Univaf Sandbox
Last compiled at 2021-07-16 17:12
In this report we take cursory look at slot-level COVID appointment availability data as gathered by USDR’s UNIVAF project. The main question we’re looking at is whether this kind of data can be used to identify issues in vaccine access by different demographics. Another question is whether this can help us understand the role of barriers to access, as opposed to more ideological barriers, can help us understand trailing vaccination rates.
- This report pertains to slot-level appointment data only. We often don’t have slot-level data, so this covers a subset of appointments. Of the 39052 locations in our national DB, 17210 have slot-level data.
- We take as the interval a slot is available for as the time between the first time we saw it being available to the last time.
- ‘work time’ is defined as Mon-Fri 9am-5pm.
- For slots we compute the following statistics:
- how many slots are ever seen to be available for that hour
- how long (in number of hours) are they available for on average
- what is the average latest time (in number of hours ahead for the appointment time) you can still book the slot
Number of slots per state per day
Slot stats
by state 
Visualization of range of checked_at time
- sample 10 appointments per day (5 from walgreens and 5 other ones)
- Mondays are highlighted
- Wallgreens is overrepresented and releases their appointments on Friday apparently.
- some appointments are availale after the slot time
Availability by vulnerability (national, by zip)
We can compare appointment availability by the demographics of the area that the distribution location is in. The most granual level is the zip-code. Of course, people can travel to locations outside of their zip-code, and so a better analysis would, for every zip code, take an average of the nearby locations weighted by their distance.
We do univariate correlations between vulnerability indices like SVI and availability statistics. In this plot, the points represent individual zip-codes. The red line is a linear fit of the relationship between the vulnerability statistic and the availbility statistic.
Note that for ‘last available’ (min lead time), less is better. For some indices, we see a slight negative correlation between vulnerability and availability.
Note: because we don’t have vaccination data on the zip-code level, we can’t (unlike below for counties) adjust slots/person by coverage, nor can we plot the SVI statistics against vaccination rates on a zip-code level.
We can combine these factors by running linear regressions for the outcome variable we’ve looked at so far (fixed effects for state not shown):
##
## ================================================================
## Dependent variable:
## -----------------------------------------------
## slots_per_person hours_available last_available
## (1) (2) (3)
## ----------------------------------------------------------------
## Population (log) -0.097*** 1.389*** 0.653***
## (0.003) (0.506) (0.132)
## Share Black 0.001 20.953*** -2.326***
## (0.014) (2.319) (0.606)
## Socioeonomic SVI 0.007 -3.608 0.050
## (0.016) (2.613) (0.683)
## Household SVI -0.061*** -15.185*** -6.483***
## (0.013) (2.181) (0.570)
## Minority SVI 0.071*** 1.860 10.699***
## (0.013) (2.122) (0.555)
## Housing SVI -0.029** -10.248*** -0.419
## (0.013) (2.148) (0.561)
## Intercept 1.062*** 97.463*** -3.438
## (0.054) (8.774) (2.293)
## ----------------------------------------------------------------
## Observations 8,027 8,027 8,027
## Adjusted R2 0.135 0.133 0.292
## ================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01- Unlike in the next section, we can’t account for our coverage by dividing slots/person by bookings/vaccinations, so slots/person is at least partially a measure of coverage in this regression and shouldn’t be taken very seriously.
- Accounting for all five measures of vulnerability, zip-codes with a higher percentage Black residents have better booking time-based access, while zipcodes that are high on any of the SVI metrics have worse access (shorter booking windows and less time to book).
