Back to contents index

Issues and unexpected findings in Census ’96, and possible explanations for them

The results from the final PES-adjusted count differ in important ways from what past demographic modelling would have expected. In this chapter, we examine these differences, and give possible explanations for them.

Table 14 gives the final PES-adjusted estimate of the population size of South Africa broken down by gender, urban or non-urban place of enumeration and population group. Census ’96 (the count) is then compared with the previous 1996 mid-year estimates based on projections from the demographic model used to adjust the 1991 population census (the model).

* All numbers given in this report are rounded to whole numbers.
        The totals may therefore differ slightly.

**    The percentages are rounded to the first decimal place, therefore they may not always add up to exactly 100.

***  This category includes approximately 8 000 people, mainly found in Northern Cape, classifying themselves as Griquas.

The table shows the following chief differences between the model and the count:

The model predicted that there would be a population of 42,1 million people in South Africa in the middle of 1996, but the PES-adjusted census count is 40,6 million.

According to the model, 49,5% of the population should be male, and 50,5% female, but the count found that 48,1% of the population was male, and 51,9% female.

.In the model, there was a lower proportion of urban compared to non-urban dwellers in the country in 1996 (48,3% urban as against 51,7% non-urban). The count found a higher proportion of urban dwellers (53,7% urban, as against 46,3% non-urban).

The model predicted that 12,7% of the population would be white, while the count found that 10,9% were white.

In addition to what is shown in the table, when comparing the model and the count there is a smaller than expected percentage of children aged 0-4 years, and a lower male to female sex ratio. The number of people designating themselves as foreigners is also lower than expected.

It would have been extremely useful to consider our final PES-adjusted estimates from Census ’96 against data from existing administrative registers such as records of births, deaths and migration. However, previous studies (Sadie, 1988) have found that such records are characterised by a high degree of under-reporting. In particular, Sadie (n.d.) further noted that, for Africans, ‘we have, in fact, no vital statistics at all, and consequently, have to make do with incomplete census data only’ (p. 4).

School attendance records, which were used by the Bureau of Market Research as part of their validation process of the 1991 census (Nel, Loubser and Van Wyk, 1993), are also incomplete and unreliable, particularly in the former TBVC states and self-governing territories. This is because enrolment numbers have been conflated in a few provinces, perhaps to increase subsidies to be claimed. The proportion of children in any given age group actually attending school is also unknown. It therefore remains to consider demographic modelling.

Data collected for any population census conducted after 1970 but prior to 1996 was regarded with suspicion, for the following reasons:

Censuses conducted since 1970 were increasingly incomplete, since they excluded the former TBVC states from the time when each of them was declared an independent state by the former government. Separate censuses were conducted in the TBVC states, but no attempt was made to use these counts to estimate the size of the total South African population. We shall see below that projections from 1970 were applied instead.

Within South Africa excluding TBVC, certain parts of the country where mainly Africans live were not clearly demarcated into EAs. This was very evident in 1991, where rural areas in the former self-governing territories (KwaZulu, Gazankulu, Lebowa, KaNgwane, KwaNdebele and Qwaqwa) were not demarcated. Instead, ‘sweeps’ were done in which groups of enumerators were entrusted with counting larger areas as best they could, without the controls of demarcated boundaries or lists.

In 1991 in certain other parts of the country, particularly informal settlements, aerial photographs were used to count the number of dwellings; and then the average household size per dwelling was obtained from small-scale surveys on the ground, and used to estimate the size of the population.

In 1991 and in the earlier censuses, the then Central Statistical Service (CSS), as a result of the difficulties described above, used demographic modelling ‘to establish a set of population numbers by age, sex and demographic group, as a base independently of the 1991 Census’ (Sadie, 1992, p. 1 – emphasis added).

For whites, Indians and coloureds, a comparison of the actual count in the 1991 census with his modelled results led Sadie to conclude that the undercount had been 10,8% for whites, 12,4% for Indians and 10,8% for coloureds. The upward adjustment of whites to fit the model is important to remember, when one considers the extent to which the white population measured by the PES-adjusted count in 1996 may be too small.

Sadie noted that, for Africans, the 1980 and 1985 censuses ‘for the most part have been conspicuous for their deficiencies’ (Sadie, 1992, p. 6). For this group he used the 1970 census as the baseline, since that census included the TBVC states. Taking the model as correct, this yielded an overall undercount for Africans in the RSA of 16,8% (Sadie, 1992, p. 50; Nel, Loubser and Van Wyk, 1993, p. 5).

This undercount among Africans was discussed in more detail in the validation study for the RSA conducted by the Bureau of Market Research (Nel, Loubser and Van Wyk, 1993). It used a post-enumeration survey in some areas, and also drew comparisons between census counts of children aged 8-12 years and primary school enrollments. Regarding the areas enumerated through ‘sweeps’ on the ground in 1991 as discussed previously, it found that ‘the CSS missed roughly 24% of the black (i.e. African) population’ (p. 8). Regarding the areas handled for the CSS by aerial photographs and surveys, it found that, among Africans, ‘the count was very complete or possibly over-counted’ (p. 101) but had anomalies in the sex ratios (p. 101).

