Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
Effect of a Child's Death on Birth Spacing: A Cross-National Analysis Laurence M. Grummer-Strawn, Paul W. Stupp, and Zuguo Mei INTRODUCTION Our purpose in this chapter is to examine the potential pathways through which infant and child mortality can affect the interval between births. Specifi- cally, we examine the extent to which the death of a child tends to shorten the time until the birth of the next child. To the extent that the shortening of the birth interval is attributable to direct physiological mechanisms (i.e., premature return of menses due to breastfeeding cessation), the linkage between mortality decline and fertility decline can be considered directly causal. On the other hand, if the linkage between child death and birth interval is related more to maternal behav- ior, such as cessation of contraceptive use, then the linkage depends on the widespread use of contraceptives. By teasing out the mechanisms for the linkage, we will better understand the future course of fertility and mortality in developing countries. The analysis uses data from phases 1 and 2 of the Demographic and Health Surveys (DHS), which collect an extensive set of fertility-related variables for women of reproductive age and for live births that occurred in the 5 years before the date of interview. We first present evidence of the magnitude of the effects on the birth interval for a variety of countries and then address the mechanisms through which these effects may operate. PATHWAYS OF INFLUENCE The effects of infant and child mortality on the intervals between births 39
40 BIRTH SPACING: A CROSS-NATIONAL ANALYSIS probably act through both physiological mechanisms and volitional behaviors as noted by Preston (1978~. Most prominent among the physiological mechanisms is lactational amenorrhea, by which breastfeeding inhibits the return of ovulation after a birth. If a child dies, the mother will stop breastfeeding (except perhaps in the case of a multiple birth), and she will become susceptible to pregnancy more quickly than if the child had survived. This effect is expected to be most pro- nounced in societies such as those in sub-Saharan Africa in which breastfeeding is nearly universal and is performed for an extended time. Studies in the early 1980s showed that the anovulatory effect of breastfeeding is related to its fre- quency, intensity, and duration (Konner and Worthman, 1980; McNeilly et al., 1980; Howie et al., 1982~. For this analysis, duration of breastfeeding was the only variable for which data were collected for all births during a 5-year period. To explore the role of lactational amenorrhea in spacing of births, we used data from the DHSs to examine the effect of death of a child on the length of time until return of menses and until the next live birth. We treat time to return of menses, which is among the data collected in the DHS, as a proxy for time to return of ovulation, which is not collected in the DHS. Beyond its effect in delaying the return of ovulation, breastfeeding may also reduce the probability of conception among ovulating, sexually active women. We investigated the effect of the premature truncation of breastfeeding on the interval from return of menses and resumption of sexual relations to a subsequent pregnancy. The DHS do not collect information on pregnancies that do not result in a live birth (i.e., spontaneous abortions and stillbirths). This restriction prevents us from considering another potential physiological mechanism by which the death of a child could lengthen the interval to the subsequent birth. If infant mortality is correlated with fetal mortality in the same woman it could lengthen the subse- quent interval due to fetal losses. This effect related to maternal health or genetic endowment appears realistic, especially in the case of neonatal mortality, and could be substantial for a subset of women. Unfortunately, we cannot use DHS data to measure this effect. The primary behavioral pathways through which death of a child can affect the interval before a subsequent pregnancy involve sexual activity and contracep- tive use. If a society observes a norm of traditional sexual abstinence for a period after a birth, then the death of a child may reduce that period. The DHS contain data on duration of postpartum abstinence and thus make it possible to investigate directly whether the death of a child affects this component of the birth interval. Once a woman has resumed sexual relations and is ovulating after a birth, the behavioral factors affecting the time to conception are frequency of sexual inter- course and contraceptive use. We do not have good measures of either of these variables during the pertinent periods for all of the surveys used in this analysis. We presume, however, that the shortening of this portion of the interval, once we control for breastfeeding status, must be related to one or both of these factors.
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI 41 We predict that the effect of the death of a child during this period of susceptibil- ity to pregnancy will be more pronounced in countries with more frequent contra- ceptive use. Unfortunately, the DHS do not generally provide appropriate data on changes in contraceptive use throughout the intervals of interest, so it was not possible to examine the effect of contraceptive use. Use of a contraceptive calendar would be appropriate for this purpose, but only in high prevalence countries DHS included a contraceptive use calendar. Frequency of sexual relations is another factor that may be related to the effect of death of an infant on the length of the susceptible period. A couple could respond to a child's death by increasing the frequency of sexual activity in an effort to replace a dead child. Unfortunately, this hypothesis could not be tested because DHS do not provide data on coital frequency during each period of the interval. Induced abortion is another factor we could not address. This is probably an important factor only in very low fertility settings where child mortality is more rare. Such countries have generally been excluded from the DHS program. INTERVENING VARIABLES It is well documented that the length of the birth interval affects child mortal- ity and that shorter previous intervals are associated with higher infant and child mortality rates (Sullivan et al., 1994~. Some of the association between a child's death and a short subsequent birth interval could therefore be confounded by the association between short previous intervals and a subsequent child death. If some women are more prone to shorter intervals than average, because of differ- entials in fecundity, sexual activity, or contraceptive use, then child mortality may be a spurious factor. It is therefore important that we control for the length of the previous interval in studying the association between child mortality and the length of subsequent intervals. Similarly, child mortality may be higher if there is a shorter subsequent interval. For a short subsequent interval to affect a child's death, the death should take place after the conception of the child who is born subsequently. Similarly, for a death to affect the length of the subsequent interval it should logically occur before the conception leading to the birth that closes the interval. To avoid the possibility of reverse causality, we took great care in the modeling to limit the effect of the death to the period before the conception leading to the subsequent live birth. Other factors that may simultaneously affect both child survival and birth interval are maternal age and education, birth order, sex of the child, whether the child was wanted, socioeconomic status of the household, and whether the woman resides in an urban or rural community. Among these factors, sex of the child, whether the child was wanted, and maternal age tend to mask the association between death and short birth intervals. If the child is female, she is less likely to
42 BIRTH SPACING: A CROSS-NATIONAL ANALYSIS die and the birth interval is likely to be shorter because of the desire to have another birth following the birth of a girl (assuming preference for a son). Simi- larly, infants who are unwanted experience higher mortality, and subsequent birth intervals are likely to be longer. Children born to older mothers are at greater risk of death, but their birth is often followed by longer birth intervals because of the mother's reduced fecundity. Other factors may tend to generate the expected association between death of a child and length of the birth interval. If maternal education or socioeconomic status is higher, if birth order is lower, or place of residence is urban, infant mortality is likely to be lower and birth interval longer. Data on these variables are available in the DHS and are thus included in the multivariate analyses. This list of intervening factors is by no means comprehensive. Other charac- teristics of the mother or child might be correlated with both interval length and the survival of the child. For example, mothers with a strong desire to have healthy children may opt for longer birth intervals (knowing that they are safer for the child) but also invest more resources in nutrition and medical care for the child. The continued presence of a male partner in the household certainly affects the probability of closing the birth interval, but also influences socioeconomic well-being of the household and allocation of resources within the household. Other characteristics that affect birth intervals but that are not necessarily corre- lated with child mortality, such as the woman's underlying fecundability, can bias estimation of the parameter estimates as well (Heckman and Singer, 1982, 1984~. This problem of unobserved heterogeneity has been addressed by a vari- ety of complex statistical procedures (Trussell and Richards, 1985; Trussell and Rodriguez, 1990; Guo and Rodriguez, 1992~. Unfortunately, we were unable to find a suitable method that accounts for left- and right-censoring, time-varying covariates and time-varying effects in a manner that could be applied easily to 45 surveys. Our hope is that the variables controlled in the analysis are reasonable proxies for the unobserved characteristics, such that the degree of bias is mini- mized. DATA AND METHODS For this analysis, we used data from the DHS phases 1 and 2 (Lapham and Westoff, 1986~. All data sets available at the time this study was started were included. Recoded data sets for 46 surveys were provided by Macro Interna- tional, Inc. (Calverton, Maryland). Conducted in developing countries in the late 1980s and early 1990s, DHS were based on household interviews of women of reproductive age (age 15-49~. Standard core questionnaires were used, encom- passing topics on family planning, fertility, child mortality, and maternal and child health. Additional questions and slight modifications were implemented in each country. Results are comparable across countries because the DHS use
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI 43 standard questionnaire modules and standardized sampling procedures, field- work, and data processing. The core questionnaire included data on the timing of a woman's live births, the survival status for these births, and the interval to death of a child, where applicable. For births occurring in the 5 years before the survey, data were obtained on the duration of breastfeeding, amenorrhea, and sexual abstinence.] Our analysis is restricted to births in the previous 5 years because a major aim is to examine the mechanisms through which the death of a child affects the subse- quent birth interval, specifically the premature truncation of breastfeeding lead- ing to earlier return of menses and resumption of sexual activity. Countries included in the analysis are listed in Table 2-1 and the number of births for each country is shown. Outcome Measures Interval Between Births We examined the relationship between the birth interval and the death of the child that started the interval. Several problems are apparent. First, many of the birth intervals (a majority in many countries) are open intervals, that is, last births. For these births, the time to the next birth is censored; thus, we had to use life table techniques in the data analyses. Second, because a short subsequent birth interval can contribute to the premature death of the index child (Hobcraft et al., 1985), we must account for the possibility of reverse causality by ensuring that the death occurs before the conception of the child whose birth closes the birth interval. Thus, death must be accounted for as a time-varying variable. To handle these problems, we have chosen to model birth intervals by using proportional hazards models; the baseline hazards are modeled with a piecewise exponential (Trussell and Guinnane, 1993~. This model is specified as \(tlX)= X0 (t) exp (X 3), 1Data on the duration of breastfeeding, postpartum amenorrhea, and postpartum abstinence are based on the mother's retrospective reports. It is well known that the quality of retrospectively reported data on duration is rather poor. Durations are frequently reported as multiples of 3 or 6 months and thus do not reflect accurate recollection of the events. To test whether this rounding or "heaping" problem affected our results, we tested models in which we excluded women who re- ported amenorrhea or breastfeeding durations of 3, 6, 9, 12, 15, 18, 24, 30, or 36 months. In all cases, the differences were so small that the results described here are not substantively changed. Data quality could be related to the main variable of interest here (child death). Interviewers may be reluctant to ask about dead children because of cultural norms. Mothers may inaccurately report breastfeeding the dead child out of a feeling of guilt.
