**Purpose**:
To compare methods to assess visual field (VF) progression in glaucoma.

**Methods**:
4,950 VFs of 253 primary open angle-glaucoma patients were evaluated for progression with the following methods: clinical evaluation, guided progression analysis (GPA), mean deviation (MD), and visual field index (VFI) rates, Advanced Glaucoma Intervention Study (AGIS) and Collaborative Initial Glaucoma Treatment Study (CIGTS) scores, pointwise linear regression (PLR), permutation of PLR (PoPLR), and glaucoma rate index (GRI). A separate simulated series of longitudinal VFs was assessed with all methods except for GPA and clinical evaluation.

**Results**:
The average (±SD) age of the patients at baseline was 65.4 (±11.5) years. The average (±SD) follow-up was 11.8 (±4.6) years, and the mean (±SD) number of VFs was 16.8 (±7.0). Proportion of series detected as progressing was 65% for PoPLR, 58% for GRI, 41% for GPA, 40% for PLR, 36% for CIGTS, 35% for clinicians, 31% for MD rate, 29% for AGIS, and 22% for VFI rate. Median times to detection of progression were 7.3 years for PoPLR, 7.5 years for GRI, 11 years for clinicians, 14 years for GPA, 16 years for PLR, 17 years for CIGTS, 19 years for AGIS, and more than 20 years for MD and VFI rates. In simulated VF series, GRI had the highest partial area under the receiver operator characteristic curve (0.040) to distinguish between glaucoma progression and aging/cataract decay, followed by VFI rate (0.028), MD rate (0.024), and PoPLR (0.006).

**Conclusions**:
GRI and PoPLR showed the highest proportion of series detected as progressing and shortest times to progression detection. GRI exhibited the best ability to detect progression in the simulated VF series.

**Translational Relevance**:
Knowledge of the properties of every method would allow tailoring application in both clinical and research settings.

^{1}Determination of the rates of progression is important to discriminate fast from slow progressors, because the former group of patients may require more aggressive treatment and more frequent follow-up. Most clinical trials have defined perimetric deterioration as their primary study outcome.

^{2–5}

^{6}Age, eccentricity, initial sensitivity, presence of other ocular diseases, and test strategy are other pertinent variables, which affect long-term fluctuation.

^{7}A real VF change must exceed the expected amount of noise of the test series and be replicable.

^{8}Numerous approaches have been implemented to detect and measure VF deterioration. Subjective judgment is still commonly used in the clinical setting, although its interobserver reliability is unsatisfactory.

^{9,10}Several statistical models have been proposed to aid the clinician in evaluating perimetric progression objectively, including guided progression analysis (GPA),

^{11}rates of change of global indices (i.e., mean deviation [MD], visual field index [VFI], or pattern standard deviation [PSD]),

^{12}multivariable regression analysis,

^{13}algorithms designed for clinical trials (e.g., the Advanced Glaucoma Intervention Study [AGIS]

^{2}and Collaborative Initial Glaucoma Treatment Study [CIGTS]),

^{5}pointwise regression analysis (linear or exponential),

^{14,15}permutation of pointwise linear regression (PoPLR),

^{16}and the recently introduced glaucoma rate index (GRI),

^{17}Several studies have provided a comparison among some of these methods, but there has been no consensus on which is the single best approach.

^{7,9,18,19}

^{20}All VF series were assessed for progression with the following methods: qualitative clinical evaluation, GPA, MD rates of change, VFI rates of change the AGIS scoring system, the CIGTS scoring system, PLR, PoPLR, and GRI.

