The lift in persistence is essentially determined by comparing the persistence outcomes for the students who participated in the initiative who were matched through PPSM with the persistence outcomes for the matched students who were eligible but did not participate in the initiative. Note that Illume Impact uses the same definition of persistence (e.g. fall to spring, term to term) for your institution as Illume Students.
When students are matched through prediction-based propensity score matching, their persistence predictions should be similar to ensure that they were equally likely to persist independent of initiative participation. Persistence predictions are calculated for each student who is currently enrolled in at least one course. The variables that have historically been predictive of persistence at your institution are used to generate prediction scores for current students, showing their individual likelihood to persist. The persistence lift calculation considers the predicted persistence rates for the matched participant and comparison groups to remove any bias that may have remained after matching, especially in the case of small data sets.
First, the predicted persistence rates for the participant group and the comparison group are determined, informed by your institution's persistence model. Next, the actual persistence rates for the two groups are measured. The lift in persistence is calculated by first taking the difference between the actual and predicted persistence rates for the two groups:
(actual persistence rate of comparison group - predicted persistence rate of comparisongroup) = C
(actual persistence rate of participant group - predicted persistence rate of participantgroup) = P
The difference of the resulting values is the lift in persistence that you see on the initiative card:
Lift in persistence = P - C
For large data sets, this calculation is often very similar to the result of measuring the difference in the actual persistence rates of the participant and comparison groups. The difference in predicted persistence rates between the two groups should be negligible if the sample size is very large.