Limitations of the incidence density ratio as approximation of the hazard ratio
Abstract
Background
Incidence density ratios (IDRs) are frequently used to account for varying followup times when comparing the risks of adverse events in two treatment groups. The validity of the IDR as approximation of the hazard ratio (HR) is unknown in the situation of differential average follow up by treatment group and nonconstant hazard functions. Thus, the use of the IDR when individual patient data are not available might be questionable.
Methods
A simulation study was performed using various survivaltime distributions with increasing and decreasing hazard functions and various situations of differential follow up by treatment group. HRs and IDRs were estimated from the simulated survival times and compared with the true HR. A rule of thumb was derived to decide in which data situations the IDR can be used as approximation of the HR.
Results
The results show that the validity of the IDR depends on the survivaltime distribution, the difference between the average followup durations, the baseline risk, and the sample size. For nonconstant hazard functions, the IDR is only an adequate approximation of the HR if the average followup durations of the groups are equal and the baseline risk is not larger than 25%. In the case of large differences in the average followup durations between the groups and nonconstant hazard functions, the IDR represents no valid approximation of the HR.
Conclusions
The proposed rule of thumb allows the use of the IDR as approximation of the HR in specific data situations, when it is not possible to estimate the HR by means of adequate survivaltime methods because the required individual patient data are not available. However, in general, adequate survivaltime methods should be used to analyze adverse events rather than the simple IDR.
Keywords
Hazard function Incidence rate Incidence density ratio Randomized controlled trials Simulation Timetoevent dataAbbreviations
 BLR
Baseline risk
 CI
Confidence interval
 CP
Coverage probability
 EAIR
Exposureadjusted incidence rate
 HR
Hazard ratio
 ID
Incidence density
 IDR
Incidence density ratio
 IQWiG
Institut für Qualität und Wirtschaftlichkeit im Gesundheitswesen
 MPE
Mean percent error
 MSE
Mean square error
 RCT
Randomized controlled trial
 RR
Risk ratio
 SE
Standard error
Background
Adverse events play an important role in the assessment of medical interventions. Simple standard methods for contingency tables are frequently applied for the analysis of adverse events. However, the application of simple, standard methods may be misleading if observations are censored at the time of discontinuation due to, for example, treatment switching or noncompliance, resulting in varying followup times, which sometimes differ remarkably between treatment groups [1]. Incidence densities (IDs), i.e., events per patient years, are frequently used to account for varying followup times when quantifying the risk of adverse events [2, 3, 4]. IDs are also called exposureadjusted incidence rates (EAIRs) to underline that varying followup times are taken into account [2, 3, 4, 5]. For comparisons between groups, incidence density ratios (IDRs) are used together with confidence intervals (CIs) based upon the assumption that the corresponding timetoevent variables follow an exponential distribution. The corresponding results are interpreted in the same way as hazard ratios (HRs).
An example is given by the benefit assessment of the Institute for Quality and Efficiency in Health Care (IQWiG) in which the added benefit of abiraterone acetate (abiraterone for short) in comparison with watchful waiting was investigated in men with metastatic prostate cancer that is not susceptible to hormoneblocking therapy, who have no symptoms or only mild ones, and in whom chemotherapy is not yet indicated [6]. In this report the IDR was used to compare the risks of cardiac failure in the abiraterone group and the control group of the corresponding approval study. The result was IDR = 4.20, 95% CI 0.94, 18.76; P = 0.060. It is questionable whether the use of the IDR is adequate in this data situation because the median followup duration was 14.8 months in the abiraterone group but only 9.3 months in the control group. The reason for this large difference was the discontinuation of treatment after disease progression with stopping of the monitoring of adverse events 30 days later. In the situation of constant hazard functions, i.e., if the timetoevent data follow an exponential distribution, the IDR accounts for the differential follow up by treatment group. However, if the hazard functions are not constant, the effect of differential follow up by treatment group on the behavior of the IDR is unknown. Appropriate methods should be used for analysis of survival data if access to the individual patient data is available. However, access to the individual patient data is not available in the assessment of dossiers or publications with aggregatelevel data. In this situation, a decision has to be made on the situations in which the IDR can or cannot be used as adequate approximation for the HR.