^{21}and Gardiner and Crabb,

^{22}we developed an algorithm in the software environment R (R Foundation for Statistical Computing, Vienna, Austria) to simulate 24-2 VF series. Key steps of the process were as follows:

- The user specifies the baseline threshold sensitivities, length of follow-up, annual pointwise rates of progression, and number of VF examinations. Tests are equally spaced over the follow-up period, and their frequency is derived as the ratio between the number of VFs and length of simulation;
- Linear regression analysis is applied at every test location. Independently from the determined rate of progression, an additional decay of 0.1 dB/y is added to simulate age-related decline
^{23}; and - For each location, the noise-free value is replaced by a Monte-Carlo value randomly computed from a Gaussian distribution. The mean of the distribution is equal to the estimated noise-free threshold sensitivity. The standard deviation (SD) is calculated with the function proposed by Gardiner and Crabb
^{22}:\(\def\upalpha{\unicode[Times]{x3B1}}\)\(\def\upbeta{\unicode[Times]{x3B2}}\)\(\def\upgamma{\unicode[Times]{x3B3}}\)\(\def\updelta{\unicode[Times]{x3B4}}\)\(\def\upvarepsilon{\unicode[Times]{x3B5}}\)\(\def\upzeta{\unicode[Times]{x3B6}}\)\(\def\upeta{\unicode[Times]{x3B7}}\)\(\def\uptheta{\unicode[Times]{x3B8}}\)\(\def\upiota{\unicode[Times]{x3B9}}\)\(\def\upkappa{\unicode[Times]{x3BA}}\)\(\def\uplambda{\unicode[Times]{x3BB}}\)\(\def\upmu{\unicode[Times]{x3BC}}\)\(\def\upnu{\unicode[Times]{x3BD}}\)\(\def\upxi{\unicode[Times]{x3BE}}\)\(\def\upomicron{\unicode[Times]{x3BF}}\)\(\def\uppi{\unicode[Times]{x3C0}}\)\(\def\uprho{\unicode[Times]{x3C1}}\)\(\def\upsigma{\unicode[Times]{x3C3}}\)\(\def\uptau{\unicode[Times]{x3C4}}\)\(\def\upupsilon{\unicode[Times]{x3C5}}\)\(\def\upphi{\unicode[Times]{x3C6}}\)\(\def\upchi{\unicode[Times]{x3C7}}\)\(\def\uppsy{\unicode[Times]{x3C8}}\)\(\def\upomega{\unicode[Times]{x3C9}}\)\(\def\bialpha{\boldsymbol{\alpha}}\)\(\def\bibeta{\boldsymbol{\beta}}\)\(\def\bigamma{\boldsymbol{\gamma}}\)\(\def\bidelta{\boldsymbol{\delta}}\)\(\def\bivarepsilon{\boldsymbol{\varepsilon}}\)\(\def\bizeta{\boldsymbol{\zeta}}\)\(\def\bieta{\boldsymbol{\eta}}\)\(\def\bitheta{\boldsymbol{\theta}}\)\(\def\biiota{\boldsymbol{\iota}}\)\(\def\bikappa{\boldsymbol{\kappa}}\)\(\def\bilambda{\boldsymbol{\lambda}}\)\(\def\bimu{\boldsymbol{\mu}}\)\(\def\binu{\boldsymbol{\nu}}\)\(\def\bixi{\boldsymbol{\xi}}\)\(\def\biomicron{\boldsymbol{\micron}}\)\(\def\bipi{\boldsymbol{\pi}}\)\(\def\birho{\boldsymbol{\rho}}\)\(\def\bisigma{\boldsymbol{\sigma}}\)\(\def\bitau{\boldsymbol{\tau}}\)\(\def\biupsilon{\boldsymbol{\upsilon}}\)\(\def\biphi{\boldsymbol{\phi}}\)\(\def\bichi{\boldsymbol{\chi}}\)\(\def\bipsy{\boldsymbol{\psy}}\)\(\def\biomega{\boldsymbol{\omega}}\)\(\def\bupalpha{\bf{\alpha}}\)\(\def\bupbeta{\bf{\beta}}\)\(\def\bupgamma{\bf{\gamma}}\)\(\def\bupdelta{\bf{\delta}}\)\(\def\bupvarepsilon{\bf{\varepsilon}}\)\(\def\bupzeta{\bf{\zeta}}\)\(\def\bupeta{\bf{\eta}}\)\(\def\buptheta{\bf{\theta}}\)\(\def\bupiota{\bf{\iota}}\)\(\def\bupkappa{\bf{\kappa}}\)\(\def\buplambda{\bf{\lambda}}\)\(\def\bupmu{\bf{\mu}}\)\(\def\bupnu{\bf{\nu}}\)\(\def\bupxi{\bf{\xi}}\)\(\def\bupomicron{\bf{\micron}}\)\(\def\buppi{\bf{\pi}}\)\(\def\buprho{\bf{\rho}}\)\(\def\bupsigma{\bf{\sigma}}\)\(\def\buptau{\bf{\tau}}\)\(\def\bupupsilon{\bf{\upsilon}}\)\(\def\bupphi{\bf{\phi}}\)\(\def\bupchi{\bf{\chi}}\)\(\def\buppsy{\bf{\psy}}\)\(\def\bupomega{\bf{\omega}}\)\(\def\bGamma{\bf{\Gamma}}\)\(\def\bDelta{\bf{\Delta}}\)\(\def\bTheta{\bf{\Theta}}\)\(\def\bLambda{\bf{\Lambda}}\)\(\def\bXi{\bf{\Xi}}\)\(\def\bPi{\bf{\Pi}}\)\(\def\bSigma{\bf{\Sigma}}\)\(\def\bPhi{\bf{\Phi}}\)\(\def\bPsi{\bf{\Psi}}\)\(\def\bOmega{\bf{\Omega}}\)\begin{equation}\tag{1}\ln \left( {{\rm{SD}}} \right):{\rm{\ }} - 0.081{\rm{\ }} \times {\rm{\ sensitivity\ }}\left( {{\rm{dB}}} \right) + 3.27\end{equation}

^{21}Specifically,

*i*and

*j*being the coordinates of a 10 × 10 grid. According to this model, the variability is proportional to the diminution of threshold sensitivity and the point eccentricity. Values generated below 0 dB or above 36 dB are truncated at 0 or 36 dB, respectively.

^{24}

^{22,25}Simulation length was established at 9.5 years with a biannual testing frequency for a total of 20 VFs. Baseline age was set at 60 years. Two VF exams from actual glaucoma patients, one with a focal inferior nasal defect and another with a superior arcuate scotoma, were chosen as the two baselines. We then applied two different models of progression, as illustrated in Figure 1:

- Focal decay, wherein four (small scotoma), eight (medium scotoma), or 16 (large scotoma) locations significantly deteriorate. Simulated defects of small dimensions were represented by a nasal scotoma and a paracentral scotoma. Medium-sized defects were represented by a nasal step and an arcuate scotoma extending to 5° from fixation. Large defects consisted of two broad, inferior and superior, arcuate scotomas. Pattern of focal deterioration were selected by a glaucoma specialist author from VFs of actual glaucoma patients. Three different rates of progression, −0.5, −1, and −2 dB/y, were applied to deteriorating locations in addition to normal age-related decay. The hemifield involved was randomly assigned. Based on the baseline exams, rates, and patterns of progression, 18 scenarios were simulated and each of them was run 372 times, to match the number of eyes in the cohort of patients. All eyes with deterioration of simulated scotomata were considered progressing.
- Diffuse decay, wherein every location undergoes the same rate of progression. In one model, we applied only age-related decay (0.1 dB/y).
^{22}In accordance with previously published results, we specified a rate of progression of 0.29 dB/y to simulate an average rate of decline from worsening cataract.^{26}These two groups of eyes were considered nonprogressing. Based on the two baseline examinations and rate of progression, four scenarios existed for this group, and each was run 1674 times so that the number of progressing and nonprogressing series was equal. The differences between these iterations represent the random noise provided by the model.

