Determinants of Appropriate Malaria Treatment-Seeking in Rural East Nusa Tenggara, Indonesia: A Multilevel Logistic Regression Analysis

2.1 Data source

The analysis utilized secondary data obtained from a previously published community-based intervention aiming to enhance malaria knowledge among rural residents in NTT, Indonesia [33]. In that study, data were collected from four sub-districts representing a high malaria-endemic setting in the region. All 25 villages within these areas participated, with 30-40 household heads randomly sampled in proportion to village size, resulting in a total of 894 respondents. In each selected village, a systematic random sampling technique was employed. The sampling frame was constructed using the official village household registry (Data Penduduk / Kartu Keluarga) provided by the Village Head. From this list, 30 to 40 households were selected per village to ensure sufficient statistical power for the multilevel analysis while adhering to logistical constraints standard in cluster surveys. Data was collected by interviewing participants who were at least 18 years old, using a validated questionnaire. The study followed the principles of the Declaration of Helsinki, and ethical clearance was granted by the Human Ethics Committee of Nusa Cendana University (Approval No. 42/UN15.21/KEPK/2024).

2.2 Research variables

Three binary outcome variables were assessed:

(1) awareness of seeking malaria treatment within 24 hours of symptom onset $\left(y_1\right)$, (2) awareness of obtaining care at professional health facilities $\left(y_2\right)$, and (3) comprehensive understanding of reported appropriate malaria treatmentseeking $\left(y_3\right)$, defined as seeking treatment within 24 hours and at a professional facility. Each of these variables was a categorical variable with one for yes and zero for no. Moreover, nine explanatory variables consisting of six variables at the individual level and three variables at the community level. At individual level, it contained gender $\left(X_1\right)$, age category $\left(X_2\right)$, level of education $\left(X_3\right)$, income group $\left(X_4\right)$, awareness for preventing malaria $\left(X_5\right)$, hearing malaria term $\left(X_6\right)$. At community level, it consisted of health centre availability $\left(X_7\right)$, geographical condition $\left(X_8\right)$, access to village $\left(X_9\right). X_1$ was grouped as female and male; $X_2$ was categorized as $<60$ and $\geq 60. \, X_3$ had three categories, including no education at all, primary school, and at least junior high school level. For $X_4$, it was classified as greater than minimum wage province (MWP) and less than MWP. Whilst each variable of $X_5, X_6$, and $X_7$ had two categories, no and yes. To ensure sufficient statistical power and model convergence, the geographic topography variable $\left(X_8\right)$, originally categorized as coastal, hilly, and rice fields, was recoded into a binary variable: Highland (comprising hilly and mountainous terrain) and Lowland (comprising coastal and flat plains). This binary classification better captures the distinct topographical barriers present in the East Nusa Tenggara archipelago. Then, for $X_9$, it had easy access and difficult access between and within the village. Prior study shown that being a female [19], young age, and no education [34], living in lowland and far from health facilities decreased the probability to have good understanding of malaria treatment-seeking behavior [20, 23, 31, 34]. Therefore, on the univariate and multivariable analysis, these categories became a reference category.

2.3 Modelling stage

The analysis began with descriptive statistics to determine the prevalence of each outcome variable. To verify independence among predictors, a multicollinearity test was conducted through tolerance (TOL) and variance inflation factor (VIF) values. These values were defined as follows [35]:

$T O L_i=1-R_{i,}^2$ for $i=1,2, \ldots, 9$           (1)

where, $R_i^2$ denoted determination coefficient accomplished by regressing $X_i$ on all the other independent variables $X_1, X_2, \ldots, X_{12}\left(\right.$excluding $\left.X_i\right)$ and

$\begin{gathered}V I F_i=\frac{1}{\text { Tolerance}_i}=\frac{1}{1-R_i^2}  \text {for} i=1,2, \ldots, 9\end{gathered}$          (2)

When TOL value $\leq 0.1$ and VIFs $>10$, it was a strong indication of multicollinearity among independent variables [35]. Furthermore, analysis was continued by exploring the relationship between each dependent variable $\left(y_1, y_2\right.$, and $\left.y_3\right)$ and all predictors with chi-square $\left(\chi^2\right)$ test following the hypothesis:

H_o: $X_i$ was not associated with $y_1$, versus H₁: $X_i$ was associated with $y_1$;

H_o: $X_i$ was not associated with $y_2$, versus H₁: $X_i$ was associated with $y_2$;

H_o: $X_i$ was not associated with $y_3$, versus H₁: $X_i$ was associated with $y_3$, for i = 1, 2, ..., 9.

