Recommendations for the use of Taylor series confidence intervals for estimates of vaccine efficacy*

Bulletin of the World Health Organization, 66 (1): 99-105 (1988) Recommendations for the use of Taylor series confidence intervals for estimates of vaccine efficacy* A. W. HIGHTOWER,1 W. A. ORENSTEIN,2 & S. M. MARTIN1 A simple formula for calculating confidence intervals by means of a Taylor series variance approximation has been recommended for gauging the precision of estimates of vaccine efficacy. To evaluate theperformance ofTaylor series 95% confidence intervalsfor vaccine efficacy, we conducted a simulation studyfor commonly expected values ofvaccine efficacy, risk ofdisease in the unvaccinated population, and sample sizes ofthe vaccinated and unvaccinated groups. In the first simulation, the sample size in the vaccinated group was 500 or 1000, whereas that in the unvaccinated group ranged from 10 to 1000. The confidence intervals were accurate when the sample size in the unvaccinated group was >50 and the risk ofdisease was 0.3-0.9. In contrast, the intervals were too narrow when all three ofthefollowing situations occurred: the number ofunvaccinated was small (10 or 20), the true vaccine efficacy was relatively low (60% or 80%), and the risk ofdisease was 0.5-0.9. Furthermore, when the true vaccine efficacy was high (90% or 95%) and the disease risk in the unvaccinated was low (0.1 and 0.2), the confidence intervals were too broad, especially when the unvaccinated sample size was < 50. Additional simulations with a sample size in the vaccinated group of200 gave broad intervalsfor 95% vaccine efficacy (for all values ofdisease risk) andfor 90% vaccine efficacy when the disease risk was < 0.3. INTRODUCTION Immunization represents one of the most cost- effective means of disease prevention. Successful immunization programmes can save lives and reduce the severe morbidity of vaccine-preventable diseases; however, for a programme to be successful, it must be monitored closely. Generally, such monitoring involves determination of vaccine coverage and estimation of disease reduction through careful surveillance. In addition, epidemiological assessment of clinical vaccine efficacy has become a useful adjunct to coverage and surveillance for evaluating the impact of immunization programmes. Field evaluation of vaccine efficacy can be particularly useful if surveillance data suggest little impact on disease, despite increasing coverage levels, and if the proportion of reported cases that are vaccine failures appears to be higher than expected. * Requests for reprints should be addressed to Mr Hightower, Division of Bacterial Diseases, Center for Infectious Diseases, Centers for Disease Control (CDC), Atlanta, GA 30333, USA. 1 Division of Bacterial Diseases, Center for Infectious Diseases, CDC. 2 Division of Immunization, Center for Prevention Services, CDC. One common technique used to measure vaccine efficacy (such as that against measles) is a cohort investigation during an outbreak. Whereas the methodology for such studies has been well devel- oped, evaluation of the techniques used to determine the precision of the estimated vaccine efficacy, i.e., confidence intervals, has been incomplete. Infor- mation about the precision of an estimate of vaccine efficacy is important, since narrow limits permit greater confidence in the estimate. The Taylor series method has been recommended for determining 95% confidence intervals for vaccine efficacy (I ) and it has the advantage that the calculations can be done in the field using a simple formula without the need for extensive computer programs. BACKGROUND In a cohort study, it is assumed that the risks of the disease in the vaccinated and unvaccinated samples arise from two independent binomial populations. Let NV vaccinated and NU unvaccinated individuals be included in the study and let ARV and ARU be the risks of disease in the samples of vaccinated and unvaccinated, respectively. The sample estimate of 4863 -99- A. W. HIGHTOWER ET AL. the risk ratio, or relative risk (RR), for the