^{27}Three experienced clinicians (ALC, JC, KNM) independently assessed 4950 VF examinations from 372 eyes. The evaluation was carried out on the 24-2 HVF single-examination printouts, and graders were masked to all clinical data, results of other scoring systems, and other graders' judgments. Each grader determined progression of each VF series with a semiquantitative scale, with a score of 1 (definite progression), 2 (probable progression), 3 (indeterminate), 4 (probably stable), and 5 (definitely stable). For each eye, an average score from 1 to 5 was calculated. “Progression” and “no progression” were defined as an average score of less than 3 and 3 or more, respectively. Experts also indicated the time when they first judged the VF as progressing, and the average time to progression was determined.

^{28}Briefly, the pattern deviation values on each follow-up VF are compared point-by-point with those of the two baseline exams chosen by the user. If the difference for a given location is significantly higher than test–retest variability at a

*P*< 0.05, it is marked with an open triangle. If the worsening of that point is confirmed on two or three consecutive examinations, it is flagged as a half-filled or filled triangle, respectively. VF progression was defined as the deterioration of three or more locations sustained on three or more consecutive examinations, in accordance with the criteria adopted by the EMGT.

^{3}

*P*< 0.05) rate of change of −0.5 dB/y or less and −1.8%/y or less, respectively. The value of 1.8% for VFI was chosen because it corresponds approximately to the 0.5 dB/y value used for MD.

^{29}

^{2}VF grading is based on number, depth, and spatial distribution of depressed locations with a final score that ranges from 0 (normal) to 20 (end-stage disease). Progression was defined as a score that increased by four or more points compared with the baseline test, sustained in at least three consecutive VF examinations.

^{5}Like the AGIS, the CIGTS score ranges from 0 (normal) to 20 (end-stage disease). A worsening of three or more points compared with the average of the first two tests and sustained for three consecutive examinations was defined as progression.

*P*< 0.01) of −1 dB/y or less was defined as progression.

^{16}For each patient's VF sequence, PLR was performed on the total deviation data. Data belonging to every location were combined to generate a global score, called S

_{obs}, with a truncated product method, which is a generalization of the Fisher method and allows to combine

*P*values derived from each pointwise series.

^{30}The patient's original VF sequence was then randomly reordered up to 5000 times, and a global score, called S

_{p}, was obtained from each permutated series. Finally, S

_{obs}was compared with the S

_{p}distribution, and the statistical significance was derived from the ranking of S

_{obs}within the S

_{p}distribution. A

*P*value of less than 0.05 was labeled as progression.

^{17}Briefly, each pointwise sequence was classified as decaying or improving depending on its a priori linear trend. Based on this categorization, pointwise exponential regression (PER) was computed for each series. For locations with a negative trend, the following formula was applied:

*y = e*

^{(a+bx)}, where y is the threshold sensitivity (dB), a is the constant, b is the slope (regression coefficient), and x is the time (years). For locations with a positive trend, the following formula was used:

*Y−y = e*

^{(a+bx)}, where Y is the normal age-matched threshold sensitivity + 2SD, y is the threshold sensitivity (dB), a is the constant, b is the slope, x is the time (years). Outliers were removed with the sequential application of Cook's distance and the Studentized residual tests. After carrying out PER, two values were obtained, pointwise rate of change (PRC) and the 90% confidence interval (CI) of the slopes. The former indicates the rate of change of each pointwise sequence, expressed as the percentage of the entire perimetric range corrected for age and location. A GRI score is generated by summing the PRC values from locations with significant negative (decaying) and positive (improving) rates across the VF series in an individual eye. The summed value is then normalized from a maximum rate of decay (−100) to a maximum rate of improvement (+100). A GRI value of less than −6 was defined as progression.

^{17}

^{31}Time to the first detection of progression was assessed with Kaplan-Meier curves, differences across methods were compared with Cox's regression shared frailty model in R software, and multiple comparisons were adjusted with the Benjamini-Hochberg test.

^{32}

^{33}

^{34,35}

*n*= 32) were excluded from this particular analysis.

^{36}and multiple comparisons were adjusted with the Benjamini-Hochberg test.

