Maternal depression and the production of infant health.

1. Introduction

Understanding the correlates of infant health has become an issue of great importance since we have learned that it has tremendous implications for childhood and adult health and well-being. Low birth weight is a key indicator of infant health, as low birth weight babies have much higher mortality rates. In addition, caring for these low birth weight babies is substantially more costly (Currie and Gruber 1996). Perhaps even more compelling is the mounting evidence that the negative effects of being born at a low birth weight persist well into adulthood and can sometimes take many years to appear. (1) Reducing the incidence of low birth weight, possibly through improved prenatal care, is therefore a worthwhile policy goal, but what is the most effective way to do it?

The recent, dramatic expansion in the Medicaid eligibility of pregnant women has been fueled, at least in part, by the belief that it would increase prenatal care and ultimately improve infant health. However, expanded eligibility does not necessarily lead to increased participation, as the take-up rate for Medicaid is tar below 100% (Gruber 1997). Furthermore, if individuals drop their private insurance, which is presumably of superior quality, on receipt of Medicaid, then expanding Medicaid may not improve infant health. (2) Finally, even if expanding Medicaid coverage leads to increased prenatal care, those increases may not substantially improve infant health if prenatal care is not effective.

We suggest that a potentially important element has been left out of the policy discussion: the mother's mental health. Epidemiological research has shown that infants of depressed mothers show signs of poorer health (Abrams et al. 1995; Field 1995, 1998; Dawson et al. 1997a, 1997b; Jones et al. 1997; Locke et al. 1997) and are more likely to be preterm (Orr and Miller 1995; Orr, James, and Prince 2002). Maternal depression is not a rare event: The lifetime risk for depression for women is estimated at 10-25% and peaks during their childbearing years (Wisner et at. 1999; Desai and Jann 2000). Low-income women are even more susceptible to depression, and there is evidence that those on welfare suffer greater depression than other low-income women (e.g., Lennon, Blome, and English 2001). Yet current economic research completely overlooks the role that maternal depression plays in infant health.

In studies of welfare reform, there is a growing recognition of the role depression plays in preventing women from getting off welfare and finding and keeping employment. For instance, Lennon, Blome, and English (2001) provide an extensive survey of the prevalence of depression among low-income women and the consequences it has for them. They also discuss and evaluate different kinds of welfare and unemployment programs that incorporate treating mental illness. Our results suggest that perhaps the Medicaid program and, more generally, the health care providers who treat pregnant women also need to recognize the important role that maternal depression plays in infant health.

We begin with a standard infant health production model in which "depression" may affect infant health directly as part of the technology of production and also as a factor that influences the choice of inputs (e.g., prenatal care). Using the National Maternal and Infant Health Survey (NMIHS), we estimate birth weight and prenatal care equations that incorporate measures of maternal depressive symptoms. (3) Because past researchers have found significant differences between whites and blacks, we stratify by race. The NMIHS does not contain the ideal measure of maternal depression--a diagnosis of depression during the pregnancy--and so we construct several measures to verify the validity of our results. One byproduct of constructing these measures is an investigation into factors associated with maternal depressive symptoms. Across all measures and samples, we consistently find that maternal depressive symptoms have a negative direct effect on birth weight and that these symptoms may operate through other channels, such as reduced prenatal care and increased unhealthy behaviors.

2. Production of Infant Health

We extend the infant health production model of Rosenzweig and Schultz (1982, 1983) by adding "depression" as both a factor that affects the technology of production and a taste variable. The mother's utility function is defined as

(1) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII.] utility function,

where H reflects her infant's health, D is her mental state or "depression," and X is a composite commodity. The mother then maximizes this utility function by choosing prenatal care, M, subject to the infant health production function and her budget constraint,

(2) H = H(M,D,Z) infant health production

(3) PM + X = I budget constraint.

Referring to Equation 2, prenatal care, M, is expected to enhance infant health, H, whereas depression, D, is expected to adversely affect infant health. In addition, other variables (Z) such as the mother's education, age, height (as suggested by Warner 1995, 1998), and past medical history, as well as other maternal behaviors, are expected to directly affect infant health. Equation 3 represents the budget constraint, where P is the out-of-pocket price per unit of prenatal care, I is income, and the price of the composite good is normalized to 1.0.

The woman then chooses the prenatal care, [M.sup.*], that maximizes her utility subject to Equations 2 and 3, which yields the two-equation model typically estimated--the desired prenatal care demand equation,

(4) [M.sup.*] = M(D,Z,P,I),

and the desired infant health equation,

(5) [H.sup.*] = H([M.sup.*], D, Z).

This is the typical framework employed by recent studies of infant health and prenatal care (e.g., Grossman and Joyce 1990: Warner 1998; Currie and Grogger 2000). It demonstrates how variables such as age and education can have both a direct and indirect (through prenatal care) effect on infant health. We introduce maternal depression as another such factor that can have dual influences--both on the decision to get prenatal care as well as having a direct effect on the production of infant health. This view is consistent with the epidemiological literature, such as Orr and Miller (1995), which also has recognized these dual avenues. In particular, Orr and Miller (1995, p. 169) in their review of maternal depressive symptoms and pregnancy outcomes note the "two primary mechanisms by which depressive symptoms might influence birthweight": by possibly leading to more harmful behaviors (including delayed prenatal care) and by "a more direct association." as "depressed mood has been increasingly linked to biochemical/hormonal alterations in the body."

One might argue that depression may be jointly produced with infant health (health inputs affect both). However, in a recent study, Williams et al. (1999, p. 64) find that among family physicians, general internists, and obstetricians-gynecologists (OB-GYNs), OB-GYNs stand out as having higher physician barriers to treating depression (e.g., having low confidence in or incomplete knowledge of treatment). It therefore seems unlikely that the woman is seeking prenatal care to treat her depression. We also hesitate to make the assumption implied by such an argument that maternal depression is a choice variable and under the mother's control. (4) However, maternal depression may be endogenous to birth weight in that the mother's unobserved health endowment may affect both. With this caveat in mind, our estimated effects of observed maternal depressive symptoms should be viewed as measuring their value as signals of maternal health and behavioral problems that could lead to a poor birth outcome.

Despite our reservations about treating depression as a choice variable, we modify this framework to allow for such a possibility in our empirical analysis. In essence, we specify a reduced-form maternal "depression" production function, which we then use to construct an instrumental variable for D to be used in estimating the structural infant health (Eqn. 5). We therefore also estimate the model using predicted values of depression as a way of dealing with both the possible endogeneity of maternal depression and the limitations of our depression measure. One by-product of this research, then, is an empirical investigation into factors associated with maternal depressive symptoms, which have been overlooked in health economics research. (5)

This theoretical framework clarifies the identifying restrictions used in the typical infant health model; income (I) and factors affecting the price of prenatal care (P), such as insurance status and community level variables, are used to identify prenatal care. (6) Warner (1998) discusses in detail his difficulty in finding satisfactory restrictions and notes that the structural birth weight equation may be weakly identified as a result. Such weak identification has been blamed in part for the lack of effectiveness of prenatal care found by many researchers (see, for example, Currie and Grogger 2000, who use policy changes in Medicaid and welfare as identifiers).

We wish to isolate the effect of adding depression to the standard model of infant health, and so we employ typical identifying restrictions. In particular, we use a number of aggregate variables--health care price index aggregated to the state level, the ratio of Medicaid fees to private fee levels paid to OB-GYNs in the state (as in Gray 2001), population density, number of OB-GYNs per 1000 state population, n umber of general practitioners per 1000, and number of hospitals and HMOs per 1000. We also include some individual-level variables--whether the mother's prenatal care was paid for by Medicaid or private insurance and her income. As discussed in more detail in the results section, we subject the instrument sets to the usual tests, which they pass for every specification reported.

The variables common to both specifications are mother's height, age, and age squared; mother's education and education squared; number of own children living with her; whether she is living with the father; and her medical history (parity, number of prior fetal deaths). We also include the gender of the infant in both equations for consistency, although it is probably not known at the time that prenatal care is sought. This is our structural model.

As another test of the effects of depression on birth weight, we also estimate a reduced-form (or quasi-structural form, in the case of models in which depressive symptoms are viewed as endogenous) birth-weight equation in which prenatal care is eliminated and birth weight is regressed on all of the variables in the system. In this way, the combined direct and indirect effects of depressive symptoms on birth weight are captured without the need to identify prenatal care. Carrie and Grogger (2000) employ a similar exercise in determining the effects of welfare and Medicaid on infant health.

A final complication in the model are maternal behaviors such as smoking and drinking alcohol, which are strongly associated with poor birth outcomes yet are likely a maternal choice. Including these variables and treating them as endogenous further strains an already weakly identified birth-weight equation. Most studies therefore either omit these behaviors (e.g., Liu 1998; Currie and Grogger 2000) or treat them as exogenous, sometimes using information on behaviors before the pregnancy (e.g., Grossman and Joyce 1990; Warner 1998; Kaestner 1999). Because we suspect that such maternal behaviors are also strongly related to maternal depression, we estimate the model both with and without these behaviors. When they are included, we use the behavior before the pregnancy to avoid endogeneity bias. In this way, we can investigate whether maternal smoking and drinking is an important avenue for maternal depression to affect birth outcomes.

3. Maternal Depression: Cause, Effect, and Data Issues

"Maternal depression" is a term that is sometimes used to describe two distinct phenomena. "a major depressive disorder" and "high levels of depressive symptoms" (Orr and Miller 1995, p, 166). "A clinical diagnosis of depression is generally made on the basis of the presence, for at least 2 weeks, of a depressed mood and/or loss of interest or pleasure in activities that are usually a source of enjoyment, along with other symptoms such as appetite and sleep disturbance; fatigue and loss of energy; feelings of worthlessness; and difficulty in concentrating, These symptoms represent a change in functioning from that which is normal for the individual" (Orr and Miller 1995, p. 166). In contrast, many screening tools, such as the Center for Epidemiological Study of Depression (CES-D) Scale used in our analysis and described shortly, are designed to estimate the prevalence of high levels of depressive symptoms. (7) Therefore, our empirical study is actually studying the effect of depressive symptoms on infant health, not depression per se.

Ample evidence exists in the medical literature that maternal depression is associated with poor infant and child health and development. However, a persistent question is the direction of causality--do depressed mothers lead to unhealthy infants, or do unhealthy infants lead to depressed mothers? As summarized in Field (1995), the answer appears to be a combination of both options. One strong piece of evidence that maternal depression causes poor infant health is provided by Abrams et al. (1995), who look at the behavior of newborns born to depressed mothers versus those born to nondepressed but otherwise comparable mothers. They find that "infants of depressed mothers are already showing nonoptimal behaviors at birth" and that they are "reminiscent of newborns who are small for date" (p. 238). Even more relevant for our study is the evidence provided by Orr, James, and Prince (2002), who find that mothers with higher depressive symptoms (measured by the CES-D) are more likely to have a preterm birth.

