Journal of Family and Economic Issues

, Volume 32, Issue 2, pp 191–203

Can Marriage Reduce Risky Health Behavior for African-Americans?

Authors

    • Department of EconomicsUniversity of Toledo and Office of Regulations, Policy & Social Science, Food & Drug Administration
  • Olugbenga Ajilore
    • Department of EconomicsUniversity of Toledo
Original Paper

DOI: 10.1007/s10834-010-9242-z

Cite this article as:
Ali, M.M. & Ajilore, O. J Fam Econ Iss (2011) 32: 191. doi:10.1007/s10834-010-9242-z
  • 218 Views

Abstract

This paper estimates whether marriage can improve health outcomes for African-Americans through changes in risky health behaviors like smoking, drinking, and drug use. Using data from the National Longitudinal Study on Adolescent Health and propensity score matching methodology to account for the potential selection bias, the results show that marriage does lead to a reduction in risky health behaviors, specifically drinking and drug use. This question has important policy implications because if marriage has the same benefits for African-Americans as it does for the general population, social welfare programs can be re-evaluated to incorporate marriage promotion, and further support can be given to programs that decrease adverse health behaviors.

Keywords

African-AmericansMarriagePropensity score matchingRisky health behaviors

Introduction

The extensive literature on marital outcomes consistently shows married adults are generally healthier than their unmarried counterparts. Married adults were documented to have lower mortality rates, lower morbidity rates and to be in better physical and mental health (Waldron et al. 1996). The absence of isolation, social support, and economic well-being has been proposed as reasons as to why marriage is positively correlated with health (Coombs 1991; Ross et al. 1990). Because goods and services contribute to the production of health (Grossman 1972) and because marriage may involve a transition from a one- to a two-person household, this may increase the resource endowments (income, health insurance) of the family, thus contributing to better health. Skogrand et al. (2010) explore the financial habits of couples who say they have a great marriage. Their analysis shows that financial stability was a key component of their marital success. Also increased time availability in a married household, due to gains from specialization and exchange in the presence of comparative advantage, would allow for greater investments in health.

While health outcomes and marriage have been widely studied, the effects of marriage on risky health behavior have received relatively little attention (Duncan et al. 2006). Risky behaviors such as smoking, heavy drinking and drug use are associated with numerous detrimental health outcomes, and a large part of most chronic health conditions are the result of engaging in such risky behaviors. What is also missing in the literature is the effect of marriage on such risky health behaviors among African-Americans. This focus on the African-American population is an important omission from the literature because large racial disparities exist in both marriage and health. Despite the emergence of a significant African American middle class, African Americans continue to experience social and economic disparities in the social contexts in which they belong (Harris et al. 2010). For example, African Americans at all levels of socioeconomic status are more segregated than any other racial or ethnic group (Iceland et al. 2002). These segregated neighborhoods are characterized by high poverty and crime, poor schools, and fewer economic opportunities and thus have serious implications for the health outcomes among African Americans, including high prevalence of substance abuse (Kawachi and Berkman 2003). This disparity experienced by African Americans despite improvements in the overall health of the nation is an important public health concern. In addition, a 2006 report released by the Center for Disease Control and Prevention (CDC) found an increasing trend of participation in risky health behaviors among African-Americans. Thus, in this paper, we aim to estimate whether marriage can improve health outcomes for African-Americans through changes in such risky behaviors, namely smoking, drinking, and drug use.

While risky behavior encompasses many other types of actions, such as violence and delinquency, we focus on these three because they have direct long-term impacts on health outcomes. In addition, risky health behaviors such as excessive drinking are also associated with negative outcomes like violence and sexual assault. Caetano et al. (2000) reports that intimate partner violence is highest among African-American couples, and a significant portion of those couples were inebriated at the time of the assault. Lower rates of marriage combined with a higher rate of participation in health compromising behaviors among African Americans motivates our primary empirical question, which is to investigate whether a causal mechanism exists linking marriages with a reduction in risky health behavior.

We also account for cohabitation in our empirical analysis. While marriage is the most common foundation of family life in the U.S., there has been an upward trend in non-marital relationships as a process of family formation, especially cohabitation (Heiland and Liu 2006). Although cohabitation can act as a precursor to marriage, not all cohabitations transition into marriage (Jones 2010; Manning and Landale 1996; Manning and Smock 1995), and there are behavioral differences between individuals who are cohabitating and individuals who are married (Axinn and Thornton 1992; Rindfuss and VandenHeuvel 1990; Winkler 1997). For example, Rindfuss and VandenHeuvel (1990) found that cohabitators behave more like singles rather than married individuals. Also Winkler (1997) found that cohabitators don’t pool all their income together. Malone et al. (2010) find that women who cohabitate tend to have more worries about income than those in stable marriages. Thus, analyzing both cohabitation and marriage would enable us to identify whether marriage as a union exhibits benefits (a reduction in health compromising behaviors in our case) that might not be present in other forms of union.

Marriage and Risky Health Behavior

Besides the absence of isolation, social support, and economic well-being, the underlying mechanism that relates marriages to reduction in risky health behaviors can be attributed to two main perspectives: social norms and monitoring a partner’s behavior (Duncan et al. 2006; Umberson 1987). These two factors allow us to understand why marriage as a form of social role or integration might reduce participation in such behaviors.

