Do Online Courses Provide an Equal Educational Value Compared to In-Person Classroom Teaching? Evidence from US Survey Data using Quantile Regression
aa r X i v : . [ ec on . GN ] J u l Do Online Courses Provide an Equal Educational ValueCompared to In-Person Classroom Teaching? Evidence from USSurvey Data using Quantile Regression
Manini Ojha a , Mohammad Arshad Rahman b, ∗ a Jindal School of Government and Public Policy, O.P. Jindal Global University, Sonipat, India. b Department of Economic Sciences, Indian Institute of Technology Kanpur, India.
Abstract
Education has traditionally been classroom-oriented with a gradual growth of online courses inrecent times. However, the outbreak of the COVID-19 pandemic has dramatically acceleratedthe shift to online classes. Associated with this learning format is the question: what do peoplethink about the educational value of an online course compared to a course taken in-person in aclassroom? This paper addresses the question and presents a Bayesian quantile analysis of publicopinion using a nationally representative survey data from the United States. Our findings showthat previous participation in online courses and full-time employment status favor the educationalvalue of online courses. We also find that the older demographic and females have a greaterpropensity for online education. In contrast, highly educated individuals have a lower willingnesstowards online education vis--vis traditional classes. Besides, covariate effects show heterogeneityacross quantiles which cannot be captured using probit or logit models.JEL codes: A20, C11, C31, C35, I20, I29
Keywords:
Bayesian quantile regression, binary quantile regression, COVID-19, educationalvalue, Gibbs sampling, public opinion, Pew Research Center.
1. Introduction
Advancements in technology have resulted in a proliferation of online educational opportunitiesover the last two decades. Allen and Seaman (2016) report that the growth rate of enrolments inonline courses is expanding faster than the traditional classroom enrolments in the United States(US). Even as academic leaders remain far more positive about traditional and blended pedagogicalformats than fully online learning, the outbreak of the COVID-19 pandemic has acted as a catalyst
Acknowledgements: ∗ Please address correspondence to Mohammad Arshad Rahman, Department of Economic Sciences, Indian In-stitute of Technology, Kanpur. Room 672, Faculty Building, IIT Kanpur, Kanpur 208016, India. Phone: +91512-259-7010. Fax: +91 512-259-7500.
Email addresses: [email protected] (Manini Ojha), [email protected] (Mohammad Arshad Rahman) Roughly one in two individuals who have graduated in the last ten years have taken at least one online coursein their degree program (Parker et al., 2011).
Preprint submitted to arXiv.org July 15, 2020 n establishing online education as an immediate substitute for in-person classrooms. This paradigmshift in education has drawn considerable attention from the media, and researchers across the globe.However, any research based on the acceptance of online education during this period of disturbanceis likely to be a deviation from the natural relationship between education and technology. Instead,as institutions of higher learning integrate web-based tools into classroom instructions, we assertthat it is more important to assess the attitude towards digital education and its acceptance ina state of equilibrium prior to the onset of the pandemic. With that in mind, in this paper, weanalyse public opinion on the value of online education relative to traditional education using surveydata from the Pew Social Trends and Demographics Project conducted by the Princeton SurveyResearch Associates International in 2011.In the early 2000s, despite significant skepticism from academics and pushback from the public,several universities invested in and adopted Massive Open Online Course (MOOCs) as a teaching-learning format (Miller, 2014). Educational institutions today are compelled to rethink their ped-agogical philosophies to incorporate either hybrid or fully-online teaching-learning formats, as aconsequence of the ongoing pandemic. Students graduating in the current era have experiencedsome education using technology, either as a supplement to traditional classes, or as fully onlinecourses. Correspondingly, faculty is expected to have the willingness and the ability to engage inpedagogy that utilises technology (Miller, 2014). While this trend towards instructional technol-ogy expands, there are ambivalent perceptions about the quality of online education (Chen et al.,2013; Otter et al., 2013; Allen and Seaman, 2011). Therefore, we specifically address public opin-ion about the value of online education and the factors that influence it vis--vis traditional classes,using a Bayesian quantile analysis.Modelling public opinion on the value of online education presents a rich area for further study.Over the last few years, a sizeable body of literature on the demand and efficacy of online ed-ucation, its scope to lower educational costs, student and faculty perceptions, and its impact onstudent learning outcomes have emerged (Xu and Jaggars, 2013; Goodman et al., 2019; Otter et al.,2013; Cassens, 2010; Bettinger et al., 2017; Figlio et al., 2013; Alpert et al., 2016; Joyce et al.,2015; Krieg and Henson, 2016; Kirtman, 2009). However, much of the existing research focuseson one or two specific courses, or are limited within a selective college or university. For instance,Goodman et al. (2019) compare an online and in-person degree in Master of Science in ComputerScience offered at Georgia Tech and document a large demand for the online program with nearlyno overlap in the applicant pools. Analysing survey data from a community college in California,Cassens (2010) finds no significant differences in students performances in online and traditionalteaching methods. Contrary to this, using data from one large for-profit university, Bettinger et al.