Добавил:
kiopkiopkiop18@yandex.ru t.me/Prokururor I Вовсе не секретарь, но почту проверяю Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:

Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_2767_Библиотеки_им_академика_М_И_Перельмана

.pdf
Скачиваний:
0
Добавлен:
31.08.2026
Размер:
25 Мб
Скачать
10 Undertaking meta- analysis
,
.
ee
https://t.me/medicina_free
Box 10.3.a SAS output of HSROC model parameters
Parameter Estimates
Standard
Parameter
alpha
theta
beta
s2ua
s2ut
Estimate
2.6016
–0.4370
0.2267
1.3014
0.5423
Error DF t ValuePr > |t| 95% Confidence Limits Gradient
0.1862
0.1469
0.1624
0.3046
0.1237
48
48
48
48
48
13.97
–2.98
1.40
4.27
4.39
<.0001
0.0046
0.1691
<.0001
<.0001
2.2273
–0.7323
–0.09978
0.6890
0.2936
2.9759
–0.1417
0.5532
1.9138
0.7909
0.000069
0.000103
0.000039
0.000020
–0.00012
10.3.1 Fitting the HSROC model using SAS
The HSROC model was used to estimate a summary curve using Proc NLMIXED in SAS (see code in Appendix 7 of the online supplementary material (10.S1 Code for undertaking meta- analysis)) to obtain the output shown in Box10.3.a.
The parameter estimates in the red box can be entered into RevMan to draw the sum­mary curve shown in Chapter9, Figure9.4.b; the estimate of the mean for accuracy (Λ, lambda) is 2.6016, −0.4370 for the mean for threshold (Θ, theta), 0.2267 for the shape parameter (β, beta), 1.3014 for the variance of the random effects for accuracy
0.5423 for the variance of the random effects for threshold
2
2
and
Estimation of a summary sensitivity and specificity are not clinically meaningful esti­mates for RF, since the 50 studies used different thresholds for RF. However, the expected sensitivity at a chosen specificity (or vice versa) can be computed from the fitted curve by using the equation given by
logit sensitivity logit 1 specificity .
0.5
The equation can be included in an ESTIMATE statement in NLMIXED (see code in Appendix 7 of the online supplementary material (10.S1 Code for undertaking meta­analysis)). The ESTIMATE statement computes additional estimates as a function of parameter values and produces standard errors and confidence intervals using the delta method. For RF, the median (interquartile range) of specificities from the 50 stud­ies was 0.87 (0.80 to 0.93). These three values of specificity were used in ESTIMATE statements to obtain the corresponding values of sensitivity and their 95% CIs. The esti­mates of logit(sensitivity) with their 95% CIs are presented in the additional estimates table in the SAS output, as shown in the red boxes in Box10.3.b. Inverse transforma­tions of the logit estimates (see equations in Section10.2.1) give the estimates of sensi­tivity and their 95% CIs at the fixed values of specificity.
10.3.2 Bayesian estimation of the HSROC model
10.3.2.1 Specification of the HSROC model in rjags
The HSROC model equations introduced in Chapter9, Section9.4.3, are written out directly in rjags. This section covers the different components of this rjags model, which is the core of the program (comment lines separate the components in Box10.3.c).
268
10.3 Estimation ofa summary curve
)
https://t.me/medicina_free
Box 10.3.b SAS output of additional estimates produced using HSROC model parameters
Additional Estimates
Label Estimate
E(logitSe_sp80)
E(logitSe_sp87)
E(logitSe_sp93)
1.2177
0.8074
0.2608 0.1780 48 1.46 0.1494 0.05
Standard
ErrorDFt Value Pr > |t| Alpha Lower Upper
0.1697 48 7.18
0.1553 48 5.20
<.0001
<.0001
0.8765 1.5589
0.05
0.4952 1.1197
0.05
–0.09713 0.6187
Specificity Sensitivity (95% CI
0.80 0.77 (0.71 to 0.83)
0.87 0.69 (0.62 to 0.75)
0.93 0.56 (0.48 to 0.65)
Box 10.3.c Specification of the HSROC model for Bayesian estimation in rjags
model {
# === LIKELIHOOD === #
for(i in 1:n) {
TP[i] ~ dbin(TPR[i],pos[i]) FP[i] ~ dbin(FPR[i],neg[i])
# === PRIOR DISTRIBUTIONS FOR TPR AND FPR === #
logit(TPR[i]) <- (theta[i] + 0.5*alpha[i])/exp(beta/2) logit(FPR[i]) <- (theta[i] - 0.5*alpha[i])*exp(beta/2) theta[i] ~ dnorm(THETA,prec[2]) alpha[i] ~ dnorm(LAMBDA,prec[1])
}
### === HYPER PRIOR DISTRIBUTIONS === ###
THETA ~ dunif(- 10,10) LAMBDA ~ dunif(- 2,20) beta ~ dunif(- 5,5) for(i in 1:2) { prec[i] ~ dgamma(2.1,2) tau.sq[i] <-
1/prec[i]
tau[i] <- pow(tau.sq[i],0.5)
} }
Likelihood: The likelihood specifies that the observed data in each study, the TP and FP cells, follow a binomial distribution with probability TPR (true positive rate) and FPR (false positive rate), respectively (Box10.3.c).