The use of IDs makes sense in the situation of constant hazard functions in both groups [2, 3, 5, 7]. However, timetoevent data rarely follow an exponential distribution in medical research [3, 7]. In the case of low event risks, deviations from the exponential distribution may be negligible if the average follow up is comparable in both groups [2]. However, in the case of differential follow up by treatment group, deviations from the exponential distribution may have a considerable effect on the validity of the IDR and the corresponding CIs as an approximation of the HR.
Kunz et al. [8] investigated bias and coverage probability (CP) of point and interval estimates of IDR in metaanalyses and in a single study with differential follow up by treatment group when incorrectly assuming that average follow up is equal in the two groups. It was shown that bias and CP worsen rapidly with increasing difference in the average followup durations between the groups [8]. Here, we do not consider the effect of incorrectly assuming equal average followup durations. IDR is calculated correctly by using the different followup durations in the groups. The focus here is the effect of deviations from the exponential distribution of the timetoevent data.
In this paper, the validity of the IDR as approximation of the HR is investigated in the situation of differential average follow up by treatment group by means of a simulation study considering decreasing and increasing hazard functions. A rule of thumb is derived to decide in which data situations the IDR can be used as approximation of the HR. We illustrate the application of the rule by using a real data example.
Methods
Data generation
We simulated data situations with identical and with different average followup durations in the control and intervention group. The average followup duration in the control group relative to the intervention group varied from 100% to 30% (in steps of 10%, i.e., 8 scenarios). To simulate a variety of study situations, we chose 9 different baseline risks (BLRs) (BLR = 0.01, 0.02, 0.05, 0.075, 0.1, 0.15, 0.2, 0.25, and 0.3), 7 different effect sizes (HR = 0.4, 0.7, 0.9, 1, 1.11, 1.43, and 2.5), and 3 different sample sizes (N = 200, 500, and 1000, with 1:1 randomization). The BLR is the absolute risk of an event in the control group over the actual followup period in the control group. The parameters of the survivaltime distributions were chosen so that the specified baseline risks and effect sizes are valid for the corresponding followup duration in the control group and the HR for the comparison treatment versus control, respectively. We considered 1 situation with decreasing hazard function (Weibull distribution with shape parameter ν = 0.75) and 3 different situations with increasing hazard function (Gompertz distribution with shape parameter α = 0.5, 0.75, 1) because the case of increasing hazard was expected to be the more problematic one. The corresponding scale parameters λ for both the Weibull and the Gompertz distribution varied depending on the baseline risk and the followup duration in the control group.
First results showed that in some situations with relative average followup durations in the control group of 80%, 90%, and 100%, the IDR has adequate properties for all baseline risks considered. Therefore, additional simulations were performed in these cases with larger baseline risks (0.5, 0.7, 0.9, 0.95, and 0.99). In total, the combination of 4 survival distributions with 8 or 3 relative followup durations, 9 or 5 baseline risks, 7 effect sizes, and 3 sample sizes resulted in (4 × 8 × 9 × 7 × 3) + (4 × 3 × 5 × 7 × 3) = 7308 different data situations.
We included only simulation runs in which at least 1 event occurred in both groups and the estimation algorithm of the Cox proportional hazard model converged. If at least one of these conditions was violated a new simulation run was started, so that for each of the 7308 data situations 1000 simulation runs were available. This procedure leads to a bias in situations in which simulation runs frequently had to be repeated (very low baseline risk, low sample size). However, this problem concerns both IDR and HR and it was not the goal of the study to evaluate the absolute bias of the estimators.
Data analysis
The Cox proportional hazards model was used for point and interval estimation of the HR. All analyses were performed using the R statistical package [11].
Performance measures
The primary performance measure is given by the CP, which should be close to the nominal level of 95%. To identify data situations in which the IDR can be used as adequate approximation of the HR we used the criterion that the CP of the 95% CI should be at least 90%. A rule of thumb was developed depending on the relative average followup duration in the control group and the baseline risk, to decide whether or not the IDR can be used as a meaningful approximation of the HR.
