We are undertaking a transformation of our private rental prices' statistics, including making better use of existing data sources, improving methods, and developing systems.
We currently produce the Index of Private Housing Rental Prices (IPHRP) and the Private rental market statistics (PRMS); however, in the future we envisage our rental price statistics to be a new, single, monthly publication that provides a more detailed insight into the rental market.
The new methodology will use a hedonic regression model; this is similar to the approach used to calculate the UK House Price Index.
The development outlined is part of a continuous programme of improvement for consumer price statistics; our ambition is to bring in new data sources for further areas of the inflation basket and continue to improve these statistics over the coming years.
This article provides an overview of the methods that we are expecting to use for our new rental prices estimates. An overview of the project aims and timelines can be found in our Private Rental Prices Development plan.
Work began at the end of 2019, when we first received access to the data that were necessary to begin redeveloping the private rental prices' statistics.
We researched existing methodologies used for rental prices statistics and began developing our future methodology. We have engaged with stakeholders (such as the Advisory Panels for Consumer Prices) and with international experts in property price statistics, including members from the Economic Statistics Centre of Excellence (ESCoE).
We have also worked closely with Office for National Statistics (ONS) methodology experts to quality assure multiple aspects of the methodology. This engagement allowed us to decide on our methodology towards the end of 2021.
Please note, the detail within these methods could change as we build a new production system, however, the overall methodology should remain the same.Nôl i'r tabl cynnwys
Currently, the Office for National Statistics (ONS) publishes two private rental prices statistical outputs:
Index of Private Housing Rental Prices (IPHRP)
Private rental market summary statistics in England (PRMS)
Index of Private Housing Rental Prices
IPHRP measures the change in price of renting residential property from private landlords. A rental price index and its annual percentage change for the UK, its countries and English regions is published.
IPRHP's method is based on a matched-pairs approach, which splits the collected rental transaction data into a sample and a substitution pool. More detail on the methodology can be found within its Quality and Methodology Information.
Existing private rental prices' statistics are used to inform the owner occupiers' housing (OOH) costs element of the Consumer Prices Index including OOH (CPIH), the ONS's lead measure of consumer prices inflation, as well as "actual rentals for housing" aspect of Consumer Prices Index (CPI) and CPIH, and "rent" in the Retail Prices Index (RPI). It is anticipated that the outputs from this development work will eventually be used in the consumer prices statistics.
Private rental market summary statistics in England
PRMS publishes point-in-time arithmetic mean and median rental price estimates for England, English regions and English local authorities.
The current methodology uses all rental transaction data, but limitations prevent compositional changes from being taken into account, so it is not appropriate to compare PRMS estimates year-on-year to infer trends in the rental market and a price index cannot currently be produced. Further detail on the methodology can be found within the PRMS publication.Nôl i'r tabl cynnwys
Note: The detail within these methods could change as we build the system, however, the overall methodology should remain the same.
We will measure the change in price of renting residential property from private landlords. Other measures of private rental prices, such as those published by Homelet, Rightmove and Zoopla, are produced using only newly advertised rentals, whereas we will produce a measure that reflects both the newly "agreed" rents and existing rents. Therefore, we aim to reflect the stock of rents and not the "flow" of new rents. This is how the current Index of Private Housing Rental Prices (IPHRP) is measured.
The Johnson Review (2015) points to research that suggests a flow measure may be worth considering; that is, only new lets. We investigated the feasibility of measuring the flow of rents, and we concluded that we do not currently have data sources available to us to disaggregate new rents from existing rents.
