Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
Monitoring the Spread of HIV Infection Charles F. Turner and Robert E. Fay The current picture of the AIDS epidemic is clouded. In 1986, the U.S. Public Health Service (PHS, 1986) estimated that approxi- mately 1.0 to 1.5 million Americans are infected by the HIV virus. These infected persons, often unaware of their illness, can transmit the disease to others. (Incleec3, high rates of transmission have been observed tFischT et al., 1987] even when one partner had been diag- nosec3 with AIDS and both partners had been counseled about the dangers of unprotected sexual intercourse.) Current studies suggest that a large proportion of this infected population may eventually clevelop AIDS. Thus, in terms of both the spread of the epidemic and its ultimate cost, the estimate that there are 1.0 to 1.5 million infected persons presents an even grimmer picture than the actual number who have developed AIDS (72,645 reported cases in the This paper was prepared during the spring of 1987 in reaction to proposals for routine (or mandatory) testing of all hospital patients, marriage license applicants, and others, in order to obtain "better" epidemiological data on the spread of HIV. The paper was widely circulated in manuscript form during the summer and fall of 1987. This back- ground paper is the same manuscript that was originally circulated (save for updating of numbers and copyediting); thus it does not review recent experiences of the National Center for Health Statistics and the Research Triangle Institute in their efforts to im- plement a survey program of the type described in this manuscript. This manuscript is included as a background paper to the committee's report because it describes the de- sign problems that must be solved in order to use sample surveys to monitor HIV preva- lence (see Chapter 1~. Where appropriate, editorial notes have been appended to de- scribe significant events that have occurred since the paper was prepared in 1987. ED. At the time the original manuscript was prepared, Charles Turner was a Scholar in Res- idence and Robert Pay a consultant at the National Research Council. 447
448 ~ BACKGROUND PAPERS United States as of September 5, 198S, according to the Centers for Disease Control tCDC, 19884~. Regrettably, there appears to be far less certainty about na- tional estimates of the number of infected individuals than about estimates of the number of AIDS cases. At present there exists no comprehensive system for monitoring the prevalence of human im- munodeficiency virus (HIV) in the United States or any other nation. In this paper, we briefly review the uncertainties inherent in current estimates of the prevalence of HIV infection in the United States. Subsequently, we identify some of the statistical require- ments of a reliable system for monitoring the spread of HTV infection in the U.S. population and the pitfalls involved in reliance upon data clerivecl from mandatory (or routine) screening of convenient popula- tions such as applicants for marriage licenses, hospital patients, and so on. We will then outline some considerations involved in mounting a practical survey program to provide the raw data required for such a monitoring system. While this discussion draws upon the situation that prevailed in the United States in early 1987, the issues ant] methods we discuss continue to be of relevance in this country (see Chapter 1), and may have applications in other nations. CURRENT ESTIMATES Various attempts have been made to estimate the total number of persons infected with HIV. Table 1 presents three such estimates for the United States, in addition to the PHS estimates. It will be noted that these estimates vary widely; the largest estimate (2.49 million) is more than three times larger than the smallest (0.75 million). We will briefly review some of the sources of uncertainty in these estimates. We commence by observing that the estimates are highly discrepant, anal, therefore, insufficient to provide reliable monitoring of the spread of HIV infection in the population. Furthermore, it appears that the uncertainties in these estimates follow from the procedures (and data) that have been used to generate the estimates. Proportional Mocle] A key element in the calculations of Curran et al. (1985) and Sivak and Wormser (1985), for example, was the fraction NAIDS NHIV
MONITORING THE SPREAD OF HIV ~ 449 TABLE 1 Estimates of Number of Persons Infected with HIV in the United States Source Population As of Estimate Curran et al. (1985) Total population May 1985 750 thousands PHS (1986) IV drug users and June 1986 1.25 millions homosexual men PHS (1987) IV drug users and Nov. 1987 1.17 millions homosexual men Sivak and Wormser (1985) Total population July 1985 1.76 million Bees (1987) Total population Dec. 1984 2.49 million aEstimated as the interval from 0.5 million to I.0 million. The midpoint of the interval is listed in the table. bEstimated as the interval from 1.0 million to 1.5 million. The midpoint of the interval is listed in the table. CEstimated as the interval from 945,000 to 1.41 million. The midpoint of the interval is listed in the table. where NAIDS is the total number of reported AIDS cases and NHIV is the total number of persons infected with the AIDS virus (HIV). This fraction can change substantially as an epidemic progresses. For infectious diseases with Tong latency periods such as AIDS, the fraction will be zero for a long time after the infection enters a new population. This occurs because the denominator (number of infections) can increase rapidly while the numerator (number of di- agnosed cases) remains zero. Similarly, in a static population that has been saturated with infection, the denominator of the fraction may remain constant since all vulnerable members of the popula- tion have been infected, however, the numerator (cliagnosed cases) will continue to grow. The instability of this fraction makes its use problematic, particularly when estimates from one subpopulation are applied to another subpopulation in which the infection may have been established for a different length of time. Multiplicative Mode} A different procedure was used to derive the most widely quoted estimate of HIV prevalence in the Unitecl States (presented in the Public Health Service's [1986] Coolfont report). That report con- cluded that "by extrapolating all available data, we estimate that there are between 1 and 1.5 million infected persons in those groups tIV drug users and homosexual men] at present." Although explicit calculations are not shown, the Coolfont report indicates that its
450 ~ BACKGROUND PAPERS authors estimated that 2.5 million American men between the ages of 16 and 55 are exclusively homosexual throughout their lives and 5 to 10 million more have some homosexual contact. Similarly, they estimated (without explicit reference to a source) that 750,000 Amer- icans inject heroin or other intravenous (IV) drugs at least once a week and that similar numbers inject drugs less frequently. These estimates of population size were then multiplied by estimates of the prevalence of HIV infections among these groups in order to gen- erate the widely quoted estimate that there are 1.0 to 1.5 million infected persons in these two groups. (The prevalence rates used in these calculations were not published but the report states that HIV prevalence estimates range from 20 to 50 percent for homosexual men and from 10 to 50 percent for users of {V drugs.) This procedure (and a hybrid! employed in the calculations of Sivak and Wormser) is vulnerable to errors of unknown magnitude in both multiplicands. Note, for example, that the Coolfont estimate used data collected by Kinsey and coworkers (1948) in the 1940s in order to estimate the current number of male homosexuals in the United States. Even 30 years ago, Kinsey's data were widely regarded as unreliable for use in making such estimates because the research did not employ probability sampling and because the respondents were disproportionately drawn from the Midwest and from the college-eclucated segment of the population (see Terman, 1948; Wallis, 1948; Cochran et al., 1953~. Today, a further leap of faith is required since the relative size of the (self-reported) ho- mosexual population must be assumed to have remained unchanged since the 1940s. Furthermore, we note that estimates of the preva- lence of HIV infection among homosexual men were not derived from representative samples. These factors introduce substantial uncertainty into the resul- tant estimates of HIV prevalence. Indeed, the multiplication used in this estimation introduces some unique problems in that errors in the two multiplicands are related, although the form of the relationship is not known with any precision. Consider, for example, the defini- tion of what has loosely been termed the "homosexual population." If the definition of this population is restricted to persons whose sexual contacts have been exclusively (or predominantly) homosex- ual, the population will be smaller than if men who have hacI any homosexual contacts since the onset of the epidemic are included. Clearly, however, the estimates of HIV prevalence will also change (by unknown amounts) depending upon which definition is used. (May and Anderson t1987] provide useful models representing the