44 BIRTH SPACING: A CROSS-NATIONAL ANALYSIS TABLE 2-1 Sample Sizes and Years of Surveys Included in Analysis Region Number of Births and Country Survey Year in Previous 5 Years Africa Botswana DHS1 1988 3,086 Burkina Faso DHS2 1992 5,828 Burundi DHS1 1987 3,811 Cameroon DHS2 1991 3,350 Egypt DHS1 1988 8,647 Ghana DHS1 1988 4,136 Kenya DHS1 1988-1989 6,980 Liberia DHS 1 1986 5,299 Madagascar DHS2 1992 5,273 Malawi DHS2 1992 4,495 Mali DHS1 1987 3,358 Morocco DHS1 1987 6,102 Morocco DHS2 1992 5,197 Namibia DHS2 1992 3,916 Niger DHS2 1992 6,899 Nigeria DHS2 1990 7,902 Ondo State, Nigeria DHS1 1986-1987 3,280 Rwanda DHS2 1992 5,510 Senegal DHS 1 1986 4,287 Senegal DHS2 1992- 1993 5,645 Sudan DHS1 1989-1990 6,644 Togo DHS1 1988 3,134 Tunisia DHS1 1988 4,477 Uganda DHS1 1988-1989 4,959 Zambia DHS2 1992 6,299 Zimbabwe DHS1 1988-1989 3,358 Asia Indonesia DHS1 1987 8,140 Indonesia DHS2 1991 15,708 Pakistan DHS2 1990- 1991 6,428 Sri Lanka DHS1 1987 4,010 Thailand DHS 1 1987 3,627 Latin America Bolivia DHS1 1989 5,814 Brazil DHS1 1986 3,573 Brazil DHS2 1991 3,159 Colombia DHS 1 1986 2,715 Colombia DHS2 1990 3,751 Dominican Republic DHS1 1986 4,767 Dominican Republic DHS2 1991 4,164 Ecuador DHS 1 1987 3,051 E1 Salvador DHS1 1985 3,339 Guatemala DHS2 1987 4,627 Mexico DHS1 1987 5,327 Paraguay DHS2 1990 4,246 Peru DHS1 1986 3,131 Peru DHS2 1991-1992 9,362 Trinidad and Tobago DHS1 1987 1,946
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI 45 where ~ is the hazard of the birth interval being closed by a subsequent live birth at time t, X0 is the baseline hazard, and X is a set of covariates. The baseline hazard is a step function, with steps defined for age groups 9-17, 18-23, 24-29, 30-35, 36-41, and 42-59 months. The hazard of closing the birth interval is assumed to be zero in the first 9 months of a child's life, allowing for a 9-month gestation for a subsequent birth. Therefore, all exposure and events in these first 9 months are ignored. We excluded intervals in which the date of either the index birth or the subsequent birth was imputed and intervals that began with a multiple birth. We began with model 1 in which the only covariate is whether the index child was alive 10 months ago. In this way, the risk of a conception occurring after the index child had died was compared with the risk of conception while the child was living. Whether the index child is dead was treated as a time-varying covariate, with a lag of 10 months (assuming a gestation of 9 months). Thus, for example, death of the index child at the age of 6 months can only affect the birth of the next child at durations of 16 months and greater. In model 2 of the birth interval we introduced controls for place of residence (urban or rural), an index of the socioeconomic status of the household, the mother's highest level of completed education (none, primary, or secondary), the mother's age at the time of the index birth, the sex of the child, parity, length of the previous birth interval, and whether the child was wanted at that time, at a later time, or not at all. The index of socioeconomic status was a continuous covariate based on a count of the number of goods and services available in the household from among the following: automobile, motorcycle, bicycle, refrig- erator, television, radio, and electricity. The mother's age was modeled with a linear and quadratic term to account for nonlinearity in the declining fecundity with age. We considered parity and previous birth interval jointly by using a categorical variable with categories of first-births, parity of two to five with an interval less than 24 months, parity of two to five with an interval of at least 24 months, parity of six or more with an interval less than 24 months, and parity of six or more with an interval of at least 24 months. As noted, some of these factors tend to mask the association between death and short birth intervals, but others tend to generate such an association. Therefore, we had no prior expectation as to whether the effect of the index child's death would be stronger or weaker after adjustment for these factors. In model 3, we added the effect of breastfeeding on closure of the birth interval. Breastfeeding was modeled in the same way as was the index child's death, that is, as a time-varying covariate with a 10-month lag. Although a woman can continue breastfeeding during pregnancy, we focused only on breast- feeding before the time of conception to avoid any possibility of reverse causality that occurs if a woman stops breastfeeding because she is pregnant. The addition of a breastfeeding dummy variable to the model along with child death had the effect of creating three possible statuses of the index child 10 months earlier:
46 BIRTH SPACING: A CROSS-NATIONAL ANALYSIS dead, weaned but alive, and still breastfed. The parameter on the variable for death of the index child thus measured the effect of that event relative to the effect of survival of the index child who was weaned. Because we expected that a major mechanism for the association between death of the index child and a shorter subsequent birth interval was the premature truncation of breastfeeding, we anticipated a large reduction in the effect of a child's death. In fact, if breastfeeding were the only mechanism for the prolongation of the birth interval, as might be the case in countries where contraceptives are seldom used, then we would expect that the effect of child death might be reduced to zero. Interval from Birth to Menses To better understand the mechanisms through which the death of a child might affect the subsequent birth interval, it is useful to consider the various components of the birth interval. Two events must take place before the concep- tion of the subsequent birth. The couple must resume sexual relations and the woman must begin ovulating again. Either event could occur first. Thus, the birth interval can be thought of in three distinct parts: insusceptible period (amenorrheic or abstinent), susceptible period, and gestation. Because data on gestational age are not available in the DHS, the part of the interval related to gestation was not examined separately. Components of the Birth Interval Return of Menses 1 1 1 1 1 Index Sexual Birth Relations Fertile Conception Next Ovulation Birth The timing of ovulation is not directly observable; thus, no data are available on ovulation. However, data on the timing of first menses after each birth were obtained from the DHS. Menses can occur several months before ovulation, with the first few cycles being anovulatory, or ovulation can precede menses. Never- theless, the return of menses is considered to be a good proxy for ovulation. For this reason, we modeled the effect of a child's death on the resumption of menses. A series of models was estimated, as were the full models for the birth interval, by adding control variables and breastfeeding in turn. The piecewise hazards were modeled with steps placed at 0-2, 3-5, 6-8, 9-11, 12-14, 15-17, 18-23, and 24-35 months. The amount of exposure and the number of events beyond 36 months are so low that it was deemed imprudent to fit the model beyond 36 months.
