 Research
 Open access
 Published:
Correction of vaccine effectiveness derived from testnegative case–control studies
BMC Medical Research Methodology volume 23, Article number: 137 (2023)
Abstract
Background
Determining the vaccine effectiveness (VE) is an important part of studying every new vaccine. Testnegative case–control (TNCC) studies have recently been used to determine the VE. However, the estimated VE derived from a TNCC design depends on the test sensitivity and specificity. Herein, a method for correction of the value of VE derived from a TNCC study is presented.
Methods
An analytical method is presented to compute the corrected VE based on the sensitivity and specificity of the diagnostic test utilized. To show the application of the method proposed, a hypothetical TNCC study is presented. In this in silico study, 100 000 individuals referring to a healthcare system for COVID19like illness were tested with diagnostic tests with sensitivities of 0.6, 0.8, and 1.0, and specificities ranging from 0.85 to 1.00. A vaccination coverage of 60%, an attack rate of 0.05 for COVID19 in unvaccinated group, and a true VE of 0.70, were assumed. In this simulation, a COVID19like illness with an attack rate of 0.30 could also affect all the studied population regardless of their vaccination status.
Results
The observed VE ranged from 0.11 (computed for a test sensitivity of 0.60 and specificity of 0.85) to 0.71 (computed for a test sensitivity and specificity of 1.0). The mean computed corrected VE derived from the proposed method was 0.71 (the standard deviation of 0.02).
Conclusions
The observed VE derived from TNCC studies can be corrected easily. An acceptable estimate for VE can be computed regardless of the diagnostic test sensitivity and specificity used in the study.
Background
Vaccines have had a significant impact on global health. Some vaccines provide lifelong protection against infections; others confer temporary protection. Many circulating pathogens change over time, which affects the effectiveness of vaccines made against them [1, 2], which is why the effectiveness of most vaccines (e.g., influenza vaccine) needs to be updated regularly [3]. Determining and regular monitoring of vaccine effectiveness (VE) is thus an integral part of studying every vaccine.
VE is a measure reflecting how well a vaccine prevents illness, hospitalization, or death in those who were vaccinated compared to unvaccinated individuals [4]. It is commonly expressed as a percentage reduction in the risk (incidence) of the disease in vaccinated vs. unvaccinated people. For example, if the incidence of a certain disease is 0.05 in unvaccinated people and vaccination decreases it to 0.015, then the vaccination decreases the risk by 70% (i.e., 0.70), hence, the corresponding VE is 0.70. Mathematically, VE can be calculated as follows [5, 6]:
where AR_{unvac} and AR_{vac} are the attack rates of the infection in the unvaccinated and vaccinated individuals, and RR is the relative risk. Observational studies are commonly used to determine the VE [7]. A cohort study is the only study design that can accurately provide the AR of the disease of interest in the vaccinated and unvaccinated groups and thus, is the best type of observational studies for calculation of the VE. However, given the low AR for many infectious disease, the odds ratio (OR) derived from a case–control study can be considered an acceptable estimate for the RR [8]. Therefore, a case–control study can also be used for determination of the VE [7].
Over the recent years, testnegative case–control (TNCC) studies have commonly been used to determine the VE [3]. Technically, a TNCC study has a case–control design except that the way the cases and controls are recruited is different. For example, to determine the VE of a new vaccine developed against SARSCoV2 using a TNCC design, cases and controls are selected from a cohort of patients attending a healthcare center because of a COVID19like illness; those who are tested positive for SARSCoV2 are considered “cases;” the remaining with a negative test, “controls” (Table 1) [5, 9]. The VE is then [7]:
To have a valid VE, a valid OR is needed, which in turn necessitates equality of the ORs computed for those who seek medical care and for those who do not seek medical care [10], that is (Table 1):