MONITORING THE SPREAD OF HIV ~ 451 underlying processes.) Parallel uncertainties afflict estimates of the number of persons who use IV cirugs and the rates of infection in this subpopulation. Temporal Distribution Mode} The most recent estimate -(as of June 1987) of the number of persons infected with HIV in the United States was produced by Rees (1987) using a different approach. Rees fit a normal distribution to the frequency of AIDS cases as a function of time elapsed between the initial HIV infection and the diagnosis of AIDS. The parameters of this distribution fib = 15, ~ = 5) were selected on the basis of a relatively small sample (n = 144) of persons who developed AIDS from blood transfusions. The resultant estimate that 2.49 million Americans are infected with HIV is the largest of the four estimates in Table 1. Indeed, this estimate is actually more extreme than it appears since it refers to an earlier date than the other estimates (December 1984) and it includes only "HIV infections that will result in AIDS over the next 30 years or so" (Rees, 1987:345~. Although the procedures proposed by Rees open another avenue for modeling, his estimates must be treater! with caution. Note that Rees's selection of the normal distribution and its parameters (p = 15, ~ = 5) was based upon the relative timing of AIDS onset for those who have developed AIDS from transfusions. The usual appropriate denominators (i.e., the number of persons infected via transfusions or the total number of persons infected via transfusion who will eventually develop AIDS) were not available. An important implication of Rees's choice of parameters is that only 5.3 percent of those eventually (1eveloping AIDS (under his model) would be expected to do so in the first seven years after infection. This rate appears quite low. Such a low presumed rate of AIDS onset in the early years after infection implies large multipliers for each AIDS case observed during this period. It will be several years before the validity of Rees's assumptions can be fully tested;) however, the fit of Rees's moclel is particularly poor in the first year. Rees's fitted values for the first year of infection add up to 20.98 AIDS cases through 1983, compared to S.5 observed cases. (Fractional observed cases result from Rees's splitting of ambiguous cases between adjacent years.) iFor present purposes, the committee concluded that a mean of 8 may be more appro- priate (see Chapter 2, and Lui et al., 1988~. ED.
452 ~ BACKGROUND PAPERS Because of such uncertainties in estimation procedures, it is, perhaps, not surprising that the estimates shown in Table 1 vary by a factor of 3. Nor is it surprising that similar uncertainties exist in other nations. (Rees's model, for example, estimates that 109,000 U.K. residents were infected by mid-1985. Rees observes, however, that "there seems to~be a general opinion that there were about 30,000 infections in the micIdle of 1986, between a quarter and a third of my estimate for a year earlier" (Rees, 1987:345~. PROPOSAL Inaccurate estimates of the size of the population infected with HTV are potentially dangerous because (among other reasons) such es- timates may generate a false sense of security or a false sense of alarm among those charged with formulating policies to cope with the epidemic. The 1986 report of the Institute of Medicine/National Academy of Sciences (TOM/NAS, 1986) concluded that better in- formation is needed to quantify the number of persons infected with HIV. This report recommended undertaking "extensive and repeated surveys of seropositivity to determine the incidence and prevalence of infection by age, race/ethnicity, geographic area, and sex" (p. 200~. Epidemiological precedence for this approach is somewhat limited, but several factors specific to the AIDS epidemic argue for the use of national sample surveys to obtain these data. Specifically, . For many infectious diseases there is only a short period between infection and manifestation of symptoms. For HIV infection, however, the latency period is several years; thus, the size of the currently infected population will remain unknown if one relies solely upon systems that monitor the number of AIDS cases. . Sample surveys are typically ineffective at measuring very rare characteristics unless the specific subpopula- tion can be identified in advance. Regrettably, it appears that HIV infection is no longer extremely rare. Sample survey methods can now give effective estimates of the size of this population, without the need to explicitly identify high-risk individuals, such as homosexual males or intravenous drug users. Controlling the spread of the AIDS epidemic will depend critically upon public education programs and other so- cial interventions. Periodic sample surveys of the preva- lence of HIV infection in the population will provide
MONITORING THE SPREAD OF HIV ~ 453 one important indicator of the overall impact of these programs. A clearer picture of the magnitude of the epidemic in terms of the prevalence and incidence of infection and its distribution by age, sex, geographic area, ethnic group, and other distinguishing characteris- tics is a prerequisite for designing and targeting intervention efforts in a rational and cost-effective manner. Without such information, risk reduction efforts will always be targeted less than optimally. Ad- ditionally, there will be no way of monitoring on a broad scale how effective our interventions are in combatting the spread of the epi- demic. Information on trends in HIV prevalence will also be essential in planning for the provision of health care. REQUIREMENTS Given the range of estimates shown in Table 1, few thoughtful ob- servers would question the need for more reliable evidence on the current prevalence of HIV infection in the American population. With a minimal amount of reflection, most observers would also con- clude that our need for up-to-date information on HIV prevalence will continue into the future at least until such time as AIDS ceases to be a medical threat. Thus, reliable estimates of HIV prevalence must be obtained at regular intervals in order to track the spread of the epidemic. In addition to the requirement that HIV prevalence estimates be generalizable to the population and regularly updated, it is also desir- able that the estimates permit disaggregation sufficient for prevalence (and changes in prevalence over time) to be monitored in particular demographic (and other) subgroups of the population. Thus, preva- lence estimates that allow us to separately track the spread of the infection among unmarried youth, for example, may be as critical as estimates of prevalence in the entire population. Even this minimal set of desiderata constrains our data-gathering activity in several important ways. It requires that our estimates be . representative of the entire population that is at risk, . based on data collected at regular intervals, and . capable of providing estimates for subpopulations of in- terest.