Results
Simulation study
In the situations considered in the simulation study it is not problematic to use the IDR as approximation of the HR if the average followup durations in both groups are equal and the BLR is not larger than 25%. The minimum CP of the interval estimation of the IDR is 92,5% (CP for HR 93,4%) for the Weibull and 91,2% (CP for HR 93,1%) for the Gompertz distribution. There were no relevant differences between the IDR and HR estimations in bias or MSE (results not shown). This means that even in the case of nonconstant hazard functions but a constant HR, the IDR  independent of the effect size and the sample size  can be used as approximation to the HR if the average followup durations in both groups are equal and the BLR is not larger than 25%.
The situation is different in the case of unequal average followup durations in the two groups, which is the more important case in practice. In this situation, there are shortfalls in the CP and in part large relative bias values for the IDR. The CP decreases remarkably under the nominal level of 95% with increasing difference in the average followup durations between the groups. The CP improves with decreasing sample size, due to the decreasing precision. Therefore, the sample size of N = 1000 is the relevant situation for the derivation of general rules.
Results for the Gompertz distribution
BLR  True HR  CP  MPE  MSE  SE  

IDR  HR  IDR  HR  IDR  HR  IDR  HR  
0.01  0.4  0.976  0.978  −22.678  −9.860  0.595  0.580  0.026  0.027 
0.7  0.964  0.978  −45.271  −5.351  0.612  0.634  0.023  0.024  
0.9  0.983  0.989  − 128.131  1.149  0.466  0.461  0.021  0.022  
1  0.977  0.976  NA  –  0.458  0.455  0.020  0.021  
0.02  0.4  0.956  0.970  −7.247  7.156  0.369  0.404  0.018  0.019 
0.7  0.952  0.956  −34.036  1.679  0.280  0.285  0.015  0.016  
0.9  0.943  0.953  − 118.042  4.062  0.240  0.243  0.014  0.015  
1  0.956  0.973  NA  NA  0.209  0.214  0.014  0.014  
0.05  0.4  0.930  0.948  −11.000  3.343  0.145  0.149  0.011  0.012 
0.7  0.928  0.964  −35.534  1.250  0.098  0.091  0.009  0.010  
0.9  0.936  0.966  − 133.290  −13.655  0.095  0.083  0.009  0.009  
1  0.929  0.946  NA  NA  0.087  0.077  0.009  0.009  
0.075  0.4  0.931  0.970  −12.835  2.092  0.086  0.082  0.009  0.009 
0.7  0.921  0.958  − 37.182  −1.180  0.070  0.059  0.008  0.008  
0.9  0.914  0.954  − 125.979  −6.983  0.069  0.055  0.007  0.007  
1  0.916  0.945  –  –  0.065  0.053  0.007  0.007  
0.1  0.4  0.914  0.943  −11.975  2.503  0.076  0.072  0.008  0.008 
0.7  0.907  0.941  −33.896  2.140  0.061  0.052  0.007  0.007  
0.9  0.927  0.968  −102.743  13.059  0.047  0.038  0.006  0.006  
1  0.902  0.959  NA  NA  0.053  0.038  0.006  0.006  
0.15  0.4  0.885  0.942  −14.697  0.333  0.058  0.046  0.006  0.006 
0.7  0.875  0.943  −35.599  0.407  0.045  0.033  0.005  0.006  
0.9  0.888  0.953  − 115.852  0.054  0.039  0.027  0.005  0.005  
1  0.884  0.958  NA  NA  0.037  0.024  0.005  0.005  
0.2  0.4  0.851  0.949  −15.946  −1.037  0.049  0.031  0.005  0.006 
0.7  0.852  0.945  −36.576  −1.049  0.037  0.023  0.005  0.005  
0.9  0.869  0.955  − 111.602  0.545  0.031  0.019  0.004  0.004  
1  0.862  0.951  NA  NA  0.031  0.019  0.004  0.004  
0.25  0.4  0.835  0.957  −15.713  −0.142  0.043  0.025  0.005  0.005 
0.7  0.830  0.951  −36.719  −0.629  0.033  0.019  0.004  0.004  
0.9  0.854  0.950  −115.785  −5.196  0.028  0.015  0.004  0.004  
1  0.872  0.956  NA  NA  0.024  0.015  0.004  0.004  
0.3  0.4  0.829  0.950  −16.209  0.014  0.038  0.019  0.004  0.004 
0.7  0.818  0.956  −36.302  − 0.295  0.029  0.014  0.004  0.004  
0.9  0.862  0.946  − 103.272  6.879  0.023  0.013  0.004  0.004  
1  0.857  0.948  NA  NA  0.021  0.013  0.003  0.004 

The IDR can be used in the case of equal followup durations in the two groups if BLR is ≤ 25%

The IDR can be used in the case of a relative average followup duration in the control group between 90% and 100% if BLR is ≤ 10%

The IDR can be used in the case of a relative average followup duration in the control group between 50% and 90% if BLR is ≤ 1%

The IDR should not be used in the case of relative average followup durations < 50% in the control group
