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
81
81
https://t.me/medicina_free
CT
Study
Dewey 2006 Kefer 2005 Langer 2009 Maintz 2007 Pouleur 2008
MRI
Study
Dewey 2006 Kefer 2005 Langer 2009 Maintz 2007 Pouleur 2008
TP
4
5
62
2
6
32
1
2
25
1
2
15
1
7
16
TP
FN
FP
42
7
2
30
4
9
18
8
15
15
1
1
17
0
17
0.94 [0.85, 0.98]
46
0.94 [0.80, 0.99]
12
0.96 [0.80, 1.00]
40
0.94 [0.70, 1.00]
2
0.94 [0.71, 1.00]
53
Sensitivity (95% CI)
TN
0.86 [0.73, 0.94]
39
0.88 [0.73, 0.97]
9
0.69 [0.48, 0.86]
27
0.94 [0.70, 1.00]
3
1.00 [0.80, 1.00]
43
Sensitivity (95% Cl)
lN
FN
FP
Specificity (95% CI)
0.90 [0.79, 0.97]
0.67 [0.41, 0.87]
0.95 [0.84, 0.99]
0.50 [0.07, 0.93]
0.88 [0.77, 0.95]
Specificity (95% CI)
0.95 [0.83, 0.99]
0.50 [0.26, 0.74]
0.64 [0.48, 0.78]
0.75 [0.19, 0.99]
0.72 [0.59, 0.83]
Sensitivity (95% CI) Specificity (95% Cl)
00.2 0.40.6 0.8100.2 0.40.6 0.
Sensitivity (95% CI) Specificity (95% Cl)
00.2 0.40.6 0.8100.2 0.40.6 0.
Figure10.6.b Forest plot of comparative studies of CT versus MRI for coronary artery disease.
Source: Data taken from Schuetz 2010
Box 10.6.a Large gradient values in SAS for bivariate model parameters for CT and MRI comparative studies (model without test type covariate)
Parameter Estimates
Standard
Parameter
msens
mspec
s2usens
s2uspec
covsesp
Estimate
2.0000
1.0000
–111E-14
2.21E-11
–108E-15
Error DF t Value Pr > |t| 95% Confidence Limits Gradient
0.4686
0.2041
0.4240
0.01977
0.0236
4.27
3
0.0163
4.90
3
1.0000
–0.00
3
1.0000
0.00
3
3
-
-
0.5086
0.3504
–1.3493
–0.06291
-
3.4914
1.6496
1.3493
0.06291
-
–6.88008
–25.4401
–0.31355
–188.328
–44.9080
-
To assess the effect of test type on sensitivity and specificity, models with and without the covariate were fitted. The output of fitting a bivariate model without the test type covariate to the data using Proc NLMIXED in SAS showed that the convergence criterion was satisfied. However, all the parameter estimates have large gradient values, indicat­ing poor estimation (Box10.6.a). A model may satisfy a convergence criterion but may be unstable (e.g. changing the starting values results in a change in parameter esti­mates), or have missing standard errors for one or more parameter estimates. Therefore, in addition to meeting a convergence criterion, meta- analysts should check for missing standard errors and if the program output permits, also check that gradient values for all model parameters are very close to zero before concluding that the analysis has suc­cessfully converged.
Several attempts were made to improve estimation of the parameters by specifying different options for fitting the model (see code in Appendix 9 of the online supple­mentary material (10.S1 Code for undertaking meta- analysis)). The options included
298
10.6 Meta- analysis ofsparse data anda typical data sets
https://t.me/medicina_free
Box 10.6.b Improved estimation in SAS of bivariate model parameters for CT and MRI comparative studies by changing model fitting options
Parameter Estimates
Standard
Parameter
msens
mspec
s2usens
s2uspec
covsesp
Estimate
2.2554
1.3083
0.006812
0.4138
–0.03331
Error DF t Value Pr > |t| 95% Confidence Limits Gradient
0.3569
0.3375
0.5582
0.3633
0.1724
0.0080
6.32
3
0.0304
3.88
3
0.9910
0.01
3
0.3374
1.14
3
0.8592
–0.19
3
1.1194
0.2342
–1.7695
–0.7424
–0.5821
3.3913
2.3823
1.7831
1.5701
0.5155
0.000057
0.000543
1.70698
–0.02091
–0.21665
increasing the number of quadrature points; changing the optimization technique from the default quasi- Newton technique to the Newton- Raphson technique; trying different starting values, e.g. using a grid search; and setting boundary constraints for the vari­ance parameters (i.e. variance 0). Imposing boundary constraints can reduce the risk of convergence problems. If boundary constraints are triggered for variance parame­ters, then estimation of these parameters is truncated at zero. The same options explored here for fitting the bivariate model can be applied when fitting the HSROC model in SAS. For the CT versus MRI example, trying each of the options separately failed to improve estimation of the bivariate model parameters and there were error and/or warning messages. Finally, setting boundary constraints for the variance param­eters and using a different set of starting values resulted in more reliable parameter estimates (Box10.6.b).