454 ~ BACKGROUND PAPERS LIMITATIONS OF SAMPLES OF CONVENIENCE The requirement that prevalence for the American population be characterized demands that probability samples be drawn from that population. Samples of convenience even when they are extremely large are not sufficient to satisfy the preceding list of statistical needs. Prevalence ciata derived from blood donors, armed forces re- cruits, persons voluntarily seeking testing, and other special popula- tions are constitutionally incapable of providing reliable information on the key question: What is the prevalence of HIV infection in the general population? The assumption that prevalence per se or trencis in prevalence in these special populations will be mirrored in the general population requires a leap of faith. Given the seriousness of the AIDS epidemic, such leaps should be discouraged to the extent scientifically possible. What is needed are better data. With better data such assumptions can be avoidecl.2 It follows from such considerations that proposals to initiate mandatory or routine HIV-antibody testing of hospital patients, ap- plicants for marriage licenses, and so forth will not provide reliable evidence on the prevalence of HIV infection in the population.3 These convenient populations provide pieces of information from a larger puzzle, but they provide this information in a manner that does not readily permit us to reassemble that puzzle. We know, for example, that hospital patients are a population that is both substantially older and sicker than the general population. We also know that some patients may be admitted to hospitals several times during a brief period and thus may be double- or triple-counted with such an approach. Healthy aclults, on the other hand, are unlikely to be hospitalized. 2This is not to say that HIV prevalence data should no longer be collected from these convenient sources. However, careful studies are needed that document the nature of the interrelationships (if any) between trends in HIV prevalence in the general population (estimated from probability samples) and trends in HIV prevalence in convenient (but self-selected) samples such as blood donors and military recruits. Over time, such studies might provide an empirical basis for broadening the interpretation of trend data derived from such special populations. 3Mandatory (or routine) testing of hospital patients has been proposed for monitoring the course of the epidemic as well as for other purposes, such as "protecting health care workers." The deficiency of such proposals for epidemiological purposes does not, of course, imply that such testing might not serve these other purposes. It should be noted, however, that the recent report of the Institute of Medicine/National Academy of Sciences (1986:124) recommended that "the question of whether to undergo tHIV] testing should be a personal health care decision to be made by an individual, ideally following counselling by health care professionals."
MONITORING THE SPREAD OF HIV ~ 455 The fact that the hospital population is considerably older than the general population will doubtlessly introduce a downward bias in prevalence estimates for any sexually transmitted disease that has been introduced recently into the population. On the other hand, persons with AIDS are more likely to be hospitalized than members of the general population. Furthermore, persons with HIV infection (but not diagnoses! as having AIDS) may also be more likely to be hospitalized.4 Such factors would introduce an upward bias in preva- lence estimates derived from hospital populations (versus estimates derived from the general population). Finally, one cannot rule out the possibility that a program of routine screening as a condition of hospital admission would discourage persons from seeking hospital care if they believe they may be infected. Besides the terrible im- pact this would have on the individuals involved, it would introduce another source of bias into hospital-basec! prevalence measures. Adjustments of hospital prevalence data might be attempted. For example, projections to the national population might adjust the hospital data to match the age distribution of the national pop- ulation. Even the most ingenious adjustments, however, could not escape the essential uncertainties of using HIV prevalence rates for persons who pass through hospitals to estimate rates for persons who are not hospitalized. Age (or other) adjustments merely restrict the domain of our assumptions. Thus an age-adjustment procedure might have us assume that the prevalence of HIV infection among hospitalized 20- to 29-year-olds, for example, is equivalent to that for 20- to 29-year-olds in general. The validity of such assumptions would remain unknown. Doubtlessly, other restrictions could be introduced. For exam- ple, (liagnosed AIDS patients might be exclu~led from the hospital estimates (relying on the CDC case-reporting system for an estimate of the number of persons who are both HIV-antibody positive and cliagnosed with AIDS). We might also derive estimates of HIV preva- lence using only patients admitted for reasons that appear medically unrelated to HIV infection (e.g., accidental injuries). Even such seemingly attractive strategies have the potential for introducing bias. We know, for example, that the propensity of inclividuals to seek meclical care (given similar symptoms) is corre- lated with a number of factors including socioeconomic status. The 4Subsequent to the preparation of this draft (in early 1987) Des Jarlais and coworkers (1988) reported a sharp rise in non-AIDS deaths from pneumonia (not pneumocpstis carinii), tuberculosis, and endocarditis among HIV-infected IV drug users. ED.