The rejection of the null hypothesis could be done when $\chi_{\text {count}}^2>\chi_{\text {table}}^2$ or $\mathrm{p} \leq 0.05$. In this case, $\chi_{\text {count }}^2$ was computed by

$\chi_{\text {count }}^2=\sum_{i=1}^r \sum_{i=1}^c \frac{\left(O_{i j}-E_{i j}\right)^2}{E_{i j}}$           (3)

where, r represented the number of rows and c denoted the quantity of columns in a table contingency during tabulation between the dependent variable and each independent variable. The computation of Eq. (3) was conducted separately for $y_1$, $y_2$ and $y_3$.

2.4 Modeling with multilevel logistic regression for $y_1$

Variable selection for the multivariate model followed a purposeful selection algorithm [36]. Initially, bivariate analysis was performed to screen for potential predictors, using a liberal threshold of p < 0.25 to capture all potentially relevant variables. However, the final inclusion of variables in the multilevel logistic regression was not based solely on statistical significance. Variables of known theoretical importance—such as gender, age group, education level, and income group—were retained in the final model to control for confounding and ensure content validity, regardless of their individual statistical significance in the bivariate stage. This approach ensures that the model reflects the complex, hierarchical nature of treatment-seeking behavior rather than relying on automated stepwise elimination.

All independent variables showing a significant relationship with $y_1$ theoretically and statistically as shown in the chi-square test of bivariate analysis, were incorporated in the multilevel analysis for $y_1$. A two-level multilevel logistic regression model was employed because the outcome variable, for $y_1$ was binary (one for yes and zero for no) and the data had a hierarchical structure, with individuals nested within villages. This model accounts for within-village correlation and captures heterogeneity across villages. The analysis included 894 individuals (level 1) from 25 villages (level 2). The predictors consisted of six individual-level variables and three village-level variables. Therefore, the multilevel logistic regression model with a village-level random intercept was specified as:

$\ln \left(\frac{p_{i j}}{1-p_{i j}}\right)=\beta_0+\sum_{k=1}^6 \beta_k X_{k i j}+\sum_{l=1}^3 \gamma_l Z_{l j}+u_{0 j}$            (4)

where, $p_{i j}$ was the probability that individual $i$ in village $j$ sought malaria treatment, $X_{k i j}$ and $Z_{l j}$ denoted individual and village-level predictors, respectively; $\beta_k$ and $\gamma_l$ was fixedeffect coefficients, and $u_{0 j}$ was the village-specific random effect, assumed to follow a normal distribution with mean zero and variance $\sigma_u^2$. Model estimation was performed using the generalized linear mixed modeling framework with a logit link function and binomial error distribution. Parameters were estimated via maximum likelihood using the Laplace approximation as implemented in the GENLINMIXED procedure in Statistical Package for the Social Sciences (IBM Corp.). This model estimation proceeded in three steps. Firstly, an intercept-only model, yielding Model I, was fitted to assess between-village variation and to compute the ICC. The ICC was determined as follows [37, 38]:

$I C C=\frac{\sigma_u^2}{\sigma_u^2+3.29}$         (5)

Secondly, individual-level predictors, yielding Model II, were added to evaluate their associations with treatment-seeking behavior. Finally, individual and village-level predictors, Model III, were incorporated to assess contextual effects after adjustment for individual characteristics. For statistical inference, the results were reported as odds ratios (ORs) with 95% confidence interval (CI) to quantify the significant impact of each predictor on $y_1$ with formula:

$O R=\exp \left(\hat{\beta}_k\right)$, for $k=1,2, \ldots, 9$            (6)

And the computation of 95% interval estimate of OR was determined by

$\operatorname{Exp}\left[\hat{\beta}_k \pm 1.96 * S E\left(\hat{\beta}_k\right)\right]$, for $k=1,2, \ldots, 9$          (7)

The variance of the random intercept, the proportional change in variance (PCV), median odds ratio (MOR), and ICC were used to quantify between-village heterogeneity [26, 28]. Statistical significance was defined as p < 0.05. Model performance was evaluated using the Receiver Operating Characteristic (ROC) curve and the Area Under the Curve (AUC). To ensure model parsimony, two-way interactions between key individual and community-level variables (e.g., Education x Village Access) were explored but were excluded from the final model as they did not reach statistical significance (p > 0.05) or improve the Akaike Information Criterion (AIC). Finally, modeling with multilevel logistic regression, for  was conducted with the same procedure as shown for $y_1$.

To our knowledge, this is the first study to model the determinants of self-reported appropriate Malaria-treatment seeking in the rural, high-transmission setting of NTT using a two-level multilevel logistic regression framework. This study applied Andersen’s Behavioral Model to dissect the hierarchical determinants of reported appropriate malaria treatment-seeking behavior in rural NTT. The findings underscore a critical interplay between individual predisposing factors and community enabling factors. While high school education—a key predisposing factor—significantly increased the likelihood of appropriate care (AOR = 1.84), the analysis revealed that enabling factors exerted a far greater influence. The strong association with health centre availability (AOR = 7.24) and easy village access (AOR = 4.37) suggests that in this high-endemic setting, the primary barrier to elimination is not merely a lack of intent (predisposing), but a deficit in structural opportunity (enabling).

This study demonstrated that educational attainment emerged as a robust predictor, with individuals who completed at least high school exhibiting significantly higher odds of seeking appropriate treatment (AOR = 1.84). This finding aligns with the health belief model, which posits that the decision to seek treatment is influenced by an individual’s perception of illness severity and susceptibility [39]. Higher education likely serves as a proxy for health literacy, enhancing an individual's ability to recognize malaria symptoms early and navigate the healthcare system effectively [22, 37, 40-42]. Educated individuals are better equipped to understand the risks of complications if malaria is untreated, prompting quicker medical visits. This highlights the need for targeted interventions. In communities with low formal education, as shown in rural NTT [43], reliance on written materials is insufficient. Strategies must utilize community-based education, local media, and Community Health Workers to bridge the literacy gap and reduce reliance on self-medication [44, 45]. This study highlights the importance of an intervention strategy to reach the rural community with a low education level. Community-based education, the use of local media, and the involvement of health workers could assist in tackling limited health literacy and intensifying access to accurate information. Therefore, improving appropriate malaria treatment-seeking behavior depends not only on service facilities but also on an inclusive and sustainable approach to health education.

A striking finding of this study is that structural factors at the community level demonstrated substantially stronger effects than individual characteristics. The availability of a health centre within a village was the strongest predictor of appropriate treatment-seeking behavior (AOR = 7.24). This magnitude suggests that in rural NTT, service utilization is driven by structural opportunity. This supports the "Distance Decay" hypothesis, where utilization declines exponentially as the effort required to reach a facility increases [40, 46]. Our finding aligns with studies from other archipelago nations and Sub-Saharan Africa, which consistently show that physical proximity is often the "rate-limiting step" for malaria treatment [20, 27, 47, 48]. Even with high health literacy (individual factor), patients are unlikely to seek appropriate care if the supply-side logistics (facility availability) are absent. This highlights that health education campaigns must be matched with infrastructure investment to be effective.

In the context of rural NTT, where the terrain is often rugged and public transport is unreliable, as well as relying on walking to reach health facilities [49], the presence of a facility within the village effectively removes the "transaction costs" of seeking care. When a Puskesmas or Pustu (subsidiary health centre) is locally available, the direct costs (transport fares) and indirect opportunity costs (lost wages, time away from agriculture) are minimized. Consequently, the threshold for seeking care is lowered; individuals do not wait for symptoms to become severe before visiting a doctor. Conversely, in villages without a facility, the physical and financial burden of travel likely acts as a deterrent, leading residents to rely on self-medication or traditional healers as a "first line" response to their malaria, as shown in the literature [50].

