Revaluation and reform of council tax in Scotland: design considerations and potential impacts

Appendix A. Methodology

A.1 Geographical analysis

To undertake our geographical analysis, we use data from three main property-level data sources:

(1) Scottish assessors data on all residential properties in Scotland, which tells us each property’s existing council tax band, as well as some characteristics (such as dwelling type, including whether it is detached, semi-detached, terraced, etc.);

(2) energy performance certificate (EPC) data, which provides further information on property characteristics (such as size, build date and energy efficiency) for a (large) sub-set of properties; and

(3) Registers of Scotland data on the date and transactions value for properties that have transacted.

We also use data on a range of local area characteristics, including deprivation, rural/urban classification and access to local services, as well as the neighbourhood (‘data zone’ or ‘intermediate zone’ depending on sample size) in which a property is located.

Matching properties in different datasets

Where possible we match properties between datasets using their Unique Property Reference Number (UPRN). This is designed to be a unique number for all properties (both residential and non-residential) across Scotland. However, not all properties have yet been assigned UPRNs, some UPRNs actually cover multiple properties (for example, if unique UPRNs have not yet been assigned to houses converted into flats) and not all records in our input datasets contain UPRNs even if properties have had them assigned. For this reason, in order to expand the number of properties which we are able to match we match properties as follows:

• For properties with a unique UPRN available in all datasets, we match on the basis of UPRN.
• For properties where this is not the case, we first try to match on the basis of the first line of their address (which could be house number and street, or apartment number and apartment building), and their full post code.
• For properties that we still cannot match, we then attempt to match on the basis of the first and second line of their address (which might add a neighbourhood for a house, and a street name for an apartment) and the first part of the postcode (e.g. ED1).
• For properties that we still cannot match, we then attempt to match on the basis of the first two words in their address (e.g. “The Ridings”) and their full postcode, excluding flats (to avoid matching the wrong flats if, for example, there is more than one “Flat 1” in a postcode).
• For flats that we still cannot match, we then attempt to match on the basis of the first four words in their address (e.g. “Flat 1, the Maltings” and their full postcode.

We investigated using ‘fuzzy matching’ approaches which do not require exact matches but the standard approaches for this in the statistical software we use (Stata) do not distinguish between numerical differences (e.g. “10” and “15”), which mean two records highly likely represent separate properties, and textual differences (e.g. “The Gables” and “Gables”) which more likely represent the same properties. Note that we cannot guarantee that there are no incorrect matches, but the ordering and design of these approaches is designed to minimise mismatches.

Using this approach, we were able to match Assessors and EPC data for 1.45 million properties. We were also able to match to 780,000 property transactions in the Registers of Scotland data, some of which were multiple transactions for the same properties. We use properties transacting more than once to improve our estimation of our hedonic regression models (see next sub-section) but only the most recent observation for such properties when undertaking our final analysis.

We then use a UPRN-to-geography look-up file (and postcode-to-geography look-up file for properties without a UPRN) available from the Scottish Government to match properties to their intermediate and data zones, and their local area characteristics (such as deprivation, urban/rural classification and whether they are located on an island).

Hedonic regressions

We use a hedonic regression approach to estimate property values using property and area characteristics and the relationships between these and observed transactions values. Our hedonic regression includes the following property and area characteristics:

• The built-form of a property (e.g. whether it is detached or terraced).*
• The type of property (e.g. whether it is a flat, tenement or house of various kinds).
• The number of habitable rooms in a property, excluding kitchens and bathrooms.
• The size of a property in square metres.
• The number of fire places in a property.
• The type of power source used to heat a property.
• The type of windows a property has.
• The year in which a property was built (based on year bands).
• The council tax band a property is currently in.**
• The council a property is located in.*
• The data zone a property is in, provided at least 17 properties that we can match transacted in that data zone between 2015 and 2024, otherwise the intermediate zone a property is located in.
• The deprivation level of the data zone a property is located in as measured by the 2020 Scottish Index of Multiple Deprivation.
• The rural/urban classification of the data zone a property is located in as measured by the 2011 census.
• The output area classification of the output area a property is located in, based on the geographic and socio-economic characteristics of that area as of the 2011 census (an output area contained 20 – 77 households at the time of the census).