vaccinated is ARV/ARU. Relative risks that are decreasingly less than unity indicate that the vaccine has an increasingly protective effect. The data are displayed in a (2 x 2) table using the following notation: Non- Cases cases Total Risk Vaccinated a b a+b=NV ARV=a/(a+b) Unvaccinated c d c+d=NU ARU=c/(c+d) where RR=ARV/ARU=[a/(a+b)]/[c/(c+d)]. The upper and lower confidence limits on sample estimates are commonly reported in order to gauge precision; however, confidence intervals on estimates of relative risks present special problems since only approximate methods for their calculation are avail- able (2, 3 ). Katz et al. evaluated several such methods for calculating confidence intervals for rela- tive risk estimates (3 ) and recommended the use of a Taylor series approximation for determining the vari- ance of the relative risk. The Taylor series variance approximation is obtained by first noting that varploge(RR)] =var[loge(ARV) -loge(ARU)]. Then, assuming that the risk of disease in the vaccinated and unvaccinated populations are statistically independent, i.e., that the covariance between the two risks is zero, var[loge(ARV)-log,(ARU)] = var[loge(ARV)] +var[loge(ARU)]. The variances of loge (ARV) and loge (ARU) are then estimated using a Taylor,series approximation as described by Kleinbaum et al. (4). With the Taylor series approximate variance and the previous notation, the two-sided 95% confidence limits are given by: exp[logeRR± 1.96 Vaccine efficacy (VE) can be shown to be a simple transformation of the relative risk and is defined as: VE= [(ARU-ARV)/ARU] x 100=(1-RR)X 100. A confidence interval for the relative risk of vaccination can be transformed into one for vaccine efficacy in a few simple steps. First, the two-sided confidence interval for the relative risk of vaccination given by: RRL <RR< RRu is multiplied by -1 to give: -RRL> -RR> -RRu and is added to each element: I-RRL>1-RR>1-RRU. Finally, each element is multiplied by 100 (if the vaccine efficacy is to be expressed as a percentage) to give: (1-RRL) X 100> (l-RR) x 100 > (l-RRu) x 100. This gives the confidence interval for the vaccine efficacy expressed as a percentage. Notice that the lower limit of the confidence interval for the relative risk is used to compute the upper limit of the interval for the vaccine efficacy, and vice versa, i.e., VEL=(1-RRu)X 100 and VEu=(l-RRL)X 100. Typical point estimates of measles vaccine efficacy have been > 80%, which corresponds to relative risks of < 0.2. While prior evaluation of the Taylor series intervals suggests that they give a valid upper 95% confidence limit if the relative risk is >0.25, i.e., the vaccine efficacy is < 75% (3), there has been no evaluation for relative risks less than 0.25. Further- more, there has been no systematic evaluation oftwo- sided confidence intervals. Finally, no studies have taken into account the distributions of vaccinated and unvaccinated subjects that apply to cohort studies of vaccine efficacy. We therefore carried out simulation studies using a variety of conditions likely to occur in vaccine efficacy investigations to determine under what circumstances Taylor series intervals should be used to calculate 95% confidence limits for estimates of vaccine efficacy. METHODS Four factors -vaccine efficacy, disease risk in the unvaccinated population, and the sample sizes of the vaccinated and the unvaccinated populations- deter- mine the width of the 95% Taylor series confidence interval. Simulations were therefore carried out using a wide range ofcommon values for all four factors, as outlined below. The following four values of the population or true vaccine efficacy (VE) were used: 95 %, 90%, 80%, and 60%. For each of these values, disease risks (attack rates) in the unvaccinated popu- lation were selected that ranged from 0.1 to 0.9 (10-90%) in increments of 0.1 (10%). Three situations were evaluated: the first, in which both vaccinated and unvaccinated subjects were common in the population under study; the second, in which vaccinees were common but unvaccinated persons were