Prior distributions: The parameters logit TPR and logit FPR are expressed as functions of three additional parameters: the proxy for positivity threshold (theta) for each study, the lnDOR (alpha) for each study and the shape parameter (beta) (β in Chapter9, Section9.4.3), which is assumed to be common across studies. The parameters theta and alpha are assumed to follow hierarchical prior distributions, thus allowing for
269
10 Undertaking meta- analysis
1
,
1
,
Density
LAMBDA
(a)
(b)
(c)
Iteration
0
https://t.me/medicina_free
both within- study and between- study variability. The theta parameters are assumed
to have a normal prior distribution with mean THETA (Θ) and precision prec[2] (
where 2 is referred to as tau.sq[2] in Box10.3.c and is the variance for the theta
parameters). The alpha parameters are assumed to have a normal prior distribution
with mean LAMBDA (Λ) and precision prec[1] (
where 2 is referred to as tau.
2
sq[1] in Box10.3.c and is the variance for the alpha parameters). The parameters
THETA, prec[2], LAMBDA, prec[1] and beta are provided with vague prior distribu-
tions, with the intention that they have a negligible impact on the results.
2
10.3.2.2
Monitoring convergence
Detailed information on how to run the rjags program in Box10.3.c is given in Appendix 6 of the online supplementary material (10.S1 Code for undertaking meta- analysis), Section A6.2. Here, the focus is on interpreting the results of the program when applied to the rheumatoid factor data of Nishimura (2007) introduced in Chapter9. As in the case of the bivariate model (see Section10.2.4.2), begin by examining whether the MCMC algorithm converged. Figure10.3.b shows the diagnostics plots for LAMBDA, but similar plots can be obtained for all parameters. These plots, created using the mcmcplots
3.02.82.72.6 2.92.52.4
0.0 0.5 1. 01.5 2.0
2.0
3.53.0
2.5 LAMBDA
3.0 3.5
Running mean
0500010000 15000
Iteration
2000
270
2.52.0
Figure10.3.b MCMC diagnostics plots for a parameter (LAMBDA) from the HSROC model
30000 3500040000 45000
10.3 Estimation ofa summary curve
https://t.me/medicina_free
Box 10.3.d Rjags output from Bayesian estimation of the HSROC model
Iterations = 6001:16000 Thinning interval = 1 Number of chains = 3 Sample size per chain = 10000
1. Empirical mean and standard deviation for each variable,
plus standard error of the mean: Mean SD Naive SE Time-
series SE LAMBDA 2.6187 0.191119 1.103e- 03 3.787e- 03 THETA - 0.4490 0.152213 8.788e- 04 5.690e- 03 beta 0.2155 0.163026 9.412e- 04 8.912e- 03 tau.sq[1] 1.3781 0.321925 1.859e- 03 2.875e- 03 tau.sq[2] 0.6241 0.139763 8.069e- 04 1.133e- 03
2. Quantiles for each variable:
2.5% 25% 50% 75% 97.5% LAMBDA 2.2434 2.4922 2.6167 2.7448 2.9985 THETA - 0.7518 - 0.5516 - 0.4476 - 0.3460 - 0.1534 beta - 0.1066 0.1069 0.2189 0.3256 0.5348 tau.sq[1] 0.8750 1.1481 1.3378 1.5589 2.1281 tau.sq[2] 0.4034 0.5243 0.6065 0.7036 0.9472
package in R, suggest that the MCMC algorithm has converged, as the results from the three independent MCMC chains (identified by different colours) with different initial values overlap nearly perfectly. The running mean plot in panel (b) shows some discord­ance between the chains when the number of iterations is small, but the values on the y- axis indicate that the apparent differences are in fact very small.