The new measures of rental prices bring together several rich administrative data sources. The methodology will use a hedonic regression model, which will allow for mix-adjustment of the monthly price data to control the effect of the changing composition of collected rental properties; this is similar to the approach used to calculate the UK House Price Index. However, the exact detail of each stage in the methodology will be tailored to suit the rental data. These stages are:
Quality assurance checks are completed on the data
Data are cleaned and property records are linked
On an annual basis, a "fixed basket" of properties is created
On a monthly basis, data are fitted to a hedonic regression model to quantify the relationship between property characteristics and associated rental price for each calendar month
Using the coefficients from the model's output, imputed prices for properties within the fixed basket are calculated
Elementary aggregates are produced at a local authority level by taking the ratio of the geometric means of the predicted prices in the base month and the current month
Elementary aggregates are weighted together (Lowe index) and then chain-linked annually to produce a rental price index series over time
The corresponding average rental price series is derived by applying the index to a base set of rental prices from the reference period. For example, if the average rental price in the reference period was £500, and the index in the current month was 110.0, a 10% growth would be applied to the reference period average rental price. So, the average rental price in the current period would be estimated at £550. This ensures the price series is consistent with the published index.
The new measures of rental prices bring together several rich administrative data sources. The data sources used fall into two distinct categories: price data and property attributes data. Combining the detailed property attributes data with the price data provides a comprehensive dataset required for use in a hedonic regression model.
The Valuation Office Agency (VOA), Scottish Government, Welsh Government and Northern Ireland Housing Executive (NIHE) deploy rental officers to collect information on the prices paid for privately rented properties, along with some characteristics of the properties. Data for Northern Ireland also include data provided by propertynews.com.
Annually over 450,000 private rental prices are collected in England, 30,000 in Wales, 25,000 in Scotland and 15,000 in Northern Ireland, which make these sources of data rich. Further information on the data sources can be found within our Index of Private Housing Rental Prices quality and methodology information publication.
To strengthen our methods, we are now able to link these rental prices data to other property attributes data (for example, age of the property and floor area), such as from Council Tax data. Separately, we can also link to a geo-demographic segmentation, which will help control for differences in smaller areas.
When a rental price is collected, it will be assumed to be valid either for 14 months from its entry date into the system, or until an update is received. A 14-month validity period will be used as it balances typical contract lengths (which tend to be either 6, 12, 18 or 24 months) against operational practices.
On an annual basis, expenditure weights are calculated to ensure the estimates are representative of the UK. To calculate expenditure weights, dwelling stock data are multiplied with average rental prices. Dwelling stock data come from the Office for National Statistics, Department for Levelling Up, Housing and Communities (DLUHC), Scottish Government, Welsh Government, Department of Finance Northern Ireland and Northern Ireland Housing Executive (NIHE). Dwelling stock estimates are split by the proportion of property types rented privately in Wales, Scotland and the nine regions of England using data from the English Housing Survey and equivalent sources from other countries.
To calculate timely expenditure weights, the most recently available data are used. For a given year, y, the dwelling stock data are based on the period y-3, while average prices are based on the period y-1.
Valuing a rental property
A regression model is used to estimate the value of each characteristic from the set of properties during a period. For example, the model might estimate the effect that every additional room and each different location have on the rental price in a certain month. Then, the rental price of a property can be calculated by combining the values assigned to each of its features. This method allows us to estimate the prices of properties with every combination of features (such as number of rooms and local authority), even if that combination was not collected in the period.
The price-determining characteristics that we expect to use for England (this may change for other countries depending on property attribute availability) are:
number of bedrooms
property type (detached, semi-detached, terraced, flat or maisonette)
geo-demographic segmentation (ACORN)
property age bracket
Mathematically, a semi-log ordinary least squares (OLS) model will be used:
pi is the rental price of property i
K is a constant
βⱼ is the coefficient associated with characteristic j
xji indicates whether property i has the characteristic j (such as detached property); if so, it takes the value 1, otherwise it takes the value 0 (expect for floor area where it takes the floor area in square metres
ei is the statistical error term
The logarithm of the rental price paid is used because rental prices tend to be log-normally distributed, meaning the frequency distribution of the log of the rental price is bell-shaped.
The rents development methodology is mix-adjusted to control for the fact that different types of rental properties might be collected in different periods. The process of mix-adjustment requires that, in each January, a fixed basket of properties is updated to reflect changes in the composition of rental properties. This basket is then used to produce imputed rental prices for the current year, before the basket is then updated again in the subsequent January.
The fixed basket contains all the rental prices collected in the previous year. If any rental property had been collected more than once in the year, the most recently collected data would be used.