infrequent; and the third, in which all sample sizes of unvaccinated persons were considered, but only 200 subjects were included in the vaccinated sample. The first situation applies to areas where vaccine coverage is relatively low (around 50%), the 100 ESTIMATING VACCINE EFFICACY: USE OF TAYLOR SERIES CONFIDENCE INTERVALS 101 Table 1. Estimated confidence coefficients obtained by using the Taylor series method for studies with a large number of unvaccinated subjects Risk in the unvaccinated group No. No. unvaccinated vaccinated 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Vaccine efficacy 95% 500 500 0.974 0.976 0.968 0.952 0.947 0.958 0.945 0.953 0.954 500 1000 0.971 0.962 0.954 0.951 0.952 0.955 0.946 0.948 0.946 1000 500 0.971 0.976 0.966 0.959 0.957 0.952 0.948 0.951 0.954 1000 1000 0.975 0.959 0.947 0.951 0.951 0.953 0.950 0.953 0.949 Vaccine efficacy 90% 500 500 0.973 0.955 0.954 0.950 0.943 0.952 0.950 0.950 0.954 500 1000 0.954 0.954 0.953 0.956 0.954 0.949 0.947 0.946 0.948 1000 500 0.975 0.959 0.951 0.953 0.950 0.953 0.945 0.948 0.943 1000 1000 0.958 0.949 0.951 0.960 0.956 0.950 0.951 0.945 0.947 Vaccine efficacy 80% 500 500 0.952 0.949 0.957 0.950 0.945 0.952 0.951 0.946 0.950 500 1000 0.951 0.956 0.950 0.956 0.953 0.952 0.949 0.948 0.953 1000 500 0.959 0.948 0.950 0.956 0.954 0.949 0.953 0.948 0.948 1000 1000 0.951 0.955 0.955 0.950 0.957 0.954 0.947 0.949 0.954 Vaccine efficacy 60% 500 500 0.952 0.953 0.945 0.950 0.948 0.949 0.951 0.946 0.944 500 1000 0.953 0.946 0.956 0.949 0.954 0.951 0.946 0.956 0.952 1000 500 0.953 0.943 0.947 0.948 0.952 0.953 0.951 0.945 0.948 1000 1000 0.953 0.952 0.952 0.947 0.955 0.948 0.953 0.948 0.958 second to areas with high vaccine coverage (> 90%), while the third evaluates the effect of having a relatively small sample size in the group that has the lowest disease risk. For each combination of vaccine efficacy and risk in the unvaccinated, we examined sample sizes for the unvaccinated (NU) of 10, 20, 50, 100, 500, and 1000, while for the vaccinated group sample sizes (NV) of 200, 500, and 1000 were taken. For each combination of vaccine efficacy, risk in the unvaccinated, and sample size in each group we generated 5000 random (2 x2) tables. The independ- ent pairs of random binomial observations were generated on a computera using the Statistical Analy- sis System (SAS). Since the number of vaccinated (NV) and unvaccinated (NU) subjects were fixed, we generated a random binomial observation for the a and c cells of each table (the number of cases in the vaccinated and unvaccinated groups) and obtained the b and d cells (the noncases in each group) by subtraction. As a result, each table could have had an observed value of the vaccine efficacy and ARU that were different from the population parameters, i.e., the true values. We next calculated the approximate upper and lower 95% confidence limits from the a IBM 3081. observed table and checked whether it contained the population vaccine efficacy. Finally, for the 5000 tables we determined the proportion of times that the confidence interval contained the population vaccine efficacy, i.e., the observed confidence coefficient whose expectation value was 0.95 (or 95%). If the proportion was much less than 95 %, the confidence intervals were, on the average, too narrow, while if it was much greater than 95% the confidence intervals were too wide (or conservative). If an observed table had no vaccinated cases (a= 0), the upper limit of the confidence interval was taken to be 100% and the lower limit-9999, since the variance was undefined. In contrast, if the table had no unvaccinated cases (c= 0), we used c= 0.5 in the calculations. This followed the procedure described by Katz et al. (3 ). RESULTS For sample sizes that involved large numbers of both vaccinated and unvaccinated subjects, the ob- served confidence coefficients of the approximate confidence intervals are presented in Table 1. Over- 102 A. W. HIGHTOWER ET AL. all, the approximate two-sided 95 % confidence inter- On the other hand, even when the vaccine efficiency vals