Marriages are governed by social norms and one such norm includes a belief that spouses should stay together for the long term, be monogamous and love each other (Duncan et al. 2006). Another social norm that is relevant in terms of risky health behavior is the eschewing of activities that are often associated with the single life. Smoking, heavy drinking, and drug use maybe regarded as an acceptable component of the single life, but these are among things one is expected to give up after marriage (Duncan et al. 2006). Thus transitions from the single life into a marriage may be indicative of such willingness motivating an individual more in that direction. However, it is not clear whether other forms of union such as cohabitation, are governed by these same social norms because only in recent decades has other non-marital unions have become a common living arrangement.

Monitoring of a partner’s behavior is one by-product of marriage, where the cost of not complying with an entrenched norm, i.e. reduction in consumption of addictive substances in our case, exceeds the benefits of violating them. An individual accustomed to consuming addictive substances might prefer the continuation of such behavior only if his/her partner does not punish this behavior (for example via threat of a separation or dissolution of the relationship). In other words, the relative ease of monitoring a partner’s behavior in marriages increases the opportunity cost of engaging in such risky health behavior. Such an increase in the opportunity cost might also be applicable to cohabitation, but because marriages entail more engagement with one’s spouse, the cost might be higher for married individuals. As mentioned previously, accounting for cohabitation in our analysis would allow us to further identify whether marriage as a form of family life presents benefits that are not observed in other non-marital relationships.

The literature also identifies two potential theoretical perspectives to explain the observed improvements in health outcomes through marriage: the marriage protection effect and the marriage selection effect (Goldman 1993; Murray 2000). The marriage protection effect refers to the beneficial effects that can stem from marriages, such as increased social support, increase in income, healthy lifestyle, all of which contribute to better health outcomes (a reduction in risky health behaviors in our case) either through social norms or monitoring. In other words, the marriage protection theory suggests that being married is the cause of good health. The marriage selection effect on the other hand, could occur because healthy individuals are disproportionately more likely to opt into marriages. Since unhealthy people are considered less desirable marriage partners (Waldron et al. 1996), the observed correlation between marital status and health is not a function of marriage, but rather a function of the process by which partners are selected. In the presence of such selection effects, the estimates of the marriage effects on health outcomes could be biased because we might be measuring the decision by healthy individuals to get married rather than the better health outcomes that might result from marriages. From a policy perspective, it is important that such biases be purged from the estimates to empirically quantify the marriage protection effects. In this paper we control for selection bias by utilizing the propensity score matching technique.

We estimate our models using data from the National Longitudinal Study of Adolescent Health (Add Health). Being a longitudinal dataset that follows individuals from their adolescence till adulthood, it allows us to control for a wide range of variables (from both adulthood and adolescence) that potentially measures not only the propensity of marriages but also the likelihood of engagement in risky behavior. However, here we do not claim to test which of the two perspectives, social norms or monitoring, plays a larger role in reduction of risky health behavior. Testing for the two perspectives would require data pertaining to the quality of marriage, which is not available in this dataset. Our focus instead is to examine whether a large enough marriage protection effect exists after accounting for marriage selection. We hypothesize that after controlling for the potential selection bias, marriage will have a significant impact on the reduction of risky health behaviors and would exhibit more protective benefits compared to other forms of union, such as cohabitation.

Data

The data set used in our study comes from the National Longitudinal Study of Adolescent Health (Harris et al. 2009). Add Health consists of data on U.S. students in 132 schools nationwide between grades 7 to 12. Add Health data includes three waves of in-home surveys first conducted in 1994 with follow-up surveys in 1996 and in 2002, when most respondents had made a transition to adulthood. The primary data for our analysis came from all three waves (1994, 1996 and 2002) of the in-home survey portion of Add Health. The third wave of the data includes individuals of marriageable age. The sample of our analysis includes all African-Americans that were interviewed in all the three waves of study (N = 2,581).

The primary advantage of the data set is its longitudinal nature, which allows us to control for past participation in risky behavior when individuals were in their adolescence. This helps us capture the addictive nature of such risky behaviors, especially smoking. It might possibly be that individuals who were more prone to risky behavior in their adolescence continue with such participation into their adulthood. In such regards, the effect that marriage could have on risky behavior could be biased. Controlling for risky behavior during adolescence allows us to obtain a more precise estimate of the marriage protection effect. In addition, Add Health includes various measures of physical and mental health over a large period of time, allowing us to assess changes in health during the transition to adulthood.

Using Add Health also allows us to focus on how marriage affects behaviors among a relatively younger cohort, the life stage in which the marriage promotion policies have most recently been focused (Harris et al. 2010). By focusing on a younger cohort it might be possible for us to inform policy better because early experiences can substantially impact subsequent behaviors (Schoen et al. 2007). Young adulthood is also an important stage in an individual’s life course with implications for future attainments and health (Harris et al. 2010). For African-Americans, focusing on the youth is important because the cultural context in which adolescents have grown up differs from that of the older cohort. African-American youth had a more difficult time gaining stable employment relative to older African-Americans (Roy 2005). Thus, the effects of marriage will differ for a younger cohort as opposed to an older cohort. Any evidence of positive health benefits from marriage among a demographic group that has a low propensity of marriage (African-Americans in early adulthood), would imply a much larger returns to health from marriage among older African Americans.

Measures of Risky Behaviors

The dependent variables of our analysis pertains to participation in risky behaviors, particularly, smoking, drinking, and drug use. There are two smoking measures, two drinking measures, and one drug use measure.