456 ~ BACKGROUND PAPERS prevalence of HIV infection is unlikely to be the same in all so- cial classes, and prevalence estimates derived from hospital samples necessarily reflect the joint operation of socioeconomic variations in hospital admission rates and differences across social classes in the rates of HIV infection. Furthermore, although using "accidental" in- jury cases may seem to introduce an element of randomness into the sampling, the probability of accidental injury may not be indepen- dent of HIV antibody status. For example, {V drug users may be at greater risk of accidental injury due to impaired cognitive function- ing (and, therefore, more likely- to appear in a sample of hospitaTizecl accident victims). Since IV drug users are at risk of HIV infection due to needle-sharing practices, prevalence rates could be biased by the joint operation of these two factors. Inventive readers will be able- to imagine a myriad of other po- tential biases, methods that might reduce them, further biases that remain or are newly introduced, and so forth ad infinitum. Such exercises are limited only by the fertility of one's imagination. The lesson taught by such exercises is that data from hospital and other seemingly "convenient" populations can never provide completely trustworthy evidence of the prevalence of HIV infection in the gen- eral population. A concrete example of the limitations of convenience samples is provided by recent reports in the popular merlin that the prevalence of infection cletectecT among military recruits in the United States diet not increase cluring the first 15 months of the military's testing pro- gram (see, for example, Washington Post, May 15, 1987~.5 Although this result may appear encouraging, it is actually quite (lifficult to interpret. We cannot, for example, rule out the possibility that po- tential military recruits who had engaged in high-risk behaviors were discouraged from volunteering for military service by publicity about the mandatory testing of recruits. Rather than expanding the quixotic enterprise of mandatory testing to include other "convenient" populations, we believe a pro- gram of regularly conducted surveys that obtain blood for testing from probability samples of the national population should be con- sidered. Such a research program might, if properly (resigned and executecl, provide a simpler, more reliable, less controversial, and less costly way of monitoring the spread of HIV infection in the pop- ulation. In the following pages we outline some of the factors that should be consiclered and tested in planning such an endeavor. MAIDS Rate Remains Stable Among U.S. Military Recruits Since Testing Started in 1985; Statistics Puzzle Experts," Washington Post, May 15, 1987:A1.
MONITORING THE SPREAD OF HIV ~ 457 DESIGN CONSIDERATIONS Overview An Oversimple summary of such a survey program might in- volve the following steps. The population to be surveyed would be identified. An appropriate probability sampling procedure would be designedly, and survey procedures would be pretested. During this preliminary stage, a decision might be made to exclude certain seg- ments of the total population from the survey (e.g., the elderly and young children). If the survey used a household sampling frame as one component of its sample design, sampled households would sub- sequently receive a letter which would (1) advise them that an inter- viewer would be calling at their household and (2) inform them of the purpose of the survey and the safeguards ensuring their anonymity. Subsequently, a survey interviewer (together with a phIebotomist) would visit the household. The interviewer would ascertain how many eligible respondents lived in the household (excluding persons outside the target population), and a respondent would be selected at random from among the eligible respondents (if any). The survey interviewer would~ ask the designated respondent a short series of de- mographic questions and subsequently request that the respondent provide a blood sample. Results of HIV testing of these blood samples would provide a ba- sis for estimating HIV prevalence in the total population and selected demographic subgroups.6 Repeated annually (or at another suitable interval), such surveys might provide a reliable way of monitoring the spread of HIV infections in the population. With this oversimplified summary of the procedure in mind let ~ - ~ us review in greater detail some of the important considerations in the design and execution of such a survey. Sample Size and Efficiency For purposes of initial discussion, a sample of approximately 10,000 persons might be considered for HIV screening. We might, in ad- dition, restrict the sample to persons age 18 to 54. This restriction would avoid problems of consent below age 18 and reflect the prac- tical judgment that those 55 and over are likely to contribute little Gin addition to providing statistical data on prevalence, the stored sera from this program could provide a valuable resource for future biomedical and epidemiological researchers. It is important, therefore, that plans be made for long-term storage of the sera collected in these surveys.
458 ~ BACKGROUND PAPERS to the overall total of seropositive individuals. (As discussed below, samples that are substantially larger than 10,000 may be desirable in order to provide reliable estimates of prevalence in important subpopulations. ~ Because the survey estimate will be based on a sample, some random error due to the selection of the sample will be present. If approximately 1.5 million Americans are infected at the time the first survey is done, the reliability of a survey estimate based on a sample of 10,000 could be stated as a relative standard deviation of about 10 percent. In other words, if the expected sample estimate (averaged over all possible samples) were 1.5 million persons, then two times out of three the sample estimate should be within 150,000 persons of this value, that is, within the interval from 1.35 to 1.65 million persons. Furthermore, 95 percent of the time, the sample estimate should fall within the wider interval from 1.2 to 1.S million persons. This random variation is an inevitable consequence of survey work. A survey estimate with this degree of reliability would be tremendously more informative, however, than currently available estimates. As Table 1 indicates, current estimates vary over a much wider range. Such a survey could also provide separate estimates for demo- graphic subgroups (e.g., for (different age groups, marital statuses, minority groups, and genders). The relative reliability of these esti- mates, however, would be less than that for estimates based on the whole sample. To provide more reliable estimates for such subpopu- lations, larger samples (and oversampling of certain subpopulations) are required. Consider, for example, estimates of HIV prevalence for females. If 90 percent of HIV infections are assumed to be found in males, then only 150,000 of the presumed 1.5 million infected persons would be female. If half of our survey sample of 10,000 were female, estimates might have a coefficient of variation of about 30 percent. This variation is too large for most purposes (although, arguably, it might be an improvement over present estimates).7 If, however, the sample included 20,000 women, the coefficient of variation should be in the range of 15 percent, yielding a 95 percent confidence interval of 105,000 ~ n ~ 195,000 (if 150,000 is the expected value averaged over all possible samples). - / The launching in 1989 of a large-scale neonatal HIV screening program (see Chapter 1) will allow the HIV serostatus of childbearing women to be estimated directly. (Newborns carry the maternal antibodies to HIV, whether or not the newborn itself is infected.) This new development may reduce the need for such information from a sample survey. Although it should be noted that the newborn screening can provide no information on the HIV status of women who have not borne children. ED.
