Neurology -- Correspondence for Pugliatti et al., 58 (2) 277-282

Correspondence published:

Multiple sclerosis distribution in northern Sardinia: Spatial cluster analysis of prevalence: John Parrett, P Donnan, RB Forbes, SV Wilson, J O'Riordan and RJ Swingler (2 May 2002)

Reply to Letter to the Editor: Maura Pugliatti, G. Solinas, S. Sotgiu, P. Castiglia, and G. Rosati (2 May 2002)

Multiple sclerosis distribution in northern Sardinia: Spatial cluster analysis of prevalence 2 May 2002

John Parrett
Ninewells Hospital Dundee Scotland,
P Donnan, RB Forbes, SV Wilson, J O'Riordan and RJ Swingler
Send Correspondence to journal:
Re: Multiple sclerosis distribution in northern Sardinia: Spatial cluster analysis of prevalence

jdeparratt{at}hotmail.com John Parrett, et al.

We commend Pugliatti et al. on their thorough examination of MS prevalence in northern Sardinia. [1] The authors used a " spider " methodology [2] to ascertain cases of MS from multiple sources. Furthermore, the spatial variation of MS was analyzed using Bayesian methods for both residences at symptom onset and childhood.
Although Pugliatti et al are likely to have identified the majority of MS cases, no formal test of ascertainment was made. Theoretically, sampling bias may have occurred secondary to variation in geographic or demographic factors at a district/commune level. It may be prudent, therefore, when examining potential MS clusters to adjust the identified prevalence of the "cluster" and "non-cluster" cohorts for under or over ascertainment using capture-recapture methods.
Capture-recapture methods have been used in a spectrum of epidemiological disease surveys, including MS [3], to estimate completeness of ascertainment. Briefly, if a study utilizes more than one source of information to identify cases, the number of missing cases may be estimated via the overlap between sources. In human populations, certain violations of underlying statistical assumptions (notably dependency between sources) may occur, but such effects can be reduced by, for example, loglinear modeling.
Cluster analysis of MS has proven difficult either because the observation is post-hoc or the statistical methods used do not take account of multiple testing inherent in such procedures. [4] Bayesian methods reduce the effects of random variability when groups of smaller populations are analyzed by pulling extreme values towards the mean. The authors identified "hotspots" but a formal test of significance for any given cluster was not carried out. To reconcile the need for formal significance testing and account for random variability, we suggest use of the spatial scan statistic. The spatial scan statistic is a generalization of a test first proposed by Turnbull. [5] Multiple concentric windows are generated from the centroid of each commune/sector and varied in size up to a pre-specified maximum of the total regional population (50% being a sensible limit). The null hypothesis is tested that the number of cases within and out of each window is equal according to a Poisson distribution of the underlying populace. The test results in a maximum likelihood ratio, the distribution of which is generated by a large number (say 999) of random Monte Carlo replications of the dataset. Thus, the approximate location and magnitude of a potential cluster is delivered, with a priori statistical testing, whilst accounting for underlying population variance. The procedure can also take into account any number of covariates.
References:
1) Pugliatti M, Solinas G, Sotgiu S, Castiglia P, Rosati G. Multiple sclerosis distributionvin northern Sardinia: Spatial cluster analysis of prevalence. Neurology 2002;58:277-282.
2) Kurtzke JF. The eipdemiology of multiple sclerosis. In: Raine CS, McFarland HF,Tourtellotte WW, editors. Multiple sclerosis: clinical and pathogenetic basis. 1 ed. London: Chapman and Hall Medical;1997:91-139.
3) Forbes RB, Swingler RJ. Estimating the prevalence of multiple sclerosis in the United Kingdom by using capture-recapture methodology. Am J Epidemiol 1999;149:1016-1024.
4) Riise T. Cluster studies in multiple sclerosis. Neurology 1997:49 (Suppl 2):S27-S32.
5) Turnbull BW, Iwano EJ, Burnett WS, Howe HL, Clark LC. Monitoring for Clusters of disease: Application to leukemia incidence in upstate New York. Am J Epidemiol 1990;132 Suppl 1):S136-S143.

Reply to Letter to the Editor 2 May 2002

Maura Pugliatti
University of Sassari Italy,
G. Solinas, S. Sotgiu, P. Castiglia, and G. Rosati
Send Correspondence to journal:
Re: Reply to Letter to the Editor

maurap{at}uniss.it Maura Pugliatti, et al.

We appreciate the comments of Parratt et al. We agree that a capture- recapture method should be used to guarantee ascertainment completeness in epidemiological disease surveys. This methodology could not be applied to our study. In fact, case ascertainment was based on previously established databasing resulting from sources that were undistinguishable at the time of the current study. Because the "spider" methodology used could allow us to verify cases notified for previous repeated assessements, a homogeneous and complete case collection finalized to cluster studies was in our opinion guaranteed.
In particular, we felt that the key assumptions of capture-recapture methods were preserved: (1) the study was conducted on a 'close' population whose composition remained stable over time due to a low migration flow and with no patient systematic migration to seek for better care; (2) the matching of patients by unique code names and anagraphics was possible by means of accurate capillarized ascertainment, allowing to rule out for "false positive matches", (i.e, underestimation from erroneous linking of notifications of different individuals), and "false negative matches", (i.e. overestimation from failing to identify same individuals); (3) the potential for misdiagnosis or misclassification was dealt with as for (2). The two other cardinal assumptions of capture- recapture methods, i.e., each patient has the same probability of being "caught" in a specific source and source independence, could actually not be explored. These last two assumptions, however, do appear to be critical also when using capture-recapture methods for disease epidemiological surveys based on medical records, because these sources are often positively-dependent, thus leading to underestimation. In fact, even in a capture-recapture context, it is often difficult to correct for source dependence mathematically, and investigators must often rely on their intuitive judgement that the sources are actually independent. As for testing the significance of "hot spots", we thank Parratt et al. for their suggestion.
In the attempt at ruling out the potential for non significant Bayesian estimates, the model was assigned the proper prior via a sensitivity analysis as described by Bernardinelli et al [1, 2 and Pascutto et al. [3] The posterior probability (PP), obtained by the final distribution of prevalence values combining the information in the a priori model and in the data, is a Bayesian analogous of formal p value. Moreover, its mapping suggests which areas should be investigated solving the problem of multiple significance testing. The suggested Turnbull's method, developed on Openshow's idea [4], is an alternative approach to cluster detection and it is particularly useful and powerful when a hot spot localization is assumed. Our knowledge on genetic and behavioral characteristics of the Sardinian population brought us about the conviction that neighboring areas were more similar than more distant ones, regardless of hot spots localization.
References:
1. Bernardinelli L, Clayton D, Montomoli C. Bayesian estimates of disease maps: how important are priors? Stat Med 1995;14:2411-2431.
2. Bernardinelli L, Pascutto C, Montomoli C, Gilks W. Investigating the genetic association between diabetes and malaria: an application of Bayesian ecological regression models with errors in covariates. In: Elliott P, Wakefield JC, Best NG, Briggs DJ, eds. Spatial epidemiology. Methods and applications. New York: Oxford University Press Inc 2000:286- 301.
3. Pascutto C, Wakefield JC, Best NG, et al. Statistical issues in the analysis of disease mapping data. Statist Med 2000;19:2493-2519.
4. Openshaw S, Craft AW, Charlton M, Birch JM. Investigation of leukaemia clusters by use of a Geographical Analysis Machine. Lancet 1988;1:272-273.