^{32}

^{36}

*P*< 0.001), whereas the difference between these two methods was not significant (

*P*= 0.97). VFI rate and AGIS scores required significantly longer times to detect progression compared with all other methods (

*P*< 0.01), and did not significantly differ from each other (

*P*= 0.12). In addition, PLR detected progression significantly faster than MD rate (

*P*= 0.04) and CIGTS (

*p*= 0.04). None of the other pairwise comparisons were statistically significant.

*P*= 0.17).

*P*< 0.0001), except for the pairwise comparison between MD rate and PoPLR (

*P*= 0.09). The introduction of eyes with simulated cataract (Fig. 6) considerably hampered the discriminatory abilities of all the trend-based methods, with GRI having the highest partial AUROC (0.0399), followed by VFI rate (0.0279), MD rate (0.0235), and lowest by PoPLR (0.0062).

*P*< 0.05).

^{9}However, its intrinsic subjectivity and lack of standardization are drawbacks. Several studies revealed a disappointing level of agreement among experienced observers, despite good intraobserver reproducibility.

^{10,37}Also, subjective evaluation is likely driven by the fast component of VF decay.

^{27}

^{38}

^{2,5}Heijl and colleagues

^{39}compared the properties of EMGT, AGIS, and CIGTS scoring systems with expert evaluation as a reference method. They found that EMGT criteria were more sensitive and identified progression faster than AGIS and CIGTS, albeit less specifically. AGIS and CIGTS criteria had similar diagnostic properties when compared with each other. Vesti et al.

^{18}came to the same conclusions with computer-simulated VFs as a reference. Nouri-Mahdavi et al.

^{19}compared the performance of an earlier version of GPA, called Glaucoma Change Probability Analysis (GCPA), the AGIS score, and PLR to predict VF progression, and found that GCPA detected true clinical progression slightly more often than the other two methods, with false-positive prediction rates between 1% and 3%.

^{40}found that an increase of AGIS score of four or more points maintained in two rather than three consecutive tests raised the sensitivity to approximately 50% with negligible change in the specificity. Among the three event-based algorithms in this study, GPA had the highest proportion of series detected as progressing with similar prediction ability and consistency. Because GPA runs on proprietary software, it is not available in the simulated environment, and we were not able to estimate its specificity in relation to the other methods. Previous studies reported an overall high specificity for GPA, although patients with higher test–retest variability and unreliable examinations can experience higher percentages of false-positive alerts.

^{41,42}The performance of CIGTS was intermediate between AGIS and GPA. The specificity of the CIGTS score was severely affected by the introduction of the cataract-related decay in the computer simulation. This finding is not surprising; indeed, the CIGTS score is based on the total deviation probability map, which is highly influenced by media opacity.

^{43}Both AGIS and CIGTS scoring systems were developed for patients with more advanced glaucoma than the ones included in our study, and this may explain their low detection rates in this setting.

^{38}MD and VFI are well-known global VF indices and their linear rates are easily measured; VFI rates of change are provided by HFA's GPA software.

^{38}However, MD is relatively insensitive to progressive glaucomatous VF loss and has poor specificity in clinical environments.

^{44,45}The MD rate of change quantifies overall VF loss, so localized but potentially clinically important changes may be missed entirely or confounded by generalized media effects, such as worsening cataract or cataract surgery, which are common events in glaucoma patients.

^{44}Another global index, VFI, fares no better, and becomes unreliable in advanced stages of glaucoma.

^{46}Gardiner et al.

^{45}reported that as the duration of follow-up and number of VFs increase, it becomes difficult to rely solely on linear models. This is because progression often occurs in nonlinear patterns, especially as the disease severity and its treatment change. Although VFI provides predictive capability with extrapolation, it assumes a linear rate of worsening, and is affected by the same drawbacks as MD.