The first smoking variable is a binary indicator for whether the individual smokes on a daily basis (all days out of the past 30 days) or not. The second variable is a continuous variable indicating the number of cigarettes smoked on the day the individual smoked. The first drinking variable is a dichotomous variable indicating whether the individual has participated in heavy drinking or not. Heavy drinking is defined as having more than five drinks usually at one time for each time they had a drink in the last 2 weeks. The second drinking variable is also a dichotomous variable indicating whether the individual got drunk in the past 12 months. The drug use variable is a binary measure indicating whether an individual tried marijuana at least once in the last 30 days.

Union Status Variables

The primary variable of interest in our study is the union status of the individuals, i.e. the individual’s marital or cohabitation status. Our union status variables are binary indicators of whether the individual is currently married or not and also whether the individual is currently cohabitating or not.

Control Variables

Other control variables in our analysis includes the following demographic characteristics: gender (dummy variable for male), age, educational attainment (whether the individual have a college degree), nativity (United States native), and current pre-tax income. Differences on these socio-demographic variables are especially profound within the African American communities and those who marry differ along socioeconomic dimensions substantially from those who remain single (Clarkwest 2006). In addition, Waldron et al. (1996) finds income and work-status to account for part of the marriage protection effect. Also included in our analysis are variables measuring whether the individual lived in a two-parent household, whether they have a sibling, and whether they consider religion to be important to them. These factors are considered to be highly correlated with the probability of entry into marriage (Harris et al. 2010).

We also control for measures of physical and mental health during adolescence including participation in risky behaviors (from Wave II: 1996). Baseline health in adolescent impacts both selections into marriages and health in young adulthood (Harris et al. 2010). Physical health during adolescent years were accounted by controlling for obesity status measured using the CDC age and gender-adjusted cutoffs in BMI (calculated using measured height and weight) and a self-reported indicator of being in poor health. We also control for the individual’s birth weight (in pounds).

Our mental health indicator is based on a dichotomized version of the Center for Epidemiologic Studies Depression (CES-D) Scale, which is a very widely used measure of depressive symptom. Add Health administered 18 of the 20 items that typically comprise the CES-D Scale. Specifically, respondents were asked to indicate the frequency with they had experienced certain feelings or emotions during the past week, such as how often they felt “life had been a failure,” how often they felt “lonely,” and how often they “talked less than usual.” Possible responses were “rarely or none of the time” (=0); “some or a little of the time” (=1); “occasionally or a moderate amount of the time” (=2); and “most or all of the time” (=3). Responses to these 18 items were summed to produce a score of between 0 and 54 (Cronbach’s α = 0.86). From this the depression indicator was set equal to 1 if a male respondent scored above 22 on the CES-D Scale and 0 otherwise. A cut-point of 24 was utilized for the female respondents. The CES-D Scale is often dichotomized in this manner in the medical literature (Goodman and Capitman 2000; Hallfors et al. 2005), and one advantage of using a dichotomized version of the CES-D Scale is that it concentrates on the right-hand tail of the CES-D distribution where medical diagnoses of major depression are made (Ali et al. 2010; Sabia and Rees 2008).

We also control for characteristics pertaining to their parents’ education, whether their parents’ are religious, whether their parents’ were on welfare, and whether their parents’ engaged in smoking or drinking. The use of these contextual parental measures allows us to account for the role modeling effects (or the inter-generational transmission of behaviors) in the social context among the respondents (Harris et al. 2010). Finally, we include the following four factors related to attitude towards marriage: whether the individuals think it is important to be faithful in a marriage, whether they believe that lifelong commitment is important for marriage, whether they believe that financial solvency is important in a marriage, and whether they think it is ok to live with a partner that they don’t intend to marry. Table 1 reports descriptive statistics for all variables of interest among our study sample.
Table 1

Descriptive statistics

Variable (N = 2,581)

Mean

Standard deviation

Minimum

Maximum

Outcome

 Daily smoker

0.098

0.297

0

1

 Number of cigarette smoked

3.972

5.129

0

100

 Heavy drinker

0.184

0.387

0

1

 Drunk

0.285

0.951

0

1

 Drug

0.191

0.393

0

1

Union status

 Married

0.087

0.282

0

1

 Cohabit

0.129

0.335

0

1

Health measures (Wave II: 1996)

 Birth weight

6.344

1.262

3

12

 Obese

0.155

0.362

0

1

 Self-reported poor health

0.370

0.258

0

1

 Depressed

0.084

0.278

0

1

 Smoke

0.886

0.317

0

1

 Drink

0.336

0.572

0

1

 Drug

0.135

0.342

0

1

Demographics

 Male

0.433

0.496

0

1

 Age

21.56

1.656

18

27

 College

0.077

0.267

0

1

 Has siblings

0.384

0.486

0

1

 Born USA

0.767

0.423

0

1

 Log income

8.380

2.181

0

12.924

 Religious

0.859

0.348

0

1

Parental characteristics (Wave I: 1994)

