Hi folks,
I am working on a panel regression analysis of crime outcomes at the census block group level, and I am using ACS data as covariates.
My study period is 2014 - 2018, with my unit of time as meteorological season (so overall, T = 19).
Unfortunately, using fixed effects (as suggested by my model specification tests) wash out time-invariant variables. Thus, if I use 2018 5-year ACS estimates, my covariates essentially get removed from my analysis (as they would not be changing during my study period). And, unfortunately, 1-year estimates are not available at the BG level.
So my question is this: for each year in my study, would it be reasonable to use each year's respective 5-year estimate to explain my covariates? This would give me some time-related variance in my covariates, but I'm having trouble deciding if it is methodologically sensible?
Regards (and happy holidays!),
Shaun Post
My understanding of your issue is: meteorological season is that you have 5 years with 4 seasons each (and you have dropped one season). You were planning to use the average of some demographic characteristics over that period (2014 to 2018, correct?) but because you are using a fixed effects model and that average is constant over time for any particular block group, it drops out. You want to know if it is OK to use a moving average instead.
I have two pieces of advice:
1) Have you considered using a hierarchical model? That would allow you to look at the influence of the demographic context separate from the influence of time on your dependent variable. This is a model used quite a bit in education to look at students (lower level) within classrooms (higher level). In the case of your model, the higher level would be the demographic context and the lower level would be the passage of time.
2) If you absolutely must use a fixed effects model, you could use the 5-year data centered on the observation year for the first and last year and let the others have missing data. That would mean you would only have one year of overlap (2016) -- 2014 would use the 2012-2016 ACS and 2018 would use the 2016-2020 ACS. You will also note that the 2016-2020 only just wrapped up the data collection phase and the data will not be available for another 11 months.
thanks so much for your thoughtful reply Katherine Nesse!. unfortunately, one thing i did not mention was that the panel models i have been working with have spatial dependency components, and to my knowledge the use of HLM within a spatial framework is still developing (i.e. not sure there are packages to analyze such data). additionally, much of the literature i have seen regarding spatial panels assume you have a completely balanced panel (i.e. observations at all points), and i'm unfamiliar with dynamic panel models to this point.
i did end up running my analysis with the 5-year estimates representing each year of my study, so it may come down to if my justification in the manuscript is accepted as a reasonable trade-off/limitation by reviewers.