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI 47 In many cases a woman reported that her menses never returned between one birth and the next. In such cases, the return of menses was assumed to have occurred 9 months before the birth. This treatment is consistent with our assump- tion that menses is a proxy for ovulation. In other cases a woman reported that menses returned, but the duration of amenorrhea exceeded the current age of the child (for open intervals) or exceeded the interval from birth to the next concep- tion (for closed intervals). We dropped these cases from the analysis of the duration of amenorrhea. It was expected that the death of the index child would shorten the time from birth to return of menses (increase the risks) but that this effect would work exclusively through the truncation of breastfeeding. Once breastfeeding was accounted for, the effect was expected to be zero. Interval from Birth to Sexual Relations We estimated a set of models exactly parallel to those for postpartum amen- orrhea using the timing of the resumption of sexual relations as the outcome. The baseline hazard was estimated in the same way, and the same set of models was estimated. Whereas postpartum amenorrhea is related to breastfeeding through biological mechanisms, postpartum abstinence is related to breastfeeding only through social norms and cultural taboos. Thus, the degree to which death of a child affects the duration of postpartum abstinence and the degree to which this effect is explained by breastfeeding were expected to vary from country to coun- try. Duration of Susceptibility to Pregnancy A series of models was also estimated on the second part of the birth interval, that is, after both menses and resumption of sexual relations had taken place. The models were specified in the same way as the models on the full birth interval, but all exposure during the periods of postpartum amenorrhea and postpartum absti- nence was excluded. Births were excluded for which return of menses or re- sumption of sexual relations was reported in the same month as the conception because they would contribute no exposure. In these models, exposure and the occurrence of subsequent births are classified by time since the index birth rather than by time since the start of susceptibility to pregnancy. We estimated alterna- tive formulations of the model by using time since menses or time since relations, but results did not change substantively. The effect of the death of a child on this portion of the interval could be explained partially by breastfeeding because breastfeeding reduces the probabil- ity of conception even for ovulating women. After adjustment for breastfeeding, the effect was expected to be small in countries where contraceptive use was
48 BIRTH SPACING: A CROSS-NATIONAL ANALYSIS infrequent because women have few means to control their fertility in response to the death of a child. Results Interval Between Births The first hazards model estimated (model 1) considered the time to the next live birth, including only covariates of age and the survival status of the index child. The coefficients of this model can be used to calculate a survival curve. The point at which this curve reaches 50 percent represents the median duration of the birth interval corresponding to the estimated model. In Figure 2-la, the estimated median duration of the birth interval for intervals after the birth of a child who survives 5 years is compared with the estimated median birth interval after the birth of a child who dies in the neonatal period. This comparison is shown for each of the countries2 for which data were analyzed. It thus gives a sense of the amount of reduction in the median birth interval that can be associ- ated with child mortality, without taking other factors into consideration. The results have been grouped by three regions: Africa, Asia, and Latin America. For each region the data have been arranged in order from the country with the longest median duration to the country with the shortest median dura- tion. Median birth intervals after the birth of children who are still living appear to be shorter in Africa than in Asia or Latin America. However, birth intervals after a child has died are generally similar across the three regions. Table 2-2 gives the ratio of the estimated median interval for subsequent births of children who survive to the median interval for births of children who die. This ratio varies between 1.21 and 3.15, indicating that the increase in birth interval associated with eliminating an early infant death is 21 to 215 percent. The average reduction in the median interval is 60 percent for the 46 DHS used for this analysis. Figure 2-2 shows a plot of the estimated coefficient of the variable DIED for all three models of the log of the hazard of closing the birth interval (conception resulting in live birth). Because explanatory covariates are added to the models in turn (confounders, then breastfeeding), the coefficients on DIED show the degree to which the effect of DIED, independent of covariates, continues to be a predictor of interval length. In this way, we try to assess how much of the effect of the child's death can be explained away by controlling for other factors that we also expect to affect birth interval length. 2Several of the countries analyzed here were actually surveyed twice. In the ensing presentation of results we refer to "countries" as a matter of convenience, although the data points of interest in fact represent countries at a point in time.
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI a. Full Birth Interval on So c: 45 ~ 40 .= 35 In 30 ~ ~ ~ _ - 25 20 15 an 10 - 1 1 ~-~-~ t-} i- 1- ~ +--l ~ ~ 1 ~ .~.,i.1,.! 1- 1 it--- Afr~ca Asia .. A: ~ ~-t-- ~ in. ~a , ~WN . . ~ t '\~:~^ ~ t! ~ \-~-~^ .. . ... .. .. .. ~. . . y ..~2 .. .... l- I---t- TV- t-~I~-i ,-l- ~ Latln America b. Postpartum Amenorrhea 0-' 16-~- tu 12 E 10 <: 8 ~ 4 :~:;~~~~~~~~~ ~ :` O 2 - A- -4 I 1- 'Africa ~ ~ ~ ~ ~ t- tam ~~-~~-l i iis''a-'~l I I I I I " ILat~n America _ - - - -~- - - ~ - c. Postpartum Abstinence ~ . _~ l\~_ __ _ _ _ _ _ _ _ _ _ __ _. . a' 20 v 18 16 ~ 14 n 12 c 10 8 6 0 2 ~-~-. ~--~ '--~---~-. O ~. . .. . ~- ~ . _ _ _ _ _ _ . ~ ~ ^~- ~-~ - -~--d ~N---~-- ~-- ~ j ~ 1.~ ~ 1~L I ~ ~ ~ I ( ~} ~ ~I ~1 ~1 1 1 1 I I I ~ ~ i ~ I I ~ 4_4i ~ Afr~ca As~a Lat~n America ~-~ Ch'ld living --~ ~ Child died I ' ~6P ql - _ FIGURE 2-1 Predicted length of interval by survival status of the index child. 49
so BIRTH SPACING: A CROSS-NATIONAL ANALYSIS TABLE 2-2 Ratio of Predicted Median Interval Length Assuming the Child Survived to the Predicted Median Interval Length Assuming the Child Died Time Birth Time to to First Region and Country Survey Interval Menses Coitus Africa Asia Botswana DHS1 1.53 2.61 1.43 Burkina Faso DHS2 1.43 2.76 2.65 Burundi DHS 1 1.42 2.78 1.28 Cameroon DHS2 1.44 2.78 1.28 Egypt DHS 1 1.88 3.05 1.43 Ghana DHS1 1.40 2.21 1.32 Kenya DHS 1 1.37 2.12 1.30 Liberia DHS 1 1.45 n.a. n.a. Madagascar DHS2 1.46 2.68 1.21 Malawi DHS2 1.28 1.85 n.a. Mali DHS 1 1.24 2.00 1.62 Morocco DHS1 1.54 6.19 n.a. Morocco DHS2 1.96 3.49 2.33 Namibia DHS2 1.41 2.32 1.33 Niger DHS2 1.33 3.04 1.57 Nigeria DHS2 1.43 2.26 1.89 Ondo State, Nigeria DHS1 1.28 1.59 1.77 Rwanda DHS2 1.53 4.23 0.84 Senegal DHS 1 1.31 3.70 1.63 Senegal DHS2 1.35 3.41 1.51 Sudan DHS1 1.27 5.68 1.41 Togo DHS1 1.38 2.41 2.25 Tunisia DHS 1 1.69 3.39 2.29 Uganda DHS 1 1.26 3.58 1.46 Zambia DHS2 1.21 2.42 1.35 Zimbabwe DHS 1 1.49 4.18 1.25 Average 1.44 2.95 1.44 Indonesia DHS 1 1.81 2.71 1.64 Indonesia DHS2 2.22 2.53 1.72 Pakistan DHS2 1.63 2.84 1.60 Sri Lanka DHS1 1.52 2.18 1.23 Thailand DHS 1 n.a. 1.50 1.50 Average 1.79 2.35 1.54 Latin America Bolivia DHS 1 1.55 2.68 1.59 Brazil DHS 1 1.87 2.97 1.85 Brazil DHS2 1.83 2.84 1.80 Colombia DHS 1 1.54 2.78 1.05 Colombia DHS2 2.52 1.96 1.13 Dominican Republic DHS1 1.60 2.21 1.86 Dominican Republic DHS2 1.74 2.86 1.63 Ecuador DHS 1 1.62 1.80 1.97
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI TABLE 2-2 (continue]) 5 Time Birth Time to to First Region and Country Survey Interval Menses Coitus E1 Salvador DHS1 1.50 n.a. n.a. Guatemala DHS1 1.41 3.97 1.67 Mexico DHS 1 1.75 2.39 1.58 Paraguay DHS 1 1.62 3.79 1.82 Peru DHS1 1.62 3.01 1.69 Peru DHS2 2.09 3.50 1.72 Trinidad and Tobago DHS1 3.15 2.74 1.41 Average 1.83 2.63 1.52 NOTE: n.a., data not available. The results for all 46 surveys are shown by region, arranged in the figure in order from the country with the largest crude effect of DIED (no other covariates) to the country with the smallest crude DIED effect. This is the same model that was used to generate the median durations shown in Figure 2-1A. The coeffi- cients of the crude DIED effect tend to lie in a range between 0.4 and 1.0 in all three regions, which corresponds to relative risks between 1.5 and 2.7. Morocco (DHS2) and Trinidad and Tobago are outliers with relative risks greater than 3.0. Model 2, labeled "with confounders," adds to the crude model the set of variables thought to be associated with both infant mortality and birth interval. These variables are mother's age, mother's education, birth order and previous interval, wontedness, socioeconomic status, residence, and sex of the child. Be- cause some of these variables were expected to mask a real association between death and the birth interval, but others were expected to artificially create an association, we had no prior expectations about what the net effect of these variables taken as a group would be on the coefficient of DIED. As shown in Figure 2-2, the net effect of these variables on this coefficient is virtually nil. In Table 2-3, we examine the degree to which the effect of child death is "explained" by each model. In the table, we compute the percentage of reduction in excess risk associated with death of a child, where excess risk is the estimated relative risk minus 1. The next to the last column of Table 2-3 shows the percentage of reduction in the excess relative risk in moving from the crude model (model 1) to the model with confounders (model 2~. For example, in model 1 for Botswana the risk of closing the birth interval when the child has died is 1.89 times the risk when the index child is still alive. This relative risk has been reduced to 1.86 in model 2 where confounders were introduced. The excess relative risk is reduced by 2.9 percent, which was calculated as 100 x (1.89 - 1.86~/~1.89 - 1.0~. By introducing the confounders, we have moved 2.9 percent
s2 Parameter Estimate 1.4 1.2 1 0.8 0~6 0.4 0.2 O -a ~ [~:~` e.e,. I ON :~< ~- ~. -0.2 11 11 1 11 1 1 ~1!1 1 ~ i Africa Asia Latin America Cwde BIRTH SPACING: A CROSS-NATIONAL ANALYSIS Reiative Risk - - - t 4 0 - 2.0 ~ 1.0 + W/ Breastfeeding: ~ W/ Confounders FIGURE 2-2 Effect of child death on rate of closing the birth interval. of the way from the relative risk of 1.89 for the crude effect to a complete explanation of the effect of death of a child, for which the relative risk is 1.0 (i.e., a model in which there is no excess risk of closing the interval when the child has died). With some exceptions, introducing the confounding variables did not reduce excess risk. In some cases the percentage of the reduction was negative, indicat- ing that the set of confounding variables as a whole was masking an association between child death and the birth interval. Model 3, labeled "with breastfeeding" in Figure 2-2, introduces an additional variable to model 2. This new variable, BF, is modeled in the same way as DIED (BF is a time-varying covariate that is set to true if the index child was still being breastfed 10 months before the hazard of closing the interval is being estimated).
