]\gt ; INDLX SLIP .
PKARSON , Karl , and , Alice.\mdash ; On the Distribution of Frequency Variation and Correlation of the Barometric Height at divers Stations .
Phil. Trans. , , 1897 , vol. 190 , pp. 423-469 .
, Alice , and PEAIiSON , Karl .
the ibution of Vallation and Correlation ) of the Barometric Height at quencydivers btations .
Phil. Trans. , , 1897 , vol. 190 , pp. 423-469 .
Barometel -recording Frequency.barometer , by G. Yule .
PEARSON , Karl , and LEE , Alice .
Phil. Trans. , , 1897 , vol. 190 , pp. 423-469 .
Barometric Heights , -distribution of , at twenty-three Stabions in British Isles ; Correlation of ; Prediction formulae .
PEARSON , Karl , and LEfi , Alice .
Phil. Trans. , , 1897 , vol. 190 , pp. 428-469 .
Isobars , eralised , for Frequency of British Isl es .
PEARSON , Karl , and LEE .
Alice .
Phil. Tlans .
, 1897 , vol. 190 , pp. 423-469 .
XI .
the of oj rometrtC Height Divers KARL PEARSON , F.R.S. , ALICE LEE , B.A. , B.Sc ' .
Received Junc 15.\mdash ; Read June 17 , 1897 .
[ PLATES 9-17 .
] I.\mdash ; -ON THE FREQUENCY AT DrVERS STATIONS 0F THE VERA L BAROMETRIC HEIGHTS .
1 .
Introduction .
LET a curve be formed such , that if be the ordinate the abscissae and , the area ySx represents the frequency of the barometer with lying between and Sx for any locality .
This curve will be spoken of as the for the given .
locality .
The curve in any series of actual ervations will be represented by polygonal line ; in the presenb case the ele[nent Sx has been throughout taken as inch of barometric height .
Such frequency curves occur in innumerable physical , and economic investigations , and can in many cases be fairly ttely represented by the normal curve of frequency , , , LAPLACE'S curve of errol .
S. The bal.ometric frequency curve is , however , a marked exception to this rule .
The mean barometric is very ' from coiuciding with the ' mode ' or height of maximum frequency .
While barometric frequency curves are remarkably smooth when a very number of observations are dealt with , the distribution of frequency does not obey the normal la but some other law which up to the present has not been fully discussed .
In a memolr published the ' Phil. Trans. , ' , vol. 186 , pp. 343-414 , a serles of generalised frequency curves are introduced , and it is shown , pp. and 382 , that the symmetry of the barometric frequency curve can probably be dealt with by one or other of these generalised curves .
The importance of this conclusion lies the fact that the distribution of barometric frequency in any locality can then be fully described by the statement of the values of three or four well defined constants .
This seems first to have beon pointed out by DR. VENN , in a letter to ' Nature , ' September 1 , 1887 .
7.3.98 .
k ) ' OF FREQUENCY OF ] ] BAROMETRIC HEIGHT AT DIVERS STATIONS .
425 Accordingly , in order to test this theory of barometric frequency , series of observations have been reduced and the frequency distributions for divers localities fitted with generalised probability curves .
On the basis of these curves the attempt has been made to answer the following questions:\mdash ; Is there any one type of curve especially chsxacberistic of barometric frequency ?
If so , what are the constants by which the distribution of this frequency can best be described ?
Does there appear to be .any numerical or geographical relation between these constants ?
and Does a knowledge of their values for a variety of localities enable us to make any statement with regard to the physics of atmospheric pressure ?
2 .
Data of Observation .
Our data were taken from the annual ' Meteorological Observations at Stations of the Second Order , ' kindly placed at our disposal by Mr. R. H. SCOTT , F.R.S. , Secretary to the Meteorological Council .
The heights dealt with are 9 o'clock morning heights .
Twenty stations were selected , sixteen in England and Scotland and four in Ireland .
The stations selected are given on the accompanying map ( page 424 ) of the British Isles .
In their choice we were guided by the desire for a fairly uniform distribution and if possible a coast distribution , ( ii .
) the extent of data available , and ( iii .
) its continuity .
The minimum number of years of continuous observation is five , the maximum thirteen .
At one or two suitable places there were gaps in the record ; at others less than five years were available ; In ] cases the records were only dealt with for whole years , to remove any irregularity which might arise from an annual periodicity in frequency .
It is hoped that the fraction of a monthly periodicity\mdash ; if there be a lunar influence on the frequency\mdash ; will not be very sensible , when the records deal with 65 to 170lunar periods .
Any long period , such as a 19-year period , would of course render less general a frequency distribution based on five to thirteen years of observation only .
The possibility of such a periodicity would render it very desirable to recalculate the frequency distributions for the same localities after , say , another ten years of observations .
Observations at the selected stations , when unreduced , were reduced by formulae kindly provided by the Meteorological Office .
The following is what we shall hel.eafter term the phical order of selected stations , with the years of observation and dates:\mdash ; VOL. CXC .
\mdash ; A. 3 1 The frequency distribution has been calculated for every .
In reckoning the frequency , any height falling midway between two-tenths inches has been split between those two-tenths half half .
Table I. of frequencies ( p. 428 ) embraces the data from which the constants of the probability curves have been calculated .
The lower or Roman figures in this table give the total number of days in which , during the years of observation , the barometer was at each particular height .
The upper , or old-style figures , give the number of per year at which the barometer ( on the basis of these observations ) may in the given locality be expected to be at each particular height .
We do not consider that Stations of the Second Order give the best possible returns .
Far from it ; the irregularities which occur in the frequency distributions at several of these stations seem to us to indicate that either the methods of observation , instruments , or the formulce of reduction are not entirely satisctory .
It is probable that the Telegraph Stations would give far better results , and for a greater number of identical years .
The ideal data would cover twenty or more well-distributed stations for identical periods of four " " lustra or twenty years .
The observations of the Stations of the Second Order , however , were readily available in print , and our object has not been so much to make a contribution * Omissions occur in the record for 1883 , so that year is not included .
Besidcs the above twenty stations , three telegraph stations , Cambridge , London ( Brixton ) , and Dover-Dungeness will be found to have their constants rer , orded in our tables .
The very great labour , however , of copying the mannscripl records discouraged a larger selection of these stations .
Our eturns are for the thirteen years 1878-1890 inclusive .
OF FREQUENCY OF TRIC HEIGHT AT DIVERS STATIONS , ) to the meteorology of the British Isles as to illustrate what we belieye to be a novel method of dealing with barometric frequency , in the hope that it may hereafter be taken up by professed meteorologists having .
their ined computators .
We need not stay to lay stress on the large amount of arithmetical work involved in calculating the constants for even the short periods dealt with in the -present memoir , it will be obvious to any one who has undertaken any statistical work of a like chalacter .
In order to say a portion of this arithmetical labour , a self-recording frequency barometer has been devised by Mr. G. A. YULE , some account of which will be found in an Appendix to this paper .
We believe that with due precautions , all instrument would be a valuable addition to those already in use at the various meteorological stations , and for the ready calculation of the frequency constants .
Yet , admitting all the disadvantages of the non-contemporary character and shol period of some of the data which we have used , we still believe that an examination of our diagrams will serye to convince the reader that our method is really capable of considerable rvice .
Putting aside one or two anomalies , we find a continuous and gradual change of the frequency constants as we pass from station to station round the coast , and it appears as easy to draw the contour lines of equal frequency as it has been hitherto supposed to be to draw isobars . .
Theory Applied .
Let Sx be the frequency of a deviation lying between and from the mean , then may be termed the power of the mean deviation .
In the case of the aussian or normal distribution of the frequency of observations .
Here all the odd 's vanish and we have the relation . .
5 .
3 .
for the even .
In particular When , therefore , the odd 's are not nearly zero , and the relation is not nearly satisfied , it is impossible on both counts to represent the frequenc , by normal distribution .
Let , then if be not zero and be not equal to three , a normal distribution is theoretically impossible .
Now a skew frequency distribution may be deduced from the skew-binomial as a limit , in the same way as it is possible to deduce the normal frequency dis- tribution as a limit the symmetrical binomial *Notably , quent comparison with tho TABLE I.\mdash ; Observed Annual Frequencies of the Several Barometric Heights ou stylejigures , annual frequeneiss .