In one respect, the issue of causality is irrelevant for us because we are looking at birth weight. Interactions that occur after the birth should have no effect. However, the tinting of our measures of depressive symptoms raises again the issue of causality. The 1988 NMIHS surveyed the mother on average 17 months after the birth of her child, and the depressive symptoms measure refers to those experienced in the past week. (8) Therefore, our measure of depressive symptoms is taken after the delivery of the infant.

This poses two problems for us. First of all, our measure may be only weakly correlated with depressive symptoms during the pregnancy, and it may also capture postpartum depression. However, Najman et al. (2000), in a study that follows women up to 5 years after the birth, lessen this concern considerably by finding that postpartum depression is usually short-lived and mild and that many cases of depression that occur after the birth are recurrences of a previous episode. Moreover, research on postpartum depression in general consistently finds that depression during pregnancy is one of the strongest predictors of postpartum depression (e.g., Gotlib et al. 1991; Seguin et at. 1999; Beck 2001; Josefsson et al. 2001). Thus, to the extent that our measure is capturing postpartum depression, there is strong evidence that it is reasonably correlated with prenatal depression.

We can also gain some reassurance from the broader depression literature. In Kessler et al. 1997, for instance, over 70% of individuals with either "minor" depression (two to four symptoms) or "major" depression (five or more symptoms) suffer recurrent episodes of depression and have had more than eight episodes, on average. Again, this indicates that our measure of depressive symptoms may be reasonably correlated with prenatal depressive symptoms, and, to the extent that our measure is not a good proxy for prenatal depression, any estimated effects should be biased toward zero, making our estimates conservative.

The second problem, however, could lead to a falsely stronger effect via reverse causality. As noted above and in Field (1995), having a low birth weight or otherwise unhealthy infant may cause a woman to become depressed. However, Singer et al. (1999) provide two insights. First, it is not the birth weight of the infant that matters but whether they are high or low risk. Second, the effects of having a sick infant vary as the child grows older. We use these insights in developing our instrument for depressive symptoms.

Another possible problem for our measurement of depressive symptoms occurs because neither the 1988 nor 1991 NMIHS contain information about treatment. This could be a serious problem if the treatment itself has an effect on birth weight. A woman who is successfully treated for depression would exhibit few or no depressive symptoms and yet may have a poor birth outcome precisely because of the treatment. Another possible problem is if the choice by the mother to receive treatment is in some way correlated to the birth outcome (if, e.g., women who more heavily discount the health of their infants opt for pharmacological treatment). We believe, however, that it is unlikely that this bias is widespread in our data and that, to the extent it exists, it should bias our results against finding an effect of depressive symptoms on birth weight.

Existing evidence regarding the safety of pharmacological treatment of depression is far from definitive, although some suggest that exposure may lead to increased risk of premature birth and lower birth weight (e.g., Grush and Cohen 1998; Wisner et al. 1999). (9) However, this uncertainty about safety during pregnancy likely leads to low treatment rates among pregnant women, which are corroborated by existing evidence. (10) Thus, given the apparent low rate of pharmacologic treatment among pregnant women and the uncertain effects on birth weight of such treatment, we are doubtful that this has a serious effect on our results. Furthermore, it appears that whatever effect there is would tend to bias our results toward zero, as women with low apparent depressive symptoms have adverse birth outcomes resulting from the treatment they are receiving. Finally, to the extent that it is a choice the woman makes, one would expect this choice to be made by women who either expect greater costs of going untreated or more heavily discount the costs to their infants. Both cases should lead to a downward bias (toward zero) of the effect of observed depressive symptoms on birth weight.

Given the possible problems with our measure, we conduct sensitivity checks by constructing three additional measures and then using them to check the robustness of our results. The first new measure uses the depression measure taken at the 1991 NMIHS follow-up to construct an average of depressive symptoms at two points in time [i.e., (CES-D in 1988 NMIHS + CES-D in 1991 NMIHS)/ 2]. Given the recurrent nature of depression, using the level of depressive symptoms at two points in time may be a superior predictor of a general susceptibility to depressive symptoms. Even in our data there is evidence of a recurrent quality, as these two measures are fairly highly correlated. However, this measure still suffers from the possibilities of reverse causality (i.e., sick infants cause depressed mothers) and that the effects of choosing treatment are ignored.

Our other two measures use the survey measures along with other information about the mother to predict her depressive symptoms during pregnancy. This approach is like using the two-sample instrumental variables approach of Angrist and Krueger (1992, 1995), in which the mothers, at a later date, are providing instruments for themselves during the pregnancy. Specifically, we estimate the 1988 NMIHS measure of depression as a function of characteristics of both the mother and infant, as well as how much time has lapsed since the birth. We then predict her depression during the pregnancy both by using variable values during the pregnancy instead of after it (where possible) and by setting infant characteristics and the length of time since the birth to the sample averages. In this way, we hope to capture any changes in the mother's environment between the pregnancy and the time of the survey and also to control for any depressive symptoms that are induced by having a sick infant. Using a predicted measure of depressive symptoms also addresses the issue of its possible endogeneity, in that we are essentially using an instrumental variable.

The specific depressive symptoms equation we estimate is 1988 NMIHS CES-D = f [receives Aid to Families with Dependent Children [(AFDC).sup.s], receives [Medicaid.sup.s], has private [insurance.sup.s], income in the 12 months before pregnancy, cohabits with [dad.sup.s], number of kids [cohabiting.sup.s], mother's education and [education.sup.2], mother's age and [age.sup.2], age at first marriage, number of marriages, never married, number of previous fetal deaths, cigarettes [smoked.sup.s], alcoholic drinks before pregnancy, marijuana or cocaine before pregnancy, exercised before pregnancy, wanted pregnancy, urban, state mental health agency expenditures per capita, psychiatrists per 1000, psychologists per 1000, social workers per 1000, number of multiservice mental health organizations per 1000, health maintenance organizations per 1000, pregnant [now.sup.s], length of time since birth, [length.sup.2], length x number of sick baby visits per month, [length.sup.2] x number of sick visits].

The superscript's denotes that the variable's value refers to the time of the survey; these variables are all replaced with their values during the pregnancy (with the exception of smoking, which is replaced with its value before the pregnancy). For example, receipt of AFDC has been found to be highly correlated with depressive symptoms (Lanzi et al. 1999; Lennon, Blome, and English 2001), so the CES-D measure is estimated as a function of AFDC receipt at the time of the survey. We then use whether the mother received AFDC during the pregnancy to predict her depressive symptoms during the pregnancy. Other variables did not change during the period (e.g., education, age at first marriage) or are only available at one point in time (e.g., family income), so the same values are used for estimation and prediction.

Several maternal behaviors are only available before and during the pregnancy: exercise, drinking alcohol, and using drugs. For these variables we used the measures taken before the pregnancy to estimate depression because we believe they more closely approximate the mother's behavior after the birth. (11) We also use smoking, drinking, drug, and exercise behaviors before the pregnancy to predict depression to avoid their possible endogeneity with birth weight, which, as we note in section 2, is a common practice in infant health models. However, because many infant health studies omit these behaviors, we estimate an alternative specification that omits them from the depression equations; because of possible endogeneity, we also omit whether the pregnancy is wanted.

We include several state-level variables that attempt to capture the availability of treatment for depression. These include the number of treatment facilities and HMOs, state expenditures on mental health agencies, and practitioners who might treat depression. As a result of the insights from Singer et al. (1999) and Najman et al. (2000), we allow the length of time since the birth to enter in nonlinearly and affect the impact of having a sick infant, the latter of which is measured by the average number of sick visits per month for the infant. To purge our predicted measure of the effects of having a sick child and the passage of time, we set the number of sick visits and the length of time between the interview and the birth to the sample means. (12)

Therefore, the variables in our depressive symptoms equation that provide identifying variation are the mothers' receipt of AFDC, her marital history, variables measuring the availability of mental health treatment, and when included, behaviors and whether the mother wanted the pregnancy. In sum, we use four different measures of depression: the 1988 NMIHS survey measure, the average of the 1988 and 1991 NMIHS survey measures, a predicted measure using the above equation and appropriate values of the variables, and a predicted measure excluding behaviors (smoking, drinking, drugs, and exercise) and whether the mother wanted the pregnancy. Admittedly, none of these measures is "ideal"; however, finding similar results across the measures adds credence to their validity.

4. Data Description

Our primary data come from the NMIHS 1988, which consists of publicly accessible data published by the National Center for Health Statistics. NMIHS contains information on women who were pregnant in 1988, although the actual interview took place on average 17 months after the delivery (and thus after 1988). Following Kaestner (1999) and Warner (1998), only data from the livebirth sample are used. Grossman and Joyce (1990) point out that this selection causes bias because women who experience a live birth may be healthier or may have desired their child more. However, both Kaestner (1999) and Warner (1998) omit this correction, and Gray (2001) explored its effect by including the fetal death sample but found no substantial difference. (13) Therefore, for simplicity and to conform to past studies using the NMIHS, we exclude it from our model.

Another consideration is that the NMIHS is not a simple random sample. It oversamples fetal deaths and infant deaths. Within the live birth sample, it oversamples blacks and, for both blacks and nonblacks, very low and low-birth weight babies. Both because of this oversampling and because previous research finds important racial differences in birth weight (e.g., Warner 1995, 1998; Liu 1998), we stratify our sample into blacks and whites. This permits us to explore whether differences in maternal mental health can help explain some of the racial differences in birth outcomes that others have found. To control for the oversampling of very low and low-birth weight babies, we follow Warner (1998). Kaestner (1999), and Gray (2001) and use the sample weights provided by the NMIHS.

As most studies do, we eliminate multiple births because such babies tend to be born at shorter gestation and lower birth weights and because it leads to multiple birth observations that share one prenatal care observation. To construct the final samples, we begin with the 9146 live, singleton births. We then eliminate, in the following order, the small number of women who bad no prenatal care (N = 284), (14) those missing information for number of prenatal care visits (N = 10), and those missing information for parity (N = 79). We follow the typical practice of omitting teenagers (<19 years old) and much older mothers (>50 years), which eliminates 1168 observations. Following Warner (1998), gestations of fewer than 20 weeks or greater than 45 weeks (N = 48) and birth weights below 400 g or above 6000 g (N = 35) were deleted. We also eliminate the 30 observations from Hawaii and the 28 observations from Alaska. After all of these exclusions, there are (9146-284-10 -79-1168-48-35-30-28 =) 7464 observations, of which 3342 are non-Hispanic blacks and 3232 are non-Hispanic whites.