MONITORING THE SPREAD OF HIV ~ 459 Sample sizes of 10,000 (or even 50,000), it should be noted, are not large in comparison to other surveys conducted by federal statis- tical agencies. The Bureau of the Census, for example, interviews a sample of over 60,000 respondents each month to provide data for the monthly unemployment estimates. Similarly, approximately 40,000 households are interviewed each year in the Health Interview Survey program of the National Center for Health Statistics (NCHS). Design The proposed survey might profitably employ a complex sampling de- sign. Oversampling of groups with high risk, followed by appropriate weighting of these groups in forming the survey estimates can im- prove the overall reliability. So, for example, oversampling younger, unmarried persons relative to older married couples could produce substantial gains in overall sampling efficiency (on the assumption that the former will yield a higher proportion of HTV-positive cases than the latter). Oversampling couIcl also improve the reliability of estimates for specific demographic groups (e.g., blacks and Hispan- ics). To take one simple example of the nature of such a sample design, let us suppose that the geographic distribution of HIV infections mirrored the distribution of AIDS cases (as of April 6, 1987~. A complex sample design considering only this one factor might divide the United States into three geographic strata: 1. the New York ant! San Fiancisco Standard Metropolitan Statistical Areas (SMSAs), which had a population of 12.37 million in the 1980 census and 12,680 AIDS cases as of April 6, 1987; 2. the SMSAs ranked third to seventeenth in number of diagnosed AIDS cases, which bad a total population of 44.27 million and 11,663 AIDS cases; and 3. the rest of the Unitecl States, which had a population of 173.47 million in the 1980 census and 9,377 AIDS cases diagnosed as of April 6, 1987. If only this geographic distribution is consi(lerecl, an optimum sam- ple allocation would assign approximately 15 percent of the survey sample to the first stratum, 30 percent to the second, and 55 per- cent to the third. This allocation would reduce sampling variation in estimates of the prevalence rate by (disproportionately drawing sam- ples from geographical areas where there is presumptive evidence
460 ~ BACKGROUND PAPERS that the prevalence rate is high. Thus stratum 3, which has a lower (presumed) prevalence rate, would be allocated 55 percent of the sample even though it represents 75 percent of the total population. (The final survey estimates, of course, would employ an appropriate reweighting to derive national estimates that take account of the disproportional allocation of the sample to these geographic strata.) A complex sample design employing only these three geographical strata could reduce the sampling variance by as much as 24 percent (compared to a design that allocated the sample in proportion to the total population of these areas). Further refinements of the sample design (e.g., taking into account variations in AIDS prevalence by age, marital status, or race and ethnicity) might reduce the sampling error even further, perhaps by as much as 40 to 50 percent in total.8 Interview Data Such a survey should collect, at a minimum, information on age, race and ethnicity, gender, and marital status from the respondents. Respondents are almost always willing to supply such information, so these questions should not threaten respondent cooperation. The resulting information about the spread of HIV infection in different geographic areas and among various demographic subgroups could be of great practical and scientific value. Questions on sexual practices and {V drug use, on the other hand, would also be highly informative but could reduce the level of cooperation. Hence, we suggest that such sensitive questions should be omitted from the national prob- ability sample proposed here, unless pretesting indicated that these data could be obtained without decreasing the response rate for the blood test. Cost An HIV survey of the sort outlined is likely to cost less than $400 per sample case (this includes development and processing costs). Thus, the cost of an annual survey of 10,000 respondents should not scare must be exercised to ensure that decisions to decrease the sampling error of the overall estimate do not compromise the ability to obtain adequately reliable estimates of HIV prevalence in important subpopulations. So, for example, the design should not allocate sample disproportionately to male respondents, even though men make up the preponderance of current AIDS cases. (Changes over time in HIV prevalence among women can provide an important indicator of the extent of heterosexual transmission in the population. Thus, estimates must be reliable enough to track changes in the prevalence rate for women.)
MONITORING THE SPREAD OF HIV ~ 461 exceed $4 million. (Because some of the costs of the survey are fixed, increasing the sample size by a factor of four should not produce a fourfold increase in survey costs.) With the annual federal budget for AIDS projected to exceed $1 billion in the near future and with even larger costs being borne by individual AIDS victims, state and local governments, insurers, and the health care system),-the proposed expenditure does not seem prohibitively expensive. When compared to the costs of mandatory testing programs that have recently been discussed, the costs of even a vastly expan(le(1 survey program would be small. The cost of HIV testing performed at alternative sites with appropriate pre- and posttest counseling averages approximately $40 per individual (IOM/NAS, 1986:17~. Even if the cost of blood tests and counseling for persons applying for marriage licenses were only one quarter of this amount (i.e., $10 per inclividual), the total annual cost would be only approximately $50 million in the United States. Mandatory screening of all patients admitted to hospitals would be much more expensive (in 1982 there were 39 million such admissions). Auspices In most cases, government statistical agencies obtain higher levels of cooperation from the public than other survey organizations (see, for example, National Research Council, 1978:42, Table 1~. We think therefore that an effort by one or more federal statistical agencies might be a reasonable approach. The possibility of using a private survey research firm, with a well-established record for quality, to carry out this work should not be excluded. Incleec3, there are cir- cumstances (see below) in which such a strategy would be mandatory. The possibility of adding HIV testing to an ongoing survey should also not be excluded. Survey designers, however, would need to be sensitive to possible adverse reaction to such "piggybacking." Sur- veys are frequently perceived as involving an implicit social contract between the survey taker ant! the respondent. Respondents con- tacted to take part in an immunological health survey may more readily appreciate the importance of blood testing than respondents in another survey who might view an adcled request for a blood sample as irrelevant to their original commitment. Furthermore, re- searchers gathering survey data for other purposes might justifiably worry about the effects that such piggybacking might have on their own data collection.
