PreprintPDF Available

Addiction and gender norms

Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

There is a claim that households where a female earns more than a man violate gender norms, which (through the increased stress) causes partnership dissolution and domestic violence towards females. This paper tests the immediate implication of the stress -- a higher prevalence of substance abuse. There is no evidence for the female or male breadwinner effect on own or partners addiction using a large nationally representative dataset of Australian mixed-gender couples and applying regression discontinuity (RD) design, mixed-effects methods, and a variation of Bartik instrumental variable (IV). The effect is present only when parametric RD is applied; however, the paper shows that this method is inappropriate for the Australian data, and the correct nonparametric approach shows no effect. The paper then reassesses the effects on marital satisfaction. The preferred IV estimates show that males are more satisfied with breadwinning female partners. Arguably because, unlike females, males are not expected to compensate lower earnings with housework. Therefore, if the income gender norm is present, it goes in the opposite direction, and the notion of gender norms in joint earnings for developed countries needs revision.
Content may be subject to copyright.
Addiction and gender norms
Sergey Alexeev
University of New South Wales, NSW, Australia
December, 2021
Abstract: There is a claim that households where a female earns more than a man violate gender norms,
which (through the increased stress) causes partnership dissolution and domestic violence towards females.
This paper tests the immediate implication of the stress – a higher prevalence of substance abuse. There is
no evidence for the female or male breadwinner effect on own or partners addiction using a large nationally
representative dataset of Australian mixed-gender couples and applying regression discontinuity (RD)
design, mixed-effects methods, and a variation of Bartik instrumental variable (IV). The effect is present
only when parametric RD is applied; however, the paper shows that this method is inappropriate for
the Australian data, and the correct nonparametric approach shows no effect. The paper then reassesses
the effects on marital satisfaction. The preferred IV estimates show that males are more satisfied with
breadwinning female partners. Arguably because, unlike females, males are not expected to compensate
lower earnings with housework. Therefore, if the income gender norm is present, it goes in the opposite
direction, and the notion of gender norms in joint earnings for developed countries needs revision.
JEL codes: J12, J16, I31, Z13, I18
Key words: Marital Satisfaction, Economics of gender, Social norms, Earnings differentials, Addiction
Declarations of interest: none
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
The funding sources had no involvement in the conduction of the research and/or preparation of the article; in the collection,
analysis and interpretation of data; in the writing of the report; and in the decision to submit the article for publication.
1 Introduction
Bertrand, Kamenica, and Pan (2015) show that in the US, from the late 1960s throughout the early
2000s, female breadwinning is negatively associated with marital satisfaction and partnership stability.
They explain this with a general public expectation that males should earn most of the family income,
and a violation of this norm creates stress for both partners. Foster and Stratton (2021) reassess these
conclusions using Australian and American national panels covering 2001 to 2016 and show that the link
between female breadwinning and marital health in the most specification is absent in Australia and
became much weaker for the US. They conclude that this results from a gradual decline in the male
breadwinning norm. An opposite conclusion is made by another recent Australian study by Zhang and
Breunig (2021). They argue that the male breadwinning norm is still strong in Australia. They also
use Australian nationally representative data for 2005-2016 and show that female breadwinning results
in a staggering 35% increase in the likelihood of partner violence and a 20% increase in emotional abuse
against women. These opposing conclusions about the breadwinning role are surprising because both
2 DATA 2
utilize nationally representative data for the same country and period. One apparent distinction is that
Foster and Stratton (2021) uses panel data and employs the fixed effect (FE) approach, whereas Zhang
and Breunig (2021) uses independent cross-sectional data and rely solely on RD.
This study contributes to this debate in multiple ways. If female breadwinning causes excessive stress
in males or females, stress mitigating behavior should also be observed. Substance abuse is the most
well-documented way to escape stress. Therefore, the work seeks to estimate if female breadwinning
causes addiction. In a narrow technical sense, this work extends Foster and Stratton (2021) to newer
outcomes using more recent data and broader empirical approaches. A vast selection of empirical strate-
gies is explicitly utilized, as the contradictory conclusions of the breadwinning norms in contemporaneous
Australia are puzzling, and methodological inconsistencies are suspected of playing a role. For all our
strategies, I do not see female breadwinning increasing or reducing either partners’ drinking, smoking,
or illegal drug usage. I then reassess the marital dissatisfaction that should pick up presumed domestic
violence or stress. Instead of detecting male dissatisfaction with a female breadwinner, I detect the op-
posite. The males are more satisfied when their female partner is the breadwinners. Taken together, my
finding confirms the conclusions in Foster and Stratton (ibid.) that gender income roles in modern days
Australia are weak and that the results reported in Zhang and Breunig (2021) might be spurious and
warrant further scrutiny.
2 Data
The study uses 17 waves of HILDA survey pooled together covering 2001–2017. The dataset consistently
reports drinking intensity and drinking frequency for all years. Following Alexeev and Weatherburn
(2021), I recode these two variables into one polychotomous variable of average daily consumption of
standard drinks (about 10 grams of alcohol). Abstinence is coded as zero. Smoking is also reported
annually and captured with a dummy variable. The measure of partner’s satisfaction on the scale of 0
(the least satisfied) to 10 (the most satisfied) is also available for all waves and respondents.
