This study aimed to determine the longitudinal outcomes of patients with multimorbidity using HRQoL as the main outcome and associated factors among a sample of individuals attending chronic outpatient NCDs care in Bahir Dar, northwest Ethiopia.

This study was conducted in five hospitals (three public and two private) and three private specialty clinics in Bahir Dar city. These facilities provide the bulk (~80%) of chronic NCDs care for the people living in the city and surrounding areas. Although chronic NCDs care and management is presumed to be provided in a relatively uniform fashion using the national NCDs treatment guideline (46), the nature of patients vising these facilities may vary and there remains a substantial difference in the quality and affordability of NCDs care between public and private health facilities in the country.

A two-stage stratified random sampling method was employed for recruiting facilities and participants. The sample size from each facility was determined based on the notion of probability proportional to size (PPS) using the pool of chronic NCD patients (≥ 40 yrs) registered for follow-up over the year preceding our assessment (January–December 2020) in each participating facility.

Only facilities who were providing chronic NCDs care by general practitioners or specialist physicians for at least a duration of 1 year prior to the data collection were considered. Older adults (40 years or more) diagnosed with at least one NCD and were on chronic diseases follow up care for at least 6 months prior to the study period were recruited for the study. Pregnant women and individuals who were too ill to be interviewed and admitted patients were excluded.

Participants enrolled for the baseline study (from March 15 to April 30, 2021) were invited 1 year later for the follow up study from March 15 to April 30, 2022. Contact information (mobile numbers and medical registration numbers) of patients involved in the baseline study were documented to contact them for the follow-up study. Printed copies of contact addresses of patients were given to the data collectors to sort out appointment dates of patients and to also remind patients to come for the study. All the participants agreed for the baseline study were informed about our plan to contact them 1 year from the baseline assessment. We used Kobo toolbox software to accurately match the end line data (period 2) with the baseline data (period 1) (47).

Data on QoL, patient activation (PA) score and multimorbidity were collected at two points on the same individual. However, some key outcome data such as mortality, hospitalization and perceived progress over time were collected only at the end of the follow up period.

HRQoL (stated as QoL in this study) is defined as individuals’ perception of their position in life in the context of physical, psychological and social functioning and well-being (48). QoL at two points (baseline and end line) was measured using the interviewer-administered short form (SF-12 V2) assessment tool (49, 50).

The SF-12 tool is extensively validated and widely used generic tool for measuring QoL in multimorbidity across different contexts, including Sub-Saharan Africa (51–53). The tool was translated and pilot tested according to the study protocol we published (45). The tool measures eight health aspects, namely physical functioning (PF), role limitations due to physical health problems (RP), bodily pain (BP), general health perceptions (GH), vitality (VT), social functioning (SF), role limitations due to emotional problems (RE), and mental health (psychological distress and psychological well-being) (MH). Two summary measures are derived from the SF-12: physical health (Physical Component Summary-PCS) and mental health (Mental Component Summary-MCS). However, owing to the possibility of correlation (lack of uni-dimensionality) between the PCS and MCS scores, some studies criticized the use of these scoring algorithms and recommended raw sum scores instead (54, 55). The use of a single raw sum score enables a consistent assessment of the impact of multimorbidity and how this varies across a given population (56). Thus, we applied this approach for analyzing the QoL data.

First, we reverse coded the scores for items 1, 9 and 10 and computed the raw total. The overall scores were scaled from 0 to 100, with 0 representing worst health (57). Although popularly used in previous studies, the notion of fitting linear regression models to summarize categorical data such as the QoL data has been questioned (55, 58). The linear regression models may potentially lose important variability in the data particularly when the QoL data is collected by Liker-type scales such as the SF-12 tool (59, 60). Recent advances in the field recommend the interpretation of QoL rather as a categorical (group continuous) variable than as a metric variable (58). Studies suggest that ordinal regression models (OLR) are superior to other method for analyzing ordinal data, including health-related QoL data (58, 61). Hence, we ranked the scaled QoL scores into three ordered and non-overlapping categories as per the recommendation (60) as poor QoL (a scaled value <75), moderate QoL (scaled value from 75 to 89.9) and high QoL (scaled value from 90 to 100) and fitted into the longitudinal OLR and proportional odds (PPO) models.

The data from the Kobo toolbox server were downloaded into an excel spreadsheet and migrated to SPSS V. 21 for cleaning before being exported to Stata V. 17 for analysis. The end line data were linked to the baseline data to form the panel data.

The authors did not do imputations to account for the missing data due to the addition of a 20% sample for the possible loss during follow up and non-response and because of the probability that the missing were at random.

We ran descriptive statistics to characterize distributions of the study participants, patient reported outcomes and associated factors. All descriptive analyzes were weighted to account for the stratified sampling. In our analysis, age and social support scores were treated as continuous variables.

Sensitivity analysis was conducted using chi squared test to check whether the data collected by face-to-face interview (81.2%) and telephone interview (18.8%) have statistical difference. No difference was observed between the two based on multimorbidity status (χ² = 7.2, p = 0.065) and QoL (χ² = 2.29, p = 0.130) measured at baseline.