Maximum BLR for which CP of at least 90% is reached for interval estimation of IDR as approximation of the HR
Relative average followup time of the control group  Maximum BLR  

Weibull (decreasing hazard)  Gompertz (increasing hazard)  
α = 0.5  α = 0.75  α = 1  
30%  –  1%  –  – 
40%  –  1%  –  – 
50%  1%  2%  1%  – 
60%  2%  2%  1%  1% 
70%  7.5%  5%  2%  1% 
80%  30%  10%  2%  2% 
90%  30%  30%  20%  10% 
100%  30%  30%  30%  25% 
Other improved rules can be derived in certain situations if there is knowledge about the true survivaltime distribution. However, this requires new simulations with the specific survivaltime distribution. Without knowledge about the true survivaltime distribution, the rule of thumb presented above can be used for practical applications when there is no access to the individual patient data.
Example
For illustration we consider the IQWiG dossier assessment, in which the added benefit of enzalutamide in comparison with watchful waiting was investigated in men with metastatic prostate cancer that is not susceptible to hormoneblocking therapy, who have no or only mild symptoms, and in whom chemotherapy is not yet indicated [12]. According to the overall assessment, enzalutamide can prolong overall survival and delay the occurrence of disease complications. The extent of added benefit is dependent on age [12].
The benefit assessment was based upon an RCT, which was the approval study for enzalutamide in the indication described above. In this study, patients were randomized to either enzalutamide (intervention group) or placebo (control group), while the hormoneblocking therapy was continued in all patients. In each group, treatment was continued until either disease progression or safety concerns arose. Due to differential treatment discontinuation by treatment group, the median followup duration for safety endpoints was threefold longer in the intervention group (17,1 months) compared to the control group (5,4 months).
Here, we consider the endpoint hot flashes, which played a minor role in the overall conclusion of the benefit assessment. However, for the present study this endpoint is relevant, because interesting results are available for three different analyses. In the corresponding dossier submitted by the company, effect estimates with 95% CIs and P values were presented in the form of risk ratios (RRs) based upon naive proportions, as IDRs and as HRs. Additionally, KaplanMeier curves were presented. In each of the analyses only the first observed event of a patient was counted, i.e., there are no problems due to neglect of withinsubject correlation.
The following results were presented in the dossier for the endpoint “at least one hot flash”. In the intervention group 174 (20.0%) among n_{1} = 871 patients experienced one or more events compared to 67 (7.9%) among n_{0} = 844 patients, which leads to an estimated RR = 2.52 with 95% CI 1.93, 3.28; P < 0.0001. However, as correctly argued by the company, this statistically significant effect could be induced simply by the threefold longer median followup duration in the control group. To account for the differential followup duration by treatment group, events per 100 patient years were presented (14.7 in the intervention group and 12.4 in the control group) leading to the not statistically significant result of IDR = 1,19 with 95% CI 0.87, 1.63; P = 0.28. However, according to our pragmatic rules, the IDR should not be used if the relative average followup duration in the control group is below 50%, which is the case here. Therefore, the validity of the IDR results is questionable in this example. Fortunately, the results of the Cox proportional hazards model were also presented. The result was statistically significant with an estimated HR = 2.29, 95% CI 1.73, 3.05; P < 0.0001. It should be noted that censoring is possibly not independent of outcome, leading to high risk of bias. Nevertheless, the results of the Cox proportional hazards model are interpretable and were accepted in the dossier assessment with the conclusion of a considerable harm of enzalutamide for the endpoint hot flashes [12].
