RDP 2019-12: Confidence in Australian Banknotes 3. Empirical Strategy

Kim Nguyen and Anny Francis

December 2019

3.1 Survey Data Regressions

We first focus on identifying how demographic factors, personal experience with counterfeits, personal experience using banknotes, and awareness of banknote upgrades and banknote security features affect confidence in banknotes. In particular, we use answers to the three survey questions described in Section 2.3 as indicating a respondent's level of confidence, and we aim to test whether the aforementioned factors increase the perceived likelihood of receiving a counterfeit banknote within the next year, the belief that there is a counterfeiting problem in Australia, and the overall confidence in the system to remove counterfeits from circulation.

Equation (1) is the general form of the empirical models used in the paper:

1 f( Confidenc e i )=α+ β ′ Factor s i + γ ′ Control s i + θ ′ Yea r i + ε i

The first measure of confidence is the ‘perceived likelihood of receiving a counterfeit’, where respondents were asked to rate their perceived risk on a scale of 0 to 10. We could treat this measure as either an ordinal variable (that is, a variable for which the order has meaning but the magnitude does not, so that a response of ‘2’ is interpreted as being larger than a response of ‘1’, but not twice as large as a response of ‘1’), or as a binary variable by collapsing the answers into two categories: likely to get a counterfeit in the next year (rating of 5 to 10) and not likely (rating of 0 to 4). An ordered logit model would be appropriate for an ordinal dependent variable of this kind, while for the second approach one would typically use a linear probability model (LPM) or standard probit and logit models. The question of whether Australia has a counterfeiting problem provides for only yes/no responses, so an LPM, probit or logit model would be appropriate. The question on an individual's confidence in the system to remove counterfeits also has a number of possible answers, and we again construct a binary variable: the ‘not at all confident’, ‘not very confident’, and ‘neither [confident nor not confident]’ responses are classified as 0 (not confident), while ‘fairly confident’ and ‘very confident’ are classified as 1 (confident). As a robustness check, we classify ‘neither’ separately in an ordered logit model (see Appendix A).

Throughout the paper our preferred modelling technique is to use a standard probit regression, as the output from such models is generally relatively easy to interpret. The complete set of robustness checks using other models and specifications can be found in Appendix A. In each regression, we control for the survey year, residential location of the respondent, and other cash-related questions asked in the surveys.[1] For the sake of clarity, we only present factors that are of interest in the regression tables.

3.2 Media Reporting Regressions

To empirically study how the media influences confidence in banknotes, we turn to the second dataset, which includes monthly state-level variables. We use fixed-effect panel regressions to study the effects of incidents and media reports on our dependent variable of genuine banknotes mistakenly submitted as counterfeits. The underlying assumption is that an increased number of genuine banknotes mistakenly submitted as counterfeits would, all else equal, indicate less confidence in banknote security. The form of the regressions is:

1 Genuine s s,t = β ′ Reported Incident s s,t + δ ′ Report s s,t + γ ′ Control s s,t + α s + θ t + ε s,t

The entities are states in Australia (denoted as s) while time is measured in months (denoted as t). The state fixed effect, given by α s , controls for all time-invariant differences between states. The time effect θ t accounts for variation over time that is the same across states. We also control for other time-varying differences between states such as the number of actual counterfeit detections, the quality of banknotes in circulation, population, employment rates, earnings per hour, inflation, and the total number of criminal proceedings.

Footnote

Cash-related questions include whether the respondent uses cash daily as part of their job and whether a cashier or bank teller has ever checked their banknotes. [1]