Descriptive statistics were also used to characterize and compare the distributions of PA levels and QoL levels between the baseline and end line period. For the sake of clarity, we have also computed the means and SDs of both these response variables. We checked multicollinearity of independent variables while fitting multivariable models (VIF = 1.02).

Most QoL data are measured by Likert-type scales, and the scores were treated as if they are continuous (equal distance between levels) and normally distributed (61). However, evidence shows that such data possess skewed distributions and it is unknown whether the distances between two successive or alternative levels (categories) are equal (73). Hence, analyzing ordinal data as if they are metric (continuous) can systematically lead to biased effect-size estimates, inflated errors rates and inaccurate parameter estimates (55, 74). In addition, collapsing the categories to suit for binary regression is inappropriate for ordered outcomes such as the QoL (59, 60). Hence, more sensitive and comprehensive models are required. Evidence suggests that the ordinal regression models are superior to the methods commonly used to analyze data of an ordered nature (75, 76). The ordinal models provide better theoretical interpretation and numerical inference than the metric (linear) models for ordered outcomes (77, 78). Based on this, QoL was treated as an ordered outcomes and categorized as low, moderate and high, and coded as 0, 1 and 2, respectively, as described above. However, as our data did not satisfy the parallel lines regression assumptions, we fitted the partial proportional odds model. The model treats the data in two categories (as Panel 1 vs. 2 and 3 or panel 1 and 2 vs. 3).

Our data were measured at two points in time. Considering the correlation between outcomes measure at different times, we fitted an ordered logistic panel data analysis model. Panel data analysis is useful to control unobserved characteristics that do not change over time (time invariance variables) (79).

The data measured at two periods were reshaped from wide to long format. Then, the authors set Stata to handle the longitudinal panel data by using the xtset command (xtset facility type year). We had 2,246 person time data for this analysis.

Facility type was the panel variable and year (2021 to 2022) the time variable. We obtained the following output, signifying the data were strongly balanced (all individuals have data at two times).

To explore the relationship between predictor and outcome variables, we fitted both fixed and random effect models.

The model we fitted is described below.

Y_it = α + β_kX_kit + uit + εit where

i = individual and t = time (from March 2021–March 2022)

α is the intercept

Y_it is the dependent variable (either QoL)

X_kit represents the k^th independent and control variable

Β_k is the coefficient for respective independent and control variables

uit is the impact of the i^th individual (not a measured variable)

εit is the error

As we recruited a random sample of study participants, we performed both fixed effect and random effect regressions, then we compared them using the Hasuman test. The null hypothesis states that the error terms are not correlated. That means a significant test (p-value <0.05) in the Hausman test implies that the error terms are not correlated. Hence, the fixed effect model is preferred. If the test result is not significant, however, the random effect model is plausible (79). The Hausman test in our model indicates that the null hypothesis be rejected (p-value = 0.110). Hence, the random effect model is appropriate. We have further checked whether the random effect model is preferred to the simple OLS by running the Breusch-Pagan Lagrange Multiplier (LM) test. The null hypothesis is that variation across entities is zero (no panel effect). The LM test in our analysis showed a significant (p-value = 0.001) implying the random effect model is appropriate (80).

We used logistic regression analysis adjusting for age and sex to identify factors associated with development of new NCDs, multimorbidity and adverse outcomes such as hospitalization, poor progression and mortality.

Of the 1,432 participants who were enrolled in the baseline study, 1,123 (78.5%) patients agreed to participate for the follow up study. They ranged in age from 41 to 93 years (mean 57.1 ± 11.8 years, median 55 years), with a slightly higher percentage of women (50.9%) versus men (49.1%; Table 1). The primary reason for non-response were absenteeism on the date of follow-up and difficulty in tracing them via telephone calls (n = 19.06%), death of the participant (n = 22, 1.54%) and refusal (n = 9, 0.98%). There was no statistically significant difference between the attendees and the lost to follow up groups with regard to age (p-value = 0.504), gender (p-value = 0.400), multimorbidity status (p-value = 0.097) and the mean baseline QoL scores (p-value = 1.000).

During the follow up period, 44 (3.9%) patients reported to have one or more newly diagnosed NCDs, with, a higher proportion (n = 34, 72.3%) of them having multimorbidity at baseline. There was a statistically significant difference in the probability of developing new NCDs between having multimorbidity and single morbidity at baseline (p-value = 0.003). In addition, there was a 2% increase in the magnitude of multimorbidity during the follow up period. In other words, the magnitude of multimorbidity has increased from 54.8% at baseline to 56.8% as measured at the follow up period. It was also observed that about 17% of those having multimorbidity had three or more NCDs, a 2 % increase from the baseline status (15.2%).

Hypertension (19.6%), heat failure (19.6%) and chronic kidney diseases (15.2%) were the most frequently newly diagnosed NCDs during the follow up (Figure 1).

