Author: Frederic C. Shorter
Publication: Working Paper November 2003
This is the best census we have ever seen for the direct measurement of current fertility. First, there is good agreement with other sources, particularly the Hacettepe Surveys. That is, a TFR of 2.63 around 1999-2000 is consistent with the curve plotted in the earlier SIS study (SIS 1955: Tables 3-1 & 3-8); The new estimate might be a little higher than expected according to this test.
Next, I decided to check against another estimate of TFR made by an indirect method from the census itself: The indirect method using C10 and Brass infant/child mortality is not perfect, but it is a strong method in the Turkish case (SIS 1995: 103-104).. I tried it and got the same births for Istanbul and Ankara provinces as the direct count!! Dating would be 1999-2000 for direct estimate and a couple of years earlier for the C10 estimate.
I only have 6 province volumes and the national volume. So I made indirect estimates for all of them.
On the whole, this comparison makes me think that direct estimates of TFRs can be used for provinces as well as nationally -- unless something looks fishy for a particular province (e.g. Diyabakir?). Congratulations, but don't use direct-count TFRs from the 1990 or earlier censuses, because there are definite undercounts of births (26 % in 1990 overall according to the SIS 1995 study, app 2).
When I saw the ASFRs presented only as charts, I remembered Professor Ömer Gebizlioglu who convinced us back in the mid-1990s that graphics should be accompanied by the numbers. I don't think very many users will be successful when they try to add up the births for 12 months, adjust them for twins and second births, and shift the location six months due to the age of mother. Of course, they could look at SIS 1995 Appendix 2 and learn how to do it, plus UN Manual 10 (pp. 32-36, Table 8) for how to shift the 6 months (too much work!). I tried this and it works. Most likely, users of the 2000 census publications will just guess from the graphics.
I became curious about measuring reproduction in addition to fertility. That requires information about survival rates for women from birth to the mean age of childbearing. Probably you will get some information on mortality when you study the deaths reported in the 2000 census. Meanwhile, however, selection of a model life table based on the Brass estimates of infant and early childhood mortality will be good, or even better if there are problems with the death data of 2000.
Following this approach, I made an estimate of the Net Reproduction Rate during the period immediately before 2000. For Turkey it is 1.15 [East models from the red book; not the latest from Coale and Guo(1990)]. The procedure could be repeated for all the provinces, but I did not want to do that much work (lazy I guess). The data are there in the national volume for all the provinces (Brass estimates and TFRs). Probably a simple Dbase or Excel program could be written to look up and interpolate between levels for the lm values for females (where m is the mean age of childbearing). [The female lm table has to be set up by reference to the collection of East life tables. Since Coale and Guo are already incorporated in FIVFIV, I find it fastest to use the FIVFIV survival-rate option, paste the results into Excel, and derive the "l" values by computation. It is, unfortunately, hard to lay hands on the printed Coale and Guo tables.
Child -- Woman Ratios
Going further, I had a look at the 2000 census statistics on child-woman ratios, which include historical data. Actually, they refer to reproduction in a rough manner, not to fertility as the write-up in the volumes claims. I say "rough" because the time trends of change in the population structure and the long lags involved can upset the C-W index as a current measure at any date. Furthermore, any tendency for shifting persons out of the 0-4 age group into the 5-9 age group will depress the ratio, and we know that data error was serious in earlier times, but perhaps vanished by the time of the 2000 census. Nevertheless, for such a C-W look at trends in reproduction, I did some more investigation.
Here is a chart I made for Turkey that compares the trend of the C-W Ratio and TFR It shows the disparity between the two. (The fertility trend was taken from the SIS 1995 book, and the C-W trend from the 2000 census report.).
Recall that mortality was improving faster than fertility was falling in the early years. Consequently, the size of the 0-4 age group was being boosted more by improvements in survival than it was being reduced by declining fertility. If we want to understand the timing of the onset of fertility decline and its rate of change over time, the TFRs are more relevant than the C-W ratios. As to timing, fertility did start down nationally earlier than the C-W ratios seem to say. Nevertheless, there is some similarity in the trends -- they both trend downward eventually, which is saying something, but not much.
I plotted the Net Reproduction Rates for Turkey against the trend of the C-W ratios. The similarities are much stronger, as in theory they ought to be..
My conclusion is that the C-W ratios may be interpreted as approximate indicators of long-term changes in reproduction, but not in fertility by itself. Used in this way, they may provide useful information at the province level. They can not tell us when the fertility decline began, because mortality generally declined faster than fertility at the early stages in the Turkish case. However, the C-W ratios do show that reproduction was suppressed by mortality in the early years, and then gradually let go, after which the fertility decline brought reproduction down.
TFR, GRR, and NRR from past to present
I would prefer to follow trends in all three variables: TFR, GRR, and NRR. Then it becomes clear that changing mortality played a large role. We have the data for Turkey, but are not in such good luck for individual provinces.