A large number of observations do not have the 1991 NMIHS follow-up information, so that the depression measure used in the second measure is unavailable. When the 1991 depression variable is used, we therefore use smaller subsamples of 2605 non-Hispanic blacks and 2766 non-Hispanic whites in estimation. Likewise, many of the observations had missing values for number of sick baby visits and length of time between the 1988 interview and the birth. Recall that these variables are only necessary to obtain estimates for Equation 6 and are then set to sample means to predict depressive symptoms. We therefore use yet another sample to estimate the depressive symptoms. We use these different samples both to bolster sample size where possible and to minimize any bias that might be introduced by excluding observations with missing data. (15,16)

The descriptions, means, and standard deviations of the variables used in our analysis are reported in Appendix A. The key variables in our analysis are prenatal care, birth weight, and maternal depressive symptoms. Although there are a number of possible measures of prenatal care, the most widely used and least criticized measure is the onset (or delay) of prenatal care, as it is widely believed that early prenatal care is important to a healthy pregnancy. In addition, it does not suffer as much from reverse causality as do measures that include number of visits, whereby difficult pregnancies may lead to more visits.

Likewise, birth weight is the most commonly used measure of infant health in studies that use individual-level data, especially in those employing structural models. Some also estimate the incidence of low birth weight (e.g., Joyce 1999; Kaestner 1999; Gray 2001), as it is well established that low birth weight is associated with greater health risks and costs. Although birth weight (continuous) is our primary measure of infant health, we include the incidence of low birth weight (discrete) as a measure when we estimate the reduced form equation.

The last key variable in the model is the mother's mental health, which is measured in the NMIHS with the CES-D scale. This scale was developed in 1969 as part of the Community Mental Health Assessment (CMHA) program to identify the presence and severity of depressive symptomatology in the general population. However, it was not intended to discriminate among different types of depression or to distinguish primary depressive disorders from secondary depression (Radloff and Locke 1986). The CES-D scale is a self-reported survey consisting of 20 questions concerning depressed mood, feelings of guilt, worthlessness, helplessness, hopelessness, psychomotor retardation, loss of appetite, and sleep disturbance. Each response is scored from 0 to 3 on the basis of the frequency of the occurrence in the past week, yielding a range of 0 to 60. Higher scores reflect more depressive symptoms.

The literature (e.g., Roberts and Vernon 1983; Radloff and Locke 1986; Zimmerman and Coryell 1994; Locke et al. 1997; Lennon, Blome, and English 2001) has consistently used a score of 16 or greater to indicate depression. Given this agreed-on threshold, we also estimate our model using a depression dummy variable that equals 1 if the CES-D exceeds 15. The CMHA found that these scores are correlated with the presence of clinical depression as diagnosed by a psychologist (Radloff and Locke 1986). Our variable is not, therefore, a clinical diagnosis, but rather a measure of depressive symptoms. Although there are many symptom checklists, the CES-D is the most commonly used scale in studies of depressive symptoms in welfare populations (Lennon, Blome, and English 2001).

Figure 1 reports the relative frequency of the CES-D in the 1988 NMIHS survey for both samples. Two results are immediately apparent. First, the majority of women do not show depressive symptoms as classified by the CES-D. Second, black women have more depressive symptoms than white women. Whereas only 18.7% of white women have CES-D scores greater than 15, 35.5% of black women do (see Appendix A). (17) The incidence of elevated depressive symptoms is therefore almost twice as great among black women as it is for white women. Furthermore, Figure 1 reveals that this tendency continues throughout the distribution of CES-D scores: White women have higher relative frequencies than blacks at scores of 9 or less, and black women have higher frequencies than whites for scores above 9. This validates the recent concern that depression is widespread among low-income women because, as Table 1 and Appendix A reveal, our black sample has much lower incomes than our white sample. It also indicates that maternal depression could help explain at least part of the differences in birth outcomes between whites and blacks, as well as for low-income women more generally.

[FIGURE 1 OMITTED]

Table 1 reports the means of our key variables for nondepressed versus depressed (1988 NMIHS's CES-D score > 15) mothers by sample. It is immediately apparent that the depressed mothers have worse birth outcomes. Whereas the average birth weight is only about 100 g less for depressed mothers, the incidence of low birth weight (<2500 g) and very low birth weight (<1500 g) is much higher for these mothers. Likewise, prenatal care begins later for depressed mothers, and they report facing much greater barriers to obtaining prenatal care. (18) Depressed mothers appear to be disadvantaged in other ways as well: they are more likely to participate in Medicaid and AFDC and are less likely to have private insurance; they also have lower family incomes and are more likely to be never married. Consistent with Saffer and Dave (2002), they also consume higher levels of tobacco and alcohol. Finally, we also see the persistence of depressive symptoms across the two surveys. The 1991 NMIHS's follow-up CES-D scores are quite a bit higher for those who were depressed in the initial survey, and these depressed mothers are about three times more likely to be classified as depressed in the follow-up survey than those who are nondepressed.

Table 1 indicates that maternal depression may be potentially important in determining birth outcomes and may operate through several avenues, such as prenatal care and maternal behaviors. Our econometric results, presented in the next section, allow us to see whether these results remain in effect when using different depression measures and controlling for confounding factors such as age, education, economic status, and other socioeconomic variables.

Most of the other variables are also from the NMIHS and are self-explanatory, with the exception of our aggregate-level variables. Because only a woman's state of residence (not city) is identified in the NMIHS, we must aggregate all data into state-level data. The health care price index is created by the American Chamber of Commerce Researchers Association (ACCRA) Cost of Living Index, which includes price data from 256 U.S. cities. Population density is taken from the Statistical Abstracts of the United States. The ratio of Medicaid to private fees for OB-GYNs is from Loprest and Gates. StateLevel Data Book on Health Care Access and Financing (1993, p. 103). Per capita state mental health agency per capita expenditures is from Lutterman, Hirad and Poindexter (1999). The number of multiservice mental health organizations is from Center for Mental Health Services (CMHS), Mental Health, United States 1994. The number of hospitals, HMOs, OB-GYNs, general practitioners, psychiatrists, psychologists, and social workers are obtained from the Bureau of Health Professions Area Resource File. 1940-90. All such variables are converted to per 1000 state population.

5. Main Empirical Results

Table 2 reports the key results from the structural and reduced-form birth weight equations using four alternative observed measures of depression--the CES-D score in the 1988 survey, and the average of the 1991 and 1988 surveys are each used as both continuous measures and as a discrete measure of depressive symptoms. For comparison, a "typical," or baseline, birth weight structural model is also estimated and reported at the top of Table 2. (The full set of results for these models is available on request.) We estimate each variation of the model, first excluding and then including maternal smoking and drinking (before pregnancy) to investigate the relationship between maternal behaviors, depressive symptoms, and birth outcomes.

For every specification reported, we perform two tests on the quality of the instrument set: whether the variables are jointly significant in the prenatal care equation, and whether the over-identifying restrictions are supported by the data. We therefore check to be sure that the instruments have explanatory power in the prenatal care equation and that, at the same time, do not have explanatory power in the infant health equation. All specifications satisfy both tests. We also perform a joint test of the significance of our aggregate-level identifiers and find that they are jointly statistically significant as well.

Recall that the structural model allows depression to have both direct and indirect (through Onset) effects on birth weight. Because an increase in the Onset of prenatal care means less prenatal care, we therefore expect depressive symptoms to increase Onset. (19) (Likewise, we expect Onset to decrease birth weight.) We can explore the direct effect of depressive symptoms by examining its effect on birth weight when prenatal care is controlled for, In contrast, our reduced-form equations yield only an estimate of the total effect of depression.

Looking first at our baseline estimates of the structural model, it may he disconcerting to find that prenatal care has no effect on the birth weights of white infants; however, finding prenatal care to have a highly variable and often insignificant effect on birth weight is entirely consistent with past research. As noted by Currie and Grogger (2000), Warner (1998), and others, the often modest and widely variable effects of prenatal care may be in part a result of difficulties in identifying the equation. (20) Our results indicate that the onset of prenatal care has no statistically significant effect for whites. For blacks, however, delaying prenatal care by 1 week appears to decrease birth weight by anywhere from 65 to 83 grams and is statistically significant. These results are at the high end of the ranges of estimates reported by Warner (1998), who uses the same data and stratifies by race. The effects of maternal behaviors and the other (unreported) explanatory variables, such as health history, age, and education, are as typically found in the literature.

Adding depression to the model does not substantively affect the estimated effect of prenatal care (or maternal behaviors) on birth weight, which perhaps indicates that our depression measures are bringing information to the model that is new and relatively uncorrelated with the typical regressors included. For both whites and blacks there is strong evidence that depression lowers birth weight, although the direct and indirect effects of depression appear to be very different between the two groups. For black women, depressive symptoms significantly delay the onset of prenatal care, which, in turn, has a significantly negative effect on birth weight. The direct effect of depressive symptoms on birth weight, however, although negative, is usually statistically insignificant. Therefore, maternal depressive symptoms appear to affect the birth outcomes of black mothers primarily through delaying prenatal care.

The opposite is true for white mothers. Depression has no effect on the onset of prenatal care, and prenatal care, in turn, has no significant effect on birth weight. However, the direct effect of depressive symptoms on birth weight is consistently negative and often statistically significant. Our results therefore indicate that observed depressive symptoms are associated with lower birth weights for both blacks and whites, but that it operates through delayed receipt of prenatal care for black mothers.

Including maternal smoking and drinking tends to reduce the estimated effects of depressive symptoms, but the results are otherwise similar. The effect on white mothers, in particular, is affected by their inclusion: this suggests that one avenue by which maternal depressive symptoms 'directly' affect birth weight is through these behaviors. Black mothers, in contrast, are less affected, which makes sense because these behaviors are not significantly associated with prenatal care.

To avoid the problem of identifying prenatal care and to obtain estimates of the total effect of observed depression on birth weight, we next estimate a reduced-form birth equation. Perhaps even more important is whether treating maternal depression will improve poor birth outcomes; that is, reduce the incidence of low birth weight. Although it is not entirely clear that increasing birth weight will improve infant health in the general population, reducing the incidence of low birth weight undoubtedly will. To answer this question, we also estimate a reduced form probit in which a low-birth weight dummy (representing a birth weight of less than 2500 g) is the dependent variable. (21)

These reduced-form equations include all of the identifying variables from the prenatal care equation but exclude prenatal care itself; the results are reported in the second panel of Table 2. Note that income is now included in the birth weight equation. A fair criticism of our structural model is that income is a questionable identifier (even though it passed our tests) and that depressive symptoms may be capturing its effects in the birth weight equation. The reduced form estimates show that the effects of depression are operating through some mechanism other than reduced income. The same reasoning applies for Medicaid and private insurance participation. However, some might argue that these latter two variables are endogenous. To address this, we also estimate policy reduced form equations in which these two variables are replaced with the Medicaid eligibility thresholds, which vary by household size and state. The results are nearly identical and so they are not reported for brevity.

The total effects of observed depressive symptoms on birth weight are significantly negative and consistent across samples and measures. A 1-point increase in the 60-point CES-D scale translates into a 2-5-g decrease in birth weight, for instance, indicating that a woman with severely elevated symptoms (CES-D = 50) would have an infant that is 100-250 g lighter than one with no symptoms (CES-D = 0). Although not a tremendous difference in weight, this variation roughly corresponds with the observed difference in birth weights between blacks and whites in our sample (Appendix A). The discrete effect of being depressed (CES-D > 15) is to decrease birth weight by 29-73 g. It is also reassuring to find that the total effect of maternal depression obtained from the reduced-form model corresponds very well to the sum of the direct and indirect effects obtained from the structural mode. (22) Finally and perhaps more important is that both the discrete and continuous measures of depression statistically significantly increase the probability of having a low-birth weight birth for both samples.