The information on drug use is available in 2017. The respondents are asked if they ever used drugs
2 DATA 3
and first and last use age. I use this information to identify drug users in all other waves, accounting for
the declared age of first and last use. For example, in 2017, a respondent reports that (s)he has smoked
cannabis from age 18 to 26, I can identify the same person in, say, 2011, and if (s)he is, for example, 17
years old at this time (which is outside the age range of the reported smoking behavior), I flag this person
as a non-smoker. In the next year (i.e., 2012), when (s)he 18, I flag him(-er) as a cannabis smoker. This
approach creates panel data and thus permits penal data techniques but potentially introduces additional
measurement errors. Some individuals from early waves have left the survey before reporting their drug
use pattern in 2017. To address this problem, I confirm that the findings stay unchanged when I only use
waves 2016 and 2017. As a separate question in 2017, Australians are also asked if they do drugs now. I
also use this cross-sectional information to validate our panel data findings on drug outcomes.
Age, schooling, various measures of income, aboriginal status, and the country of birth are also
available for everyone for all waves. Breadwinner – the main focal variable – is defined as an indicator
function if a respondent brings more than a half of joint household income. The household income
is defined as the total income from the respondent and the partner. Individual labor earnings and
total household income are also used as control variables. However, these variables are heavily skewed
to the right, and many individuals report zero earning. To stabilize the skewed data for linear data
modeling and preserve observations, I apply inverse hyperbolic sine (IHSF) parameterization to earnings
and income. Except for minimal values, the IHSF is identical to a standard log transformation, but
it is defined for values close to zero and performs better than the known alternative transformation
techniques (e.g., Burbidge, Magee, and Robb 1988; MacKinnon and Magee 1990).1drinking would also
benefit from applying IHSF because heavy drinking is relatively rare, and most of the values for drinking
are less than 1. But I prefer unmodified outcome dues to concerns that IHSF of the dependent variable
misrepresents the condition mean function (e.g., Bellemare and Wichman 2020). However, our results
are not qualitatively different if IHSF transformation is applied to drinking or not.
All respondents are matched with their partners using unique partners identifiers. As in Foster and
Stratton (2021) the data is then limited to only couples with mixed gender and where both partners
1IHSF is defined as sinh1(𝑥) := ln(𝑥+1 + 𝑥2). Its derivative is 1
/1+𝑥2which if 𝑥is not too small approximates
/𝑥, the derivative of ln(𝑥).
2 DATA 4
Table 1: Descriptive statistics
Male Female
Variable Obs Mean Std. Dev. Min Max Obs Mean Std. Dev. Min Max
Smoking 55,223 0.190 0.392 0 1 55,223 0.156 0.363 0 1
Drinking 45,168 1.308 1.749 0 13 47,868 0.567 0.993 0 13
Cannabis 38,807 0.166 0.372 0 1 40,998 0.110 0.313 0 1
Other drugs 38,490 0.067 0.249 0 1 40,754 0.041 0.197 0 1
Satisfaction 45,273 8.386 1.792 0 10 48,069 8.192 1.951 0 10
Breadwinner 55,223 0.770 0.421 0 1 55,223 0.230 0.421 0 1
Own labor earnings 55,223 10.261 3.850 0 14.385 55,223 8.488 4.697 0 14.385
Partner’s labor earnings 55,223 8.488 4.697 0 14.385 55,223 10.261 3.850 0 14.385
Total HH income 55,223 12.013 1.205 -14.126 15.609 55,223 12.013 1.205 -14.126 15.609
All income 55,223 0.119 0.323 0 1 55,223 0.025 0.157 0 1
No income 55,223 0.025 0.157 0 1 55,223 0.119 0.323 0 1
Own age 55,223 40.429 10.143 18 58 55,223 38.319 9.971 18 58
Partner’s age 55,223 38.319 9.971 18 58 55,223 40.429 10.143 18 58
Own schooling 50,772 12.821 2.346 0 18.5 52,956 12.858 2.388 0 18.5
Partner’s schooling 52,956 12.858 2.388 0 18.5 50,772 12.821 2.346 0 18.5
Aboriginal 55,223 0.017 0.128 0 1 55,223 0.022 0.146 0 1
Aboriginal partner 55,223 0.022 0.146 0 1 55,223 0.017 0.128 0 1
Australia born 55,223 0.718 0.450 0 1 55,223 0.757 0.429 0 1
Australia born partner 55,223 0.757 0.429 0 1 55,223 0.718 0.450 0 1
Source: HILDA 2001-2017
Figure 1: Reduction in drinking and gender gap
Notes: Distribution of relative income. The sample includes self-reported mixed-gendered partnerships where both partners
are aged 18–65. Each bar captures a 0.05 relative income bin.
Source: HILDA 2001-2017
are within age 18-58. The data is then divided into male and female samples. Note that the data is
structured to show the breadwinner effect on the breadwinner, but because male and female samples are
looking at the same couple, the data can also show the effect on the breadwinner partner but with the
opposite sign. I will return to this point once again. Table 1reports the descriptive statistics. Figure 1
presents the relative income distribution within the partnerships from the perspective of both genders.
The figure reveals the discontinuous pattern at equal shares of total couple income that has been found
in other papers for Australia and other countries.
