Demographic and Health Survey 1993

QUALITY OF THE DATA: NONSAMPLING ERRORS

Non-sampling errors arise in surveys from a variety of causes, such as the failure to locate and interview the selected household, mistakes in the way questions are asked, misunderstanding on the part of either the interviewer or respondent, coding errors, data entry errors, etc. Although it is impossible to avoid non-sampling errors entirely, great efforts were expended in the GDHS to keep them under control. These efforts included very careful questionnaire design; a pretest of survey instruments to guarantee their functionality; a four-week interviewers' training course; careful fieldwork supervision, including field visits by headquarters' personnel; a swift editing process prior to data entry; and, finally, the use of interactive data entry software to keep keying errors to a minimum. Nevertheless, there is still a need to investigate content errors such as misreporting of ages, ignorance of dates of birth or marriage, and other recall problems.

Table B. 1 of the survey Report shows the distribution of the household population by single years of age. There is rather substantial heaping on ages ending in 0 and 5 throughout the distribution, and it is somewhat more pronounced for females than for males. Errors are particularly notable in the age reporting of females around the borders of eligibility for the individual questionnaire, i.e., ages 15 and 49. Both this table and Table B.2 show an evident deficit of women at ages 15-17 years and a corresponding surplus at ages 12-14. At the other extreme, many women of ages 45-49 have been classified as being 50-54 years old.

Little difference can be seen between the age distribution of women recorded in the household schedule and those interviewed with the individual questionnaire, indicating that response rates vary little across the age of respondents (Table B.2).

Information on the completeness of reporting in connection with a set of important variables is provided in Table B.3. The month of birth was missing for nearly 18 percent of births reported in the birth history for the most recent 15-year period; however, both month and year were missing for less than I percent of these births. For the remaining variables, the percentage of cases with missing information is low.

According to Table B.4, information on month and year of birth is available for 78 percent of all children included in the birth history; the figure for living children is 81 percent, and for dead children is only 62 percent. However, reporting is much better for recent births. For children born since 1990, complete dates are given for over 93 percent; the figure reaches 95 percent for living children, but falls to 75 percent for dead children. Table B.4 also shows that the overall sex ratio at birth is 106; from year to year there are random fluctuations around this value without any indication of bias. The sex ratio for dead children (119) is much higher than for surviving children (104), indicating higher mortality among male children.

Another aspect of Table B.4 is the heavy concentration of births in 1989 (965) relative to 1990 (605). This may represent displacement by some interviewers so as to avoid having to ask the long sequence of questions in Section 4 of the interview. The calendar-year ratio at 1990 is 72, versus 150 at 1989. The effect is to produce a spike in the distribution of births at the fourth year prior to the survey (see Figure B.I). Though the displacement is serious, there is no indication of omission of births from the birth history. For this reason, a 5-year reference period for the measurement of current fertility was adopted in Chapter 3. A 3-year rate would have been artificially low.

Note: See detailed tables (Table B.1, Table B.2, Table B.3 and Table B.4) in APPENDIX B of the survey report.