(2017) find negative effects of online courses on student academic success and progression relativeto in-person courses. On similar lines, Otter et al. (2013) find significant differences upon compar-ison of faculty and student perceptions of online courses versus traditional courses at a large publicuniversity in the south-eastern United States. As such, mixed evidence found owing to the narrowfocus of these papers often brings their external validity into question. To this end, we attempt to2ddress the educational value of online classes by utilising a nationally representative US surveydata, thereby drawing conclusions for a population at large. This is the first contribution of ourpaper to the existing literature on online education.Evidence finds the proportion of faculty who believe in the legitimacy of online education to berelatively low. In addition, the proportion of faculty who perceive online education as more timeintensive and requiring greater effort has seen a steady growth (Allen and Seaman, 2011). Contraryto this, students perceive such courses to be largely self-taught with minimal effort from the faculty(Chen et al., 2013; Otter et al., 2013). Such ambiguous views on the issue makes it imperative toinvestigate overall public opinion on the matter. That being the case, our paper also contributes toa second body of literature that points towards the differential adoption and acceptance of technol-ogy in higher education across different demographies (Cooper, 2006; Norum and Weagley, 2006;Chen and Fu, 2009; Cotten and Jelenewicz, 2006; Odell et al., 2000; Jones et al., 2009). Given thatonline education is a matter of individual selection, individual characteristics may vary drasticallyacross the utility derived from it. Traditional mean regression of the effects of covariates on the pref-erence about online classes may mask important heterogeneity in individual choices. Our study isthe first, of which we are aware, to offer new insights regarding the opinion on the educational valueof online courses across the quantiles and latent utility scale. These differential effects across thelatent utility scale may be of direct interest to policy makers and educationists as our methodologyprovides a more comprehensive picture.Utilising a nationally representative US survey data from the Pew Social Trends and Demo-graphics Project conducted in the year 2011, we examine individual responses about the educationalvalue derived from online classes in comparison to in-person classroom . Our paper presents an em-pirical application of binary quantile regression to a model of educational decision. More specifically,we model the latent utility differential between online classes and traditional classes. This may beinterpreted as a propensity or a willingness index, where higher propensity towards online educationare characterised by large positive values and vice versa. The results are compelling and ought toserve as a guide for future research. We find that an older demographic, individuals with full-timeemployment, individuals with previous online experience, and females display a propensity towardsonline education. Interestingly, our findings highlight that highly educated respondents have lowerwillingness for online education. We also note some amount of regional differences in the propen-sity to value online classes. All these covariates show considerable differences in covariate effectsat different quantiles. Lastly, we find no convincing evidence of race or income having an effect onthe propensity for online education.The remainder of the paper is organised as follows. Section 2 outlines the data used for ouranalysis including a descriptive summary. This is followed by Section 3 that outlines a modelof quantile regression for binary outcomes and presents a Markov chain Monte Carlo (MCMC)algorithm for its estimation. Next, we present the results of our binary quantile regressions inSection 4. Section 5 presents the concluding remarks.3 . Data
The study utilises a nationally representative US survey data from the Pew Social Trends andDemographics Project, conducted over telephone between March 15 −
29, 2011, by the PrincetonSurvey Research Associates International. The survey was primarily for higher education andhousing and contains information on 2,142 adults living in the continental US. We consider asubset of variables from this survey and upon removing missing observations from our variables ofinterest (see Table 1), we are left with 1,591 observations available for the analysis. The dependentvariable is the response to the question: “In general, do you think a course taken only onlineprovides an equal educational value compared with a course taken in person in a classroom, ornot?” . Responses are recorded either as “Yes” , “No” , or “Don’t know/Refused” . We ignore thelast response category. Of the 1,591 respondents, 505 (31.74%) respondents agree that a coursetaken online provides an equal educational value compared to in-person classroom teaching, whilethe remaining 1086 respondents (68.26%) do not agree and thus believe that online courses havelesser educational value. The survey also consists of information on an array of other variables,some of which we utilise as covariates (independent variable) in our analysis. A description of thecovariates and the response variable, along with the main characteristic of the data is presented inTable 1.In our sample, a typical individual is 44 years of age with a family income of 63 thousand USdollars. The survey recorded income as belonging to one of the following nine income categories: < k , 10 k − k , 20 k − k , 30 k − k , 40 k − k , 50 k − k , 75 k − k , 100 k − k and > k ,where k denotes a thousand dollars. We use the mid-point of each income category to represent theincome variable, where $5,000 and $1,75,000 are used as the mid-point for the first and last incomecategories, respectively. With respect to online learning, we have a little more than one-fifth of thesample who have previously taken an online course for academic credit. A sizeable proportion ofthe sample, therefore, have had prior exposure to online learning. Individuals who are aged lessthan 65 and currently enrolled in school comprise a little less than one-fifth of the sample. Here,enrolment in school implies that the respondent is either attending high school, technical school,trade or vocational school, is a college undergraduate or in graduate school.The sample has more females (51.23%) than males (48.77%), but both genders have approxi-mately equal representation. Education has been classified into four categories with ‘High School(HS) and below’ forming the largest category (33.44%) followed by ‘Below Bachelors’ (30.30%).The smallest two educational categories are ‘Bachelors’ (22.38%) and ‘Post-Bachelors’ (13.89%).So, approximately two-thirds of the sample have less than bachelors education. With respect toemployment status, about a little less than two-thirds (i.e., 62.66%) are either employed full-time orpart-time, while the remaining percentage are either unemployed, students or retired individuals.Racial classification shows that more than two-thirds are White (71.09%), followed by African-Americans (16.15%) and all other races (12.76%). In terms of rural-urban classification, most ofthe sampled individuals live in the suburban areas (47.77%), followed by the urban areas (39.35%).The lowest proportion lives in the rural areas (12.88%). Regional classification as defined by the4 able 1: Descriptive summary of the variables. variable description mean std Age/100 Age (in years) divided by 100 0.44 0.18Income/100,000 Mid-point of income category (in US dollars) divided by100,000 0.63 0.48 count percent