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI 53 Again, the 10-month lag is introduced to allow for the 9-month delay from conception to birth. By introducing the BF parameter, we change the interpreta- tion of the parameter for DIED. In models 1 and 2, the comparison group for the relative risk for this variable was all children who are still alive. In model 3, the comparison group for both variables is living children who have been weaned. Figure 2-2 shows that the relative risk of closing the birth interval when the child is dead at a given age, compared with being weaned at that age, was considerably less for all countries than when the comparison group included breastfed children. This finding gives strong evidence that much of the excess risk of a shorter birth interval after death of a child is explained by the physiologi- cal effects of breastfeeding. The relative risks of closing the interval when the child has died, as opposed to when the child has been weaned, generally are 1.1 to 1.5. Many of these effects are no longer statistically significant (Table 2-3~. There are, however, notable exceptions in which the effect of child death remains high despite adjustments for the breastfeeding status of living children. These exceptions include Morocco (DHS2) with a relative risk or 1.94, Thailand with 2.55, Peru (DHS2) with 2.03, and Trinidad and Tobago (DHS1) with 2.81. As shown in the comparison of model 3 with model 1 in Table 2-3 (last column), introducing an adjustment to control for breastfeeding generally pro- duces a large percentage of reduction in the excess risk of child mortality on the hazard of closing the birth interval. In 14 of the 25 surveys analyzed for 22 African countries, controlling for breastfeeding status explains 65 percent or more of the excess risk of child death seen in the initial crude model. The same is true for 4 of the 5 surveys analyzed for 4 countries in Asia and for 8 of the 14 surveys analyzed in 11 countries in Latin America. Latin America seems to stand out as the region where the initially large effect of child death cannot be so consistently explained by premature weaning. The statistical models used here examine the effect of a child's survival status at a point in time on the risk of an event occurring at the next point in time. These models cannot address the question of whether it matters when the death occurred. We might want to know whether the effect of death of a child is immediate, that is, whether it affects events in the next 1 or 2 months and whether the effect persists several months after the child has died. To address this ques- tion, we selected a subset of five African countries. In these countries, we estimated an alternative formulation of the DIED variable in which we consid- ered the time since death. We considered if the risk of closing the interval is affected by whether the index child had died within specific intervals of time before the duration for which the risk is being estimated. Intervals used for this analysis were 10-12, 13- 18,19-24, 25-30, and more than 30 months. In all previous models presented we only consider whether or not the index child was dead in the month before the duration for which the risk is estimated. As with the crude models discussed above, the comparison group for these models was living children. The lagged
54 Cq o Cq a' _. a' be ·_4 Cq to o a' a' o a' · _4 o C) a' Cal ¢ EM ca · ~ ca ca C) x · ~ o 4= C) ca ~ o so o ~ >> ca ~ o so o ~ >> o ~ ~ ca ~ ~ o cd ;^ so V, ·bl) ~ ~ V ~ ~ ~ Do ~ o ~ ~ ~ ~ ~ ~ ~ o Do ~ ~ ~ ~ ~ ~ ~ Do ~ o ~ oo ~ ~ oo cd ~ ~ o ~ ~ ~ o oo ~ ~ ~ ~ ~CM ~ ~ ~ oo CM O ~ CM ~ ~ O ~ ~ ~ ~ ~ ~ ~ ~ ~ CM . . . . . . . . . . . . . . . . . . . . .. . . . CM ~ ~ O oO CM O ~ ~ ~ ~ CM CM ~ CM ~ ~ CM O CM ~ ~ ~ oO 1 ~ ~ ~ ~ 1 ~ ~1 ~ 1 1 1 1 1 * ** * * * * * * * * * * * * * * * * * * * * * *. O O ~ ~ ~ CM oO ~ ~ ~ ~ ~ oO oO ~ O O CM oO ~ ~ CM Ct ~ ~ ~ ~ ~ ~ ~ ~ ~ CM O CM oO ~ ~ ~ ~ . . . . . . . . . . . . . . . . . . . . . . . . . ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ O ~ ~ ~ ~ ~ ~ ~ ~ ~ O * * * * * * * * * * * * * * * * * * * * * * * * * O CM ~ oO ~ ~ ~ ~ ~ ~ CM oO O oO ~ ~ ~ O ~ CM ~ O oo ~ ~ ~ ~ ~ oo ~ O ~ ~ ~ ~ O oo ~ ~ ~ oo . . . . . . . . . . . . . . . . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * O CM ~ ~ ~ ~ ~ ~ ~ ~ O O ~ ~ O oo ~ ~ ~ ~ ~ oo ~ ~ ~ ~ ~ ~ ~ ~ O . . . . . . . . . . . . . . . . . . . . . ~ cM cM cM cM cM ~ ~ cM ~ ~ cM ~ ~ ~ cM ~ cM ~ ~ ~ * * * * . . . . v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, ~a · ~ .~ O z ° e 3 0 0 . O ~R ~ ~ ~ o ~ .~ ~ C 3 ~ ¢ .~ ~ . O .ca 0 ~ ~ 63
~ ~ ~ ~ ~ ~ oo ~o ~ ~ ~ oo M ~ ~ ~ oo cM cM M M ~ ~ ~ ~ ~ o ~ ~ ~ oo ~ ~ ~ ~ ~ ~ ~ ~ o ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ - ~ cM ~ ~ ~ ~ ~ ~ ~ ~ ~ cM ~ ~ cM o oo ~ ~ o ~ ~ ~ ~ ~ . .. . . . . . . . . . . . . . . . . . . . . . . . . . . M ~ ~ ~ ~ oo oo ~ ~ ~ ~ cM ~ ~ ~ ~ ~ ~ ~ o ~ ~ cM 1 ~ 1 ~ 1 ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ 1 ** * * * * * * * * * * * * * * * * * * * * * * * * * M ~ ~ ~oo oo ~ oo ~ ~ ~ ~ *. cM o oo ~ ~ ~ M oo ~ ~ ~ ~ ~ ~ cM oo ~ cM ~ ~ oo ~ ~ cM ~ ~ ~ ~ cM o oo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . M * * * * * * * * * * * * M ~ o ~ ~ ~o o ~ cM oo ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ . . . . . . . . . . . . . M ~ ~ ~ cM cM cM cM ~ cM * * * * * * * * o oo ~ cM \ ~\ ~) . . . . . . . . ~ ~ ~ ~cM cM ~ ~ cM * * * * ~ cM o oo cM . . . . . ~ ~ cM ~ ~ ~ cM ~ cM cM ~ ~ ~ cM ~ cM ~ cM n cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn ~4 ~·0 .0 ~ ~ ~ ~0 ~ cn cn cn E~ .O ·0 ~Ct Cd Ct · ~ Ct Ct Ct R ~ ~R R ~ ~ · ~ ¢ ~> ~ ~ O O R R ca cd ¢ ~ 55 * * * * * * * * * * * * * * oo ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ oo oo ~ ~ oo oo oo ~ ~ ~ . . . . . . . . . . . . . . . * * * * * * * * * * * * * * ~ ~ o ~ ~ ~ ~ ~ ~ ~ ~ oo ~ ~ ~ ~ ~ ~ ~ ~ ~ o oo ~ ~ oo oo ~ ~ o . . . . . . . . . . . . . . . cM ~ ~ ~ cM ~ cM ~ ~ ~ ~ ~ ~ cM ~ cn cn cn cn cn cn cn cn o o E~ O ~ R O =t cd . >< t~ ~ ~ ~R
s6 Parameter Estimate 2 1.5 1 BIRTH SPACING: A CROSS-NATIONAL ANALYSIS Relative Risk ~ ~it, 0.5 ; -0.5 1 -1 .5 | ~ Kenya ~ Mali - ' 7 4.0 2.0 1.0 0.5 -0.25 l 2 1 1 i 1 ~1 10-12 13-18 19-24 25-30 34+ Months Before Birth J ~ Burundi ~ Ghana ^~ Zambia FIGURE 2-3 Lagged effect of child death on rate of closing the birth interval. effects of a child death are shown in Figure 2-3. In all five countries, the effects are greatest immediately after the death occurs and wane over time. If a birth has not occurred within 2 years after the death, the death of the index child is irrel- evant to the closure of the birth interval. To address whether the observed effects are stronger if a child is older or younger when it dies, we used the same five African countries to estimate models in which the DIED variable was allowed to have time-varying effects. For the full birth interval, effects of a child death are strongest for closing the birth interval within 24 months (Figure 2-4~. Effects fall off quickly beyond this point;