TNCC design is relatively cheaper and faster to conduct than cohort and traditional case–control studies [9]. Nonetheless, the estimated VE value derived from a TNCC design depends on the test sensitivity (Se), the probability that a diseased person becomes testpositive, and more seriously on the test specificity (Sp), the probability that a diseasefree person becomes testnegative [7, 11, 12]. But, the Se and Sp of most diagnostic tests are not 1.0 (i.e., 100%); there are almost always falsepositive and falsenegative results that cause misclassification problem [11, 13]. Herein, it is meant to present a method for computation of the VE based on the results of a TNCC design regardless of the Se and Sp of the diagnostic test used in the study.
Methods
The proposed correction method
Suppose that a TNCC study was conducted to determine the effectiveness of a new vaccine against SARSCoV2 and tested a group of patients attended a healthcare center with COVID19like illness. Also, suppose that the results in their parametric form are presented in Table 1. Let us focus on a single row of Table 1, for instance, those vaccinated. If all those who attended the healthcare center were considered a cohort of people with COVID19like illness, then the apparent prevalence of COVID19 in the vaccinated patients who sought medical care is [14]:
Note that pr is a true estimation of COVID19 prevalence neither in the whole population nor in patients who sought medical care, as the diagnostic test used for the diagnosis of COVID19 was presumably not perfect; there were falsepositive and falsenegative results. The true prevalence (π), the prevalence had a perfect diagnostic test with a Se and Sp of 1.0 (i.e., 100%) been used, is [14]:
where o represents odds corresponding to pr, a/b. The true odds (ω) of COVID19 in the vaccinated patients who sought medical care is then:
In the same way, the odds in the unvaccinated patients who sought medical care (c/d), can be derived. From Eq. 2, the observed VE is:
Using Eq. 6, the corrected VE is then:
where the subscripts “vac” and “unvac” represent the variable in the vaccinated and unvaccinated patients who sought medical care for the COVID19like illness, respectively. Variance of the corrected VE can be computed too (see Supplementary Materials).
A hypothetical in silico case study
Let us examine the results of the application of the above scenario in an in silico study. Suppose that we want to conduct a TNCC study on a sample of 100 000 individuals who sought medical care for a COVID19like illness. Let 60% of the study population had been vaccinated (vaccine coverage) and that the true VE be 0.70. Assume diagnostic tests with different combinations of Se (0.6, 0.8, and 1.0) and Sp values (ranging from 0.85 to 1.00) for the diagnosis of SARSCoV2 were used. Furthermore, suppose that the test Se and Sp were not different in vaccinated and unvaccinated groups, that COVID19like illness affects people living in the study community independent of whether they have already been infected with SARSCoV2 or not; that the SARSCoV2 has an AR of 5% in unvaccinated individuals [15]; and that the AR of the COVID19like illness is 30% (consistent with the AR of noninfluenza flulike illness seen during a cold season) [16]. Moreover, to make things simple, assume that the AR of the COVID19like illness does not depend on the vaccination status of studied people, duration since vaccination, age, and other variables. Severity of the COVID19 outcomes (e.g., needing hospitalization or admission in intensive care units) was not taken into account in the current in silico study, as it evidently does not significantly affect the healthcareseeking behavior of people so that Eq. 3 holds, regardless of the disease severity [17]. A piece of code developed in R (R software version 4.1.0, R Project for Statistical Computing) was used for the simulation (see Supplementary Materials).
Results and Discussion
Figure 1 shows the apparent and corrected VE values for various combinations of test Se and Sp values used in the in silico TNCC study. The observed VE ranged from 0.11 (computed for a test Se of 0.60 and Sp of 0.85) to 0.71 (computed for a test Se and Sp of 1.0, a perfect test). The mean computed corrected VE derived from the proposed method (Eq. 8) was 0.71 (the standard deviation of 0.02); the corrected value ranged from 0.67 to 0.76. The variation observed was probably attributed to the sampling error in the simulation.
The gross difference between the apparent and corrected VE was due to misclassification of patients [13]. The falsepositive rate could be decreased by increasing the test Sp by changing the cutoff value for tests with continuous results [11]. Should a perfect test (Se and Sp of 1.0) have been used instead, no false results occurred at all and the apparent VE was equal to the true VE (Fig. 1). The corrected values computed were just a little bit different from the true VE of 0.70. As it has been shown earlier [7], the test Sp is much more important than the test Se; if a test with a very high Sp is used, the observed VE is a good estimate of the correct value regardless of the test Se, whereas it is not correct for a test with even a Se of 1.0 (Fig. 1).
