February 24, 2026
Analysis of indebtedness by payroll loans in municipal public servants
Lucas Piovesan Kobayashi; Charles Gomes da Silva
Summary prepared by the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege focused on synthesis and writing.
This study uses multinomial logistic regression to identify the demographic, professional, and financial factors that influence the indebtedness profile of employees of a public hospital of the Prefecture of the Municipality of São Paulo (PMSP). The objective is to predict the probability of an individual belonging to a low, moderate, or high risk profile of indebtedness through payroll loans. The research is justified by the gap in the literature, which focuses on federal employees, with unified legislation and higher salaries (Loureiro, 2021), leaving municipal employees as an underexplored area, despite their representativeness.
Municipal indebtedness is complex due to varied local legislations, economic pressures, and generally more modest incomes. The absence of a national regulatory standard for municipal payroll loans creates disparities and obscures the dimension of the problem (Macedo, 2023). In São Paulo, Law nº 8.989/1979 limited the payroll loan margin to 30% of net remuneration. Decree nº 58.890/2019 expanded the possibilities by instituting the payroll loan card and the benefits card, with additional margins of 5% and 10%, respectively. This normative evolution reflects the tension between the demand for credit and the concern with over-indebtedness.
A paradox of public service is that functional stability, instead of guaranteeing financial security, makes civil servants a preferential target for credit institutions. The payroll deduction guarantee allows for the offering of loans with attractive interest rates, which in São Paulo range from 1.98% to 2.61%, and terms of up to 120 months. During the COVID-19 pandemic, Decree No. 60.316/2021 temporarily extended the margin to 35%, an emergency measure that potentiated the risk of a debt cycle. Such adjustments, without financial education programs, can aggravate the vulnerability of civil servants.
Indebtedness is not solely due to easy access to credit or poor individual management, being influenced by social and behavioral factors. The literature on financial education, although recent in Brazil, emphasizes the importance of budget planning for financial sustainability (Cordeiro, Costa e Silva, 2018). For São Paulo’s municipal employees, the combination of stability, facilitated credit, and portability creates an environment conducive to continuous indebtedness. The finding that up to 45% of their salary can be consumed by payroll deductions highlights the urgency in understanding the variables that modulate this behavior.
Multinomial logistic regression is the appropriate methodological approach, as it allows specifying a model to explain and predict a categorical dependent variable — the indebtedness risk profile — from explanatory variables (Favero and Belfiore, 2017). The dependent variable, operationalized into three categories (low, moderate, and high risk), justifies the choice of the multinomial model for granular analysis. The study therefore seeks to identify the correlates of indebtedness and provide a predictive model that can support municipal administration’s personnel management policies and financial well-being programs.
The methodology started with data collection from an IT system of the PMSP, with a temporal scope from August 2023 to August 2025, covering hospital servers with some payroll deduction. To ensure data quality, the ETL (Extract, Transform, Load) technique was employed. The initial database was transactional, with each row representing a contract, meaning a server could appear multiple times. A preliminary analysis revealed that approximately 41% of servers, active and retired, had at least one payroll deduction credit contract.
The data transformation process was executed with the pandas library in Python. The steps included converting variables to numeric and date formats, handling inconsistent values, and removing rows with missing data. The most critical step was data aggregation: using the unique server identifier (“RF/VINC”), the transactional base was consolidated into an individual-centered analytical base; each row represents the complete profile of a server. Columns with a high incidence of missing values and low predictive relevance, such as “ATIVIDADE”, were discarded.
Data engineering was crucial for deriving significant indicators. Six explanatory variables were created: Age, TimeofService, TotalLoans (number of contracts), TotalInstallmentsValue (sum of monthly installments), NumUniqueConsignees (number of distinct financial institutions) and the DebttoIncome Ratio (TotalInstallmentsValue divided by gross income). To operationalize indebtedness risk, theRiskIndex was developed. This index was calculated for each server by summing the scores (from 1 to 4) obtained in three metrics: TotalLoans, NumUniqueConsignees and DebttoIncome Ratio, after classifying each server into quartiles for each of these metrics.
The analysis of the results began with the evaluation of the model’s overall fit. The Likelihood Ratio Test, comparing the full model against a null model, resulted in a Chi-squared (χ²) statistic of 245.7, with 28 degrees of freedom and an extremely low p-value (p < 0.001). This result allows us to reject the null hypothesis that the coefficients are zero, providing evidence that the independent variables have explanatory power. The quality of fit was corroborated by McFadden’s Pseudo R² of 0.31, indicating that the model represents a 31% improvement in fit compared to the null model. A value in this range is considered indicative of an excellent fit in social sciences (Greene, 2018).
The significance of each predictor was evaluated by the Wald z-test, with the “Low Risk” category as the reference. For the “High Risk vs. Low Risk” comparison, four predictors were highly significant. The most impactful factor was retirement status (STATUS_Retired), with an Odds Ratio (OR) of 17.29 (95% CI: [5.85, 51.05]). This indicates that the odds of a retired employee being classified as ‘High Risk’ rather than ‘Low Risk’ are more than 17 times higher than for an active employee, ceteris paribus. This finding quantifies the hypothesis that the stability of benefits may facilitate more aggressive indebtedness.
The Age variable also showed a positive and significant effect: for each additional year, the chances of belonging to the ‘High Risk’ group increase by 8% (OR = 1.08, 95% CI: [1.04, 1.12]). In contrast, Gross Income (BRUTO) showed a protective effect, with a negative and significant coefficient (OR slightly less than 1.00), indicating that for each increase in income, the chances of being classified as ‘High Risk’ decrease. This suggests that the relative income burden is lower among employees with higher salaries.
