Banking prompts trial 2

There are 2 treatment arms in this trial: a no-intervention control and one intervention arm. The control group will not receive any communication. For the treatment group, participants will be shown a prompt through two channels and be prompted by home loan staff they interact with.
The prompt advises consumers that they may be able to reprice their home loan interest rate and suggests they contact their bank. Consumers can call directly by selecting a link or make contact by calling the number provided.

2025-11-06

2025-11-21

Consumer contact – measured at the individual level: Measured 16 days post intervention start (21 November 2025, primary endpoint).

Note. The primary outcome evaluates the effect of the full multi-channel intervention package (consumer prompts + staff alert), not consumer prompts alone. Treatment group consumers who contact the bank via the contact number provided will trigger the automated staff-sided alert that may facilitate interest rate discussions. Control group consumers receive neither consumer prompts and staff interacting with these consumers will not receive the alert.
Each individual is treated as an independent actor and assumes that there is coordination among individuals in the cluster, but that it is imperfect. That is, we will use cluster-robust standard errors to account for correlations between individuals within each cluster while measuring each person’s separate action.
At an individual level the outcome will be a binary indicator, flagging whether a consumer telephoned the bank on their specialist home loan line in the trial period (0 = did not make contact, 1 = made contact). The variable will be constructed from administrative data measuring the count of calls made to the bank in the trial period. This will be averaged within treatment and control group to give the proportion of consumers who contacted the bank by arm.
Our primary hypothesis is that the proportion of consumers who contacted the bank will be higher in the treatment group than the control group (treatment > control) at T1.
The analysis of the effect of the intervention will consist of a covariate-adjusted comparison of our primary outcome. This estimate, confidence intervals and p-values will be derived from an ordinary least squares (OLS) model using cluster-robust (CR2) standard errors with the following mean-centred baseline covariates:
• Number of times the consumer called the bank in the last 12 months (individual, or mean for the cluster)
• The number of person-loan pairs in the cluster (all outcomes)
• Mean interest rate by cluster
• Mean maturity of loan by cluster
• Flag for any loan in a cluster with loan-to-value ratio (LVR) or less than 80% at origination
• Over 60 (older person) flag either at the individual or at the cluster level (1 if any person in the cluster have the flag, 0 if no one in the cluster have the flag)
• Flag for any loan in a cluster with a balance over $600,000

Interest rate, repricing, consumer contact rate – measured at the cluster level: Measured through Day 21 (26 November 2025, 5 days post intervention end) to allow for administrative processing time between rate offer and system update.

Note.
• All secondary analyses are contingent on statistically significant findings (including Time 2 data collection) and will only proceed if significant treatment effects are observed on the primary outcome.
• All the secondary outcomes will be measured at the cluster level.
All secondary outcomes evaluate the effect of the full multi-channel intervention package (consumer prompts + staff alert), not consumer prompts alone. Treatment group consumers who contact the bank via the contact number provided will trigger the automated staff-sided alert that may facilitate interest rate discussions. Control group consumers receive neither consumer prompts and staff interacting with these consumers will not receive the alert.

