RDP 2020-04: The Apartment Shortage Appendix A: Data

A.1 Prices

Table A1 shows details of the price estimates shown in Table 2. Specifically, it shows the effect various adjustments have on the number of sales and average prices.

CoreLogic data on sale prices are often reported separately for houses and units. Within the latter category we make an effort to exclude townhouses, villas, estates and other types of strata dwellings that have a substantial land component. Since data on these characteristics are not always available, we exclude buildings where at least 10 per cent of sales are labelled as ‘townhouse’, ‘triplex’, ‘quadraplex’ or ‘boarding house’. This is done for comparability with the ABS construction estimates which are for apartments. We additionally exclude some outliers and other implausible data entries, such as duplicate sales or sales occurring more than three years before the date of construction. We spot check these rules against photographs on real estate websites and they generally seem to rule in and rule out the right properties.

The profitability of supplying extra apartments is the difference between the cost and price of new dwellings. Accordingly, we exclude properties sold more than five years after construction. This filter is perhaps the most important step in Table A1 and we discuss its implications in Appendix E.

A small proportion of sales are anomalously high – in the tens or hundreds of millions of dollars – even though the building characteristics and location are little different from nearby apartments. We expect this occurs when an entire building is sold and its price is entered for individual apartments. To protect against data entry mistakes like this we exclude the top and bottom 1 per cent from our sample. In comparison, CoreLogic winsorise the top and bottom 5 per cent from many of the variables entering their indices.

Even after trimming, sale prices are heavily skewed. The median new sale price in Sydney or Melbourne is 11 per cent lower than the trimmed mean. Although some other research focuses on median housing prices, the mean is appropriate for calculating excess profits. Moreover, our cost estimates are only available on an average basis and presumably reflect the same skew, so consistency requires taking the difference between averages, not medians.

Finally, we multiply these estimates by the change in CoreLogic's unit sales price index for each city, to express in 2018 prices. This increases prices in Sydney and Melbourne and decreases them in Brisbane.

A troubling feature of our data is that the number of new apartment sales sometimes differs substantially from the number of apartment completions, especially at the end of our sample. We assume that the discrepancies between sales and completions are not systematically related to prices but were not able to verify this.

A.2 Unpublished Cost Estimates

Given that our estimates of average construction costs from the ABS Building Activity Survey are unpublished, Table A2 presents some summary statistics for the data, which may be of interest. Multiplication of units per storey by gross floor area per unit, and assuming that floor area per storey is constant, provides an estimate of building footprint, used in Section 7.

Figure A1 shows unpublished ABS estimates of building heights by year. As can be seen, these have fluctuated about rising trends. Measuring cost at actual building heights would result in transient movements in our cost estimates. This volatility does not seem relevant to building decisions, which are based on expected, rather than historic costs. Instead, we value both average and marginal cost at the trend building height. For Section 5.2, we extrapolate the estimated trends from 2003, when the height data begins, back to 1997. An alternative approach of holding each city's average apartment height constant at its estimated 2003 level makes little difference.

A.3 Government Charges

Charges for infrastructure and public goods are a private cost, but it is debateable whether they should be counted as a social cost. Assuming that planning regulations do not change overall population, an increase in infrastructure use in one area will mean a reduction in infrastructure use in the areas from where the new residents come. It seems inappropriate to include the extra use as a cost without also allowing for the offsetting savings elsewhere.

Our estimates come from Urbis (2011) and CIE (2011), which are the most recent estimates of which we are aware. Developers tell us that government charges have increased substantially since these estimates were published. In particular, Voluntary Planning Agreements (including for parks and affordable housing) often increase costs by more than the charges we allow for. Moreover, developers also suggest that current charges are much greater than needed to fund marginal increases in infrastructure and that they represent a large element of value capture.

Government charges are not a major cost, so a simple approach is to use the available published estimates. This judgement recognises that the considerations mentioned above are difficult to quantify and that some would imply higher estimates and some smaller.

A.4 Comparison to Other Cost Estimates

Table A3 compares our estimates of construction costs with estimates from Kendall and Tulip (2018), Urbis (2011) and CIE (2011). Each of these provide an estimate of the average construction cost of a typical Sydney apartment, shown in the first row. Definitions for these estimates differ and the subsequent rows attempt to include various components so that the alternative estimates represent the same concept. These estimates, shown in the final row, are of the average supply cost excluding the cost of land, finance and developer's profit.

Allowing for conceptual and timing differences, the ABS construction cost estimates are slightly smaller than the RLB-based estimates used by Kendall and Tulip (2018), offset by inclusion of infrastructure contributions, discussed in Appendix A.3. The ABS-based estimates are noticeably lower than those of Urbis (2011) and CIE (2011), perhaps reflecting the more representative ABS sample.