10.3.2.3 Summary statistics and SROC plot
Following successful convergence of the MCMC algorithm, the summary statistics of the parameters of interest can be calculated using a sample from the posterior distribution. Output from the rjags program (see red box in Box10.3.d) gives the estimates for LAMBDA, THETA and beta, and the between- study variance of alpha (tau.sq[1]) and theta (tau.sq[2]). For comparison, see the results in Section10.3.1 that were obtained using a frequentist approach in SAS.
The summary curve on the SROC plot (Figure10.3.c) from the Bayesian estimation using the DTAplots package is very similar to the curve from the frequentist estimation (Chapter9, Figure9.4.b). Study points on the plot were scaled according to sample size. This plot can also be obtained in RevMan by providing the parameter estimates in the red box in Box10.3.d.
271
10 Undertaking meta- analysis
0
Sensitivity
Specificity
0.0
https://t.me/medicina_free
1.
0.80.60.4
0.20.0
1. 0 0.8 0.6 0.4 0.2
Figure10.3.c SROC plot of rheumatoid factor for rheumatoid arthritis
10.3.2.4 Sensitivity analyses
As already noted, Bayesian estimation of the HSROC model typically relies on vague prior distribution functions. Since there is no unique way to specify a vague prior distri­bution, it is important to verify the impact of using an alternative vague prior distribu­tion, as indicated in Section10.2.4.5. In Box10.3.c Gamma priors were used over the precision parameters. An alternative approach would be to use a uniform or half- normal prior distribution over the standard deviation parameters.
10.4 Comparison ofsummary points
The bivariate model is a regression model that can be extended to incorporate covariates (i.e. meta- regression) to compare summary points (see Chapter9, Section 9.4.6.2 and Section9.4.7.2). In Chapter9, Section9.4.6.3, a bivariate meta- regression was used to investigate heterogeneity by assessing differences in the sensitivity and specificity of two different generations of the anti- CCP test for diagnosis of rheumatoid arthritis. Similarly, a bivariate meta- regression was used in Chapter9, Section 9.4.7.3, to compare the accuracy of two imaging modalities – multislice computed tomography (CT) and magnetic resonance imaging (MRI)– for diagnosis of coronary artery disease (CAD). For both analyses, the variances of the random effects were assumed to be equal for the logit sensitivities and the logit specificities.
The assumption of equal variances for the random effects of the logit sensitivities and the logit specificities of different subgroups may be reasonable in many situations when investigating heterogeneity in the accuracy of a single test, but less so when comparing the accuracy of different tests. For ease of reference, subgroups or tests will simply be
272
10.4 Comparison ofsummary points
,,
2
AK
2
Bk
00
https://t.me/medicina_free
referred to as ‘group’ in the model specifications. The model that allows for unequal vari­ances for each group can be written as
2
ABk Bk
(10.1)
2
where μ
and μ
Aik
Aik A A k AK ABk
N
Bik B B k
are the logit(sensitivity) and logit(specificity) for the kth group in the
Bik
vZ
vZ
ith study; Zk represents the study- level covariate; μA estimates the mean logit(sensitivity)
for the referent group (note that for test comparisons this is not the reference standard but another index test); μB estimates the mean logit(specificity) for the referent group;
μA+vAZk estimates the mean logit(sensitivity) for the kth group; and μB+vBZk estimates
the mean logit(specificity) for the kth group. Hence, exp(v
) and exp(vB) estimate the
A
odds ratio for sensitivity and specificity in the kth group relative to the referent group. The parameters cities for the kth group; and σ
and
are the variances for the logit sensitivities and logit specifi-
is the covariance between the logits across studies
ABk
evaluating the group (Takwoingi 2016). The variance–covariance structure in equa­tion10.1 is typically modelled assuming independence between groups. For a binary covariate (e.g. two tests), this variance–covariance matrix can be expressed as
2 A1 A1B1
2
0
2 B1
(10.2)
0
2 B2
The means
Ai
1
N
and
A
B
12B
B
are column vectors of the means of logit sensi-
B
Ai
2 A2 A2B2
~ , with .