462 ~ BACKGROUND PAPERS Pretesting Most surveys require a period of "pretesting" to evaluate the effect of the questionnaire and other survey procedures on the quality of the data to be collected. Such testing would be especially critical in developing an effective strategy for this survey, given the emotionally charged atmosphere surrounding public discussions of AIDS. Thus, greeting respondents at the door with "GoocT afternoon. May T please have a sample of your blood?" is unlikely to yield the desired re- sults. A carefully structured interview, conclucted by a well-trained interviewer and followed by blood sampling by a quaTifiecl health pro- fessional, may succeed. A letter stating the purposes an importance of the survey and detailing privacy ant! confidentiality provisions may help to set the stage for the interview, if mailect to the respondents in advance of the initial contact. Supplemental reading material (or a videotaped presentation) on the importance of the survey may also help to inform respondents of their expecter! roles. We would antic- ipate that many details of the survey strategy would be refined and tested in a pilot phase (prior to the formal data-gathering phase of the survey). We will not attempt to anticipate all of those details here. However, three issues are sufficiently important to merit fur- ther discussion: (1) assuring the completeness of survey coverage, (2) guaranteeing the confidentiality of survey data, and (3) providing blood test results to interested respondents. We briefly discuss each of these issues below. Aclequacy of Survey Coverage Some readers may wonder whether it is feasible to achieve high levels of respondent cooperation in surveys that ask randomly se- lectec! respondents to provide blood samples. Although it may seem unusual to request that survey respondents provide blood samples, precedent suggests that such surveys are possible and do produce relatively high levels of cooperation. The National Health and Nutri- tion Examination Survey (NHANES), for example, performs physical examinations and administers a range of tests (including blood tests) to large probability samples of the U.S. population (approximately 21,000 examinations were performed in the most recent surveys). The recent experience of the NHANES program has been that ap- proximately 91 percent of the public consent to a lengthy household health interview, and approximately 73 percent subsequently report to the survey's medical facility for examinations. (It should be noted that although NHANES could provide HIV prevalence estimates, its
MONITORING THE SPREAD OF HIV ~ 463 sampling schedule would not yield timely estimates. According to NCHS personnel, national estimates from the next NHANES survey will not be available until 1991.) Nonresponse to the NHANES medical examination occurs for various reasons. Empirical studies suggest that nonresponse rates do not vary substantially by sex or race. Examination rates were 71.8 percent for females, 74.4 percent for mates, 72.7 percent for whites, 75.6 percent for blacks, and 74.3 percent for other racial groups. There is, however, a monotonic decrease in examination rates with in- creasing age. Medical examinations were completed with 81 percent of designated respondents aged 6 months to 17 years, 74 percent of 18- to 34-year-olds, 69 percent of 35- to 54-year-olds, and 64 percent of persons aged 55 or oilier. The NCHS (1982:Appendix 1) reports, nonetheless, that various comparative studies indicate that respon- dents who do not undergo medical examinations in the NHANES pro- gram have self-reported health characteristics similar to respondents who do undergo those examinations. Similarly, Forthofer (1983:507) finds that after standard NCHS nonresponse and posts/ratification adjustments, "there is excellent agreement in the marginal distribu- tion of variables between NHANES-~] for examined persons and the 1976 National Health Interview Survey (NHIS)" (which achieved a 96 percent response rate). The response rates from the NHANES demonstrate that most respondents are willing to provide blood samples in a national health survey. Three features of the proposed HIV survey might encourage even higher response rates than those obtained in the NHANES. First, the survey focuses on a health problem of pressing national importance. Second, restricting the survey interview to a few ques- tions will minimize the burclen on respondents. Finally, by cirawing blood samples in the respondents' homes, the survey eliminates the need for respondents to travel to a central site for testing. Although these factors encourage the belief that blood samples can be obtained from a substantial proportion of eligible respondents, it must be recognized that HIV blood testing is an emotionally and politically charged issue. As noted above, careful pretesting (and other measures) will be required to explore the feasibility and refine the details of any such measurement program. It is vital that sufficient time and resources be devoted to these preliminary research activities because they will have a major impact on the quality and usefulness of the resultant data. Several potential impediments to adequate survey coverage can be identified at the outset. First, household-based sampling frames
464 ~ BACKGROUND PAPERS will not obtain data from persons who are not attached to a house- hold. Such sampling frames will miss persons who are homeless or who live in prisons, hospitals, and other institutions. An HTV prevalence survey using a household sampling frame will, therefore, underrepresent some important subgroups (e.g., current or past IV drug users) if those subgroups are disproportionately homeless or institutionalized. In designing a survey, separate estimates must be obtained for important subpopulations that will be missed or un- derrepresented in household surveys (e.g., by drawing samples from prisoners or persons in drug treatment centers). More troubling than the Toss of identifiable segments of the popu- lation from the sampling frame is the loss of an unidentifiahie fraction of the population because some respondents choose not to cooperate with the survey. As this fraction increases, the claim that the survey "represents" the general population is weakened. For our purposes, this threat is particularly serious because we are seeking to assess a relatively uncommon characteristic (HIV seropositivity) in a situa- tion where one might expect noncooperation to be correlates! with the respondent's known or suspected seropositivity. That is to say, persons who know or suspect that they are infected may be more likely to refuse to participate in the survey.9 Minimizing such refusals will be major challenge. Indeed, it must be recognized at the outset that it is both impossible to eliminate refusals entirely and likely, even under the best of circumstances, that these refusals will introduce some bias. As a consequence, estimation of HIV prevalence will require some imputation of missing data.~° Furthermore, it is important to recognize that the survey effort could fad! completely because of substantial noncooperation by mem- bers of the public. Hopefully, this failure would be detected during pretesting, so that the cost of failing wouicT be minimized. Guaranteeing Anonymity A first step in reducing noncooperation is to ensure that both in fact and in the perceptions of potential respondents, the survey poses no actual or potential threat. Given the level of public concern about 9Subsequent to the drafting of this paper, Hull and coworkers (1988) Reported results demonstrating this phenomenon in a situation in which confidential testing was offered to clients of an STD clinic ED. 1OTo assist in this process, the suggestion has been made that a recent sample of re- spondents from another survey (e.g., the Health Interview Survey) might profitably be used. Data from the previous interview would assist in the analysis of the biases caused by nonresponse to the HIV survey.
