Student success benchmarking in higher education

BACKGROUND

This interactive serves as a one-stop source for colleges and universities to benchmark how they fare against their peers on different student outcome and equity measures across four distinct student segments identified by Pearson.⁷ We have grouped our analysis into four broad categories: access and affordability, student outcomes, flexibility and student services, and other indicators. Provided below are some of the key characteristics for each Pearson student segment that serve as the foundation of our analysis.

Learner segments

1. Traditional learners: 18–24-year-olds attending brick-and-mortar institutions for the traditional college experience. This segment is less price-sensitive, passionate about learning, prefers in-person instruction, and places a high value on education.

2. Career leaners: These learners place a high value on education, see college as a steppingstone to professional success, and prefer to learn digitally.

3. Reluctant learners: This segment has little passion for learning, favors flexibility, is price-sensitive, and places a low value on higher education. These learners attend college because they have to, not because they want to.

4. Skeptical learners: These learners self-identify as average or below-average students, are very price-sensitive, and favor digital modes of learning over in-person.

For each learner segment, we identified a set of relevant variables from the following datasets: IPEDS, College Scorecard, Georgetown University’s NPV database, and Raj Chetty’s “Income segregation and intergenerational mobility across colleges in the United States” database available on the Opportunity Insights Data Library.⁸ While many of the variables selected are common across learner segments, others are unique to segments given differences in attitudes and preferences (see table a).

Table a. Master sheet of all variables included in dataset

METHODOLOGY AND ASSUMPTIONS

This dataset includes public, private not-for-profit, and private for-profit four-year postsecondary institutions in the United States.

At the outset, a total of 2,882 four-institutions were extracted from IPEDS (note: a small minority of the 2,882 institutions are community colleges that have bachelor’s degree as the highest level of offering). This number was reduced to 1,585 institutions through the following steps:

1. All institutions with undergraduate enrollment of fewer than 500 students or institutions with blank values were eliminated from this analysis.

2. Similarly, institutions with blank observations for graduation rates for students from underrepresented minority groups (Black or African American and Hispanic students) and Pell Grant recipients were eliminated, yielding a total of 1,709 institutions.

3. Last, this reduced set of institutions was mapped against Raj Chetty’s social mobility rate (joint probability of parents in the bottom quintile and child in the top quintile of the income distribution) for the 1991 birth cohort available in the Opportunity Insights Data Library. Institutions that were marked as “colleges with insufficient data” in the Opportunity Insights’ “Income segregation and intergenerational mobility across colleges in the United States” database, and institutions that were not part of the database were excluded, yielding a total of 1,585 institutions.

THREE METHODS OF DATA PRESENTATION

Institutions can compare their performance across variables with their Carnegie peers, with other four-year institutions included in the dataset, and with institutional clusters that emerge algorithmically based on all standardized variables included in this dataset.

Comparison with Carnegie peers

This approach allows institutions to compare their performance across variables with that of their Carnegie peers. R1 institutions can, for example, compare their values with other doctoral universities with very high research activity and see where they fall on a quartile basis relative to their R1 Carnegie peers. Variables have been standardized using the averages and standard deviation for a specific Carnegie group in order to facilitate apples-to-apples comparison with similar institutions. None of the binary variables were standardized in this analysis.

The number of institutions that fall in a given Carnegie group represents a limitation to this approach. Carnegie groups with fewer than 27 institutions cannot be split into quartiles since the data is not sufficient to carry out calculations.

For Carnegie groups with fewer than 27 institutions, a density curve is not displayed.
Moreover, where values were blank in IPEDS, they were either replaced with a value of zero or treated as blank values based on guidance from IPEDS staff (see table b).

Table b. Variables with blank values

Comparison with other four-year institutions

This approach allows institutions to compare their performance with all other four-year institutions in our sample. To facilitate meaningful comparison, variables were transformed using the methodology presented in table c. To control for demographic variation across states, for example, enrollment variables (with the exception of Pell Grant recipients) were standardized using the averages and standard deviation for the state in which an institution is located. Note that all state-wide averages were calculated using the data included in the sample of 1,585 institutions, rather than the broader set of four-year institutions. These standardization techniques were applied to all four learner segments.

Table c. Standardization methods for all variables

Cluster analysis

To identify clusters of higher education institutions, we used the k-means clustering algorithm. The algorithm works by allocating institutions into the cluster that minimizes the difference between each member in the cluster and the mean value of that cluster. The algorithm iterates this process and shuffles institutions around until further improvements are not possible. To determine the optimal number of clusters, we used the elbow method that indicates when an additional partition of the data yields a marginal decreasing improvement on the partition. To improve the performance of the model, we also used a dimensionality reduction technique known as principal component analysis (PCA). PCA generates new variables that capture the variation of the original data in a more compact set of new variables. We selected the number of components that preserved 85% of the variation of the original data and applied the clustering algorithm to those variables. We repeated this process for each of the four learner segments.

DATA LIMITATIONS

1. Raj Chetty’s mobility rate (joint probability of parents in the bottom quintile and child in the top quintile of the income distribution) for 1991 birth cohort

• The mobility rate for the 1991 birth cohort was included in this dataset. Where observations were missing for the 1991 birth cohort, we used the closest available observation from cohorts spanning 1985–1991. Observations before the 1985 birth cohort were excluded from this analysis.

• Many institutions have undergone a name change since the time period considered in Raj Chetty’s analysis. In those instances, we manually identified the name change and used the mobility rate for the previous moniker in the database.

• Lastly, all institutions labelled “colleges with insufficient data” in Crosswalk from College-Level OPEIDs to Super-OPEID Groups were excluded from this dataset.⁹

• For additional details regarding the Opportunity Insights’ “Income segregation and intergenerational mobility across colleges in the United States” database, please visit this link.

Note: This analysis was last updated in March 2021. Any updates to the relevant datasets after this date are not reflected in the analysis.