3.1. Data collection

Since 1999, demographic and anthropological data have been collected from four villages in the lowland areas. Demographic census surveys were undertaken in 1999 and 2003, during which birth histories were collected from all reproductive aged women (15–50 years) using calendars that recorded the monthly timing of births and deaths over the preceding 6 years. In 2003, a sample of 700 children (<16 years) was included in an anthropometric survey across the four villages. Measures of height and weight were performed based on the guidelines set out by Lohman, Roche, and Martorell (1988). The median age of children measured during the survey was 5 years and 7 months. The sample included an equal number of both sexes (50.4% females and 49.6% males).

Additionally, a time allocation study was undertaken for a subsample of 58 households across the villages. The weekly activity budgets of 58 women and 200 children aged 0–15 years were obtained using spot observation techniques (Borgerhoff Mulder & Caro, 1985, Hames, 1992). This was considered to be the least intrusive method of collecting time budget data, which would generate the largest sample sizes. Each household was visited five times a day between sunrise and sunset (7 a.m. and 6:30 p.m.) for 1 to 2 weeks, and spot observations for the presence and activity of each individual within the household were recorded. If the individual was not present in the compound, then activities outside the household were recorded, according to information provided by neighbors or other household members present.

3.2. Mortality analyses

The demographic data were analyzed using event history analysis, a powerful tool for dealing with censoring biases and isolating the precise effects of time-varying events, for example, the introduction of village level tap stands (Allison, 1984, Gibson & Mace, 2002a, Gibson & Mace, 2002b). Multivariate logistic regression was used to control for sex, season, and year of birth; mother's age, parity, education, and household socioeconomic status (cattle herd size) when assessing the effect of kin on the probabilities of child death over time. The dependent variable was the monthly probability of death, which was modeled up to the third year of life, because the greatest risk of child death for this population occurs during the first few years of life (Gibson, 2002). To control for any hierarchical structures in the data relating to the household (genetic and/or environmental effects within family), only the last two births per woman were included within the mortality analyses. A total of 1652 women contributed 2746 births to the final analyses. SAS software Version 8.2 was used to perform the statistical analyses.

Postmarital residence of the household at the survey date was recorded as the dichotomous covariate “locality.” The reference category was “patrilocal residence.” The survival status of each category of grandparent at the survey date was entered into the model as a dichotomous variable. Additionally, analyses were performed to identify the effect of constellation or composition of grandparental support, which some researchers have indicated may be an important predictor of child survival (Voland & Beise, 2002). In this case, three dummy variables were entered into the model (both grandparents alive, only one grandparent alive, and both grandparents dead). The maximum amount of time that had elapsed between any reproductive event included in the analyses and the survey date was 5 years.

3.3. Growth analyses

In the analyses of child nutritional status, multivariate general linear models (GLM) were used to assess the partial effects of child's age, square of child's age, sex, maternal age, parity, family size, and wealth (household cattle herd size), as well as kin effects on child height and weight for height. A household identifier was included in the analyses to control for hierarchical structures of the data relating to family characteristics. A total of 652 children (<15 years) was included in the final analyses. SPSS Version 12.0 was used to perform the statistical analyses. Z score values (standard deviations), reflecting the width of the distribution around the mean of the NCHS/WHO international reference population (WHO, 1983), were calculated using the Epi-Info 3.2.2 software. By convention, children −2Z scores below the median of the reference group, based on U.S. growth data collated in 1978, are considered to be malnourished (WHO, 1983).