6. Results with Predicted Depressive Symptoms

As described in section 3, we construct instruments for our observed measure of depressive symptoms taken from the 1988 NMIHS survey, with the goal of purging our measures of any endogeneity and adjusting for the fact that our observed measure is taken after the birth. The empirical results from these regressions are included in Appendix B, both with and without maternal smoking, drinking, and drug use and whether the mother wanted the pregnancy as regressors. For the most part, the results are sensible and are fairly stable across the samples, although our equation fits white mothers' depressive symptoms better than black mothers', perhaps in part because the mother's age and education are only important factors for white mothers. In general, being economically disadvantaged, experiencing poor maternal behaviors, not wanting the pregnancy, and being pregnant (again) at the time of the survey are all associated with higher symptoms. Having a sick baby also increased depressive symptoms, so there is some weak evidence of possible reverse causality. In terms of our aggregate variables, the availability of psychiatrists is associated with lower levels of depressive symptoms for both blacks and whites. Although many of the aggregated variables are not individually significant, they are jointly statistically significant at the 5% level in all specifications, thereby helping to add meaningful variation to our depressive symptoms instrument.

Using these coefficient estimates, we predict the mother's depressive symptoms during the pregnancy, assigning all observations a 1.0 for the "pregnant now'" dummy variable and the sample averages for the average number of sick visits and length of time since the birth. When available, we substitute the values of the variables during the pregnancy for the ones at the time of the survey. Table 3 repeats the exercises reported in Table 2 using the two predicted depression measures (plus their dummy variable counterparts). Because the estimated behavior coefficients are similar, we only report the Onset and depression coefficients. In addition, for simplicity and to be logically consistent, we do not report models that include maternal behaviors in the infant health model but not in the CES-D regressions.

Given the rather poor explanatory power of the depressive symptoms regressions, the main message from this exercise is surprisingly similar. Depressive symptoms decrease the birth weights of both whites and blacks, but for black mothers, they appear to operate only through delaying prenatal care. The total effects of the reduced-form models again correspond well to the sum of the direct and indirect effects of the structural-form models. The importance of maternal behaviors is even more pronounced here, which is as expected because they are important predictors of depressive symptoms. Either including maternal behaviors in the infant health model or excluding them as predictors of depressive symptoms reduces the effect of depressive symptoms, especially for black mothers. This could be a result of the poor fit of the depression regressions for black mothers, especially when the behavior and "wanted" variables are omitted.

What are we to conclude then? For whites, the evidence is strong: The direct effect of depressive symptoms on birth weight is consistently negative, whether we use an observed measure or an instrument, a continuous or discrete measure, or a structural- or reduced-form model. Of the 42 depression coefficients estimated in a birth-weight equation or low-birth weight probit, all indicate a negative effect on birth weight, and 29 are statistically significant. The magnitude varies quite a bit across the different measures. When the observed measures are used. the estimated effect is quite modest--approximately 2-3 g per point on the CES-D scale or a 29-62-g decrease for being depressed. In contrast, when the instruments are used, the effect is much more substantial and ranges from a 10- to a 44-g decrease per point on the CES-D scale, or from a 90- to a 176-g decrease for being depressed.

For blacks, the evidence is more mixed. Three out of four measures (either continuous or discrete) strongly suggest that depression delays prenatal care and leads to lower birth weights. For the observed measures, the magnitudes are again modest--about 4 g per point on the CES-D scale or a 58-73-g decrease for being depressed. When the instruments are used, the results become volatile, perhaps because our CES-D regressions have a poorer fit for this group.

7. Concluding Remarks

Studies of depression show that its economic costs, in terms of treatment costs and lost earnings resulting from mortality and morbidity, are quite large (Greenberg et al. 1993a, 1993b). More recently, there has been a growing recognition that treating depression may play an important role in helping the low-income and unemployed population, especially those on welfare. Our research suggests that addressing the problem of depression within the low-income population has benefits that extend well beyond employment into the production of infant health. Furthermore, it also offers some cost-effective ways to improve infant health through the Medicaid program--by augmenting the recent Medicaid expansions with greater attention to screening for and treatment of maternal depression. Our results also have a lesson for health care providers. A recent study by Williams et al. (1999, p. 64) find that among family physicians, general internists, and OB-GYNs, the OB-GYNs stand out as having higher physician barriers to caring for depression (e.g., having low confidence in or incomplete knowledge of treatment). Our research provides evidence of the potential costs to the infant of failing to detect and treat depression during pregnancy and suggests that greater effort should be directed at improving physician recognition and treatment. Our results provide compelling evidence that maternal depressive symptoms are associated with lower birth weight, and it highlights some of the ways these symptoms may affect birth outcomes. For black mothers, they appear to operate through delaying prenatal care, whereas for white mothers they have more of a direct effect. For both groups, there is evidence that it may also operate through increased maternal smoking and drinking.

There remain, however, some unanswered questions. The results for black women are empirically fragile when we instrument maternal depression. This is disappointing because this is the group that displays more depressive symptoms in our sample and that has a higher incidence of low birth weight and very low birth weight births. For both samples, the magnitudes range from quite modest to substantial depending on the measure used. It is possible that depressive symptoms are acting as a signal for the mother's health endowment and behaviors. We therefore may not have isolated the independent effect of depressive symptoms on birth outcomes: however, we can conclude that it is a meaningful signal of a potentially troubled pregnancy. This is useful especially given the relative ease of using the CES-D as a diagnostic tool.

Yet there are also reasons to believe that our estimates may be conservative. First, our measures of depression are not ideal and may be only weakly correlated with depression during the pregnancy. Second, we have eliminated the poorest birth outcomes by focusing only on live births (to remain comparable with other studies of birth weight). Finally and perhaps most important, we have ignored many possible avenues for depression to have an effect on infant health. For example, maternal smoking has consistently been found to be one of the most important determinants of infant health, and our depression regressions confirm the casual observation that smoking is associated with maternal depression. Indeed, as Saffer and Dave (2002) note, mental illness leads individuals to consume higher amounts of addictive goods and may also affect their price-responsiveness. It is therefore possible that treating depression may be a more effective way of reducing maternal smoking than tax increases or other policies. We explore this idea by reestimating our model both including and excluding maternal smoking (and drinking) as factors affecting birth weight and as predictors of depression. But there are other avenues as well. As noted by Lennon, Blome, and English (2001) and others, depression can make it more difficult to obtain and retain quality employment. As a result, depression may affect income, insurance status, and even family structure, all of which can possibly have both direct and indirect effects on birth weight. Taken together, our research suggests that treating maternal depression is a promising new approach to improving infant health, and one that warrants further investigation.

Appendix A
Variable Definitions, Sample-Weighted Means, and Standard Deviations

Name                                     Variable Definition

Variables in structural birth weight equation
  Birth weight                Infant's birth weight in grams
  Low birth weight            Dummy variable for birth weight less than
                                2500 g
  Onset                       Week of the first prenatal care visit
  CES-D in 1988               CES-D measure taken in 1988 survey
                                (ranges from 0 to 60)
  CES-D > 15                  Dummy variable if CES-D in 1988 greater
                                than 15
  Mother's age                Mother's age at delivery
  Education                   Years of education
  Male                        Dummy variable if infant is male
  Parity                      Dummy variable if any prior pregnancies
  Number of fetal deaths      Number of prior fetal deaths
  Number of kids cohabiting   Number of own children living with mother
                                during most of pregnancy
  Dad cohabits                Dummy variable for if baby's father lived
                                with mother during the pregnancy
  Never married               Dummy variable for never married
  Urban                       Dummy variable if live in urban county
  Mother's height             Mother's height in inches

Identifying variables in prenatal care equation
  ACCRA                       American Chamber of Commerce Researchers
                                Association (ACCRA) price index for
                                health care items, state level
  Population density          Population density for the state
  Income                      Total family income in 12 months before
                                delivery, in thousands
  Medicaid                    Dummy variable for if Medicaid paid for
                                prenatal care
  Private insurance           Dummy variable for if private insurance
                                paid for prenatal care
  OB Medicaid fee ratio       Ratio of maximum Medicaid fees to private
                                fee levels for obstetrician-
                                gynecologists
  OB-GYN per 1000             Number of obstetrician-gynecologists per
                                1000 state population
  General practice per 1000   Number of MDs in general practice per
                                1000 state population
  Hospitals per 1000          Total hospital per 1000 state population
  HMOs per 1000               Total health maintenance organizations
                                per 1000 state population
                              Largest sample size used

1988 and 1991 NMIHS Depression Average
  AvgCES-D                    Average of the 1988 and 1991 CES-D
                                Measures
  Depressed avg.              Dummy variable if average is 16 or
                                greater
                              Largest sample size used

Variables used only in the depression regressions
  Receives AFDC now           Dummy variable for receiving Aid to
                                Families with Dependent Children at
                                time of 1988 NMIHS survey ("now")
  Receives Medicaid now       Dummy variable if baby's health care is
                                paid for by Medicaid "now"
  Has private insurance now   Dummy variable if baby's health care is
                                paid for by private insurance "now"
  Smokes now                  How many cigarettes smoked per day "now"
  Drinks before pregnancy     Number of alcoholic drinks per week
                                before pregnancy
  Drugs before pregnancy      Dummy variable if reported smoking or
                                using cocaine/crack in the three months
                                before pregnant
  Exercise before pregnancy   Dummy variable if exercised/played sports
                                at least 3 times per week before
                                pregnant
  Dad cohabits now            Dummy variable if baby's father is with
                                mother "now"
  Number of kids cohabit now  Number of own children living with mother
                                "now"
  Age at first marriage       Mother's age at first marriage (only for
                                those who have married)
  Number of marriages         Number of marriages. =1 if one, =2 if two
                                or more
  Wanted pregnancy            Dummy variable if wanted pregnancy at an
                                earlier time and at that time, or
                                wanted pregnancy at that time
  Pregnancy now               Dummy variable if pregnant "now"
  Length of time since        Number of months between interview and
    delivery                    delivery
  Length x sick visits        Number of months X average number sick
                                baby visits per month
  Per capita MH expenditures  State mental health agency per capita
                                expenditures
  Psychiatrists per 1000      Number of psychiatrists per 1000 state
                                population
  Psychologists per 1000      Number of psychologists per 1000 state
                                population
  Social workers per 1000     Number of social workers per 1000 state
                                population
  MH services organizations   Number of multiservice mental health
    per 1000                    organizations per 1000 state population
                              Largest sample size used

Name                               Black             White

Variables in structural birth weight equation
  Birth weight                3162.09 (619.95)  3443.53 (552.61)
  Low birth weight             0.1125 (0.316)     0.045 (0.21)
  Onset                         9.886 (5.84)       8.52 (4.46)
  CES-D in 1988                 13.62 (10.84)      8.91 (9.38)

  CES-D > 15                    0.355 (0.479)     0.187 (0.39)
  Mother's age                 25.646 (5.05)      27.36 (5.02)
  Education                     12.49 (1.95)      13.28 (2.18)
  Male                          0.514 (0.50)      0.523 (0.50)
  Parity                        0.705 (0.456)     0.671 (0.47)
  Number of fetal deaths        0.408 (0.854)     0.336 (0.71)
  Number of kids cohabiting      1.21 (1.3)        0.90 (l.03)

  Dad cohabits                  0.515 (0.50)      0.908 (0.29)

  Never married                 0.499 (0.50)      0.072 (0.26)
  Urban                         0.795 (0.40)      0.743 (0.44)
  Mother's height               64.67 (2.85)      64.88 (2.61)

Identifying variables in prenatal care equation
  ACCRA                        99.180 (11.3)     101.76 (12.1)

  Population density           315.48 (1043.95)  217.41 (362.50)
  Income                       17.197 (16.896)   32.739 (20.94)

  Medicaid                      0.453 (0.50)      0.110 (0.32)
  Private insurance             0.339 (0.473)     0.709 (0.45)

  OB Medicaid fee ratio         0.587 (0.18)      0.589 (0.17)

  OB-GYN per 1000               0.134 (0.04)      0.125 (0.03)

  General practice per 1000     0.026 (0.01)      0.030 (0.03)

  Hospitals per 1000           0.0029 (0.002)    0.0033 (0.004)
  HMOs per 1000                0.0002 (0.0002)   0.0003 (0.0004)

                              3342              3232

1988 and 1991 NMIHS Depression Average
  AvgCES-D                     12.388 (8.4)        8.52 (7.2)
  Depressed avg.                0.305 (0.46)      0.158 (0.36)
                              2605              2766

Variables used only in the depression regressions
  Receives AFDC now             0.395 (0.49)      0.075 (0.26)

  Receives Medicaid now         0.424 (0.49)      0.091 (0.29)

  Has private insurance now     0.297 (0.46)      0.607 (0.49)

  Smokes now                     3.08 (6.54)       4.95 (9.22)
  Drinks before pregnancy       0.685 (2.52)       1.15 (2.72)
  Drugs before pregnancy        0.057 (0.23)      0.063 (0.24)

  Exercise before pregnancy     0.494 (0.50)      0.464 (0.50)

  Dad cohabits now              0.453 (0.50)      0.876 (0.33)
  Number of kids cohabit now     2.10 (1.49)       1.66 (1.21)
  Age at first marriage         21.88 (3.9)       21.64 (3.6)

  Number of marriages           0.545 (0.58)       1.06 (0.45)
  Wanted pregnancy              0.376 (0.48)      0.648 (0.48)

  Pregnancy now                 0.084 (0.28)      0.083 (0.28)
  Length of time since deliver  17.29 (5.05)      16.35 (4.89)
  Length x sick visits           2.46 (3.53)       2.86 (3.55)

  Per capita MH expenditures    49.80 (33.45)     47.66 (24.78)
  Psychiatrists per 1000       0.0006 (0.0008)   0.0006 (0.0006)
  Psychologists per 1000        0.015 (0.01)      0.016 (0.01)
  Social workers per 1000       0.093 (0.04)      0.096 (0.03)
  MH services organizations     0.005 (0.002)     0.006 (0.002)
    per 1000 state
    population
                              2654               2635

CES-D: Center for Epidemiological Study of Depression.

Appendix B
Center for Epidemiological Study of Depression Score Ordinary Least
Squares Regression Coefficient Estimates

Variable                                        Black

Intercept                        26.18 ** (2.54)       23.99 ** (2.35)
Receives Aid to Families
  with Dependent Children now       0.468 (0.78)          0.642 (1.16)
Receives Medicaid now               0.645 (1.03)          0.771 (1.19)
Has private insurance now        -0.97 ** (-1.98)      -1.22 ** (-2.28)
Income                             -0.022 (-1.42)        -0.027 (-1.64)
Smokes now                       0.084 ** (2.26)
Drinks before pregnancy          0.429 ** (4.17)
Drugs before pregnancy               0.99 (1.35)
Exercise before pregnancy           0.610 (1.42)
Dad cohabits now                 -1.55 ** (-2.38)      -2.04 ** (-3.47)
Number of kids cohabiting now      -0.178 (-1.59)         -0.07 (-0.67)
Years of education                  -1.03 (-0.90)        -0.936 (-0.81)
Education (2)                       0.020 (0.46)          0.015 (0.34)
Mother's age                        0.017 (0.03)          0.184 (0.36)
Mother's age (2)                   -0.003 (-0.31)        -0.005 (-0.59)
Age at first marriage               0.063 (1.21)          0.039 (0.730
Number of marriages                -0.503 (-0.69)        -0.543 (-0.69)
Never married                      -0.072 (-0.04)        -0.421 (-0.23)
Number of prior fetal deaths       -0.196 (-0.81)        -0.164 (-0.71)
Wanted pregnancy                  -2.20 ** (-5.10)
Pregnant now                      1.782 ** (2.68)      1.798 ** (2.66)
Length of time since delivery       -0.121 (-0.26)       -0.174 (-0.37)
Length of time (2)                   0.004 (0.38)         0.006 (0.50)
Length x sick visits              0.788 ** (3.18)      0.816 ** (3.30)
Length (2) x sick visits          -0.026 * (-1.89)     -0.027 * (-1.86)
Urban                               -0.552 (-0.86)       -0.476 (-0.76)
Per capita MH expenditures        0.052 ** (3.02)      0.044 ** (2.60)
Psychiatrists per 1000         -1671.24 ** (-2.35)  -1573.34 ** (-2.10)
Psychologists per 1M)                 8.09 (0.20)         15.53 (0.37)
Social workers per 1000             -26.23 (-1.41)       -23.53 (-1.20)
MH services organizations per
  1000                              -92.70 (-0.77)       -64.39 (-0.47)
HMOs per 1000                       241.07 (1.08)        258.39 (1.09)
[R.sup.2]                           0.1214                0.094
F-Statistic                         144.25                56.55
Sample Size                           2654                2654

Variable                                         White

Intercept                         47.94 ** (6.16)      50.62 ** (6.70)
Receives Aid to Families
  with Dependent Children now         1.73 (1.15)          1.70 (1.12)
Receives Medicaid now                 1.28 (0.94)          1.62 (1.14)
Has private insurance now           -0.432 (-1.21)       -0.484 (-1.29)
Income                           -0.031 ** (-3.17)    -0.037 ** (-3.37)
Smokes now                         0.039 * (1.83)
Drinks before pregnancy              0.131 (1.51)
Drugs before pregnancy             2.65 ** (2.67)
Exercise before pregnancy           -0.439 (-1.05)
Dad cohabits now                  -2.00 ** (-2.37)     -2.79 ** (-3.19)
Number of kids cohabiting now        -0.06 (-0.36)        0.025 (0.14)
Years of education                -2.51 ** (-3.41)     -2.54 ** (-3.70)
Education (2)                     0.082 ** (3.21)      0.082 ** (3.46)
Mother's age                      -1.00 ** (-2.27)    -1.210 ** (-2.75)
Mother's age (2)                  0.016 ** (2.15)      0.019 ** (2.58)
Age at first marriage               -0.062 (-1.35)       -0.054 (-1.05)
Number of marriages                  0.136 (0.20)         0.452 (0.63)
Never married                      -2.44 * (-1.69)        -1.49 (0.97)
Number of prior fetal deaths         0.134 (0.51)         0.276 (1.01)
Wanted pregnancy                 -1.994 ** (-5.10)
Pregnant now                       1.90 ** (2.30)       1.72 ** (2.09)
Length of time since delivery       -0.130 (-0.58)       -0.098 (-0.43)
Length of time (2)                  0.0049 (0.73)         0.004 (0.53)
Length x sick visits              0.425 ** (2.21)      0.427 ** (2.20)
Length (2) x sick visits           -0.0069 (-0.58)       -0.007 (0.58)
Urban                                0.190 (0.43)          0.40 (0.93)
Per capita MH expenditures           0.018 (1.57)         0.015 (1.26)
Psychiatrists per 1000           -643.37 * (-1.88)      -507.70 (-1.43)
Psychologists per 1M)                15.35 (0.65)         11.53 (0.44)
Social workers per 1000              -6.34 (-0.59)        -7.89 (-0.68)
MH services organizations per
  1000                           -143.44 * (-1.91)      -120.60 (-1.58)
HMOs per 1000                       -63.43 (0.36)        -63.64 (-0.35)
[R.sup.2]                           0.1502                0.130
F-Statistic                          77.03                80.92
Sample Size                           2635                 2635

All estimated standard enrors are robust and adjusted for clustering at
the state level.

* Denotes statistical significance at 90% level.

** Denotes statistical significance at 95% level.

Table 1. Key Characteristics of Depressed (Center for
Epidemiological Study of Depression (CES-D)
in 1988 > 15) and Nondepressed Mothers, by
Group and Sample-Weighted

                                                        Black

                                             Nondepressed   Depressed

CES-D in 1991 (a)                                9.09          15.4
Percentage depressed in 1991 (a)                 15.6          41.6
Birth weight (grams)                           3197.96       3096.95
Percentage low birth weight (<2500 g)            10.1          13.3
Percentage very low birth weight (<1500 g)        2             2.5
Week prenatal care (PNC) began                    9.4          10.7
Percentage with barriers to PNC                   9.7          18.9
Income                                        19130.85      13687.76
Percentage Medicaid paid for PNC                 39.9          55
Percentage private insurance paid for PNC        39.1          24.5
Percentage on AFDC during pregnancy              32           43.9
Percentage never married                         46           56.9
Cigarettes smoked per day during pregnancy        1.73         2.6
Alcoholic drinks per week during pregnancy        0.119        0.432
Sample size                                    2125 [1681]   1217 [924]

                                                        White

                                             Nondepressed   Depressed

CES-D in 1991 (a)                                7.02          13.99
Percentage depressed in 1991 (a)                 8.3           34.6
Birth weight (grams)                          3458.87        3376.8
Percentage low birth weight (<2500 g)            4              6.7
Percentage very low birth weight (<1500 g)       0.6            0.9
Week prenatal care (PNC) began                   8.4            9.2
Percentage with barriers to PNC                  7.8           24.8
Income                                       34647.74       24436.34
Percentage Medicaid paid for PNC                 8.8           20.4
Percentage private insurance paid for PNC       74.2           56.6
Percentage on AFDC during pregnancy              6             12.9
Percentage never married                         5.7           13.7
Cigarettes smoked per day during pregnancy       2.85           5.04
Alcoholic drinks per week during pregnancy       0.172          0.264
Sample size                                   2580 [2244]     652 [522]

(a) The sample size for this variable
is reported in brackets at the bottom of the table.

Table 2. Key Estimated Coefficients when Observed Center
for Epidemiological Study of Depression (CES-D Measures
Used (t-Statistics in Parentheses)

                                   Black

                                          Average
                       Observed            1988
                         1988            and 1991
Structural Models (a)
  Baseline, no depression
    Birth weight
      Onset            -81.69 **         -70.55 **
                       (-2.95)           (-2.53)
    Birth weight with behaviors
      Onset            -76.86 **         -63.79 **
                       (-2.85)           (-2.35)
      Smoke             -8.36 **          -7.56 **
                       (-3.75)           (-3.36)
      Drink             -5.76             -3.60
                        (0.81)           (-0.44)
  Continuous depression
      Onset

      CES-D Score        0.039 **          0.023 *
                        (4.34)             1.94
    Birth weight
      Onset            -82.49 **         -64.71 **
                        (2.70)            -0.34
      CES-D score       -0.552            -3.212 **
                        -0.34            (-2.13)
  Continuous depression with behaviors
    Onset
      CES-D score        0.038 **          0.021 *
                        (4.43)            (1.80)
      Smoke             -0.008            -0.021
                       (-0.46)           (-1.18)
      Drink              0.044             0.085
                        (0.88)            (1.54)
    Birth weight
      Onset            -79.47 **         -59.27 **
                       (-2.68)           (-2.22)
      CES-D score       -0.136            -2.76 **
                       (-0.09)           (-1.99)
      Smoke             -8.37 **          -7.32 **
                       (-3.66)           (-3.27)
      Drink             -5.52             -2.83
                       (-0.79)           (-0.36)
  Depression dummy
    Onset
      CES-D > 15         0.796 **          0.296
                        (4.04)            (1.04)
    Birth weight
      Onset            -82.574 **        -66.973 **
                       (-2.72)           (-2.40)
      CES-D > 15        -5.51            -44.079
                       (-0.17)           (-1.64)
  Depression dummy with behaviors
    Onset
      CES-D > 15         0.777 **          0.274
                        (4.02)            (0.99)
      Smoke             -0.007            -0.02
                       (-0.40)           (-1.15)
      Drink              0.051             0.09 *
                        (1.02)            (1.66)
    Birth weight
      Onset            -78.474 **        -60.789 **
                       (-2.66)           (-2.24)
      CES-D > 15        -1.46            -39.34 *
                       (-0.05)           (-1.65)
      Smoke             -8.37 **          -7.42 **
                       (-3.72)           (-3.31)
      Drink             -5.63             -3.34
                       (-0.79)           (-0.42)
Reduced-Form Models
  Continuous depression
    Birth weight
      CES-D score       -3.91 **          -4.84 **
                        (4.40)           (-4.04)
    Low birth weight probit derivative
      CES-D score        0.0013 **         0.0013 **
                        (3.56)            (2.87)

  Continuous depression with behaviors
    Birth weight
      CES-D score       -3.336 **         -4.187 **
                       (-3.92)           (-3.32)
      Smoke             -7.72 **          -5.99 **
                       (-4.21)           (-2.75)
      Drink             -8.24             -6.84
                       (-1.74)           (-1.18)
    Low birth weight
      CES-D score        0.001 **          0.001 **
                        (3.01)            (2.15)
      Smoke              0.0026 **         0.002 **
                        (4.39)            (2.82)
      Drink              0.005 **          0.0049 **
                        (3.14)            (2.18)
  Depression dummy
    Birth weight
      CES-D > 15       -72.848 **        -65.936 **
                       (-3.60)           (-3.54)
    Low birth weight probit derivative
      CES-D > 15         0.0229 **         0.014 *
                        (2.10)            (1.70)
  Depression dummy with behaviors
    Birth weight
      CES-D > 15       -64.46 **         -58.39 **
                       (-3.44)           (-3.08)
      Smoke             -7.82 **          -6.09 **
                       (-4.25)           (-2.79)
      Drink             -8.96 *           -7.86
                       (-1.90)           (-1.35)
    Low birth weight probit derivative
      CES-D > 15         0.018 *           0.001
                        (1.84)            (1.20)
      Smoke              0.0027 **         0.002 **
                        (4.37)            (2.84)
      Drink              0.005 **          0.005 **
                        (3.32)            (2.29)
Sample Size           3342              2605

                         White
                                          Average
                       Observed            1988
                         1988            and 1991
Structural Models (a)
  Baseline, no depression
    Birth weight
      Onset             -11.76              1.842
                        (-0.86)           (-0.14)
    Birth weight with behaviors
      Onset               3.79             13.68
                         (0.25)            (0.94)
      Smoke              -9.32 **          -9.64 **
                       (-11.10)          (-10.31)
      Drink              -6.29 **          -5.99 *
                        (-2.13)           (-1.78)
  Continuous depression
    Onset
      CES-D Score        -0.0033           -0.0076
                        (-0.41)           (-0.61)
    Birth weight
      Onset              -8.25              1.375
                        (-0.62)            (0.10)
      CES-D score        -2.890 **         -2.849 **
                        (-3.15)           (-1.96)
  Continuous depression with behaviors
    Onset
      CES-D score        -0.003            -0.007
                        (-0.44)           (-0.59)
      Smoke              -0.013            -0.011
                        (-1.20)           (-0.97)
      Drink               0.092 **          0.09 **
                         (2.97)            (2.68)
    Birth weight
      Onset              -5.77             15.18
                         (0.39)            (1.04)
      CES-D score        -1.85 *           -1.46
                        (-1.90)           (-0.95)
      Smoke              -9.127 **         -9.517 **
                       (-10.61)           (-9.97)
      Drink              -6.28 **          -6.105 *
                        (-2.12)           (-1.78)
  Depression dummy
    Onset
      CES-D > 15          0.171             0.061
                         (0.89)            (0.21)
    Birth weight
      Onset             -10.277             0.945
                        (-0.75)            (0.07)
      CES-D > 15        -45.189 *         -61.583 **
                        (-1.92)           (-2.26)
  Depression dummy with behaviors
    Onset
      CES-D > 15          0.173             0.068
                         (0.94)            (0.24)
      Smoke              -0.014            -0.012
                        (-1.25)           (-1.02)
      Drink               0.091 **          0.090 **
                         (2.93)            (2.67)
    Birth weight
      Onset               4.65              15.63
                         (0.31)            (1.06)
      CES-D > 15        -30.38            -47.11
                        (-1.19)           (-1.57)
      Smoke              -9.22 **          -9.51 **
                       (-10.79)           (-9.52)
      Drink              -6.29 **          -6.13 *
                        (-2.13)           (-1.79)
Reduced-Form Models
  Continuous depression
    Birth weight
      CES-D score        -2.892 **         -2.90 **
                        (-3.21)           (-2.01)
    Low birth weight probit derivative
      CES-D score        0.00082 **        0.00063 **
                        (5.53)            (2.38)

  Continuous depression with behaviors
    Birth weight
      CES-D score       -1.906 **          -1.62
                       (-1.99)            (-1.07)
      Smoke             -9.18 **           -9.65 **
                      (-10.31)           (-10.05)
      Drink             -5.806 **          -4.71
                       (-2.26)            (-1.50)
    Low birth weight
      CES-D score        0.0006 **          0.0003
                        (4.37)             (1.48)
      Smoke              0.0013 **          0.0014 **
                        (6.59)             (6.58)
      Drink              0.0004             0.0004
                        (0.49)             (0.48)
  Depression dummy
    Birth weight
      CES-D > 15       -46.721 **         -61.985 **
                       (-1.96)            (-2.23)
    Low birth weight probit derivative
      CES-D > 15         0.0178 **          0.0055
                        (3.56)             (1.05)
  Depression dummy with behaviors
    Birth weight
      CES-D > 15       -29.483            -46.50
                       (-1.16)            (-1.56)
      Smoke             -9.27 **           -9.66 **
                      (-10.45)           (-10.06)
      Drink             -5.93 **           -4.70
                        (2.33)            (-1.50)
    Low birth weight probit derivative
      CES-D > 15         0.014 **           0.002
                        (2.91)             (0.41)
      Smoke              0.0013 **          0.0014 **
                        (6.73)             (6.68)
      Drink              0.0005             0.0005
                        (0.55)             (0.50)
Sample Size           3232               2766

(a) Other variables included in the structural model birth
weight equation are mothers age and age squared, mother's
education and education squared, infant gender. parity,
number of previous fetal deaths, number of children
cohabiting, whether cohabits with father, whether ever
married. urban, and mothers height. See Appendix A for a
list of the identifying variables.

* Denotes statistical significance at 90% level.

** Denotes statistical significance at 95% level.
All estimated standard errors are robust and adjusted
for clustering at the state level.

Table 3. Key Estimated Coefficients when Predicted Center for
Epidemiological Study of Depression (CES-D) Used (t-Statistics is in
Parentheses)

                                               Black

                                        Predicted    Predicted
                                           with       without
                                        Behaviors    Behaviors

Structural Models
  Continuous depression
    Onset
      CES-D                                0.249 **     0.072
                                          (4.81)       (0.66)
    Birth weight
      Onset                              -91.143 *    -96.967 **
                                         (-1.81)      (-2.20)
      CES-D                                1.19         8.86
                                          (0.08)       (0.65)
  Continuous depression with behaviors
    Onset
      CES-D                                0.437 **
                                          (5.59)
    Birth weight
      Onset                              -82.82 *
                                         (-1.85)
      CES-D                               23.64
                                          (1.21)
  Depression dummy
    Onset
      CES-D > 15                           0.422       -0.010
                                          (1.35)      (-0.23)
    Birth weight
      Onset                              -82.293 **   -82.562 **
                                         (-2.44)      (-2.68)
      CES-D > 15                          -5.84         4.53
                                         (-0.13)       (0.12)
  Depression dummy with behaviors
    Onset
      CES-D                                0.385
                                          (1.33)
    Birth weight
      Onset                              -86.96 **
                                         (-2.58)
      CES-D > 15                          42.48
                                          (0.97)
Reduced-form models
  Continuous depression
    Birth weight
      CES-D                              -23.16 **      6.782
                                         (-3.88)       (0.57)
    Low birth weight probit derivative
      CES-D                                0.010 **    -0.002
                                          (5.0)       (-0.67)

  Continuous depression with behaviors
    Birth weight
      CES-D                               -3.41
                                         (-0.47)
  Low birth weight probit derivative
    CES-D                                  0.0002
                                          (0.11)
  Depression dummy
    Birth weight
      CES-D > 15                         -47.42 *       9.27
                                         (-1.67)       (0.31)
    Low birth weight probit derivative
      CES-D > 15                           0.013       -0.018
                                          (1.07)      (-1.36)
  Depression dummy with behaviors
    Birth weight
      CES-D > 15                           1.59
                                          (0.60)
    Low birth weight probit derivative
      CES-D > 15                          -0.011
                                         (-0.85)