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI Parameter Estimate 2 1.5 T.----- t~ \ \W 1 0.5 o -0.5 -1 1 ~\,\~ ~ a\ ~ \ \\ ~ \ . . . ~ - 57 Relative Risk .. . ... 4.0 2.0 ----I ~ 1.0 ..... ~ 9-23 24-29 30-41 4~-5g Months Since Index Birth l - Kenya ~ Mali 0.5 ~ Burundi ~ Gha nil : ~ Zambia FIGURE 2-4 Age-dependent effect of child death on rate of closing the birth interval. for intervals that are still open at 2.5 years postpartum (30 months), the occur- rence of a child's death does not affect the propensity to close the interval. Interval from Birth to Menses The major physiological pathway through which the death of a child is hypothesized to shorten the subsequent birth interval is by hastening the return of menses and ovulation because the mother is not breastfeeding. In Figure 2-lb, the predicted duration of postpartum amenorrhea after the birth of a child who survives until menses returns is compared with the duration after the death of a neonate for the full set of countries in the DHS. Figure 2-lb indicates the
58 BIRTH SPACING: A CROSS-NATIONAL ANALYSIS magnitude of the overall crude effect of an early infant death on the duration of postpartum amenorrhea. As in Figure 2-la, the intervals were calculated from the coefficients of a hazards model of time to resumption of menses in which the only covariates are age and whether the child is dead at the beginning of each month. Africa stands out as having considerably longer median durations of postpar- tum amenorrhea, both when the index child who started the interval dies and when the child survives until return of menses (Figure 2-lb). The percent in- crease in the median duration of amenorrhea associated with the child's survival is considerably greater than was the case for the effect on the overall birth interval (see Table 2-2~. The percent increase in the time until return of menses when the child survives over when the child dies is 50-468 percent and usually is 100-250 percent. Analogous to the models of the hazard of conception resulting in a live birth, three separate models have been estimated for the hazard of return of menses. Figure 2-5 shows the estimated coefficient of DIED for each of the three models for all of the countries. As before, the countries are ordered, within regions, by the size of the DIED coefficient estimated in model 1. As expected, the relative risks for the duration of amenorrhea in mothers of children who died compared with that in mothers of those who lived are highest in Africa (range, 2.07 to 5.14), lowest in Asia (range, 1.58 to 2.69), and interme- diate in Latin America (range, 1.95 to 4.65~. As in the models for overall birth interval, there was little change in the estimated coefficients of DIED when the confounder variables were added for model 2. In Latin America there was a tendency for the addition of the confound- ing variables in model 2 to increase the effect of DIED over the effect seen in the crude model. This increase indicates that, as a whole, the confounders actually mask an association between death of a child and return of menses in women in Latin America. In model 3 of the hazard of the return of menses, the addition of the variable on breastfeeding status considerably reduced the effect of DIED for all but a few countries (Table 2-4~. But surprisingly, the risk of return of menses after a death, relative to the risk after weaning a living child (Table 2-4, model 3), remains high, especially in Africa where it is greater than 2.0 for 16 of the 25 countries. Also, the percent reduction in the excess risk of DIED that was achieved by adding the variable for breastfeeding (Table 2-4, model 3 versus model 1) is not as great as the reduction in the analogous model estimated for the birth interval (Table 2-3, model 2 versus model 1~. The percent reduction in excess risk for Africa and Asia rarely exceeded 60 percent, whereas the percent reduction was greater than 60 percent for all but three countries in Latin America. As with the full birth interval, we examined the data from the five African countries to determine the lagged effect of a child's death on the risk of menses
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI Parameter Estimate 2.0 T 1.B ~------------------------------~-~ 1.6 1.4 i 1.0 0.8 0.6 0.4 0.2 0.0 Relative Risk 14: .. -a ~ f ~ ~ ~ -me t; -0.2 ~1; ~1111 I ' it"` its.,` no, :~ '~ In' lye '.~- - 1-1---~1 1 1 ~1 1 ~ ~--~ - Africa ~ 4.0 . Asia Latin America 2~0 1.0 · - Crude ~ W/Confounders ~ W/Breastteeding FIGURE 2-5 Effect of child death on rate of resuming menses postpartum. 59 returning (Figure 2-6~. We considered whether the death took place 1, 2-3, 4-5, 6-7, or more than 7 months previously. These intervals are shorter than those for the full birth interval because the duration of amenorrhea is so much shorter than the birth interval and there are more events on which the hazards can be esti- mated. Again, risks are estimated relative to intervals in which the index child was still alive. The largest effect of death of a child on return of menses occurs immediately after the death (Figure 2-6~. Effects wane with time and are negli- gible beyond 7 months after the death. We also considered age dependencies in the effect of a child's death on return of menses for these five countries. The effect was essentially confined to the first year of life (Figure 2-7~. Except in Burundi, effects were minimal in the second year. In the third year of life (24-35 months), the parameter estimates
60 . ca ca C a' To o ~ ·~O be 3 to o Cq a' a' be Cq o a' a' o a' · _4 o C) a' Cal ¢ EM ca .~ ·4~= ca ~ o so o ca o so o o ~ ~ ca - °. o ~ ~ ;^ V, ·bl) ~ V ~ ~ ~ Do ~ ~ ~ ~ ~ Do ~ ~ ~ ~ o ~ o ~ o ~ ~ ~ ~ ~ ~ o CM ~ ~ CM ~ ~1 o ~ ~ ~ o oo ~ ~ oo oo CM ~ CM ~ O ~ O . . . . . . . . . . . . . . . . . . . . O ~ ~ ~ O ~ O ~ CM ~ ~ ~ O 1 ~1 1 1 1 ~1 1 ~1 ~1 1 1 1 1 1 1 * * * * * * * * * * * * * * ~ ~ O ~ 0 ~ ~ oo oo ~ oo CM oo ~ O ~ oo ~ oo . . . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * O ~ ~ ~ oo ~ ~ oo ~ CM ~ oo ~ O ~ oo CM ~ ~ ~ CM ~ ~ ~ O ~ CM ~ ~ CM ~ ~ ~ CM CM . . . . . . . . . . . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * ~ CM ~ ~ ~ CM ~ CM ~ O ~ CM ~ O CM O ~ ~ ~ . . . . . . . . . . . . . . . . . . . CM CM ~ CM ~ ~ ~ CM CM CM ~ CM CM ~ CM ~ CM ~ ~ ~ CM CM ~ ~ CM CM CM CM ~ CM ~ CM ~ V, V, V, V, V, V, V, V, V, V, V, V, V, V, V, V, V, V, V, V, ·_4 .~ O Z Cd s~ g ° O U rLi ~ ~ ~ ~ ~ ~ ~ .~ .~ ~ 3 c ¢ o ~ oo o ~ . . . 1 1 1 ~ * * * * * * * * * 1 ~ _I O ~ _I O 00 ~n ~ ~ 1 -I ~ ~ -I ~ ~ ~_I ~ O . . . . . . . . . . , * * * oo ~ oo ~ o ~ . . . * * * o o . . . V~ V~ V ·0 o . o