The gradient for the variance of the random effects for logit(sensitivity) is large (1.70698), and there is concern about the analysis. Since the estimate of the variance of logit(sensitivity) is close to zero (s2usens = 0.006812) and given the previous observation about limited heterogeneity in the sensitivity estimates for CT, simplifying the model by removing either a variance or covariance parameter may be a reasonable strategy. Such an analysis will be explored in Section10.6.2.
When the bivariate model was fitted to this data set in Stata using meqrlogit (see Appendix 11 of the online supplementary material (10.S1 Code for undertaking meta­analysis)), the results were similar to those obtained using SAS and all the gradient val­ues were close to zero (Box10.6.c). In Stata, the default optimization technique for meqrlogit and xtmelogit is a Newton- Raphson technique. Starting values are determined by the commands and are not specified by the user.
As stated in Section10.2.3, it is not possible to increase the number of quadrature points used by glmer to fit a bivariate model in R. Different optimization options can be explored simultaneously using the allFit function (see code in Appendix 14 of the online supplementary material (10.S1 Code for undertaking meta- analysis)). The options are (1) bound optimization by quadratic approximation (BOBYQA); (2) an algorithm derived from BOBYQA named NLOPT_LN_BOBYQA in Box10.6.d; (3) limited memory Broyden– Fletcher–Goldfarb–Shanno bound- constrained (L- BFGS- B) optimization; (4) Nelder- Mead
299
10 Undertaking meta- analysis
.2104002
2.335581
5
nterval]
1.972165
2
]
=0
0
=1
0
4
0
4
=5
=20
-53.173887
1
1.528322
338.5475
study_id:
nterval]
https://t.me/medicina_free
Box 10.6.c Estimation in Stata of bivariate model parameters for CT and MRI comparative studies (model without test type covariate)
Iteration 3:
loglikelihood=
Gradientvector (length= .000245):
eq1:eq1:lns1_1_1: lns1_1_2: atr1_1_1_2:
sens spec _cons_cons _cons
r1 .0000374 .0000781 .0000129 .0002278 -.0000219
Mixed-effectslogistic regressionNumberofobs Binomialvariable:n Group variable: study_id Number of groups
Obs per group:
min= avg= 4. max=
Integration points= 5Waldchi2(2) Log likelihood =-53.173887Prob>chi2
true Coef.Std.Err.zP>|z|[95%Conf. Interval
sens 2.249856.204119 11.020.000 1.849792.64992 spec 1.308397.33866343.860.000 .6446286
Random-effectsParameters Estimate Std. Err. [95% Conf.I
study_id: Unstructured
var(sens) .001594.01471322.22e-11114638. var(spec) .4178528 .366883.074757
cov(sens,spec) -.0258082.1205167-.2620166
Using option stddev to output standard deviations and correlation instead of the variances and covariance above gives the following output.
Random-effectsParameters Estimate Std. Err. [95% Conf.I
Unstructured
sd(sens) .0399259 .1842611 4.71e-06 sd(spec) .6464237 .2837953 .2734133
corr(sens,spec) -1 .0011101 -1
45.1 .000
300
10.6 Meta- analysis ofsparse data anda typical data sets
2
~ , with .