Online Course Indicates that the respondent has previously taken an onlinecourse for academic credit 352 22.12(Age < ∗ Enroll Indicates that the respondent is of age below 65 and currentlyenrolled in school 291 18.29Female Indicator variable for female gender 815 51.23Post-Bachelors Respondent’s highest qualification is Masters, Professionalor Doctorate 221 13.89Bachelors Respondent’s highest qualification is Bachelors 356 22.38Below Bachelors Respondent holds a 2-year associate degree, went to somecollege with no degree, or attended technical, trade or voca-tional school after high school 482 30.30HS and below Respondent is a high school (HS) graduate or below 532 33.44Full-time Indicator for full-time employment 757 47.58Part-time Indicator for part-time employment 240 15.08Unemployed Indicator for either unemployed, student or retired 594 37.34White Indicator for a White respondent 1131 71.09African-American Indicator for an African-American respondent 257 16.15Other Races Indicator for a respondent who is either an Asian, Asian-American or belongs to some other race 203 12.76Urban Lives in an urban region 626 39.35Suburban Lives in a suburban region 760 47.77Rural Lives in a rural region 205 12.88Northeast Lives in the Northeast 220 13.83West Lives in the West 362 22.75South Lives in the South 724 45.51Midwest Lives in the Midwest 285 17.91Opinion Respondent answered ‘Yes’ to our question of interest 505 31.74Respondent answered ‘No’ to our question of interest 1086 68.26
US Census Bureau shows that the largest percentage of the sample live in the South (47.77%).This is followed by the West (22.75%), Midwest (17.91%), and Northeast (13.83%) regions.Before we formally delve into modelling the dependent variable (i.e., public opinion on educa-tional value of online learning relative to in-person classroom teaching), we explore its relationshipwith some selected independent variables or covariates (see Parker et al., 2011, for a report ondata summary). To explore this association, we present a stacked bar graph in Figure 1 with fourpanels, each portraying the relationship between the dependent variable and a single covariate.Each bar within a panel corresponds to a category of the covariate and displays the percentage ofobservations that says ‘Yes’ and ‘No’ to our question of interest. For example, the upper (lower)bar in Panel 1 shows that for people aged greater than (less than equal to) 30, 32.8% (29.5%)of the sample agree that online courses have the same educational value as in-person classroom5 ge Category29.532.8 70.567.20 20 40 60 80 100Age<=30Age>30 YesNo Online Course (OC)38.129.9 61.970.10 20 40 60 80 100No Previous OCPrevious OC YesNoEducation2425.335.935.5 7674.764.164.50 20 40 60 80 100HS and belowBelow BachelorsBachelorsPost-Bachelors Race3730.930 6369.1700 20 40 60 80 100Other RacesWhiteAfrican-American
Figure 1: Stacked bar graph displaying the percentage of observations corresponding to the two categories of publicopinion (Yes and No) for each category of some selected covariates. teaching, while the remaining 67.2% (70.5%) do not agree. The other three panels of Figure 1 canbe interpreted analogously.We see from the first panel of Figure 1 that the percentage of sample who says ‘Yes’ (and thus‘No’) is approximately equal amongst the younger (Age < = 30) and older (Age >
30) population.From Panel 2 we note that, amongst the sample who have taken an online course for academiccredit, a higher percentage (38.1%) says ‘Yes’ compared to those (at 29.9%) who have no previousonline learning experience. Panel 3 suggests that the highly educated group (Bachelors and Post-Bachelors) are less likely to agree (1 in every 4 individual) about the equal educational value ofonline learning and classroom teaching, as compared to the lower educated group (where 1 in every3 agrees). Similarly, the racial classification of response shows that the African-Americans are morelikely to agree (37%) as compared to White (30.9%) and Other Races (30%).The discussion involving the stacked bar graph only presents an association between the publicopinion on the educational value of online learning relative to in-person classroom teaching and onecovariate at a time, namely, age, previous participation in online course, education, and race. Such6n association can be captured by regressing the dependent variable on a chosen covariate/regressor.However, inference based on such an analysis is unlikely to present the true relationship becausethere may be other determinants of the dependent variable which are correlated with the chosencovariate. If ignored, this may lead to estimation bias and incorrect inferences. For instance, letus suppose we are interested in the relationship between public opinion on online learning relativeto in-person classroom teaching and the age category. To this end, we regress the dependentvariable on age category. However, this relationship is likely to change when we control for previousparticipation in online course owing to the correlation between previous participation in onlinecourse and age. To net out such effects and understand the actual impact of a covariate on thedependent variable, we next turn to some formal econometric modelling.
3. Quantile Regression for Binary Outcomes
Quantile regression, as introduced by Koenker and Bassett (1978), looks at quantiles of the (con-tinuous) response variable conditional on the covariates and thus provides, amongst other things, acomprehensive picture (as compared to traditional mean regression) of the effect of covariates on theresponse variable. Estimation involves minimizing the quantile loss function using linear program-ming techniques (Koenker, 2005). Interestingly, the quantile loss function appears in the exponentof the asymmetric Laplace (AL) distribution (Yu and Zhang, 2005), which makes minimization ofthe quantile loss function equivalent to maximization of the AL likelihood. This characteristicallowed Yu and Moyeed (2001) to construct a working likelihood and propose Bayesian quantileregression. However, when outcomes are discrete (e.g., binary, ordinal) estimation becomes chal-lenging because quantiles for discrete outcomes are not readily defined. With discrete outcomes,the concern is to model the latent utility differential (say, between making a choice versus notmaking it or occurrence of an event versus its non-occurrence) facilitated through the introductionof a latent variable (Albert and Chib, 1993; Greenberg, 2012; Rahman, 