METHODOLOGY Year : 2019  Volume : 2  Issue : 2  Page : 5154 Analyzing educational interventions without random assignment Samuel C Karpen Office of Academic Affairs, University of Georgia College of Veterinary Medicine, Athens, Georgia, USA Correspondence Address: Since educational researchers rarely have the luxury of random assignment, confounding variables are a common concern. This manuscript introduces readers to methods for statistically controlling confounding variables, namely propensity score matching, propensity score weighting, and doubly robust estimation. These techniques allow researchers to accurately estimate the effect of an intervention (e.g. a new teaching method's effect on course grades) even when the groups being compared differ on other relevant variables (e.g. one group has a higher preDVM GPA than the other). Example analysis are included to aid researchers hoping to conduct their own analyses.
Introduction Fluctuations in admission criteria, curricular content, and curricular delivery can make it difficult to attribute changes in students' performance to specific interventions. If an instructor in a graduate course finds that students who experienced a new teaching technique earned higher final grades than previous students who did not experience the technique, is she/he to conclude that the intervention worked? Perhaps, current students earned higher undergraduate grade point average (GPA) or higher graduate record examination (GRE) scores than previous students, and this is driving the difference in performance. Similarly, in 2016, the National Association of Boards of Pharmacy witnessed an overall drop in the North American Pharmacist Licensure Examination (NAPLEX) pass rates after changing the examination format from adaptive delivery to fixed form, increasing the number of questions from 185 to 250, and allowing 6 h instead of 4 h and 15 min.[1] Since 2012, however, applications to pharmacy schools have been steadily decreasing, encouraging colleges/schools of pharmacy to admit previously unacceptable students.[2] Was the drop in NAPLEX scores due to relaxed admissions standards starting in 2012 or changes to the examination in 2016? In both examples, variables outside of the researchers' control, known as confounding variables, made it difficult to determine a change's source. Propensity score techniques can mitigate both the NAPLEX and new teaching technique dilemmas by minimizing the effect of confounding variables. Karpen and Welch 2018 found that the changes to the 2016 NAPLEX did not significantly affect pass rates after accounting for the fact that 2016 examinees had significantly lower GPA and Pharmacy College Admissions Test scores than previous years' examinees. While the teaching technique scenario was hypothetical, it is reasonable to assume that no two classes are academically or demographically identical before entering a graduate program. These differences must be accounted for when comparing classes within the graduate program. Propensity score matching (PSM) attempts to account for confounding variables by comparing treatment cases to control cases who are as similar as possible in terms of the confounding variables. [Table 1] presents a hypothetical scenario in which some students received an intervention intended to boost the final examination scores and others did not. In addition, students who received the intervention had lower undergraduate GPAs than students who did not receive the intervention (Mnointervention = 3.52 vs. Mintervention = 3.34). If researchers compared the groups' mean final examination scores without considering GPA, they would conclude that the intervention harmed examination performance by 2.6 points; however, the nonintervention group should outperform the intervention since it is composed of students with higher preadmission GPAs. By comparing nonintervention cases to intervention cases with similar preadmission GPAs, the researchers can estimate how the nonintervention group would have performed without a GPA disadvantage. In this example, the new control group's mean final examination score was 77.6, indicating that when preadmission GPA was accounted for, the intervention increased the final examination performance by 3.8 points.{Table 1} Manual matching is feasible with [Table 1] data; however, most studies include hundreds of cases and multiple confounders. Rather than painstakingly matching each case on multiple confounding variables, researchers summarize participants' scores on the confounders using logistic regression or generalized boosted models.[3],[4] The summary is called a propensity score, and it represents the odds of treatment group membership, given a case's confounder scores. Nonintervention cases that are most similar to the intervention cases in terms of the confounders will have the highest propensity scores. A new control group is then created by matching individuals in the nonintervention group to individuals in the intervention group who are most similar in terms of propensity score and therefore most similar in terms of the confounders. The most common PSM methods are nearestneighbor and inverse probability weighting. In nearestneighbor matching, each treatment case is matched to a control case with an acceptably close propensity score. The range of acceptability is called the caliper. It is generally advised to start with a very narrow caliper and increase its size until every treatment case has a match. In inverse probability weighting, each treatment case is weighted by (1/its propensity score) and each control case is weighted by (1/1 minus its propensity score). Control observations that are similar to treatment observations in terms of the confounders will be weighted more heavily than those that are dissimilar, and treatment observations that are similar to control observations will be weighted more heavily than those that are dissimilar. One advantage of inverse probability weighting is that cases without adequate matches are retained, thereby limiting data loss.