Send letter to journal:
Re: Re: No provable effects

Dear Editor,

We thank Wolf and Vana for their comments on our article "Subjective symptoms, sleeping problems and cognitive performance in subjects living near mobile phone base-stations" (OEM 63:307-313). We appreciate their regret that due to methodological problems results may not be as clear-cut as they desire.

The study of potential effects of emissions from mobile phone base- stations is indeed fraught with intriguing methodological problems, however, we admit that those raised by Wolf and Vana are not among those we were primarily concerned about.

1. Concerning the problem of random clustering that could reduce effective sample size via a joint influence of an unknown factor on one or several outcome variables on all subjects within that cluster, we have taken substantial precaution, in the statistical and material sense. We did all analyses first by inclusion of the base-station as a random factor (the variance component of which is equivalent to the intra-class correlation). These analyses revealed no indication of clustering. Furthermore, we measured in homes of participants in the urban area volatile organic compounds and did ambient noise measurements in the rural area. There were no differences between base-stations in these environmental factors and no correlation between exposure to electromagnetic fields and indoor VOC concentrations was detected.

2. Strong concerns were expressed by 8% of participants and there was indeed a correlation with distance to the base station and, to a lesser degree, also with exposure category (mean ratings on a scale 0 to 3 were 0.7, 0.9, and 1.3 for the exposure categories, respectively). Hence concerns about adverse effects of base-stations is indeed a confounding variable because, as we stated in the discussion, it was also correlated with many symptoms. Likert type scale of concerns was area transformed before inclusion into ANCOVA.

3. Dichotomisation of symptoms of the Zerssen Scale was done applying the only objective cut-off, i.e. discriminating between those that did not experience this symptom from those that did (thereby neglecting strength of symptoms). The reasons for applying a dichotomisation were manifold, but the most important one was to enable usage of the same method of evaluation for all symptoms. Some symptoms were never or only rarely categorized into the highest category and therefore applying ordinal regression led to parameter identification problems. Scoring several symptoms together into groups of symptoms has not found approval by reviewers.

4. It is correct that the model of analysis must be set beforehand. Therefore, results were given for the model as specified. However, to give an idea about the effect of loss of power by inclusion of insignificant covariables, following reviewers' suggestions, also the result after their omission was presented in one case. Application of the Bonferroni or Bonferroni-Holm method is up to the reader. He/she should only be aware that this method is utterly conservative and will substantially increase type II error in the case of correlations between dependent variables (which is the rule for symptoms, performance variables, and sleep quality indices). In exploratory studies that cannot start from evidence based hypotheses such correction will render even much bigger trials practically powerless and is therefore not recommended.

Send letter to journal:
Re: No provable effects

Dear Editor,

In their publication, the authors postulate that effects upon health and performance cannot be ruled out despite a low exposure to high-frequency electro-magnetic fields, effects far below the WHO threshold values. Unfortunately, this paper has substantial methodological problems.

1. There may well have been a clustering in the choice of test- subjects’ addresses in relation to the location of the mobile phone base stations. For example, in the sense that other environmental influences may have affected everyone in the same cluster. In this case, the independence of the people within the cluster may no longer be given, which in turn may lead to an overestimation of the effective sample size, and therefore to progressive (results are too often significant) decisions. In order to control this effect, the Intra-Class-Correlation should have been calculated to give an impression of the effective sample size. In the light of the few, tightly significant results, such effects may have played a decisive role.

2. It still remains unclear as to whether there are a priori differences as to the degree of “concerns about base station“ in relation to the expositional groups. This was unfortunately not presented in table 1. What was the percentage of test-persons who were concerned, and to what extend? How is this distributed throughout the groups? In addition, the degree of concern was measured on a ranking-scale, despite the fact that the co-variance analysis demands an interval scaled co-variant.

3. The Zerssen Scale was evaluated dichotomously, whereby even very slight complaints (“1“) were already registered as complaints. This then raises the question as to whether the named effects in the Zerssen scale would still be evident if one had pooled the groups 0 and 1 (no, and/or light complaints), and the complaints-group 2 and 3. Who doesn’t sometimes suffer of headaches, dizziness, and appetite disorders, etc? It would be interesting to see how the complaints - before a dichotomization – were distributed amongst the exposition-groups.

4. In Table 2, 13(!) covariance analyses were carried out, from which the p-value of one main effect reaches tendency level (p<0.1). After the implementation of a Bonferroni-Holm-Correction, which strictly seen should also include the p-values inside the covariance analyses, there are no effects to be found whatsoever. In this case, it is then not permitted to omit the non- significant co-variables, even in the co-variance analysis. The model for the covariance analysis must always be set, a priori, before testing; otherwise a Type I-error-inflation will ensue. The attained significance would only then be credible if confirmed on the basis of a further sample.

5. If one would adjust the p-values in table 4 according to the Bonferroni- or Bonferroni-Holm-convention, then none of the symptoms would be significant anymore. The same would apply to the effects shown in table 5.