Principal component analysis with autocorrelated data

Bartolomeu Zamprogno (1) , Valdério Reisen (1, 2) , Pascal Bondon (2) , Higor Cotta (1, 2) , Neyval Costa Reis Júnior (1)

1 UFES - Universidade Federal do Espirito Santo
2 L2S - Laboratoire des signaux et systèmes

This paper contributes to the analysis, interpretation and the use of the principal component analysis in a multivariate time-correlated linear process. The effect of ignoring the autocorrelation structure of the vector process is investigated. The results show a spurious impact of the time-correlation on the eigenvalues. To mitigate this impact, a pre-filtering procedure to whiten the data is applied. The methodology is used to identify redundant particulate matter measurements in a region in Brazil. Among the eight considered monitoring stations, it is found that three are needed to characterize the dynamic of the pollutant in the region.

Keywords: Principal component analysis; autocorrelation; cross-correlation; eigenvalue; air pollution

Domains: Statistics [stat]; Methodology [stat.ME]; Statistics [stat]; Applications [stat.AP]

Bartolomeu Zamprogno, Valdério Reisen, Pascal Bondon, Higor Cotta, Neyval Costa Reis Júnior. Principal component analysis with autocorrelated data. Journal of Statistical Computation and Simulation, 2020, 90 (12), pp.2117-2135. ⟨10.1080/00949655.2020.1764556⟩. ⟨hal-02560885⟩