Roman jigures , gross OF FREQUENCY OF THE BAROMETRIC HETGHT AT DIVERS STATIONS .
429 at Twenty Stations in the British Isles , and Three Supplementary Stations .
style ures , amnual , gross Tventy Statione in the British Isl Referred to its modal ( or maximum ) ordinate this frequency distribution is iven by ( ii .
) Vhere .
( iii .
) In order that this curve should fit a skew frequency distribution it is theoretically that should be zero , or at least that it should in statistical practice be small .
The theory of this curve is discussed , ' Phil. Trans vol. , p. 373 .
It is shown in the same memoir thaf if the genelalised probability curve . . . . .
will be found suitable , whereas if the generalised probability curve will , in a variety of cases , represent the frequency .
Methods for determining constants of these curves are fully desclibed , and these methods have been adopted in the present paper .
Of the curves ( ii .
) , ( iv .
) , and , it may be noted that ( ii .
) has a range of frequency limited in one direction , ( iv .
) a range limited in both directions , while , like the normal distribution , has a range unlimited in both directions .
When a frequency distribution can be satisfactorily described by a curve of form , then it is generally possible to obtain a skew binomial representing , with a fair degree of accuracy , this distribution .
In the ' Phil. Trans. ' memoir above referred to , some Cambridge barometric data are dealt with , and it is shown for this case : That is , and this barometric frequency therefore fits a curve of limited range .
meall ordinate is given by The rule is not general , just as it is sometimes , but not always , possible to represent a nol.mttl ibution by a symmctrical binomial .
OF FREQUENCY OF THE BAROMETRIC HEIGHT AT DIVERS STATIONS .
481 ( ii .
) That this expression not being large , a curve like above describes with an equally small percentage error ( p. 383 ) the frequenc distribution .
( iii .
) That a point-binomial could be found to fit the observations also with a close degree of accuracy .
The discussion of the thirteen years of Cambridge observations was not therefore conclusive as to the best means of describing barometric frequency , and more than one set of constants seemed capable of performing this function with an equal degl.ee of accuracy .
It was sufficient to indicate that the generalised probability curve in one type or another was fully capable of supplementing the obvious and admitted inadequacy of the normal curve .
The first stage of the present investigation was accordingly an inquiry as to the best type of generalised probability curve for barometric frequency .
4 .
On the of the Criterion for Divers Localities .
TABLE II .
This table shows us at once that there is a comparatively small of values for the constants and for very diverse localities .
The mean values of the constants , weighted with the number of years over which the observations extend , are as follows:\mdash ; It will thus be seen that the mean value of the criterion is small .
Actually , it would mark a frequency ibution corresponding to a curve like , but for practical istical purposes a curve like ( ii .
) , which fits closely when the criterion is even large as ( see Plate 10 , fig. 6 , of ' Phil. Trans. ' memoir cited above ) , will suffice to describe the distribution of barometric frequency .
The standard deyiations , and ] are , and ; these give for the probable errors of the means of the same three quantities ( duly weighted ) , , and respectively .
Thus the probable errors are about , and 71 per cent. of the observed means .
Thus , while it is exceedingly improbable that the mean values of and differ much from their observed values , it is very possible that the true mean value of the criterion is zero and not comparatively large and opposite values of the criterion at Markree Castle and Stonyhurst must , of course , be duly regarded , and a longer series of observations at these two stations may some day serve to indicate how far their barometric conditions are peculiar to local conditions of climate or of observation .
As a matter of fact , the curves corresponding to equations ( iv .
) or above were originally calculated for all the stations , and drawn upon the diaglams as well as those corresponding to ( ii .
) , but the mean percentage error in frequency , when tested by a planimeter , showed no very sensible improvement .
An examination of Plates 9-17 , giving the graphical representation{ .
of the observed frequency by the theoretical distribution ( the negative direction of being towards high barometer ) amply demonstrates that this limit to the skew binomial suffices to satisfactorily describe barometric frequency .
The closeness of the approximation is one rarely met with in the usual representation of variation by the normal curve of errors .
To illustrate how closely curve ( ii .
) corresponds to curve , even with a value of the criterion relatively large compared with the majority of those with which we ' are dealing , we give in the following table the frequencies per annum for Babbacombe as observed , and as calculated from ( ii .
) and respectively : *The ) error of the critel'ion for skew variation will be discussed in a forthcoming memoir by one of the authors , and it will there be seen that this probable error is frequently large .
We have to thank Mr. C. , Demonstrator in University College , for much aid in the ation of our diagrams .
Anothe instance of the same closeness is exhibited in the ' Phil. Trans. ' memoir above cited for a like with ( i-i .
) .
Cf .
Curves I. and II .
of , Plate 10 .
The criterion is here still larger , e.g. , OF FREQUENCY OF THE BAROMETRIC HEIGHT AT DIVERS STATIONS .
433 In view of the above remarks we shall confine our attention to curves of the form ( ii .
) .
We can now discuss the relations between the constants of this curve and the physical quantities associated with barometry .
5 .
Constants of Local Distribution of rometric Frequency .
It is convenient to term the height of barometer , corresponding to the maximum frequency , the ? ? .
This mode never coincides with the mean , and we shall represent them by respectively .
The divergence of the mode from the mean marks the skewness of the frequency distribution , but it is fitting to measure this divergence in terms of some standard of variation of the local distribution .
It is convenient to select as this standard , , the standard deviation of the distribution , or .
We shall represent the skewness of the frequency by .
Another interesting physical constant of the distribution is indicated by the theoretical law of frequency\mdash ; this is the possibile of barometric height in the locality .
It is directly obtained by adding to the mode or mean the possible range above either .
This maximum height we will indicate by , while shall indicate the maximum rved height .
The following relations hold between the constants of the theoretical distribution and the above physical quantities , . . . . .
, . . . .
, .
VOL. CXC.\mdash ; A. 3 Table III .
contains the values of the mean and the moments calculated directly from the observed frequencies .
In determining the moments , the observation were arded as trapezia Trans .
Table IV .
then gives he constants of the theoretical distribution and the physical constants of the frequency deduced from them .
The constants and the mean suffice to determine the form of the distribution .
But any other three constants would equally well .
, we might take the three physical constants , the frequency of the modal height , the variation , and .
the skewness of the distribution .
we might use the frequency of the mean instead of the frequency of the mode .
In a memoir , which one of us hopes to shortly publish , probable errors are worked out for the constants of any frequency curye , and it is shown that these errors are not , as in the case of a normal frequency distribution , uncorrelated .
An error , for example , in , marks a correlated error in and in the , while in the case of the normal frequency curve there is correlation between its position ( mean ) alJd shape standard deviation .
It becomes accordingly of considerable importance to determine which are the quantities in barometric frequency which have the least percentage probable errors , and then to adopt these as our standard constants of barometric frequency .
TABLE III .
unit * See and FILON , " " On the Probable Errors of Frequency Constants ract ) 'Roy . .
Proc vol. 62 , p. 173 .
OF @FRQQUNCC OF TH@ BAROMETRIC IHHI HT DIVERS STATIONS .
435 The following are some of the formulae given in the memoir aboye referred to : Probable error of mean .
Probable .
of Probable error of Probable error of the modal frequency Probable error of the mean frequency Probable error of .
Probable error ofskew-ness Correlation of errors in and Correlation of errors in and mean Correlation of errors in and mean ) .
Here is the series \mdash ; .
and the 's being the BERNOULLI numbers .
Now a consideration of these formulae gives the following results\mdash ; which will be found to be fully verified by the Table V. of calculated values\mdash ; The errors in the mean , standard deviation and skewness are very small .
The errors in both mean and modal frequencies are small , but the error in the modal frequency is considerably smaller than that in the mean frequency .
This follows from the fact that the frequency curve is horizontal at the mode , but ] opes at the mean , and accordingly any bodily shift of the curve produces far more effect on the mean than on the modal frequency .
The errors in and are large in cases very large .
The question then arises how is it possible to determine the frequency curve correctly .
The answer lies ir$ the fact that and are so highly correlated , that given a large error in , therQ OF FREQUENCY OF THE BAROMETRIC ]IGHT AT DIyERS STATIONS .
437 is little variation possible in ; it also has a large error , and these errors tend very closely to balance each other , , to retain the same shape for the frequency curve .
