In microeconometrics, the time component is usually short (meaning that $T$ is fixed in $t=1,\ldots,T$). Serial correlation is here usually just seen as a negligible issue affecting the standard errors of a standard linear regression model (and not the point estimates) because it can be easily correct via robust standard error (both adjusting for possible heteroscedasticity and serial correlation, a similar approach as to use the Newey-West correction).
A lot of studies have however apparently calculated biased standard error in the difference-and-difference context as demonstrated by Bertrand, Duflo and Mullainathan (2003). They recommend, among others, to use block bootstrap for such an analyses. Yet, their study focuses mainly on the validity of different correction mechanisms with respect to the number of available groups ($n=1,\ldots,N$ in a panel context) but less on the length of the time component $T$.
I have some of questions based on my limited understanding:
- Are there other good overviews I should be aware of?
- Can you recommend an introduction or a paper about how to assure correct standard error in the time series literature for real-world time series data (not focusing on entirely on theoretical asymptotics but taking into account, for example, time series with multiple seasonalities such as hourly data; or long time series in a slowing but changing world / the effects of a slowly changing data generating process such as climate change). Here is a recent slightly related question which inspired me to raise these questions.
- Is there a similar overview available differentiating more between shorter and longer panels?
- Are the Bertrand et al. conclusions still considered as current state of the literature after more than 15 years?
An answer to any of these question is welcome!