Another relevant finding was the effect of sex. Being male (SEX_M) significantly reduced the chances of belonging to the ‘High Risk’ profile, with an OR of 0.43 (95% CI: [0.22, 0.84]). This means that the chances for a man are 57% lower than for a woman, ceteris paribus, which may reflect differences in financial behavior or family responsibilities (Johnson & Smith, 2020). For the “Moderate Risk vs. Low” comparison, the effects were in the same direction but with smaller magnitudes. Retired status (OR = 3.49) and age (OR = 1.04) continued to increase the chances of risk, while gross income showed a protective effect, all with statistical significance (p < 0.05). The Sex variable, however, was not significant for this comparison (p = 0.067).
The visual interpretation of the results, through forest plots, highlighted the STATUS_Retired variable as the strongest predictor for ‘High Risk’. Additionally, predicted probability graphs illustrated that, with advancing age, the probability of belonging to the ‘Low Risk’ group decreases, while that of belonging to the ‘High Risk’ group increases sharply, especially after 60 years of age, corroborating the association between aging, retirement, and a higher risk of indebtedness.
The discussion of the results points to a complex interaction of factors. The vulnerability of retirees can be explained by the stability of their income, which makes them an attractive target, added to a possible reduction in benefits, health expenses, and support for family members. Municipal legislation on pension calculation may result in lower income in retirement, increasing dependence on credit. Stability functions as a guarantee for the lender, but not necessarily as protection for the debtor.
The protective effect of higher income is expected, as it offers greater room to absorb debt. However, initial bivariate analysis suggested that median income was slightly higher in the higher-risk groups, indicating that borrowing capacity also increases with income. The regression model, by controlling for multiple factors, clarifies that, for a given borrowing behavior, higher income indeed reduces relative risk. The association of the female sex with higher risk may be linked to socioeconomic factors, such as greater female representation in lower-paying positions and the social role of managing the family budget.
In summary, the model’s results provide a detailed and quantifiable diagnosis of the factors that modulate indebtedness through payroll loans in the sample. The identification of clear risk profiles, especially the high vulnerability of retired and older servers, offers concrete subsidies for the formulation of public policies and institutional actions, such as targeted financial education programs, review of credit limits for specific populations, and active monitoring of indebtedness by public administration.
Multinomial logistic regression analysis successfully identified demographic and functional factors that discriminate indebtedness profiles, with the model demonstrating a robust overall fit. The profile of the server most likely to be classified as ‘High Risk’ is characterized by being retired, having advanced age, being female, and having lower gross income relative to their indebtedness. Retired status emerged as the strongest predictor. Study limitations include the cross-sectional nature of the analysis, which prevents causal inference, and the fact that the dependent variable is a methodological construct that approximates risk, not an outcome such as default.
For future work, a longitudinal perspective is suggested to analyze the evolution of indebtedness, as well as the investigation of loan renegotiation dynamics. The analysis of the impact of the negative margin and the study of behavioral and financial literacy variables could add new dimensions to the understanding of the phenomenon. It is concluded that the objective was achieved: it was demonstrated that retiree status, age, sex, and gross income are factors that significantly influence the probability of a municipal public servant belonging to a certain risk profile of indebtedness through payroll loans.
References:
Agresti, A. (2013). Categorical Data Analysis. John Wiley & Sons.
Banco Central do Brasil (BACEN). (2022). Relatório de Estabilidade Financeira. Brasília, DF: BACEN.
Cordeiro, A., Costa, F., & Silva, J. (2018). A importância da educação financeira no contexto brasileiro: desafios e perspectivas. Revista de Finanças e Contabilidade, 9(2), 45-62.
Fávero, L. P.; Belfiore, P. (2017). Manual de análise de dados. 1ed. Elsevier, Rio de Janeiro, RJ, Brasil.
Freitas, L. (2019). Testes de hipótese em Modelos Multivariados de Covariância Linear Generalizada (McGLM). Recuperado de https://lineu96. github. io/mscpage/2wald_test. html
Gomes, P. (2024). Regressão Logística: Um Guia Detalhado. Recuperado de https://www. datageeks. com. br/regressao-logistica/
Greene, W. H. (2018). Econometric Analysis. 8th ed. Pearson.
Hair, J. F., Black, W. C., Babin, B. J., & Anderson, R. E. (2019). Multivariate Data Analysis. 8th ed. Cengage Learning.
Johnson, R., & Smith, L. (2020). Gender Disparities in Financial Literacy and Indebtedness. Journal of Economic Psychology, 41, 102-115.
Kimball, R., & Ross, M. (2013). The Data Warehouse Toolkit: The Definitive Guide to Dimensional Modeling. 3rd ed. John Wiley & Sons.
Loureiro, D. (2021). Endividamento Do Servidor Público: A Possibilidade de Estudo da Realidade do MPDFT sob a Perspectiva do PES. Dissertação de Mestrado, Instituto de Pesquisa Econômica Aplicada – IPEA, Brasília, DF, Brasil.
Macedo, T. (2023). Endividamento e uso de Crédito Consignado: Um Olhar Sobre a Realidade do Servidor Público Federal da UNIVASF. Dissertação de Mestrado, Universidade Federal do Vale do São Francisco, Bahia, BA, Brasil.
Prefeitura do Município de São Paulo (PMSP). (2022). Anuário Estatístico do Servidor Público. São Paulo, SP: Secretaria Municipal de Gestão.
Silva, P. (2024). Aplicação De Modelo de Regressão Logística Politômica. Trabalho de Conclusão de Curso, Universidade Estadual Paulista Júlio Mesquita Filho, São Paulo, SP, Brasil.
Executive summary from the Final Course Work of the Specialization in Data Science and Analytics of the MBA USP/Esalq
Learn more about the course; click here:
