Anders Pape Møller
Decision on Scientific Dishonesty
In the autumn of 2003 Anders Pape Møller (APM) was found guilty in scientific dishonesty/misconduct by the Danish Committees on Scientific Dishonesty, designated UVVU in Danish and DCSD in English.
In 2005 APM published a remarkable commentary in "ISBE Newsletter". Comments and correspondence in relation to this commentary are included.
In January 2007 the magazine "The Scientist" published a critical article of the Anders Pape Møller case The fluctuating reality - from superstar to pariah. The article also lists an easy-to-view time line of key documents and milestones.
Other presumed cases of scientific misconduct by APM
Dutch Elm Disease
The following is meant as a debate article to one or another scientific journal. OIKOS was the obvious choice as Moller (1999) was published in this journal. However, OIKOS still protects and favors APM. I already sent this Opinion contribution to OIKOS but they never answered.
Fluctuating asymmetry and Dutch elm disease (Møller 1999)
As mentioned in Nature (29 Jan. 2004) and Science (30 Jan. 2004) Anders Pape Møller was found guilty in scientific misconduct (data constructions) of Danish UVVU in the now retracted paper by Møller & de Lope (1998).
Obviously, it must be important for APM to make probable that this misconduct was an atypical single case in an otherwise outstanding and clean career.
Now, the UVVU case against APM started about 5 years when some obvious errors and improbabilities appeared in the paper of Møller & de Lope (1998). Such errors and improbabilities are not uncommon in papers by APM. In the following we have a question concerning Møller (1999) - with reference also to Møller (1995).
Møller (1999, Tables 1 to 3) investigated 39 pairs of elm trees where all 78 trees in the first year (1993) were healthy, i.e. not attacked by Dutch elm disease. In course of a four years period a single tree in each pair developed disease. According to Møller (1999) "the disease is still relatively uncommon at the Danish study site with less than 5% of all elm trees being infected". I.e. less than 5% were infected when the investigation was finished in 1997. Apparently, the 39 pairs of trees was a sub-sample in a sample of 43 pairs (or perhaps a larger sample of 43 pairs + 25 extra pairs = 68 pairs) of trees (Møller 1995). We have problems understanding how the 39 pairs (consisting of one healthy tree and one diseased tree) arose. If 5% of the trees in the population developed disease in course of the period the following frequencies of non-diseased (0/0) pairs, one tree diseased (1/0), and two trees diseased (1/1) pairs should be expected: 0.9025, 0.095, and 0.0025, respectively. In a sample of 43 or 68 pairs this means 4.09 or 6.46 1/0 pairs, respectively. Møller reported 39 pairs! Clearly we need a plausible explanation.
Møller, A.P. 1995: Leaf mining insects and fluctuating asymmetry in elm (Ulmus glabra Huds.) leaves. - J. Anim. Ecol. 64, 697-707.
Møller, A.P. 1999: Elm, Ulmus glabra, leaf asymmetry and Dutch elm disease. - Oikos 85, 109-116.
Møller, A.P. & F. de Lope 1998: Herbivory affects developmental instability of Stone oak, Quercus rotundifolia. - Oikos 82, 246-252.
Preferable APM should not know these notes. However, knowledge will not help him much.
First, I am convinced that APM made use of data constructions in the 1999-paper. The task is to make it highly probable that he did so.
After about five years of experience with APM I know that he lies if necessary (and also sometimes when not necessary) and adapt his answer to the question if answering the question at all. Therefore it is important to lock his answer along a certain line, then in the next step his degrees of freedom are restricted.
Four possible answers of APM could be something like:
- In 1993 I started measuring (about) 400 pairs of non-diseased elm trees and in course of the four year period - measuring 10 leaves of (about) 800 trees every year - 39 pairs appeared in which one tree was diseased and the other tree was (still) healthy.
- I followed the 43 + 25 pairs over the four years, and though the probability for obtaining 39 one diseased/one healthy pairs is very small such a low-probability event sometimes occurs.
- The 43 pairs of elm trees consisted of one tree heavily infested with leaf mines and there is a very strong positive coupling between (susceptibility to) leaf mines and Dutch elm disease, i.e. (almost) all trees heavily infested with leaf mines also developed Dutch elm disease in course of the next few years.
- Less than 5 were infected is wrong. It is an error of writing. Less than 50 is the correct notion. 50 of infection and 43 25 68 pairs mean, that we should expect about the half, i.e. 34 pairs in which one tree is infected and the other tree uninfected. There is no significant difference between the 39 pairs observed and the 34 pairs expected.
Comments to these answers:
- Møller (1999) gives no number of start pairs, but according to Møller (1995) the number of trees investigated over the first two years were 43 pairs + 25 pairs + 60 single trees. If the latter is transformed to pairs this make an absolute but not reasonable maximum of 98 pairs. According to Moller (1999, p.110) A total of 10 randomly chosen leaves without mines where chosen for measurements from each tree during a study of the relationship between foliar asymmetry and leaf mining (Moller 1995b). Pairs of nearest neighbouring trees were used for the study ensuring that trees were relatively similar with respect to growth conditions and hence size. Therefore Moller cannot with any kind of reliability maintain that he started with about 400 pairs of elm trees in 1993. Clearly, Moller measured no more than 196 trees.
- Møller (in litt.) used such kind of argumentation in his correspondence to UVVU.
- Møller (1999) mentioned nothing about such a positive correlation and causal relation (which also would be rather fatal for his paper focusing alone on Dutch elm disease). Furthermore, in the second of the FA-measures (inter-rib-distance) there is no significant correlation between FA and number of leaf mines (the insignificant correlation is even negative). Therefore, it is extremely unlikely that number of leaf mines is a very strong indicator of susceptibility for Dutch elm disease. Finally, the 39 pairs of trees (1999) cannot be considered a subsample of the 43 pairs (1995) as most t test comparisons (absolute FA) between lower part of Tab.2 Moller (1995) and Tabs.1 to 3 Moller (1999) are statistically significant.
- If 50 of the trees were infected then the phrase still relatively uncommon makes no sense.
Tail length and tail asymmetry in male Barn Swallows
Nachman & Heller (1999) were sceptical about fluctuating asymmetry as an index of fitness, and focused on three examples; one of these tail asymmetry of male Barn Swallows, because this trait is considered to be a secondary sexual character involved in female mate preference (Møller 1991, 1994). As a basis for their investigation and critique Nachman & Heller (N & H, pers. comm.) of course wanted to make use of real measurements and therefore as the logical step contacted Anders Pape Møller (APM) for releasing relevant material. After a lengthy correspondence APM finally sent Gøsta Nachman (GN) data for the seven years 1990 to 1996 and for 560 individual birds ordered after tail asymmetry, i.e. it was not possible to distinguish between the years (Tail.xls, copy of a paper outprint sent from APM to GN).
When GN omitted males with a very asymmetrical tail (more than 12 mm) there was no influence of tail asymmetry in the logistic regression (SAS programme) based on 542 individuals (P was 0.0175, 0.0130 and 0.4099 for tail length, tail length squared and tail asymmetry, respectively).
Faced with this unpleasant scenario APM interrupted the cooperation, and told GN that the heterogeneity between the years was too high to permit a pooling of all years.
N & H then had to write their paper based on simulated data and known properties, and as the reaction a commentary from Møller et al. (1999) appeared.
In Møller et al. (1999) there is a Table 1 from where the following data are taken from the years 1990 to 1996.
|Year||Unmated males (%)||N||Tail||Tail squared||Tail asymmetry|
N is total number of males and on the basis of the percentages the number of unmated males could be calculated as 4, 3, 10, 6, 4, 1, and 6, respectively (8.0% should probably be 8.5% in order to get a whole number of unmated males). The sum is 34 unmated males, out of a total of 560 for the 7 years.
Tail is tail length, and what is depicted is the "results of logistic regression analyses between male mating success (dependent variable) and tail length, tail length squared and tail asymmetry (independent variables) in male barn swallows. Values are chi-square values for the different independent variables, with the sign of the regression coefficient being indicated" (Møller et al. 1999). The underlining is mine because the sign of the regression coefficients is important for the following discussion.
Compared with the material sent from APM to GN, the total number of males is the same, 560. However, the number of unmated males in Table 1 is 34, whereas this number is 40 in the material delivered by APM to GN (Tail.xls). The last 6 birds presented in this file were all unmated and had the very most asymmetrical tails of all the birds (22, 40, 45, 46, 49 and 74 mm). Clearly, 34 are 40 minus 6 but then the total sample size for the seven years of Table 1 should be 554. So in some way the material in Table1 in Møller et al. (1999) is adjusted compared with the material sent from APM to GN.