3.4. Time budget analyses

The observational data were analyzed using logistic regression models to identify the activities undertaken by non-nuclear relatives within the household. For each spot observation, the likelihood of a behavior taking place was estimated as a function of the various covariates. The analysis was carried out separately for each category of kin. These relatives included grandparents and both husband's and his wife's siblings and their families, but excluded the nuclear family such as husband, children, and cowives. The observed behaviors were grouped into six categories of activity: agriculture (working in fields, winnowing, or threshing), light domestic (light workload tasks within the compound, e.g., food preparation, cooking, and cleaning), heavy domestic (heavy workload tasks, e.g., grinding maize, collecting firewood, or water), social activities (preparing/drinking coffee, talking with friends/neighbors), resting (lying down, listening to the radio), eating, and child contact (any child contact, e.g., holding, breast feeding, and washing a child). A household identifier was included in the analyses to control for hierarchical structures of the data relating to household characteristics because not all households contributed the same number of observations. The samples of individuals observed in 58 households were 58 mothers, 22 grandmothers, and 75 other kin; 44 were patrilocal households and 14 were matrilocal households. SPSS Version 12.0 was used to perform all the statistical analyses.

4.1. Mortality

Overall, 14.6% of the children in the community died before the age of five, including 13.7% before the age of three; this is comparable with regional rates (in the 2000 national census, 19.4% of children in Oromiya region died before the age of five; Central Statistics Authority & Macro, 2001). In our sample, the first month of life represented the most risky period of early childhood; 11.5% of all new births died during this neonatal period.

The results of the multivariate discrete-time event-history model, which was used to analyze the predictors of early child death, are presented in Table 1. The results are presented in the form of exponentiated coefficients for the log odds, known as odds ratios (OR). OR represent the relative risk of other groups in relation to the baseline group, or reference category (which is indicated with an OR of 1). An OR value greater (or lesser) than unity indicates that the relative odds of dying is greater (or lesser) compared with that of the reference group, taking account of the other factors included in the model.

Table 1. Relative risk (OR) and 95% confidence intervals (CI) of child death under 3 years of age for both and each sex (a) by locality and grandparent status and (b) by combination of grandparents alive

		Male death			Female death			Any child death
		n	OR	CI	n	OR	CI	n	OR	CI
	(a)
	Locality
	Matrilocal	331	0.733	0.48–1.11	349	0.492⁎⁎	0.30–0.81	650	0.617⁎⁎	0.45–0.85
	Patrilocal	1144	1.000		1088	1.000		2234	1.000
	Maternal grandmother
	Alive	1077	0.714	0.51–0.99	1065	0.940	0.64–1.37	2143	0.802	0.62–1.03
	Dead	397	1.000		370	1.000		768	1.000
	Maternal grandfather
	Alive	858	0.889	0.65–1.22	833	0.763	0.54–1.08	1691	0.847	0.67–1.07
	Dead	615	1.000		602	1.000		1219	1.000
	Paternal grandmother
	Alive	563	1.096	0.78–1.54	522	0.674⁎	0.45–1.00	1086	0.880	0.68–1.15
	Dead	624	1.000		622	1.000		1247	1.000
	Paternal grandfather
	Alive	430	0.841	0.57–1.23	423	1.037	0.69–1.57	854	0.923	0.70–1.22
	Dead	643	1.000		621	1.000		1265	1.000
	(b)
	Both grandmothers alive	415	0.801	0.51–1.26	392	0.679	0.68–1.16	808	0.748⁎	0.52–1.05
	One grandmother alive	516	0.638⁎	0.42–0.96	529	0.952	0.61–1.48	1045	0.777⁎	0.58–1.05
	Both grandmothers dead	177	1.00		169	1.00		397	1.00
	Both grandfathers alive	293	0.922	0.56–1.52	280	0.944	0.57–1.57	573	0.969	0.69–1.36
	One grandfather alive	412	0.953	0.60–1.37	431	0.630	0.4–1.00	846	0.787	0.59–1.06
	Both grandfathers dead	294	1.00		286	1.00		581	1.00

Other variables controlled are sex, season and year of birth, mother's age at birth, parity, mother's education, cattle herd size, and household identifier. [Bold type and asterisks indicate where the child's relative risk of mortality was significantly different from the reference category (OR=1.00)].

⁎ p<.05. ⁎⁎ p<.01.

