The Road to Retention Passes through First Year Academic Performance: A Meta-Analytic Path Analysis of Academic Performance and Persistence

ABSTRACT

This meta-analytic path analysis presents evidence that first-year academic performance (FYAP), measured by first-year grade point average (FYGPA) plays the major role in determining second-year student retention and that socioeconomic status (SES), measured by parental income, plays a negligible role. Based on large sample data used in a previous study, we conducted additional analyses that included corrections for measurement error and created path models using precollege academic achievement, measured by ACT Composite scores and high school GPA (HSGPA), and SES to predict FYAP and then second-year retention. Precollege academic performances had direct effects on FYAP, and FYAP fully mediated their effects on second-year retention. SES did not contribute to the prediction of FYAP, and it had only a trivial effect on second-year retention. The results of this study point to the criticality of FYAP, and supporting first-year student academic success may serve as the central vehicle for retention efforts.

Notes

¹ Although constructs and true scores can sometimes not be co-linear, research shows that under most circumstances in research the two are highly correlated (.98 on average), indicating that construct and true score dimensions are close to perfectly co-linear (Schmidt et al., 2013).

² Selective, Traditional, and Liberal institutions typically drew students from the top 25%, top 50%, and bottom 50% of their high school class, respectively. Open institutions are typically open to all students with a high school diploma or equivalent. Interquartile ranges for ACT Composite scores at Selective, Traditional, Liberal, and Open institutions were 21–26, 18–24, 17–22, and 16–21, respectively (ACT, 2010). The characteristics of the institutions included in the current study differ somewhat from those previously reported for a nationally-representative sample of ACT-tested students who immediately enrolled in college in fall 2003 (see Table 3 on page 8 of Radunzel and Noble (2012)).

³ To examine the stability of the three predictor variables across seven cohorts, we calculated the correlation for each possible comparison between cohort years for each predictor (21 each, 63 overall). First, we calculated the mean ACT Composite score, HSGPA, and PI at each institution within each year. Next, we calculated the correlations between the institution-level means for each predictor over the seven years. (For example, if 30 institutions had data for years one and two, the 30 mean institution-level ACT Composite scores from the first year were correlated with the mean institution-level ACT Composites scores from the second year.) The average number of institutions in each correlation was 25.2, and the average correlations across time for mean ACT Composite scores was .96; for mean HSGPA the average correlation was .96; and for mean PI the average correlation was .95. Given the high correlations, we concluded that the student populations within the schools were stable enough to justify pooling data across cohorts within each school. This made it possible to conduct one validity study per school rather than conducting multiple studies for each school.

⁴ Research has shown that admission test scores such as ACT scores measure more than just general intelligence (Coyle & Pillow, 2008). Others consider admission tests as measures of developed ability (Zwick, 2002), and research has shown that the scores are strongly related to the rigor of the courses taken in high school (Ferguson, 2004; Sawyer, 2008).

⁵ The ACT College and Career Readiness standards provide educators with descriptors of what students within given score ranges are expected to know and be able to do. This is information that colleges can use to help with score interpretation and for placement purposes (ACT, 2007a, 2017).

⁶ One may question why corrections were not made for multivariate range restriction. Using corrections for multivariate range restriction assumes that range restriction on FYGPA was only due to explicit selections on HSGPA, ACT Composite scores, and PI. In fact, students may drop out of colleges in their first year due to many other reasons not captured by these predictors. The indirect range restriction correction procedure introduced by Hunter et al. (2006), on the other hand, does not require that assumption. The correction only assumes that selection occurred on an unknown variable (“suitability”), which can be considered as representing a combined effect of all factors potentially influencing student drop-outs. As such, correcting for indirect range restriction is actually more appropriate here.

⁷ In classical test theory, the squared correlation between observed scores and true scores would be the reliability estimate, as ${\rho }_{XT}^2$ = ${\sigma }_T^2/{\sigma }_X^2$ (Lord & Novick, 1968). For the current study, we considered the correlation between self-reported parental income and actual parental income as a correlation between observed scores and true scores. Though not an ideal estimate, we argue this estimate is better than assuming that students reported their parents’ income with perfect accuracy and that the reliability of student-reported parental income was 1.00.

⁸ Using harmonic N in path-analysis/SEM based on meta-analysis results is a well-accepted procedure in research in applied psychology. It was originally suggested by Viswesvaran and Ones (1995). Many studies published in major journals in applied psychology (e.g., Christian, Garza, & Slaughter, 2011; Colquitt, LePine, & Noe, 2000; Judge & Piccolo, 2004; Podsakoff, Whiting, Podsakoff, & Blume, 2009) have used the procedure. Use of the arithmetic mean produces essentially identical results.

⁹ Acceptable ranges for the fit indices are CFI >.90; GFI >.90; RMSEA <.10, and 90% confidence intervals <.10; and SRMR <.10 (Kline, 2005).