Table 2A shows one example of the data generated in the simulation. Using Eq. 2, the apparent VE is 0.18. Plugging in the values in Eq. 8 gives a corrected VE of 0.69. The gross difference between the apparent and corrected VE is due to the presence of a large number of people with falsepositive results (for using a test with lower Sp); 5683 in vaccinated and 3206 in unvaccinated groups (Table 2A). The falsepositive rate could be decreased by increasing the test Sp by changing the cutoff value for tests with continuous results [11]. Should a perfect test have been used instead, no false results occurred at all and the apparent VE was equal to the true VE (Table 2B). Using Eq. 8, the corrected VE, the value if a perfect test would have been used, can be calculated.
The corrected value of 0.71 in this case, as well as the mean corrected VE of 0.71 derived from the simulation, is a little bit higher than the true VE of 0.70. This is because of using OR derived from TNCC design, which is only an estimation of RR (Eq. 2). This is in fact true for any case–control studies [8]. However, as long as the AR of the disease of interest is small, OR is an acceptable estimate for RR.
The vaccine coverage in the population was 0.60 (Table 2). The disease AR was 3% in the whole study population (Table 2B). This estimation was correct, because the AR is different in unvaccinated and vaccinated group. Given the coverage of vaccination, the true VE, and the AR in unvaccinated and vaccinated groups, the AR in the whole population is:
where n is the population size, and Coverage, the vaccination coverage. Plugging in the values (AR_{unvac} of 0.05, Coverage of 0.60, and VE of 0.70) gives an observed AR of 3%.
This study had some limitations. Assuming that the test Se and Sp were not different in vaccinated and unvaccinated groups, that COVID19like illness affects people living in the study community independent of whether they have already been infected with SARSCoV2 or not, that vaccination does not affect the disease severity in COVID19 breakthrough infections [18], and that the AR of the COVID19like illness does not depend on the vaccination status of studied people, duration since vaccination, age, and other variables, might be oversimplification of the situation. More complex simulations should be designed to assess the possible effects of these factors.
Although TNCC design may diminish the effect of many confounding variables and selection bias attributable to differential recall of the exposure compared with traditional case–control design, it cannot completely eliminate the effects of all confounders [9, 19]. TNCC has the advantage over cohort and traditional case–control studies in that it requires fewer resources and can be conducted within a short period [7]. TNCC and traditional case–control study basically share the same design and thus expectedly have similar biases — for instance, both designs provide OR, as an estimate for RR, a presumption that is only true with low AR values [8].
Conclusions
It was shown that the correct value of VE can be computed regardless of the Se and Sp of the diagnostic test to be used in a TNCC study. The computed value and its precision depend on the odds of a positive test in vaccinated and unvaccinated patients who sought medical care and the test Se and Sp (Eq. 8) as well as their variance (see Supplementary Materials). Therefore, although the Se and Sp of the test utilized might not be important, their precisions are. The Se and Sp of the test to be used in TNCC studies are better to be estimated in large validity studies. So far, researchers had to utilize highly specific (and most often, sensitive) tests in TNCC designs to come up with an acceptable estimate for the VE. Employing the proposed method, it is just enough to use tests with known Se and Sp values, no matter how much they are. TNCC is not only used to determine VE, but also has other applications including risk assessment in other settings such as antibiotic resistance [20], and venous thrombosis [21], to name only a few. The proposed correction method may be applied to the results of these studies too.
Availability of data and materials
All data generated or analyzed during this study as well as the R codes are included in this published article and its supplementary information files.
Abbreviations
 AR :

