The Pre-Pandemic Growth in Online Public Education and the Factors that Predict It

ABSTRACT

While spring of 2020 introduced virtual instruction to all public schools, virtual schooling had already been growing in most states. We focus on pre-COVID-19 changes to full-time virtual school enrollment in public schools, and provide evidence on the relationship between virtual school enrollment, internet speed, community demographics, and traditional K–12 school achievement levels. We find negative associations between online enrollment and test achievement in brick-and-mortar schools, and low internet speeds. There is some evidence that students are less likely to enroll in virtual schools as the share of students of their own demographic in brick-and-mortar schools increases.

KEYWORDS:

• public education
• school choice
• virtual instruction

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1. As described in Gemin et al. (2015), “Virtual schools are full-time online schools, sometimes referred to as cyber schools, which do not serve students at a physical facility. Teachers and students are geographically remote from one another, and all or most of the instruction is provided online. These may be virtual charter schools or non-charter virtual schools” (p. 6).

2. There is no census of full-time virtual public school students before the first year of tracking within the Common Core of Data, 2014. Wicks (2010) bases his best estimate of 2010 full-time virtual school enrollments off of survey data, and classifies it as “somewhat conservative”. Such comparisons are challenging because definitions of full-time virtual school may be different. For instance, if “primarily virtual” schools are considered full-time virtual schools in the CCD then enrollment increased 120% since 2010. Regardless, the rise in virtual school enrollments is much higher than increases in overall enrollment (three percent) (National Center for Education Statistics, 2021a).

3. On the other hand, some states, such as New Jersey, will not allow school districts to offer virtual options in the 2021–2022 school year (Tully, 2021).

4. To be clear, we are not attempting to assess the quality of virtual K–12 education, rather whether achievement in brick-and-mortar schools predicts virtual school enrollment. For more on the limited evidence on the student achievement effects of online schooling, see Patrick and Powell (2009), Ahn and McEachin (2017), Molnar et al. (2019), and Sahni et al. (2021).

5. Estimates come from the annual National Education Policy Center on virtual schools. During this time no national data was collected in a standardized way. Instead, the authors estimate enrollment figures from reports on education management organizations and an annual report, Keeping Pace, paid for by the K–12 virtual schooling industry.

6. For a review of virtual charter enrollment by state see, Gill et al. (2015).

7. This is noteworthy as Stride is proactive in growing its enrollment via recruitment and advertising, and as such is expansion process may affect the makeup of online enrollments.

8. Mann (2019) uses the 2015–2016 CCD to examine racial segregation in online charter schools (not all of which are full-time virtual schools). He finds that 66% of students in virtual charter schools are White. Note that because of the difference in definitions between online schools and all online charter schools, it is impossible to tell if this four percent differential between Gulosino and Miron’s study and Mann’s reflects differences amongst virtual charter schools or differences in the definitions of virtual.

9. Clotfelter finds that there are tipping points, that is, White attrition to private schools becomes much more marked after the proportion of the nonwhite share reaches a certain threshold; in the case of Mississippi in the late 60s that threshold was approximately 57%. More recent work suggests that such tipping points exist in neighborhood composition (as opposed to school composition at much lower thresholds. Using a regression discontinuity design, Card et al. (2008) find that, depending on the city, White residents are more likely to leave a neighborhood when the minority share reaches 5–20%. School segregation tends to track housing segregation and this relationship has only gotten stronger over the past few decades (Frankenberg, 2013).

10. Recent research by Cordes and Laurito (2022) suggest that charter expasion has important consequences for school diversity and patterns are in flux.

11. For a full discussion on the complexity of measuring school segregation/integration, please see Ritter et al. (2016).

12. An important distinction considering the rate of change of information technology; for example, between the two periods of study, broadband access at home has climbed from 57% in 2008 to 70% in 2016 (Pew, 2021).

13. Murnane and Reardon argue that declines in private school enrollment between 1990 and 2010 are due in part to a narrowing gap in student achievement between public and private schools.

14. There is limited evidence on the academic effects of attending virtual schools, but the existing evidence suggests that student performance is lower in a virtual setting; Fitzpatrick, Berends, Ferrare, and Waddington (2020), for instance, find that on average students who switched from traditional schools to virtual charter schools in Indiana experienced negative effects on mathematics and ELA achievement.

15. There is limited evidence on the performance of students in virtual schools, but the available evidence shows that virtual schools had lower performance ratings and lower graduation rates than brick-and-mortar schools (Ahn & McEachin, 2017; Molnar et al., 2019); for instance, the graduation rate for virtual schools in 2019 was 50% compared to 84% for traditional public schools (Molnar et al., 2019).