Variables with a * are interacted with either the year a transaction took place in (for built form) or the quarter and year a transaction took place in (local authority), to allow for different time trends for property values over the period of data used for our estimation (2015 to 2024). Council tax bands (labelled with **) are interacted with the council a property is located in to allow the effect of existing council tax bands on property values to vary by local authority.

The hedonic regression is run in log-log form, whereby the log of the property transaction price is regressed on the above variables in either indicator form (for example, separate dummy variables indicating whether a property has 1, 2, 3, etc. habitable rooms) or in log form (for example, log of the property size).

The number of data zones and various interaction effects mean there are thousands of explanatory variables in the regressions, with the coefficients on them capturing the effect of that variable on transactions values, conditional upon all other variables.

Our final specification uses 10 years of transactions data, and predicted values for Q3 2024. During the course of analysis, we tested the sensitivity of our results to the use of fewer years of transaction data (e.g. the last 4 or 5 years) and predicted values for an earlier period (Q2 2024). Overall, the results were similar. However, using fewer years of transaction data meant that we lacked sufficient matched transactions volumes for more data zones and so had to pool more data zones together into intermediate zones. This resulted in less spatially-precise estimates for value, especially for areas where intermediate zones cover large geographical areas. Estimated Q2 values differed somewhat by council compared to Q3 values, but did not differ so far as to indicate instability in our estimates. Q3 was therefore chosen as the most appropriate period to estimate values for, balancing being up-to-date, while still ensuring sufficient transactions had been reported in time for incorporation into the Registers of Scotland data provided to us for this project.

Predicted property values

The use of a log-log hedonic regression specification means that if we used the (exponential of) the central (log) predicted value for each property we would, on average, under-estimate property values. Mathematically, this is because the mean of the log value is not the same as the log of the mean value. So that our predicted property values do not, on average, under-estimate property values in this way, we draw five random error terms from the estimated error distribution for our hedonic regression. Our final predicted property values are then the average of these five predicted values with added errors. These error terms have mean zero in log terms, but a positive mean in cash terms, addressing the under-estimation problem.

Our approach bears some similarity with the approach the Value Office Agency (VOA) has taken for valuing properties in Wales. Here rather than take five random error draws, they average over the errors between the hedonic regression’s predicted value and actual transaction value for the five most similar properties that actually transacted. This approach allows for the hedonic regression to systematically under- and over-estimate values by different amounts for different kinds of properties. Our approach, based on random error draws does not allow for this. It was not feasible in the time or resources available for this project to adopt the approach taken by the VOA.

Sample weighting

While the hedonic regressions and predictions (with error) provide estimated Q3 2024 property values for the 1.45 million matched properties in our sample, these are not fully reflective of all properties in Scotland. In particular because EPCs are required for rental properties, rental properties are somewhat over-represented in this sample, and hence so too are lower-value and lower-band properties. For example, whereas a combined 58.4% of properties (excluding garages) are bands A – C according to the Assessors data, 61.9% of our matched sample are. We therefore weight our sample so that the number of properties by first part of a postcode (e.g. ED1), existing council tax band and property type (e.g. detached house, terraced house, flat, tenement), as well as by local authority and existing council tax band match that reported in the full Assessors data.

A.2 Household-level analysis

To undertake our household-level analysis, we use data from four waves (waves 8–10, and 14) of Understanding Society, a representative household panel survey. This covers households interviewed between 2016 and 2019, and between 2022 and 2023. We omit waves 11 to 13 which were adversely affected by the Covid-19 pandemic. Since it is a panel, there are some households that appear more than once, although we treat each household–wave observation individually. This gives us an initial sample of 7,577 household observations in Scotland.