* Indeed it can be shown , both theoretically and empirically , that very considerable changes may be made in the constant , and the frequency curve will not sensibly change its shape provided the correlated error be made in But these results show us at once while the frequency curves as determined from and may fit\mdash ; as indeed they do\mdash ; the observations very accurately , still and are not good constants on which to base ( when treated separately ) any discussion of the relative distribution of barometric frequency with geographical position .
The best constants are , as we might hope they would be , the physical constants , namely the mean or modal height , the mean or modal frequency , the standard deviation or variability , and the skewness .
It is accordingly to these constants , as the constants , that we have specially directed our attention .
All the other quantities involved can be expressed in terms of ) .
If ( 1 be written for the skewness Sk .
have : ' In Table V. the whole system of probable and of correlation between errors is given for the 23 stations , and an inspection of them will show at a glance the degree of accuracy which may be expected from any number of years of barometric observation .
Curves will be found drawn on Plate 9 , giving the probable percentage errors of the modal frequency , and the standard-deviation with the probable errors of the mean and the skewness for any number of years up to 20 , for a station having the mean values of and .
They will , we believe , be found of service by the practical meteorologist for estimating how closely any series of observations will suffice to determine the theoretical frequency of the station .
It is true that these errors are not merely functions of the number of years of observation ; they depend also upon the local valnes of and , as an inspection of Table V. will illustrate .
Compare , for example , Southampton with Babbacombe , or Armagh with Glasgow .
The deviations of the probable errors from the graphical values owing to the variation in and may , however , be looked upon as a second deviation , and a reasonable , if rough , appreciation of the magnitude of the errors made may be obtained from Plate 9 .
It will be seen from the Table .
as well as from Plate 9 , that the value of the physical frequency constants are , for the years under consideration , given by the calculated values with a very close degree of accuracy .
* It should be noticed that the term involving in the frequency , i.e. , , becomes more and more independent of as increases .
OF FREQUENCy OF THE TRIC HEIGHT AT DIVERS STATIONS .
439 6 .
On the ) for the British Isles .
In order to reach a fair appreciatiol ] of the .
in which the distributions of frequency differ at any two local stations , it is desirable to have a standard frequency curve with which either of them may be compared .
This curve was obtained in the following manner .
The mean values of and were found ; they are recorded with their standard deviat , ions , probable errors and coefficients of variation below " " selected tions \ldquo ; in Table .
Naturally these values did not satisfy the relation , which is necessary if the curve is to be of the required type .
Accordingl small alterations Sp , S , Sa ' made in their values , proportional in each case to the corresponding probable error of , so that the relation was satisfied .
The alterations in no case amount to 1 per cent. of the corresponding value , being of the probable errors of those values .
The curve thus determined has its equation , the unit of being a day , and of one-tenth inch : the being at the mean height .
This curve we shall in future speak of as the Standard Frequency Curve for the British Isles .
It may be at first looked upon as an arbitrary curve , artificially obtained with a view to having some standard of comparison , lout when it has once been plotted on the diagrams of the barometric frequency , we can give the standard curve a physical meaning .
On examining Plate 15 , fig. xvii .
; Plate 1.1 , fig. .
and Plate 12 , fig. , it will be at once seen that the Standard Frequency Curve expresses the barometric frequency of places on a line running somewhat south of Scarborough , a trifle south of Llandudno , and slightly north of Parsonstown .
oximately we may state that the climate of Llandudno very nearly represents the standard of barometric frequency for the British Isles .
Probably there would be hardly any sensible deviation at all between the standard frequency curve and the distribution of barometric frequency for places like Hull and Chester .
A comparison of the mean frequency contour as sketched in on the map , page 424 , with Plate 13 of the Meteoro ] ogical Atlas of the British Isles , ' issued in 1883 by the Meteorological Office , shows that this contour is not very divergent from the isobar for ; it lies sensibly above the isobar given in that plate for The important suggestion now made is that a series of contour lines , generalised isobars , could be constructed , along which not only the mean barometric height would be the same , but practically all the constants which determine the distribution of ba , rometric frequency .
Our are very far from sufficient to enable us to draw such a series satisfactorily , but an examination of Plate 10 will show that another such generalised isobar passes very close to St. Leonards , mpton and Babbacombe , where in each case the fitted distribution diverges almost to the same extent from the standard ; Plate 13 , figs. .
and , and Plate 14 , fig. xiii .
shows Londonderry and some place between Glasgow and Scaleby marking a third , while Laudale , and a place about midway between Dunrobin and Dundee give a fourth ; lastly , Geldeston and Carmarthen mark a fairly satisfactory fifth generalised isobar .
These isobars are roughly sketched in on the ( page 424 ) to indicate the general grouping of stations with approximately identical frequency distributions .
The accurate determination of these isobars is a problem for the future ; it will require a knowledge of the frequency at a much larger number of stations and a delicate process of interpolation .
Turning now to the diagrams on Plates 10-17 , the reader will find the standard frequency curve marked in strokes and dots , and the local frequency curve given by a broken line .
ring these two curves , and neglecting , for the time , the irregular observation-polygon on which the latter is based , a continuous and reguIar change in the divergence of these two will be observed as the diagrams are taken in succession round , say , the coast line of gland and Scotland .
There are slight local deviations from uniformity , but on whole there can be no doubt that the distribution of barometric frequency is a perfectly uniform and continuous phenomenon over the district treated .
A accurate distribution for any station not included among the twenty dealt with could be obtained by graphical interpolation from the generalised isobars we have roughly sketched .
7 .
On the Modal Height of the Barometer : Attention has already been drawn by , MAZELLE , H. MEYER in a series of papers to the importance of modal heights , which they term Scheit lwerthe.* It seeins , however , impossible to accurate]y determine the modal heights , even if the observations be grouped in very small ranges , until a theory of skew frequency for the barometer has been adopted .
A glance at the diagrams for Churchstoke , Carmarthen , Dunrobin , \amp ; c. , will sufficiently ustrate how delusive is the peak of the observation polygon .
The true position of the mode depends , like that of the mean , on the whole series of observations , and not on the observational lnaximum only , which must always be largely the result of the grouping selected , and the elementary range taken as basis of the grouping .
In Table IV .
, under the heading obsPrved value , will be found nearly all our observation polygon by itself could tell us of the modal height .
We should be able to determine the nearest tenth to * See FECHNER : " " Ueber den Ausgangswerth der kleinsten gssumme , \ldquo ; \amp ; c. 'Abhandlungen der math.-phys .
CIasse der .
Sachsiscben ells .
der Wissenschaften , ' vol. 11 , No. 1 , 1874 ; MAZELLE : 'Wiener Denkscbriften , ' vol. 60 , p. 433 , 1893 , vol. 62 , p. 57 , 1895 ; dealing with air temperatures ; H. 'Anleitung sir Bearbeitung meteorologischer Beobachtungen , ' Berlin , 1891 , pp. 12-27 ; but compare .
HANN : 'Die Ausgabe , Einleitun OF FRLQUENCY OF THE IETRIC HLIGHT ATION S. 441 mode and to say on side of this tenth the mode lies , in Table IV .
, for means that the Glasgow modal has for its nearest tenth , but lies on the side of this .
A closer approximation to the position of the mode may be ) by dealing with the three chief fiequellcies and ffiding the vertex of a with vertical axis passing their tops , or of a normal curve of errors with vertical axis passing through the three points , the latter , according to our experience , giving the better approximation to the true mode .
Let be the unit of grouping , let be the maximum frequency , and and the frequencies on of it .
Let be distance of the mode from tcwal'ds , then we have For the parabola : , where For the normal curve : , where A third method , which is far accurate , as it depends on ali the observations , has been given the memoir on skew variation in the ' Phil. ' already cited ( see pp. 375-6 ) .
This depends upon the principie the distance of the mode from the mean is , with a close degree of , thrice the distance of the median* from mean .
It may be as to illustrate hese methods on an actual example .
Modal of the Barometer By inspection of observation polygon . . .
By using a parabola through three ordinates By using a normal curve through tilree ordinates the principle of the modal third , as above By actual determination of the frequency curve .
By adding up the frequencies for Southampton in I. , and then interpolating , it will be that the median crht of barometer there exactly But by Table III .
, the is 981LJ , of the yeen m and mode or third whence the mode .