16. Across the four years of this study the CCD records schools are coded as one of nine types: exclusively virtual, fully virtual, primarily virtual, supplemental virtual, virtual with face-to-face options, not virtual, no virtual instruction, missing, and not reported. Prior to the 2020 school year, “fully virtual” was used instead of “Exclusively Virtual”, we code “fully virtual” schools in these earlier years as virtual schools. NCES collects data for a reference date of October 1^st for a given school year, i.e., data reflect Fall enrollments. As such, all CCD utilized in this study predate potential COVID-19 induced enrollment pattern changes.

17. For more information on Stride K12 please visit https://www.k12.com/.

18. The CCD classifies some Stride schools in some years as partially virtual. For comparison, we calculate the above statistic only from CCD fully virtual CCD schools and fully virtual, per the CCD, Stride schools. While most Stride schools are fully virtual, because Stride schools can be partially virtual we refer to Stride schools as Stride schools from here on, as opposed to fully virtual schools.

19. Similarly, if students only from a single county are eligible to attend a Stride school, then all zip codes from that county constitute the catchment zone.

20. Students attending Stride schools were linked to Common Core District data by merging U.S. Census Bureau TIGER/Line shapefiles of zip codes to the National Center for Education Statistic’s Education Demographic and Geographic Estimates for district boundaries. When a zip code intersected multiple districts, the zip code was assigned to the district with the most overlapping area.

21. Median Income as opposed to Free-or-Reduced-Price Lunch Eligibility was elected for two primary reasons. Median Income is available for all observations in our sample at the level our primary outcome is: enrollment in online education at the zip code level. Second, recent work recent by Fazlul, Koedel, and Parsons (2021) shows that FRPL is a crude proxy at the school-level for estimating school-level poverty, and measures of family income from the ACS are better at depicting poverty.

22. Data on a zip code’s urbanicity come from the 2010 Rural-Urban Commuting Area Codes (RUCA) maintained by the U.S. Department of Agriculture. While not shown in the summary statistics, the breakdown of urbanicity in the analytic data is 56% Urban, 15% Micropolitan, 11% Town, and 18% Rural.

23. FCC data was linked to Stride K12 data (described below) through a zip code to census track crosswalk maintained by the Department of Housing and Urban Development (HUD): the USPS Zip crosswalk file. Because many census tracts can be matched to the same zip code, the HUD data also contains the percent of residential addresses that a particular census tract comprises of a zip code’s total residential addresses. To get one observation per zip code, we took the mean of download speeds offered across intersecting census tracts and weighted by the share of residential addresses. Stride K12 and FCC data were merged by zip codes and year.

24. A note concerning these data from the FCC, “A provider that reports deployment of a particular technology and bandwidth in a census block may not necessarily offer that service everywhere in the block. Accordingly, a list of providers deployed in a census block does not necessarily reflect the number of choices available to any particular household or business location in that block, and the number of such providers in the census block does not purport to measure competition.”

25. For more information on the granularity of the data see: https://broadbandmap.fcc.gov/#/

26. All analyses assume that students are not changing their residential zip codes so as to be able to enroll in online education. We believe that such changes are possible, but likely small because most online schools in this sample allow students from an entire state to enroll.

27. Specifically, the district cohort standardized datafile was used. The most recent available year was used for each school district. SEDA data was standardized within the sample and merged to residential zip codes in the Stride data by first merging SEDA data to district catchment boundaries maintained by NCES, and then spatially merging SEDA and zip code data.

28. The correlation between district level math and reading achievement was 0.89.

29. The code used to build and model the data is available at https://github.com/gratzt/Online_education.

30. This comports with prior research, who offer a descriptive snapshot of virtual schools in the 2014–15 school year (Gulosino & Miron, 2017).

31. We say “likely” because there are some zip codes that intersect multiple school districts. In these cases, we assume the neighborhood school district is the one most overlapping in area with the zip code. The median zip code has 89% of its area covered by one neighborhood school district. Zip codes with 0 children under the age of 18 (per the ACS data), zip codes with too few residents to warrant data collection in the ACS, and/or zip codes that were recorded as having zero residential addresses in the USPS zip code crosswalk dataset maintained by the U.S. Housing and Urban Development were dropped.

32. Statistical significance was calculated following equation 2.11 of Cameron, Gelbach, and Miller (2011) for testing the statistical significance of mean differences with clustered standard errors on more than one variable. For this analysis, observations were clustered at the Stride school and neighborhood school district level. All mean differences are weighted by the total enrollment of Stride schools or for particular student groups from the Stride schools.

33. Stride data provided information on enrollment by race/ethnicity and students residential zip codes. For this reason, it is difficult to compare variables other than these between Stride schools and neighborhood school districts.