Table 1 indicates that the proximity of kin, and specifically maternal kin, contributed to improved rates of child survival. Children born into matrilocal households had significantly lower risk of dying in early childhood than do those born into patrilocal households (OR=0.617, CI=0.5–0.8, p=.002). While both sexes benefited from proximity to maternal kin, when analyses were performed independently on the sexes, the relative advantage of matrilocal residence pattern was statistically significant only among females (OR=0.492, CI=0.3–0.8 p=.008).

In general, grandparents and especially maternal grandmothers, exerted a positive influence on child survival. Those children whose maternal grandmother's were alive had a lower probability of dying in early childhood (OR=0.802, CI=0.6–1.0, p=.09). Male grandchildren were more likely to survive when the maternal grandmother was alive (OR=0.714, CI=0.5–0.9, p=.06); however, the reverse was true for paternal grandmothers. An OR greater than unity, although not statistically significant, is notable, because it suggests that paternal grandmothers may have elevated a male child's risk of death. Other studies have attributed similar findings to the high cost of having many male offspring under certain socioecological conditions, such as brideprice paying societies and those with primogeniture (Jamison et al., 2002, Mace, 1996).

While the individual effect of either grandmother falls just short of statistical significance, the second model described in Table 1 shows that children that had lost both grandmothers had a significantly higher risk of dying before the age of three, compared with those with either one or two grandmothers alive. In both models, grandfather status did not appear to influence child survival. The results support previous findings that grandfathers' survival have a lesser role in determining child well-being (Sear et al., 2000, Sear et al., 2002).

4.2. Growth

Overall, 49.6% of children in the villages were seriously stunted for their age [height for age (HAZ) Z score <−2.00; i.e., more than two standard deviations below the median of the WHO international reference population], and 10.3% were seriously undernourished for their height [or wasted; weight for height (WHZ) Z score <−2.00]. These figures are comparable with levels of childhood malnutrition recorded in the 2000 Demographic and Health Survey for the entire Oromiya region (47.2% stunting and 10.4% wasting; Central Statistics Authority & Macro, 2001).

The results of the multivariate GLM, used to analyze the determinants of child nutritional status, are presented in Table 2. The results are presented as estimated marginal mean values of height (cm) and weight for height (kg) according to postmarital residence and grandparent survival, controlling for age, family size, birth order, and wealth.

Table 2. Estimated marginal means of (a) height in centimeters and (b) weight for height in kilograms of children (±S.E.) for both and each sex by locality and grandparent status

	(a) Height	Males		Females		All children
		n	±S.E. (cm)	n	±S.E. (cm)	n	±S.E. (cm)
	Locality
	Matrilocal	99	101.31±0.70	108	99.55±0.57⁎⁎	207	99.73±0.62⁎⁎
	Patrilocal	231	102.57±0.46	232	102.18±0.39	463	101.52±0.43
	Maternal grandmother
	Alive	204	102.79±0.80	208	101.75±0.58⁎	412	101.24±0.48⁎
	Dead	120	101.29±1.00	121	100.84±0.83	241	99.86±0.62
	Maternal grandfather
	Alive	150	102.42±0.89	143	101.06±0.77	293	101.57±0.5
	Dead	174	101.76±0.87	186	101.57±0.67	360	101.68±0.55
	Paternal grandmother
	Alive	133	102.39±0.78⁎⁎	138	101.51±0.71	271	101.82±0.53
	Dead	191	101.74±0.96	191	101.08±0.68	382	101.42±0.58
	Paternal grandfather
	Alive	85	102.67±1.16	98	101.12±0.86	183	101.79±0.71
	Dead	239	101.58±0.56	231	101.47±0.57	470	101.48±0.42
	(b) Weight for height	Males		Females		All children
		n	±S.E. (kg)	n	±S.E. (kg)	n	±S.E. (kg)
	Locality
	Matrilocal	99	16.68±0.20	108	16.42±0.20	207	16.54±0.14
	Patrilocal	231	17.01±0.13	232	16.58±0.13	463	16.80±0.09
	Maternal grandmother
	Alive	204	17.24±0.23	208	16.77±0.19	412	16.95±0.15
	Dead	120	16.63±0.29	121	16.49±0.28	241	16.59±0.20
	Maternal grandfather
	Alive	150	17.01±0.25	143	16.89±0.26	293	16.90±0.18
	Dead	174	16.90±0.25	186	16.37±0.23	360	16.66±0.17
	Paternal grandmother
	Alive	133	16.89±0.22	138	16.95±0.24	271	16.90±0.17
	Dead	191	17.03±0.27	191	16.30±0.23	382	16.65±0.18
	Paternal grandfather
	Alive	85	17.31±0.33	98	16.53±0.29	183	16.88±0.22
	Dead	239	16.65±0.17	231	16.73±0.19	470	16.69±0.13

