Joseph LARMARANGE, Roselyne VALLO, Seydou YARO, Philippe MSELLATI, Nicolas MÉDA, Benoît FERRY.
In most countries in Sub-Saharan Africa, Demographic and Health Surveys (DHS) with HIV testing became the only measure of HIV prevalence in general population. Significant non response rate were often cited to explain differences between DHS results and estimations from sentinel surveillance in antenatal clinics. The objective of this presentation consists to predict with multivariate models the prevalence of non tested persons in order to estimate the effect of non response on national estimates.
Method / Issue
We used data from 9 DHS surveys (Burkina Faso 2003, Cameroon 2004, Ethiopia 2005, Ghana 2003, Kenya 2003, Lesotho 2004, Malawi 2004, Senegal 2005 and Tanzanie 2003) where HIV results could be linked with data from household and individual questionnaires. Logistic regression were calculated for each country, separately for men and women 15-49 years old, with common predictor variables : region, place of residence, age group, education, wealth index, marital status, work status, having radio or television, age at first sexual intercourse, recent sexual activities, using condom at last sexual intercourse, number of partners in last 12 months, smoking, STI in last 12 months, female and male circumcision and willing to care for relative with AIDS. For each group, adjusted prevalence was calculated by using observed prevalence for tested people and estimated prevalence for non tested people.
Results / Comments
The non response rates in these 9 studies vary from 7.9% to 34.2%. Estimated prevalence of non tested persons is usually higher than observed prevalence of tested persons : in 15 groups on 18, the ratio exceeds 1 (it vary from 0.820 to 2.424). Nevertheless, ratios of adjusted prevalence to observed prevalence remain relatively small (from 0.956 to 1.251). Except for men in Lesotho and women in Malawi, differences between adjusted and observed prevalence is less than 0.5 points. In both cases, number of tested persons was small (less than 3’000). No relation was found between non response rate and ratio of non tested to tested or ratio of adjusted prevalence to observed prevalence. Nevertheless, highest ratio of adjusted prevalence to observed prevalence were found for groups with smallest prevalence (<3%). But this effect is probably a consequence of a small statistical power.
If differences between adjusted and observed prevalence are more important than in a precedent survey conducted by Mishra et al. in 2006 on 5 DHS, the overall effect of non response bias on national HIV estimates tend to be small. Adjustments need to be interpreted with caution due to the limited information available to predict the prevalence of non tested people, in particular for people who did not answer the individual questionnaire and for whom only household questionnaire data were used.