0
B
## nloptwrap.NLOPT_LN_BOBYQA
0.03977226
0.6429241
–0.9999665
https://t.me/medicina_free
Box 10.6.d Estimation in R of bivariate model parameters for CT and MRI compara­tive studies (model without test type covariate)
ss$fixef
##
## bobyqa
## Nelder_Mead
## nlminbwrap
## nmkbw
## optimx.L-BFGS-B
## nloptwrap.NLOPT_LN_NELDERMEAD
## nloptwrap.NLOPT_LN_BOBYQA
ss$sdcor ## table of random effect SDs and correlations
##
## bobyqa
## Nelder_Mead
## nlminbwrap
## nmkbw
## optimx.L-BFGS-B
## nloptwrap.NLOPT_LN_NELDERMEAD
## table of fixed effects
sens
2.249866
2.249865
2.249866
2.249880
2.249866
2.249860
2.249818
Study_ID.sens
1.308946
1.308946
1.308946
1.309040
1.308946
1.308935
1.308969
0.03989542
0.03989567
0.03989607
0.03977192
0.03989560
0.03990032
spec
Study_ID.spec.sens
0.6429813
0.6429822
0.6429808
0.6429379
0.6429822
0.6429883
Study_ID.spec
–1.0000000
–1.0000000
–1.0000000
–0.9999521
–1.0000000
–0.9999998
method; (5) Nelder- Mead simplex algorithm (NLOPT_LN_NELDERMEAD in Box10.6.d); (6) nonlinear minimization with box constraints (nlminb); and (7) Nelder- Mead algorithm for derivative- free optimization (nmkb). In this example, the seven optimization options in the red box produced similar results (Box10.6.d). The help files of the functions and commands in different software packages are a useful source of information on the quad­rature and optimization options available for model estimation.
10.6.2 Simplifying hierarchical models
Hierarchical models can be simplified by removing parameters from the regression equa­tion and editing the covariance structure. Prior to simplifying the models, plot the data on forest and SROC plots to assess heterogeneity visually and to gain a better understanding of the data (e.g. all or most studies report 100% sensitivity and/or specificity with no indi­cation of a threshold effect). If the bivariate model described in Chapter9, Section9.4.1, is simplified by assuming that the covariance or correlation is zero (i.e. an independent vari­ance–covariance structure), the model reduces to two univariate random- effects logistic regression models for sensitivity and specificity as follows:
Ai A A
N
Bi B
0
(10.7)
2
301
10 Undertaking meta- analysis
81
https://t.me/medicina_free
Study
Zar 2019 Nicol 2018 Sabi 2018
FP
TP
0
108
31
0
206
64
2
22
TN
FN
51 97
105
67
Sensitivity (95% Cl)
0.22 [0.16, 0.30]
0.24 [0.19, 0.29]
0.25 [0.16, 0.35]
Specificity (95% Cl)
1. 00 [0.93, 1. 00]
1. 00 [0.96, 1. 00]
0.98 [0.93, 1. 00]
Sensitivity (95% CI) Specificity (95% Cl)
00.2 0.40.6 0.8100.2 0.40.6 0.
Figure10.6.c Forest plot of Xpert Ultra against a composite reference standard. Studies on the
plot are sorted by sensitivity and then by specificity. Source: Adapted from Kay 2020
If there is little or no observed variation in sensitivity and/or specificity, these univari­ate models can be further simplified by dropping one or both variance parameters. Without the variance parameters, the models reduce to fixed- effect logistic regression models. For example, a Cochrane Review included three studies that assessed Xpert Ultra against a composite reference standard using sputum specimens from children (Kay 2020). The estimates of sensitivity and specificity were similar in the three studies (Figure10.6.c). Given the small number of studies and lack of heterogeneity, a fixed-
effect meta- analysis was considered appropriate. In Stata the meqrlogit or xtmelogit command cannot be used for fitting fixed- effect logistic regression models; instead use the blogit command (see code in Appendix 15 of the online supplementary material (10.S1 Code for undertaking meta- analysis)).
For the IOC example introduced in Section10.6, when the covariance matrix for the random effects for logit(sensitivity) and logit(specificity) was unstructured (model A)– i.e. no constraints imposed so that the variances and covariance were uniquely esti­mated– the variances were poorly estimated, especially the variance parameter for the logit(sensitivity) of IOC. It is unsurprising that the variances were poorly estimated given the small number of studies and sparse data. Alternative models were investi­gated, as shown in Table10.6.a. In model B, the exchangeable covariance structure esti­mated a common variance and a covariance for the random effects of the logit(sensitivity) and logit(specificity). The independent covariance structure used in model C estimated distinct variances for the random effects, but the covariance was assumed to be zero (i.e. two separate univariate random- effects logistic regression models). In model D the variance parameter for logit(sensitivity) was dropped and so there was no covariance. Stata code for the four models is in Appendix 16 of the online supplementary material (10.S1 Code for undertaking meta- analysis).
