Sparse HP Filter: Finding Kinks in the COVID-19 Contact Rate

Sokbae Lee, Yuan Liao, Myung Hwan Seo, Youngki ShinJournal of Econometrics 2020

TL;DR

This paper introduces a novel "sparse HP filter" to estimate the time-varying COVID-19 contact rate, a crucial parameter in epidemiological models. By adapting the traditional Hodrick-Prescott filter with a constraint on the number of "kinks" (sudden changes in trend), the method effectively identifies significant shifts in transmission dynamics that align with real-world events, offering a more interpretable and parsimonious approach than existing methods.

What Problem Does This Paper Address?

Understanding and accurately estimating the COVID-19 contact rate (often denoted as $\beta_t$) is fundamental for modeling the pandemic's spread and informing public health policies. The contact rate, representing the average number of adequate contacts a person has per unit of time that can lead to transmission, is not static. It's influenced by a myriad of factors including government interventions (lockdowns, social distancing mandates), individual behavior (self-isolation), and the evolving nature of the virus. Traditional SIR (Susceptible-Infected-Recovered) models often assume a constant contact rate, which is a poor approximation for a dynamic pandemic. While methods exist to estimate time-varying contact rates, raw data is often too noisy and volatile to draw reliable inferences. Existing trend filtering techniques, like the $l_1$ trend filter, can identify changes but may produce an excessive number of "kinks," making interpretation difficult. The challenge is to develop a method that can reliably detect significant shifts in the contact rate trend, corresponding to real-world events, without being overly sensitive to noise or generating too many spurious changes.

Methods

The core contribution is the sparse HP filter, a new trend filtering method. It builds upon the classic Hodrick-Prescott (HP) filter, which is widely used in macroeconomics to decompose time series into trend and cycle components. The standard HP filter minimizes a sum of squared errors (fidelity term) and a penalty on the second difference of the trend (smoothness term). The authors' innovation is to introduce a constraint on the number of "kinks" in the trend, effectively limiting the degree of piecewise linearity. This is achieved by adding an $l_0$ constraint to the HP filter's objective function, which penalizes the number of times the second difference of the trend is non-zero.

Specifically, the sparse HP filter minimizes the standard HP objective function subject to a constraint on the number of kinks ($\kappa$). This formulation is solved using mixed integer quadratic programming (MIQP). The authors also compare their method to the $l_1$ trend filter, which is known for its ability to produce piecewise linear trends and identify structural breaks. While the $l_1$ filter is effective, the authors note it can sometimes produce too many kinks, similar to how LASSO can select too many covariates. The sparse HP filter aims to strike a balance by allowing for kinks but controlling their number.

Datasets

The study utilizes daily data on confirmed, recovered, and deceased cases from publicly available sources. Specifically, they use data from the Johns Hopkins University Center for Systems Science and Engineering (JHU CSSE). The analysis is conducted on data from five countries: Canada, China, South Korea, the UK, and the US. The data is processed to derive a measure of the time-varying contact rate, denoted as $Y_t$, which is then log-transformed to $y_t$ for filtering. The analysis period generally spans from early March to early June 2020, with specific end dates varying by country based on data availability and pandemic status.

Key Insights and Ideas

  • The sparse HP filter offers a parsimonious yet effective way to identify significant shifts in the COVID-19 contact rate. By directly controlling the number of kinks, it provides a more interpretable trend than methods that might over-segment the data.
  • Kinks identified by the sparse HP filter align remarkably well with actual historical events. This suggests the filter is capturing meaningful changes in transmission dynamics driven by policy interventions and behavioral shifts.
  • The $l_1$ trend filter, while capable of identifying breaks, tends to produce a significantly larger number of kinks compared to the sparse HP filter, even when fitting the data equally well. This highlights the potential for over-segmentation and reduced interpretability with the $l_1$ approach.
  • The proposed method allows for the calculation of "contact growth rates," which are essentially the growth rates of the estimated contact rate (or the basic reproduction number $R_0(t)$). These rates serve as a valuable surveillance tool to monitor the epidemic's trajectory and the effectiveness of interventions.

Key Results

  • The sparse HP filter successfully identified kink dates in the contact rate that correspond to significant events such as lockdowns, policy changes, and the initial stages of outbreaks in the five studied countries.
  • For the US, the sparse HP filter identified four key kink dates (March 16, March 20, April 14, and May 13), dividing the period into five distinct phases with varying contact rate trends. These dates were well-aligned with major policy announcements and observed changes in epidemic dynamics.
  • In comparisons, the sparse HP filter produced fewer kinks than the $l_1$ trend filter across all countries, while achieving comparable data-fitting performance. For instance, the $l_1$ filter identified 10 kinks for the US, whereas the sparse HP filter identified 4.
  • The study also theoretically establishes the risk consistency of both the sparse HP and $l_1$ trend filters, providing a solid statistical foundation for their use.

Limitations and Future Work

  • Data Quality and Under-reporting: The authors acknowledge that the data used (confirmed, recovered, and deceased cases) are subject to measurement errors and potential under-reporting, especially in the early stages of the pandemic. While they show their model is robust to some degree of under-reporting, time-varying under-reporting could still bias estimates.
  • Real-time Analysis Caveat: The empirical analysis was conducted using data up to early June 2020. The authors note that this means some remarks might be out of sync with the real-time progression of the pandemic.
  • Future Research Directions:
    • Distinguishing Methods: Further theoretical work is suggested to distinguish between the sparse HP and $l_1$ filters, particularly regarding kink selection consistency.
  • Inference and Testing: Developing methods for testing the presence of kinks and for inferring their locations and magnitudes would be valuable.
  • Forecasting: While the linear trend specification is suitable for short-term forecasting, long-term prediction remains challenging.
  • Panel Regression: Extending the analysis to a panel regression model for the contact rate at city, state, or country levels is proposed as an interesting avenue for future research.

Summary updated

Abstract

In this paper, we estimate the time-varying COVID-19 contact rate of a Susceptible-Infected-Recovered (SIR) model. Our measurement of the contact rate is constructed using data on actively infected, recovered and deceased cases. We propose a new trend filtering method that is a variant of the Hodrick-Prescott (HP) filter, constrained by the number of possible kinks. We term it the sparse HP filter and apply it to daily data from five countries: Canada, China, South Korea, the UK and the US. Our new method yields the kinks that are well aligned with actual events in each country. We find that the sparse HP filter provides a fewer kinks than the l1 trend filter, while both methods fitting data equally well. Theoretically, we establish risk consistency of both the sparse HP and l1 trend filters. Ultimately, we propose to use time-varying contact growth rates to document and monitor outbreaks of COVID-19.
BibTeX
@article{lee2020sparse,
  author = {Sokbae Lee and Yuan Liao and Myung Hwan Seo and Youngki Shin},
  title = {{Sparse HP Filter: Finding Kinks in the COVID-19 Contact Rate}},
  journal = {Journal of Econometrics},
  year = {2020},
}
Figure