Data and Methods Macro-Micro Interactions in Birth Rate Models

4. Main Effects of Aggregate Education on Birth Rates

According to models where only age, parity and duration since previous birth were included along with individual education, a complete primary education does not reduce first-birth rates significantly at the 0.05 level. However, the effect on higher-order birth rates was found to be significant. The point estimates (log odds) were -0.17 and -0.19, respectively (Model 1, Table 1). This is for the period 1990-1994 and is quite consistent with the difference of 0.5 in TFR between uneducated women and those with a primary education that is shown in univariate tabulations (covering the last few years before the survey) from the 1994 ZDHS [Central Statistical Office 1995] [Note 9].

The effects of secondary education were significant for both first- and higher-order birth rates. If short and long secondary education were combined, the point estimates were about -0.70 and -0.45, respectively (not shown). This accords reasonably well with the TFR difference of 1.9 between uneducated women and those with at least some secondary education that was reported from the 1994 ZDHS.

(Table 1 about here)

When average length of education was included, the effects of individual educational level were reduced (Model 3). The effects of aggregate education were significant in these simple models. In order to get an impression of the relative importance of aggregate and individual effects, the consequences of changing the educational distribution for women aged 15-50 were predicted. Such distributions are shown in Table 2 for some groups of districts. For example, it is easy to calculate that, if the educational distribution is changed from that in the bottom-10 districts to that in the top-10 districts, the individual contributions will be -0.32 for first births and -0.14 for higher-order births [Note 10], whereas the aggregate contribution associated with this increase of about 2.5 years in average education will be -0.16 and -0.21, respectively. It should also be noted that the sums of individual and aggregate contributions are markedly higher than the corresponding predictions of -0.35 and -0.22 based on models with only individual education included. In other words, when it is taken into account that educated women tend to live in areas with many other educated women, the effect of individual education is reduced, but this is more than compensated for by the aggregate effect. This is the same as saying that, when only individual education is included in the model, part of the aggregate effect is also captured, but not the entire [Note 11].

(Table 2 about here)

It should also be noted from Model 3 that first-birth rates were more clearly influenced by the depth of education than its breadth (the latter effect was almost significant at the 0.10 level). For higher-order births, however, both effects were significant.

The estimates presented above grossly overstate the causal impact on fertility of individual investments in education. When it was taken into account that educated women tend to have a Christian background, and that their education, as well as that of other women in the district, must be determined partly by the urban/rural character of the enumeration area they currently live in, effects of education were markedly reduced. Individual effects were left almost unchanged for first births and slightly reduced for higher-order births, whereas aggregate effects no longer were significant (Model 4). Also the point estimates of the latter were very small. An increase of, say, 2.5 years in average education would increase first-birth rates by 0.08 and decrease higher-order birth rates by 0.02, which is completely negligible compared to individual effects. In fact, aggregate education might just as well have been left out of the models. When only urban/rural and religion were included, and not aggregate education, one was left with the same impression of effects of investments in education (Model 2). (As further illustration of the importance of these control variables, the impact of the above-mentioned change in the educational distribution would be about 1/3 weaker according to Model 4 than according to the simpler Model 1, which corresponds to univariate TFR tabulations).

Moreover, experiments with various specifications clearly showed that the results depend markedly on the kind of urban/rural variable that is used. Because a woman's individual education may have been determined more by the place of residence where she grew up than by the place where she currently lives, it would seem quite reasonable to include the urban/rural nature of the former. However, adding this variable to the simplest model had much less impact on individual education effects (not shown) than adding an urban/rural variable referring to the current place of residence. Moreover, aggregate effects remained almost unchanged (Model 5), because this urban/rural variable is less closely linked with the aggregate education of the district in which she currently lives.

The aggregate education is a district-level variable. Living in an urban enumeration area does not necessarily mean that the district is generally very urbanized. For that reason, one might alternatively have preferred to include the proportion urban residents in the district. This would have reduced aggregate effects of education very markedly (Model 6), although not quite as much as the inclusion of an urban/rural variable referring to the enumeration area (Model 4).

As yet another alternative, the character of the current place of residence was grouped into more categories by considering the rural enumeration areas' distance to cities. This gave only slightly smaller effects of education than in Model 4 (not shown).

Because of the close link between urbanization and aggregate education (Pearson correlation coefficient of about 0.8), one might suspect that it would be difficult to identify separate effects. Some models were therefore estimated separately for rural areas. No significant effects of aggregate education appeared in these models either, except that depth of education was almost significant at the 0.10 level for higher-order births (not shown).

To summarize, the results provide very little support for the idea that there may be an extra stimulus to fertility decline through effects of aggregate education (i.e. at a given individual educational level). The lack of access to control variables should not weaken this conclusion much. If the general educational level is strongly determined by structural factors that have a clear stimulating effect on fertility, its true effect, net of urbanization, would be negative, but this is not very plausible. Such background factors are more likely to have a fertility-inhibiting effect, if any. A negative effect of aggregate education only seems reasonable to the extent that investments in education in a region contribute to stimulate the urbanization of that region, which must be a very slow process.

The latter comment may need an elaboration: It was explained above that the urban/rural character of the community is likely to be a determinant of aggregate education. This means that the estimated effect of aggregate education, when both variables are included, can be interpreted as the total effect of that variable (assuming also that none of the other regressors are causally intermediate). If it had been more appropriate to consider urbanization a consequence of education, the interpretation would have been different. In that case, the aggregate education effect estimate would reflect the direct effect of that variable, while there would also be an indirect one operating through urbanization. The size of the latter would be determined by the effect of aggregate education on urbanization (not estimated) and the effect of urbanization on fertility. The sum of these contributions, i.e. the total effect, could, of course, more easily have been estimated by leaving the urban/rural variable out. This reverse causality may hold some truth, without being very dominant. It is certainly possible that investments in education may stimulate urbanization, although it is not a result that is likely to be seen very soon after the school system is expanded. In other words, the total effect of aggregate education may, in the long run, be slightly more negative than suggested by the aggregate education effect estimate from a model where both variables are included.

When such very weak effects are estimated, there is no need to relax the independence assumptions by estimating MLwiN models. These would have given larger standard errors of the estimates and even less reason to reject an hypothesis of no aggregate effect of education.


Data and Methods Macro-Micro Interactions in Birth Rate Models

A Search for Aggregate-Level Effects of Education on Fertility, Using Data from Zimbabwe
Øystein Kravdal
© 2000 Max-Planck-Gesellschaft ISSN 1435-9871