It is clear that this method gives a close tion to true result , and is one which can be used any ordinary observer .
The ained wilL bc , median rrht of the baromeber is the height by that out of } ' tions , which has the heights of observations less and the of obscrvntions its owll height , , ib is middle hoight of the of Vation s nrranged in of 11 tItudc .
VOL. as a rule , remarkably closer to the true mode than any application of methods such as ( ii , ) or ( iii .
) .
If we invert the process and calculate from Table IV .
the median values at our three southern stations , we find : Median , St. Leonards .
Southampton .
Babbacombe . . . .
Thus we see that the median height of the barometer along the south coast of ngland approaches extremely closely to the , so commonly adopted by physicists as a measure of the " " standard atmosphere .
There is another conyenient method of looking at this standard atmosphere of If we turn to the modal heights in Table IV .
, we notice that the mean modal height for all our stations is , with the very small probable error of Hence the mean modal height for the British Isles differs by less than twice its probable error from the customary standard atmosphere .
On the other hand , the mean mean height differs by more than 12 times its probable error from standard atmosphere .
We may accordingly look upon the standard atmosphere either as corres o very closely to the mean modal height of the barometer for the itisl ) Isles , or as the median height of the barometer along the English southern coast .
Another interesting feature of the modal height is that it is less variable than the mean as we pass from station to station , and the probable error in the determination of mean of the modes is accordingly less than that of the mean of the means .
Whereas up and down the British Isles we find a coefficient of variation of per cent. for the mean barometric height , the corresponding quantity is only per cent. for the modes .
On this account , and because the mode\mdash ; as the most frequent barometric height\mdash ; has a more direct physical interpretation than the mean , it seems to us that a record of local modes would be of greater significance , than a record of local means .
Owing to the property we have already noticed , i. e. , that the distribution of barometric frequency is constant along contour lines , differing , at any rate , not very widely from the isobars in the narrower sense of word , it follows that mean and mode oscillate , not without deviations , in general accordance .
Thus , both mean and mode are least at Laudale and Dunrobin Castle , the most stations we have dealt with ; both means and es are greatest at the four southernmost stations , NIargate , St. Leonards , Southampton , and bacombe , although it is characteristic of some peculiarity of observation or climatological individuality , that while out of the four the means are greatest at St. Leonards and Southampton , the modes are greatest at Southampton and Margate .
In particular we look upon the modal value Margate ( 30 0493 ) as standing geographically between Geldeston ( 30 0047 ) and St. Leonards OF FREQUENCY OF THE BAROMETRIC HEIGHT AT DIVERS STATIONS .
443 as not beyond suspicion , and accordingly open to revision when a wider range of data is available .
The great value of the " " skewness\ldquo ; in the Margate distribution is also tisfactory .
8 .
On the Variability of the Barometric Height .
The only method hitherto used by meteorologists to express 'oriefly the variability of the atmospheric pressure is , so far as we are aware , the statement of the maximum and minimum heights reached during any given period .
The fallacy of this method has been illustrated by one of us elsewhere .
It gives no real impyession whatever of the manner in which the bulk of the variation is distributed , yet , for most climatological purposes , this is precise]y what we require .
by such a test as this ( namely , the range from maximum to minimum height observed ) Hillington has a mere variable climate than Scarborough , and Dton than Babbacombe , but , as a matter of fact , Hillington is less variable than rough , and Southampton is slightly less variable than Babbacombe .
Another striking illustration of the defects of this method of easuring the variability has been mentioned to us by Mr. R. H. SCOTT , namely , that in 23 years of barometric observations at Valencia , the maximum was only reached iu the year .
Whatever be the form of the frequency distribution , the problem of determining how the bulk of the variability is distributed about either mean or mode , is exactly similar in character to the problem of determining how the inertia of a plate is distributed about any axis in its plane .
One satisfactory and useful measure in both cases is the swing-radius , or radius of gyration .
This is the quantity which , for distribution about the mean , appears under the heading the standard deviation in our Table Iy .
Judged by this test the following is the order of variability in barometric pressure at our 20 stations:\mdash ; ( 1 ) .
St. Leonards [ 1 ] .
( 2 ) .
Margate [ 4 ] .
( 3 ) .
Southampton [ 2 ] .
( 4 ) .
Babbacombe [ S ] .
( 5 ) .
Geldeston [ 7 ] .
( 6 ) .
Hillington [ 8 ] .
( 7 ) .
Carmarthen [ 6 ] .
( 8 ) .
Stonyhurst [ 9 ] .
( 9 ) .
Churchstoke [ 5 ] .
( 10 ) .
Llandudno [ 10 ] .
( 11 ) .
Scarborough [ 13 ] .
( 12 ) .
Parsonstown [ 11 ] .
( 13 ) .
Armagh [ 12 ] .
( 14 ) .
Scaloby ( 15 ) .
Dunrobin Castle ] .
( 16 ) .
Gla , sgow [ 16 ] .
( 17 ) .
Dundee [ 18 ] .
( 18 ) .
Londonderry [ 14 ] .
( 19 ) .
Laudale [ 19 ] . .
Markree Castle [ 17 In an essay on : ' Variation in Man and Woman ; " " see ' The nces of , and other Studies in volution , ' vol. 1 , p. 275 .
In square brackets we have inserted the order of mean barometric pressures .
It will be seen at once that the 10 stations of least yariability are tlJe 1 stations of highest pressure .
Thus , there is a correlation between high pressure and small variabiiiby Some changes in the two ordel .
S may well be due to the doubt which attaches to the reduction to sea level , but taken as a whole the list illustrates the local character of the climate at the various stations , so as it depends upon the height and variability in height of the barometer .
This method of appreciating variability seems to us more satisfactory than a mere measurement of maximum to imum ranges , which , with our data , while leaving St. Leonards first steadiness of climate , would place Geldeston quite close to it , and make both that town and Scarborough superior to Southampton !
Instead of taking the variability about the mean , we might equally well have taken it about the mode , only difference being that we should now have to calculate instead of : see p. 433 , Equations and ( ii .
) .
The comparatively large values of , however , do not allow of any widely divergent differences in the results .
The more interesting problem of the valiabilities in excess and defect , which , owing to the skewness of barometric frequency curves , are not the , will be dealt with in the next section .
9 .
On of rometric F The comparative eness of the mean to the mode enables us to easily find a formula for the oility that the barometric height in any locality shall be in excess or defect of the modal heigbt .
Using the property of the modal third have to integrate to Hen ce the area as far as terms of the order . .
Thus the total area on the mean side of the mode has been shown for this skew curve ( PEARSON and FILO ' Roy .
Soc. Proc vol. 62 , p. 175 ) that the is negativoly col.reJated with the staudard deviation .
Tbus we have a theoretical indication that high pressure is related with small variability .
The actual correlation for the meall value of is 25 , approximately , this a random variation from the standard OF EQUENCY OF THE TRIC H HT AT RS STATIONS .
445 where is the skewness recorded in Table IV .
Since the skewness is in our case small , it is sufficiently close for barometry to take the chance of a reading of the barometer less than the mode , as : and greater than the mode , as The corresponding probabilities for the baromebric height less and greater than mean will be found to be : and respectively .
The followino is the of stations arranged according to their chances of the barometer exceeding its modal heights:\mdash ; This list would present in general the same features , as we have ready noted\mdash ; of a continuous change , distributed in contour lines running a little north-east to south-west of the parallels of latitude\mdash ; were it not for the anomalous positions of Llandudno , Stonyhurst , and Margate .
These stations appear to us , especially the last , to have anomalies in the values of their constants , which can hardly be entirely due to local peculiarities in climate .
Tt will be observed that the mode is here again more suitable than the mean as a method of high barometer .
It might , if the point were only superficially considered , be deemed a climatological advantage to have the frequency of the barometer above its mean value as great as possible .
But this is not eally so , for the simple eason that climates which have an extreme range of low barometer have a low mean , , other things being equal , places with low mean have most frequency above the mean .
On the other hand , places with high means , as a rule , give the greatest frequency above the mode .
Of course , this relationship is not invariable ; it follows from the mode being more steady than the mean .
Thus , there is a greater chance of the barometer standing above the mode in St. Leonards than Laudale , but a less chance of its standing above the mean .
The inequality of the frequencies above and below the mode is not the only point of interest connected with the skewness of the barometric frequency-curves .