1
Bi
2
Bi
12A
A
A
tivities and logit specificities for the two groups. Since test comparisons may include only studies that used a paired design in which individuals received all index tests, the bivariate model can allow for correlation in test performance between tests by estimat­ing all between- study and between- test variability using the following unstructured variance- covariance matrix:
Ai
Ai
Bi
Bi
Note that potential within- study correlation between tests is not taken into account in equation10.3– this would require individual participant data or aggregate data in the form of 2×4 tables of the results of two index tests cross- classified for individuals with and for those without the target condition. Such tables are seldom reported in primary studies. There may be biological/clinical justification for other variants of the variance–covariance matrix in equation10.3, such as assuming independence between the sensitivity of one test and the specificity of another test (i.e.σ
2
1
2 A2 A2B1 A2B2
~ , with .
1
2
A
N
B
A1 A1A2 A1B1 A1B2
2
2 B1 B1B 2
2 B2
A1B2
(10.3)
=0 and σ
A2B1
=0).
273
10 Undertaking meta- analysis
https://t.me/medicina_free
Sections10.4.1,10.4.2 and10.4.3 revisit and extend the analyses presented in sections
9.4.6.3 and 9.4.7.3 to show how to fit models that allow for unequal variances using the variance–covariance matrix expressed in equation 10.2. Section 10.4.4 presents Bayesian estimation of the meta- regression analyses.
10.4.1 Fitting the bivariate model in SAS to compare summary points
The SAS code for the investigation of heterogeneity presented in Chapter 9, Section9.4.6.3, is in Part I of Appendix 8 of the online supplementary material (10.S1 Code for undertaking meta- analysis). This appendix also contains code for meta­regression using MetaDAS. The parameter estimates required to draw the summary points and regions in RevMan can be extracted from the Proc NLMIXED output in Box10.4.a. The variances of the random effects for logit(sensitivity) and logit(specificity), and their covariance, are common for both generations of CCP (see blue box in out­put). For the referent group (CCP1in this example) the mean logit(sensitivity), mean logit(specificity), the corresponding standard errors and covariance are shown in the red boxes in the output. The covariate parameter estimates, se2 and sp2 (i.e. νA and νB), give the expected change in logit(sensitivity) and logit(specificity) for CCP2 rela­tive to CCP1. The mean logit(sensitivity) for CCP2 is thus estimated by msens+se2, and the mean logit(specificity) is estimated by mspec+sp2. These additional estimates and their standard errors and covariance can be obtained as described in Chapter 9, Section9.4.6.3, or using the ESTIMATE statement in Proc NLMIXED, as shown in the
Box 10.4.a SAS output of bivariate meta- regression of CCP generation: equal variances (model 1)
Fit Statistics
274
Parameter Gradient95% Confidence LimitsPr > |t|t ValueDFEstimate
msens
mspec
s2usens
s2uspec
covsesp
se2
sp2
–0.09653
3.4467
0.3598
0.5399
–0.1969
0.9626
–0.4302
-2 Log Likelihood
AIC (smaller is better)
AICC (smaller is better)
BIC (smaller is better)
Parameter Estimates
Standard
Error
0.2203
0.2982
0.1022
0.1802
0.09836
0.2513
0.3377
35
35
35
35
35
35
35
–0.44
11.56
3.52
3.00
–2.00
3.83
–1.27
0.6640
<.0001
0.0012
0.0050
0.0532
0.0005
0.2111
533.4
547.4
549.1
558.6
–0.5438
2.8412
0.1524
0.1742
0.4523
–1.1158
0.3507
4.0522
0.5673
0.9057
0.002825
1.4728
0.2554
0.000317
–0.00005
8.325E-6
0.000159
–0.00004
0.000319
–0.00004
msens
https://t.me/medicina_free
mspec
s2usens
s2uspec
covsesp
se2
sp2
Label
logitsens CCP2
logitspec CCP2
logLR+ CCP1
logLR-CCP1
logLR+ CCP2
logLR-CCP2
–0.02464
–0.00012
–0.00001
–0.00003
–0.04855
Covariance Matrix of Parameter Estimates
0.04854
0.02465
0.004772
Estimate
0.8660
3.0165