3 Methods
3.1 Econometric model
Consider the following liner specification that relates the individuals’ breadwinner status to the individ-
uals’ outcomes of interest.
𝑌𝑖𝑠𝑡 =𝛽Breadwinner𝑖𝑠𝑡 +𝑋𝑖𝑠𝑡 ·𝛿+𝜆𝑠+𝛾𝑡+𝜀𝑖𝑠𝑡.(1)
Here 𝑌𝑖𝑡 is an outcome of respondent 𝑖in wave 𝑡and in area 𝑠. The variable Breadwinner𝑖𝑠𝑡 is a dummy
if 𝑖in 𝑡and 𝑠owns most income in the partnership. The equation includes a vector of control variables
𝑋𝑖𝑠𝑡, which is comprised of IHSF of individual labor earnings of both partners, IHSF of total household
income, quadratics in each partner’s age, schooling of both partners, each partners’ dummies for being
aboriginal, for being born in Australia, for being no or all income earner. These variable are chosen to
match Foster and Stratton (2021) as close as possible. This choice, in turn, matches the variables used
in Bertrand, Kamenica, and Pan (2015). Variable 𝛾𝑡stands for time effects; 𝜆𝑠for the SA4 area effects
and 𝜀𝑖𝑠𝑡 is the econometric error.
The parameter 𝛽is of interest. It shows the mean difference in outcomes between breadwinners and
non-breadwinners. When the model is estimated on a sample of females matched with the males, the
parameter shows the effect of the violation of the income gender norms on the female’s outcomes. When
the model is estimated on a sample of males matched with females, the parameter shows the effect of the
income gender norm compliance on male outcomes.
Since both samples look at the same mixed gendered couples and the parameters show the mean
difference, the estimands are also informative about the effects on the breadwinner’s partner. For example,
I show that if the male is a breadwinner, he is less satisfied with his female partner. Note that a tautology
to this is that the male is more satisfied when the female is the breadwinner. This shows in the model. I
get the same results (but with the opposite sign) if I instead estimate the effect of the female breadwinner
on the marital satisfaction of this female’s male partner.
Since the same person is observed over time, standard errors are clustered on a partnership level, and
a robust covariance matrix is used Model (1) is estimated with a large number of different methods that
I now discuss.
3.2 Estimation methods
Model (1) estimated with ordinary least square (OLS) delivers causal results under the RD assumptions.
The couples around the equal income split are identical, and there is no preemptive self-selection into those
shares. Even if those assumptions hold (at least conditionally as treated by the previous author), linear
parametrization is another assumption. If the true relationship is non-linear, the model may misrepresent
the treatment effect. Therefore, I also estimate the model with polynomial RD point estimators with
robust bias-corrected confidence intervals and inference procedures developed in Calonico, Cattaneo, and
Farrell (2018) and Calonico, Cattaneo, and Titiunik (2014). The estimation is implements with the
package described in Calonico, Cattaneo, Farrell, and Titiunik (2017).
As an alternative to OLS, I also follow Foster and Stratton (2021) and include the household FE. This
method controls for unobserved time-fixed confounders. One drawback of including household fixed effects
is that the breadwinner is a time-invariant dummy variable by contraction. The identifying variation
comes from the change of the individual’s breadwinner status over time. This highlights another related
problem. Theoretically, the interest is in the permanent long-term income, which is not directly observed.
The observed income has a transitory component that behaves differently depending on the partners’ and
partnership age. This misclassifies some breadwinners.2In the FE approach, this misclassification feeds
2Hypothetically, one could follow from income mobility literature (e.g., Grawe 2006) and look at the partnerships when
both partners achieve their highest income potential (the late 30s or early 40s in developed economies). Then the transitory
into the sources of identification. For example, a person could be a breadwinner for some years because
the partner is finishing school, but the breadwinner may change after the school completion. Controlling
for age alleviates the problem, but FE can not incorporate own or partner’s age if year effects are added
(aboriginal status or place of birth also drop out with FE, but these are time-invariant characteristics). I
also estimate our model with correlated random effects (CRE) to address these concerns. This approach
combines random and FE and somewhat improve upon FE by allowing the inclusion of time-invariant
variables. CRE is essentially a random effect model, but instead of assuming that random effects are
exogenous, it assumes that invariant unobservables follow a linear relationship (e.g., Wooldridge 2011).
The mixed-effects techniques are susceptible to simultaneity and defenseless if partners self-select each
other based on their views on the importance of the gender role in the household. Ignoring this self-select
nature of observed partners’ income shares, the regression of outcomes showing the breadwinner effect
would be biased, regardless of whether the regression model accounts for all relevant covariates using
FE or any other technique. To overcome this and other potential sources of endogeneity, I also follow
Bertrand, Kamenica, and Pan (2015) and instrument the focal variable in Model (1) with potential value
rather than the realized one. This is the approach originally suggested by Bartik and Bartik Timothy
(1991) and it entails characterization of local labor market conditions adjusted for gender, age, and
education and using that as an IV for breadwinning status (e.g., Goldsmith-Pinkham, Sorkin, and Swift
2020). An apparent limitation of this approach is that locations are not random. Therefore I use the
survey interviewers.