have observed confidence coefficients that are is 90% or 95 %, if the risk in the unvaccinated group quite close to 95%. If the population vaccine efficacy is > 0.3, the observed confidence coefficient is still is 80% or 60%, the confidence coefficient is always always within 2% of 95%. within 0.7% of the expected value of95 %. However, Table 2 shows the observed confidence coefficients as the vaccine efficacy approaches 100%, and the risk for samples with large numbers of vaccinated but in the unvaccinated decreases, the observed con- small numbers of unvaccinated subjects, a situation fidence coefficients tend to exceed the expected which is common in developed countries, where confidence coefficient. Under these conditions, the immunization levels are high. Regardless of the true confidence intervals are, on the average, too broad. vaccine efficacy, if the risk in the unvaccinated is Table 2. Estimated confidence coefficients obtained by using the Taylor series method for studies with a small number of unvaccinated subjects Risk in the unvaccinated group No. No. unvaccinated vaccinated 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Vaccine efficacy 95% 10 500 0.993 0.976 0.974 0.972 0.962 0.956 0.958 0.951 0.951 10 1000 0.975 0.966 0.957 0.952 0.952 0.947 0.948 0.951 0.956 20 500 1.000 0.992 0.968 0.960 0.958 0.953 0.961 0.956 0.950 20 1000 0.982 0.971 0.960 0.956 0.954 0.959 0.953 0.952 0.955 50 500 0.989 0.971 0.961 0.956 0.954 0.953 0.953 0.961 0.945 50 1000 0.984 0.957 0.951 0.953 0.951 0.952 0.950 0.951 0.949 100 500 0.982 0.972 0.963 0.958 0.956 0.948 0.951 0.949 0.954 100 1000 0.971 0.952 0.948 0.952 0.950 0.946 0.953 0.946 0.950 Vaccine efficacy 90% 10 500 0.978 0.966 0.951 0.953 0.948 0.955 0.944 0.949 0.955 10 1000 0.953 0.952 0.949 0.945 0.943 0.934 0.929 0.919 0.937 20 500 0.987 0.972 0.961 0.957 0.954 0.951 0.959 0.957 0.951 20 1000 0.964 0.962 0.960 0.955 0.954 0.948 0.950 0.948 0.953 50 500 0.986 0.951 0.957 0.951 0.949 0.958 0.955 0.957 0.955 50 1000 0.971 0.958 0.952 0.950 0.954 0.954 0.944 0.954 0.953 100 500 0.976 0.952 0.952 0.953 0.952 0.953 0.955 0.955 0.949 100 1000 0.959 0.950 0.952 0.950 0.959 0.954 0.946 0.946 0.949 Vaccine efficacy 80% 10 500 0.959 0.954 0.949 0.941 0.938 0.933 0.930 0.924 0.923 10 1000 0.944 0.944 0.942 0.938 0.932 0.928 0.913 0.894 0.875 20 500 0.961 0.960 0.960 0.953 0.949 0.944 0.945 0.948 0.951 20 1000 0.959 0.960 0.957 0.942 0.946 0.945 0.940 0.935 0.936 50 500 0.973 0.955 0.951 0.954 0.950 0.950 0.958 0.952 0.950 50 1000 0.968 0.954 0.952 0.953 0.949 0.951 0.939 0.950 0.955 100 500 0.955 0.957 0.951 0.948 0.955 0.953 0.946 0.950 0.953 100 1000 0.952 0.953 0.954 0.951 0.956 0.954 0.957 0.948 0.954 Vaccine efficacy 60% 10 500 0.949 0.947 0.948 0.938 0.938 0.935 0.906 0.891 0.856 10 1000 0.945 0.954 0.948 0.942 0.933 0.935 0.890 0.899 0.753 20 500 0.952 0.963 0.952 0.946 0.946 0.936 0.938 0.932 0.928 20 1000 0.959 0.956 0.955 0.946 0.941 0.942 0.932 0.922 0.897 50 500 0.966 0.955 0.956 0.948 0.956 0.950 0.944 0.955 0.952 50 1000 0.963 0.952 0.949 0.949 0.949 0.949 0.946 0.945 0.948 100 500 0.959 0.951 0.952 0.950 0.958 0.950 0.953 0.948 0.949 100 1000 0.948 0.947 0.952 0.950 0.951 0.950 0.946 0.948 0.945 ESTIMATING VACCINE EFFICACY: USE OF TAYLOR SERIES CONFIDENCE INTERVALS >0.3 and NU >50, the observed confidence coef- nated group has the lowest disease risk, the number of ficient is reasonably close, i.e., within 2 %, to 95%. expected cases is very low for the higher vaccine However, the observed confidence coefficients efficacies. In the most extreme case, when the true become considerably greater than 95 % as the true vaccine efficacy is 95 % and the disease risk is 0.1 in vaccine efficacy increases and the disease risk in the unvaccinated group, only one case is expected in the unvaccinated decreases, and both NU and NV the vaccinated sample of 200, since their disease risk become small. Under these circumstances, the con- is only 0.005. This situation becomes less decisive fidence intervals are too broad. In contrast, the and the performance of the Taylor series