We have already noticed that the standard deviation about the mean is a good measure of the local variability of the barometer , and may well be used to replace the maximum to minimum range But a further question owing to the skewness , namely , what is the variability above , and what is the variability below , the modal height ?
These two will Ilot be equal , and an appreciation of their value is of considerable importance .
Some idea of the range above the mode may be obtained by considering the columns marked and in Table IV .
, which give the theoretical maximum height , the observed maximum height and the total theoretical above the mode .
The disadvantage of using has already been referred to ; it Jnay make the lalJge above the mode depend on a single observation in the last year of tho whole series .
is calculated from the whole sweep of observations , and hence , although it ma never be reached in actuality , it is a far better measure of range .
The order of the st.ations is given in bhe following :\mdash ; * An arbitrary muItiple of the standard deviation , , may be conveniently takeu as the , if required .
Thus practically covers the whole range of -frequency at any ation .
OF FREQUENCY OF THE BAROMETRIC HEIGHT AT DIVERS STATIONS .
447 Range above the mode according to K. PhARSON AND AIISS ON THL RIBUTION An examination of these lists seems to show that , eyen allowing for some di lrbances in Stonyhurst , Llandudno , and Margate returns , there is no closely represented by range bove the mode thus measured .
There but few and small changes introduced into the order of , whether we consider or Hp only .
Bnt so widely Hp that we must consider one or other of them as of little value in the determination of the nature of the frequency above the , In general the range above the mode dealt with in this manner appears to be more closely correlated with local conditions than with geographical position .
It would undoubtedly be most satisfactory in order to appreciate the skewness of the range of frequency to calculate the values of the standard deviation from the mode for two portions of the frequency curve , which fall respectively above and below the mean .
The extremely slow convergence , however , of the series which express incomplete -functions , renders this calculation extremely tedious , and such methods as AMSLER'S Integrator do not , in .
experience , give very good results , the curves for which the second moment is to be found have , as is the case with nearly all frequency-curves , considerable " " tails Fortunately , .
DE ]ORREST , in a paper published in .
Analyst , ' ( vol. 10 , p. 69 ; Iowa , 1883 ) , has found for a series of values of the probable deviations in excess and defect of the mode , , the values of on either side of the mode for which corresponding , cut off the half areas .
Unfortunately , although he has interpolated fcr a considerable number of values of , his values of the probable are only calculated for the small series , 5 , 7 , 10 , 20 , 50 , and 200 , doubtless on account of the great amount of arithmetic involved .
Accordingly his table is only as strong as these seven values , which are as follows:\mdash ; DE FORREST'S Table of Probable Deviations .
probable deviation in excess of mode , i.e. , along negative , , , , defect , , , , positive , , completion of the Table would be a usefu ] , if laborious , piece of work .
We have calculated the values of and the twenty stations by interpolation An An , especially ucfed f me , to , and is fairly satisctory f oval\ldquo ; sections , it ves a alue o , bnt is accu enough give working valucs of for " " tailed\ldquo ; areas.\mdash ; K.P. ] OF FREQUENCY OF THE BAROMETRIC HEIGHT AT ] VERSE SATIONS .
449 from this Table .
The results will , of course , only be correct to a corresponding degree of accuracy .
The positive sign is given to TABLE VII.\mdash ; Skewness in Variability and Range .
Unit , one inch .
It will be found that the following empirical formulae give the values of and ) with an accuracy quite as great as that of their determination by interpolation from DE FORREST'S table:\mdash ; , .
These formulae must , of course , only be applied with caution beyond the British Isles , and still less to other problems in skew frequency , when the skewness does not fall within the range of barometric skewness considered in this paper .
Within this range , however , they give remarkably good results .
The above formulae , or the table , show at once that and are quantities which follow the system of generalised isobars , and thus and , are good measures of the range and its skewness .
It is , accol'dingly , these which ought to be calculated for the purpose of obtaining an ] preciation of the range above and below the mode at any station .
For example : St. Leonards half the frequency of the barometer falls into a range of , namely , from to , i.e. , the mode being , from to .
Further , the relative scattering of the frequency above and below the mode is given by the ratio of : or : .
At Markree Castle , on the other hand , it requires more than half-an-inch to cover half the frequency , i.e. , to , and the ratio of gives the relative ranges of half the frequencies above and below the mode .
The comparatively small amount of labour necessary to determine and VOL. CXC.\mdash ; A. 3 PROFESSOR K. PEARSON AND blISS A. LEE ON THE DTSTRIBUTION in fact , Sk may , by the rule of the modal third be determined from the median and the mean\mdash ; and , thence , and by the above formulae , make this a very convenient , as well as scientifically accurate , method of appreciating the range and the skewness of the variability at any station .
Our discussion has now led us to the conclusions:\mdash ; ( 1 .
) The mode , the standard deviation , and the skewness fully define barometric frequency .
These three constants depend , in the first place , on geographical position , and appear to be constant along certain lines\mdash ; the generalised isobars .
( 2 .
) A knowledge of toese three constants enables us , by means of very simple formulae , to describe the chief physical features of the barometric frequency at any station .
( 3 .
) By aid of the twenty stations dealt with in this paper , a fair appreciation can be obtained of the barometric frequency at any place whatever in the British Isles by means of interpolation .
For example : line from Hillington on the Wash to a point midway between St. Leonards and Southampton strikes the south coast between Littlehampton and Wortlnng , cuts the generalised isobars so that they make approximately equal angles with it , and passes very nearly at one-third and two-thirds distances through Cambridge and London .
Thus , if the St. Leonards constants were based on a longer period and so somewhat more satisfactory , might fairly accurately predict the constants of the Cambridge and London frequencies from those of Hillington , Southampton , and St. Leonards .
We find , as a matter of fact , by interpolation : Actual calculation for thirteen years at these stations corresponding to the years used for ampton and Hilliugton gives:\mdash ; Mean height .
viation .
attempt was made to replace the St. Leonards returns by the same thirteen years as have been dealt with for the other stations recorded at a suitable " " station Unfortunately , during these thirteen years the most suitable station , namely , that at Dover , was changed to Dungeness\mdash ; a station some considerable distance off and subject to probably somewhat different conditions .
A short missing period between the two sets of observations was most kindly supplied by Mr. R. H. SCOTT , by intelpolation from the Meterological Office Records .
The result of the calculatious showed a considerable increase of variation , probably due to the combination of two stations , and it was very doubtful whether the result was of greater weight than the six years returns from St. Leonards .
OF FREQUENCY OF THE BAROMETRIC HEIGHT AT DIVERS STATIONS .
451 These values are not widely divergent from the previous interpolated values , and would for many climatological purposes suffice to describe the barometric frequency ; indeed , graphically it is hardly possible to show the difference between the frequencies curves corresponding to these two sets of constants on the scale of our diagrams .
The chief difference is in the skewness of the London distribution , but as we have seen in the case of Llandudno , Stonyhurst and Margate , it is the skewness which is most influenced by local conditions .
The constants and the observed and theoretical frequencies for the three telegraph stations LondoIl , Cambridge and Dover-Dungeness , are given in the form of supplements to our tables .
They are of very considerable interest , but they have not been included in the general returns based on the selected distribution of venty stations of the second order , as they would weight too much the eastern side of the British Isles unless an additional series of western stations had also been included .
* \mdash ; ON THE CORRELATION 0F HEIeHTS 0F THE BAROMETER AT 10 .
So far as we are aware , no tabulations have hitherto been made of the barometric heights at pairs of stations , and yet the degree of correlation between stations in different situations is one of extreme interest and importance .
We have shown that the constants of barometric frequency vary continuously and from one end to another of the British Isles .
We should accordingly expect a close degree of correlation between the heights at different stations .
This degree will probably be found to vary wibh the distance at a different along and perpendicular to the generalised It may also be greater when a interval is allowed between the observations the two stations .
Our present object , however , being only to illustrate the general treatment of barometric correlation , we have dealt only with three pairs of stations and with contemporaneous observations .
The stations are the following:\mdash ; ( 1 .
) Babbacombe and Churchstoke for the eight years 1878 to 1885 .
( 2 .
) Southampton and Laudale for the eight years 1880 to 1887 .
( 3 .
) Hillington and Churchstoke for the eight years 1878 to 1885 .