2016). This applies to bothmean and quantile regressions and is useful for estimation and inference.Quantile regression for binary outcomes (or binary quantile regression ) was introduced inKordas (2006) and the Bayesian framework was presented in Benoit and Poel (2012). The binaryquantile model can be conveniently expressed using the latent variable z i as follows, z i = x ′ i β p + ǫ i , ∀ i = 1 , · · · , n,y i = ( z i > , . (1)where x i is a k × β p is a k × p -th quantile Binary quantile regression is a special of ordinal quantile regression considered in Rahman (2016) and canbe linked to the random utility theory in economics (Train, 2009; Jeliazkov and Rahman, 2012). For otherdevelopments on Bayesian quantile regression with discrete outcomes, please see Alhamzawi and Ali (2018a),Alhamzawi and Ali (2018b), Ghasemzadeh et al. (2018b), Ghasemzadeh et al. (2018a), Rahman and Vossmeyer(2019), Rahman and Karnawat (2019), Bresson et al. (2020). p is dropped for notational convenience), ǫ i follows an AL distributioni.e., ǫ i ∼ AL (0 , , p ), and n denotes the number of observations. In our study, the latent variable z i can be interpreted as the latent utility differential between online learning relative to in-personclassroom learning. Whenever the observed response y i = 1 (i.e., the respondent answers ‘Yes’ toour question of interest), propensity to online learning is likely to be high and z i takes a value inthe positive part of the real line. Similarly, when y i = 0 (i.e., the respondent answers ‘No’ to ourquestion of interest), the propensity to online learning is low and z i takes a value in the negativepart of the real line.We can form a working likelihood from equation (1) and directly use it to construct the posteriordistribution, but this is not convenient for MCMC sampling. A preferred alternative is to employthe normal-exponential mixture of the AL distribution (Kozumi and Kobayashi, 2011). In thisformulation, ǫ i = θw i + τ √ w i u i , and the binary quantile model is re-expressed as, z i = x ′ i β + + θw i + τ √ w i u i , ∀ i = 1 , · · · , n,y i = ( z i > , . (2)where θ = (1 − p ) p (1 − p ) and τ = q p (1 − p ) are constants, and w i ∼ E (1) is independently distributed of u i ∼ N (0 , E and N denote exponential and normal distributions, respectively. Itis clear from formulation (2) that the latent variable z i | β, w i ∼ N ( x ′ i β + θw i , τ w i ), thus allowingaccess to the properties of normal distribution.By the Bayes’ theorem, the complete data likelihood from equation (2) is combined with anormal prior distribution on β (i.e., β ∼ N ( β , B )) to form the complete data posterior. This Algorithm 1 (MCMC Algorithm for Binary Quantile Regression)
1. Sample β | z, w ∼ N ( ˜ β, ˜ B ), where,˜ B − = (cid:18) P ni =1 x i x ′ i τ w i + B − (cid:19) and ˜ β = ˜ B (cid:18) P ni =1 x i ( z i − θw i ) τ w i + B − β (cid:19) .2. Sample w i | β, z i ∼ GIG (0 . , ˜ λ i , ˜ η ), for i = 1 , · · · , n , where,˜ λ i = (cid:16) z i − x ′ i βτ (cid:17) and ˜ η = (cid:16) θ τ + 2 (cid:17) .3. Sample the latent variable z | y, β, w for all values of i = 1 , · · · , n from an univariate truncated normal(TN) distribution as follows, z i | y, β, w ∼ T N ( −∞ , (cid:18) x ′ i β + θw i , τ w i (cid:19) if y i = 0 ,T N (0 , ∞ ) (cid:18) x ′ i β + θw i , τ w i (cid:19) if y i = 1 . π ( z, β, w | y ) ∝ (cid:26) n Y i =1 (cid:2) I ( z i > I ( y i = 1) + I ( z i ≤ I ( y i = 0) (cid:3) N ( z i | x ′ i β + θw i , τ w i ) × E ( w i | (cid:27) N ( β , B ) . (3)The full conditional posterior densities for ( z, β, w ) can be derived from equation (3) and themodel can be estimated using the Gibbs sampler (Geman and Geman, 1984) − a well knownMCMC technique − presented in Algorithm 1. The sampling algorithm is straightforward andinvolves sampling β conditional on ( z, w ) from an updated normal distribution. The latent weight w conditional on ( β, z ) is sampled from a Generalized Inverse Gaussian (GIG) distribution (Devroye,2014). Finally, the latent variable z conditional on ( y, β, w ) is sampled from a truncated normaldistribution (Robert, 1995).
4. Results
Table 2 presents the posterior means, and standard deviations of the parameters from theBayesian estimation of probit model (Albert and Chib, 1993), and the binary quantile regressionat the 10th, 25th, 50th, 75th and 90th quantiles. We assume the following diffuse prior distribution: β ∼ N (0 k , ∗ I k ), where N and I denote a multivariate normal distribution and an identitymatrix of dimension k , respectively. The results are based on 20,000 MCMC iterations after aburn-in of 5,000 iterations. The inefficiency factors were calculated using the batch-means method(Greenberg, 2012; Chib, 2013). For the five chosen quantiles, they lie in the range (6 . , . . , . . , . . , . . , . − . , . , . − . , . , . − . , . , . − . , . , . − . , . , . − . , . , . and are calculated marginally of the remaining covariatesand the parameters (Chib and Jeliazkov, 2006; Jeliazkov et al., 2008; Jeliazkov and Rahman, 2012;Jeliazkov and Vossmeyer, 2018; Rahman and Vossmeyer, 2019; Bresson et al., 2020).While many of the results are in line with extant literature, our results provide some usefulinsights into the differences across quantiles. As previously noted, we are modelling the latent The covariate effects for previous online course, full-time employment, post-bachelors, bachelors, Northeastand South are calculated on the respective sub-samples and are a discrete change compared to their base groupsrespectively. The covariate effect for female is calculated on the full sample and is a discrete change compared tomale. able 2: Posterior mean ( mean ) and standard deviation ( std ) of the parameters from the Bayesian estimation ofprobit regression and binary quantile regression. quantileprobit 10th 25th 50th 75th 90thmean std mean std mean std mean std mean std mean std Intercept − .
80 0 . − .
02 2 . − .
03 0 . − .
69 0 . − .
23 0 .
46 1 .
58 0 . .
58 0 .
23 5 .
39 2 .
44 2 .
24 0 .
92 1 .
17 0 .
49 1 .
34 0 .
52 3 .
35 1 . .
37 0 .
26 4 .
20 2 .
73 1 .
70 1 .
13 0 .
92 0 .
58 0 .
87 0 .
59 1 .
02 1 . − .
28 0 . − .
18 1 . − .
30 0 . − .
69 0 . − .
63 0 . − .
86 0 . .
31 0 .
09 2 .
80 0 .
83 1 .
14 0 .
35 0 .
64 0 .
19 0 .
75 0 .
21 1 .
74 0 . < ∗ Enroll 0 .
02 0 .
11 0 .
25 1 .
13 0 .
12 0 .
43 0 .
07 0 .
22 0 .
00 0 .
24 0 .
00 0 . .
14 0 .
07 1 .
72 0 .
71 0 .
71 0 .
28 0 .
36 0 .
15 0 .
27 0 .
15 0 .
53 0 . − .
45 0 . − .
76 1 . − .
95 0 . − .
02 0 . − .
98 0 . − .
09 0 . − .
40 0 . − .
19 1 . − .
69 0 . − .
90 0 . − .
87 0 . − .
81 0 . − .
09 0 . − .
03 0 . − .
44 0 . − .
26 0 . − .
17 0 . − .
27 0 . .
27 0 .
08 2 .
76 0 .
94 1 .
13 0 .
37 0 .
58 0 .
19 0 .
58 0 .
19 1 .
28 0 . .
17 0 .
11 1 .
67 1 .
15 0 .
68 0 .
49 0 .