To check the robustness of the assumptions about the variances of the random effects, the estimates of sensitivity and specificity were also compared between models (Table10.6.a). The analytical approach adopted was based on the reasoning that it is inappropriate to overfit models by estimating too many parameters from few studies, and to simplify models when parameter estimates cannot be reliably estimated. The validity of this approach has been investigated in a simulation study that concluded that simpler hierarchical models are valid in situations with few studies or sparse data (Takwoingi 2017). For estimating summary sensitivity and specificity, simplifying the bivariate model to univariate random- effects logistic regression models is appropriate. For estimating a summary curve, an HSROC model without the shape parameter (i.e. a symmetrical curve) can be used. If there is very little or no observed heterogeneity, fixed- effect equivalents of the models can be applied.
The output of the bivariate model fitted to the CT and MRI comparative studies in Stata (Box10.6.c) and R (Box10.6.d) show that the correlation of the logits was
302
10.6 Meta- analysis ofsparse data anda typical data sets
https://t.me/medicina_free
Table10.6.a Parameter andsummary estimates forIOC frommodels withdierent variance–covariance
structure
Parameter Models
A: Unstructured variance–covariance structure
Mean logit sensitivity (SE)
Mean logit specificity (SE)
Variance of random effects for logit sensitivity (SE)
Variance of random effects for logit specificity (SE)
Correlation of the logits (SE)
Sensitivity (95% CI) 99.9 (14.1, to 100) 98.8 (83.1 to 99.9) 99.8 (42.3 to 100) 94.4 (88.7 to 97.3)
Specificity (95% CI) 98.5 (95.8 to 99.4) 99.2 (94.5 to 99.9) 98.5 (95.6 to 99.5) 98.5 (95.6 to 99.5)
7.06 (4.53) 4.43 (1.45) 6.12 (3.2) 2.82 (0.39)
4.15 (0.52) 4.80 (1.00) 4.19 (0.57) 4.19 (0.57)
16.7 (26.9) 2.85 (2.71) 9.74 (12.7) 0
0.25 (0.54) 2.85 (2.71) 0.34 (0.67) 0.34 (0.67)
–0.73 (0.98) 0.03 (0.76) 0 0
B: Exchangeable variance–covariance structure
C: Independent variance–covariance structure
D: Fixed effects for sensitivity of IOC
estimated on the boundary of the parameter space as –1 (see Section10.6.1). Therefore, simplifying the model to two univariate random- effects logistic regression models for sensitivity and specificity, by removing the correlation or covariance parameter, seems a reasonable first step (see SAS code in Appendix 9 of the online supplementary mate­rial (10.S1 Code for undertaking meta- analysis)). The output of fitting this model in SAS shows that the boundary constraint for the variance of logit(sensitivity) was triggered (see Active BC column in (i) in Box10.6.e). Thus the next step is to simplify further by also removing this variance parameter, i.e. assume a fixed effect for sensitivity. Covariate terms can now be added to this model to compare the sensitivity and specificity of CT and MRI (see SAS code in Appendix 9 of the online supplementary material (10.S1 Code for undertaking meta-
analysis)). The difficulty encountered in the analysis of CT and MRI when using Proc NLMIXED in SAS was not experienced when using Stata (see code in Appendix 11 of the online supplementary material (10.S1 Code for undertaking meta- analysis)) and R (Appendix 14 of the online supplementary material (10.S1 Code for undertaking meta- analysis)). It is possible for the analysis of sparse data to be prob­lematic in one program and not as problematic in another program due to differences in the implementation of the programs.
Univariate models can also be used for investigating heterogeneity or for test com­parisons when the rationale for simplifying hierarchical models is not due to the num­ber of studies or convergence issues. A univariate random- effects logistic meta- regression model that allowed for a separate variance term for the random effects of logit(sensitivity) for each test was used in the main meta- analysis comparing the accuracy of nine first­trimester serum test strategies for Down syndrome screening (Alldred 2015). The analysis
303
10 Undertaking meta- analysis
2
~ ,.