A single univariate decision tree imputation routine is used to impute missing property characteristics in the fixed basket, as recommended by the Editing and Imputation Expert Group within the Office for National Statistics (ONS). Several reasons were given for choosing to impute using a univariate decision tree including that they are fast to implement and re-train, and they are easy to interpret.
Calculating an index
The Ordinary Least Squares model creates coefficients, which are used to calculate an imputed rental price for each property within the fixed basket. These imputed rental prices are then averaged using a geometric mean, which involves multiplying the "n" imputed prices together, and then taking the nth root.
A fuller description of this method and other alternative methods for calculating residential property prices can be found in the Handbook on Residential Property Price Index.
We aim for these measures to be available for the UK, its constituent countries, English regions and local authorities. The new publication will contain:
an index of private rental prices
annual rates of change
average private rental prices
a breakdown of private rental prices by geography and bedroom category (studio, one bedroom, two bedrooms, three bedrooms and four or more bedrooms)
Within our data, sometimes a rental property can have missing property characteristics. In the fixed basket, a single univariate decision tree imputation routine is used to impute these missing property characteristics. On a monthly basis, any properties with missing characteristics are dropped from the model.
We treat missing property characteristics differently in both scenarios because when testing different imputation options, we found that imputing the sample that is used in our hedonic regression each month using a single univariate decision tree gave differing outputs to other imputation methods. This meant the differences in the indices were driven by the imputation method. The imputation rate is currently low and dropping data points with missing property characteristics is unlikely to introduce bias into the data. This decision will be monitored and reviewed as necessary.
Following engagement with experts (Advisory Panel for Consumer Prices - Technical, and international property prices experts), we considered multiple models in our methodological research phase. These included Ordinary Least Squares (OLS), Weighted Least Squares (WLS) and Random Forest approaches. We also considered the use of interaction terms in the model.
By using K-fold cross-validation and generalised variance inflation factor, we decided to use an OLS model without interaction terms. As well as being methodologically robust, the chosen model must also balance different aspects of quality.
Accessibility and clarity
The behaviour of some models are easier to describe, understand and extend. Linear models are well established in existing literature, while others currently require specialised expertise in machine learning.
Coherence and comparability
The House Price Index uses a WLS, and so although we propose using OLS for the rental prices, this does have internal similarity with other housing statistics produced by the Office for National Statistics (ONS). Separately, StatCan recently moved from a matched pairs model to a hedonic model using a log-linear regression when calculating their rental prices.
Timeliness and punctuality
Models that are easier to troubleshoot and computationally less intensive tend to be easier to implement in the production of official statistics. The monthly production round provides us with around one week to produce, quality assure and evaluate our statistics. More complex models can take more time to run and evaluate.
It is usually easier to find support for, and to maintain, simpler, well-understood models.Nôl i'r tabl cynnwys
Throughout 2022 we worked to build a sustainable system for use in a monthly production environment. In 2023 we have been thoroughly testing to ensure we are producing a high-quality production system.
Future dates are estimates and are subject to continued systems development, research, and impact analysis to ensure the quality of our statistics, which is our priority. Decisions will be made through continuous engagement with our stakeholders. An outline of the future project milestones of the statistical development work includes:
in November 2023, we plan to produce an impact analysis release, covering data from 2015 up to July 2022 for Great Britain, published following feedback from the Advisory Panels for Consumer Prices in October 2023
following the publication of data in November 2023, we will be engaging with users
in early March 2024, we aim to publish a decision on the incorporation of these measures into official statistics
we plan to publish the first consumer price statistics and price index for private rents statistical bulletins using the new methodology for Great Britain in March 2024, subject to the decision on the incorporation of these measures into official statistics
we plan to update the consumer price statistics and price index for private rents statistical bulletins using the new methodology for Northern Ireland by March 2025
Administrative data are data that people have already provided to the government through day-to-day activities, for example, health records, social security payments or educational attainment information.
Interaction effects occur when the effect of one variable depends on another variable. An interaction term within a model accounts for these interaction effects.Nôl i'r tabl cynnwys
Manylion cyswllt ar gyfer y Erthygl
Ffôn: +44 1633 456400