In order to model reforms to council tax at the household level, we need (a) up-to-date property values, (b) current council tax bands and (c) council tax liabilities, taking into account council-specific tax rates, eligibility for discounts and exemptions (such as the single-person discount and student exemptions) and the CTRS. We do not model empty home discounts, as our data only capture information on primary residences, nor disability-related discounts, which cannot be identified in the data. We are also unable to model whether households meet asset requirements for CTRS due to a lack of information on assets in the Understanding Society data. For our main results we do not model the specific Bands E to H CTRS introduced in 2017, this decision is explained in more detail below.

The process for deriving up-to-date property values is described in detail at the end of this appendix.

We use linked administrative data to determine households’ current council tax bands. The Understanding Society data also contain self-reported council tax bands. However, we consider these to be less reliable than the council tax bands from the administrative data: they differ from the administrative data in around a third of all cases, and the distribution of self-reported council tax bands differs from the administrative data on all properties in Scotland. (Specifically, self-reports tend to overstate the share of properties in band D, which may reflect the fact that band D is the reference band and the band D rate is therefore often listed at the top of council tax bills).

Administrative data are not available for 23 of the households in our data. In these cases, we use the households’ self-reported council tax band. If we have no linked band or self-reported band, we impute their council tax band using their reported house value or rent, council and property characteristics. This is done using an ordered logistic regression, run separately for homeowners, private renters and social renters. For each tenure type, we regress administrative-linked council tax band on (log) self-reported house price or monthly rent (whichever is relevant), housing characteristics (house type interacted with number of rooms), location characteristics (rurality and Scottish Index of Multiple Deprivation (SIMD) decile) and which council the property is located in.^[21] We then randomly select a council tax band for those with missing values from the predicted probability distribution.

Because of the small sample size in Scotland, the imputation is done jointly for all of Great Britain, controlling for country and upper-tier council and allowing the effects of Index of Multiple Deprivation (IMD/SIMD) deciles to differ by country (because they are separately defined). The results are robust to alternative imputation methods, including an ordered probit regression and nearest-neighbour matching based on reported house value or rent, dwelling type, upper-tier council (in England, where some areas have two tiers of local government) and the number of rooms.

Table A.1 shows the distribution of council tax bands using different data sources. It shows that the distribution of council tax bands in Understanding Society (USoc), using linked administrative data and including imputations (row 4), closely matches the distribution of council tax bands in Scotland as a whole (row 1). We then further reweight our data so that they match exactly the distribution of council tax bands in the full administrative data. The final sample closely matches the (representative) overall USoc sample in terms of the distributions of income, local area deprivation (SIMD), age of oldest household member and household size. That said, the distribution by age of oldest household member differs from the distribution in other data sources (such as the Labour Force Survey and Family Resources Survey), with fewer younger households.

Table A.1. Distribution of council tax bands in different data sources (%)

Note: All Scotland figures are for 2024. They differ from the band distribution in the geographical analysis as some properties in the assessors’ data are not part of the tax base. USoc figures (2) to (4) are weighted using sample weights.

Source: Authors’ calculations using Scottish Government data (via statistics.gov.scot) and Understanding Society waves 8–10 and 14.

To calculate estimated council tax liabilities, the impacts of reforms are modelled using the IFS tax and benefit microsimulation model, TAXBEN. This contains council tax rates for each council, as well as information on the Scottish CTRS. We model reforms under the 2025–26 tax and benefit system, assuming that changes being phased in, such as the roll-out of universal credit, are fully in place.^[22] This allows us to estimate the long-run effect of revaluation and reform, once the roll-out of universal credit and other transitions under way are complete in the next few years. We drop 1,931 households with incomplete information on incomes and household characteristics. This leaves us with a final sample of 5,646 household observations in Scotland.