35 0 .
24 0 .
41 0 .
24 0 .
93 0 . − .
01 0 .
11 0 .
08 1 .
12 0 .
06 0 .
48 0 .
03 0 . − .
05 0 . − .
18 0 . .
21 0 .
13 2 .
05 1 .
26 0 .
93 0 .
55 0 .
48 0 .
28 0 .
43 0 .
31 0 .
87 0 . − .
10 0 . − .
00 1 . − .
36 0 . − .
23 0 . − .
25 0 . − .
35 0 . .
06 0 .
11 0 .
58 1 .
04 0 .
32 0 .
44 0 .
11 0 .
23 0 .
12 0 .
24 0 .
43 0 . − .
29 0 . − .
09 1 . − .
23 0 . − .
63 0 . − .
63 0 . − .
53 0 . − .
06 0 . − .
44 1 . − .
18 0 . − .
10 0 . − .
18 0 . − .
58 0 . − .
22 0 . − .
34 0 . − .
93 0 . − .
48 0 . − .
46 0 . − .
14 0 . utility differential between online and in-person classes. The results therefore, may be interpretedas a utility index of online education. Large positive (negative) values of this index signify high(low) propensity to favour online classes, and values around zero would indicate relative indifferencebetween the two alternatives. A bird’s-eye view of the results shows that age, past online experience,full time employment and gender have a positive effect on the propensity to favour online education.Higher level of educational degree, on the other hand, has a negative effect on the willingness towardsonline education. We also note some amount of regional variation in the propensity to favour onlineclasses. To better understand the results, we focus on each variable separately.The coefficient for age is positive across all quantiles. This is not surprising as online coursesinvariably attract an older demographic (Crain and Ragan, 2017). Goodman et al. (2019) findsimilar results highlighting that on average, the online applicants were 34 years of age comparedto 24 years for inperson applicants in their study. Besides, our result is perhaps indicative of mid-career professionals favouring online classes since several online courses cater to those active inthe workforce, requiring professional development or retaining by employers (Kizilcec et al., 2019).From the calculated covariate effects in Table 3, we see that the covariate effect of age is between1.7 to 2.2 percentage points across the quantiles. Stronger effects are visible in the upper part ofthe latent index. 10 able 3: Covariate Effect. quantileprobit . . . . . . . . . . . . . . . . . . − . − . − . − . − . − . − . − . − . − . − . − . . . . . . . − . − . − . − . − . − . − . − . − . − . − . − . Next, we note that educational value of online classes is favoured positively by individualswith a full-time employment status compared to base category (unemployed, students or retiredindividuals). The coefficient for full-time employment, compared to the base category, is statisticallypositive across the quantiles. The coefficients for part-time employment are positive but the effectsare not statistically different, implying that regardless of the latent utility for online education,part-time employment does not impact the decision. Our result for full-time employment is inagreement with the evidence that demand for online education is high for employed mid-careerprofessionals, or those who seek professional development (Simmons, 2014; Kizilcec et al., 2019). Itappears to be commonplace for employers to sponsor their employees enrolment into online coursesfor training purposes as observed by Goodman et al. (2019) and Deming et al. (2015). In fact,from Table 3, the covariate effect of full-time employment increases the willingness for online classby 7.8 percentage points in the 10th quantile and consistently increases across quantiles to about 8.7percentage points in the 90th quantile. For individuals who are in the lower part of the latent index,employment impacts their valuation for online education less than those in the upper quantile.Turning to previous exposure to digital learning, we find that individuals propensity of valuingonline education is higher for those who have had past participation in online classes for academiccredit than those who have not. We find positive effects of previous exposure to online educationacross the quantiles of the utility scale. Astani et al. (2010); Williams (2006); Goode (2010) showsimilar evidence that previous online experience changes the perceptions about an online learningenvironment. A positive stance towards online education is therefore undeniably linked to previousexposure and use of technology. The covariate effect of previous online class ranges between 9.5 to11.1 percentage points across the quantiles (See Table 3). Although the effect somewhat plateausat the 75th quantile, our findings suggest that past online experience increases the probability ofvaluing online classes most for those with higher utility for online education. The National Post Secondary Student Aid Study (NPSAS) for 2011-12 that includes a nationally representativecross-section of institutions and students shows that online students are older and more likely to be working full-timewhile enrolled (Deming et al. (2015).
11e also find that females are more in favour of online education relative to males. This findingis in consonance with Fortson et al. (2007), who propose that female college students are morelikely to go online for communicative and educational purposes while male college students aremore likely to use the internet as a source of entertainment. Perhaps the noted gender differentialcould also be a result of differences in past usage of internet. Furthermore, online education allowsfor flexible schedules that individuals can customise around their family and job constraints moreeasily (Goodman et al., 2019). This greater flexibility in schedule likely implies greater willingnessfor online education for females. The covariate effect of females, displayed in Table 3, shows thatbeing female increases the probability of valuing online education by 5.3 to 3.5 percentage pointsfrom 10th to 90th quantiles respectively. Strongest effect of female is found in the 25th quantileand the effect reduces at the 90th quantile. At higher utility, females are more similar to malesthan at lower utility.Next, we find that the coefficient for different levels of education are consistently negative, rela-tive to the base category (HS and below), across the quantiles. In each quantile, the post-bachelorscategory shows a large negative propensity for online classes vis--vis traditional learning. The ef-fects are also negative for those with a bachelors degree compared to those with HS education orbelow. While the effects are negative for below-bachelors degree, they are not statistically differentin comparison to the base category. This is useful in understanding the differences in preferencesbetween individuals with different educational qualifications. Highly educated respondents reportdiminished value of online classes in comparison to those with a HS degree or below. This pointsto some degree of stigma towards online education as the level of educational qualification rises(Kizilcec et al., 2019). The decrease in utility from online classes is likely driven by greater inten-sity of learning and teaching at graduate or post graduate levels. Students perceive better learningfrom face-to-face interactions, and visualising materials. The self-regulatory nature of physicalclassroom teaching perhaps enables students to track their understanding of the course. Our re-sult finds support in Chen et al. (2013) who note more favourable student outcomes for traditionalclassrooms versus an online mode for advanced accounting courses, and Otter et al. (2013) whonote that students believe that they must do the teaching and learning on their own in onlinecourses in contrast to what they feel about the time and effort from faculty for traditional courses.This perhaps reduces the value they attach to online education at higher degree levels. Accordingto ONeill and