Strengths: Well-motivated analysis that develops a model for COVID dynamics and demonstrates by simulation that a lockdown orders can explain the Fall 2020 wave in PA.
Points for consideration:
“We observe no significant evidence that the ARIMA model performs better than white noise” is an error: The AIC table shows that ARMA(1,1) is a big improvement over white noise, with some small potential advantage from larger models.
The authors identify model misspecification as a likely cause of the high variability when numerically evaluating and maximizing the likelihood. In the time allowed, it is hard to resolve such issues. I think the high weekly variability in measurememt (likely not present in the actual transmission dynamics) may be relevant. Also, perhaps the noise modeling in the process and/or measurement model is a misfit.
The project’s interpretation of model misspecification issues is reasonable. However, a simulation study to test the optimization on simulated data could have confirmed that the inference methodology was working correctly. Modeling COVID is not easy: see Projects 13 & 15 for a successful approach.
The initial ACF plots are unpolished: it can be unclear what we learn from an ACF of data with substantial trend, and attention is needed to graph labels.
What is the red horizontal line in the log_test_positive_ratio plot?