Implications of Linear and Nonlinear Models of Family Planning Diffusion with Social Interactions Estimates and Interpretation

3. Data and Context of Empirical Estimation

Program efforts began in Kenya in the late 1950s in urban areas, and were expanded after President Kenyatta's government adopted a family planning policy in 1967. Support for the program was quite modest, however, until Moi became president in 1978. Moi's public support of efforts to reduce population growth rates may have acted as a "shock" to the system, legitimizing family planning and permitting much greater donor activity [Robinson 1992]. Family planning messages were regularly distributed over the radio and through posters and routine family planning talks in the clinics, and by the early 1990s family planning was widely accessible in local clinics in rural areas. The methods and messages are widely and frequently discussed in informal conversations, which thus may multiply program efforts. The overall picture from our household surveys as well as from qualitative data is of conversations about family planning that occur frequently and casually [Rutenberg and Watkins 1997, Watkins and Warriner 1999]. Because transportation is limited and expensive and telephones very few, frequent interaction is largely restricted to members of the local community.

For our empirical illustrations of how the estimation of program effects is affected by incorporating the social interaction, we use the Kenya Diffusion and Ideational Change Project (KDICP) of Watkins and colleagues and the Kenya Demographic and Health Survey (KDHS 1989 and 1993). Using two data sets permits us to evaluate whether our results hold across time and across data sets, and across different measures of social interaction based on cluster (village) aggregates and reported social networks. In addition, the KDHS data have the advantage of being similar to data that are widely available, so others can replicate our analysis in other settings. The KDICP data were collected in South Nyanza District (subsequently subdivided and renamed) in Nyanza Province, a predominantly rural area. To maximize comparability with the KDICP data, we restrict the KDHS to Nyanza Province. The characteristics of rural Nyanza Province are quite similar to those of South Nyanza District [Ayiemba 1986, Blount 1972, Ndisi 1974, Ocholla Ayayo 1976, Olenja 1991, Reynar 2000] and the KDICP and the KDHS for Nyanza give similar results for a variety of measures. (In preliminary analysis we further restricted the KDHS to rural South Nyanza, but this reduced the sample size substantially)

The KDHS was aimed at providing accurate measurement of basic demographic characteristics and to permit evaluation of the national population program, with few direct measures of social interactions. The KDICP was intended to examine the role of social interactions in fertility change and has many measures of social interactions. In particular, respondents were asked with how many people they had talked about family planning, followed by questions about the names and characteristics of a maximum of four of these individuals, including where the network partner lives and whether the network partner uses family planning [see Watkins and Warriner 1999 and for a detailed description of the available network data]. However, the KDICP data include only few measures of contact with the formal family planning program and a limited range of basic demographic information. Even though the overlap of comparable variables between these data sets is limited, it includes enough of the standard variables for analysis of contraceptive use to make the comparisons that we present in Section 4 of interest. Because the KDICP 1994 data were gathered only one year after the KDHS 1993 data, there are not likely to be large differences between these two data sets due to events associated with the passage of time.

In the analyses in this paper, the unit in which social interaction is assumed primarily to take place is the sample cluster in the KDHS and the village in the KDICP. The KDHS clusters correspond to larger geographical units than the KDICP villages, and thus are likely to approximate less well than villages the area of local social interactions. Thus, from the point of view of each respondent's local social interactions, a cluster average is likely to be a noisy indicator of the average for the respondent's (smaller) village. This means that the coefficient estimates of the cluster averages are likely to be biased towards zero in comparison with the coefficient estimates that would have been obtained were village averages available from the DHS; the bias is smaller in clusters that have villages with similar family planning use rates than clusters in which villages have different family planning use rates. In the absence of data on both, however, it is not possible to know how large these biases are.

By comparing the estimates based on village aggregates with those that use individual-level social network data in the KDICP, we are able to assess the extent to which these aggregate measures of social interaction lead to different results than the individual-level information on the interaction about family planning with members of the respondent's social network.

We restrict the analyses in this paper to currently married women respondents, whether or not their husband was present, to avoid complications related to absent husbands (27.4% of the husbands were not present at the time of the first wave of the KDICP, most of them because they were working elsewhere), which are unimportant for the purposes of this paper. Moreover, there is evidence from the KDICP data and elsewhere that in fact women may be the final arbiters of contraceptive use, at least in part because the availability of modern methods permits them to use family planning without their husband's agreement [Bawah et al. 1999, Biddlecom and Fapohunda 1998, Pictet and Ouedraogo 1999, Reynar 1998, Reynar 2000, Watkins, Rutenberg and Wilkinson 1997].

Table 1 gives basic sample characteristics and variable means for the variables used in this paper. Panel A gives the total number of respondents, the number of clusters/villages, and the average number of respondents in each cluster/village. Panel B summarizes the distributions of the main variables in the samples used for our estimates.

Dependent variable: Our dependent variable is whether the respondent has ever used family planning. We think that ever-use of contraception is a better indicator of innovative behavior than is current use because discontinuation rates are high and modern contraception is frequently used temporarily for spacing. The average level of this dependent variable is somewhat higher for the 1993 KDHS than for the KDICP, which may be due to the fact that the former is based on a larger geographic area (Nyanza Province instead of South Nyanza District) [Note 6].

Right-side variables: The only measure of the population program that was the same in both surveys concerned hearing a family planning message on the radio. We measure program effect through the proportion of women (other than the respondent) in a cluster or village who have heard a family planning message on the radio [Note 7]. Because this program representation basically is an aggregate variable in the statistical sense of having almost the same representation of a right-side variable for every respondent in a given village or cluster, we use the Huber-White correction for our standard errors. The proportion reporting that they heard a family planning message on the radio is notably larger for the KDICP than for the KDHS. This may be due to the fact that the KDHS made reference to specific family planning programs, whereas the KDICP questionnaire contained only one general question about exposure to family planning messages on the radio. We use two measures of interaction. For both the KDHS and the KDICP data, we use the cluster (village) aggregate of the dependent variable for individuals, excluding the respondent -- e.g. the proportion of others in the cluster (village) that reported having ever used family planning. For the KDICP data, we additionally use the proportion of network partners who have ever used family planning.


Implications of Linear and Nonlinear Models of Family Planning Diffusion with Social Interactions Estimates and Interpretation

Empirical Assessments of Social Networks, Fertility and Family Planning Programs: Nonlinearities and their Implications
Hans-Peter Kohler, Jere R. Behrman, Susan Cotts Watkins
© 2000 Max-Planck-Gesellschaft ISSN 1435-9871