 Lived with both biological parents

0.322

0.467

0

1

 Parent black

0.802

0.398

0

1

 Parent college

0.129

0.336

0

1

 Parent religious

0.129

0.336

0

1

 Parent smoke

0.239

0.427

0

1

 Parent drink

0.522

0.499

0

1

 Welfare

0.359

0.480

0

1

Attitude towards marriage

 Faithful

0.862

0.345

0

1

 Life long

0.651

0.476

0

1

 Finance

0.369

0.483

0

1

 Live together

0.651

0.477

0

1

Estimation Methodology

As mentioned previously, because marriage is a non-random event, a simple regression about the effects of marriage on risky behavior may be ignoring potential selection biases. It is quite possible that individuals who engage less in risky behavior are more prone to being married. The selection issue may also have implications for African-Americans specifically because African-Americans are less likely to be in any relationship (Harris et al. 2010). Therefore, it is likely that the non-random selection into marriage could complicate the estimation of the marriage effect. Simple comparisons of risky behavior outcomes by marital status can be misleading if individuals who get married are different from those who remain unmarried. A method of correcting this selection bias is to use propensity score matching (Rosenbaum and Rubin 1983). Propensity score matching is a technique that is used to adjust for pre-treatment observable differences between a treatment group and a control group; thus the primary purpose of this methodology is “to replicate conditions of an experiment such that the treatment variable, in this case marriage, can be treated as though it occurred at random and that the individuals under analysis are homogenous on all other factors except the treatment variable” (King et al. 2007, p. 43). Propensity score matching allows us to formulate the relationship between marriage and risky health behavior in a framework similar to a social experiment in which the treatment is randomly assigned. In our context, the treatment (marriage) is defined in terms of the potential outcomes for those who married (treated). We estimate matching methods to construct the counterfactual outcomes for the treated in the absence of a treatment by matching the treated with controls (individuals who did not marry) who share identical characteristics that rule selection into treatment.

Although this methodology addresses selection on observables, it does not extend to selection on unobservables; thus, like the literature, we also rely on the richness of our data set to reduce such biases generated by unobservables. The variables selected for inclusion are those thought to be necessary to improve the quality of the match between the treated and control groups, also based on marriage theory and prior research. As indicated in “Control Variables” section, our models include an extensive set of control variables pertaining to demographic measures, socioeconomic characteristics, parental background, baseline indicator of physical and mental health, and attitudes towards marriage and baseline participation in risky health behaviors to account for their addictive nature.

Empirical Framework

Closely following the notations used by Liu and Heiland (2010) we provide an intuitive exposition of our estimation framework; consider an individual i who engages in risky health behaviors Ri. The interrelation of the risky behavior and marital status can be presented as:
$$ \begin{gathered} R_{i} = \beta M{}_{i} + \alpha X{}_{i} + \varepsilon_{i} \hfill \\ M_{i} = \eta X_{i} + v_{i} \hfill \\ \end{gathered} $$
where Mi equals to 1 if the individual is married and 0 otherwise. Characteristics of the individual that influence their engagement in health risky activities and marital outcome are represented by Xi. Unobservable characteristics affecting Ri and Mi are captured by εi and vi. The effect of marriage on risky behavior is measured by β. However, estimating Ri directly may yield a biased estimate of β if Mi and εi are statistically dependent. Two main sources can be attributed to this dependency (Heckman and Robb 1985; Rosenbaum and Rubin 1983): first, Xi and εi may be correlated, (the individuals’ characteristics may be correlated with unmeasured addictive propensities); and second, εi and vi may be correlated (unobserved factors may affect both risky behaviors and marital status). The existence of an either source of bias would likely show that married individuals have different outcomes compared to their non-married counterparts, independent of any causal effect of marriage. Selection bias may arise in the regression analysis because these estimators would utilize data from all observations to be combined into one estimate of the marriage effect. The validity of the estimate would be suspect, if individuals who marry are different from those who don’t. In the presence of any factors that affect the individuals’ decision to marry as well as their engagement in risky behavior, the estimate will reflect both the marriage protection effect (the “true” marriage effect we want to identify) and the marriage selection effect (the effect that influences the individual’s decision to marry in the first place).
In our analysis the treatment is marriage, thus Mi = 1 denotes the treatment group and Mi = 0 denotes the control group (individuals who do not marry). Let Ri(1) denote the potential outcome of individual i under the treatment state (Mi = 1) and Ri(0) the potential outcome if the same individual i receives no treatment (Mi = 0). Thus \( R_{i} = M_{i} R_{i} (1) + (1 - M_{i} )R_{i} (0) \)is the observed outcome of individual i. The individual treatment effect \( \beta_{i} = R_{i} (1) - R_{i} (0) \)is unobserved because either Ri(1) or Ri(0) is missing. Standard parametric models (e.g. OLS) estimate the average treatment effect (ATE) by taking the average outcome difference between the treatment groups: \( \beta_{\text{OLS}} = E[R_{i} (1)|M_{i} = 1] - E[R_{i} (0)|M_{i} = 0] \). If individuals who remained unmarried are unlikely to ever marry, the ATE may not be particularly helpful in understanding how marriage affects participation in risky behaviors. An alternative is to estimate the average treatment effect on the treated (ATT):
$$ \beta_{{M_{i = 1} }} = E[\beta_{i} |M_{i} = 1] = E[R_{i} (1)|M_{i} = 1] - E[R_{i} (0)|M_{i} = 1] $$
which is the difference between the expected outcome of an individual who marries and the expected outcome of the same individual if he/she were to remain unmarried.

While we observe the outcomes of the married individuals and thus are able to construct the first expectation \( E[R_{i} (1)|M_{i} = 1] \), we cannot identify the counterfactual expectation \( E[R_{i} (0)|M_{i} = 1] \)without invoking further assumptions. To overcome this problem, we have to rely on the individuals who remain unmarried to obtain information on the counterfactual outcome. A way to construct a sample counterpart for the counterfactual outcomes of the treated had they not received treatment is to use statistical matching. The matching estimators can be devised to reconstruct the condition of an experiment by stratifying the sample with respect to covariates Xi that rule selection into treatment. Selection bias is eliminated provided all variables in Xiare measured and balanced (comparable) between the two treatment groups within each stratum. In this case, each stratum represents a separate randomized experiment and simple outcome difference between the treated, and controls provide an unbiased estimate of the treatment effect (Liu and Heiland 2010).