This example shows that the use of IDR is invalid in the present case of differential followup duration by treatment group and nonconstant hazard functions. From the KaplanMeier curves presented in the dossier it can be concluded that the hazard function of the endpoint hot flashes is decreasing. This situation can be illustrated as follows.
Discussion
The IDR represents a valid estimator of the HR if the true hazard function is constant. However, for nonconstant hazard functions we found that in the simulated data situations with decreasing and increasing hazard functions, the IDR is only an adequate approximation of the HR if the average followup durations in the groups are equal and the baseline risk is not larger than 25%. In the case of differential follow up by treatment group, the validity of the IDR depends on the true survivaltime distribution, the difference between the average followup durations, the baseline risk, and the sample size. As a rule of thumb, the IDR can be used as approximation of the HR if the relative average followup duration in the control group is between 90% and 100% and BLR is ≤ 10, and in the situation where the average followup duration in the control group is between 50% and 90% and BLR is ≤ 1%. The IDR should not be used for relative average followup durations in the control group below 50%, because in general the IDR represents no valid approximation of the HR and the meaning of the IDR is unclear. The usefulness of this rule of thumb was illustrated by means of a real data example.
The results and the conclusions of our simulation study are limited in the first instance to the data situations considered. We considered a wide range of effect sizes (HR 0.4–2.5), three total sample sizes (N = 200, 500, 1000) with balanced design, and four survivaltime distributions with deceasing (Weibull distribution) and increasing hazard functions (Gompertz distribution). For the baseline risk, we considered almost the complete range (0.01–0.99) in the simulations. We derived practical rules to decide in which data situations the IDR can be used as approximation of the HR. These rules should also be approximately valid for other data situations. If detailed knowledge of the underlying survivaltime distribution is available, more simulations can be performed to find improved rules for the specific data situation.
We have not investigated the amount of bias associated with different patterns of dependent censoring. In this context, the framework of estimands offers additional possibilities to deal with competing events, leading to censoring mechanisms that are not independent of the considered timetoevent endpoint [13]. We have also not considered the data situations with recurrent events. Extensions of the Cox proportional hazards model, such as the AndersenGill, the PrenticeWilliamsPeterson, the WeiLinWeissfeld, and frailty models [14, 15] have been developed for analysis of recurrent event data. The application of methods for analysis of recurrent event data to analysis of adverse events in RCTs is discussed by Hengelbrock et al. [16]. Further research is required for the investigation of the impact of dependent censoring and multiple events on the validity of the IDR.
Conclusions
In summary, in the case of large differences in the average followup durations between groups, the IDR represents no valid approximation of the HR if the true hazard functions are not constant. As constant hazard functions are rarely justified in practice, adequate survivaltime methods accounting for different followup times should be used to analyze adverse events rather than the simple IDR, including methods for competing risks [17]. However, the proposed rule of thumb allows the application of IDR as approximation of the HR in specific data situations, when it is not possible to estimate the HR by means of adequate survivaltime methods because the required individual patient data are not available.
Notes
Acknowledgements
We thank Ulrich Grouven for editorial support.
Authors’ contributions
RB conceived the concept of the study. LB carried out the simulations. Both authors drafted and reviewed the manuscript. Both authors have been involved in revisions and read and approved the final manuscript.
Funding
Not applicable.
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
The authors declare that they have no competing interests.
References
 1.Bender R, Beckmann L, Lange S. Biometrical issues in the analysis of adverse events within the benefit assessment of drugs. Pharm Stat. 2016;15(4):292–6. https://doi.org/10.1002/pst.1740.CrossRefPubMedPubMedCentralGoogle Scholar
 2.Liu GF, Wang J, Liu K, Snavely DB. Confidence intervals for an exposure adjusted incidence rate difference with applications to clinical trials. Stat Med. 2006;25(8):1275–86. https://doi.org/10.1002/sim.2335.CrossRefPubMedGoogle Scholar
 3.Siddiqui O. Statistical methods to analyze adverse events data of randomized clinical trials. J Biopharm Stat. 2009;19(5):889–99. https://doi.org/10.1080/10543400903105463.CrossRefPubMedGoogle Scholar
 4.Stein AS, Larson RA, Schuh AC, Stevenson W, LechMaranda E, Tran Q, Zimmerman Z, Kormany W, Topp MS. Exposureadjusted adverse events comparing blinatumomab with chemotherapy in advanced acute lymphoblastic leukemia. Blood Adv. 2018;2(13):1522–31. https://doi.org/10.1182/bloodadvances.2018019034.CrossRefPubMedPubMedCentralGoogle Scholar
 5.Zink RC, Marchenko O, SanchezKam M, Ma H, Jiang Q. Sources of safety data and statistical strategies for design and analysis: clinical trials. Ther Innov Regul Sci. 2018;52(2):141–58. https://doi.org/10.1177/2168479017738980.CrossRefPubMedGoogle Scholar
 6.IQWiG. Abirateronacetat (neues Anwendungsgebiet) – Nutzenbewertung gemäß § 35a SGB V, Auftrag A13–06, Version 1.0 vom 11.04.2013. https://www.iqwig.de/download/A1306_Abirateronacetat_neues_Anwendungsgebiet_Nutzenbewertung_35a_SGB_V.pdf. Accessed 27 June 2019.