Sai (2014), traditional classes also allow for better relationship and lines of com-munication with the instructor. Other studies have shown that students face difficulty in keepingup motivation in online classes. This is likely to become more prominent at higher levels of ed-ucation. Findings documented in Anstine and Skidmore (2005) and Fendler et al. (2011), suggestthat students perform worse in online courses compared to face-to-face classes at undergraduateand graduate levels also support our results of lower willingness for online education by them. The proportion of individuals with previous experiences of online education is higher for females in our sample,with 55% of females having taken an online course for credit before. Individuals prior experience with online courses and their performance likely play a role in determining the value . Regional differences inthe popularity of online courses are likely driven by use of technology, student population, coursedesign and support provided by the universities, as well as the philosophies of universities in theregion. Educational value of online courses are thus considered higher in regions where onlineeducation is more popular. Xu and Jaggars (2013) highlight the importance of the institutionalstate in determining the cultural capital around technology. Allen and Seaman (2007) also suggestthat the Southern states represented over one-third of total online enrolments in 2005-06 and theproportion of Southern institutions with fully online programs is steadily rising. Our covariateeffect calculations indicate that in the Northeast, the probability of favouring online classes reducesby 9.1 to 10.2 percentage points relative to the Midwest, across the quantiles. The highest negativeeffect is found for those in the 90th quantile. The propensity to value online education is reducesby 6.9 to 8 percentage points in the Southern regions compared to the Midwest.We also examine the effect of race and find no noteworthy racial differences in the willingnesstowards online education relative to in-person education. Specifically, the coefficients for White,relative to the base category (Other Races), are statistically not different from zero. Similarly,the coefficients for African-Americans across the quantiles are statistically equivalent to zero. Thislikely indicates that, after controlling for different educational levels and previous exposure to onlinelearning, individuals across racial groups seem to hold similar attitudes about the online classesas an educational tool. Our results fall in line with (Cotten and Jelenewicz, 2006; Odell et al.,2000; Jones et al., 2009; Bowen et al., 2014), who note that the digital divide upheld by race maybe narrowing, and in some cases negligible among college campuses in the US. Contrary to this,Figlio et al. (2013), find negative outcomes for Hispanic students. they attach to online education. As per Parker et al. (2011), roughly 39% of those who have taken an online coursebefore respond favourably to online educational value whereas about 27% of those with no prior online educationfavourably value online classes. Use of technology in the Mid-west and the South is higher compared to the East. In fact, the East coasters arefound to lag behind the rest of the country in some aspects of technology adoption as per Parker et al. (2011). Miller (2014) states that universities in Arizona are considered to be early adopters of online teaching techniques,in fact preferring faculty with experience in technology. Regions with students having high levels of technological proficiency are more likely to take courses whichintegrate technology, major in technology-rich disciplines, and pursue technology-rich careers (Xu and Jaggars, 2013).
5. Conclusion
Technological advancements and the rising cost of higher education have rendered online edu-cation as an attractive substitute or a complementary technique for teaching and learning. Withthe online enrolment growth rate in the US at 9.3 percent, over 6.7 million students were estimatedto have taken at least one online course in 2012 (Allen and Seaman, 2013). Considering this trend,in this paper, we examine public opinion about the value of online learning methods in comparisonto in-person education across United States. Evaluating public opinion can not only serve as aguide for policy, but also help design the transformative shift away from physical classrooms as thedominant paradigm of teaching and learning. Some scholars argue that public policies should infact be guided by public opinion, so that mass opinion and democracy is upheld (Monroe, 1998;Paletz et al., 2013).Extant literature points to uncertainty regarding the quality and rigour of online education.While there is some degree of adaptation to specific online courses offered by traditional univer-sities as blended learning, reservations about fully online degree programs remain. As such, weassert that the approach provided in this paper leads to a richer view of how the demographic co-variates may influence public opinion about the educational value of online classes, thereby, betterinforming future educational policies. We find important effects of employment status, previousonline experience, age and gender on the propensity towards online education across the latentutility scale. We also note that willingness towards online classes versus in-person classes is lowerfor highly educated individuals. While interesting regional variations exist, we find no evidence ofrace and income on the propensity for online education.We conclude with three main questions for future work. First, creating an in-depth, systematicsupport for both faculty and students, in transitioning from traditional to online teaching-learningplatforms, is not an inexpensive venture. With considerable fixed costs incurred in training, coursecreations and delivery methods for online education (Ginn and Hammond, 2012; Xu and Jaggars,2013), what would be the incentives to switch back to in-person classrooms or blended formats in apost-pandemic general equilibrium? Second, evidence suggests that online education democratisesand improves access to education (Goodman et al., 2019). Given this, it may be interesting toexamine the trade-off between a perceived decrease in outcomes and efficacy of online educationversus the increase in the exposure of education to previously inaccessible population from a policyperspective. Finally, it is well established that upward mobility in teaching colleges are largely14nfluenced by student feedback and evaluations (Chen and Hoshower, 2003; McClain et al., 2018;Krautmann and Sander, 1999). If we consider the current health landscape across the globe as aperiod of deviation from the true nature of dynamics between education and technology, one oughtto think about how student feedback during this phase of aberration, will contribute to upwardmobility of faculty. 15 eferences
Albert, J. H. and Chib, S. (1993), “Bayesian Analysis of Binary and Polychotomous ResponseData,”
Journal of the American Statistical Association , 88, 669–679.Alhamzawi, R. and Ali, H. T. M. (2018a), “Bayesian Quantile Regression for Ordinal LongitudinalData,”
Journal of Applied Statistics , 45, 815–828.Alhamzawi, R. and Ali, H. T. M. (2018b), “Bayesian Single-Index Quantile Regression for OrdinalData,”
Communications in Statistics − Simulation and Computation , pp. 1–15.Allen, I. E. and Seaman, J. (2007), “Making the Grade: Online Education in the United States,2006,”
Sloan Consortium , Feb 2007. Available: https://files.eric.ed.gov/fulltext/ED530101.pdf[Last accessed: 11 July 2020].Allen, I. E. and Seaman, J. (2011), “Going the Distance: Online Education in the United States,2011.”
