Kunming, the capital of Yunnan Province, lies between 102°10′–103°40′E and 24°23′–26°22′N. Administratively, Kunming comprises seven municipal districts, three counties, one county-level city, and three autonomous counties; its resident population was 8.46 million in the 2020 census (12). In 2020, Kunming reported 1,453 new HIV cases and 54 AIDS-related deaths. During data preprocessing, we excluded decedents and non-mainland Chinese cases (foreign nationals and residents of Hong Kong, Macao, and Taiwan). The analytical dataset included laboratory-confirmed cases with validated case report forms. Figure 1 shows the spatial distribution of HIV/AIDS case counts across district- and county-level administrative units in Kunming in 2020.

In regression analyses, substantial intercorrelations among predictors can induce multicollinearity. Accordingly, prior to model estimation, we applied principal component analysis (PCA) to mitigate this issue and subsequently incorporated the resulting orthogonal components into geographically weighted regression (GWR) to quantify their spatially varying effects.

Principal component analysis (PCA) reduces a set of variables to a small number of composite factors—principal components—that retain as much information as possible from the original variables (11). The PCA framework is mathematically defined as follows: Consider n spatial units (street-level divisions), each characterized by p AIDS prevalence influence factors ( X 1 , X 2 , ⋯ X p ) , which collectively form an n × p dimensional matrix. As shown in Equation 1.

where ( X 1 , X 2 , ⋯ X P ) are original HIV/AIDS prevalence indicators. As the different dimensions of the raw HIV/AIDS epidemic factors, when performing principal component extraction, the original factors are first standardized. The correlation coefficient matrix was computed for the standardized variables, from which the eigenvalues ( λ 1 , λ 2 , ⋯ λ p ) and corresponding eigenvectors were derived. These eigenvalues were subsequently sorted in descending order ( λ 1 > λ 2 > ⋯ λ p ) to determine the principal components. The eigenvector F 1 corresponding to the largest eigenvalue λ 1 constitutes the first principal component, while F 2 associated with λ 2 represents the second principal component, with subsequent components following analogously in descending order of their explained variance. F 1 , F 2 , ⋯ F P are new variable indicators. As shown in Equation 2.

The formula satisfies, a 1 i 2 + a 2 i 2 ⋯ + a ip 2 = 1 ; F i , F j ( i ≠ j , i , j = 1 , 2 , ⋯ , p ) , are uncorrelated and the variance of F 1 , F 2 , ⋯ , F p is gradually decreasing, i.e., Var ( F 1 ) > Var ( F 2 ) > ⋯ > Var ( F p ) , the greater the variance of the principal components contains more information about the original variables.

The optimal number of principal components is determined by cumulative explained variance, with the common practice of retaining components that together account for 80–90% of the total variance in the original data (13). The formulas for the contribution rate and the cumulative contribution rate are provided in Equations 3 and 4.

where λ is the eigenvalue of each factor and k is the first k principal components.

Using PCA, the 19 selected factors were synthesized into new composite indicators, which were then incorporated into a geographically weighted regression (GWR) model to analyze determinants of HIV/AIDS prevalence in Kunming. This approach mitigates multicollinearity arising from numerous predictors while simplifying the set of influencing factors.

The spread of HIV/AIDS is closely associated with economic conditions, population mobility, and healthcare provision. Differences in regional socioeconomic development and natural environments lead to spatial heterogeneity in the determinants of HIV/AIDS prevalence in Kunming. Geographically weighted regression (GWR), proposed by Fotheringham, is an extension of ordinary least squares (OLS). Based on the principle of local regression, it embeds the spatial locations of samples into the regression equation to enable location-specific parameter estimation and to quantify the effects of individual factors at different geographic locations (14). The expression is given in Equation 5.

Where x i denote the HIV/AIDS prevalence rate in Kunming’s ith street; ( u i , v i ) is the geographical coordinate of the ith street; β 0 ( u i , v i ) is the regression constant of the i th street; k denotes the k th principal component extracted; β m ( u i , v i ) as the regression coefficient for the m th principal component at the i th street; F mi as the value of the m th principal component of the i th street; ε i as the spatially distributed random error term of the i th street.

In the GWR model, regression coefficients vary with the geographic location of each observation, and the degree of influence is expressed by a distance-based kernel function (15). Common kernels include the distance-threshold, inverse-distance, Gaussian, and bisquare functions, among which the Gaussian kernel is widely used for its general applicability (8). The mathematical form of the Gaussian kernel is given in Equation 6.

where w ij is the weight of the data point; d ij is the distance between data point i and regression j ; and b is the bandwidth. In GWR, two key choices are the kernel function and the optimal bandwidth. An excessively large bandwidth inflates regression parameter estimates, whereas an excessively small bandwidth deflates them. To avoid errors due to inappropriate bandwidths, we determine the optimal bandwidth using the Akaike Information Criterion (AIC), selecting the bandwidth that yields the minimum AIC value. Accordingly, this study employs a Gaussian kernel with AIC-based bandwidth selection and incorporates the newly extracted principal components into the GWR model to analyze the determinants of the HIV/AIDS epidemic in Kunming.