MONITORING THE SPREAD OF HIV ~ 465 AIDS and the particular vuInerabilities of the groups presently ex- periencing the bulk of AIDS cases, we believe it is essential that bloocT test data be collected in a way that ensures the anonymity (not merely the confidentiality) of these data. Even if persons fa- miliar with the excellent record of the federal statistical system in preserving confidentiality are reassured by past history, it is unlikely that their trust will be shared by all members of the public. This may be particularly true for respondents who feel vulnerable to dis- criminatory actions as a result of being identified as carriers of the AIDS virus. Obviously, as a practical matter one must know the addresses of respondents in order to draw a sample. Thus, total anonymity of the respondents is impossible. We would propose, however, that the blood test results be obtained in a manner that guarantees those results will be anonymous. Procedures to achieve this end might include the following: . To the fullest extent possible, no linkage should be allowed at any stage of the fieldwork or data processing between addresses and the identification numbers assigned to respondents. As soon as practical after completion of the relevant portion of fieldwork, all materials that might allow the identification of actresses should be destroyed. If practicable, this should be done on a continual basis at the local sites. No information allowing linkages between addresses and iclenti- fication numbers should be transmitted from the fieldwork sites . to the central site. All blood samples should be identified only by encrypted i(lenti- fication numbers. All sampling records should be destroyed as soon as practical after completion of the relevant portion of the fieldwork. Blood samples should be tested by one central laboratory, and its testing procedures shouIcI be subject to strict security and quality control measures. Tests should not be performed until sampling, fieldwork, and other survey records have been destroyed. The results of the blood testing should be transmitted only to the central site (not to the fieldwork sites). All interview ciata obtained from respondents (e.g., age, sex, race) when entered into a data file and combined with all remain- ing geographic information, must conform to present stanciards for public use tapes released by the Bureau of the Census. .
466 ~ BACKGROUND PAPERS The foregoing are initial suggestions for the types of precautions that will be required to guarantee anc Amity of the data records. The ultimate test will be whether or not the most knowledgeable members of the sampling, fieldwork, and statistical staffs of the survey agency can state with certainty that they would be unable to identify the ciata derived from any individual respondent (using all records that remained available from the survey, together with the data file). This litmus test should be satisfied prior to the merging of the encrypted HIV test results with the information obtained in the · , ~ survey interview. Perception of Anonymity In addition to guaranteeing the anonymity of the survey data records, it is crucial that the public at large, all respondents to the survey, and key social groups (especially those presently bearing the brunt of the AIDS epidemic) be convinced that the design of the survey preclucles any threat to the respondents in the survey. Moreover, each of these audiences must be convinced that the resultant data will play an important role in the U.S. attempt to cope with the AIDS epidemic. We suggest that representatives of these different audiences shouIcl be appointed to review the design and execution of the sur- vey. This group shouicl be provided with access to survey sites and appropriate staff to allow them to undertake an independent review to ensure that the anonymity of survey respondents has been fully protected. Their certification of this fact should be made a prereq- uisite for the merging of the encrypted blood test results with the other survey data. To further ensure that the anonymity of the data gathering would not be compromised (anal to further reassure the respondents on this point), it may be clesirable for the survey to be performed in cooperation with the Bureau of the Census so that the protections of Title 13 of the U.S. Code would apply (Title 13 provides crimi- nal penalties under federal law for violations of the confidentiality of census data). Careful pretesting shout be undertaken, however, to ascertain whether the "reassurance" provided to respondents by such legal protections will outweigh any reluctance of respondents to provide sensitive information directly to employees of a federal statistical agency. If a substantial fraction of the population (partic- ularly groups with atypical HIV prevalence rates) feels threatened by the request to provide blood samples to interviewers from a federal
MONITORING THE SPREAD OF HIV ~ 467 statistical agency, it may be advisable for data gathering to be done by a nongovernmental survey research organization. Informing Respondents of Test Results There is an inevitable tension between the desire to provide fail-safe protection of a respondent's anonymity and the desire to provide test results to individuals who want this information. The design we have outlined precludes direct notification of indiviclual survey respondents (because names are never obtained and acldress records would be destroyed). Nonetheless, there are ways in which test re- sults might be provided without directly recontacting respondents. We believe, however, that it would be inadvisable to provicle such information without appropriate counseling and support services. Counseling, of course, requires personal contacts that would jeopar- dize anonymity. Alternative procedures should be considered that would protect the anonymity of the survey data while permitting respondents to use the blood samples drawn in the survey to obtain information about their HIV status (together with appropriate counseling). Any proposed procedure should have the following characteristics: . It should eliminate the need to ask respondents for their names. · It should eliminate the need to retain survey records that might permit the personal identification of blood samples for any respondent. It thereby ensures that all such records can be destroyed prior to testing of the blood samples. It should ensure that respondents who wish to know their HIV status can use the sample drawn during the survey to find that out, and it should do so without requiring that respondents reveal to the interviewer their interest in learning the test results. . It should not force blood test results upon respondents who may not wish to know their status. . It should guarantee that appropriate counseling will be provicled along with blood test results. . This list of requirements might be user! as a preliminary screening device for vetting proposed procedures. Survey development and pretesting are likely to reveal other requirements that should be met. One approach that satisfies these preliminary requirements would