Other variables controlled are child's age, age squared, sex, mother's age at birth, parity, mother's education, cattle herd size, family size, and household identifier. Age adjusted for mean age: 5 years and 10 months. (Bold type and asterisks indicate where child's nutritional status was significantly different from the reference category, patrilocal or dead).

⁎ p<.05. ⁎⁎ p<.01.

Table 2 indicates that children living in matrilocal households, thus with access to matrilineal and patrilineal relatives, were statistically more likely to be shorter than children born into patrilocal households (β=−6.08±2.49, t=−2.42, p=.01). On average, children were 2 cm shorter in matrilocal households. This was especially true among girls (β=−6.138±2.28, t=−2.67, p=.008). Although not statistically significant, in a similar analysis of weight for height, the direction of the effect indicated that children in matrilocal households had lower weight for height.

The general impact of grandparents was positive for both child height and weight; however, grandfathers were less important than grandmothers were. Maternal grandmothers were a significant positive predictor of child height (β=3.323±1.76, t=−1.98, p=.048), particularly among girls (β=3.361±1.543, t=−2.178, p=.03), and also had a positive (although nonsignificant) effect on their weight. Paternal grandmothers were associated with improved height only among male children (β=11.668±2.82, t=−4.131, p<.001) but had a negative (although nonsignificant effect) on their weight.

4.3. Time budgets

Fig. 1 illustrates the categories of extended family relatives who were present in the household during the time allocation survey. Overall, relatives were equally likely to be observed in patrilocal and matrilocal households (17.2% vs. 15.5% of spot observations), and grandmothers was the category of relative most likely to be present in the household (13.3% of spot observations). However chi-squared analyses revealed an association between postmarital residence and category of kin in the household (Pearson χ2=158.15, df=3, p<.01). Patrilineal kin were found to be present in the patrilocal households in 12.2% of observations, but in only 4.5% of observations in matrilocal households. Even paternal grandmothers almost never visited their son and daughter-in-law's household if that household was matrilocal (see Fig. 1). The absence of paternal kin in matrilocal households seems surprising, given the fact that matrilocal postmarital residence among the Arsi Oromo ensures close access to both sets of kin.

Fig. 1. Presence during spot sampling of grandmothers and other kin by lineage and postmarital residence in 58 households (% of spot observations, n=3033).

Matrilineal kin, particularly maternal grandmothers, were more likely to be present in matrilocal households (10.8% of spot observations); however, they also visited patrilocal households (7.2%). Thus, grandmothers were more likely to be spotted in their daughter's household rather than in their son's household, even if they had to travel to another village to visit their daughter (see Fig. 1).

Table 3 describes the activities undertaken by mothers and visiting grandmothers and other relatives present in 58 households. It includes the parameter estimated values from the logistic regression time budget analyses, which identify the statistical likelihood that one set of kin was more likely to take part in each activity, taking account of household level variation in the number of spot observations. A positive (or negative) coefficient represents an increased (or decreased) relative risk of the maternal/matrilocal kin being observed performing an activity compared with the paternal/patrilocal reference group. Percentages refer to the percentages of observations that a category of relative allocated to a particular category.