Because the survey is a multi purse, the interviewers’ assignment is not conditioned on any particular
household characteristic. Therefore, using an average (and demographically adjusted) value reported to
the interviewer as an IV is possible. One concern is that HILDA is a two-stage survey where the random
territories are selected, and families are chosen from those areas. Thus, two sampled households within
a cluster might be served by the same pool of interviewers. Fortunately, this is already corrected by
including the SA4 area fixed effect.
The following procedure is performed separately for the male and female samples to construct the
component is most similar to classical measurement errors, and the permanent income can be measured by averaging income
over several years. This, however, is not the currently accepted practice for the research question at hand.
average value reported to the interviewer. Let the residual of focal variable after removing the interacted
SA4-interviewer fixed effects, 𝜆𝑠×𝜔𝑗, and the control variable vector 𝑋𝑖𝑠𝑡 be denoted by
𝑖𝑠𝑡 =Breadwinner𝑖𝑠𝑡 𝑋𝑖𝑠𝑡 ·𝜅𝜆𝑠×𝜔𝑗=𝑧𝑖𝑠𝑡 +𝜌𝑖𝑠𝑡.(2)
Here, the new subscribe 𝑗refers to an interviewer, and 𝜔𝑗is interviewer fixed effects. Note that without
adding SA4 fixed effects and controls, the IV is just an average of the focal variable reported to an
interviewer. Additional interaction by statistical area corrects for potential non-randomness of interviewer
within the statistical areas, and the control variable adjusts the residual for the observables.
These adjusted residuals are then used to construct the IV
𝑧𝑖𝑠𝑡(𝑗):= 1
𝑛𝑗𝑛𝑖𝑗 𝑛𝑗
Here, 𝑛𝑗is the total number of respondents seen by interviewer 𝑗and 𝑛𝑖𝑗 is the number of occasions
respondent 𝑖is seen by 𝑗. The first summation inside the brakes refers to all residualized focal variables
reported to 𝑗; the second summation leaves out the variable of 𝑖. Effectively, Equation (3) removes the
residual 𝜌𝑖𝑠𝑡 from Equation (2) by aggregating the residualised measure to the 𝑗cell level and excluding
the respondent whose focal variable is being instrumented.
The residualized leave-out IV from (3) is then used within the two-stage least squares (2SLS) frame-
work where the second stage is as in (1), and the first stage is the projection of the endogenous variable
with all exogenous variables used in Model (1) into the IV space.
For the IV to be valid, it must be uncorrelated with the household observed and unobserved char-
acteristics. Table 2verifies that the IV is random. Column (1) regresses the focal variable on control
variables (the area and year effects are included but not shown). The controls are highly correlated with
the status of the breadwinner. On the contrary, Column (2) shows that the IV is independent of the
controls. The estimates are all close to zero, with only one exception. Given the number of variables,
having one weakly significant coefficient is not unusual since the probability of observing one 𝑝-value at
this level by chance alone is large. Importantly the variables are not jointly significant. The same holds
Table 2: The IV randomness
Male sample Female sample
(1) (2) (3) (4)
Breadwinner IV Breadwinner IV
Own labor earnings 0.320*** -0.00831* 0.132*** -0.00182
(0.0112) (0.00347) (0.00500) (0.00153)
Partner’s labor earnings -0.132*** 0.00129 -0.320*** 0.00863*
(0.00500) (0.00152) (0.0112) (0.00351)
Total HH income -0.135*** 0.00333 0.135*** -0.00289
(0.0181) (0.00549) (0.0181) (0.00551)
All income -0.0703 0.00157 0.684*** -0.0454
(0.0853) (0.00882) (0.138) (0.0454)
No income -0.684*** 0.0462 0.0703 -0.00158
(0.138) (0.0452) (0.0853) (0.00866)
Own age 0.00569 0.000728 0.00990** -0.00147
(0.00307) (0.000986) (0.00320) (0.00130)
Own age squared -0.0000306 -0.00000344 -0.0000860* 0.0000232
(0.0000361) (0.0000120) (0.0000397) (0.0000166)
Partner’s age -0.00990** 0.00136 -0.00569 -0.000371
(0.00318) (0.00127) (0.00308) (0.00103)
Partner’s age squared 0.0000860* -0.0000226 0.0000306 0.000000389
(0.0000396) (0.0000163) (0.0000362) (0.0000124)
Own schooling 0.00764*** -0.000522 0.0169*** 0.00115
(0.00159) (0.000596) (0.00167) (0.000691)
Partner’s schooling -0.0169*** -0.00112 -0.00764*** 0.000535
(0.00166) (0.000683) (0.00158) (0.000620)
Own aboriginal status -0.00812 -0.00904 -0.00985 0.00589
(0.0246) (0.00735) (0.0201) (0.00408)
Partner’s aboriginal status 0.00985 -0.00395 0.00812 0.00840
(0.0198) (0.00428) (0.0245) (0.00765)
Born in Australia 0.0184* 0.00393 0.0109 0.000941
(0.00870) (0.00347) (0.00910) (0.00359)
Partner born in Australia -0.0109 -0.00148 -0.0184* -0.00546
(0.00911) (0.00360) (0.00870) (0.00352)
P-value on joint significance <0.000 0.158 <0.000 0.117
Observations 37,945 37,561 37,945 37,553
Notes: Table demonstrates the non randomness of the focal variable, but randomness of the IV.