Attack rate
 RR :

Relative risk
 OR :

Odds ratio
 VE :

Vaccine effectiveness
 TNCC:

Testnegative case–control
 SARSCoV2:

Severe acute respiratory syndrome coronavirus 2
 COVID19:

Coronavirus disease 2019
 Se :

Sensitivity
 Sp :

Specificity
 pr :

Prevalence
 π :

True prevalence
References
Eguia RT, Crawford KHD, StevensAyers T, KelnhoferMillevolte L, Greninger AL, Englund JA, Boeckh MJ, Bloom JD. A human coronavirus evolves antigenically to escape antibody immunity. PLoS Pathog. 2021;17(4): e1009453.
Volz E, Hill V, McCrone JT, Price A, Jorgensen D, O’Toole A, Southgate J, Johnson R, Jackson B, Nascimento FF, et al. Evaluating the effects of SARSCoV2 spike mutation D614G on transmissibility and pathogenicity. Cell. 2021;184(1):6475 e11.
Chua H, Feng S, Lewnard JA, Sullivan SG, Blyth CC, Lipsitch M, Cowling BJ. The use of testnegative controls to monitor vaccine effectiveness: a systematic review of methodology. Epidemiology. 2020;31(1):43–64.
Vaccine efficacy, effectiveness and protection [https://www.who.int/newsroom/featurestories/detail/vaccineefficacyeffectivenessandprotection].
De Serres G, Skowronski DM, Wu XW, Ambrose CS. The testnegative design: validity, accuracy and precision of vaccine efficacy estimates compared to the gold standard of randomised placebocontrolled clinical trials. Euro Surveill. 2013;18(37):20585.
Jackson ML, Rothman KJ. Effects of imperfect test sensitivity and specificity on observational studies of influenza vaccine effectiveness. Vaccine. 2015;33(11):1313–6.
Habibzadeh F, Habibzadeh P, Yadollahie M. On measuring vaccine effectiveness with observational study designs. Acta Med Acad. 2022;51(2):134–46.
Schulz KF, Grimes DA. Casecontrol studies: research in reverse. Lancet. 2002;359(9304):431–4.
Sullivan SG, Tchetgen Tchetgen EJ, Cowling BJ. Theoretical basis of the testnegative study design for assessment of influenza vaccine effectiveness. Am J Epidemiol. 2016;184(5):345–53.
Jackson ML, Nelson JC. The testnegative design for estimating influenza vaccine effectiveness. Vaccine. 2013;31(17):2165–8.
Habibzadeh F, Habibzadeh P, Yadollahie M. On determining the most appropriate test cutoff value: the case of tests with continuous results. Biochem Med (Zagreb). 2016;26(3):297–307.
Altman DG, Bland JM. Diagnostic tests. 1: sensitivity and specificity. BMJ. 1994;308(6943):1552.
Shi M, An Q, Ainslie KEC, Haber M, Orenstein WA. A comparison of the testnegative and the traditional casecontrol study designs for estimation of influenza vaccine effectiveness under nonrandom vaccination. BMC Infect Dis. 2017;17(1):757.
Habibzadeh F, Habibzadeh P, Yadollahie M. The apparent prevalence, the true prevalence. Biochem Med (Zagreb). 2022;32(2): 020101.
Orenstein EW, De Serres G, Haber MJ, Shay DK, Bridges CB, Gargiullo P, Orenstein WA. Methodologic issues regarding the use of three observational study designs to assess influenza vaccine effectiveness. Int J Epidemiol. 2007;36(3):623–31.
Hoberman A, Greenberg DP, Paradise JL, Rockette HE, Lave JR, Kearney DH, Colborn DK, KursLasky M, Haralam MA, Byers CJ, et al. Effectiveness of inactivated influenza vaccine in preventing acute otitis media in young children: a randomized controlled trial. JAMA. 2003;290(12):1608–16.
Lewnard JA, Patel MM, Jewell NP, Verani JR, Kobayashi M, Tenforde MW, Dean NE, Cowling BJ, Lopman BA. Theoretical framework for retrospective studies of the effectiveness of SARSCoV2 Vaccines. Epidemiology. 2021;32(4):508–17.
Gupta RK, Topol EJ. COVID19 vaccine breakthrough infections. Science. 2021;374(6575):1561–2.
Vandenbroucke JP, Pearce N. Testnegative designs: differences and commonalities with other casecontrol studies with “Other Patient” controls. Epidemiology. 2019;30(6):838–44.
Lipsitch M. Measuring and interpreting associations between antibiotic use and penicillin resistance in Streptococcus pneumoniae. Clin Infect Dis. 2001;32(7):1044–54.
Bloemenkamp KW, Rosendaal FR, Buller HR, Helmerhorst FM, Colly LP, Vandenbroucke JP. Risk of venous thrombosis with use of current lowdose oral contraceptives is not explained by diagnostic suspicion and referral bias. Arch Intern Med. 1999;159(1):65–70.
Acknowledgements
Not applicable.
Funding
No financial support.
Author information
Authors and Affiliations
Contributions
FH has done all the relevant things from the conception of idea, to design, data analysis, drafting and substantial editing of the manuscript, and developing the computer codes.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
This study was a methodology study proposing a new mathematical technique for the correction of an index. It did not involve any humans or animals. The dataset used in the study was generated in silico and thus ethical approval and obtaining consent were not applicable.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
About this article
Cite this article
Habibzadeh, F. Correction of vaccine effectiveness derived from testnegative case–control studies. BMC Med Res Methodol 23, 137 (2023). https://doi.org/10.1186/s12874023019620
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s12874023019620