Sloan Consortium , Nov 2011. Available: https://files.eric.ed.gov/fulltext/ED529948.pdf[Last accessed: 11 July 2020].Allen, I. E. and Seaman, J. (2013), “Changing Course: Ten Years of TrackingOnline Education in the United States,”
Sloan Consortium , Jan 2013. Available:https://files.eric.ed.gov/fulltext/ED541571.pdf [Last accessed: 11 July 2020].Allen, I. E. and Seaman, J. (2016), “Online Report Card: Tracking Online Edu-cation in the United States,”
Babson Survey Research Group , Feb 2016. Available:https://files.eric.ed.gov/fulltext/ED572777.pdf [Last accessed: 11 July 2020].Alpert, W. T., Couch, K. A., and Harmon, O. R. (2016), “A Randomized Assessment of OnlineLearning,”
American Economic Review , 106, 378–82.Anstine, J. and Skidmore, M. (2005), “A Small Sample Study of Traditional and Online Courseswith Sample Selection Adjustment,”
The Journal of Economic Education , 36, 107–127.Astani, M., Ready, K. J., and Duplaga, E. A. (2010), “Online Course Experience Matters: Inves-tigating Students Perceptions of Online Learning,”
Issues in Information Systems , 11, 14–21.Benoit, D. F. and Poel, D. V. D. (2012), “Binary Quantile Regression: A Bayesian Approach basedon the Asymmetric Laplace Distribution,”
Journal of Applied Econometrics , 27, 1174–1188.Bettinger, E. P., Fox, L., Loeb, S., and Taylor, E. S. (2017), “Virtual Classrooms: How OnlineCollege Courses Affect Student Success,”
American Economic Review , 107, 2855–2875.Bowen, W. G., Chingos, M. M., Lack, K. A., and Nygren, T. I. (2014), “Interactive LearningOnline at Public Universities: Evidence from a Six-campus Randomized Trial,”
Journal of PolicyAnalysis and Management , 33, 94–111. 16resson, G., Lacroix, G., and Rahman, M. A. (2020), “Bayesian Panel Quantile Regression forBinary Outcomes with Correlated Random Effects: An Application on Crime Recidivism inCanada,”
Empirical Economics , (forthcoming).Cassens, T. S. (2010), “Comparing the Effectiveness of Online and Face-to-Face Classes amongCalifornia Community College Students,” Ph.D. thesis, University of Southern California.Chen, C. C., Jones, K. T., and Moreland, K. A. (2013), “Online Accounting Education versusIn-class Delivery: Does Course Level Matter?”
Issues in Accounting Education , 28, 1–16.Chen, S.-Y. and Fu, Y.-C. (2009), “Internet Use and Academic Achievement: Gender Differencesin Early Adolescence,”
Adolescence , 44, 797–812.Chen, Y. and Hoshower, L. B. (2003), “Student Evaluation of Teaching Effectiveness: An Assess-ment of Student Perception and Motivation,”
Assessment & Evaluation in Higher Education , 28,71–88.Chib, S. (2013), “Introduction to Simulation and MCMC Methods,” in
The Oxford Handbook ofBayesian Econometrics , eds. J. Geweke, G. Koop, and H. V. Dijk, pp. 183–218, Oxford UniversityPress, Oxford.Chib, S. and Jeliazkov, I. (2006), “Inference in Semiparametric Dynamic Models for Binary Longi-tudinal Data,”
Journal of the American Statistical Association , 101, 685–700.Cooper, J. (2006), “The Digital Divide: The Special case of Gender,”
Journal of Computer AssistedLearning , 22, 320–334.Cotten, S. R. and Jelenewicz, S. M. (2006), “A Disappearing Digital Divide among College Stu-dents? Peeling away the Layers of the Digital Divide,”
Social Science Computer Review , 24,497–506.Crain, S. J. and Ragan, K. P. (2017), “Online Versus Face-to-Face Course Learning Effectiveness:Measured Outcomes for Intermediate Financial Management,”
Journal of Financial Education ,43, 243–261.Deming, D. J., Goldin, C., Katz, L. F., and Yuchtman, N. (2015), “Can Online Learning bend theHigher Education Cost Curve?”
American Economic Review , 105, 496–501.Devroye, L. (2014), “Random Variate Generation for the Generalized Inverse Gaussian Distribu-tion,”
Statistics and Computing , 24, 239–246.Fendler, R. J., Ruff, C., and Shrikhande, M. (2011), “Online versus In-class Teaching: LearningLevels Explain Student Performance,”
Journal of Financial Education , 37, 45–63.Figlio, D., Rush, M., and Yin, L. (2013), “Is it Live or is it Internet? Experimental Estimates of theEffects of Online Instruction on Student Learning,”
Journal of Labor Economics , 31, 763–784.17ortson, B. L., Scotti, J. R., Chen, Y.-C., Malone, J., and Del Ben, K. S. (2007), “Internet Use,Abuse, and Dependence Among Students at a Southeastern Regional University,”
Journal ofAmerican College Health , 56, 137–144.Geman, S. and Geman, D. (1984), “Stochastic Relaxation, Gibbs Distributions, and the BayesianRestoration of Images,”
IEEE Transactions on Pattern Analysis and Machine Intelligence , 6,721–741.Ghasemzadeh, S., Ganjali, M., and Baghfalaki, T. (2018a), “Bayesian Quantile Regression for Ana-lyzing Ordinal Longitudinal Responses in the Presence of Non-ignorable Missingness,”
METRON ,76, 321–348.Ghasemzadeh, S., Ganjali, M., and Baghfalaki, T. (2018b), “Bayesian Quantile Regression forJoint Modeling of Longitudinal Mixed Ordinal Continuous Data,”
Communications in Statistics − Simulation and Computation , pp. 1–21.Ginn, M. H. and Hammond, A. (2012), “Online Education in Public Affairs: Current State andEmerging Issues,”
Online Journal of Distance Learning Administration , 18, 247–270.Goode, J. (2010), “Mind the Gap: The Digital Dimension of College Access,”
The Journal ofHigher Education , 81, 583–618.Goodman, J., Melkers, J., and Pallais, A. (2019), “Can Online Delivery Increase Access to Educa-tion?”
