00000ntm a22000004a 4500 UP-100084294495528776 Buklod 20230814113137.0 m |o d | ta 230814s2022 xx grm 000 0 eng d (Buklod)UP-99796217607232816 DSTC rda eng DMLUC LG 995 2022 S8 B55 Bilon, Xavier Javines author Bootstrap-based parameter estimation for networked hard-to-reach populations under respondent-driven sampling Xavier Javines Bilon. Quezon City University of the Philippines Diliman 2022 xxii, 238 leaves illustrations 28 cm text rdacontent unmediated rdamedia volume rdacarrier Includes UPSS OGS FORM 2c (STAT 300 FINAL DEFENSE). Thesis Master of Statistics University of the Philippines Diliman 2022 Hard-to-reach populations present a methodological challenge in obtaining samples and in drawing inference about population characteristics from the samples obtained. Such challenge follows from the undefined population size and boundaries and the involvement of stigmatized or illegal behavior in one's membership in the population that limit the usefulness of standard sampling methods and the corresponding inference procedures in the types of population. Respondent-driven sampling (RDS) which views the sample as a subset of the social network under study can mitigate these issues. However, despite the practicality of RDS procedure, and while the current estimation procedures are simple and analytically tractable, they depend on several strong assumptions. In this study, we proposed employing bootstrap-based procedures --- nonparametric bootstrapping for homogenous means, and model-based and residual-based bootstrapping for heterogenous means --- in estimating parameters of networked hard-to-reach populations. Simulation studies show that, in estimating homogenous means, the proposed procedure --- nonparametric bootstrapping estimation --- was not only asymptotically unbiased, but also performed better than the existing procedure --- the RDS II estimation (Volz & Heckathorn, 2008) --- across most network configurations and parameter values considered. Additionally, the proposed RDS estimation procedure for homogenous means was robust against the presence of few and small or average-sized clusters. On the other hand, in estimating heterogenous means, the three RDS estimation procedures --- least squares estimation (LSE), and model-based and residual-based bootstrap estimation --- performed similarly across most different network configurations and parameter values considered. although, in some cases, the model-based or residual-based bootstrap procedure appears to have performed better than the others. Results also show that, in constructing confidence intervals, the residual-based bootstrap estimation was more robust against the presence of clusters than the other two procedures. Furthermore, simulation studies conducted demonstrate that the bootstrap-based procedures performed well when not too few and not too many seeds (e.g., 4-10 initial participants) and recruits (e.g., five recruits) were used in obtaining the respondents-driven samples. Lastly, the proposed estimation procedures were used on actual RDS samples from hard-to-reach populations (i.e., people who inject drugs in a region in ESTONIA and SYRIAN activist-refugees in Jordan) to illustrate the applicability and advantages of our approach. For both the homogenous and heterogenous means, point estimates obtained using the proposed procedures and their corresponding estimated standard errors supported the results of the simulation studies. Computer bootstrapping. Monte Carlo method. Social networks. Respondent-driven sampling. Thesis FI UP UPD DSTC LG 995 2022 S8 B55 Book