The results are exhibited in the following three tables:\mdash ; * We have to cordially thank Mr. R. H. SCOTT for allowing us to copy the nuscript reccrds of the MeteoroIogical Office for these three tclegraph stations .
We believe that the discussion of the frequency of twenty contemporaneous years of the whole system of telegraph stations would give most interesting results , but the labour of copying and reducing only thirteen years for but three stations has convinced us that it could hardly be llndertaken by private individuals largely occupied with other work .
A most interesting estigation wouId be degree of corrclation between suitable NorthAmerican and British stations , when the interval betwcen the observations was varied from one up to six or seven days .
Similar for tish and Continental stations ve important results bearing directly on the prediction of OF QUENC Y OF TH BAROMETRIC HEIGHT AT DIVERS .
453 PROFESSOR K. PEARSON AND MISS A. ON THhl DISTRIBUTION OF FREQUENCY OF THE BAROMETRIC HEIGHT AT STATIONS .
455 Now the distribution of frequency at each station being skew , the correlation surfaces as based upon these tables will also be skew , and the curves giving the mean height at one station for given height at the other are no longer straight lines .
We do not propose on the present occasion to discuss at length the propertios of this type of skew correlation , but to refer the reader to a paper by Mr. G. U. YULE , published in the ' Boy .
Soc. Proceedings , ' vol. , pp. et seq. , 1897 .
In that paper it is shown that the coefficient of regression is still significant in the case of skew correlation , it gives the slope of the line of closest fit to the cJrve of sion , or the locus of the mean heights of one station for successive heights ab the other .
Since the locus is not very far loved from a straight line in any of the cases dealt with , it follows the line of closest fit will very approximately represent it .
Calculating the coefficients of correlation and the regressions for the three pairs of stations by the usual formulae ( see " " Mathematical Contributions to the Theory of Evolution , III 'Phil .
Trans vol. , pp. 265-6 , 275-7 ) , we have the folLowing results:\mdash ; TABLE XI.\mdash ; Barometric The application of this table to predict the height of the at one station from a knowledge of the contemporaneous height at a second will be clear to readers familiar with the mathematical theory of correlation .
For example , if the height of the barometer at Babbacombe be above the mean Babbacombe height , then the height to be predicted at Churchstoke is above the Churchstoke mean , with a probable error of .
We may give a numerical illustration .
The barometer at Churchstoke stands at ; what are its probable heights at Babbacombe and Hillington ?
The mean height at Churchstoke ( see Table IV .
) ; hence ths observed height at Churchstoke is above the mean .
The most probable heights at Babbacombe and ] lington will accordingly be and or and above the respective means of those places .
Extracting these means from Table , we find : and for the probable heights at Babbacombe and Hillington .
The probable deviations in the two cases are and .
Thus , the prediction would probably be correct to within of an inch .
We should not propose , however , to base the prediction of the barotnetric height at one station on its correlation with the height at a single other station .
We should think it desirable to apply the principles of multiple correlation , and endeavour by a suitable selection of stations to decrease the probable deviation of the array at the given station which corresponds to observed heights at the selected stations .
We shall cite here the general formula for the prediction of the height of the barometer from a knowledge of its heights at correlated stations .
Let be the probable height of the barometer at the station above its mean value for that station .
Let be the coefficient of correlation between the and stations , where in finding the correlated heights may be taken at different times , or if it seems desirable on different days .
Let be the observed height at the station and the standard deviation , both measured from the mean of that station .
11 .
Case iction from one Correlated only .
, with a probable deviation of It is clear that unless be very nearly unity , i.e. , the stations very close , the predicted height will be subject to a large probable deviation .
For ver .
close stations , such as London , Cambridge , Dover-Dungeness , where a rough investigation leads me to the conclusion that bhe correlation is as high as , we have the probable deviation about , or about .
In such cases the above formula will give fairly closely the height to be predicted at the first station .
It may be used for purposes of polation .
The approximate linearity of the " " regression ' ' leads us to an interesting property of barometric correlation .
There is a certain height of the barometer which , if it occurs at one station , will itself be the most probable height at the correlated station .
This may be termed the height .
This height may be easily found from the equation : whence , if , and the balance height .
Above and below the balance height the relative heights 01 the two stations are reversed .
If above the balance height the first station has a probable height of its rometer invariably higher than the observed height at the second station , then below the balance height the probable height at the first station will be invariably lower than the observed height at the second station , and vice versed .
FREQUENC OF TRICt IGHT AT The are the balance heights of the three pairs of stations dealt with in Table XI .
:\mdash ; bbacombe Above this value rebstoke , below Babbacombe , stoke usually the higher barometer .
Southampton Above this value mpton , below Laudale , Laudale has most probably the rometer .
Hillington Churchstoke Above this value Churchstoke , below has most probably the higher In the case of close and highly correlated stations , the balance no , the probable deviation is small , may be ] found from a mere in spection of the barometer records for the two stations .
Thus , we it is about for London and Cambridge .
Perhaps a better approximation be obtained estimating , Diagram illustrating application of Theory of Corl'elation to pl.edict ights .
\mdash ; Stations approximately lthogonal to isobars .
Stations .
isobars .
from their known distance , the coefficient of correlation for the two , stations means of the diagram above , and then calculating , since given Table IV .
, or can be found by interpolation from that table .
The diagram indicates graphically , wljat is fairly clear Table XI .
, that correlation is not a lued function of the distance between the two stnf , ions .
It VOL. CXC .
) suggests\mdash ; we wish especially to note it does not correlat , ion differs in its variation with the distance ording as the two stations have their distance practically or orthogonal to the generalised isobars .
Unfortunately , without crossing the Irish Channel , it was impossible to find two stations a generalised isobar so apart as Southampton and Laudale .
There can be little doubt , we think , ) such tions would , however , give a less correlation than the north and south stations .
the ( owing to the very considerable labour of calculating the ] ation for a pair of stations even for years ) is upon a very inadequate number of measurements , yet we should expect , with continuity in the correlation coefficient , which can hardly fail to be the case , that it would give fairly approximate values .
For example , we should anticipate that the correlation between Southampton and Stonyhurst would be about to .
Since Stonyhurst is , roughly , about equally distant from Southampton and Laudale , the correlatiolt between Stonyhurst and Laudale may with somewhat smaller probability be also put at to the general relationship between correlation , distance , and direction of distance will only be determined when a very great number of of stations have been worked out , and these stations ought to be ributed , not over a small area like he British Isles , but over a large continental area .
12 .
as ( ii.).\mdash ; Prediction from ) Here : with a probable Hence if we to predict the of the barometer as closely as ossible a one station from either an earlier or contemporary observation of the at two other , we to choose out intervals of time and the distribution of the three so that may be as small as possible .
It would thus seem , considering the great variety of and places available , within our power to predict almost exactly the height at any selected station from a knowledge of the two other selected tions at selected intervals of lime .
The importance of testing this principle seems to us very great ; it might lead to quite novel methods of barometric change .
Unfortunately the needful knowledge of the correlation coefficients of widely separated pairs of stations for divers intervals of observation is still wholly wanting .
The following example is merely illustrative .
Suppose the second and stations , so selected that they have equal correlation with the first , and a tion with each other .
Then would be absolute]y zero if OF FREQUENCY OF THE BAROM TRIC HEIGHT AT DIVERb1 ATIONS .
45 or if that is , if .
For example , if the relafio between the second and third st tions be 0 .
In other words , if three stations could be found with coefficients and ; then the romefric h , at the first would be exactly a linear function contemporaneous heights at the other two stations .
There seems no reason why stations should not be found with these colrelations , or similarly related correlations .
The correlation between Sout'nampton Lauclale is ; the correlation between Stonyhurst and Laudale and between and Southampton muQt be about to .
We incl.ease the distance ftnd ke it less perpendicuiar to the generalised isobars by moving across towards the coast .
Probably somewhere near the required correlation of } old be and at such a station we should expect the barometric height to be very nearly a linear function of the ights at Southampton and Laudale Supposing the relation ?
) to hold , then it is easy to deduce fi om the expression for above , that if and , be the absolute of the barometer at the three stations , This passage of correlation into causal relationship is of such extreme nce , that it is worth while to see what approach we can find to it , even within the somewhat narrow limits of the British Isles .