confidence observed confidence coefficient drops considerably intervals improves as either the disease risk increases below 95 %, which implies that the confidence inter- or the true vaccine efficacy decreases. vals are too narrow, as the true vaccine efficacy decreases, and the disease risk in the unvaccinated increases, and NU becomes small. The problem is more marked the larger the value of NV. DISCUSSION The effect of decreasing the number of vaccinated subjects in the study to 200 is shown for various The goal of the study was to evaluate the perform- sample sizes in Table 3; not surprisingly, the con- ance of the Taylor series confidence intervals for fidence intervals become broader. Since the vacci- determining vaccine efficacy (and the corresponding Table 3. Estimated confidence coefficients obtained by using the Taylor series method for studies with 200 vaccinated subjects Risk in the unvaccinated group No. unvaccinated 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Vaccine efficacy 95% 10 1.000 1.000 0.999 0.995 0.990 0.981 0.974 0.974 0.960 20 1.000 0.990 0.983 0.975 0.981 0.976 0.972 0.962 0.960 50 0.981 0.979 0.976 0.976 0.972 0.972 0.967 0.956 0.956 100 0.976 0.972 0.975 0.971 0.973 0.974 0.968 0.965 0.960 500 0.973 0.969 0.980 0.973 0.975 0.971 0.975 0.961 0.960 1000 0.974 0.968 0.979 0.967 0.978 0.972 0.972 0.961 0.958 Vaccine efficacy 90% 10 0.999 0.993 0.977 0.973 0.969 0.958 0.960 0.949 0.957 20 1.000 0.992 0.979 0.965 0.956 0.962 0.955 0.957 0.961 50 0.983 0.976 0.970 0.960 0.961 0.954 0.952 0.954 0.947 100 0.978 0.972 0.973 0.963 0.956 0.961 0.951 0.952 0.950 500 0.977 0.971 0.974 0.959 0.956 0.953 0.952 0.950 0.957 1000 0.976 0.972 0.970 0.959 0.955 0.955 0.952 0.944 0.954 Vaccine efficacy 80% 10 0.977 0.962 0.960 0.957 0.963 0.951 0.953 0.949 0.950 20 0.989 0.973 0.962 0.960 0.953 0.953 0.951 0.945 0.950 50 0.989 0.962 0.955 0.957 0.955 0.959 0.953 0.950 0.948 100 0.973 0.960 0.958 0.957 0.946 0.953 0.952 0.950 0.953 500 0.974 0.959 0.955 0.949 0.954 0.951 0.949 0.951 0.946 1000 0.974 0.961 0.957 0.955 0.948 0.949 0.951 0.952 0.952 Vaccine efficacy 60% 10 0.956 0.957 0.958 0.951 0.947 0.936 0.930 0.930 0.937 20 0.969 0.963 0.956 0.953 0.947 0.946 0.951 0.947 0.951 50 0.970 0.959 0.953 0.956 0.952 0.956 0.952 0.949 0.951 100 0.966 0.954 0.951 0.955 0.957 0.956 0.954 0.953 0.954 500 0.957 0.948 0.947 0.951 0.950 0.949 0.949 0.951 0.948 1000 0.958 0.959 0.945 0.950 0.951 0.950 0.956 0.954 0.942 103 A. W. HIGHTOWER ET AL. relative risks) under a wide variety of circumstances in order to define the conditions under which the tech- nique will perform reliably. Investigators should consider the methodology to be used for calculating confidence intervals during the planning stages of a study. Once sample size estimates are obtained using the desired Type I and Type II errors, along with some estimate of the true vaccine efficacy, all of the information needed to make a decision is available. Unfortunately, in the absence of data an investigator is obligated during the planning stages to "guess" the true vaccine efficacy in order to assess how many subjects are needed and decide what methodology will be used to assess the precision of the estimates. Fortunately, however, under most circumstances the guess need only be approximate, since the confidence intervals give misleading results only under combinations of extreme values of the study factors, e.g., very small unvaccinated group combined with a low vaccine efficacy and a risk that exceeds 0.5. Most investiga- tors will probably know intuitively whether their study could potentially fall into one of these "danger zones. " For applications involving large numbers of both vaccinated and unvaccinated subjects, reliable confi- dence intervals are produced whenever the disease risk in the unvaccinated group is > 0.3 ( > 0.2 for true vaccine efficacies of 90% and >0.1 with true effi- cacies of 80% or 60%). These intervals are therefore widely applicable in such studies. However, the combination of low risk in the unvaccinated group and a high vaccine efficacy results