Journal of Labor Economics , 37, 1–34.Greenberg, E. (2012),
Introduction to Bayesian Econometrics , 2nd Edition, Cambridge UniversityPress, New York.Horrigan, J. B. and Rainie, L. (2006), “The Internet’s growing Role inLife’s Major Moments,”
Pew Internet & American Life Project, Pew Re-search Center
Mathematical Modeling with Multidisciplinary Applications , ed. X. S.Yang, pp. 123–150, John Wiley & Sons Inc., New Jersey.Jeliazkov, I. and Vossmeyer, A. (2018), “The Impact of Estimation Uncertainty on Covariate Effectsin Nonlinear Models,”
Statistical Papers , 59, 1031–1042.Jeliazkov, I., Graves, J., and Kutzbach, M. (2008), “Fitting and Comparison of Models for Multi-variate Ordinal Outcomes,”
Advances in Econometrics: Bayesian Econometrics , 23, 115–156.18ones, S., Johnson-Yale, C., Millermaier, S., and P´erez, F. S. (2009), “US College Students InternetUse: Race, Gender and Digital Divides,”
Journal of Computer-Mediated Communication , 14,244–264.Joyce, T., Crockett, S., Jaeger, D. A., Altindag, O., and O’Connell, S. D. (2015), “Does ClassroomTime Matter?”
Economics of Education Review , 46, 64–77.Kirtman, L. (2009), “Online versus In-class Courses: An Examination of Differences in LearningOutcomes,”
Issues in Teacher Education , 18, 103–116.Kizilcec, R., Davis, D., and Wang, E. (2019), “Online Degree Stigma and Stereotypes: A newInstrument and Implications for Diversity in Higher Education,”
Available at SSRN 3339768 .Koenker, R. (2005),
Quantile Regression , Cambridge University Press, Cambridge.Koenker, R. and Bassett, G. (1978), “Regression Quantiles,”
Econometrica , 46, 33–50.Kordas, G. (2006), “Smoothed Binary Regression Quantiles,”
Journal of Applied Econometrics , 21,387–407.Kozumi, H. and Kobayashi, G. (2011), “Gibbs Sampling Methods for Bayesian Quantile Regres-sion,”
Journal of Statistical Computation and Simulation , 81, 1565–1578.Krautmann, A. C. and Sander, W. (1999), “Grades and Student Evaluations of Teachers,”
Eco-nomics of Education Review , 18, 59–63.Krieg, J. M. and Henson, S. E. (2016), “The Educational Impact of Online Learning: How doUniversity Students perform in Subsequent Courses?”
Education Finance and Policy , 11, 426–448.McClain, L., Gulbis, A., and Hays, D. (2018), “Honesty on Student Evaluations of Teaching:Effectiveness, Purpose, and Timing Matter!”
Assessment & Evaluation in Higher Education , 43,369–385.Miller, M. D. (2014),
Minds online , Harvard University Press, Cambridge.Monroe, A. D. (1998), “Public Opinion and Public Policy, 1980-1993,”
Public Opinion Quarterly ,62, 6–28.Norum, P. S. and Weagley, R. O. (2006), “College Students, Internet Use, and Protection fromOnline Identity Theft,”
Journal of Educational Technology Systems , 35, 45–63.Odell, P. M., Korgen, K. O., Schumacher, P., and Delucchi, M. (2000), “Internet Use among Femaleand Male College Students,”
CyberPsychology & Behavior , 3, 855–862.ONeill, D. K. and Sai, T. H. (2014), “Why not? Examining College Students Reasons for Avoidingan Online Course,”
Higher Education , 68, 1–14.19tter, R. R., Seipel, S., Graeff, T., Alexander, B., Boraiko, C., Gray, J., Petersen, K., and Sadler,K. (2013), “Comparing Student and Faculty Perceptions of Online and Traditional Courses,”
The Internet and Higher Education , 19, 27–35.Paletz, D. L., Owen, D. M., and Cook, T. E. (2013),
American Government and Politics in theInformation Age , Flat World Knowledge.Parker, K., Lenhart, A., and Moore, K. (2011), “The Digital Revolution and HigherEducation: College Presidents, Public Differ on Value of Online Learning.”
Pew In-ternet & American Life Project, Pew Research Center , 28 Aug 2011. Available:https://files.eric.ed.gov/fulltext/ED524306.pdf [Last accessed: 11 July 2020].Rahman, M. A. (2016), “Bayesian Quantile Regression for Ordinal Models,”
Bayesian Analysis ,11, 1–24.Rahman, M. A. and Karnawat, S. (2019), “Flexible Bayesian Quantile Regression in Ordinal Mod-els,”
Advances in Econometrics , 40B, 211–251.Rahman, M. A. and Vossmeyer, A. (2019), “Estimation and Applications of Quantile Regressionfor Binary Longitudinal Data,”
Advances in Econometrics , 40B, 157–191.Robert, C. P. (1995), “Simulation of Truncated Normal Variables,”
Statistics and Computing , 5,121–125.Simmons, G. R. (2014), “Business statistics: A Comparison of Student Performance in ThreeLearning Modes,”
Journal of Education for Business , 89, 186–195.Train, K. (2009),
Discrete Choice Methods with Simulation , Cambridge University Press, Cam-bridge.Williams, S. L. (2006), “The Effectiveness of Distance Education in Allied Health Science Programs:A Meta-analysis of Outcomes,”
The American Journal of Distance Education , 20, 127–141.Xu, D. and Jaggars, S. S. (2013), “The Impact of Online Learning on Students Course Outcomes:Evidence from a Large Community and Technical College System,”
Economics of EducationReview , 37, 46–57.Yu, K. and Moyeed, R. A. (2001), “Bayesian Quantile Regression,”
Statistics and Probability Let-ters , 54, 437–447.Yu, K. and Zhang, J. (2005), “A Three Parameter Asymmetric Laplace Distribution and its Ex-tensions,”