We have no details available for barometer * Let the reader imagiJle the heights at Southampton and Laudale replaced by the heights at ( or if necessary more ) stations on the continent , say , along the Easb coast , and tho at the Whitby station replaced by the heig.ht , after an interval of time , a British station , say VaIencin , and then he will the sort of possibility\mdash ; not the feasibility\mdash ; of prediction , which the authors wish lay stress upon .
[ This expression is used advisedly to draw attention to the importance of the limit when a lation passes into a causal relationship .
If the unit A be always preceded , accompanied or followed by , and without does not take place , then we are accustomed to speak of a ationship between A and B. On the other hand , if hcll A occurs amounts , of are found with frequencies . .
per cent. of the reuces of , we speak of a corrclatio]l A and B. Now , if we approach tho case , of . . . .
and per cent. , it is clear that more and ) will occur whene ver A occnrs , i.e. , fixed amount of on rence of A. This is transition of correlation into relationship .
It is the construction of a -dimensioned correIation surface , of which a ) dimensioned section approaches indetinitely close a frequency curve of zero standard tion .
] K. ANY ) A. ON THE DISTRIBUTION at Wbitby , if we had them and worked out the Whitby-Southainpton and udal eions , it is unlikely that they old 1ooth be exactly equal , equad to fact the colrelation a number of Yorkshire stations , with both Laudale , would have first to be worked out , and the true station with Laudale and Southampton would require to ) found by interpolatiol ] .
But should expect stations not so far remoyed from the right position to have rometric given in terms of those of Laudade and Southampton by an imalely linear relation .
Hence , to indicate to the reader that our conclusion \mdash ; namely , that barometric correlation may pass into a causal relationship\mdash ; is not paradoxical may pear at firsc sight , we have endeavoured to test how far the Stonyhurst heigl is a linear function of the heights at Laudale and Soutbampton .
In order to do this have assumed nor worked out the and the Southampton-Stonyhurst correlations .
The former was too risky , ae the latter { oo laborious for he end to be desired .
We have assumed a linear ionship 1veell the heights } the three stations , i. e. , yhere are nerical constants , and is a number of inches .
To determine we chose twelve observations , taking the 15th day of each month one year , by the method of least squares .
Unfortunately the resulting equations for , throw back their determination on decimal figures , which are the limit of vbat is usually tabulated in barometric observations .
The resulting equations were solution of these equations is wlIere the values and tainly not correct to the second place of decimals .
The resulting formula ives attempt to pproximate to the coefficients of correlation , gave the Southampton a somewhat higher value than the Laudale factor , and this is probably the case .
But with the data available we shall hardly do better than the above formula .
tesb its degree of accuracy 50 values were taken out of the returns for Southampton , Stonyhurst at ightly .
intervals , and the observed and calculated values at Stonyhurst given in .
Ths differences distributed fairly evenly , ) vely and their mean value is about .
We consider noted that the ) are not , our opinio1l , very ctory : see p. 446 .
OF FREQUENCY AT THE BAROMIBTRIC HEIGHT AT DIVERS .
461 them sufficiently satisfactory to justify the view that a station could be found in Yorkshire for which correlation would pass into a causal relationship .
TABLE XII .
Approach from Correlation to Causal Relationship .
Observation .
Calculation .
29.9 29 .
29.94 0 .
30 .
30.48 29 .
29 .
29.94 0 .
30 .
30.48 29 .
29 .
29.94 0 .
30 .
30.48 29 .
29 .
29.94 0 .
30 .
30.48 29 .
29 .
29.94 0 .
30 .
30.48 29 .
29 .
29.94 0 .
30 .
30.48 29 .
29 .
29.94 0 .
30 .
30.48 29 .
29 .
29.94 0 .
30 .
30.48 29 .
29 .
29.94 0 .
30 .
30.48 29 .
29 .
29.94 0 .
30 .
30.48 SSOR K. P ARSON AND MISS A. LEht ON THE DISTRIBUTION As a Llle there will exist for most triplets of stations a certain height , which may be termed the height , by which we are to understand that if the barometer stand at the height at two of the stations , its most probable value at the third correlated station is ) ce height also .
The balance height at once found by putting in the form u for regression and .
If we put , in the formula on p. are led to for the balance height ; this is probably very far from being the true balance height , and for the simple reason that the factors , of and on p. 460 , are only approximate .
If we found a Yorkshire atio which had the correlation with and Southampton , its balance would be given the formula , on p. 459 ; and supposing it to lie on , or nearly on , the same generalised isobar as Stonyhurst , i.e. , to have nearly same mean and stan then its balance height with Laudale and mpton would be .
An error of between and in one of the factors , of the Stonyhurst-Laudale-Southampton linear relationship , on p. 460 , would thus have reduced the balance heigbt of those stations from to about The general ) ressio for the balance height of a station 1 , with regard to stations 2 and , is given by ance ) Hence if the three stations lie on the same generalised isobar , since and have tely the same value , the balance height will be the mean height , and the same for every station with regard to the other pair .
One further general } ) osition may be noted before we leave the special case of prediction frcm two correlated stations .
Suppose the barometer constant at one station , then the coefficient of correlation between the heights at the other two is given .
, a selected height at the first station .
, for a selected height at the second station .
, for a selected height at the third statio1l .
* These values have been termed by Mr. G. U. neit coefficients of correlation , to distinguish them from , and , which he terms gross coefficients .
The diflerence would , perhaps , be best expressed mathematically by tho use of such as partial col'relation coefficient and total correlation coefficient , the former being the value of the coefficient when one val'iable is not allowed to vary , and the Iatter when it is .
They at once able from the general expl.ession on p. 287 of bho Memoir , ' Phil. rans .
, vol. 187 , by putting , say const .
and emembering that the coefficient of tion f iu is OF TH17 , TRTC H HT AT DlVHiRS IONS .
463 The values of these essions a peculiarly interesting , for they lead us to the general theorem that whenever the total correlation between two stations is 1 than the product of the total correlations at the other two of stations , then the partial correlation yeen the latter two stations will be ) egative , i.e. , for a given value of the barometer at the first station , a barometer at the second station the mark a falling barometer at the third .
station , and a barometer at the second station a rising barometer at tbird .
Owing to the generally high values of } ) arometric correlation it is comparatively easy for , for example , to be greater than .
Thus in the case where two of the stations have equal correlation with a third we have , and will always be negative if be .
For example , we }lave Laudale and Southampton with a correlation , while Stonyhurst and both these stations must have a correlation of about to .
It follows , , that the partial correlation of Laudale and Southampton with regard to yhurst is negative , 0 for a constant value of the barometer at Stonyhurst rising at woulci in general denote falling at Southampton , and vice versed .
This is best illustrated by noting that being the mean height at Southampton for heights and at Stonyhurst and Laudale and being measured from the respective of the stations ) , or , in the special case approximately , Hence , if be or , an increase of , for a constant means decrease of In the particular instance of lation P into to by us on p. 459 , i. e. , when we have Thus the partial correlation between the three stations is " " perfect between the second and third it is .
The ratio of the fall at the second to the rise at third , for stationary barometer at the first , is This principle , which at first sight appears rather paradoxical , , that at three stations , , a at A will , on the average , be accompanied by a rise at , but a rise at A for a constant barometric height at , on be accompat ) by a fall at , appears capable , when extended , developed , and illustrated by actual numerical examples , of considerable on the nature of barometric variation .
* To show the importance and truth of the principle , a table has been formed for the deviations from the at Southampton and Laudale , when the barometer at Stonyhurst does not differ by more than inch from its mean value , In course of two years 31 values were found , and ?
oith only one exception , possibly , the barometer at always above or below the according it ?
) below or tloe at Southampton : see Table XIII .
TABLE XIII.\mdash ; Deviations of Barometer from the Means at Southampton and Laudale ( 29 86 ) when the Barometer does not differ more than of an incl ] at yhurst from its Mean .
corresponding apparent paradox in the theory of heredity is referred to ( Phil. Trans , vol. 187 , p. 289 .
is an exception to the general rule , that for the mean at Stonyhurst a high barometer at iNCY OTRlC H AT ( iii , ) .
rhree Stattons .
The general are given on p. 294 of the memoir eviously c ) not reproduce them here , we have no data available at which to illustrate them .
We will , vever , consider special application similar to that dealt with under Case ( ii .
) .