in confidence intervals that are too broad. If a study is to be conducted under these conditions, another method of computing the confidence intervals should be used, such as that described by Bailey (5) (which can also be carried out using a pocket calculator) or by Thomas & Gart (6) (or the refinement in the computing algorithm reported by Mehta et al. (7)). None of these methods has been evaluated under con- ditions appropriate for vaccine efficacy confidence intervals, but of the two, the calculations are easier to perform using Bailey's method (although they are more difficult than the Taylor series approach) and the intervals obtained are very accurate (5). It should be noted, however, that the 95% Taylor series con- fidence interval has at least a 95% chance of contain- ing the true vaccine efficacy. For applications that involve large numbers of vaccinated and smaller numbers of unvaccinated subjects, very reliable confidence intervals are ob- tained when the number of unvaccinated subjects is > 50 and the disease risk in the unvaccinated group is >0.3, regardless of the efficacy of the vaccine. Of course, many other conditions give reliable confi- dence intervals (see Table 2); however, as the risk and the sample size in the unvaccinated group decrease and the vaccine efficacy increases, the con- fidence coefficient becomes considerably greater than 95 %, indicating that the confidence intervals are too broad. Under these conditions, the true variation in the point estimate of the vaccine efficacy will be overestimated. Only rarely were the confidence intervals obtained too narrow. This occurred when the number of unvaccinated persons in a sample was < 20, the number vaccinated was > 500, the attack rate in the unvaccinated was 50-90%, and the true vaccine efficacy was < 80%. Under these conditions, other methods for calculating 95% confidence intervals should be considered (5-7). As shown in Table 3, reducing the sample size in the vaccinated group tends to increase the breadth of the confidence intervals. Ideally, the estimated con- fidence coefficient should not be a function of sample size, but that this is not the case for an approximate procedure under extreme conditions is not surprising. Confidence intervals calculated using the Taylor series method should be used very carefully for studies that involve a true vaccine efficacy of > 95% or a true efficacy of approximately 90% and a disease risk of < 0.3, since the true variation will then be overestimated. In contrast, for studies that involve a true vaccine efflcacy of 80%, the intervals will be too broad when the disease risk is < 0. 1. Otherwise, the confidence intervals calculated using the Taylor series method are quite accurate. Table 4 summarizes the recommendations for the use of Taylor series 95% confidence intervals. Shown are the general conditions under which the intervals obtained are usable or specific circumstances under which alternative methods should be considered. The performance of the confidence intervals under con- ditions not covered in Table 4 were too variable to make blanket recommendations and, if a study is planned under such conditions, Tables 1, 2, or 3 should be consulted to assess whether or not the Taylor series intervals should be used. The results we have reported show that, under virtually all conditions examined, the 95% Taylor series confidence limits determined are never too narrow if the unvaccinated group has > 50 subjects. Investigators should therefore note that the intervals obtained will not underestimate the variation in their estimates and that the calculated interval has at least a 95% chance of including the true vaccine efficacy. However, under extreme conditions, e.g., small sample size for the unvaccinated combined with low risk and a high true vaccine efficacy or a high risk and a low true vaccine efficacy, the intervals may be either too broad or too narrow, respectively. 