Why the CPI matched models method may fail us: results from an hedonic and matched experiment using scanner data

Statistical offices use the matched models method to compile consumer price indices (CPIs) to measure inflation. Price collectors record the prices of a sample of models or varieties of selected products existing in period t, and then continue to record the prices of these same matched models in subsequent periods t+n. The matched models method is designed to control for quality changes by ensuring that like is always compared with like. But new, unmatched models launched in periods t+1,.. t+n have their prices ignored unless they are used as replacements. The matched models method also fails when an old model is no longer available and a matched comparison of like with like cannot take place. In markets where models change rapidly due to technological innovations the matched sample may, by the end of a year, be quite unrepresentative of what is, and what was, bought. We show that serious sample degradation is a necessary condition for the matched approach to fail and then consider how the prices of unmatched models differ from matched ones and estimate their impact on the index. The paper thus argues that a form of sampling bias may exist as a by-product of the method used to control for quality changes. The bias is one not recognised in major studies of CPIs including Boskin (1996). An alternative approach to measuring quality adjusted price changes is the use of hedonic indices which can use data on matched and unmatched models. This study is in part motivated by evidence of differences between the results of the two approaches. A replication of CPI procedures is attempted on scanner data. Bar-code scanner data is particularly suited to this purpose covering the universe of transactions. The data allows us to identify the extent of sample degradation, compare unmatched (new and old) prices and matched prices and their effects on price measurement and demonstrate the effects of more frequent sample rotation, such as chaining. Hedonic indices can use the whole sample and chained indices refresh the sample on a regular basis. They are argued to be more useful to price measurement in markets with a rapid turnover of models. JEL classification: C43, C81, D12, E31, L15, L68, O47