All income measure are in logs. Year and area fixed effects included but not shown.
*𝑝 < 0.05, ** 𝑝 < 0.01, *** 𝑝 < 0.001.
Source: HILDA 2001-2017
for the female sample. This provides evidence that interviewers are random and, therefore, the IV is also
independent of the unobserved characteristics.
Interpretability of the 2SLS estimand as LATE relies on a monotonicity assumption. The IV design
compares groups of otherwise similar individuals (a mix of IV always takers and compliers and a mix of IV
never takers and compliers) who have reported a breadwinner status to a randomly assigned interviewer.
Therefore, the design must ensure the absence of the IV defiers. Table 3reports the regression of the
breadwinner status on the IV (controlling for all covariates including year and area fixed effects) when the
sample is limited to different age groups. The presence of the IV defier within the age subsample can be
detected if the first stage changes the sign for that subsample. No defiers can be seen. For all subgroups,
Table 3: The IV monotonicity
Male sample Female sample
Subsample (1) N (2) N
Age 25-30 -0.114*** 4,473 -0.135*** 4,924
(0.0279) (0.0310)
Age 30-35 -0.0895** 4,934 -0.0654* 5,067
(0.0297) (0.0330)
Age 35-40 -0.0917* 4,781 -0.116*** 5,110
(0.0403) (0.0281)
Age 40-45 -0.0533 5,043 -0.0942** 5,195
(0.0382) (0.0338)
Age 45-50 -0.0843* 4,986 -0.0961*** 4,849
(0.0340) (0.0273)
Age 50-55 -0.112*** 4,393 -0.143*** 3,923
(0.0290) (0.0368)
Notes: Table reports the first stage on age subsample and
demonstrates absense of the IV defier. All controls including
area and year effects are present.
*𝑝 < 0.05, ** 𝑝 < 0.01, *** 𝑝 < 0.001.
Source: HILDA 2001-2017
the signs are the same. This provides grounds for the LATE interpretation of the IV estimates. In
addition to what is shown in Table 3, I also tested the correlation on various other subgroups. The first
stage estimates never turn negative. However, it is important to keep in mind the local nature of the IV
results. The difference in the strength of the first stage for different subsamples reported in Column (1)
in Table 3may suggest that the number of compliers varies in the subsamples. The coefficient at the first
stage is higher for the subgroups with more compliers. Although, the difference in the estimate and its
precious may also reflect the data sparsity of the underlying age subsample.
4 Results
Using the female sample Figure 2plots the outcome variable against the share on the female share in
income. The vertical dashed line in the middle shows the location of the anticipated effect. The global
polynomial fit of order four constructed with a uniform kernel approximates the population conditional
mean functions for control and treated units. No obvious discontinuity can be seen for any outcome, but
non-linearity is present, suggesting that linear extrapolation may misrepresent the effect.
Table 4reports the results from Model (1) estimated with five different methods using male and
female samples. The top line reports OLS results. This model corresponds to the one used Zhang and
Breunig (2021). Female breadwinning causes females to drink more and do more non-cannabis illicit
Figure 2: Reduction in drinking and gender gap
Notes: Female outcomes against female share of total both-partners income. Vertical line show 0.5 split.
Source: HILDA 2001-2017
drugs. This might appear plausible because modest drinking is known to increase earning, so some of
this effect might transmit into the family setting (e.g., Auld 2005; Barrett 2002). It is also in line with
the argument that violating gender norms stress females (including being abused by the male partner
who retaliates for not being a breadwinner) and they cope by drinking and doing drugs. The results on
the male sample are the opposite. If they earn more, thus complying with the gender norms, they drink
and consume less illicit drugs (presumably, they save energy to retaliate for earning less).
Other lines in the table report result using other estimation methods. None of them show that
breadwinning has any effect on addiction. The IV estimates on the male sample show that males are less
happy when they are the breadwinners. This is the opposite of the presumed mechanism.
Table 4: Estimate on income gender norms, addiction and marital satisfaction
Female sample Male sample
Dependent variable (DV)
Satisfaction Smoking Drinking Cannabis Other drugs Satisfaction Smoking Drinking Cannabis Other drugs
Ordinary least squares estimates
Breadwinner -0.0547 0.0115 0.0465* 0.00773 0.00830* -0.0423 -0.0230*** -0.0324 -0.0146 -0.0106*
(0.0325) (0.00589) (0.0205) (0.00592) (0.00333) (0.0300) (0.00634) (0.0357) (0.00753) (0.00483)
Local polynomial estimates with robust bias-corrected standard errors
Breadwinner 0.0673 0.00365 0.0386 -0.0144 -0.00671 -0.00817 -0.00247 0.0435 0.0281 0.00586
(0.0752) (0.0134) (0.0428) (0.0128) (0.00848) (0.0669) (0.0160) (0.0774) (0.0173) (0.0104)
Fixed effects estimates
Breadwinner 0.0123 0.00531 0.0173 -0.00183 0.00131 -0.0254 -0.00448 -0.00155 -0.00399 -0.00207
(0.0209) (0.00309) (0.00919) (0.00251) (0.00171) (0.0197) (0.00354) (0.0150) (0.00236) (0.00209)
Correlated random effects
Breadwinner 0.0123 0.00541 0.0155 -0.00199 0.00117 -0.0245 -0.00455 0.000137 -0.00394 -0.00186
(0.0206) (0.00309) (0.00915) (0.00251) (0.00171) (0.0195) (0.00354) (0.0149) (0.00236) (0.00209)
Two-stage least squares estimates
Breadwinner 0.265 0.0106 0.136 0.0534 0.0123 -0.677** -0.0222 -0.105 0.0141 0.00411
(0.264) (0.0480) (0.144) (0.0434) (0.0269) (0.249) (0.0526) (0.247) (0.0502) (0.0331)
F test 359.985 341.472 337.328 270.7 270.333 349.591 345.925 344.007 285.782 282.206
Interviewers 702 704 702 692 692 700 704 700 689 689
Observations 40827 45574 40400 31659 31376 39862 45574 39666 29507 29161
Individuals 8966 9329 8969 5789 5764 8770 9229 8778 5339 5297
DV mean 8.271 0.139 0.595 0.0853 0.0307 8.511 0.179 1.330 0.131 0.0509
DV min 0 0 0 0 0 0 0 0 0 0
DV max 10 1 13 1 1 10 1 13 1 1
Notes: All models include all controls and year and area effects. The IV is the p otential breadwinning estimated using quasi-random allocation of interviewers.
*𝑝 < 0.05, ** 𝑝 < 0.01, *** 𝑝 < 0.001.
Source: HILDA 2001-2017
5 Discussion
Some authors argue that couples, where female earns more are under stress because society believes it
is wrong for a woman to make more than a man. In these couples, males are also believed to retaliate
against females. I hypothesized that stress should induce addictive behaviors. It could have explained a
decline in drinking among the younger generation (most prominent but still unanswered research ques-
tions in addiction) because income parity in modern partnerships is more common. I initially replicated
a methodology of the study that shows that in contemporaneous Australia, males physically and emo-
tionally abuse females that earn more and found that indeed the violation of the norm causes an increase
in drinking and drug use in both males and females. However, this result is strongly dependent on the
linearity assumption, and it does not survive a more flexible parametrization. I then also tried mixed
effects and IV methods. There is no causal connection between gender income norms and addiction.
Failing to see the connection, I then reassessed the marital satisfaction, and I could not confirm that
the family’s violation of income norms causes stress. The IV estimates, which depend on ultimately
untestable assumptions and at best show the effect on compliers, show that males are happier if the
female is the breadwinner. The results cast doubts on the recent claim that breadwinning norms play an
important role in modern-day Australia.
Alexeev, S. and D. Weatherburn (2021). “The Australian ready-to-drink beverages tax missed its target
age group”. In: International Journal of Drug Policy 95, p. 103399. issn: 0955-3959. doi:https :
S0955395921003042 (cit. on p. 2).
Auld, M. C. (2005). “Smoking, drinking, and income”. In: Journal of Human Resources 40.2, pp. 505–518
(cit. on p. 12).
Barrett, G. F. (2002). “The Effect of Alcohol Consumption on Earnings”. In: Economic Record 78.240,
pp. 79–96. doi: eprint:
doi/ pdf /10.1111 /1475-4932. 00041.url:https:/ /onlinelibrary.wiley. com/doi/abs/10 .1111/1475 -
4932.00041 (cit. on p. 12).
Bartik, T. J. and J. Bartik Timothy (1991). “Who benefits from state and local economic development
policies?” In: (cit. on p. 7).
Bellemare, M. F. and C. J. Wichman (2020). “Elasticities and the Inverse Hyperbolic Sine Transfor-
mation”. In: Oxford Bulletin of Economics and Statistics 82.1, pp. 50–61. doi:https : / / doi . org /
10.1111/obes.12325. eprint: https: / / onlinelibrary. doi / pdf/10.1111/obes.12325.url: (cit. on p. 3).
Bertrand, M., E. Kamenica, and J. Pan (Jan. 2015). “ Gender Identity and Relative Income within
Households *”. In: The Quarterly Journal of Economics 130.2, pp. 571–614. issn: 0033-5533. doi:
10.1093/qje/qjv001. eprint:
pdf.url: (cit. on pp. 1,5,7).
Burbidge, J. B., L. Magee, and A. L. Robb (1988). “Alternative Transformations to Handle Extreme
Values of the Dependent Variable”. In: Journal of the American Statistical Association 83.401, pp. 123–
127. issn: 01621459. url: (cit. on p. 3).
Calonico, S., M. D. Cattaneo, and M. H. Farrell (2018). “On the Effect of Bias Estimation on Coverage
Accuracy in Nonparametric Inference”. In: Journal of the American Statistical Association 113.522,
pp. 767–779. doi:10.1080/01621459.2017.1285776. eprint:
1285776.url: (cit. on p. 6).
Calonico, S., M. D. Cattaneo, M. H. Farrell, and R. Titiunik (2017). “Rdrobust: Software for Regression-
discontinuity Designs”. In: The Stata Journal 17.2, pp. 372–404. doi:10.1177/1536867X1701700208.
(cit. on p. 6).