468 ~ BACKGROUND PAPERS provide respondents with an encrypted identification number to be presenter! at a local medical facility if they wished to learn their blood test results. (They wouIc3, of course, be counseled at that time.) The encrypted number could be given to respondents in a seated envelope so that it would not be seen by the interviewer. Obviously, this number shout be different from the serial number on the survey interview form or the encrypted! identification number placed on the blood sample. This number might also be printed on a distinctive form so that it could not be copied, and the respondent might be instructed to place some personal information (e.g., height, eye color, sex) on the form so that it could not be used by anyone else to obtain the respondent's blood test results. An alternative strategy might involve dividing blood samples in half. One half of the sample wouIcT be taken for testing by the survey organization. The other half of the blood sample would be left with respondents, who would be given the address of a local medical facility that would provide free testing of the blood sample together with appropriate counseling. Respondents conic! then choose for themselves whether they wished to know their HIV test results. (HIV tests presently under development may make this alternative quite practical because future tests may require only droplets of blood lanced from a finger and blotted on filter paper.) The foregoing suggestions are only two of a large number of possible strategies. There is no need to commit to any particular strategy prior to beginning the design and pretesting of the survey. However, the five attributes listed above are a reasonable starting point for vetting any strategies proposed for informing respondents of their test results. Further consideration might also be given to the option of not providing test result data in order to provide more assured protection of the anonymity of the blood test results. CONCI`USION Public and scientific awareness of the impending magnitude of the AIDS epidemic has increased substantially in recent years. Yet an as- pect of unreality remains. Thus, many individuals and whole nations must face critical decisions without trustworthy information on the prevalence of the AIDS infection in the population. If the infection is as pervasive as the present fragmentary evidence suggests, a se- ries of annual estimates based upon representative national samples can serve to remind us all of how real and threatening the AIDS epidemic will become in the near future. Such estimates will permit
MONITORING THE SPREAD OF HIV ~ 469 more reliable tracking of the spread of the infection, and they will facilitate evaluation of the overall impact of educational and other interventions clesignecl to retard the spread of the AIDS epidemic. REFERENCES Centers for Disease Control (CDC). (1987a) Human immunodeficiency virus infection in the United States: A review of current knowledge. Morbidity and Mortality Weekly Report 36(Suppl. S-6): 1-48. Centers for Disease Control (CDC). (1987b) Human Immunodeficiency Virus Infections in the United States: Review of Current Knowledge and Plans for Expansion of HIV Surveillance Activities. Report to the Domestic Policy Council. Atlanta, Gal, November 30, 1987. Centers for Disease Control (CDC). (1988) AIDS Weekly Surveillance Report United States. Atlanta: AIDS Program, Center for Infectious Diseases, September 5. Cochran, W. G., Mosteller, F., and Tukey, J. W. (1953) Statistical problems of the Kinsey Report. Journal of the American Statistical Association 48:673-716. Curran, J. W., Morgan, W. M., Hardy, A. M., Jafte, H. W., Darrow, W. W., and Dowdle, W. R. (1985) The epidemiology of AIDS: Current status and future prospects. Science 229:1352-1357. Des Jarlais, D. C., Friedman, S. R., and Stoneburner, R. L. (1988) HIV infection and intravenous drug use: Critical issues in transmission dynamics, infection outcomes, and prevention. Reviews of Infectious Diseases 10:155. Fischl, M. A., Dickinson, G. M., Scott, G. D., Kitmas, N., Fletcher, M. A., and Perk, W. (1987) Evaluation of heterosexual partners, children, and household contacts of adults with AIDS. Journal of the American Medical Association 257:640-644. Forthofer, R. N. (1983) Investigation of non-response bias in NHANES-II. American Journal of Epidemiology 117:507-515. Hull, H. F., Bettinger, C. J., Gallaher, M. M., Keller, N. M., Wilson, J., and Mertz, G. J. (1988) Comparison of HIV-antibody prevalence in patients consenting to and declining HIV-antibody testing in an STD clinic. Journal of the American Medical Association 260:935-938. Institute of Medicine/National Academy of Sciences (IOM/NAS). (1986) Confronting AIDS: Directions for Public Health, Health Care, and Research. Washington, D.C.: National Academy Press. Kinsey, A. C., Pomeroy, W. B., and Martin, C. E. (1948) Sexual Behavior in the Human Male. Philadelphia: Saunders. Lui, K., Darrow, W. W., and Rutherford, G. W. (1988) A model based estimate of the mean incubation period for AIDS in homosexual men. Science 240:1333-1335. May, R. M., and Anderson, R. M. (1987) Transmission dynamics of HIV infection. Nature 326:137-142. National Center for Health Statistics. (1982) Hematological and nutritional bi~ chemistry reference data for persons 6 months-74 years of age: United States, 1976-80. Vital and Health Statistics, Series 11, No. 22, DHHS Pub. No. 83-1682. Washington, D.C.: U.S. Government Printing Office. National Research Council. (1978) Privacy and Confidentiality as Factors in Survey Response. Washington, D.C.: National Academy Press. Public Health Service. (1986) Coolfont Report: A PHS plan for the prevention and control of AIDS and AIDS virus. Public Health Reports 101:341-348. Rees, M. (1987) The sombre view of AIDS. Nature 326:343-345.
470 ~ BACKGROUND PAPERS Sivak, S. L., and Wormser, G. P. (1985) How common is HTLV-III infection in the United States? New England Journal of Medicine, 313:1352. Terman, L. M. (1948) Kinsey's "Sexual Behavior in the Human Male": Some comments and criticisms. Psychological Bulletin 45:443-459. Wallis, W. A. (1948) Statistics of the Kinsey Report. Journal of the American Statistical Association 44:463-484.