I carried out multiple logistic regressions based on 1) all 560 birds in the material, 2) 554 birds (tail asymmetry less than 20 mm included), and 3) 552 birds (tail asymmetry less than 18 mm included).
The probabilities based on tail length, tail length squared and tail asymmetry, respectively, were
- 0.0150, 0.0112, and < 0.0001
- 0.0153, 0.0114, and 0.0353
- 0.0173, 0.0128, and 0.0519
As already mentioned the probabilities of the analysis of GN based on 542 birds (tail asymmetries more than 12 mm omitted) were 4) 0.0175, 0.0130, and 0.4099, respectively.
It is easy to see how the influence of asymmetry disappears when the very asymmetrical individuals are omitted. On the other hand the influences of the two other variables are rather constant.
Compared with the analyses of Møller et al. (1999, Tab.1) the sign of the regression coefficient of tail length was negative in all four (sub)samples mentioned above, whereas the sign of this coefficient was positive for all seven years in the Møller et al. analyses. Is this a possible real world scenario: a negative coefficient for the total sample, and positive coefficients for all seven sub-samples of the total sample?
May be this has something to do with the claim of APM of heterogeneity between the years, but to me (in symphony with the uncertainties about the total sample size and the number of un-paired males) it looks like a strong indication that APM made use of fabricated data (as done previously in Møller & de Lope 1998) because he intuitively supposed that the sign of both tail length and tail length squared should be positive (if tail length squared is omitted from the logistic regression in our analyses the sign of the regression coefficient of tail length certainly changes to positive).
Far from being a mathematician, I needed an expert for answering the question about the real world scenario. GN had no further intentions moving into the case, and therefore I contacted another person familiar with logistic regression in biology. This person at least at the present stage wants to be anonymous.
I asked the expert the simple question whether the seven sub-samples of Tab.1 (Møller et al. 1999) could originate (with a reasonable probability) from the sample sent from APM to GN. In particular I wanted to know whether it was possible that the sign of the regression coefficients of tail length could be positive in all seven sub-samples in Tab.1 but negative in the total sample as calculated by me and GN.
As always, experts have problems focusing on and answering such simple questions because they see much more depths and complexity than the ignorant person asking the question.
First, the expert had some general considerations not directly relating to my question. 1) The answer of Møller et al. (1999) was largely meaningless in relation to the rather profound critic raised by N & H. 2) The application of logistic regression with four parameters on the data presented is highly problematic - in particular on the years separated where the number of non-paired males varied between 1 and 10. The risk of over-fitting is very high. According to the expert even the total sample (Tail.xls, n = 560 and 40 un-paired males) are not fulfilling the conditions for carrying out logistic regression with three independent variables, nor constituting the basis for a reasonable control of the model.
Next, applying logistic regression with the variables considered the expert investigated many sub-sets of the total sample mimicking the yearly sub-samples of Møller et al. In all these analyses the sign of variables tail length and tail length squared was opposite - as already mentioned for the analyses performed by me and GN. Furthermore, the influence of tail length squared was always statistically significant - contrary to the yearly probabilities indicated by Møller et al. (Tab.1).
Summing up, the expert considers it extremely unlikely that the seven sub-samples presented by Møller et al. originated in the total sample (Tail.xls) delivered by APM to GN, and the expert recommends APM to carry the burden of proof by presenting the raw data of the yearly sub-samples used for the analyses.
Møller, A.P. 1991. Sexual selection in the monogamous barn swallow (Hirundo rustica). I. Determinants of tail ornament
size. - Evolution 45: 1823-1836.
Møller, A.P. 1994. Sexual selection in the barn swallow (Hirundo rustica). IV. Patterns of fluctuating asymmetry and selection against asymmetry. - Evolution 48: 658-670.
Møller, A.P. & F. de Lope 1998. Herbivory affects developmental instability of stone oak, Quercus rotundifolia. - Oikos 82: 246-252.
Møller, A.P., S.W. Gangesstad & R. Thornhill 1999. Nonlinearity and the importance of fluctuating asymmetry as a predictor of fitness. - Oikos 86: 366-368.
Nachman, G. & K.E. Heller 1999. Fluctuating asymmetry as an index of fitness: causality or statistical artifact? - Oikos 86: 357-365.
The measurements of 560 male Barn Swallows delivered by APM to GN.