Compared with traditional OLS, GWR reveals spatial variation in relationships between variables across regions. Given the substantial disparities in economic, educational, and healthcare conditions within Kunming, a single global regression cannot accurately capture how these factors vary in influence across areas. Accordingly, we first applied principal component analysis (PCA) to reduce the dimensionality of 19 socioeconomic and healthcare variables and to extract independent composite factors, thereby minimizing multicollinearity. We then incorporated these components into a geographically weighted regression (GWR) to conduct spatially weighted analysis of HIV/AIDS prevalence and identify key determinants across regions. This combined PCA–GWR framework more effectively characterizes spatial heterogeneity and provides a scientific basis for region-specific HIV/AIDS prevention and control.

To evaluate the suitability of the selected determinants of HIV/AIDS prevalence for factor analysis, we conducted Kaiser–Meyer–Olkin (KMO) and Bartlett’s test of sphericity in SPSS. The results are reported in Table 2.

As presented in Table 2, the KMO measure for the selected influencing factors is 0.804, suggesting adequate similarity in the strength of correlations among the variables and supporting the suitability of the data for factor analysis. Furthermore, Bartlett’s test of sphericity shows a significance level of 0.000, indicating that the null hypothesis of sphericity is rejected and significant correlations exist among the variables, thus verifying the appropriateness of applying factor analysis.

The principal components selected by PCA capture the key information contained in the influencing factors; however, variables that are strongly correlated with prevalence but account for only a small proportion within the principal components may be overlooked (13). To address this issue, we first conducted correlation analyses between each factor and disease prevalence before performing PCA, and only factors showing strong correlations with prevalence were retained for PCA. Pearson correlation coefficients were used to characterize the strength of association between influencing factors and prevalence, with the results reported in Table 3.

As shown in Table 3, no statistically significant associations (p > 0.05) were observed between street-level HIV/AIDS prevalence in Kunming and (a) the annual registered unemployment rate, (b) the completion rate of nine-year compulsory education, (c) PITC spatial accessibility, or (d) population density. To ensure the robustness of the analysis, these four factors were excluded, and the remaining 15 factors that exhibited significant correlations with prevalence were retained for PCA. The explained variance of the retained components is reported in Table 4.

As shown in Table 4, the first, second, and third principal components account for 61.829, 16.958, and 7.928% of the total variance, respectively, yielding a cumulative variance explained of 86.715%. Thus, the first three components capture 86.715% of the information in the original variables. Accordingly, we extracted these three components, transforming the 15 correlated HIV/AIDS prevalence variables into three mutually uncorrelated (orthogonal) components.

As summarized in Table 5, the characteristic variables corresponding to the three principal components are as follows. The first principal component (PC1) is dominated by four education-related variables—population with tertiary education (college degree or above), population with upper secondary education (including vocational school), illiteracy rate, and population with lower secondary education—with factor loadings all exceeding 0.90; we therefore label PC1 as education level. The second principal component (PC2) is characterized by annual GDP and road network density, both with loadings greater than 0.80; because the remaining variables exhibit substantially lower loadings (<0.80), PC2 is interpreted as economic development. The third principal component (PC3) is dominated by the spatial accessibility of voluntary counseling and testing (VCT) services, which shows the highest loading; accordingly, PC3 is labeled healthcare.

The three principal components derived from PCA were incorporated into the GWR model to estimate local regression coefficients. We then performed street-level spatial interpolation to visualize the 2020 spatial distribution of determinants of HIV/AIDS prevalence in Kunming.

Economic level is widely regarded as a key social determinant of the AIDS epidemic, particularly in low- and middle-income countries (16–18). In Kunming, the effect of economic development on HIV/AIDS prevalence shows clear spatial variation. Areas with lower development—such as Luquan Yi and Miao Autonomous County, Xundian Hui and Yi Autonomous County, Dongchuan District, and Songming County in northern Kunming—exhibit higher prevalence. In these settings, health awareness is relatively weak and limited resources constrain the coverage of prevention and treatment measures. Moreover, low-income populations often experience greater mobility (e.g., labor migration and prolonged separation from family or spouses), which increases the likelihood of engaging in high-risk behaviors such as commercial sex or drug use, thereby elevating HIV transmission risk.

Healthcare capacity plays a pivotal role in HIV/AIDS prevention and control (18, 19). In Kunming, southern districts with stronger medical and public health conditions (e.g., Wuhua, Panlong, Guandu) show lower HIV/AIDS prevalence, reflecting the significant impact of healthcare resources on epidemic control. Higher levels of care not only facilitate early diagnosis and treatment but also promote earlier intervention by increasing testing and reporting rates. However, in areas with better health infrastructure, the continued expansion of VCT and PITC coverage leads to the identification and reporting of more infections, and these areas may therefore exhibit apparently higher prevalence.