An identifying assumption of the matching method is the Conditional Independence Assumption (CIA), i.e. all relevant outcome differences between the matched treated and controls are captured in their observed characteristics. Hence, conditional on X, the outcomes of those who remained unmarried are what the outcomes of those who married would have been if they had remained unmarried. The conditional response of the treated under no treatment could thus be estimated by the conditional mean response of the matched untreated. To estimate the ATT, one is first required to take the outcome difference between the two treatment groups conditional on X and then average over the distribution of the observables in the treated population. Rosenbaum and Rubin (1983) proposed using the conditional probability of selection into treatment (propensity score) to stratify the sample. They demonstrate that by definition the treated and the non-treated with the same propensity score have the same distribution of X. This is also called the balancing property of the propensity score. Matching treated and untreated units using their estimated propensity score and placing them into one block (i.e. observations with propensity score falling within a specified range) means that selection into treatment within each block is random and the probability of receiving treatment within this block equals the propensity score. However, the probability of finding an exact match is theoretically zero. Thus, a certain distance between the treated and the untreated has to be accepted (Becker and Ichino 2002). A variety of matching algorithms have been used in the literature, including Gaussian, Epanechinikov, and Uniform (radius) Kernel matching, with none a priori superior than the other. Because there is no consensus in the existing literature about what the appropriate or the most efficient matching algorithm is, we utilize all of the mentioned algorithms and compare our estimates. This use of all the algorithms also provides a way to check the robustness of our results.

Results

We begin first by estimating the propensity score for selection into the treatment, by using a probit model. An important issue in implementing the probit model is to decide on the covariates to be included. Here we rely on the proposition by Rosenbaum and Rubin (1983) and Dehejia and Wahba (1999): for any given specification, group observations into blocks defined by the estimated propensity score and verify whether it succeeds in balancing the covariates between the treated and the controls within each block. If a particular structure that balances the covariates are not found (indicating that the specification does not capture the differences between the treated and the controls), we include additional covariates until this condition is satisfied. We begin by including the simplest set of controls (age, gender, education, and religion) and finally succeed in balancing the covariate means when we include initial health endowments for the individuals’ adolescent years. It is important to note here that the extensive array of control variables contribute more towards satisfying the balancing property and producing better quality matches (as evidenced in Fig. 1), rather than illuminating which aspect of marriage might contribute more towards the decline in risky health behavior.
https://static-content.springer.com/image/art%3A10.1007%2Fs10834-010-9242-z/MediaObjects/10834_2010_9242_Fig1_HTML.gif
Fig. 1

Box plot of the estimated propensity score for the treated units (1) and the control units (0) within the common support region

Table 2 presents results of the balancing test between the treated and the control groups after stratifying the sample into blocks based on their estimated propensity score. From the table we can see that the characteristics of the matched control within each block resemble the treated group, showing that the balancing condition is satisfied. Matching based on the full set of controls result in a sample of 2,581 observations with propensity score falling within the region of common support (0.009, 0.463). Figure 1 also shows that the treated and the control are comparable because there is sufficient overlap in the propensity score within each block. In Table 3 we present the probit estimate of the propensity score for the fully specified model. The results are consistent with our conjecture and the previous literature (Harris et al. 2010) that better baseline health status are positively correlated with propensity score of marriage, whereas baseline participation in risky health behaviors are negatively correlated. Attitude towards marriage such as whether the individuals think it to be important to be faithful in a marriage and whether they believe that lifelong commitment is important for marriage is positively correlated; whether they believe that financial solvency is important in a marriage and whether they think it is ok to live with a partner that they do not intend to marry are negatively correlated. Once again it is important to note here that the control variables serve to provide a better match between the treated and the control group. We also note that although the pseudo R2 is not that large, it is still very similar to the literature that has utilized propensity score matching method to provide causal evidence of the effect of marriage (for example see Liu and Heiland 2010). In addition, because Fig. 1 exhibits a sufficient overlap between the two groups, the pseudo R2 is less of a concern.
Table 2

Test of balancing properties between the control and treated group (two-sample T test of means): T-statistic reported

 

Block 1

Block 2

Block 3

Block 4

Block 5

N treated

18

59

96

48

4

N control

839

694

530

163

4

range of the propensity score

[0.009, 0.050]

[0.050, 0.010]

[0.100, 0.200]

[0.200, 0.400]

[0.400, 0.463]