104 ESTIMATING VACCINE EFFICACY: USE OF TAYLOR SERIES CONFIDENCE INTERVALS 105 Table 4. Recommendations for the use of 95% Taylor series confidence levels Sample category' Risk in the unvaccinated True vaccine efficacy (%) Results and recommendations 1. NU,> 50, NV>, 500 0.2-0.9 60-90 Taylor series approximation NU, 20, NV>500 0.3-0.9 95 provides accurate intervals All NU, NV = 200 0.2-0.9 60-80 All NU, NV= 200 0.4-0.9 90 2. All NU, NV> 500 0.1 and 0.2 95 Intervals are too broad; will All NU, NV,> 500 0.1 90 affect ability to show good All NU, NV-200 0.1-0.7 95 precision All NU, NV= 200 0.1-0.3 90 3. NU = 10, NV >500 0.5-0.9 80 Intervals are too narrow; NU < 20, NV > 500 0.5-0.9 60 consider alternative methods ' NU - number unvaccinated; NV - number vaccinated. RESUME RECOMMANDATIONS RELATIVES A L'EMPLOI DES INTERVALLES DE CONFIANCE OBTENUS A PARTIR DES SERIES DE TAYLOR POUR L'ESTIMATION DE L'EFFICACITt DES VACCINS Les dtudes d'efficacitd des vaccins effectu6es sur des cohortes constituent un moyen important pour surveiller l'impact des programmes de vaccination. Pour d6terminer la precision des estimations de 1'efficacite des vaccins, il a etd recommande d'appliquer une formule simple pour le calcul des intervalles de confiance, a partir d'une approxi- mation de la variance obtenue au moyen des series de Taylor. Pour 6valuer la qualit6 des intervalles de confiance ainsi calcules, nous avons effectue une simulation des valeurs de l'efficacit6 des vaccins auxquelles on peut normalement s'attendre, du risque de contracter la maladie dans la population non vaccinee, ainsi que de la taille des echantillons des groupes vaccines et non vaccines a etudier. Dans le premier cas, l'echantillon du groupe vaccine representait 500 ou 1000 personnes, tandis que celui du groupe non vaccind allait de 10 a 1000 personnes. Les intervalles de confiance calcules etaient en general assez exacts pour toutes les valeurs r6elles de l'efficacite du vaccin examinees, lorsque la taille de l'echantillon du groupe non vaccine etait > 50 et le risque de maladie situe entre 0,3 et 0,9. En revanche ces intervalles etaient trop dtroits lorsque les trois situations suivantes s'ajoutaient: le nombre de personnes non vaccinees etait faible (10 ou 20), l'efficacit6 rdelle du vaccin 6tait relativement faible (60% ou 80%) et le risque de contracter la maladie se situait entre 0,5 et 0,9. D'autre part, avec une efficacite reelle du vaccin elevee (90% ou 95%) et un risque de contracter la maladie faible chez les non vaccin6s (0,1 ou 0,2), ces intervalles de confiance dtaient trop grands, en particulier lorsque l'echantillon de population non vaccinee etait inferieur a 50. D'autres simulations dans lesquelles on a utilise un echantillon de population vaccin6e plus petit (<200) ont donnd des intervalles trop grands pour un vaccin efficace a 95% (quelles que soient les valeurs du risque de contracter la maladie) et a 90% lorsque le risque etait <0,3. Dans le present article figurent des recommandations concernant les circonstances dans lesquelles les intervalles de confiance obtenus a partir des series de Taylor sont fiables, et des propositions d'autres mdthodologies pour les situations dans lesquelles ces intervalles ne conviennent pas. REFERENCES 1. ORENSTEIN, W. A. ET AL. Field evaluation of vaccine efficacy. Bulletin of the World Health Organization, 63: 1055-1068 (1985). 2. GART, J. J. Statistical analyses of the relative risk. Environmental health perspectives, 32: 157-167 (1979). 3. KATZ, D. ET AL. Obtaining confidence intervals for the risk ratio in cohort studies. Biometrics, 34: 469-474 (1978). 4. KLEINBAUM, D. G. ET AL. Epidemiologic research, principles and quantitative methods. Lifetime Learning Publications, Belmont, CA, 1982. 5. BAILEY, B. J. R. Confidence limits to the risk ratio. Biometrics, 43: 201-206 (1987). 6. THOMAS, D. G. & GART, J. J. A table of exact confidence limits for differences and ratios of two proportions and their odds ratio. Journal of the American Statistical Association, 72: 73-76 (1977). 7. MEHTA, C. R. ET AL. Computing an exact confidence interval for the common odds ratio in several 2x2 contingency tables. Journal ofthe American Statistical Association, 80: 969-973 (1985).