Table 3. Time allocation for family members observed in the household (% of spot observations) and β parameter estimates (±S.E.) from logistic regression models identifying the likelihood of undertaking certain household activities

	Activity	Mother		Grandmother		Other kin
		Matrilocal	Patrilocal	Maternal	Paternal	Maternal	Paternal
	Agriculture (%)	7.0	6.6	3.3	17.5	10.2	22.4
	β±S.E.	0.07 (0.17)		−1.89⁎(0.78)		0.14 (0.54)
	Light domestic (%)	27.9	29.1	29.5	20	8.2	15.8
	β±S.E.	−0.08 (0.09)		0.51 (0.40)		0.74 (0.56)
	Heavy domestic (%)	16.6	14.6	14.8	2.5	4.1	10.4
	β±S.E.	0.16 (0.11)		1.91⁎(0.82)		1.12 (0.76)
	Social activity (%)	12.8	12.5	16.4	17.5	22.4	14.5
	β±S.E.	0.01 (0.13)		−0.07 (0.46)		−0.33 (0.40)
	Resting (%)	5.8	8.4	23	23.8	38.8	21.6
	β±S.E.	−0.39⁎(0.1)		0.19 (0.44)		−0.55 (0.34)
	Eating (%)	2.2	2.2	4.9	3.8	14.3	6.2
	β±S.E.	−0.08 (0.29)		0.27 (0.84)		−0.69 (0.50)
	Child contact (%)	22.4	19.7	3.3	8.8	2.0	2.1
	β±S.E.	0.23⁎(0.1)		−1.07 (0.82)		0.04 (1.14)
	Observations (n)	789	2244	61	80	49	241
	Households (n)	14	44	4	18	14	34

Other variables controlled are household identifier. (Bold type and asterisks indicate where matrilocal/maternal relative activity was significantly different from the reference category, patrilocal/paternal activity.

⁎ p<.05.

Over 50% of a mother's time was spent in work activities (heavy domestic tasks, such as collecting fire wood and water; light domestic tasks; and agricultural activities); 20% in childcare-related activities (e.g., breast feeding, washing children, and taking a child to the clinic); 12% in social activities (e.g., drinking coffee with visitors); and 8% resting. There were no large differences between patrilocal and matrilocal households in how mothers divided their time between work, social activities, or resting. However, mothers living in matrilocal households were statistically more likely to spend time in childcare-related activities (β=0.229±0.102, OR=1.205, p=.025) and less time resting (β=−0.392±0.171, OR=0.52, p=.02) than did mothers in patrilocal households.

When grandmothers and other kin visited, they took part in a wide variety of activities; the majority of their time in the household was spent in domestic tasks, social activities, and resting. However, clear differences did exist between the activities undertaken by grandmothers visiting their daughters and their daughters-in-law. Maternal grandmothers were more likely to be employed in heavy domestic tasks, for example, collecting fuel wood and grinding maize (14.7% of observations) than were paternal grandmothers (2.5%; β=1.91±0.82, OR=6.75, p=.019). Paternal grandmothers were significantly more likely to spend their visiting time in agriculture-related tasks, for example, threshing (17.5%), than were maternal grandmothers (3.3%; β=−1.89±0.78, OR=0.151, p=.016). Other paternal relatives also appeared to take part in productive, work-related activities (48.6% of observations), particularly agriculture (22.4%); however, this does not necessarily imply altruistic helping behavior. Any agricultural work may have direct payoffs for paternal kin, since they can claim a percentage of their paternal relative's harvest or may have been farming fields together. Other maternal relatives were significantly more likely to be observed resting, socializing, or eating during their visits (75.5% of observations; β=1.03±0.37, OR=1.643, p=.005), possibly providing social and psychological support for female relatives (Euler, Hoier, & Rohde, 2001).