Two-sample test of means: |T| statistics

 Propensity score

0.1217

0.3337

0.1304

2.0830

0.2888

 Male

0.6515

1.0761

0.9038

1.5924

0.6547

 Age

0.3282

0.2965

0.7870

1.3983

0.8783

 College

1.2585

0.6503

1.4739

0.5417

0.6000

 Has siblings

1.6678

0.4569

0.9700

1.0080

1.7321

 Born USA

0.7641

0.7057

2.2056

2.1837

0.6077

 Log income

0.9564

0.0366

0.1881

0.6583

0.2015

 Religious

1.8965

0.4307

1.2103

0.6583

0.6006

 Obese

0.3905

1.6009

0.8649

1.2711

0.6547

 Birth weight

1.3188

0.3149

1.0424

0.9969

0.0493

 Self-reported poor health

0.0716

0.7378

1.2458

1.6867

1.7321

 Depressed

0.9969

1.2485

0.6636

0.8020

0.6547

 Smoke

1.4535

0.9419

0.5613

0.7941

1.0006

 Drink

1.0196

0.9764

0.4064

1.4555

0.2064

 Drug

0.3170

0.7596

0.0340

0.8548

1.0006

 Lived with both biological parents

0.1086

0.3602

1.3488

1.7236

0.6007

 Parent black

0.1705

0.7058

0.7489

0.5830

1.0086

 Parent college

1.7228

2.2937

1.6666

2.5438

0.0514

 Parent religious

0.8419

0.1203

0.1384

1.4171

1.0054

 Parent smoke

0.3997

1.4740

1.0769

0.9778

0.0947

 Parent drink

0.0714

0.5818

0.6468

0.3531

0.3871

 Welfare

0.2952

0.0448

0.4008

1.1980

0.0006

 Faithful

0.5262

0.0809

0.8303

1.6178

0.0600

 Life long

1.1343

0.5889

0.5481

0.5417

0.5017

 Finance

0.0992

1.7737

0.7168

2.0731

1.7321

 Live together

0.2014

0.3418

1.5616

2.1004

0.6547

Table 3

Probit estimates of the propensity score

Variables

Coefficient

Standard Error

p > |z|

Male

−0.112

0.079

0.159

Age

0.221

0.025

0.000

College

−0.431

0.153

0.005

Has siblings

0.084

0.079

0.289

Born USA

0.026

0.092

0.775

Log income

0.019

0.017

0.282

Religious

0.099

0.116

0.395

Obese

−0.079

0.108

0.467

Birth weight

0.073

0.031

0.018

Self-reported poor health

−0.050

0.141

0.723

Depressed

0.152

0.124

0.221

Smoke

−0.019

0.122

0.877

Drink

0.009

0.007

0.229

Drug

−0.088

0.113

0.438

Lived with both biological parents

−0.031

0.087

0.718

Parent black

−0.127

0.132

0.336

Parent college

0.010

0.117

0.930

Parent religious

0.133

0.148

0.369

Parent smoke

0.050

0.094

0.596

Parent drink

−0.039

0.085

0.641

Welfare

−0.061

0.087

0.480

Faithful

0.191

0.151

0.207

Life long

0.531

0.126

0.000

Finance

−0.184

0.080

0.022

Live together

−0.013

0.079

0.873

 

Log likelihood = −686

Pseudo R2 = 0.101

N = 2,581 (treated = 225, control = 2,356)

Table 4 presents our OLS estimates for each risky health behavior along with Gaussian, Epanechinikov and Uniform (radius) Kernel matching estimates. To assess the sensitivity of the matching estimates to the choice of bandwidth (or radius) we report results using different bandwidths. We report estimates for our main variable of interest only. However, it is important to note that our model controls for past participation in risky behaviors besides other demographic, health, and parental characteristics. An important result from our estimation is that participation in risky behavior at the baseline is positively related to current risky behaviors (not reported), and these effects are the largest in magnitude among all the covariates. This highlights the needs for controlling past risky behavior because for most individuals, engagement in risky behavior could be habitual. Not controlling for this could have provided us with an overestimation of the effect of marriage.
Table 4

Estimated effect of marriage on risky health behaviors (ATT)

 

Parametric Estimate

Matching

  

Epanechnikov

Uniform

OLS

Gaussian

h = 0.01

h = 0.005

r = 0.01

r = 0.005

Daily smoker

−0.011 (0.020)

−0.011 (0.021)

−0.008 (0.022)

−0.004 (0.023)

−0.007 (0.019)

−0.005 (0.019)

Number of cigarettes

0.439 (0.478)

0.360 (0.419)

0.418 (0.437)

0.432 (0.502)

0.561 (0.436)

0.567 (0.431)

Heavy drinker

−0.044 (0.026)

−0.043 (0.026)

−0.051 (0.029)

−0.045 (0.030)

−0.037 (0.026)

−0.038 (0.025)

Drunk

−0.100** (0.028)

−0.099** (0.027)

−0.106** (0.030)

−0.097** (0.033)

−0.106** (0.031)

−0.104** (0.031)

Drug

−0.053* (0.024)

−0.059* (0.023)

−0.053* (0.026)

−0.054* (0.028)

−0.060* (0.025)

−0.062* (0.026)

N treated

225

225

225

225

225

225

N controls

2,356

2,356

2,356

2,356

2,356

2,356

Each cell represents a separate regression; Control Group: All Unmarried; Robust standard errors (OLS) and bootstrapped standard errors (Matching) are reported in parenthesis. Standard errors for the matching estimators are obtained by bootstrap with 500 replications; Statistical significance: ** p < 0.01, * p < 0.05,  p < 0.1; Sets of controls include all variables from Table 1

