using decision tree analysis to identify population groups at risk of subjective unmet need for assistance with activities of daily living

The aim of this work was to use conditional inference trees to consider complex interactions between a large set of potential predictors for subjective unmet need for assistance with ADL and subjective unmet need for assistance with IADL. Potential predictors were a selection of socio-demographic factors and impairments in body function or structure.

We used the scientific use file of the cross-sectional study German Health Update (GEDA 2019/2020-EHIS) which has been carried out by the Robert Koch-Institute in 2019 and 2020 [24]. The study was conducted among the population aged 15 or older with a usual residence in a German private single or multi-person household. Care and residential homes were not included in the survey. The survey was based on a telephone sample that has been drawn using the dual-frame method, which considers landline and mobile phone numbers. After obtaining informed consent, Computer Assisted Telephone Interviews (CATI) were applied to collect the data. No proxy interviews have been conducted among people with cognitive or sensory impairment. The “Response Rate 3” of the American Association for Public Opinion Research (AAPOR) classification was 21.6%. Finally, survey weights were calculated by the principal investigators of the study using external information on federal state, residential structure, age, sex, and education to reduce a possible non-response bias. These survey weights were also provided to the authors in the scientific use file. More details about the survey methodology and data quality assurance procedures can be obtained elsewhere [25]. Overall, 23,001 participants were interviewed in GEDA 2019/2020-EHIS. We included the population aged 55 or older (N = 12,985, 56.5% of the original sample), because items on ADL and IADL were only applied in this age group. GEDA 2019/2020-EHIS included an ADL score according to Katz et al. and an IADL score according to Lawton and Brody [4, 5]. The entire GEDA 2019/2020-EHIS questionnaire can be found elsewhere [26]. To assess limitations in ADL, five items asking for limitations in feeding, transferring, dressing, using toilets, and bathing were used. Response options to each item were “no difficulties”, “some difficulties”, “severe difficulties”, and “it´s not possible for me”. In the IADL score, seven items asked for limitations in food preparation, using the telephone, shopping, taking medication, light housework, heavy housework, and taking care of finances. For IADL items, response options were “no difficulty”, “some difficulty”, “severe difficulty”, “it is not possible for me”, and “have never tried or done it”. The latter response option was considered missing in the IADL items.

In the analysis of subjective unmet need for assistance with ADL, we only included participants who reported “some difficulty”, “severe difficulty”, or “it is not possible for me” for at least one activity of the ADL score. We excluded all participants who only responded with”no difficulty” in all available ADL items as well as participants who had missing observations in all available ADL items (final N = 1,957, 8.5% of the original sample). In the analysis of subjective unmet need for assistance with IADL, we applied similar inclusion and exclusion criteria. Participants indicating “some difficulty”, “severe difficulty”, or “it is not possible for me” for at least one activity of the IADL score were included. Participants responding only”no difficulty” in available IADL items or having missing values in all available IADL items were excluded (final N = 3,801, 16.5% of the original sample).

The two outcomes of interest in our study were subjective unmet need for assistance with ADL and subjective unmet need for assistance with IADL. In the questionnaire of the GEDA 2019/2020-EHIS study, questions on subjective unmet need for assistance were asked directly after the items of the ADL and IADL scores, respectively. Concerning the first outcome, participants were asked if they had any help with ADL tasks in general (“Do you usually have help with these activities? Now think about activities related to personal hygiene and other personal needs that you have difficulty performing without help.”). Response options were “Yes, for at least one activity” and “No”. After this question, all participants who stated that they had no help were asked if they needed any help (“Do you need help?”). All participants who answered that they had help were asked if they needed more help (“Do you need more help?”). Both questions could be answered with “yes” and “no”. In our study, all individuals who stated that they needed any help or that they needed more help were classified as having subjective unmet need for assistance with ADL. Subjective unmet need for assistance with IADL was assessed in the same manner. Only the phrasing of the initial question for subjective unmet need for assistance with IADL tasks differed from the ADL section (“Do you usually have help with these activities? Now think of all the activities in the household that you have difficulty doing without help.”).