61 ~ ~ o ~ ~ oo ~ ~ ~ ~ oo ~ * * * . . . o ~ oo ~ oo ~ ~ oo ~ ~ M ~ o ~ ~ ~ . . . . . cM cM cM ~ 1 1 1 oo oo ~ ~ ~ . . . . . M ~ 1 1 1 ~ ~ ~ ~ ~ ~ o o ~ ~ ~ ~ ~ ~ o ~ ~ cM ~ ~ ~ ~ ~ ~ o cM ~ ~ ~ ~ o ~ ~ ~ ~ ~ ~ ~ ~ ~ M . . . . . . . . . . . . . . M o ~ cM ~ cM 1 ~ ~ 1 1 ~ ~ ~ ~ 1 1 1 1 1 1 1 1 1 1 * * * * ** * * * ** * * * * * * * * * * * * M cM ~ ~ ~ ~ ~ ~ o ~ oo cM ~ ~ o ~ ~ ~ ~ ~ ~o ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ . . . . .. . . . .. . . . . . . . . . . . . . ~ ~ cM cM cMcM ~ ~ ~ ~cM ~ ~ ~ o ~ ~ ~ cM * * * * ** * * * ** * * * * * * * * * * * * * oo ~ oo ~ ~ cM ~ ~ ~ o ~ ~ ~ o ~ ~ ~ o ~ oo ~ ~ ~ O ~ oO ~ ~ c~ ~ ~ ~ O ~ ~ oo ~ O ~ oo ~ ~ ~ oo . . . . .. . . . .. . . . . . . . . . . . . . ~ ~ ~ c~ ~c~ c~ ~ ~ ~ c~ ~ c~ ~ c~ ~ ~ ~ c~ * * * * * * * * * * * * * * * * * * * * * * * * * * * O ~ ~ ~ ~ ~ ~ oo ~ ~ ~ ~ ~ oo O ~ ~ O ~ ~ ~ ~ ~O ~ oO ~ ~ c~ oO ~ ~ c~ . . . . . .. .. . . ... . . .. . . . . . . . . . c~ ~c ~ c ~ c~ c~ ~ ~ c~ c~ ~ ~ c~ c~ n cn cn cn ;~ ~ ~ ~ . cn cn cn ~ ~ ~ CM ~ CM CM ~ ~ ~ CM ~ CM ~ CM ~ ~ ~ CM ~ CM ~ cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn cn O . v, R ~ , ~R R v, ca ¢ .O .O O ~ ~E-° .O cd cd R R ~d ~ ~ R R ,~ R 0 4= c' 4= . . ca ;^ c' ·4= ca ·4= ca ca ·_4 c' O c' s~ *
62 Parameter Estimate 2~5 I- 2 1.5 1 0.5 o -0.5 BIRTH SPACING: A CROSS-NATIONAL ANALYSIS Relative Risk W- e - - - tic ~ \ __~__~_ \ ),'~,\\~ ado \ '~- ~ \ ~ Am. .: I .. ........ ~ 8.0 4.0 2~0 1 .0 l 2-3 4-5 6-7 8+ Months Before Menses ~ Kenya ~ Mali ~ Burundi ~ Ghana - ~ Zambia FIGURE 2-6 Lagged effect of child death on rate of menses returning postpartum. vary considerably across the countries, most likely because of the small amount of exposure in this group. Interval from Birth to Sexual Relations A behavioral pathway through which a child's death may shorten the subse- quent birth interval is by shortening the period of postpartum abstinence. In Figure 2-lc, the estimated duration of postpartum abstinence after the birth of a child who survives until sexual relations are resumed is compared with the dura- tion after the death of a neonate. The death of a child consistently results in a shorter duration of abstinence in all countries.
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI Parameter Estimate 2.5 1 ~ . ~ ~1 ~ ~_; ~ ~ \ ~ \ Relative Risk 8.0 0 1 ~- Em, 1 2.0 0.5 ~ ~\' ~ -0 . 5 -I - - - - -- - - - \. l -1 i .,,, ~ _ ~,, _ 1 0-1 1 12-23 24-35 Months Since Index Birth -~.0 - 0.5 ~ Kenya ~ Mali ~ Burundi ~ Ghana ~ Zambia 63 FIGURE 2-7 Age-dependent effect of child death on rate of menses returning postpar tum. Table 2-2 (column 3) gives the ratio of the estimated median duration of postpartum abstinence for births of infants who were still alive, to the duration following births of infants who died as neonates. Except in Rwanda, where the relationship is oddly reversed, the ratios are 1.05 to 2.65. This variation indicates that the lengthened abstinence associated with eliminating an early infant death is 5-165 percent. The relative effect of a child's death is not as great as that for postpartum amenorrhea or for the overall birth interval. On average, across the 46 surveys, child survival increased the birth interval by 60 percent, postpartum amenorrhea by 178 percent, and postpartum abstinence by 47 percent. Again, three separate models were estimated for the hazard of resuming
64 BIRTH SPACING: A CROSS-NATIONAL ANALYSIS sexual relations. Figure 2-8 shows that the effect of DIED estimated in the crude model is virtually unchanged by the addition of either the confounding variables or the variable for breastfeeding status. We conclude that the effect of a child's death on increasing the risk of resuming sexual relations is direct and does not appear to operate through breastfeeding. However, there does appear to be a group of Afncan countries for which an addition of the breastfeeding status variable actually increases the effect of DIED over that in the initial crude model. In Table 2-5 (model 3 versus model 1), the percentage of reduction in the excess risk is negative for most Afncan countnes. This finding is puzzling and requires further investigation. Parameter Estimate 4.6 41.4 ].2 1 0~8 0.6 0.4 0.2 o -0.2 0 4 t I | c | | | | ~I I ~I I I I I I | - -A ~ _ ~___~__ . _ ~ _ Africa Relative Risk _ 4.0 2.0 1.0 Asia Latin America a Crude ~ W/ Confounders · W/ Breastfeeding FIGURE 2-8 Effect of child death on rate of resuming sexual relations.
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI Interval from Menses and Sexual Relations to Birth 65 For the first set of models for the entire birth interval, some countries showed a general excess effect of a child's death on the risk of conception resulting in a live birth, even in comparison to the effect of a living child who was weaned (model 3 in Figure 2-2 and Table 2-3~. This finding was more likely in Latin America than in either Africa or Asia. We decided to estimate an analogous series of three models of the risk of conception resulting in a live birth. In these models, all the exposure before resumption of menses and sexual relations has been excluded. In this way, we eliminated postpartum amenorrhea and abstinence from being explanatory mechanisms for the effect of DIED. Figure 2-9 and Table 2-6 show the results of these three models. The results of the first model show that, with just the crude effect of DIED, for all the countries there is an excess risk for death of a child when conception of a subse- quent child results in a live birth. This excess risk exists even when postpartum abstinence and amenorrhea are removed as possible mechanisms. In Africa, the crude relative risk associated with a child's death is greater than 1.5 in only 8 of the 23 surveys analyzed in 21 countries, whereas it is greater than 1.5 in 3 of the 5 surveys analyzed in 4 Asian countries and in 8 of the 14 surveys analyzed in 10 Latin American countries. We can thus see that in Africa much of the effect of DIED is operating through postpartum amenorrhea and abstinence, as expected. Even within Africa an effect larger than 1.5 is primarily limited to two regions: North Africa (Morocco, Egypt, and Tunisia) and East Africa (Burundi, Rwanda, and Zimbabwe). As before, a second set of models was estimated in which the confounding variables were added to the first model. This generally had little effect on the coefficients for the effect of DIED in Africa and Asia, but did substantially reduce the effect in most Latin American countries. In Table 2-6, the addition of the confounding variables in Latin America generally explained 30-60 percent of the excess risk of conception resulting in a live birth, which is associated with death of a child in the first model. When a control for breastfeeding status was added to the model, the effect of DIED was essentially eliminated for most countries in Africa and Asia. In Africa, the effect of DIED is insignificant in all countries except Morocco; in Asia, only Thailand has a relative risk (2.31) greater than 1.17 (Table 2-6~. In Latin America, five countries continue to show a significant effect of a child's death, but there is a substantial reduction relative to the crude model. This effect of breastfeeding is not acting through lactational amenorrhea because menses has already returned for these women. We examined interactions of the effect of DIED with the sex of the child, birth order, previous birth interval, and whether the child was wanted at that time. We hypothesized that the effect of death of a child would be greatest if the child were male, were of a low parity (especially first), or were wanted at that time.