Our OLS estimates show that African-Americans who are married are less likely to engage in excessive alcohol use, getting drunk, and drug use. On average individuals who are married are 4.4% less likely to engage in heavy drinking, 10% less likely to get drunk, and 5.3% less likely to use any drugs. While the matching estimates confirm the direction of the effect implied by the parametric results, they suggest that being married reduces the individual’s likelihood of engaging in risky behavior relative to if the individuals remained unmarried (marriage protection effect). The coefficients are also larger in magnitude compared to the OLS estimates; however, the effect of marriage on heavy drinking is only significant in two of the matching algorithms. The finding, that on average, the outcome difference between a given treated individual and an individual in the control group is smaller (OLS) than the outcome difference between the same treated individual and an individual in the control group (PSM) who exhibit similar characteristics implies that marriage yields some protective benefits. Similar to the models estimated without propensity scores, marriage did not have a statistically significant effect on either the probability of being a regular smoker or on the number of cigarettes smoked. The insignificant effect of marriage on smoking could possibly be due to the fact that smoking may not be considered necessarily as a detrimental behavioral outcome that requires immediate curtailing upon marriage as opposed to heavy drinking, getting drunk or drug use. Also the negative health outcomes of smoking tend to unfold more in the long run in addition to it being related to a high level of dependency or addiction. The effects of marriage on smoking may be a longer run phenomenon and may require a longer panel of data to be estimated empirically.

As discussed previously, marriage may induce better health outcomes (reduction in risky behavior) for two primary reasons: social norms and monitoring of a partner. Differences in health endowments or background characteristics are less plausible explanations because we match married individuals with unmarried individuals who are similar in these characteristics. However, given that some of the benefits to marriage could extend to cohabitating individuals as well (e.g. monitoring of a partner), we re-estimate the models with cohabitation as the control group. The objective is to compare individuals with similar characteristics, while indicating who made different choices on their union status. If we find a significant effect then this would imply that marriage as a union could potentially exhibit more protective benefits compared to cohabitation. These results reported in Table 5 indicate better outcomes for the treated (married) in terms of heavy drinking, getting drunk, and drug use. Specifically, our results indicate that compared with those who cohabit, married individuals are on average 10% less likely to engage in heavy drinking, 8% less likely to get drunk, and 10% less likely to use any drug. This pattern is consistent with the idea that in cohabitation the norms of non-engagement in health compromising behaviors are less clear (due to greater instability), and also individuals in a cohabitating relationship behaves more like singles (Rindfuss and VandenHeuvel 1990). Thus gains from non-marital unions are less compared to marital unions.
Table 5

Estimated effect of marriage on risky health behaviors (ATT)

 

Matching

 

Epanechnikov

Uniform

Gaussian

h = 0.01

h = 0.005

r = 0.01

r = 0.005

Daily smoker

−0.040 (0.029)

−0.035 (0.034)

−0.026 (0.038)

−0.063 (0.033)

−0.055 (0.037)

Number of cigarettes

0.531 (0.476)

0.569 (0.524)

−0.783 (0.649)

0.433 (0.566)

0.586 (0.612)

Heavy drinker

−0.097* (0.038)

−0.094* (0.046)

−0.085 (0.049)

−0.098* (0.042)

−0.104* (0.043)

Drunk

−0.083* (0.039)

−0.081 (0.046)

−0.081 (0.057)

−0.075* (0.044)

−0.095* (0.050)

Drug

−0.114** (0.039)

−0.109* (0.044)

−0.085 (0.057)

−0.104** (0.040)

−0.100* (0.047)

N treated

225

225

225

225

225

N controls

333

333

333

333

333

Each cell represents a separate regression; Control Group: All Cohabitators; Bootstrapped standard errors (Matching) are reported in parenthesis. Standard errors for the matching estimators are obtained by bootstrap with 500 replications; Statistical significance: ** p < 0.01, * p < 0.05,  p < 0.1; Sets of controls include all variables from Table 1

As a further robustness check we ran our analysis only on Non-Hispanic Whites. A similar pattern of marriage effects in the reduction of health compromising behavior among Non-Hispanic Whites would imply that the marriage protection effect pertains to other demographic group as well, and this will provide further evidence regarding the beneficial effect of marriage among the African American population. We conduct our analysis only on the Non-Hispanic Whites because the Add Health data does not have enough samples in other race categories to obtain better quality matches. Additionally, there is concern on how we define other race categories given small sample sizes and the implications for the results (Thornton and White-Means 2000). From our results in Table 6 we can see that compared to unmarried Non-Hispanic Whites, married Non-Hispanic Whites are on average less likely to engage in heavy drinking, getting drunk, and using drugs. This is very similar to our results for African Americans (Table 4), but the magnitudes of the coefficients are higher. Also similar to our results in Table 4, marriage among the Non-Hispanic Whites did not exhibit a significant effect on smoking.
Table 6

Estimated effect of marriage on risky health behaviors (ATT)—only Non-Hispanic Whites

 

Matching

 

Epanechnikov

Uniform

Gaussian

h = 0.01

h = 0.005

r = 0.01

r = 0.005

Daily smoker

0.005 (0.014)

0.006 (0.014)

0.009 (0.016)

0.011 (0.014)

0.011 (0.015)

Number of cigarettes

0.264 (0.234)

0.258 (0.257)

0.300 (0.266)

0.353 (0.252)

0.367 (0.249)

Heavy drinker

−0.140** (0.014)

−0.129** (0.014)

−0.124** (0.014)

−0.179** (0.016)

−0.178** (0.013)

Drunk

−0.189** (0.015)

−0.176** (0.016)

−0.172** (0.016)

−0.230** (0.016)

−0.229** (0.014)

Drug

−0.108** (0.011)

−0.101** (0.012)

−0.101** (0.011)

−0.136** (0.011)

−0.134* (0.011)

N treated

1258

1258

1258

1258

1258

N controls

5948

5948

5948

5948

5948

Each cell represents a separate regression; Control Group: All Unmarried; Bootstrapped standard errors (Matching) are reported in parenthesis. Standard errors for the matching estimators are obtained by bootstrap with 500 replications; Statistical significance: ** p < 0.01, * p < 0.05, Sets of controls include all variables from Table 1

Discussion

In this paper we sought to estimate whether marriage can result in a reduction in risky health behaviors among African-Americans. Our results indicate that marriage among African Americans may lead to a reduction in risky health behaviors, especially for alcohol consumption and drug use. Such benefits are also greater compared to those who are in a cohabiting relationship. Our models are analyzed after accounting for selection bias, thus it can be suggested that marriage may exhibit a protection effect, which might be larger than the protection effect in other union statuses, such as cohabitation.