[5] After matching or weighting, researchers should ensure that the newly matched groups are similar in terms of the confounders. If matching/weighting is effective, the treatment and new control group will have similar scores on the confounders and similar propensity scores. Researchers can examine the effect sizes for the confounders by the group before and after matching/weighting. Effect sizes should be smaller (ideally [6] Example To illustrate propensity score methods, the authors created a dataset resembling one that educators may encounter. It included GPA, GRE, endofyear comprehensive examination score, and an indicator of whether a student completed a study skills course during orientation. One group completed the course and the other did not. The dataset was designed so that GRE, GPA, and comprehensive examination score were correlated with one another such that students with lower GPAs had lower GRE and lower comprehensive examination scores. In addition, a higher proportion of students with low comprehensive examination scores were placed in the study skills group. Descriptive statistics and correlations are shown in [Table 2] and [Table 3], respectively.{Table 2}{Table 3} Unweighted analysis An unadjusted comparison indicated that the study skills course harmed comprehensive examination performance by 1.1 points (95% confidence interval (CI) = 0.79–1.56). This conclusion, however, ignored the fact that the study skills group also had lower GRE scores and GPA than the nonstudy skills group and that both of these variables correlated with comprehensive examination performance. Weighting Propensity scores for each participant were generated with a generalized boosted model. These values represent the inverse probability of being in the study skills group, and when applied to the data, they adjust for the effects of GRE and GPA. Each treatment case was weighted by (1/its propensity score) and each control case was weighted by (1/1 minus its propensity score), thereby making the groups more similar in terms of GPA and GRE score [Appendix 1 for R code]. Balance check Prior to weighting, both GPA and GRE differed markedly between the study skills and no study skills groups. Specifically, the effect size for GRE was 0.328, and the effect size for GPA was 0.863. After weighting, the effect sizes for GPA and GRE were reduced to 0.062 and 0.302, respectively. While 0.302 was still larger than ideal, it was less than half of the effect size observed in the unweighted data. In addition, the mean GPA and GRE for the weighted control group were 3.67 and 310, respectively, which were much closer to the means of the treatment group (3.62 and 310) than were observed in the unweighted data (3.79 and 313) [Appendix 2 for R code]. Doubly robust estimation The final step was to build a model using the weighted data in which the treatment group was used to predict comprehensive examination scores while including GRE and GPA as covariates to account for any confounding missed by the inverse probability weighting. As indicated by the positive coefficient, the study skills class improved comprehensive examination scores by 0.39 points after accounting for GPA and GRE scores (95% CI = 0.18–1.03) [[Table 4] and Appendix 3 for R code].{Table 4} Summary Since random assignment is sometimes impossible, and since curricula and admissions metrics often fluctuate, it is important for researchers to have tools to establish statistical control. This article introduces educators to one such tool. We hope that this review will enable and encourage educators to analyze previously problematic data and to produce more accurate estimates of the effects of their interventions, as both have implications for program quality improvement. Financial support and sponsorship Nil. Conflicts of interest There are no conflicts of interest. Appendixes Appendix 1 #Computing Inverse Probability Weights install.packages('twang') install.packages('ggplot2') install.packages('dplyr') install.packages('survey') install.packages('knitr') library (twang) library (ggplot2) library (dplyr) library (survey) library (knitr) ps.Data < ps (Treatment ~ GRE + GPA, data = Data, n.trees = 15000, interaction.depth = 3, shrinkage = 0.001, perm.test.iters = 0, stop.method = c(”es.mean”, “ks.max”), estimand = “ATT”, verbose = FALSE) Appendix 2 #Assessing Balance After Weighting Data$pscores < ps.Data$ps$es.mean.ATT ggplot (data = Data, aes (pscores, fill = factor (Treatment)))+geom_density (alpha=0.5)+ theme_classic() Data.balance < bal.table (ps.Data) kable (Data.balance$unw) kable (Data.balance$es.mean.ATT) Appendix 3 #Doubly robust estimation Data$weights < get.weights (ps.Data, stop.method = “es.mean”) design.ps < svydesign (ids = ~1, weights = ~weights, data = Data) summary (svyglm (Exam ~ Treatment + GRE + GPA, design = design.ps)) References