Calonico, S., M. D. Cattaneo, and R. Titiunik (2014). “Robust Nonparametric Confidence Intervals for
Regression-Discontinuity Designs”. In: Econometrica 82.6, pp. 2295–2326. doi:https:/ /
3982 / ECTA11757. eprint: https: / / onlinelibrary . wiley. com / doi / pdf / 10 . 3982 / ECTA11757.url: (cit. on p. 6).
Foster, G. and L. S. Stratton (Jan. 2021). “Does female breadwinning make partnerships less healthy or
less stable?” In: Journal of Population Economics 34.1, pp. 63–96. issn: 1432-1475. doi:10.1007/
s00148-020-00783-5.url: (cit. on pp. 13,5,6).
Goldsmith-Pinkham, P., I. Sorkin, and H. Swift (Aug. 2020). “Bartik Instruments: What, When, Why,
and How”. In: American Economic Review 110.8, pp. 2586–2624. doi:10.1257 /aer. 20181047.url: (cit. on p. 7).
Grawe, N. D. (2006). “Lifecycle bias in estimates of intergenerational earnings persistence”. In: Labour
Economics 13.5, pp. 551–570. issn: 0927-5371. doi:
url: (cit. on p. 6).
MacKinnon, J. G. and L. Magee (1990). “Transforming the Dependent Variable in Regression Models”.
In: International Economic Review 31.2, pp. 315–339. issn: 00206598, 14682354. url:http://www. (cit. on p. 3).
Wooldridge, J. M. (2011). “A simple method for estimating unconditional heterogeneity distributions
in correlated random effects models”. In: Economics Letters 113.1, pp. 12–15. issn: 0165-1765. doi:
pii/S0165176511001960 (cit. on p. 7).
Zhang, Y. and R. V. Breunig (2021). “Gender norms and domestic abuse: Evidence from Australia”.
In: Tax and Transfer Policy Institute-working paper 5. url:https: / / taxpolicy.crawford.anu. edu .
au/ publication/ttpi - working-papers/18650 /gender- norms- and- domestic- abuse- evidence- australia
(cit. on pp. 1,2,10).
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
Background During 2008 and 2009, the Australian Government increased the tax on ready-to-drink alcoholic beverages by 70% to discourage drinking among adolescents. Methods To evaluate the tax, we use the difference-in-difference and comparative interrupted time series estimators, where age is used to define the control and treatment groups. This methodology is applied to the Household Income and Labour Dynamics in Australia survey. Results We show that the tax did not affect the alcohol consumption of those aged under 25 (the tax targeted age group) but substantially reduced drinking among those aged 25–69, reducing their average daily consumption of standard drinks by 8.9% from 2010 to 2018. Conclusion The age group under 25 did not respond to the tax likely because of product substitution. Alcohol price policy may need to acknowledge complex substitute/complement relationships between beverages and consider a floor price on alcohol or a uniform volumetric tax per standard drink.
Full-text available
Social norms can have a persistent influence on outcomes. Since the end of World War II, men have been the primary breadwinner in most households in the developed world, and US data from the late twentieth century suggests violation of this norm stresses partnerships. Is this still true? We examine whether female breadwinning makes partnerships less healthy or less stable using more recent US and Australian data. We find a much more modest association in both countries between female breadwinning and measures of relationship health or stability in OLS models for mixed-gender couples than has been found in prior studies. Transitions into female breadwinning are problematic mainly for cohabiting couples and especially so for younger people and less-educated men. These results suggest that social norms may be weakening, but mating market dynamics may also play a role. We find some evidence that cohabiting women in Australia who out-earn their partners subsequently re-partner with men who have higher earnings relative to themselves.
Background Evidence suggests adolescent alcohol consumption has declined since the turn of the millennium in almost all high-income countries. However, differences in the timing and magnitude of the decline have not been explored across countries. Methods We examined trends in adolescent past month or monthly alcohol consumption prevalence from cross-national or national survey reports for 39 countries and four US territories. For each country, we calculated the magnitude of the decline in youth drinking as the relative change in prevalence from the peak year to the most recent year available. Heat maps were utilized to present the timing and magnitudes of these declines. Results The timing and extent of youth drinking declines have varied markedly across countries. The decline began in the USA before 1999, followed by Northern European countries in the early 2000s; Western Europe and Australasia in the mid-2000s. The steepest declines were found for Northern Europe and the UK, and the shallowest declines were observed in Eastern and Southern European countries. Conclusions Previous analyses of the decline in adolescent drinking have emphasized the wide reach of the changes and their near-coincidence in time. Our analysis points to the other side of the picture that there were limits to the wide reach, and that there was considerable variation in timing. These findings suggest that as well as broader explanations that stretch across countries, efforts to explain recent trends in adolescent drinking should also consider factors specific to countries and regions.
Background: Adolescent drinking has declined across many developed countries from the turn of the century. The aim of this review is to explore existing evidence examining possible reasons for this decline. Methods: We conducted systematic searches across five databases: Medline, PsycINFO, CINAHL, Informit Health and Scopus. Studies were included if association between declining alcohol consumption and potential explanatory factors were measured over time. Narrative synthesis was undertaken due to substantial methodological heterogeneity in these studies. Results: 17 studies met the inclusion criteria. Five studies found moderate evidence for changes in parental practices as a potential cause for the decline. Five studies that examined whether alcohol policy changes influenced the decline found weak evidence of association. Three studies explored whether alcohol use has been substituted by illicit substances but no evidence was found. Two studies examined the effect of a weaker economy; both identified increase in adolescent alcohol use during times of economic crisis. One study indicated that changes in exposure to alcohol advertising were positively associated with the decline and another examined the role of immigration of non-drinking populations but found no evidence of association. One study tested participation in organised sports and party lifestyle as a potential cause but did not use robust analytical methods and therefore did not provide strong evidence of association for the decline. Conclusions: The most robust and consistent evidence was identified for shifts in parental practices. Further research is required using robust analytical methods such as ARIMA modelling techniques and utilising cross-national data.
