Impact

Does Peer-Supported Learning work?

For the last 16 years, the PSLC has kept anonymised records of attendances and outcomes of students attending the various modules supported. In that time, 30,305 students were targeted for support. 

Having this big set of data, we decided to run some simple analysis on it, generating a predictive model using all data up to AY2025/6 which we then applied to the AY2025/6 data to generate predicted grades and then compare to actual grades to check the efficacy of the prediction.

We used 21,430 rows of data for the training. Of these 19,471 were standard CAO entry, and 1,959 were non-traditional entry (mature, international, erasmus and other non-traditional routes).

Of the 21,430 students in the training set, 4,508 (21%) attended at least one support session. 16,922 (79%) didn't attend any sessions. Of those students who withdrew from their course 96% didn't attend any PSLG support as was the case for 85% of those who didn't sit the exam.

This suggests a possible correlation between non-engagement with PSLG and non-completion.

So most don't attend, and those who do come to less than half of the supports offered in a semester. So what do we find when we look at the final grade outcome? 

The average QPV for those attending at least one session across all the modules supported was 2.871. 

The same average for those who didn't attend was 2.568, a difference of 0.302 or more succinctly, here the difference between a C1/C2 and a B2/B3. (Mann-Whitney U test: p < 0.001, effect size r = +0.172(small). 

The effect is statistically robust across all 21,430 graded student records.

That's quite a jump for those not attending much. The difference becomes more pronounced as attendance increases though peaks at the highest end of the attendance bands would suggest that there is a maximum impact threshold.

The argument can of course be made that stronger students are more motivated and engaged in their studies, which may contribute to their higher performance. 

So we then examined the CAO entry points to see if the students' incoming CAO points suggested a link to their attendance rates at supports. Here's what we found.

Analysis restricted to standard CAO entry students (n=19,471). Spearman correlation: CAO vs attendances rho=0.107 (p<0.001) — weak but significant overall. The relationship is U-shaped rather than linear.

Both the lowest (<300 points) and highest (550-625 points) CAO bands attend at substantially higher rates than the middle bands. 

The middle band (400-499) represent the largest group but have the lowest attendance rate.

So the next step was to examine attendance and non-attendance patterns across the bands. 

We took a simplistic approach here and just looked at high CAO (above the median of 466 points) and low CAO (on or below the median) and their associated attendance rates and final outcomes.

Key finding 1

Low CAO students benefit more from attendance (+0.295 vs +0.194) than non-attendance than their high CAO counterparts.

Key finding 2

Low CAO attenders (2.542) move towards the performance of High CAO non-attenders (2.942) a gap of only 0.4 QPV points.

Key finding 3

Where CAO would normally act as a predictor of final outcome, for attenders (rho=0.406) vs non-attenders (rho-0.432) the attendance partially decouples grade from CAO entry points.

So we have now a situation where those attending do better than those who don't. 

The vast majority of the 'stronger' students (based on CAO entry points) are not attending at the same rate as the 'weaker' students, and attendance helps those 'weaker' students more than their 'stronger' classmates suggesting the ability of the support to decouple entry position from grade outcome. 

We were now ready to create a predictive model to see if the results are consistent across years.

We decided on 4 models to allow for the different types of student entry and to allow for the impact of the COVID-19 lockdown on the model training data. Our models were as follows and used Gradient Boosting regression model training on the 2010 to 2024 data sets. For the no-COVID models 2019 and 2020 data was excluded.

MAE: Mean Absolute Error, R²: Coefficient of Determination

Of the 4 models, Model A (standard entry, all years) performed best in terms of MAE and R². It's most important features were the CAO points, and the course of study. When we applied the training model to the holdout data (AY2025-26) some interesting patterns emerged. 11 modules were supported that year and the accuracy was within 0.2 QPV points for 7 of those 11 modules. For the 4 where the results varied greatly:

• Two changed their assessment method this year (AY2025/26) over the previous 15 years.
• One had a brand-new lecturer
• The last had significant timetabling issues and ran both late and for a shorter duration as a result.

In other words, the model was able to detect anomalies with the modules external to the data for these 4 modules.

For the two where there was a change in assessment method, the module underpredicted the performance of the students. For those where the supported late or the lecturer changed, the model overpredicted the performance of the students.

The Equity gap also changed as a result for the modules this year with the low CAO high attending students (four or more sessions in a semester) gaining 0.555 QPV points over their low or no attending counterparts, and compared to a difference of 0.265 QPV points for their High CAO classmates. The mean QPV for the Low CAO High Attending students (2.804) also almost matched the mean QPV for the High CAO Low/No attending students (2.882).

Altogether, this suggests attendance at PSLG support (particularly at higher and consistent rates of attendance) shows a correlation with stronger final module outcomes and seems to benefit those students entering with lower CAO points more than their higher CAO point classmates.

But of course, the support has more than just an academic impact as we'll see in our testimonials page.

Caveats and limitations

Correlation not causation

All associations are correlational. Motivated students may self-select into support attendance. A controlled experiment would be needed to establish causality.

No student identifier

Each row is treated independently. A student taking three modules appears three times. Results describe module-enrolment outcomes, not individual trajectories.

Model accuracy

CV MAE 0.631 QPV points; holdout MAE 0.740. Suitable for group-level analysis but not for high-stakes individual decisions.

Holdout generalisation

The ~17% increase in MAE from CV to holdout reflects cohort shift and module-level structural changes, not purely overfitting.

COVID years

2019–2020 show elevated QPV from grade inflation. Downstream effects are real and retained in Model A; year-level comparisons should note this context.

Non-traditional entry

Modelled separately with lower accuracy (CV R²=0.127). May warrant separate outreach and support strategies.