34. We also estimate Equation (1)(1) $\begin{array}{rl}& \mathrm{l}\mathrm{n}\left(E_{STZ}\right)={\beta }_{1}AI{S}_{\left(T-1\right)Z}+{\beta }_{2}Tes{t}_Z+{\beta }_3{X}_{TZ}+{\theta }_T+U_Z+{\delta }_S+1\ast \mathrm{l}\mathrm{o}\mathrm{g}\left(Chil{d}_z\right)\\ & +{\in }_{STZ}\end{array}$(1) using ordinary least squares (OLS). Results are directionally consistent with those presented below, but we prefer the Poisson model given that OLS can produced biased estimates with low count data (Coxe, West, & Aiken, 2009); indeed we find the magnitudes of some key variables of interest to be significantly larger in OLS models than Poisson models. Lastly, we note that residual analysis suggests Poisson models fit the data substantially better than OLS. Results are available upon request.

35. We also estimated estimate Equation (1)(1) $\begin{array}{rl}& \mathrm{l}\mathrm{n}\left(E_{STZ}\right)={\beta }_{1}AI{S}_{\left(T-1\right)Z}+{\beta }_{2}Tes{t}_Z+{\beta }_3{X}_{TZ}+{\theta }_T+U_Z+{\delta }_S+1\ast \mathrm{l}\mathrm{o}\mathrm{g}\left(Chil{d}_z\right)\\ & +{\in }_{STZ}\end{array}$(1) with a cubic in available internet speeds. Results are qualitatively similar and available upon request.

36. Missing covariates were recorded as 0 and a missing dummy was added to the regression. Achievement data is missing for 1/16^th of the sample and income data is missing for 1/50^th of the sample. Models with listwise deletion are qualitatively similar and available upon request.

37. We also estimate models without Stride school fixed effects, but prefer the models with Stride effects given that Stride schools can focus on particular grades (and school levels) and/or different sub-populations of students. Models with state fixed effects are qualitatively similar to models with Stride school fixed effects and are available upon request.

38. Conditional Fixed Effect Poisson regression models cannot directly estimate clustered standard errors. To account for this, we follow Allison (2009) and bootstrap our models by sampling at the cluster level 1,000 times with replacement and run each model. Reported standard errors are the standard deviations of coefficients across the 1,000 models, and significance tests come from constructing the 95% confidence intervals using the 2.5 and 97.5 percentiles of coefficient estimates.

39. Unfortunately, school district panel student achievement data is not available for this period of study. Hence, no time series analyses were done using student achievement data.

40. This assumes that the relationship between wealthy areas and online enrollment is not fully accounted for by the median income of the zip code.

41. In some specifications, we replace school by zip code fixed effects with zip code fixed effects. We might expect differences between these models if new virtual schools open up or existing schools close. That is, school by zip code models capture changes in enrollment for existing schools, while zip code fixed effects capture changes in enrollment for existing schools and the entry or exit of schools. Results are qualitatively similar and available upon request.

42. Within this period of study, median income is time-invariant as it is only available at the zip code level through the American Community Survey’s 5-year rolling average data.

43. There are an average of 2.7 zip codes per neighborhood school district.

44. Recall Poisson models control for the total number of students that could have enrolled in online school through the inclusion of the under 18 population as the exposure parameter. In raw differences, urban zip codes enroll more students in online education because the average number of children living in an urban zip code is higher than the number living in a rural zip code.

45. Relative to zero private schools being located in a neighborhood school district, five private schools or more is negatively associated with enrollment in virtual schools. These data come from Private School Universe Survey maintained by NCES. Results are available upon request.

46. Note that we elected not to estimate the model with Stride school by neighborhood school district fixed effects because the SEDA data only runs through the 2017–2018 school years, and are incomplete.

47. We also ran a model replacing average test achievement of students’ neighborhood schools with the change in students’ average test achievement between their most recent pre-study year and their second most recent pre-study year. The coefficient on the change in standardized test achievement is positive, but not statistically significant. Results are available upon request.

48. Note these under 18 population counts come from the American Communities Survey, which does not disaggregate data by race (and age). Instead, we use the zip codes overall racial composition to estimate the race specific under 18 population.

49. For visual clarity in the scatter plot, neighborhood school district compositions were averaged over 1% increments, however, the cubic regression is based off of the un-collapsed data.

50. The data was collapsed to the neighborhood school district level and all regressions include the same set of controls present in column (2) of Table (3).

51. While the number of students of a particular race enrolled in online schools is mechanically related to the percent of students from that race going to the neighborhood school districts, in practice Stride school enrollments are roughly 0.3% of the total neighborhood school district enrollments and likely do not strongly influence the coefficients on the racial composition of neighborhood school districts.

52. Results are available upon request.

53. Results are from a model similar to column (1) of Table 3, where the neighborhood school district percent Black, Hispanic, and Other have been replaced by percent White.

Additional information

Funding

This work was supported by the National Center for Analysis of Longitudinal Data in Education Research.