We included 14 potential sociodemographic predictors of subjective unmet need for assistance: gender identity (female, male, other) [27], 5-year age group, migrant status (two-sided migration background, no or one-sided migration background), rurality of the district of residence (large city, urban district, rural district, sparsely populated rural district), educational level according to the International Standard Classification of Education (ISCED, levels 1–8), household income (quintiles), employment status (employed, unemployed, retired, unable to work due to long-term illness, doing housework, not working for other reasons), main earner in the household (myself, there is no main breadwinner, my partner), marital status (single, married, widowed, divorced), living with partner (yes, no), number of people in the household (1, 2, 3, 4, 5, 6, 7 or more), health insurance type (statutory health insurance, private insurance, other), Oslo-3 scale for social support (low, medium, high) [28], time spent caring for others (none, less than 10 h per week, at least 10 but less than 20 h per week, 20 h per week or more).

Twenty two variables were used to mirror a selection of impairments in body function or structure. We used single questions with the response options yes and no to measure the following self-reported diseases that were prevalent in the last 12 months: hypertension, myocardial infarction, coronary artery disease, stroke, diabetes, asthma, hypolipoproteinaemia, chronic bronchitis, liver cirrhosis, chronic kidney disease, urinary incontinence, any allergy, arthrosis, complaints of neck or cervical spine, complaints of lower back, injury due to traffic accident, injury due to home accident, injury due to accident in free time. Furthermore, the Patient Health Questionnaire 8 (PHQ-8) [29] score for depressive symptoms in the past 2 weeks was included alongside the severity of pain in the last 4 weeks (no pain, very mild, mild, moderate, severe, very severe). The total score of the PHQ-8 is determined by summing up four single items (each on a scale 0–3). The total PHQ-8 score ranges from 0 to 12 and represents a metric on the ordinal scale [29]. Moreover, visual difficulty (none without visual aid, none with visual aid, difficulties with visual aid, difficulties without visual aid) and hearing difficulty (none, moderate, severe) were considered.

We used conditional inference trees to select predictors of subjective unmet need for assistance with ADL and IADL tasks, respectively. Conditional inference trees use a procedure based on a regression coefficient and a hypothesis test of a global null hypothesis. The global null hypothesis is the hypothesis that none of the covariates has a univariate association with the outcome. If the algorithm selects a first split of the data, it assesses subsequent splits in the resulting two subsets of the data. This process represents a decision algorithm, which produces a sequence of hierarchical binary decisions that can be graphically displayed as a tree. To deal with multiple testing, an overall Bonferroni-corrected type on error rate (alpha) is chosen in conditional inference trees and serves as stopping criterion in the tree-building process. Generated subsets of the data are called nodes and final subsets represent terminal nodes. Each observation can be allocated to a single terminal node [23, 30].

Missing observations in the two outcomes were not included in our analyses. The first outcome, subjective unmet need for assistance with ADL, had 8 (0.4%) missing observations while the variable subjective unmet need for assistance with IADL had 22 (0.6%) missings. Our selection of 14 socio-demographic variables and 22 variables indicating impairments of body function were used together as independent variables in the analyses of both outcomes. Missing observations in independent variables were not excluded. It is possible to maintain missing values of independent variables in conditional inference tree analyses since the algorithm uses so-called surrogate splits to deal with missing observations [23]. Survey weights that were provided in the scientific use file were applied in both decision trees to correct for non-response. The minimum node size in a terminal node was restricted to 1% of the weight of the overall population. Trees were grown using the package partykit (version 1.2–15) in R (version 4.0.2) [31].

Finally, the area under the receiver operating characteristics curve (AUC) was calculated for each tree to assess discriminatory accuracy. The AUC combines the true positive and the false positive rate in a single metric for discriminatory accuracy of a binary outcome. The metric has a range from 0.5 to 1 where 1 represents perfect discrimination [32]. In our case, the AUC shows the ability of a decision tree to predict unmet need for assistance. Finally, trees were very large when using alpha = 5% which impacted interpretability of our results. Hence, we used cross-validation (CV), to evaluate if there are values for alpha that lead to smaller trees with no or little loss of discriminatory accuracy as measured by the AUC [33]. Tenfold CV showed that alpha = 0.5% did not result in a substantive loss of AUC in the analysis of subjective unmet need for assistance with ADL. For the analysis of subjective unmet need for assistance with IADL, alpha = 0.1% was chosen based on CV. The impact of this choice was shown by calculating AUC for all trees with alpha = 5% and for the lower levels of alpha that were chosen based on CV.