2
Ak
(i) Model without covariance parameter
(ii) Model without
https://t.me/medicina_free
Box 10.6.e Simplifying the bivariate model to univariate random- effects logistic regression models for CT and MRI comparative studies (model without covariate)
Parameter
msens
mspec
s2usens
s2uspec
Estimate
2.2385
1.3122
0
0.4111
covariance parameter and variance parameter for logit (sensitivity)
Fit Statistics
-2 Log Likelihood
AIC (smaller is better)
AICC (smaller is better)
BIC (smaller is better)
Parameter Estimates
Standard
Error DF t Value Pr > |t|95% Confidence Limits Gradient Active BC
0.0014
11.46
0.1953
0.3363
0.3602
3
0.0299
3.90
3
1.14
-
0.3366
3
-
3
106.4
114.4
117.1
112.9
0.000407
1.6168
0.2418
-
–0.7353
2.8602
–0.00025
2.3825
1.81131
1.5575
-
–8.11E-6
-
Lower BC
included all studies that used a 5% FPR threshold. Equation10.8 expressed only for logit(sensitivity) was thus extended as follows:
where μ kth test; μA estimates the mean logit(sensitivity) for the index test used as the referent test (note not the reference standard); μA+vAZk estimates the mean logit(sensitivity) for the kth test; and
Fit Statistics
-2 Log Likelihood
AIC (smaller is better)
AICC (smaller is better)
BIC (smaller is better)
Parameter Estimates
2.2385
1.3122
0.4111
Standard
Error DF t Value Pr > |t|95% Confidence Limits Gradient
0.0003
11.46
0.1953
0.3363
0.3602
4
0.0175
3.90
4
0.3174
1.14
4
N vZ
Aik A A k Ak
Parameter
msens
mspec
s2uspec
Estimate
is the logit(sensitivity) for the kth test within the ith study; Zk represents the
Aik
106.4
112.4
113.9
111.3
0.000012
1.6961
0.3784
–0.5890
2.7809
2.2460
1.4112
–0.00012
–0.00004
(10.8)
is the variance of logit(sensitivity) for the kth test. The Stata
304
10.7 Meta- analysis withmultiple thresholds per study
90
Sensitivity (95% CI)
https://t.me/medicina_free
A
80
70
60
50
40
30
20
10
2 (160/1144)
Figure10.6.d Sensitivity at a 5% false positive rate for nine first- trimester serum test strategies
for Down syndrome screening. Sensitivity is presented as percentages. Each circle represents the summary sensitivity for a test strategy and the size of each circle is proportional to the number of Down syndrome cases. The estimates are shown with 95% confidence intervals. The test strategies are ordered on the plot according to decreasing detection rate. The number of studies, cases and women included for each test strategy are shown on the horizontal axis. A = Age, PlGF, PAPP­free ßhCG; B = Age, PAPP- A, free ßhCG and AFP; C = Age, ADAM 12, PAPP- A and freeßhCG; D = Age, PAPP- A and free ßhCG; E = Age, PAPP- A; F = PAPP- A; G = Age, free ßhCG and AFP; H = Age, free ßhCG; I = Free ßhCG. Source: Adapted from Alldred 2015
B
2 (116 /2705)
C
2 (74/1222)
D
(1037/49827)
E
5 (359/3491)
F
4 (325/2837)
G
3 (157/2992)
H
7 (460/5893)
A and
I
4 (390/4280)
code and output of the parameter estimates are in Appendix 17 of the online supple­mentary material (10.S1 Code for undertaking meta­the summary estimates of sensitivity obtained, including their 95% confidence inter­vals, at the 5% FPR. The test combinations were ordered on the plot according to decreasing sensitivity. The plot shows that the single- test strategies with and without maternal age (PAPP- A alone, free βhCG alone, PAPP- A and maternal age, and free βhCG and maternal age) have the worst performance, whereas thetriple- test strategies (ADAM 12, PAPP- A, free βhCG and maternal age; PAPP- A, free βhCG, AFP and maternal age) have the highest performance.
10.7 Meta- analysis withmultiple thresholds per study
As noted in Chapter9, Section9.4.5, some studies may report sensitivity and specificity at more than one threshold, because the results of the index test are ordinal catego­ries or continuous measurements (see Chapter4). The development of methods that
analysis). Figure 10.6.d shows
305
10 Undertaking meta- analysis
i i it it
i i it it
sp
it
se
it
https://t.me/medicina_free
allow for the inclusion of multiple thresholds from each included study is an active area ofresearch (Zapf 2021). This section describes and demonstrates implementa­tion of the Steinhauser (2016) and Jones(2019) approaches introduced in Chapter9, Section9.4.5. Both models performed similarly when evaluated empirically using data for a subset of published thresholds as well as all relevant thresholds from an individual participant data set of 45 studies (Benedetti 2020). The models are for tests that produce a continuous, numerical result. It is assumed that each study reports sensitivity and specificity at one or more numerical threshold values.