2.7355
–0.6147
2.7132
–1.1693
–0.02464
0.08895
–0.00002
–0.00065
0.02463
–0.08834
–0.00012
–0.00002
0.01044
0.002118
–0.00440
0.000693
–0.00005
Additional Estimates
Standard
Error DF t Value Pr > |t| Alpha Lower
0.1209
0.1622
0.2681
0.1018
0.1458
0.08270
35
35
35
35
35
35
–0.00001
0.004772
0.002118
0.03246
–0.00860
–0.00006
–0.00039
7.16
18.59
10.21
–6.04
18.60
–14.14
10.4 Comparison ofsummary points
–0.00003
–0.00065
–0.00440
–0.00860
0.009674
0.000100
–0.00091
<.0001
<.0001
<.0001
<.0001
<.0001
<.0001
–0.04855
0.02463
0.000693
–0.00006
0.000100
0.06317
–0.03160
0.05
0.05
0.05
0.05
0.05
0.05
–0.08834
–0.00005
–0.00039
–0.00091
–0.03160
0.6207
2.6871
2.1913
–0.8213
2.4171
–1.3372
sp2se2covsesps2uspecs2usensmspecmsens
0.02465
0.1141
Upper
1.1114
3.3459
3.2797
–0.4081
3.0093
–1.0014
Covariance Matrix of Additional Estimates
Label
logitsens CCP2
logitspec CCP2
logLR+ CCP1
logLR-CCP1
logLR+ CCP2
logLR-CCP2
Cov1 Cov2 Cov3 Cov4 Cov5 Cov6
0.01461
–0.00697
–0.00001
5.023E-6
–0.00231
–0.00996
–0.00697
0.02632
0.000598
–0.00002
0.02303
0.003674
–0.00001
0.000598
0.07185
–0.00301
0.000566
–0.00002
5.023E-6
–0.00002
–0.00301
0.01035
–0.00002
–2.45E-6
–0.00231
0.02303
0.000566
–0.00002
0.02127
0.000554
–0.00996
0.003674
–0.00002
–2.45E-6
0.000554
0.006840
program in Appendix 8 of the online supplementary material (10.S1 Code for undertak­ing meta- analysis). The additional estimates and their covariance matrix are shown in the green boxes in Box10.4.a. The resulting SROC plot from entering parameter esti­mates, standard errors and covariances into RevMan is shown in Chapter9, Figure9.4.c, and panel (a) of Figure10.4.a.
Part II of Appendix 8 of the online supplementary material (10.S1 Code for undertak­ing meta- analysis) fits the more complex model that allows for unequal variances for generations of anti- CCP (model 1). The regression equation specified in Proc NLMIXED is not in the same format as for model 1 because dummy variables have been created
275
10 Undertaking meta- analysis
https://t.me/medicina_free
(a)(b)
1
0.9
0.8
0.7
0.6
0.5
Sensitivity
0.4
0.3
0.2
0.1
0
10.9
0.8 0.70.6 0.5
Legend
Generation: CCP1 Generation: CCP2
Generation Summary
CCP1 0.48 (0.37 to 0.58) 0.97 (0.95 to 0.98)CCP1 0.48 (0.43 to 0.53)0.97 (0.95 to 0.98)
CCP2 0.70 (0.65 to 0.75) 0.95 (0.94 to 0.97)CCP2 0.71 (0.64 to 0.76)0.95 (0.94 to 0.97)
These results indicate an improvement in sensitivity (P < 0.001), without loss of specificity (P = 0.21).
sensitivity (95% CI)
0.40.3 0.2 0.10
Specificity
Summary specificity (95% CI)
1
0.9
0.8
0.7
0.6
0.5
Sensitivity
0.4
0.3
0.2
0.1
0
0.8 0.70.6 0.5
10.9
Legend
Generation: CCP1 Generation: CCP2
Generation Summary
These results indicate an improvement in sensitivity (P < 0.0001), without loss of specificity (P = 0.19).
sensitivity (95% CI)
0.4 0.3 0.2 0.10
Specificity
Summary specificity (95% CI)
Figure10.4.a SROC plots produced using estimates from (a) model 1: equal variances and
(b)model 2: unequal variances. The dotted region around each summary point (solid circle) is the 95% confidence region while the dashed region is the 95% prediction region
and included for both CCP1 and CCP2. This approach is more straightforward for directly obtaining all parameter estimates and for specifying the variance–covariance structure. The statistical significance of differences between the model that assumed equal vari­ances (model 1) and the model that allowed for unequal variances (model 2) can be assessed using a likelihood ratio test (see Chapter9, Section9.4.6.1). The likelihood ratio Chi2 statistic is the difference in the −2Log likelihood when covariate terms are added or removed from a model. The degrees of freedom used along with the Chi statistic to obtain a P value are the difference in the number of parameters fitted in the two models being compared.