We describe a major upgrade to the Stata (and R) rdrobust package, which provides a wide array of estimation, inference, and falsification methods for the analysis and interpretation of regression-discontinuity designs. The main new features of this upgraded version are as follows: i) covariate-adjusted bandwidth selection, point estimation, and robust bias-corrected inference, ii) cluster–robust bandwidth selection, point estimation, and robust bias-corrected inference, iii) weighted global polynomial fits and pointwise confidence bands in regression-discontinuity plots, and iv) several new bandwidth selection methods, including different bandwidths for control and treatment groups, coverage error-rate optimal bandwidths, and optimal bandwidths for fuzzy designs. In addition, the upgraded package has superior performance because of several numerical and implementation improvements. We also discuss issues of backward compatibility and provide a companion R package with the same syntax and capabilities.
All human societies exhibit some degree of division of labour by gender. These divisions continue to exist as participation in paid work has increased over time. Gender divisions occur between household tasks, between unpaid and paid work, and within paid work. Economists have explained these divisions through reliance on essentialist arguments and/or the fundamental economic concepts of efficiency of specialization and division of labour, and investment in human capital. However, gender discrimination can also cause division of labour, and the feedback effects of such discrimination make it difficult to untangle the causes of the gender division of labour.
Economic theory concerning inequality between the sexes focuses upon inequality in wages, job recruitment, promotion and dismissal, for women and men with similar qualifications and availability. Neoclassical theory explains these inequalities as a result of free and rational choice, based upon the biological differences between the sexes. According to Becker (1981), women’s role in reproduction makes it rational for women to specialize more in family skills, and men more in labour market skills, and parents make a rational choice for their children by preparing them for different careers. When women’s reproductive role is reduced due to the decline of birth rates, women’s availability for the labour market increases, and they begin to invest more in labour market skills than is the case in countries with continued high fertility. So sex-related differences in level and types of human investment and availability provide the explanation for the differences in wages, types of work and promotion.
Nonparametric methods play a central role in modern empirical work. While they provide inference procedures that are more robust to parametric misspecification bias, they may be quite sensitive to tuning parameter choices. We study the effects of bias correction on confidence interval coverage in the context of kernel density and local polynomial regression estimation, and prove that bias correction can be preferred to undersmoothing for minimizing coverage error. This result is achieved using a novel, yet simple, Studentization, which leads to a new way of constructing kernel-based bias-corrected confidence intervals. We also derive coverage error optimal bandwidths, and discuss easy-to-implement bandwidth selection procedures. In particular, we show that the MSE-optimal bandwidth delivers the fastest coverage error decay rate only at interior points when second-order (equivalent) kernels are employed, but is otherwise suboptimal both at interior and boundary points. All the results are established using valid Edgeworth expansions and illustrated with simulated data. Our findings have important consequences for empirical work as they indicate that bias-corrected confidence intervals, coupled with appropriate standard errors, have smaller coverage errors and therefore are less sensitive to tuning parameter choices. To illustrate the applicability of our results, we study inference in regression discontinuity (RD) designs, where we establish the same coverage error and robustness improvements for bias-corrected confidence intervals, and also give a simple rule-of-thumb bandwidth choice for their implementation based on correcting the MSE-optimal bandwidth. For example, for the popular local-linear RD estimator and a sample size of $n=500$, shrinking the MSE-optimal bandwidth by $27\%$ leads to bias-corrected confidence intervals with the fastest coverage error decay rate.
Transformations that could be used to reduce the influence of extreme observations of dependent variables, which can assume either sign, on regression coefficient estimates are studied in this article. Two that seem reasonable on a priori grounds—the extended Box—Cox (BC) and the inverse hyperbolic sine (IHS)—are evaluated in detail. One feature is that the log-likelihood function for IHS is defined for zero values of the dependent variable, which is not true of BC. The double-length regression technique (Davidson and MacKinnon 1984) is used to perform hypothesis tests of one transformation against the other using Canadian data on household net worth. These tests support the use of IHS instead of BC for this data set. Empirical investigators in economics often work with a logged dependent variable (taking the natural logarithm of a data series is, of course, a special case of BC) to reduce the weight their particular estimation procedure might otherwise attach to extreme values of the dependent variable. Logging dependent and independent variables has the added attraction that slope coefficients may be interpreted as elasticities. In the event that the dependent variable assumes nonpositive values, some researchers (e.g., Diamond and Hausman 1984; King and Dicks-Mireaux 1982) have dropped the nonpositive values and others have added a constant so that each observation is positive. An alternative transformation, which is defined for any real number, is the IHS transformation, sinh(x) = log(x + (x + l)). This was proposed in Johnson (1949) and is just as easy to employ as the BC transformation.