                                               White

                                        Predicted    Predicted
                                           with       without
                                        Behaviors    Behaviors

Structural Models
  Continuous depression
    Onset
      CES-D                                0.243 **    -0.395 **
                                          (3.05)       (2.08)
    Birth weight
      Onset                               29.492       18.55
                                          (1.19)       (1.19)
      CES-D                              -36.15 **    -20.86 **
                                          (4.07)      (-2.03)
  Continuous depression with behaviors
    Onset
      CES-D                                0.295 **
                                          (3.49)
    Birth weight
      Onset                                9.38
                                          (0.41)
      CES-D                               -7.51
                                         (-0.84)
  Depression dummy
    Onset
      CES-D > 15                           1.09 **      0.594
                                          (2.34)       (1.11)
    Birth weight
      Onset                               13.483        2.66
                                          (0.80)       (0.19)
      CES-D > 15                        -182.17 **   -118.91 **
                                         (-4.64)      (-2.22)
  Depression dummy with behaviors
    Onset
      CES-D                                0.043 **
                                          (2.42)
    Birth weight
      Onset                               17.12
                                          (0.97)
      CES-D > 15                        -103.98 **
                                         (-2.68)
Reduced-form models
  Continuous depression
    Birth weight
      CES-D                              -36.38 **    -44.88 **
                                         (-7.44)      (-2.36)
    Low birth weight probit derivative
      CES-D                                0.0062 **    0.0109 **
                                          (4.61)       (3.31)

  Continuous depression with behaviors
    Birth weight
      CES-D                               -9.61
                                         (-1.46)
  Low birth weight probit derivative
    CES-D                                  0.0022
                                          (1.40)
  Depression dummy
    Birth weight
      CES-D > 15                        -172.64 **   -122.95 **
                                         (-5.53)      (-2.44)
    Low birth weight probit derivative
      CES-D > 15                           0.030 **     0.012
                                          (3.96)       (0.99)
  Depression dummy with behaviors
    Birth weight
      CES-D > 15                         -89.685 **
                                         (-2.90)
    Low birth weight probit derivative
      CES-D > 15                           0.013 *
                                          (1.67)

* Denotes statistical at 90% level.

** Denotes statistical significance at 95% level. All estimated
standard errors are robust and adjusted for clustering at the state
level.

(1) See, for example, Currie and Hyson (1999), who find that negative health (and other) effects of being born at a low birth weight persist well into adulthood, even after controlling for socioeconomic and other background characteristics. The popular press has also discovered this link; the cover story of the September 27, 1999, issue of Newsweek reports that low birth weight is associated with a greater risk of heart disease and obesity.

(2) Dropping private insurance to participate in Medicaid is referred to as "crowd out." For further discussion and estimates of crowd out. see Cutler and Gruber (1996) and Dubay and Kenney (1997).

(3) As we discuss shortly, there is a clear distinction between clinically diagnosed depression and measures of depressive symptoms or mood. Our empirical study is investigating the latter, which we refer to as "depressive symptoms." In our more general discussions, which may apply to either or both, such as in our theoretical framework, we use the term "depression" for simplicity.

(4) Saffer and Dave (2002) also make this assumption in their theoretical model, which looks at the effects of mental illness on the consumption of addictive goods. However, to the extent that women choose whether and how to be treated, especially when they are pregnant and the medications offered may pose risk to the fetus, one might be able to argue that it is a choice. Another consideration for us is the argument that some types of depression may be the direct result of the infant health production process--postpartum depression or depression that results from having a sick infant. As discussed further in section 3, this may be relevant for us because of the way in which depression is measured.

(5) To our knowledge, Chatterji and Markowitz (2004) is the only other economic study to estimate factors associated with maternal depression, and their focus is on the effects of maternity leave and returning to work on maternal health after the birth. These authors also use the NMIHS.

(6) For example, Grossman and Joyce (1990) use insurance status, the availability of Women, Infants, and Children (WIC) centers and prenatal clinics in the area, and the percentage in poverty to identify prenatal care.

(7) However, recent research by Kessler et al. (1997) indicates that the demarcation between "major" depression (with five or more symptoms) and "minor" depression (two to four symptoms) may not be as clear.

(8) To our knowledge, the only other publicly available data set with all of the necessary information to conduct this study is the Pregnancy Risk Assessment Monitoring System (PRAMS), which has a state-developed question asking the respondent whether she ever felt "depressed" during her pregnancy. Not only do these measures require the woman to define "depressed" (unlike the CES-D measure used in the NMIHS, which we discuss shortly), but it is only available for one state. Georgia. Although imperfect, the NMIHS appears to be the best data set to address this issue.

(9) For a discussion of alternative treatments and their risks during pregnancy see also Lamberg (1999) and Gold (2003). All report few risks associated with most treatments but caution that more research still needs to be done. For a discussion of the different kinds of depression and alternative treatments, as well as a study that reveals how frequently it goes undiagnosed (and inappropriately treated when diagnosed), see Wells et al. (1996).

(10) For instance, Nonacs and Cohen (2003, p. 548) state "the wish to avoid pharmacologic treatment during pregnancy is common." and "women commonly choose or are counseled to discontinue antidepressant treatment during pregnancy." Another example is Marcus et al. (2003), who use the CES-D to detect depressive symptoms in pregnant women and find that only 13.8% of women exceeding the cutoff score for depression reported receiving any formal treatment for depression. Of those who reported depression in the last 6 months, only 24.6% reported currently receiving treatment. Recognizing that "any treatment" includes many treatments other than pharmacologic, especially during pregnancy, indicates that the rate of such treatment is low. In addition, Kessler et al. (1997) find the use of medications to be low among the general depressed population, with only 10-20% reporting that "they ever took medications more than once for their depression" (Table 5, p. 24).

(11) For example, a pregnant woman may change her exercise routine or limit her drinking while pregnant and then return to her old ways after the birth. Smoking is available at the time of the survey, and so we use that variable to predict depressive symptoms. However. treating it like the other behaviors by instead using smoking before pregnancy did not substantively affect the results.

(12) In preliminary work, we instead used birth weight or the number of sick visits in the past month (each interacted with the length of time quadratic), which produced similar results but had slightly less explanatory power. Note that it does not matter what values we choose for these variables as long as they are uniformly chosen for all observations, as it is the variation in the depression measure across observations that identifies the depression coefficient in the birth weight and prenatal care equations.

(13) The NMIHS actually consists of three different sample--live births, infant deaths, and fetal deaths--and oversamples fetal and infant deaths. Gray (2001) estimates his model alternately including and excluding the fetal deaths sample and finds it makes no material difference.

(14) Although this could be another potential source of self-selection, the small number of observations made controlling for it unfeasible.

(15) For instance, women who attrite or who do not know how many sick visits their child has had may differ in systematic ways from those who do. The sizes of the black and white samples used to estimate Equation 6 are 2654 and 2635. respectively.

(16) Maternal birth weight has been found to be an important predictor of the infant's birth weight (Warner 1998). However, once again a large number of observations have missing data for this variable, especially for the black sample. Estimating the model both with and without maternal birth weight leads to similar results, and so we report only those estimates from the larger samples that exclude it.

(17) These percentages are somewhat higher than the usually reported incidence and are constructed using the sample weights. For instance, Wisner et al. (1999) report a lifetime risk of 10-25% for women, and Lennon, Blome, and English (2001) report a range of 12-36% for the prevalence of a major depressive disorder (during a 12-month period) among welfare populations. However, the CES-D is a measure of depressive symptoms and. as such, likely has a higher prevalence. In studies that use the CES-D. Lennon, Blome, and English (2001) report a range of estimates of 24.9-56.7% among welfare populations, which corresponds very well to the estimated prevalence in our black sample.

(18) This variable is based on questions in the NMIHS that asks the woman what barriers (e.g., transportation, economic, psychological), if any, she faced in obtaining prenatal care. In preliminary work, we used this variable as another identifier, and the results were quite similar. Concerned that it might be endogenous, we dropped it from the final model.

(19) Joyce el al. (1983), in a case study of women who received no prenatal care, suggest that "internal" barriers such as depression, fear, and denial are more to blame than "external" ones such as financial, transportation, and child-care problems. This is also evident in Table 1, in which depressed mothers sought prenatal care later and reported more barriers to care than did nondepressed mothers.

(20) For example, Grossman and Joyce (1990) find a decrease in birth weight of 12-37 g per month delay in prenatal care for blacks and 4-23 grams for whites. Liu (1998) reports a range of 6-197 grams per month, and Warner reports a range of -18 to +50 per week. Recent work by Conway and Deb (2003) offers an alternative explanation for the weak effects found for prenatal care: that researchers have failed to adequately account for the different kinds of pregnancies (i.e., "complicated" vs. "normal").

(21) Recall that the 1988 NMIHS sample weights are used to adjust for the oversampling of poor birth outcomes. We also considered estimating the incidence of very low birth weight (< 1500 g), but the weighted incidence is so low (the highest is 2.1% for the black sample) that we concluded such estimation is implausible.

(22) For example, using the average CES-D measure for blacks (column 2) in the model using continuous depression (without behaviors)--a model with statistically significant indirect and direct effects--yields an indirect effect of 1.49 g and a direct effect of 3.212 g per CES-D point, for a total of 4.702 g. Including maternal behaviors in the same specification yields a total effect of 4.0 g (1.24 + 2.76). The results from the other models, measures, and samples yield similar results that are very much in line with the reduced form total effects. This verification lends additional support to the structural instrumental variable model results.

References

Abrams, Sonya M.. Tiffany Field, Frank Scafidi, and Margarita Prodromidis. 1995. Newborns of depressed mothers. In]ant Mental Health Journal 16:23_%-9.

Angrist, Joshua D., and Alan B. Krueger. 1992. The effect of age at school entry on educational attainment: An application of instrumental variables with moments from two samples. Journal of the American Statistical Association 87:328-36.

Angrist, Joshua D., and Alan B. Krueger. 1995. Split-sample instrumental variables estimates of the return to schooling. Journal of Business and Economic Statistics 13:225-35.

Beck, Cheryl Tatano. 200l. Predictors of postpartum depression--an update. Nursing Research 50:275-83.

Chatterji, Pinka, and Sara Markowitz. 2004. Does the length of maternity leave affect maternal health? NBER Working Paper No. 10206.

Center for Mental Health Services. 1994. Mental Health, United States, 1994, edited by R. W. Manderscheid and M. A. Sonnenschein. Washington, DC: Supt. Of Docs., Government Printing Office.

Conway, Karen Smith, and Partha Deb. 2003. Is prenatal care really ineffective? Or is the "'devil" in the distribution? Unpublished paper. University of New Hampshire.

Currie, Janet, and Jeffrey Grogger. 2000. Medicaid expansions and welfare contractions: Offsetting effects on prenatal care and infant health? Journal of Health Economics 21:313-35.

Currie, Janet, and Jonathan Gruber. 1996. Saving babies: The efficacy and cost of recent changes in the Medicaid eligibility of pregnant women. Journal of Political Economy 111:431-66.