Table 15 shows the actual and the adjusted counts by demographic modelling in 1991. The population in the former TBVC states was simply arrived at by subtraction of the various modelled amounts from the modelled total, and is therefore excluded from the table.

Source: Nel, Loubser and Van Wyk (1993).

* Excluding the former TBVC states.

The table shows that the PES-adjusted counts differ significantly from those modelled, and used to adjust the 1991 census. For example, only 8,8 million African males were counted, but the count was adjusted upwards by 2,1 million by the model to reach 10,9 million. The total white population was adjusted upwards by 500 000 from 4,5 million to reach 5,1 million. The model needs to be examined in more detail to understand these appreciable differences.

Altogether, Table 15 shows that, in 1991, the part of the population that was counted was raised by some 4,7 million people. This needs to be borne in mind when considering whether the forward projection of the model to 1996 is too high at 42,1 million, or whether the PES adjusted count is too low at 40,6 million.

This measure reflects the number of men per 100 women. In Table 14 it was pointed out that the sex ratio in 1996 was lower than the one that was modelled in 1991.

But the 1991 model-based adjustments to the count in this regard were rather large. For example, in the mainly urban informal areas in which aerial photography was used backed by information from small samples (HSRC-counted area), a sex ratio of 93,6 men per 100 women was found. Altogether, 1 209 000 more women than men were counted in the HSRC-counted areas. The conclusion drawn at the time was as follows: ‘Taking into account that 81,5% of the black population counted by the HSRC was resident in urban areas and in metropolitan areas in particular, where men have traditionally been in the majority, such a low sex ratio seems unlikely’ (Nel, Loubser and Van Wyk, 1993, p. 10, emphasis added). Consequently, a large proportion of women (approximately 250 000) were reclassified as men (making a shift of 500 000)!

Overall in 1991, an upward adjustment factor of 1,24 was applied to African males, and of 1,17 to African females. This of course increased the sex ratio implied by the modelled totals, and should be borne in mind when considering whether those implied by the PES-adjusted Census ’96 totals are lower than those expected demographically.

Among Africans in the 1970 census, when using the unadjusted count and excluding the foreign born, a sex ratio of 91,5 was found. Although the respective results are not strictly comparable, this is interestingly close to the sex ratio of 89,6 found among Africans in 1996 after excluding the foreign born.

Shell (1998) surmises that the low proportion of males to females in South Africa reflected in these measurements may be due to an unusually low sex composition at birth of the African sector of the South African population. He hypothesizes that that inferior social and economic conditions impact more negatively on male than female foetuses from the moment of conception. The relationship between poverty and its effects on low sex ratios at birth and in the early years of life have been confirmed in other countries (for example in the United Kingdom in the 1930s).

Census ’96 shows that more than half the population overall (53,7%) was living in urban areas (see Table 14), compared with fewer than half (48,3%) estimated to be in urban areas in the forward projection from 1991.

In 1991, however, adjustments were made among Africans living in the former ‘self-governing territories’, which were mainly rural, to bring estimates more in line with demographically-modelled expectations. Thus 46,7% of the African population was actually counted as living in these areas, but this figure was adjusted upwards to reach 49,6%. Overall, among the African population, 54,1% was actually counted as living in non-urban areas, but this was adjusted upwards to reach 57,3%. These adjustments had ‘a significant effect on the geographic distribution of the black (i.e. African) population as counted’ (Nel, Loubser and Van Wyk, 1993, p. 16). The authors further state that ‘without the adjustments the importance of the self-governing and the non-urban areas would … have been underestimated’ (p. 16).

Amongst whites in 1991, the unadjusted count was 4,5 million. It needs to be noted that the white population as enumerated in 1991 was reported to have declined by 2% per annum between 1985 and 1991 (Nel, Loubser and Van Wyk, 1993, p. 1).

This raises the question of whether the modelled size of the white population, rather than the actual count, was not too high in 1991, and whether the actual size has continued to decline since then. The lower-than-expected count of adjusted whites in 1996 may be due in part to the effect of increased undeclared emigration. The question of migration of whites will be discussed in a later section of this chapter. It may also be due in part to lower fertility and higher mortality amongst whites than modelled. However, it may also be due in part to under-enumeration amongst the white population being higher than estimated, even after the adjustments by the post-enumeration survey. In Census ’96, whites may have been difficult to reach because of suspicion of the process and the difficulty in finding a sufficient number of white enumerators.

In addition, in Census ’96 the method of classification into population groups had changed compared to earlier years. In common with other countries, classification depended upon self-perception, rather than an apartheid-based legal racial classification. It is possible that some people who were previously classified as white either did not specify their race or else classified themselves as African. This may also account, in part, for the lower-than expected tally of whites.

Table 16 indicates the actual count and the adjustments applied in 1991 for male and female children aged 0-4 years (excluding TBVC).

Source: Sadie, 1992.