 7.Kraemer HC. Events per persontime (incidence rate): a misleading statistic? Stat Med. 2009;28(6):1028–39. https://doi.org/10.1002/sim.3525.CrossRefPubMedGoogle Scholar
 8.Kunz LM, Normand SL, Sedrakyan A. Metaanalysis of rate ratios with differential followup by treatment arm: inferring comparative effectiveness of medical devices. Stat Med. 2015;34(21):2913–25. https://doi.org/10.1002/sim.6530.CrossRefPubMedPubMedCentralGoogle Scholar
 9.Bender R, Augustin T, Blettner M. Generating survival times to simulate Cox proportional hazards models. Stat Med. 2005;24(11):1713–23. https://doi.org/10.1002/sim.2059.CrossRefPubMedGoogle Scholar
 10.Deeks J, Higgins JPT, Altman D (editors) on behalf of the Cochrane Statistical Methods Group. Chapter 9: Analysing data and undertaking metaanalyses. In: Higgins JPT, Churchill R, Chandler J, Cumpston MS, editors. Cochrane handbook for systematic reviews of interventions, Version 5.2.0 (updated June 2017). http://www.training.cochrane.org/handbook. Accessed 02 February 2018.
 11.The R Foundation. The R project for statistical computing. https://www.Rproject.org. Accessed 27 June 2019.
 12.IQWiG. Enzalutamid (neues Anwendungsgebiet) – Nutzenbewertung gemäß § 35a SGB V, Auftrag A14–48, Version 1.0 vom 30.03.2015. https://www.iqwig.de/download/A1448_EnzalutamidneuesAnwendungsgebiet_Nutzenbewertung35aSGBV.pdf. Accessed 27 June 2019.
 13.Unkel S, Amiri M, Benda N, Beyersmann J, Knoerzer D, Kupas K, Langer F, Leverkus F, Loos A, Ose C, et al. On estimands and the analysis of adverse events in the presence of varying followup times within the benefit assessment of therapies. Pharm Stat. 2019;18(2):166–83. https://doi.org/10.1002/pst.1915.CrossRefPubMedGoogle Scholar
 14.Amorim LDAF, Cai J. Modelling recurrent events: a tutorial for analysis in epidemiology. Int J Epidemiol. 2015;44(1):324–33. https://doi.org/10.1093/ije/dyu222.CrossRefPubMedGoogle Scholar
 15.Wei LJ, Glidden DV. An overview of statistical methods for multiple failure time data in clinical trials. Stat Med. 1997;16(8):833–9. https://doi.org/10.1002/(SICI)10970258(19970430)16:8<833::AIDSIM538>3.0.CO;22.CrossRefPubMedGoogle Scholar
 16.Hengelbrock J, Gillhaus J, Kloss S, Leverkus F. Safety data from randomized controlled trials: applying models for recurrent events. Pharm Stat. 2016;15(4):315–23. https://doi.org/10.1002/pst.1757.CrossRefPubMedGoogle Scholar
 17.Schmoor C, Bender R, Beyersmann J, Kieser M, Schumacher M. Adverse event development in clinical oncology trials. Lancet Oncol. 2016;17(7):e263–4. https://doi.org/10.1016/S14702045(16)302236.CrossRefPubMedGoogle Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. 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.