66 O ca Cq X o .~ ·_4 be 3 o o Cq o .~ a' x VO be Cq o a' o a' · _4 o C) Cal ¢ EM ca 4~= ca ~ o so o ca ~ o so o o ca - °. o ~ ~ ;^ so V, two ~ V o ~ ~ ~ ~ ~ ~ ~ ~ o ~ o o ~ ~ ~ ~ ~ ~ o ~ ~ ~ ~ ~ CM ~ ~ 0 oo oo ~ oo ~ ~ ~ ~ ~ ~ 00 ~ ~ t ~ I ~ ~ ~I I I I ~ ~ ~ ~ o I I ~ ~ y o * * * * * * * * * * * * * * * * * * * oo ~ ~ ~ ~ o ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~oo ~ ~ ~ o ~ oo ~ ~ ~ o ~ ~ CM ~ o oo CM CM oo o . . . . . . . . . . . . . . . . . . . . . . ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ _I ~1 * * * * * * * * * * * * * * * * * CM ~ ~ o CM ~ oo ~ ~ ~ ~ ~ ~ CM o ~ ~ ~ ~ ~ ~ ~ CM oo . . . . . . . . . . . . . . . . . . ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ o * * * * * * * * * * * * * * * * * oo oo oo ~ ~ ~ ~ oo . . . . . . . . . . . . . . . . . . ~ CM ~ CM ~ ~ ~ CM ~ ~ CM ~ ~ ~ o ~ ~ ~ * * * * ~: CM . . . . * * * * o ~ ~ ~ . . . . ~ ~ ~ _ v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, · - ~ . o z ca S~ ~ cd 3 c ~ ,,, c ~ :^ ~ .~ ° ~ ,,, b,, m m m v ~ ~ ~ ~ ~ ~ Z Z Z O ¢
67 ~ ~ ~ ~ ~ o CM ~ . . . .. . . . . ~ ~ ~ ~ oo o ~ oo ~ ~ ~ o 1 1 11 ~1 CM~ ~ ~ CM . . . .. . . . . . . . . . o ~ o ~ o CM ~ oo 11 ~ 1 1 1 * * ** * * * * * * * * * ~ ', ~ ~oo ~ ~ CM oooo ~ ~ ~ o o oo CM CM oo o ~ ~ ~oo ~ ~ ~ ~ . . . . . . . . .. . . . . ~ ~ CM ~CM CM ~ CM ~ ~ CM ~ ~ o o ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ oo ~ ~ ~ ~ ~ CM ~ ~ CM ~ oo oo 1 1 1 ~ ~ 0 1 ~ ~ ~ 1 1 ~ ~ ~D ~ ~ ~ oo ~ ~ cn v~ ~ c~ x cn . . . . . . . . . . . . . . oo ~ ~ ~ ~ ~ o ~ o CM 1 oo ~ ~ 1 c~ 1 1 1 1 * * * * * * * * * * * * oo ~ ~ ~ o oo ~ ~ ~ oo oo ~ o o oo ~ oo . . . . . . . . . . . . . . * * * * * * * * * , o ~ ~ ~ ~ o oo ~ ~ oo o ~ ~ ~ ~ ~ oo . . . . . . . . . . . . ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ 4= C) 4= 4= C) ca ;^ O ca ·4= ca ca 4= ·C) O S~ * * * * * * * * * * * * * ** * * CM ~ ~ ~ ~ ~ ~O ~ ~ ~ ~00 00 ~ ~ ~ CM ~ ~ ~00 ~ O . . . . . . . . . . . . . .. . . O ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ * * * * * * * * * * * * ** * * * * * * * * * * * ~ ~ ~ ~O ~ ~ ~ O ~ CM ~ O ~O ~ O ~ ~ ~ ~ ~ ~ 00 ~ CM CM oo ~n ~ ~ ~ ~ ~n ~ ~ 0 ~ ~ oo oo 0 _I oo ~ ~ ~ ~n ~ ~ oo . . . . . . . . . . . . . . . . . . . . . . . . . . . . O CM ~ CM ~ ~ ~ CM ~ CM CM ~ ~ ~ CM ~ CM ~ CM ~ CM ~ CM ~ CM V~ V~ V~ V ~V~ V~ V~ V~ V ~V~ V~ V~ V~ V ~V~ V~ V~ V~ V~ V~ V~ V~ V~ V~ V~ V~ V~ V 3 R R ~04 8 ~ R R v~ v~ v ~E~ E~ Ct ~ ·Ct . ·~ ~ ~ ~ ~ ~ ca ¢ .O .O O E-° . ~ ~ ~ ~ ~ R R · ~ · ~ ~ ~ .
68 Parameter Estimate 1.2 1 0.8 06 0.4 0.2 O -0.2 -0.4 l -0.6 Crude BIRTH SPACING: A CROSS-NATIONAL ANALYSIS llelati~'e Risk :\ 5~-~: W/ y ~11~` a;,., . ~ . ~-I 1 1 1 1 1 1 1 1 I I~ - Africa Asia Latin America ~ 3.0 - 2.0 1.0 ~ ~. . ~- ~ W/Confounders · W/Breasifeeding . . . ... ... ... . . FIGURE 2-9 Effect of child death on rate of closing the birth interval after menses and sexual relations have resumed. For none of these interactions did we find a consistent pattern of effect across the countries. DISCUSSION In this analysis, we have demonstrated that the death of a child has a substan- tial effect on the birth interval. We estimated that the median birth interval is 60 percent longer when a child lives than when it dies in early infancy. Preston (1978) estimated that the death of a child would shorten the average birth interval by 30 percent. However, the methods we used here are quite different from those Preston used. First, for our estimate, we use the median birth interval associated
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI 69 with the death of a child, rather than the overall mean birth interval, as the denominator. Second, the 1978 estimate does not account for the timing of the events of death of a child and the end of postpartum sterility and therefore allows for a potential reverse causality. If menses returns early while the child is still living, the subsequent interval to the next birth will likely be short, which may itself be a risk factor for infant death (Hobcraft et al., 1985~. Furthermore, early weaning could lead to both the death of the child and the early return of menses. In our analysis, the survival status of the child is considered in the month prior to the month for which risks are being estimated, so reverse causality is not a problem. Third, our predicted birth intervals are based on creating a survival function in which the child's status is dead or living for all durations. Thus, the 60 percent estimate presented here is the reduction associated with elimination of all child death, not just infant death. Finally, whereas Preston's figure was based only on the mechanism of shortening the period of postpartum sterility, our estimate includes postpartum sterility as well as the interval when the woman is susceptible to pregnancy because of not using contraceptives and coital frequency. Our analysis has demonstrated that a major portion of the effect of a child's death on shortening the subsequent birth interval operates through the premature truncation of breastfeeding. Overall, of the excess risk associated with an early infant death on closing the birth interval, 64 percent is explained by breastfeeding. This percentage is nearly identical in all three continents. As expected, we found that the duration of amenorrhea is longest in the African countries. Death of a child has a substantial effect on the return of menses. Although premature weaning explains part of this effect, we were sur- prised to find a large effect remaining even after adjustment for breastfeeding. The substantial effect of child death on the duration of postpartum abstinence was also surprising, as was the finding that the effect does not seem to operate through breastfeeding. We had expected that a major reason for prolonged peri- ods of postpartum abstinence was a social taboo against intercourse during breastfeeding and a belief that intercourse would poison the milk (Aborampah, 1985~. Instead, the norms for sexual activity appear to relate to the presence of the child, rather than to feeding patterns. The duration of postpartum abstinence from sexual activity is generally not an important determinant of the length of birth interval because the period of abstinence is much shorter than the period of amenorrhea. With the exception of a few countries in Africa where the duration of abstinence is exceptionally long, the median duration of amenorrhea is longer than the duration of abstinence in every country. Particularly long durations of abstinence exist in the countries on the Gulf of Guinea: Ondo State, Burkina Faso, Togo, Ghana, Cameroon, and Nigeria. Notably, the effect of a child's death is far greater in Togo and Ondo State than in any other countries, and a large portion of the effect is explained by breastfeeding. The effect of a child's death continues to operate even after menses and