Cluster-level consumer contact rate
This will be a binary indicator at the cluster level, taking the value of 1 if at least one consumer in a cluster contacted the bank on their specialist home loan line in the trial period, and 0 otherwise. This definition captures cluster-level engagement with the intervention, recognising that for shared loans, one borrower making contact is sufficient. The variable will be constructed from administrative data measuring the count of calls made to the bank in the trial period. This will be averaged within the treatment and control groups to give the proportion of clusters in which at least one person contacted the bank in each arm.
The cluster-level analysis is consistent with a model of joint decision-making by people within each group. This model assumes perfect coordination. Under this model all members of a group collaborate on decisions and therefore action taken by one member represents a coordinated action.
This assumption is likely to be true of many clusters, especially as most clusters have only one person. However, other clusters represent more disparate groups that may be less coordinated. These include clusters where not all people are borrowers on all loans in the cluster. As repricing does not require any application nor consent of all parties, a call from more than one person in these clusters is possible and represents separate decision-making. The likely level of cooperation or influence within these clusters is unknown.
Hypothesis: The proportion of clusters with at least one person who contacted the bank will be higher in the treatment group than the control group (treatment > control) at T1.
Interest rate
A variable at the cluster level, the interest rate outcome will be a continuous variable representing the mean interest rate of all loans in a cluster expressed as a percentage. This will be averaged within treatment groups to give the mean interest rate by arm. This outcome will be measured 5 days post intervention end day 21 (26 November 2025) as well as at T2 (26 March 2026).
Hypothesis: The mean interest rate will be lower in the treatment group than the control group (treatment < control) at T1.
Repricing
A binary variable at the cluster level of whether at least one loan in a cluster was repriced during the intervention period. This will be constructed from administrative data, with 1 indicating at least one loan was repriced and 0 if it was not. For shared loans, the outcome captures whether the shared loan was repriced regardless of which borrower initiated the change. This will be averaged within treatment and control groups to give the proportion of clusters with reduced rates by arm. This outcome will be measured 5 days post intervention end day 21 (26 November 2025) as well as at T2 (26 March 2026).
Hypothesis: The proportion of clusters with reduced rates will be higher in the treatment group than the control group (treatment > control) at T1.
Note. As with interest rate outcome variable, this outcome evaluates the full multi-channel intervention package (consumer prompts + staff alert). Effects cannot be attributed to consumer prompts alone.
All the secondary outcomes will be measured at the cluster level. These estimates, confidence intervals and p-values will be derived from an ordinary least squares (OLS) model using robust (HC2) standard errors with the following mean-centred baseline covariates:
• Number of times the consumer called the bank in the last 12 months (individual, or mean for the cluster)
• The number of person-loan pairs in the cluster (all outcomes)
• Mean interest rate by cluster
• Mean maturity of loan by cluster
• Flag for any loan in a cluster with loan-to-value ratio (LVR) or less than 80% at origination
• Over 60 (older person) flag either at the individual or at the cluster level (1 if any person in the cluster have the flag, 0 if no one in the cluster have the flag)
• Flag for any loan in a cluster with a balance over $600,000

This project will be a 2-arm cluster randomised controlled field trial. Clusters will be defined in the administrative data as people who share loans, defined through inter-related person-loan pairs. This is because the intervention will be delivered to individuals, and the outcomes will relate to both people (e.g. calling the bank) and loans (e.g. Interest rate reductions).
We will test an omnibus intervention as described above. We will use bank administrative data to measure the primary and secondary outcomes.

Not available

Randomisation will be at the cluster level. Clusters of connected consumers will be randomly assigned to treatment and control groups in a 1:1 ratio. The clusters are flexible structures able to manage a multitude of individual arrangements and the many-to-many relationship between people and loans.

Randomisation will be at the cluster level.

Yes

The planned number of clusters for this trial is 10,419 (defined as unique groups of individuals and their shared loans).

12,701, representing the total number of unique consumers in the dataset

Treatment group: 5,210 clusters
Control group: 5,209 clusters
We excluded clusters where any member participated in the treatment group of Trial 1. Trial 1 control participants remain eligible. Trial 1 control participants remain eligible.

Since randomisation occurs at the cluster level but our primary outcome is measured at the individual level, we account for within-cluster correlation using a design effect adjustment: Design Effect = 1 + (m-1) x ICC Where m is the mean cluster size and ICC is the intra-cluster correlation coefficient. We use ICC = 0.90 as a very conservative assumption, representing extremely high within-cluster correlation. With our small mean cluster size, this produces a design effect of 1.20, yielding an effective sample of 10,610 individuals. In Trial 1, the individual-level contact rate in the control group was 2.16%. With our sample size, design effect adjustment, and using a one-sided test with alpha = 5% and power = 90%, we can detect a minimum effect of 1.32 percentage points. This represents the ability to detect an increase from 2.16% to 3.48%, representing a 61% relative increase. We chose an alpha level of 5% and power of 90% to reduce Type 1 error to maintain high sensitivity to detect effects whilst using standard significance thresholds.