Currie, Janet, and Rosemary Hyson. 1999. Is the impact of health shocks cushioned by socioeconomic status'? The case of low birthweight. American Economic Review 89:245-50.

Cutler, David M., and Jonathan Gruber. 1996. Does public insurance crowd out private insurance'? The Quarterly Journal of Economics 111:391-430.

Dawson. Geraldine, Karin Frey, Hercules Panagiotides, Julie Osterling, and David Hessl. 1997a. Infants of depressed mothers exhibit atypical frontal brain activity: A replication and extension of previous findings. Journal of Child Psychology and Psychiatry 38:179-86.

Dawson, Geraldine, Hercules Panagiotides, Laura Grofer Klinger, and Susan Spieker. 1997b. Infants of depressed and nondepressed mothers exhibit differences in frontal brain electrical activity during the expression of negative emotions. Developmental Psychology 33:650-6.

Desai, Hirai, and Michael W. Jann. 2000. Major depression in women: A review of the literature. Journal of the American Pharmaceutical Association 40:525-37.

Dubay, Lisa. and Genevieve Kenney. 1997. Did Medicaid expansions for pregnant women crowd out private coverage? Health Affairs 16:185-93.

Field, Tiffany. 1995. Infants of depressed mothers. Infant Behavior and Development 18:1-13.

Field, Tiffany. 1998. Maternal depression effects on infants and early interventions. Preventative Medicine 27:200-3.

Gold, Liza H. 2003. Psychopharmacologic treatment of depression during pregnancy, Current Women's Health Reports 3: 236-41.

Gotlib, Ian H., Valerie E. Whiffen, Pamela M. Wallace, and John H. Mount. 1991. Prospective investigation of postpartum depression: Factors involved in onset and recovery. Journal of Abnormal Psychology 100:122-32.

Gray, Bradley. 2001. Do Medicaid physician fees for prenatal services affect birth outcomes? Journal of Health Economics 20:571-90.

Greenberg, Paul E., Laura E. Stiglin, Start N. Finkelstein, and Ernst R. Berndt. 1993a. The economic burden of depression in 1990. The Journal of Clinical Psychiatry 54:405-18.

Greenberg, Paul E., Laura E. Stiglin, Stan N. Finkelstein, and Ernst R. Berndt. 1993b. Depression: A neglected major illness. The Journal of Clinical Psychiatry 54:419-24.

Grossman, Michael, and Theodore J. Joyce. 1990. Unobservables, pregnancy resolutions, and birth weight production functions in New York City. Journal of Political Economy 98:983-1007.

Gruber, Jonathan. 1997. Policy watch: Medicaid and uninsured women and children. Journal of Economic Perspectives 11: 199-208.

Grush, Lynn R., and Lee S. Cohen. 1998. Treatment of depression during pregnancy: Balancing the risks. Harvard Review of Psychiatry 6:105-9.

Jones, Nancy Aaron, Tiffany Field, Nathan A. Fox, Brenda Lundy, and Marisabel Davalos. 1997. EEG activation in 1-month-old infants of depressed mothers. Development and Psychopathology 9:491-505.

Josefsson, Anne, Goran Berg, Conny Nordin, and Gunilla Sydsjo. 2001. Prevalence of depressive symptoms in late pregnancy and postpartum. Acta Obstetrica et Gynecologica Scandinavica 80:251-5.

Joyce, Kathleen, Geraldine Diffenbacher, Janet Greene, and Yoram Sorokin. 1983. Internal and external barriers to obtaining prenatal care. Social Work in Health Care 9:89-96.

Joyce, Theodore. 1999. Impact of augmented prenatal care on birth outcomes of Medicaid recipients in New York City. Journal of Health Economics 18:31-67. Kaestner, Robert. 1999. Health insurance, the quantity and quality of prenatal care, and infant health. Inquiry 36:162-75.

Kessler, Ronald C., Shanyang Zhao, Dan G. Blazer, and Marvin Swartz. 1997. Prevalence, correlates, and course of minor depression and major depression in the national comorbidity survey. Journal of Affective Disorders 45:19-30.

Lamberg, Lynne. 1999. Safety of antidepressant use in pregnant and nursing women. Journal of American Medical Association 281:222-3.

Lanzi, Robin Gaines, John M. Pascoe, Bette Keltner, and Sharon Landesman Ramey. 1999. Correlates of maternal depressive symptoms in a national head start program sample. Archives of Pediatric and Adolescent Medicine 153:801-7.

Lennon, Mary Clare, Juliana Blome, and Kevin English, 2001. Depression and low income women: Challenges for TANF and Welfare-to-Work policies and programs, Unpublished paper, Research Forum on Children, Families and the New Federalism, National Center for Children in Poverty.

Liu, Gordon G. 1998. Birth outcomes and the effectiveness of prenatal care. Health Services Research 32:805-23.

Locke, R., S. Baumgart, Kenneth M. Locke, M. Goodstein. C. Thies, and J. Greenspan. 1997. Effect of maternal depression on premature infant health during initial hospitalization. Journal of the American Osteopathic Association 97:145-9.

Loprest, Pamela and Michael Gates. 1993. State-level data book on health care access and financing. Washington, DC: The Urban Institute Press.

Lutterman, Theodore, Abdi Hirad, and Bethanne Poindexter. 1999. Funding Sources and Expenditures of State Mental Health Agencies, Fiscal Year 1997, Final Report. Alexandria, VA: National Association of State Mental Health Program Directors (NASMHPD) Research Institute.

Marcus, Sheila M., Heather A. Flynn, Frederic C. Blow, and Kristen L. Barry. 2003. Depressive symptoms among pregnant women screened in obstetrics settings. Journal of Women's Health 12:373-80.

Najman, J. M., M. J. Andersen, W. Bor, M. J. O'Callaghan, and G. M. Williams. 2000. Postnatal depression--myth and reality: Maternal depression before and after the birth of a child. Social Psychiatry and Psychiatric Epidemiology 35: 19-27.

Nonacs, Ruta, and Lee S. Cohen. 2003. Assessment and treatment of depression during pregnancy: An update. Psychiatric Clinics of North America 26:547-62.

Orr, Suezanne T., Sherman A. James, and Cheryl Blackmore Prince. 2002, Maternal prenatal depressive symptoms and spontaneous preterm births among African-American women in Baltimore, Maryland. American Journal of Epidemiology 156:797-802.

Orr, Suezanne T., and C. Arden Miller. 1995. Maternal depressive symptoms and the risk of poor pregnancy outcome. Epidemiologic Reviews 17:165-71.

Radloff, L. S.. and B. Z. Locke. 1986. The community mental health assessment survey and the CES-D scale. In Community Surveys of Psychiatric Disorders, edited by Myrna M. Weissman, Jerome K. Myers, and Catherine E. Ross. New Brunswick, NJ: Rutgers University Press, pp. 177-89.

Roberts, R. E., and S. W. Vernon. 1983. The Center for Epidemiologic Studies Depression Scale: Its use in a community sample. American Journal of Psychiatry 140:41-46.

Rosenzweig, Mark R., and T. Paul Schultz. 1982. The behavior of mothers as inputs to child health: The determinants of birth weight, gestation, and rate of fetal growth. In Economic Aspects of Health, edited by Victor R. Fuchs. Chicago: University of Chicago Press, pp. 53-92.

Rosenzweig, Mark R., and T. Paul Schultz. 1983. Estimating a household production function: Heterogeneity, the demand for health inputs, and their effects on birth weight. Journal of Political Economy 91:723-46.

Saffer, Henry, and Dhaval Dave. 2002. Mental illness and the demand for alcohol, cocaine and cigarettes. NBER Working Paper No. 8699.

Seguin, Louise, Louise Potvin, Michele St-Denis, and Jacinthe Loiselle. 1999. Depressive symptoms in the late postpartum among low socioeconomic status women. Birth 26:157-63.

Singer, Lynn T., Ann Salvator, Shenyang Guo, Marc Collin, Lawrence Lilien, and Jill Baley. 1999. Maternal psychological distress and parenting stress after the birth of a very low-birth-weight infant. Journal of the American Medical Association 281:799-805.

U.S. Dept. of Health and Human Services, National Center for Health Statistics. 1991. National Maternal and Infant Health Survey, 1988. Hyattsville, MD: Department of Health and Human Services, National Center for Health Statistics.

U.S. Dept. of Health and Human Services, National Center for Health Statistics. 1993. National Maternal and Infant Health Survey, 1988: Longitudinal Follow-up, 1991. Hyattsville, MD: U.S. Dept. of Health and Human Services, National Center for Health Statistics.

U.S. Dept. of Health and Human Services. Health Resources and Services Administration, Bureau of Health Professions. 1991. Bureau of Health Professions Area Resource File, 1940-1990: [United States]. 2nd release. Rockville. MD: U.S. Dept. of Health and Human Services, Office of Data Analysis and Management.

Warner, Geoffrey L. 1995. Prenatal care demand and birthweight production of black mothers. American Economic' Review 85:132-7.

Warner, Geoffrey L. 1998. Birthweight productivity of prenatal care. Southern Economic Journal 65:42-63.

Wells, Kenneth B., Roland Sturm, Cathy D. Sherbourne, and Lisa S. Meredith. 1996. Caring for Depression. Cambridge, MA: Harvard University Press.

Williams, John W. Jr., Kathryn Rost, Alien J. Dietrich, Mary C. Ciotti, Stephen J. Zyzanski, and John Cornell. 1999, Primary care physicians' approach to depressive disorders--effects of physician specialty and practice structure. Archives of Family Medicine 8:58-67.

Wisner, Katherine L.. Alan J, Gelenberg., Henrietta Leonard, Deborah Zarin, and Ellen Frank, 1999. Pharmacologic treatment of depression during pregnancy. Journal of the American Medical Association 282:1264-9.

Zimmerman, Mark, and William Coryell. 1994. Screening for major depressive disorder in the community: A comparison of measures. Psychological Assessment 6:71-4.

Karen Smith Conway * and Lisa DeFelice Kennedy ([dagger])

* University of New Hampshire, 15 College Road, Durham, NH 03824, USA: E-mail [e-mail]; corresponding author.

([dagger]) Bachrodt Academy, San Jose Unified School District, 855 Lenzen Ave. San Jose, CA 95126, USA.

We are grateful to W. David Bradford for his suggestions throughout this research and to Thomas Croghan, Janet Currie, David Cutler, Partha Deb, Bradley Gray, Thomas J. Kniesner, Steven Lehrer, Anthony LoSasso, Regina Powers, Geoff Warner, Bob Woodward, and participants in the UNH Economics Seminar and a session at the Southern Economic Association Meetings for their helpful comments on previous drains of this research. This article has also benefited from the comments of three anonymous referees and a coeditor of this journal. We thank Minghua Li for her fine research assistance. This article has grown out of the research undertaken in Dr. DeFelice Kennedy's dissertation.

Received June 2003; accepted Marcia 2004.