Does availability predict vaccinations? (national, by county)
- data is for 4 weeks: 2021-05-24 to 2021-06-20
- vaccination counts from covid act now
How does appointment availability relate to vaccinations? Unfortunately we don’t have vaccination data on the zip-code level, so we’ll do this analysis on a county level. Vaccination counts by county by day are pulled from CovicActNow.org. The outcome we look at is the number of vaccinations that happened in the two weeks that we’re looking at. We run a simple linear regression with counties as units:
##
## ============================================================================
## Dependent variable:
## ----------------------------------------------------
## n_vax/population log(n_vax)
## (1) (2) (3) (4) (5)
## ----------------------------------------------------------------------------
## % hesitant (CDC) -0.036*** 0.487
## (0.013) (0.432)
## Trump vote share -0.024*** -0.552***
## (0.002) (0.064)
## Prior vax rate 0.0001*** 0.007***
## (0.00003) (0.001)
## Slots/Person (workday) 0.228*** 0.161*** 7.791*** 4.993***
## (0.012) (0.011) (0.495) (0.356)
## Slots/Person (weekend) 0.195*** 0.223*** 4.392*** 4.435***
## (0.013) (0.014) (0.557) (0.462)
## Slots/Person (evenings) -0.329*** -0.228*** -10.714*** -6.973***
## (0.018) (0.018) (0.775) (0.574)
## Avg range (hrs) -0.00001 -0.00003*** -0.001*** -0.001***
## (0.00001) (0.00001) (0.0003) (0.0002)
## Avg lead time (hrs) 0.001*** 0.0004*** 0.009*** 0.008***
## (0.00004) (0.00004) (0.002) (0.001)
## Population (log) 1.165*** 1.147*** 1.089***
## (0.008) (0.008) (0.007)
## Intercept 0.016*** 0.029*** -5.707*** -5.565*** -5.179***
## (0.003) (0.008) (0.169) (0.156) (0.289)
## ----------------------------------------------------------------------------
## Observations 2,122 1,560 2,122 2,122 1,560
## Adjusted R2 0.706 0.743 0.938 0.948 0.976
## ============================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01- Availability stats are predictive of the number of vaccinations, and explain about 2% of the variation depending on the specification.
- There are two specifications of the outcome: vaccinations per individual (aka the difference in vaccination rate over the three weeks), and the log number of vaccinations (with the log population as a covariate).
- Hesitancy (taken from this CDC survey) and the prior vaccination rate are strongly correlated with vaccinations, but don’t change the outcomes.
Week over week
Now, week over week (only two weeks since we don’t have vaccination numbers for this week until end of week):
##
## =========================================================================================================
## Dependent variable:
## --------------------------------------------------------------------------------
## n_vax/population log(n_vax)
## (1) (2) (3) (4) (5) (6) (7) (8)
## ---------------------------------------------------------------------------------------------------------
## Slots/Person 0.024*** 0.022*** 1.923*** 1.788***
## (0.001) (0.001) (0.098) (0.096)
## Average range (hrs) 0.00000 -0.00001*** -0.0004** -0.001***
## (0.00000) (0.00000) (0.0002) (0.0002)
## Average time ahead (hrs) 0.0002*** 0.0002*** 0.011*** 0.011***
## (0.00001) (0.00001) (0.001) (0.001)
## % hesitant (CDC survey) -0.046*** -0.037*** -1.109*** -0.670*
## (0.004) (0.004) (0.356) (0.348)
## Prior vax rate 0.013*** 0.012*** 1.465*** 1.431***
## (0.001) (0.001) (0.088) (0.086)
## Population (log) 1.170*** 1.154*** 1.130*** 1.121***
## (0.005) (0.005) (0.005) (0.005)
## Intercept 0.008*** 0.006*** 0.012*** 0.009*** -6.765*** -6.700*** -6.715*** -6.748***
## (0.002) (0.002) (0.002) (0.002) (0.181) (0.177) (0.210) (0.205)
## ---------------------------------------------------------------------------------------------------------
## Observations 9,204 9,204 9,000 9,000 9,204 9,204 9,000 9,000
## Adjusted R2 0.299 0.363 0.328 0.379 0.912 0.917 0.917 0.921
## =========================================================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01- avg range negative now
- explained variance is 5-12% now.
Maps