Assumptions on grant adjustment

As discussed in Section 3.2, the impact of revaluation and reform of council tax will depend crucially on whether and how grant funding for local councils is adjusted to reflect changes in the tax bases of different councils. We are unable to explicitly account for this as samples at the council level are too small to be properly representative. Instead, we adjust the council tax rates that all households in Scotland face by the same proportion so that reforms are revenue-neutral across Scotland as a whole. When tax rates are fairly similar across council areas, as is the case in Scotland, this approach will lead to estimates closer to what we would obtain if we were able to model full grant adjustment, rather than no grant adjustment.

Modelling the Band E–H council tax reduction scheme

When the Scottish Government increased the relative tax rates applied to properties in Bands E–H in 2017, they also introduced a specific CTRS for households with below-median income in these bands but not eligible for usual CTRS. For our main results we do not model this specific scheme because published figures imply very few households who are entitled to the scheme actually take it up. As a result, including the scheme in our modelling which assumes full take-up of CTRS would take us further from reality.

When we do model the scheme, we must make assumptions regarding how the system would operate under the various reforms. For the pure revaluation we assume eligible households in Bands E to H receive the same reductions as apply in the current system. For the 12-band reforms we assume that those in Bands E1 and above could potentially be eligible, with households in Bands E1 and E2 eligible for support to bring their bill down to what it would be under the pre-2017 Band E, Band F brought down to a pre-2017 Band F, Bands G1 and G2 brought down to a pre-2017 Band G, and Bands H and I brought down to a pre-2017 Band H. For the 14-band reform we similarly map bands to their pre-2017 equivalent, where G1 and G2 can be brought down to pre-2017 G, and Bands H and above can be brought down to pre-2017 H. For the continuous proportional system, we assign the potential support for each household based on which band they would fall into under a pure revaluation. For example, a household whose property value would put them in Band E under a pure revaluation would be eligible for support to bring their bill down to pre-2017 Band E.

Hedonic regressions for property values

The Understanding Society data contain self-reported property values for homeowners, which we uprate to 2024 Q3 using the HM Land Registry (2025) council-level House Price Index for the appropriate dwelling type (detached, semi-detached, terraced, etc.). This leaves us needing to estimate property values for renters.

To do this, we regress property values for homeowners on property characteristics (dwelling type, number of bedrooms and other rooms, existing council tax band), location characteristics (council, rurality, population density, Data Zone deprivation levels^[23]) and household characteristics (income, household composition and demographics^[24]). The estimated coefficients from this regression are then used to predict property values for renters. Note that the aim of this exercise is to predict property prices as closely as possible, not to model the price of specific housing amenities – it is not a ‘hedonic regression’ in the traditional sense of the term. As such, characteristics that do not directly affect property values but are nonetheless predictive of property values, such as household income and the number of children in the household, are included in the regression.

The regression explains 69% of the variation in property values for homeowners in Scotland. Property prices are regressed in log form. Regression coefficients for the main characteristics are listed in Table A.2. Dwelling type, number of bedrooms, number of other rooms and council tax band all have a statistically significant impact on price. Conditional on other characteristics, detached properties, those with more rooms, and those in higher council tax bands tend to be more valuable. To impute values for rental properties, a random error (drawn from the distribution of prediction errors among homeowners) is added to the predicted log property price, which is then converted back into pound values. This ensures we have an appropriate degree of variation in property values conditional on observed characteristics. To ensure that our results are robust to these random draws, we impute 20 property values for each household based on 20 randomly drawn error terms. The results we present are based on averaging 5 of these error terms four times for each household.

It is possible that the approach of imputing property values for renters based on a regression for owner-occupiers could lead us to overstate the values of rented properties, if they are systematically less desirable than owner-occupied properties with the same observed characteristics. This would in turn lead us to overestimate the council tax liabilities of households that rent after revaluation and reform.^[25] However, adequately controlling for unobserved differences would require advanced statistical techniques beyond the scope of this work.

Table A.2. Regression of log property prices: selected coefficients

Continues

Note: *** and ** indicate statistical significance at the 1% and 5% levels respectively.

Source: Authors’ calculations using Understanding Society waves 8–10, 14