The distinction between Highland and Lowland communities in NTT extends beyond topography to cultural and behavioral domains [51]. Highland communities often maintain stronger adherence to traditional customs (adat), where illness is frequently interpreted through spiritual lenses requiring ancestral appeasement before medical intervention [52]. Furthermore, the agricultural practice of residing in semi-permanent 'garden houses' (Rumah Kebun) during planting and harvest seasons is more prevalent in Highland areas [53]. This practice physically isolates families from village centers where health information is disseminated, creating a 'double burden' of geographic remoteness and cultural insulation that delays appropriate treatment-seeking.

It is important to acknowledge the wide confidence interval associated with this finding (95% CI: 1.14-45.99). This width is likely a statistical artifact resulting from the sample size at the second level of analysis (25 villages) [54]. While the small number of clusters reduces the precision of the point estimate, the lower bound of the interval (1.14) remains above 1.0, confirming that the positive association is statistically significant and robust. This suggests that while we cannot pinpoint the exact magnitude of the benefit (whether it is 7-fold or merely 2-fold), we can confidently conclude that the presence of local infrastructure is a critical determinant of service uptake.

Furthermore, the multilevel analysis showed that residents in villages with easy access were over four times more likely to seek appropriate treatment (AOR = 4.37) highlights the critical role of transport infrastructure in malaria control. This is particularly relevant in the context of NTT, where the topography is dominated by mountainous terrain and scattered islands. Many remote villages in NTT are connected only by poorly maintained or unpaved roads that are difficult for motor vehicles to navigate, especially during the rainy season. As a result, residents often walk long distances to reach the nearest passable road and rely on limited and relatively expensive private informal transport services (e.g., motorcycle taxis) to access health facilities [55, 56]. Furthermore, this infrastructure deficit creates a "seasonal trap." Malaria transmission in NTT typically peaks during the rainy season; however, this is precisely when unpaved village roads are most likely to become impassable due to mud or flooding. This suggests that the barrier is not merely the static distance to a clinic, but the dynamic difficulty of travel, which worsens exactly when the need for treatment is highest. These physical barriers likely contribute to the "Second Delay" (delay in reaching the health facility), discouraging patients from seeking formal care until symptoms become severe.

From a methodological standpoint, the findings validate the use of multilevel modeling to disentangle individual decision-making processes from structural village-level influences. Ignoring this hierarchical structure would likely have led to biased effect estimates and an underestimation of contextual impacts. In practical terms, the results imply that interventions aimed at improving malaria treatment outcomes should prioritize both educational advancement and community-level infrastructure development. While education enhances individual capacity for informed health decisions, the presence and accessibility of health facilities fundamentally condition whether such decisions can be acted upon.

Finally, it was important to acknowledge that the outcome measure in this study was based on self-reported behavior obtained through survey interviews rather than directly observed clinical actions or medical records. While self-reports were a standard method in demographic health surveys, they were subject to recall bias (participants might not remember details accurately) and social desirability bias (participants may report 'appropriate' behavior to please the interviewer). Therefore, the findings should be interpreted as determinants of the reported intent or recalled practice of treatment-seeking, which serves as a proxy for actual health utilization behavior. Then, the use of official village registries as the sampling frame ensured the representativeness of the settled population but may have excluded unregistered individuals or those residing temporarily in farming locations (kebun) away from the main village settlement. Consequently, the findings primarily reflect the treatment-seeking behavior of the permanent, registered population. Moreover, while the study focused on a high-endemic rural setting in NTT, the results might not be fully generalizable to urban settings or regions with different topographical challenges. However, these findings were likely highly relevant to other remote, high-burden regions in Eastern Indonesia. Future research could incorporate longitudinal designs or spatial modeling to capture these dynamic geographic effects more explicitly.