Among participants who reported difficulty with at least one ADL (N = 1,957), 22.6% reported a need for support with ADL, 41.1% reported a male gender identity and 0.4% reported an identity that was not female or male (Table 1). 73.8% were over 65 years. Furthermore, 6.9% had a two-sided migration background, 11.9% lived in a sparsely populated rural district, and 23.4% had low social support according to the Oslo-3 scale. Moreover, 6.8% experienced a stroke in the past 12 months. Among participants who reported a difficulty in IADL (N = 3,801), 22.8% reported a need for support with IADL, 35.2% reported a male gender identity, and 0.4% reported an identity that was not female or male. 72.2% were over 65 years of age, 7.0% had a two-sided migration background, 11.0% lived in a sparsely populated rural district, and 20.1% reported low social support. Finally, 5.3% had a stroke in the past 12 months. All included independent variables are summarised in the additional file (Additional file 1).

In the analysis of predictors for subjective unmet need for assistance with ADL, the first split occurred for depressive symptoms (PHQ-8 higher seven) (Fig. 1). Among participants with a PHQ-8 lower or equal to seven, the proportion of subjective unmet need for assistance was 18.0%, while it was 34.1% for participants who showed a PHQ-8 higher seven. A subsequent split occurred according to gender identity in the subset of participants with a PHQ-8 lower or equal to seven. The proportion of subjective unmet need for assistance was 19.9% among people with female or other gender identity compared to 15.6% among people identifying as male in this subset. In the subset with male gender identity, the next split was induced for severity of pain, with a proportion of subjective unmet need for assistance of 23.9% among the subset with a severe or very severe intensity of pain. On the other hand, in the subset with moderate or less severe intensity of pain, 13.0% reported an unmet need for assistance. In this node, a final split occurred for employment status. Turning back to the node of people with a PHQ-8 lower or equal to seven and a female or other gender identity, the next split was induced for marital status. Single people had a higher proportion of subjective unmet need for assistance (23.2%) compared to people with another marital status (19.6%) in this subset. In this node with a non-single marital status, the final split was induced for employment status. In the node with a marital status “single”, the final split occurred for level of education. Finally, just one split occurred on the other half of the tree. The node with a PHQ-8 higher than seven was split according to employment status. Subjective unmet need for assistance was lower in the subset who were employed or unemployed (13.7%) compared to the subset in retirement or in another form of employment (38.2%). The AUC of the decision tree was 0.66 when using alpha = 0.5% and 0.66 when using alpha = 5%.

Among participants reporting difficulties in IADL, depressive symptoms also induced the first split (PHQ-8 higher six) (Fig. 2). The proportion of subjective unmet need for assistance was 22.8% among participants with a PHQ-8 score lower or equal to six compared to 41.3% among participants with a score higher than six. The former node (PHQ-8 score lower or equal to six) was split further according to severity of pain. The subset with a moderate or less severe intensity of pain showed a lower proportion of subjective unmet need for assistance (20.5%) compared to the subset with severe or very severe intensity (33.3%). In the latter node, the next split was induced for prevalent self-reported asthma. In the subset of participants reporting an asthma diagnosis in the past 12 months, 50.7% reported having an unmet need for assistance. In the subset with no self-reported asthma diagnosis, 29.3% had a subjective unmet need for assistance. Finally, a last split occurred in the subgroup with no asthma diagnosis according to gender identity with a higher proportion of subjective unmet need for assistance among people with female identity (29.4%) compared to people with male or other identity (29.2%). On the other half of the tree (PHQ-8 score higher than six), the subsequent split after depressive symptoms occurred for social support. In the node with high social support according to Oslo-3, the proportion of subjective unmet need for assistance was 33.8%. In contrast, 43.3% reported a subjective unmet need for assistance in the node with medium or less social support. A final split occurred in this subset. The terminal node with severe or less intense pain showed a lower proportion of subjective unmet need for assistance (40.1%) compared to the terminal node with very severe pain (65.4%). The AUC in the final tree for the prediction of subjective unmet need for assistance in IADL was 0.63 when using alpha = 0.1% and 0.64 when using alpha = 5%.