The table shows that the actual count for children aged 0-4 years in 1991 was raised by as many as 1,1 million, i.e., from 2,7 million to 3,8 million, and that 1,0 million of the increment were African (the African count was raised from 2,0 million to 3,0 million). An adjustment factor of 1,51 was applied to the overall African population of children.

If we look at African male and female children separately, we see that the number of males was raised by 520 000, i.e., from 991 000 to 1,5 million, to adjust the count to the model. However, the number of females was raised by a smaller amount. A slightly higher adjustment factor was applied to African male (1,53) than to African female (1,49) children. These adjustments may be based on an over-estimation of fertility rates, and possibly incorrect assumptions regarding sex ratios at birth.

By contrast, in Census ’96 there were 4,4 million children aged 0-4 years, after PES-adjustments, of which 3,6 million were African, with an overall sex ratio of 99,2 for both unweighted and weighted data. A PES-based adjustment factor of 1,17 was applied in the 1996 census to children younger than one year of age and to those aged one year, while for those aged two or three years, an adjustment factor of 1,13 was applied, and for four year olds it was 1,12. This is considerably lower than adjustments made in 1991.

It may be argued that the reported proportion of children in the age category 0-4 years (12,3% among males, and 11,4% among females) is too low, since it would correspond to a too-great recent drop in the fertility rate (the total number of children a woman can expect to have in her lifetime). This under-reporting is a known problem in censuses in Africa, and some ‘smoothing’ across this and the higher age categories may be desirable after further research on the Census ’96 data regarding fertility rates. On the other hand, the drop in fertility rate may also be somewhat faster than was previously predicted.

For instance, recent evidence suggests that the estimates of age-specific and total fertility rate among African women, may have been too high in 1991 (Mostert, 1990; Rossouw and Jordaan, 1997).

As far as sex ratios are concerned, in the unadjusted census in 1970 (the last time there was a census covering the entire country) the overall sex ratio for Africans among the 0-1 year olds was found to be 99,1, interestingly close to the value of 98,9 found in Census ’96.

We have recalled how the 1991 population estimates, in total and by population group, reflected a demographic model. The actual counts from enumeration or from aerial photography and sample surveys were adjusted to fit the model; not the other way round. Subsequent mid-year population projections similarly relied upon modelled population estimates. There are no census data subsequent to 1970 which include the whole country, against which to compare the results of Census ’96.

By contrast, the estimates of Census ’96 represent a countrywide count adjusted on the basis of a countrywide PES. They take the empirical data, rather than the match to a demographic model, as primary.

There are two aspects of migration which are relevant. The first concerns the number of people who have migrated into South Africa from elsewhere, particularly Africans; and the second concerns the number of emigrants, particularly whites.

According to findings from Census ’96, 814 000 people of a total of 39,8 million were not born in South Africa (this excludes those found in institutions on census night and those who did not specify where they were born). Of the total population, 405 000 (excluding unspecified) declared themselves as not holding South African citizenship. This may be an under-statement, since there is likely to be a degree of ‘self-naturalisation’, i.e., respondents representing themselves as South African, particularly in respect of illegal immigrants.

It is very difficult to estimate, either from the census or from demographic modelling, how many immigrants there were in South Africa on census night in 1996. Published estimates tend to vary widely, particularly in view of the suspicion held by citizens of South Africa that a large proportion, particularly those from other African countries, are illegally in the country.

It is likely that some proportion of the difference between the previously modelled and the PES-adjusted outcome for whites is due to emigration. Among potential emigrants, the information available from those leaving the country is likely to be unreliable. For example, when comparing the data obtained from the five main countries of destination about the

number of South Africans who have emigrated there, a different picture is obtained compared to emigration figures based on returns of control forms by the Department of Home Affairs and published by Stats SA.

For the United Kingdom in 1994, there were 8 100 South African immigrants, compared to 2 880 reported emigrants from the South African side.

For Canada in 1994, there were 2 910 South African immigrants, compared to 679 reported emigrants from the South African side.

A similar picture is found in Australia, New Zealand and the United States of America.

Figures for earlier years suggest under-reporting of emigration by a factor of two to three.

People may prefer not to declare their emigration status at the time of departure for various reasons, including the possibility that they may return if they find it difficult to make a success in their new country of destination.

Demographic modelling was recently undertaken by Udjo (1998), using parameters based upon the unadjusted counts from Census ’96, and going back as usual to 1970 as the baseline for forward projections. The aim of this model was to examine fertility and mortality trends in South Africa. In the first, preparatory stage of modelling, migration was not taken into account and neither was population group.

Udjo’s population projections were based on three different starting points using the 1970 census: the unadjusted 1970 count of 21,8 million; Sadie’s adjusted estimates of 22,1 million; and the adjustments of the former CSS of 22,8 million. These were his starting points for the low, medium and high variants respectively.

Calculations yielded population estimates in October 1996 of 38,3, 38,8 and 40,0 million for the three variants, excluding net migration. These may be compared with the PES-adjusted Census ’96 estimate of 40,6 million, in which 958 000 foreign-born people are included.

A further step in this modelling process will be to ensure that differential fertility rates and net migration broken down by population group are taken into account.