The significant decline in alcohol consumption and drug use implies that the social norms associated with “cleaning up one’s act” (Duncan et al. 2006) that surrounds marriage might facilitate such a reduction. The “monitoring of a partner” entailed by marriage might also be a contributing factor to such a decline. The absence of such declines in cohabitation could be attributed to the fact that compared with marriage cohabitation may be viewed as an incomplete institution, where the norms are less clear. So even if cohabitation can lead to some reduction in risky behavior, the norms are less agreed upon (Hofferth 2006). This could potentially offset the “monitoring of partners” potential of the co-residence because cohabitation, unlike marriage, does not involve more engagement with one’s partner in addition to not being a long-term commitment like marriage.

The insignificant effect of marriage on smoking is similar to the previous literature (Duncan et al. 2006) and could be a result of the addictive nature of cigarettes, or it might be because smoking is not yet considered as something that has to be given up after marriage like heavy drinking and drug use. It could also be a result of smokers selecting other smokers as partners (Clark and Etile 2006) to a higher degree than alcohol or drug use. This brings us to a potential shortcoming of the paper, which is the possibility of homophily or positive assortative mating on risky behaviors. Individuals who smoke or like to drink or use drugs might be more likely to marry or live with a person with similar preferences. However, because Add Health does not contain information on spouse’s behavior, we are unable to control for this. But this positive assortative mating also means that the findings reported in the paper are conservative estimates of the impact of marriage on risky behaviors.

Our data set also does not allow us to account for the quality of marriage, much like the previous literature. A better quality marriage may exhibit a larger effect in the reduction of risky health behaviors, whereas lower quality marriages might not have any significant effect on reduction of such behaviors. Another limitation of our study is the low sample size of our treated group. Although the sample size is large enough (and similar to the previous studies using propensity score matching–see for example Liu and Heiland 2010) to conduct and obtain efficient statistical matching—it is not large enough to conduct our analysis separately by gender. Our gender analysis did not reveal any differential outcomes among males and females. We attribute such similar effects to the low sample size. We were also unable to control for cohabitation history in this study. Although current cohabitation status is important, it is plausible that cohabitation history would give better insight into the marital selection process of African Americans. This should be examined in future studies. In relation to the data, we did not utilize the newly released Wave IV (2007–2008) data of the Add Health survey since we wanted to focus on younger African-American; thus, future studies could utilize the Wave IV data to examine whether the health benefits documented here for young African-Americans can also be found for an older cohort.

From a policy perspective, our results indicate the positive effect that marriage can have among African Americans in terms of reducing participation in risky health behaviors. This positive benefit was not evident in other forms of union, like cohabitation. As part of welfare reform reauthorization in 2002, marriage promotion policies have been put in place. These policies are also focused on divorce reduction and responsible fatherhood. Wherry and Finegold (2004) show that these policies, though beneficial, have minimal impacts on African-Americans because African-Americans are the least likely to be married. They suggest that policies should expand the range of its focus on promoting marriages more. However, one needs to be careful when promoting marriage policy because the emphasis needs to be on the quality of marriages and not just the encouragement of union formation. Because of the shortage of African American men relative to women, marriage market conditions are considered one possible explanation for low marriage rates among African Americans. A combination of social conditions, including imbalanced sex ratios at birth as well as excess mortality and high rates of incarceration among young African American men, has contributed towards such shortages (Harknett and McLanahan 2004). In this regards, a policy focus on educational attainment might also be beneficial for African Americans in terms of not only health improvement but also to increase the supply of marriageable men.

A possible extension of this study would be to analyze the effect of marriage on other health-related behaviors such as diet, exercise, and risky sexual behaviors. This is especially relevant among African Americans given a higher percentage of diabetes, heart disease, obesity, and sexually transmitted diseases are prevalent among them. Another possible extension would be to analyze whether marriage plays any role in the reduction of participation in violent behaviors because it is a major contributor to mortality among African American males.

Acknowledgments

The authors would like to thank Frank Heiland, CUNY-Baruch College and Shirley Liu, Institute for Defense Analyses, for helpful comments and suggestions. The views expressed here are those of the authors and do not necessarily reflect the views of the Food & Drug Administration. This research uses data from Add Health, a program project designed by J. Richard Udry Peter S. Bearman, and Kathleen Mullan Harris, and funded by a grant P01-HD31921 from the National Institute of Child Health and Human Development, with cooperative funding from 17 other agencies. Special acknowledgment is due Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. Persons interested in obtaining data files from Add Health should contact Add Health, Carolina Population Center, 123 W. Franklin Street, Chapel Hill, NC 27516-2524 (addhealth@unc.edu).

Copyright information

© Springer Science+Business Media, LLC 2010