^{46}Global indices have the potential to provide an estimate of VF decay at the expense of loss of spatial information that may be important to clinical decision-making. In the current study, both MD and VFI rates were characterized by low proportion of series detected as progressing and were the slowest to detect progression, although they had better prediction ability than any other method. VFI rate exhibited even lower proportion of series detected as progressing in the cohort of patients, and it performed worse than MD rate in the simulated environment. Gardiner and Demirel

^{12}compared the performances of three global indices (i.e., MD, VFI, PSD), and found similar results to ours, with MD rate detecting progression sooner and more frequently than the VFI rate. When we included the cataract group in our computer simulations, however, an opposite scenario occurred with VFI rate performing considerably better than MD rate. This finding is not unexpected because VFI would be expected to be less influenced by cataract and cataract surgery.

^{47}

^{7}Another positive aspect of pointwise trend-based methods is the preservation of the spatial information, which provide insight in the patterns, in addition to the rates, of progression.

^{7}PLR is undoubtedly the most commonly used pointwise trend-based method and the first to be commercially available in the Progressor software (OBF Labs UK Ltd, Wiltshire, UK). In our study, PLR had a performance similar to GPA in the cohort of patients, and exhibited intermediate sensitivity, but high specificity, in the simulated series. PoPLR is an evolution of PLR that may outperform simple PLR.

^{16}PoPLR is independent of data format, and allows for comparison of different instruments, follow-up protocols, and test strategies.

^{48,49}Our results indicate that PoPLR has a high proportion of series detected as progressing and detects progression early. In the computer simulations, PoPLR was the poorest performing trend-based method, especially when simulated cataract was included in the analysis. These drawbacks are likely related to the assumptions made by the algorithm. PoPLR is generated by comparing a series of VFs of a patient with a maximum of 5000 randomly permutated sequences, assuming that a nonprogressing eye should not differ from the null distribution. However, healthy eyes experience physiological age-related decline and decay from cataract development, and this could be detected as significant progression when compared with the randomly permutated sequences by the algorithm. In an attempt to limit such phenomenon, PoPLR is calculated from the total deviation values, which indicate the difference between the patient's VF and a normal reference VF based on the patient's age. Age-corrected normal thresholds, however, have high interindividual variation, especially in the midperipheral and peripheral locations and follow a non-Gaussian distribution.

^{23}O'Leary et al.

^{16}reported that PoPLR had a low percentage of false-positives, but these results were obtained with randomly permutated sequences as true nonprogressing sequences, not taking into account the aforementioned factors. In a prospective study by Redmond and colleagues,

^{48}PoPLR labeled as falsely progressing almost one-third of healthy subjects followed over a mean time of 5 years; when PoPLR was calculated from the pattern deviation values (rather than the total deviation ones), the false-positive rate was null, reinforcing the idea that this method is severely impaired by the diffuse, paraphysiological VF decay caused by aging and cataract development. Although calculating PoPLR on the pattern deviation values may seem a potential solution to increase its specificity, it is well known that pattern deviation values tend to underestimate VF progression, and may be misleading in the case of very early glaucoma because of a ceiling effect, as well as in severe glaucoma where it can underestimate diffuse generalized damage.

^{29,41}

^{17}In contrast with PLR and its derivatives, GRI is based on a pointwise exponential regression, in accordance with previous findings that the exponential model fits perimetric measurements and predicts future changes better than linear models.

^{14,50}GRI has potential advantages, including discriminating ability for fast-progressors, assessment of improvement, and an intuitively interpretable display. In our cohort of patients, GRI had high proportion of series detected as progressing and detected progression faster than other methods and as frequently as PoPLR. On the computer simulation, GRI had the highest partial AUROC both for detection of perimetric progression and discrimination of fast progression compared with the other trend-based methods.

^{51}defined fast progression and catastrophic progression as a MD worsening rate between 1 and 2 dB/y and more than 2 dB/y, respectively, and they found these conditions in 4.3% and 1.5% in patients under clinical care, respectively. In their study, they identified fast-progressors with MD rate rather than VFI rate, because ceiling effects may potentially hamper the latter's loss of sensitivity because it relies on pattern deviation, and there is discontinuity in case of severe VF damage.