Educational attainment is strongly associated with HIV/AIDS awareness and preventive behaviors. Findings from the Kunming study indicate that areas with a higher proportion of low-educated populations—such as Xundian Hui and Yi Autonomous County, Luquan Yi and Miao Autonomous County, and Dongchuan District in the northern part of the city—tend to have higher HIV/AIDS prevalence rates. Populations with limited education generally have lower HIV/AIDS knowledge and weaker self-protective behaviors, making them more susceptible to infection and onward transmission. Evidence from low-income, ethnically diverse border regions—such as Sichuan Province and Dehong Prefecture in Yunnan—shows that the rate of late HIV/AIDS diagnosis is highest among illiterate populations (20, 21). Limited knowledge and weak self-protection among low-educated groups substantially increase their risk of infection.

In addition, we found that the unique cultural background of ethnic minority groups is an important factor contributing to the prevalence of HIV/AIDS in multi-ethnic regions, a finding that has also been supported by related studies (22–24). Kunming is a modern, multi-ethnic city that is home to 11 minority groups and is characterized by rich cultural diversity. Ethnic minority populations are more susceptible to HIV/AIDS infection due to limited proficiency in Mandarin, lower levels of education, insufficient awareness of HIV/AIDS, and weaker self-protection consciousness and capacity. Most ethnic minority populations tend to embrace a more naturalistic worldview and hold relatively open attitudes toward sexuality, showing greater tolerance of premarital sex. They generally experience sexual initiation at an earlier age, have multiple sexual partners, and lack both the awareness and access to condom use, leading to a higher prevalence of high-risk sexual behaviors associated with HIV transmission.

The prevalence of HIV/AIDS is shaped by the complex interplay of multiple contributing factors. The study found that northern areas of Kunming—such as Luquan Yi and Miao Autonomous County, Xundian Hui and Yi Autonomous County, and Dongchuan District—are affected by a combination of economic poverty, limited educational resources, and inadequate healthcare services. In future HIV/AIDS prevention and control efforts, resources should be allocated rationally, and public health infrastructure should be strengthened through increased government investment and the active participation of social organizations. Particular attention should be given to enhancing the supply of HIV/AIDS prevention materials and improving the accessibility of medical services. Second, in the healthcare sector, efforts should be made to strengthen its role in early screening, diagnosis, and standardized treatment of HIV/AIDS. Greater investment from both the government and society is needed to improve the testing and treatment service system, thereby enhancing the accessibility and effectiveness of HIV prevention and control at the primary level. In the field of education, greater emphasis should be placed on disseminating HIV/AIDS-related knowledge, particularly by strengthening sex and health education, to enhance public awareness of HIV/AIDS prevention and control—especially among populations in remote areas and those with lower levels of educational attainment. In the southern areas of Kunming, management should be strengthened to ensure regular follow-up of people living with HIV/AIDS, improve their quality of life, and reduce the risk of HIV transmission from high-risk groups to the general population.

This study has certain limitations. First, prior research suggests that differences in sex and age may influence infection risk and behavioral patterns (25), yet these demographic characteristics were not incorporated into our analysis. Because all influencing factors were derived from district/county–level data and downscaled to street-level points via spatial interpolation, precise spatial distributions for populations by sex or specific age groups are unavailable. Given the high population mobility in Kunming, existing data cannot accurately characterize the spatial features of stratified populations, thereby limiting in-depth analysis of the potential moderating effects of demographic factors on the HIV/AIDS epidemic.

In addition, the range of variables in this study was limited. Certain potentially important factors closely related to sex and age—such as illicit blood trading and social media use—were not incorporated due to data confidentiality and restricted access. Future work that integrates more detailed census or migrant-population datasets, together with additional variables capturing individual behaviors, would enable a more in-depth exploration of the mechanisms driving the HIV/AIDS epidemic.

Second, 2020 coincided with the COVID-19 pandemic, which disrupted healthcare systems and public health behaviors. On one hand, the pandemic likely reduced the willingness of some populations to seek voluntary testing, leading to undetected infections; on the other hand, the reallocation of attention and resources from HIV programs to COVID-19 response weakened HIV prevention and treatment capacity. Consequently, the 2020 data may underestimate the true level of HIV prevalence.

The integration of PCA and GWR effectively alleviated issues of multicollinearity and spatial non-stationarity, thereby improving the model’s ability to capture spatial heterogeneity in HIV/AIDS prevalence. Because the analysis relies on data aggregated at the administrative-unit level, the findings may still be affected by the modifiable areal unit problem (MAUP). Differences in spatial scale and boundary delineation may exert certain influences on the estimation of model parameters. Although the GWR model can partially mitigate this issue, it cannot completely eliminate its effects. Future research could incorporate higher-resolution raster data or individual-level datasets to further validate the findings and enhance the robustness and generalizability of the results.