70 a' vO He a' a' c) · - ~ 3 ·_4 a' o Cq ·_4 a' _. a' be ·_4 Cq o o a' a' o a' o 1 Cal Cq o ~ ·~ ¢ a' EM ~ ca .~ ca ca X o ·4= C) ca .~ ·4~= ~ V, ~ ca ~ o so o ~ V, ~ ca ~ o so o · ~ ,= ca 4= o ~ ~ ca o 4._, o to ~ ;^ o ~ ~ o ~ V ~ ~ o cM ~ oo ~ o ~ ~ cM oo ~ ~ ~ ~ ooo cM cM . . . . . . . . . . . . . . . . .. . . ~ ~ ~ ~ oo oo oo ~ ~ ~ ~ ~ oo o ~ o ~ o ~ M ~ ~ ~ ~ ~ ~ ~ ~ cM ~o~ ~ ~ ~cM o ~ ~ ~ ~ ~ . . . . . . ~ ~ ~ ~ ~ oo 1 oo ~ ~ ~ ~ ~ ~ cM o . . . . . . . . . M 1 1 1 ~ ~ ~ ~ ~ ~ ~ ~ ~ oo o ~ ~ oo ~ oo oo o o o ~ ~ cM ~ ~ ~ ~ cM o ~ o ~ ~ ~ . . . . . . . . . . . . . . . . . ~ ~ ~ ~ ~ ~ o ~ o ~ ~ ~ ~ ~ o o o * * * * * * * * * * * * M ~ ~ ~ ~ ~ ~ ~ ~ o cM oo ~ ~ ~ ') ~ ~t ~ ~ cM ~ ~ ~ ~ ~ ~ oo ~ o oo . . . . . . . . . . . . . . . . . ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ o * * * * * * * * * * * * ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ oo ~ cM ~ ~ ~ oo ~ cM ~ o o ~ ~ ~ oo cM o o . . . . . . . . . . . . . . . . . * ~ cM ~ cM ~ ~ ~ cM ~ cM cM cM ~ cM ~ cM ~ ~v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, s ~=~= V, ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ P ~ ~ ~ ~ . . . oo 1 ~ 0 ~ . . . O ~ 0 * * . . . * * ~ 0 ~ . . . V~ V~ V~ . 0 Z ca S ~^ R c;~ .= O ~, ~ ,c,V) ~ ~ t~o R o V) R ~ m m m v ¢
71 . . . . ~ ~ o ~ ~ ~ o . . . ~ oo ~ 1 ~ ~ ~ ~ oo oo . . . . ~ o o o o cM cM ~ cMoo ~ ~ ~ ~ . . . . .. . . . . cM ~ oo ~ ~ ~ oo oo o ~ 1 M oo cM . . . . .. . . . . cM o oo ~ ~oo ~ ~ ~t 1 ~ ~ ~ c~1 ~ O c~ 1~ 1 * * * ~ o ~ ~ o ~ ~ ~ ~ ~ M . . . . . . . . . . o ~ o o ~o ~ ~ o ~ ~ ~ cM ~ ~ o ~ ~ ~ ~ o ~ o ~ . . . . . . . . . . . . . . ~ oo cM ~ ~ ~ ~ ~ ~ oo ~ ~ ~ oo oo ~ ~ oo ~ ~ ~ ~ ~ c~ ~ ~ ~ 1 M . . . . . . . . . . . . . . M cM cM ~ ~ ~ oo ~ ~ ~ ~ cM 1 c~ ~ ~ * * * * 7~ ~ ~ ~ ~ cM o ~ ~ o oo ~ ~ ~ oo ~ ~ ~ ~ ~ . . . . . . . . . . . . . . ~ ~ ~ ~ ~ ~ ~ o o ~ ~ ~ ~ ~ * * * * * ** * * ** * * * * * * * * ~- ~ ~ ~n ~ c~ O oo~ ~ ~ ~ ~ O ~ ~ ~ ~ ~ ~ ~ ~ ~ O . .. .. . . . .. . . . .. . . . . . . . . . . . . . O~ ~ ~ ~ ~ ~ ~ O * * * * ** * * ** * * * * * * * * * * * oo ~ ~ ~ ~ ~ ~ 0 ~ ~ ~ ~ ~ ~ oo oo ~ 0 0 CM O O CM O CM ~ ~ CM ~ O ~ ~ ~ ~ ~ CM ~ CM ~ ~ ~ ~ CM .. . . . . . . . . . . . . . . . . . . . . . . . . . . Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q (Q ;- ~ ~ ~ ~ (Q (Q (Q o R t~o ~ ~O O ~ ~ ·= C caCt ¢~ .O .O O ~ Ct · ~ _ - R o ~ ~ ~ ~ ~ ~ ~ ~ R C) 4= 4= C) ca ;^ C) ·4= ca ._4 4= ca ca ·_4 C) O C) s~ *
72 BIRTH SPACING: A CROSS-NATIONAL ANALYSIS sexual relations have resumed. Part of the effect operates through the truncation of breastfeeding when the child dies. It is not clear whether direct physiological mechanisms are at work here (such as anovulatory cycles after return of menses or reduced fecundity) or whether indirect effects of the breastfeeding experience are more important (such as fatigue and nighttime feedings which reduce coital frequency). But whatever the reason, breastfeeding does seem to confer a protec- tive effect against closing the birth interval, even after return of menses and sexual relations. Breastfeeding explains most of the susceptible period DIED effect in Af- rica after controlling for breastfeeding, the effect of DIED is insignificant in every country in this region except Morocco (DUSK. This finding is not surpris- ing since contraceptive use is generally low in Africa. The effect of child death is virtually explained away by breastfeeding because women have few mechanisms to increase their fertility, such as stopping contraception after the death of a child. On the other hand, the effect is largely explained by confounders in much of Latin America. One hypothesis for this finding is that, in Latin America, the confounders (especially mother's education, residence, socioeconomic status, and parity) are acting as proxies for the propensity to use contraceptives. As expected, the effect of a child's death on the length of the susceptible period appears to operate primarily through breastfeeding and contraception. We were surprised to find that the effect of DIED did not depend on whether the child was wanted. We would have expected a much stronger effect of the child's death if the child had been wanted, in that there would be a desire to replace the lost child. It is unclear whether this result reflects the difficulties mothers have in describing a child (even one who has died) as unwanted or whether the wantedness of a previous child is simply not a good predictor of subsequent behavior. ACKNOWLEDGMENT We thank Bridgette James for help in obtaining the Demographic and Health Survey data sets and Sandy Jewell and Ellen Borland for their assistance in managing the original data sets. REFERENCES Aborampah, O.M. 1985 Determinants of breast-feeding and postpartum abstinence: Analysis of a sample of Yoruba women, Western Nigeria. Journal of Biosocial Science 17(4):461-469. Guo, G., and G. Rodriguez 1992 Estimating a multivariate proportional hazards model for clustered data using the EM algorithm, with an application to child survival in Guatemala. Journal of the American Statistical Association 87:969-976.
LAURENCE M. GRUMMER-STRAWN, PAUL W. STUPP, AND ZUGUO MEI 73 Heckman, J., and B. Singer 1982 Population heterogeneity in demographic models. Pp. 567-599 in K. Land and A. Rogers, eds., Multidimensional Mathematical Demography. New York: Academic Press. 1984 A method for minimizing the impact of distributional assumptions in econometric models for duration data. Econometrica 52(2):271-320. Hobcraft, J.N., J.W. McDonald, and S.O. Rutstein 1985 Demographic determinants of infant and early child mortality: A comparative analysis. Population Studies 39(3):363-385. Howie, P.W., A.S. McNeilly, M.J. Houston, A. Cook, and H. Boyle 1982 Fertility after childbirth: Infant feeding patterns, basal PRL levels and postpartum ovula- tion. Clinical Endocrinology 17(4):323-332. Konner, M., and C. Worthman 1980 Nursing frequency, gonadal function, and birth spacing among Mung hunter-gatherers. Science 207:788-791. Lapham, R.J., and C.F. Westoff 1986 Demographic and Health Surveys: Population and health information for the late 1980s. Population Index 52:28-34. McNeilly, A.S., P.W. Howie, and M.J. Houston 1980 Relationship of feeding patterns, prolactin and resumption of ovulation postpartum. Pp. 102-116 in G.I. Zatuchni, M.H. Labbok, and J.J. Sciarra, eds., Research Frontiers in Fertility Regulation. New York: Harper & Row. Preston, S.H., ed. 1978 The Elects of Infant and Child Mortality on Fertility. New York: Academic Press. Sullivan, J., S. Rutstein, and G. Bicego 1994 Infant and Child Mortality. Demographic and Health Surveys Comparative Studies no. 15. Columbia, Md.: Macro International, Inc. Trussell J., and T. Guinnane 1993 Techniques of event history analysis. Pp. 181-205 in D. Reher and R. Schofield, eds., Old and New Methods in Historical Demography. Oxford, England: Clarendon Press. Trussell, J., and T. Richards 1985 Correcting for unmeasured heterogeneity in hazard models using the Heckman-Singer procedure. Pp. 242-276 in N. Tuma, ea., Sociological Methodology. San Francisco, Calif.: Jossey-Bass. Trussell, J., and G. Rodriguez 1990 Heterogeneity in demographic research. Pp. 111-132 in J. Adams, D. Lam, A. Hermalin, P. Smouse, eds., Convergent Issues in Genetics and Demography. Oxford, England: Oxford University Press.