^{29,51}In our study, VFI rate performed better than MD rate to identify fast-progressors. Neither approach, however, retains spatial information and may be quite insensitive to fast, but focal, deterioration.

^{52}In the current study, GRI displayed the best capability to identify fast progressing eyes. PoPLR, which also retains spatial discrimination, performed considerably worse in this regard.

^{19,53,54}Our data are consistent with the previous findings. Only 8% of the eyes were judged unanimously as progressing, whereas 20% were deemed as stable by all the methods. In a recent study, Saeedi and colleagues

^{55}evaluated the agreement among six methods (PLR, PoPLR, MD, and VFI rates, CIGTS, and AGIS scoring systems) to detect glaucomatous VF progression on a large cohort of patients, and found that eyes labeled as progressing and stable by all the methods were 2.5% and 41.5% of all series, respectively. These results are largely different from ours, and may be explained by various factors, such the shorter follow-up length and more stringent progression criteria for some of the trend-based methods used by Saeedi et al.

^{55}Furthermore, the authors have chosen questionable cutoffs for some of the trend-based methods.

^{56}For MD rate, Saeedi and colleagues

^{55}used a cutoff of −1 dB/y, so that eyes having a rate of progression faster or slower than this threshold value were categorized as progressing or stable, respectively. However, this is quite a high cutoff value, previously used to distinguish between fast progressing and slower progressing eyes, rather than between progressing and stable eyes.

^{51}Because MD and VFI are highly correlated, a corresponding cutoff for VFI rate would have been −5.4%/y, which is considerably higher than the one used by Saeedi and colleagues (−1%/y).

^{29}It is evident that progression detection strongly depends on the method employed (and cutoff used for each individual method). In our study, MD rate and VFI rate were then only pair to show substantial agreement; this finding is not surprising because the two indices are highly correlated, and we chose equivalent decay rates to define progression.

^{29}GRI and PoPLR, the two most sensitive methods in the current study, revealed one of the highest agreement, as exemplified by their conspicuous intersection in the UpSet graph (Fig. 3). On the other hand, the two most specific methods (AGIS and VFI rate) did not exhibit such agreement. Once again, this is not unexpected because several studies have shown discrepancies between event- and trend-based analyses, suggesting that they identify distinct aspects of perimetric change.

^{11,57}Medeiros and colleagues

^{11}proposed a Bayesian hierarchical model to combine event- and trend-based approaches, and they reported that the combined approach outperformed each method used alone. The combination of more than one method may represent a viable option to integrate complementary information from individual algorithms, possibly mitigating their drawbacks.

^{58}Many methods to detect glaucoma progression have been published, and the relationship between untested methods remains undetermined. Nevertheless, we evaluated a considerable number of methods with different strategies (i.e., subjective evaluation, event-based analysis, trend-based analyses), established methods (i.e., AGIS, CIGTS, GPA, MD rate, VFI rate, PLR), as well as novel and promising ones (i.e., PoPLR, GRI). In the simulations, some indices (i.e., MD, VFI, AGIS, CIGTS) were based on calculations carried out with the normative database of the ‘VisualFields' package, and they might differ slightly from the values generated by HVF's software.

^{24}Computer simulation is a strategy to obtain an external gold standard, but may oversimplify real and more complex disease progression. Additionally, only a few of all the myriad of possible baseline examinations, patterns, and rates of progression are included in these simulations. GPA and expert evaluation were tested only in the cohort of patients, and were not available in the simulated environment.

**A. Rabiolo**, Santen Italy srl (C);

**E. Morales**, None;

**L. Mohamed**, None;

**V. Capistrano**, None;

**J.H. Kim**, None;

**A. Afifi**, None;

**F. Yu**, None;

**A.L. Coleman**, None;

**K. Nouri-Mahdavi**, Heidelberg Engineering (C);

**J. Caprioli**, Aerie, Alcon, Allergan, Glaukos, and New World Medical (C)

*Am J Ophthalmol*. 2008; 145: 191–192.