Having multimorbidity and being in the overweight or obese BMI category at baseline were the factors predicting development of new NCDs. However, only the presence of multimorbidity at baseline remained a statistically significant factor in the adjusted model. The odds of having a new NCD diagnosis was 2.5 times higher among patients that had multimorbidity at baseline compared to those who had single morbidity. However, participants’ age and sex were not associated with development of new NCDs during the one-year follow up (Table 2).

The majority (n = 659, 58.6%) of the participants reported that they had a good progress compared to their status at baseline. While about one third (n = 374, 33.3%) reported to have a fair progress and 91 patients (8.1%) had poor and deteriorating progresses.

Nearly 10% of the patients were admitted during the follow-up because of one or more of the NCDs they were living with. Presence of three or more chronic NCDs (AOR: 3.64, 95%CI = 2.15, 6.17) compared with having single NCD, and those who reported to have a deteriorating progress (AOR: 12.42, 95%CI = 6.97, 22.14) or fair progress (AOR: 2.82, 95%CI = 1.72, 4.60) since the baseline were more likely than their counterparts to have hospital admission (Table 3).

Of the total number of patients whose status was known (n = 1,160, 82%), 22 (2%) patients were reported to have died during the course of the one-year follow up.

The PA scores were summed and scaled into a 100-point scale to compute the mean PA value. The mean score was 76.0 ± 23.2. According to the standard classification (81), the majority (45.8%) were classified under the highest activation category (score ≥ 67.1), followed by those in level two 26.1% (score 47.1–55.1), 18.4% in level one (score < 47.0) and 9.7% in level three (score 55.2–67.0; Table 4). For the sake of simplifying the interpretation of its effect on QoL on the ordered logistic regression (proportional odds) model, we classified the PA score into two groups: high (≥55.2) and low (≤55.1). In this sense, the proportion of participants in the high and low category was 55.5 and 44.5%, respectively. We observed that activation level fluctuated over time. Based on the Pearson correlation test, the correlation between the baseline and end line PA score was 0.065, implying no strong correlation between the two measurements.

The majority (37.9%) of participants had a lower QoL and there was a modest increase from the baseline level (33.5%). In this study, about one-third (33.6) of the patients had better QoL and 28.5% of them had moderate QoL.

We assessed the effect of key independent variables through fitting an ordered longitudinal panel data analysis model. As indicated in the method section above the random effect model is appropriate for the data at hand and the output of the model is shown below (Table 4).

Compared to males, females had lower odds of being in the higher level of QoL versus the combined lower and moderate QoL and in the combined higher and moderate levels of QoL versus the lower level of QoL, given the other variables are held constant in the model [AOR1 = AOR2: 0.99, (95%CI: 0.82, 1.22)]. Similarly, a person living with multimorbidity had a lower odds of being in the higher level of QoL versus the combined lower and moderate QoL and in the combined higher and moderate levels of QoL versus the lower level of QoL, given the other variables are held constant in the model [AOR1 = AOR2: 0.89, (95%CI: 0.73, 1.09)]. However, none of these association are statistically significant. Further, age has no statistically significant association with QoL in our analysis.

On the other hand, PA score has shown a statistically significant association with QoL in both panels. For individuals having a higher levels of PA score, the odd of being in the higher category of QoL versus the combined moderate and lower QoL was 2.4 times higher [AOR1: 2.35, (95%CI: 1.93, 2.87)]. Likewise, the odds of being in the combined higher and moderate QoL versus lower QoL was 1.5 times higher for individuals having a higher level of PA score [AOR2: 1.53, (95%CI: 1.25, 1.88)] (Table 4).

The main goal of health care for the people living with chronic conditions and multimorbidity is to support them achieve a better QoL, improved wellbeing and survival (28, 96). However, a significant number of patients attending chronic care in the study area experienced a range of adverse outcomes, including development of new NCDs and multimorbidity, poor disease progression, poor QoL, hospitalization and mortality. This implies living with chronic NCDs and associated multimorbidity has profound impacts on individuals, and that the health system does not seem to be well prepared to adequately respond to individual patient needs. The provision of patient-centered care in which all healthcare providers work together with patients to ensure coordination, consistency and continuity of care over time is essential (97). This will in turn improve the wellbeing and survival of the people with multimorbidity in the study area.

Given the positive association between PA level and QoL, it is desirable to determine and devise strategies to increase activation status. The notion of patient activation is relatively new in the study context and there is a need to experiment its effects in improving outcomes of patients in the chronic care landscape and beyond.

Our study has the advantage of involving a broad group of health facilities and patients receiving chronic NCDs care. Guided by a published study protocol, this longitudinal study provided strong insight on the course and patterns of disease progression, the impact of multimorbidity on important patient outcomes such as hospitalization and mortality, and the level and predictors of QoL using robust methodologies. Our ability to determine activation levels of patients would encourage service providers to measure and intervene with mechanisms to increase PA. However, the findings of this facility-based study may not exactly represent the underlying epidemiology of multimorbidity and its impact in the general population. In addition, although the sensitivity analysis does not show variations, the data collected from a portion of patients by telephone interview might not be an exact replica of the data obtained from face-to-face interviews. However, adequacy of the sample size and parsimony of the methods we employed would make our study robust.