Suppose it possible to select three stations so three have equal correlation ?
with each othel each a elatioJ l with the fourth .
The latter being with the in the , tl following are the proper elations which may be obtained after ) reductions the general case above refel.red to .
, with a probable deviation \mdash ; , with a probable deviation ) .
Here the closest prediction will be obtained , if we select if possible , such that , which is the point at which correlation } ) asses into causation .
In bhis case we find , , with vanishing probable deviations .
In the first case , if , which can easily be true for correlated stations within 300 miles of each other ; in ) second case always , a rise of the barometer : two of " " outer\ldquo ; stati at the " " inner\ldquo ; or station station and Now , it is not contended that jour st tions can fur and , slill less it is ) to ) ) it is suggested the values 0 the barometeric correlation } we have found in the Islc approxnnations to these relations can be found selected stations , and that such stations are what we require for prediction interpolation .
Further , such principles as we have noted with regard to the relation Laudale is a low barometer at Southampton , and vice versed .
Or , the rule ( by differcntiation ) may bc stated steady barometer at Stouyhurst a riso Laudnlc iudicntcs a ab Southamp an , versed .
VOL. PROFESSOR K. PEARSON AND MISS A. LEE ON THE DISTRIBUTION of rise and fall at correlated stations are independent of the special relations between the coefficients , which we have selected to illustrate them , they are really a deduction from the sign of the regression coefficient , or of the coefficient of partial correlation , which has the same sign .
14 .
Case ( iv.)\mdash ; Prediction from any Nurnber of Correlated Stations .
The general formulae are given , p. 302 of the memoir above , cited , namely : with a probable deviation of where is the determinant below and is the minor formed by leaving out the column and row .
In order to obtain close prediction , it might be supposed that all that is necessary is to take a sufficiency of correlated stations .
This is very far from ) eing the case .
The true test of closeness of prediction is the smallness of , and this .
can often be obtained by a few well-selected stations better than a great number .
In order to roughly illustrate this , suppose the correlation coefficients of the stations to be all of the same order of magnitude , i.e. , about , then the order of for 1 , 2 , 3 , 4 .
stations is given by , , or the prediction is only increased iu certitude in the ratio of to 1 by taking an indefinitely great number of ions , or , since cannot be greater than unity , it can only be increased in the ratio of to 1 .
Our object should accordingly be to make as small as by a fit selection of comparatively few stations * It is easy to illustrate its vanishing by taking one station equally correlated with others ( ) , which are equally correlated among themselves .
In this case we have OF FREQUENCY OF THE BAROMETRIC HEIGHT AT DIVERS STATIONS .
467 at different intervals of time .
Nor should it be forgotten that theoretically we have much power of the magnitude of our correlation coefficients .
They decrease from about 1 at the distance zero to zero as the distance increases .
We have thus a zone round every station of zero-correlated stations ; beyond this it is highly probable that the correlation becomes negative as we reach places which have cyclones corre- sponding to anticyclones at the given station .
This zone is of course not reached in such a area as the British Isles .
Passing through an area of negalive correlation , we should in all probability ultimately reach an extended area of zero correlation .
In the next place the time interval is under our control , and the coefficient of correlation can be reduced by increasing this .
Lastly , it varies although probably to a much less extent , by varying the direction in which the distance between two stations is taken .
We can imagine no more useful piece of work than the determination of the correlation for a of 10 or 20 years of a series of stations taken so far as possible round , say , a parallel of latitude .
We believe that the future of barometric prediction , i.e. , the accurate foretelling of the arrival of depressions , \amp ; c. , lies in an extended knowledge of the correlation between a system of barometric stations widely diffused over the surface of the earth ; special attention being paid to the changes of the correlation with intervals of time .
The object of the present writers has not been to make an elaborate investigation of the numerical values of barometric variation or correlation , but rather to indicate to those more directly occupied with meteorological investigations how the mathematical theory of statistics may be applied to barometry with novel and , they believe , valuable results .
APPENDIX .
On Barorneter by G. U. In all ordinary forms of registering barometer the resulting diagram shows the height of the barometer at each instant of time .
To construct a frequency curve from such a diagram , the heights must be read ofl for all the times desired , corrected if necessary , grouped , and replotted in the manner described in the preceding paper .
This procedure is somewhat tedious , and it may be obviated by so constructing the barometer that it shall give the frequency record automatically .
, with a probable deviation 1 \mdash ; .
Hence we should have absolute prediction if .
Similar propositions follow for partial correlation and for rise corresponding to fall .
It seems doubtful , however , whether such a system of correlation could possibly be arranged for more than four stations .
, K. AND MISS A. ON DISTRIBUTION barometer devised for this purpose is illustrated in figs. A and B. Fig. A is a atic plan of the instrument ; fig. is from a photograph of a rou model . .
made by Cambridge Instrument Company .
The barometer must be an aneroid , or some in-trument working with a needle over a horizontal dial .
This needle is removed , and replaced by a light -shaped gutter of paper or metal-foil , sloping downwards from the centre to the periphery of the dial .
This gutter needle is indiby in fig. A. Its outer extremity projects slightly beyond the dial over the tops of a series of smull vertical tubes , , which are equally spaced and separated by small wedge-shaped divisions , running along the lines .
If we suppose a ball to be dropped into the gutter at , it will roll down to , drop over , OF FREQUENC OF BAROMETRIC AT STATIONS .
469 and fall either into one of the tubes or on to one of the wedges , which will guide it into the corresponding tube .
The shot then be a record that the barometer needle stood once over the section corresponding to the tube .
If balls be dropped in at regular intervals , a frequency record is obtained by simply counting up the number of alls in each tube .
The record , of course , suffers from the disadvantage that the usual corrections cannot be made .
In practice , of course , the wedge edges must be arranged so as to correspond exactly to the divisions between , say , the tenth inches .
Any one tube will then record the number of times that the needle stood over a particular tenth when the record was taken .
In the model itself a fixed gutter , , is brought over the centre of the needle gutter , above it , axis ; it stands clear of it , and runs baclc to reserve of balls , R. A trigger arrangement , , is fired by a clock-driven cam often as desired .
The trigger gear be seen in the second figure .
The 1 cam , , turns round once in the twenty-four hours , one of the saw teeth being pushed on each hour by the hand of the clock .
The channel , containing the reserve of balls in single file , is normally closed by the forked , which is oted near , and bears at its lower end against the cam , being held forward by a spring .
Only when the lever is fully cocked , i.e. , pushed over to its furthest reach by the cam , does the fork of the lever stand in line with the reserve channel .
One ball then drops into the fork , which isjust of the thickness of a ball .
When the lever is freed again , the fork comes opposite the fixed gutter , and the ball drops first into this , then ffills into the needle gutter , and ultimately into a vertical tube .
may be desirable to take two or more frequency records for different hours of the day , and keep these records separate , so as to admit of the study of systematic differences .
Mr. DARWIN , to whom the details of the vorking model due , devised a very simple ement for doing this , which be seen in the general view of fig. B. The balls do nob drop from the needle-gutter into the collecting tubes , but run into radial gutters fixed to the of the barometer .
At their outer terminals these fixed adial gutters stand so far ] they only deliver balls into alternate tubes , , to the set marked on figure with numbers not with black spots .
Before the alternate records are made , the whole barometer , with the radial gutters attached , is given a slight turn by the same clock that serves to release the trigger .
The radial gutters now stand opposite the tubes marked with black spots , and the record is made in them .
The working model was kept running for rather more a month , and worked very satisfactorily , considering its somewhat rough construction , There were one two " " misfires due to the balls hanging up , but such accidents could be ensilv remedied by a slight alteration .
* It musb be remembered , however , grouping in tellths of an inch giveq smooth and detailed returns for the purpose of calculating frequency curves .
Bicycle bearing balls were actually used , as shot were fouud too egular in shnpe ftnd size .
Pearson Lee .
Phil. Trans. , , vol. 190 , 1897 , Plate 10 .
and Le Phil. , 1897 , Plate 11 .
and Phil. Trans. , , 1897 , Plate 12 .
and Lee .
, 1897 , and , vol. 190 , 189 Plate 14 .
ctnii Lec . .
Trans. , ) , 1897 , Lee .
Phil. Trans. , , vol. 190 , 1897 , Plate 1 and Lee .
Phil. , vol. 190 , 1897 , Plate

