Cannabis
The estimates stay close.
Around 4% lower hourly pay with either approach: −4.2% ordinary and −4.0% interviewer-based.
Interviewers as instruments
estimating the wage penalty of drug use
Journal of Population Economics · Accepted / forthcoming
A trace in survey reporting offers a second way to estimate the link between drug use and pay.
95 seconds · Narrated · English captions
01The question
Does drug use affect wages?
Ordinary comparisons are hard to interpret. People who use drugs differ in other ways, use is self-reported, and wages are observed only for people in paid work.
Australia · 2017 & 2021HILDA, the Household, Income and Labour Dynamics in Australia Survey.
02The idea
A second source of evidence.
The drug questions were completed privately, on a separate questionnaire. Interviewers did not ask them. Yet reported use varies with the household’s interviewer. The study uses this trace to construct an instrument.
Ifassignment is quasi-random
Within the same area and year, interviewer–household matching is as good as random after accounting for observed characteristics.
Ifthe exclusion assumption holds
Interviewer identity affects wages only through reported drug use.
These are identifying assumptions. The data do not prove them.
03What changes?
Some estimates hold.
Others diverge.
Estimated differences in hourly pay.
Monthly-or-more use, compared with no reported use.
Ordinary comparisonInterviewer-based
Methamphetamine
The estimate changes sign.
Suggestive, not definitive.Romano–Wolf adjusted p = 0.058, narrowly above the usual 5% threshold.
Inhalants
The estimated penalty doubles.
From −3.4% ordinary to −6.8% interviewer-based, for monthly-or-more use.
Wage and salary employees, HILDA 2017 and 2021. Ordinary comparison: weighted least squares (WLS). Interviewer-based: instrumental variables (IV). Reference: no reported use of the same substance in the last 12 months. The adjusted p-value tests both frequency coefficients jointly, across six substance-level hypotheses.
04Why it matters
No single universal
drug–wage effect.
Both the substance and the source of evidence matter. Some estimates remain close across approaches; others change substantially.
Methodologically, interviewer-related variation in reporting can provide useful information under the stated assumptions. The study does not establish why wages differ.
05The paper
Interviewers as instruments:
estimating the wage penalty of drug use
JournalForthcoming
Post-printPending final PDF
Watch film ↑95 seconds
Download film ↓MP4 · 36.9 MB
CitationAccepted / forthcoming
06Authors
The researchers
07Read further
Transcript & details
Film transcript
Every year, the HILDA Survey asks thousands of Australians about their lives.
Some questions, they answer alone.
One is about drugs.
For decades, economists have asked whether drug use costs people at work.
But people who use drugs differ in other ways. Some don’t say. And only workers have wages.
Every household also meets an interviewer.
They don’t ask about drugs. Yet what people report shifts with who the interviewer was.
Survey researchers call that error.
This study uses it.
Within an area and year, who gets which interviewer comes down to fieldwork rosters.
If that’s as good as random, and interviewers affect pay only through reporting, the trace becomes a lever:
a second way to ask the same question.
For cannabis, both ways agree: about four percent lower hourly pay.
For frequent methamphetamine use, they split: slightly higher pay, if anything, one way; almost ten percent lower the other.
Corrected for testing six drugs, it narrowly misses the usual bar: a lead, not a verdict.
For frequent inhalant use, the estimated penalty roughly doubles.
Six drugs. No single answer.
Some findings hold either way. Others depend on where the evidence comes from.
The noise was never just noise.
Method & interpretation
The paper constructs an optimal formula instrument from interviewer-driven predicted use, with permutation recentring within Statistical Area Level 3 (SA3) × year clusters, residualisation and inverse-variance weighting. Estimation uses weighted limited-information maximum likelihood and a Heckman-type correction for selection into wage or salary employment.
- Assignment is assumed to be as good as random within area–year, conditional on observed characteristics. This is not a randomised experiment.
- The exclusion assumption requires interviewer identity to affect wages only through reported drug use. Among non-users, interviewer effects on wages are jointly significant (p = 0.043), but add negligible explanatory power. Exclusion remains an assumption.
- The estimates are local to respondents whose reporting responds to interviewer assignment, under further assumptions including monotonicity and the selection model. These respondents may differ across substances.
- Frequent use of several substances is rare, and some estimates are imprecise. The outcome is hourly pay among employees; the study does not identify effects on employment, hours or total earnings.
Displayed percentages are 100 × (exp(β) − 1), from Table 8. Frequent use means monthly or more in the last 12 months. Each comparison uses no reported use of the same substance as its reference.
For frequent methamphetamine use, the IV estimate is −9.6%, with a pointwise, unadjusted 95% interval of −18.0% to −0.3%. The ordinary estimate is +2.4%, with an interval including zero. The Romano–Wolf adjusted p = 0.058 is for the substance-level joint test of both frequency coefficients, controlling family-wise error across six substances (Table 9). It is not a p-value for the frequent-use coefficient alone.
The needles show point estimates on a common scale, with no uncertainty intervals drawn. They do not establish a mechanism or prove causality.
Citation
Alexeev, S., & Wooden, M. (forthcoming). Interviewers as instruments: estimating the wage penalty of drug use. Journal of Population Economics.
Journal link and final post-print will be added when available.
Captions & accessibility
The film is narrated in English. English captions are enabled by default; use the video player’s captions control to turn them off or on. The full transcript is available above.
The lights, household groups and interviewer routes are schematic illustrations, not identifiable survey participants or reconstructed participant data. The three needle pairs show the paper’s estimates.