Prior to fitting model 1, it is useful to perform separate meta- analyses for CCP1 and CCP2 to explore whether or not the values of the variances are close or differ substan­tially between the two CCP generations. The output of these subgroup analyses are shown in Box10.4.b. The variances of the random effects for the logit sensitivities (s2usens) are 0.04277 and 0.4830 for CCP1 and CCP2, respectively (see red boxes). This suggests much greater variation in sensitivity across CCP2 studies compared to CCP1 studies, which is also apparent from the scatter of the study points in Figure10.4.a. The variances of the random effects for the logit specificities (s2uspec) also differ, but per­haps not to the extent that alone would justify fitting a model with unequal variances. It was evident from Box10.4.a that the variances of the random effects have reduced, particularly for sensitivity, compared to the analysis without the covariate in Box10.2.a. Nevertheless, it is worth investigating the assumption of equal variances, especially as
2
276
Meta-analysis of CCP1
https://t.me/medicina_free
10.4 Comparison ofsummary points
Box 10.4.b Subgroup analyses of CCP1 and CCP2 generations of anti- CCP test for rheumatoid arthritis
Parameter Estimates
Standard
Parameter
msens
mspec
s2usens
s2uspec
covsesp
Meta-analysis of CCP2
Parameter
msens
mspec
s2usens
s2uspec
covsesp
Estimate
–0.07773
3.4586
0.04277
0.4793
–0.06258
Estimate
0.8725
3.0147
0.4830
0.5661
–0.2326
Error DF t Value Pr > |t| 95% Confidence Limits Gradient
–0.83
0.09375
0.2910
0.03305
0.3201
0.07902
Standard
Error DF t Value Pr > |t| 95% Confidence Limits Gradient
0.1377
0.1655
0.1530
0.2185
0.1303
6
6
6
6
6
Parameter Estimates
27
27
27
27
27
11.89
1.29
1.50
–0.79
6.33
18.21
3.16
2.59
–1.79
0.4387
<.0001
0.2432
0.1850
0.4585
<.0001
<.0001
0.0039
0.0153
0.0854
–0.3071
2.7466
–0.03811
–0.3040
–0.2559
0.5899
2.6750
0.1691
0.1177
–0.4999
0.1517
4.1706
0.1237
1.2626
0.1308
1.1551
3.3543
0.7968
1.0145
0.03468
0.000033
0.000043
–0.00019
–0.00002
–0.00008
0.000025
0.000489
–0.00004
0.000055
0.000114
there are many included studies and so model overfitting is not a concern. In addition to providing insight into whether or not an assumption of equal variances for the two generations is justifiable, the subgroup analyses can guide the choice of starting values. It is advisable not to use the exact values of the parameter estimates from the analyses, but rather values close to the estimates, e.g. in Box10.4.b the estimate for ‘msens’ for CCP1 is −0.07773 and was used to inform a starting value of 0.1 for ‘msens’ in the meta-
regression analysis (see Part II in Appendix 8 of the online supplementary material
(10.S1 Code for undertaking meta- analysis)).
Box10.4.c shows the parameter estimates obtained for CCP1 (red box) and for CCP2 (blue box) and their covariances (green boxes) from fitting model 2. The fit statistics table gives a –2Log likelihood of 524.6, a reduction of 8.8 (533.4−524.6) com­pared with model 1 that assumed equal variances. Hence, there is statistical evidence (Chi2 = 8.8, 3df, P = 0.032) that the variances are unequal for CCP1 and CCP2. The differ­ences in mean logit sensitivity and mean logit specificity (i.e. νA and νB) are shown in Box10.4.d (see red box). The P values based on Wald statistics indicate an improvement in sensitivity (P < 0.0001), without loss of specificity (P = 0.19); there is a similar conclu­sion based on the results of model 1 (Figure10.4.a). Comparing the summary estimates
277