*Ophthalmology*. 1994; 101: 1445–1455.

*Acta Ophthalmol Scand*. 2003; 81: 286–293.

*Ophthalmology*. 1992; 99: 1468–1470.

*Ophthalmology*. 1999; 106: 653–662.

*Arch Ophthalmol*. 1984; 102: 704–706.

*Surv Ophthalmol*. 2002; 47: 158–173.

*Am J Ophthalmol*. 1984; 97: 328–331.

*Ophthalmology*. 2017; 124: S51–S56.

*Ophthalmology*. 2011; 118: 60–65.

*Ophthalmology*. 2012; 119: 458–467.

*Am J Ophthalmol*. 2017; 176: 148–156.

*Invest Ophthalmol Vis Sci*. 1993; 34: 1907–1916.

*Invest Ophthalmol Vis Sci*. 2014; 55: 7881–7887.

*Br J Ophthalmol*. 1997; 81: 1037–1042.

*Invest Ophthalmol Vis Sci*. 2012; 53: 6776–6784.

*Transl Vis Sci Technol*. 2018; 7 (6): 14.

*Invest Ophthalmol Vis Sci*. 2003; 44: 3873–3879.

*Arch Ophthalmol*. 2007; 125: 1176–1181.

*Am J Ophthalmol*. 2019; 201: 19–30.

*Invest Ophthalmol Vis Sci*. 2000; 41: 2192–2200.

*Invest Ophthalmol Vis Sci*. 2002; 43: 1400–1407.

*Arch Ophthalmol*. 1987; 105: 1544–1549.

*J Vis*. 2013; 13 (4).

*Ophthalmic Physiol Opt*. 2015; 35: 155–169.

*JAMA Ophthalmol*. 2014; 132: 1296–1302.

*J Glaucoma*. 2017; 26: 902–910.

*Arch Ophthalmol*. 1991; 109: 1690–1692.

*Invest Ophthalmol Vis Sci*. 2011; 52: 4030–4038.

*Genet Epidemiol*. 2002; 22: 170–185.

*Bioinformatics*. 2017; 33: 2938–2940.

*J R Stat Soc Series B Stat Methodol*. 1995; 57: 289–300.

*Biometrics*. 1977; 33: 159–174.

*J Am Stat Assoc*. 1937; 32: 675–701.

*Technometrics*. 1964; 6: 241–252.

*BMC Bioinformatics*. 2011; 12: 77.

*Br J Ophthalmol*. 2003; 87: 726–730.

*Prog Retin Eye Res*. 2017; 56: 107–147.

*Ophthalmology*. 2008; 115: 1557–1565.

*Ophthalmology*. 2004; 111: 2117–2125.

*Invest Ophthalmol Vis Sci*. 2005; 46: 4600–4606.

*Ophthalmology*. 2014; 121: 2023–2027.

*Acta Ophthalmol Scand*. 1997; 75: 184–188.

*Invest Ophthalmol Vis Sci*. 1990; 31: 512–520.

*Invest Ophthalmol Vis Sci*. 2013; 54: 1252–1259.

*Arch Ophthalmol*. 2009; 127: 1610–1615.

*Am J Ophthalmol*. 2008; 145: 343–353.

*JAMA Ophthalmol*. 2013; 131: 1565–1572.

*Invest Ophthalmol Vis Sci*. 2018; 59: 439–445.

*Invest Ophthalmol Vis Sci*. 2012; 53: 5403–5409.

*Invest Ophthalmol Vis Sci*. 2014; 55: 4135–4143.

*Ophthalmology*. 2002; 109: 766–770.

*Ophthalmology*. 1997; 104: 1228–1236.

*Ophthalmology*2012; 119: 468–473.

*Ophthalmology*. 2019; 126: 822–828.

*Ophthalmology*. In press.

*Eur J Ophthalmol*. 2013; 23: 187–195.

*Arch Ophthalmol*. 2011; 129: 1521–1527.