We first define a common notation for both models. Study i=1, ..., I provides data on sensitivity and specificity at threshold values Cit, where t=1, ..., Ti. Both models accommodate varying numbers of, and different sets of, thresholds across studies and the possibility that some studies report accuracy data at only a single threshold (i.e.T
equals 1).
i
10.7.1 Meta- analysis ofmultiple thresholds withR
The Steinhauser approach for meta- analysis of multiple thresholds creates a link between the range of thresholds and the respective pairs of sensitivity and specificity (Steinhauser
2016). The model is a two- stage random- effects model. At the study level, the reported specificity estimates across thresholds provide an estimate of the cumulative distribution function (cdf) of continuous test results among individuals without the target condition. Likewise, the set of sensitivity estimates across reported thresholds provides an estimate of the cdf of test results among individuals with the target condition. At the meta- analytical level, the model fits the data for both groups and all available thresholds over all studies. Based on a chosen parametric model, for example a logistic model, it provides estimates of the two cdfs for the two groups across all studies, accounting for the between- study heterogeneity and across- study correlation between groups.
The general model, here assuming an underlying logistic distribution for the log­transformed continuous test results, is given by
where
and
denote the observed values of specificity and sensitivity at threshold Cit in study i, α1 and α0 are fixed intercepts, and β1 and β0 are fixed slopes for the individuals with and without the target condition. The terms a0i, a1i, b0i, b1i denote random intercepts and slopes. These are assumed to follow a multivariate normal distribution, reflecting the correlation across studies. The terms ϵit and δit represent within- study random errors. The parameters are estimated using weighted least squares, where each data point is weighted with the inverse variance of the respective logit- transformed proportion.
The model provides estimates of the average distribution functions for both study groups. This allows derivation of a model- based SROC curve. If, in addition, criteria for selecting a threshold can be specified– in this case maximization of the Youden index (defined as sensitivity + specificity– 1), which applies if false negatives and false positives are of equal importance – an estimate of this threshold can be
306
sp a b C
00 00
it
se a b C
it
11 11
10.7 Meta- analysis withmultiple thresholds per study
https://t.me/medicina_free
obtained from the model. Alternatively, a larger or smaller weight can be specified (for details, see later), implemented in the R package diagmeta (Rücker 2020).
The practical use of diagmeta version 0.5-0 is demonstrated using an example provided with the diagmeta package. The data are from a meta- analysis of the diagnos­tic accuracy of fractional exhaled nitric oxide (FeNO) for diagnosis of asthma (Schneider
2017). First, the diagmeta package must be installed on the user’s platform from the CRAN repository:
install.packages("diagmeta")
After the package has been installed, it must be made available in the working space using the command library(diagmeta). Load the FENO data using
data(Schneider2017)and examine the data structure using the command View(Schneider2017), which provides the full data table. Alternatively, the first
few lines of the table can be examined using the command head(Schneider2017), which produces the output in Box10.7.a.
The first number in each row is a row number automatically provided by R (not given in the data set). The second column (study_id) is the study identifier, here a number; the third and fourth columns provide study author and year of publication. All the visible lines belong to the same study, Arora 2006, with a study ID of 1. The ‘group’ column is empty for this study (it refers to a later study, Malinovschi 2012, which provides data in three sub­studies). The ‘cutpoint’ column contains the threshold for FeNO measured in ppb. The last four columns (tpos, fneg, fpos and tneg) give the numbers of true positive (TP), false nega­tive (FN), false positive (FP) and true negative (TN) results (2×2data), if the value in the ‘cutpoint’ column is used to determine test positivity. Because larger values indicate asthma (thetarget condition), the numbers of (true or false) positives must decrease with increasing threshold, whereas the numbers of (true or false) negatives increase with the threshold within a study. Note that it is also possible to start from individual participant data with a given study ID, true status of the participant and individual marker value. This type of data can be transformed to the format needed using the function IPD2diag().
The main function of diagmeta is also called diagmeta(). To see the arguments required by diagmeta(), look at the help file using help(diagmeta) or look directly at the arguments using args(diagmeta)to give the
output presented in Box10.7.b.
Box 10.7.a Subset of fractional exhaled nitric oxide data for diagnosis of asthma
study_id author year group cutpoint tpos fneg fpos tneg
1 1 Arora 2006 6 133 5 34 0
2 1 Arora 2006 7 131 7 33 1
3 1 Arora 2006 8 130 8 31 3
4 1 Arora 2006 9 127 11 30 4
5 1 Arora 2006 10 119 19 28 6
6 1 Arora 2006 11 115 23 26 8
307