- 10:00 – 10:05 – Introduction – Sir Bernard Silverman
- 10: 20 – 10:50 – Census update – Jon Wroth-Smith
- 10:55 – 11:50 – EAP173 – Alternative Household Estimate – Tanita Barnett
- 11:55 – 12:50 – EAP171 – Statistical Contingencies – Jon Wroth-Smith, Cal Ghee
- 13:30 – 14:25 – EAP172 – Model selection for Coverage Estimation – Viktor Račinskij
- 14:30 – 14:45 – Annual report sign-off – Bernard Silverman
- 14:45 – 15:00 – Actions update – Gareth Powell
- 15:00 – 15:15 – AOB – Bernard Silverman
- Sir Bernard Silverman (Chair)
- Prof Ana Basiri
- Dr Oliver Duke-Williams
- Dr Nik Lomax
- Prof David Martin
- Prof Natalie Shlomo
Office for National Statistics
- Cal Ghee (ONS Attendee)
- Gareth Powell (ONS secretariat)
- Jon Wroth-Smith (ONS Attendee)
- Owen Abbott (ONS Deputy Lead and Presenter)
- Tanita Barnett (Presenter)
- Bethany Fitzgibbon (ONS Attendee)
- Harry Evans (ONS Attendee)
- Viktor Racinskij (Presenter)
- Abu Hossain (ONS Attendee)
- The chair welcomed attendees to the session. The session was conducted by each in-person attendee using their own laptop camera, but with a single microphone and loudspeaker in the meeting room. This allowed all participants to engage in the meeting on equal terms.
2 – Census update
- ONS provided an update on Census operations.
3 – EAP173 – Alternative Household Estimate
- Tanita Barnett presented the AHE work and listed some areas for particular feedback.
- The panel advised that the choice between the “2011+” method and the modelling approach was mostly one of prioritising variance against bias. As long as the priorities are clear and the approach documented, either choice was valid.
- The panel also advised that a minimum of five for a stratum to be used was very low. Although issues would be likely to cancel out overall, usually a lower-limit of 20 or even 50 would be more usual. They suggested sensitivity analysis to provide assurance of what level should be chosen. ONS said that sensitivity analysis, especially once the full data is available, would be used to examine this problem.
- ONS asked whether the key assumption – that the ratio of truly ‘empty’ households was the same in responding households as non-responding households – was felt to be valid. The panel advised that while there may be some issues, they were probably not too serious as long as the issues were properly documented.
4 – EAP171 – Statistical Contingencies
- Jon Wroth-Smith and Cal Ghee spoke about some of the issues so far and the contingencies used. Discussion centred around the three main potential issues being looked at – students, within household missingness and the CCS return rate. Contingencies have been put in place for all three and these were discussed.
- The panel discussed the imputation of students from out-of-term addresses and the relative benefits of imputing missing variables normally, versus using information from others in the donor household.
- The panel also asked about whether the issues being discussed tending to have a strong geographical component, meaning that it was possible that where issues exist, they might be particularly strong in some areas. ONS agreed that there was a risk of this, but no evidence of an issue so far; at the moment coverage does not seem worse than in 2011.
5 – EAP172 – Model selection for Coverage Estimation
- Viktor Račinskij gave an overview of the proposed model selection process.
- The panel raised concerns about not including local authority (LA) as a main effect in the model. ONS responded that while region and various interactions with region are included, including LA as a main effect leads to overfitting, and the models perform better using LA as a random effect.
- The panel had concerns that many issues had spatial elements and that different areas of E&W will have different behaviours. ONS said that although including LA as a main effect was impractical for the modelling, regional effects were captured directly by conditioning on region variable while local authority effect reflected via random residuals; testing on 2011 data and simulations have shown better results with LA included as a random effect rather as a main effect.
- To allay the panel’s concerns, it was suggested that ONS could build a branch of model selection with LA as main effect once data is available; this would give direct evidence on whether it works or not. The panel also suggested a middle way of using LA classifications, rather than LAs themselves.
- The panel raised no issues around the other key features of the paper, though ONS agreed to include Classification and Regression Trees (CART) into the selection system (data exploration step).
6 – Annual Report Sign-off
- Some small further changes were suggested. The panel then agreed the report and that the chair would submit it to the National Statistician. It was agreed that the MARP minutes and papers for the year would be regarded as annexes to the report and therefore as far as possible they should be put into the public domain at the same time as, or before, the report itself is published.
ACTION: ONS to publish the Annual Report and all associated MARP papers.
7 – Actions Update
- The panel reviewed the actions and asked for a verbal update on A79 at the next meeting. They also asked for a written report on Variance Estimation, rather than the scheduled presentation that had been planned.
8 – Any other business
- The panel discussed the format and frequency of future meetings. ONS will need to discuss this further given that meetings will probably centre around the 2023 recommendation in future.