What about province-level statistics over time? It would be challenging to track the history of fertility, infant mortality, and net reproduction by provinces (or regional aggregates), but it could be done. If you want to see an interesting map of the TFRs by provinces in 1960, look at Figure 13 in the book by the Panel on Turkey published in 1982 -- methodology explained pp. 49-52. There is also a province-level table and map of infant mortality rates, 1967, based on the 1970 census (pp. 81-85). Those are hints of what might be done for years between about 1960 and the 2000 census on a province level. We also did some of the history in the SIS 1995 book, focusing on regions and large cities.
Once given the TFRs and eo values by provinces at various dates, good approximations of NRR could be constructed for the same dates. It is a big job and would require access to the provincial volumes of the censuses for 1970 through 2000. I have not thought of doing this myself -- lazy, and the province volumes up to 1990 are still in my library archive in Istanbul. I have only a few of those volumes for 2000, although I might use tables of age structure and eo listed by province in the new 2000 national volume. That was not done for earlier censuses as far as I recall. Is it worthwhile to try to fill in these missing comparisons over time by provinces, or just get along with the regional information?
Overcount in 2000 ?
I did one more thing, which is interesting but not really decisive: compare 2000 with a projection from 1990. Actually, I know that this is not a very good way to check a census for over or under-enumeration. What about international in- and out-migration? And it could be 1990 that is wrong, not 2000. All the work SIS did to cancel the bogus extra questionnaires was no doubt the best way to try for accuracy in the total count.
Subtracting the projection from the 2000 count gives the results in the next figure. The projection assumptions were the mortality we used in the 1994 assessment projections, and an age-sex distribution of the population at midyear 1990 adjusted for age mis-reporting.
The extra population in 2000 relative to 1990 is 1,573 thousand persons for ages 10 and over -- a lot less than the 3 million that eliminating bogus returns cancelled! I did not make fresh assumptions about fertility between 1990 and 2000, so an estimate of the "extra" population under age 10 is not shown. Explanations of the relative overcount might be given as follows:
- Errors of coverage in 1990 or in 2000.
- There is a small appearance in the chart of over-enumeration in 2000 due to dating -- the projection is for mid-year and the 2000 census is for October. However, this discrepancy is only 0.6 percent (compare mid-period and census-date population in Table 3.1 of the 2000 census report).
- The up and down jiggles above age 25 might be due to differential age mis-reporting, particularly for females, in the two censuses. This is not, however, an issue of relative over-enumeration in 2000.
- Something happened to the population counts for "youth" in the ages between 10 and 25. Since it happened to both males and females, it would not be due to coverage differences for young men in the armed forces at the time of the census. Nothing suggests that they were over- or under-enumerated in either census.
- More likely, some of the young extra population, in 2000 relative to 1990, could be net immigration. This would include foreign-born persons, some of them tourists and some returned temporarily or permanently from places where they were born abroad. There might also have been "leaky" borders during all the disturbances that took place around Turkey during those 10 years, and other returns of citizens of Turkey. The foreign-born rose by 130,000 between the two census dates. An age and sex distribution of these individuals from both censuses would be helpful to look at this point.
After all is said and done, the 2000 census is likely to go down in history with the count that is published, following as it does the painstaking review of the original questionnaires.
Concerning the historical data in the 2000 census report, I think users should beware. On the other hand, data for 2000 by themselves are really good.
The historical section gives TFRs and other variables without evaluation from earlier censuses. (Remember all that work we did for the SIS 1995 publication?) For example, the TFRs in the census volume (Table 3.10) for dates prior to 2000 are the old undercounts of earlier censuses, not the updated and corrected estimates from SIS 1995: Chapter 3. There are similar errors in other variables in the history section of the report.
The infant and child mortality rates of Table 3.10 in the census report were not corrected for 1975; that was a census with certain reporting errors. In SIS 1995, we substituted the 1980 census instead of 1975 to make the retrospective Brass estimates relating to ca. 1973
In general, users can think of the 2000 data as the result of a good census, but should beware of using the historical time-series variables in the report, because they are from prior censuses without evaluation or adjustment for errors. That even applies to trends in the C-W ratios that would be different if age mis-reporting were taken into consideration. For history, it is advisable to use SIS 1995.
Panel on Turkey 1982. Trends in Fertility and Mortality in Turkey,
State Institute of Statistics 1995. The Population of Turkey,
State Institute of Statistics 2003. 2000 Census of Population: Social and Economic Characteristics of Population.
United Nations. 1983. Manual 10: Indirect Techniques for Demographic Estimation. New York: United Nations. Sales
Estimates of Crude Birth Rates from Place-of-Birth Data: 67 Provinces,
Annex: Plot points for some charts