Review and Evaluation of Dendrochronological Methods Synthesis and Integration Report Number 10 St United Statu? Wrl A Environmental Protection Agency ------- Review and Evaluation of Dendrochronological Methods Synthesis and Integration Report Number 10 October, 1987 Prepared By: William G. Warren Research Associate Department of Forest Management Oregon State University Corvallis, OR 97331 Prepared For: Synthesis and Integration Project Forest Response Program US EPA Environmental Research Laboratory - Corvallis 200 SW 35th St. Corvallis, OR 97333 Notice This document is an internal report. It has neither been peer reviewed nor approved by the U.S. Environmental Protection Agency. It is being circulated for comment on technical merit and policy implications. Do not release. Do not quote or cite. ------- Review and Evaluation of Dendrochronological Methods. General Concepts There is no question that the history of a tree's response to its environment is embodied in its sequence of tree rings. There are, however, a multitude of environmental factors that are acting simultaneously, both natural and anthropogenic. Natural factors include climate, pests and pathogens, and, in some instances, fire; examples of anthropogenic factors are thinning and fertilization and, possibly, atmospheric pollution. Cook (1985b) formally describes the situation in terms of the so- called linear aggregate model, an earlier version of which was given by Graybill (1982)? specifically Rt = Gt + Ct + 5D1, + 6D2t + 6Pt + E( where Rt is the observed ring-width series measured along a single radius, the subscript t referring to time, i.e. year t. Gt is the growth trend associated with increasing age and size of the tree, Ct is the climatically-related growth variation common to a stand of trees, Dl, is the variation due to endogenous disturbances, that is those that affect a small subset of trees in a stand, D2t is the variation due to natural exogenous disturbances, that is those that have a standwide impact on radial growth, Pt is the variation due to anthropogenic pollutants that again have a standwide impact, and Et is a random component, unique to each tree or radius. The 6 associated with each Dlt, D2t and Pt is a binary indicator of the presence (6 = 1) or absence (5 = 0) of that component for a particular year or group of years. Conceptually there is little, if any, difference between this and Sundberg's (1974) seemingly neglected representation Ar(t) = Mt)r,(t) [1 - 0(t) ]C(t) in which Ar(t) corresponds to Rt, K(t) is an individual growth level function (cf. Gt) , ri(t) is a climatic factor which "for fixed t can probably be regarded as common to all trees in a region, or at most as having a slight and smooth spatial trend" (cf. ct) . The [1 - 0(t)] introduces the effect of external agents with 0(t) si (and may be negative) and 0(t) ¦ 0 for t < t0 (cf. Pt) . Finally g(t) is a random process and "with satisfactory accuracy we may probably assume that log{?(t)) is a stationary Gaussian process centered to have expectation zero, with independence between realizations for not very closely 1 ------- adjacent trees" (cf. Et) . The linear aggregate model is, perhaps, the more_sophis- ticated in that it separates out the effects of endogenous and exogenous disturbances which Sundberg more or less embodies in 6(t). Sundberg's representation, however, preceded the linear aggregate model by approximately a decade. The other noticeable difference is that Sundberg deliberately chose a multiplicative model, in contrast to an additive model, although there is evidence to the effect that the linear additivity of the components in the latter was intended to simplify the exposition and not to imply any structural relationship. Any formal development should not, therefore, involve linear additivity as a necessary component. The challenging problem is, thus, to isolate any pollution signal from the various other synchronous components. It seems reasonable to address this problem through an extension of the methodology developed for dendrochronology and its outgrowths, namely dendroclimatology and dendroecology. Dendrochronology has its roots in antiquity. The relationship between time and the growth of trees dates back more than two millennia to Theophrastus, the student and successor to Aristotle. That rings of trees are annual in nature was well understood by a number of botanists, foresters and astronomers in Europe and America in the 18th and 19th centuries; indeed, some understood the principles of crossdating (Studhalter, 1956). Fritts (1976), quoting Dean (1978) observes that the "French naturalists Duhamel and Buffon in 1737 discovered that a conspicuous frost-damaged ring occurred 29 rings in from the bark on each of several newly felled trees. Other investigators confirmed their observation and this feature was subsequently used as a marker for the 1709 ring. Crossdating based on relative ring widths was independently recognized in 1827 by A.C. Twining in Connecticut, in 1838 by the mathematician Charles Babbage in England, in 1859 by Jacob Keuchler in Texas, and in 1904 by A.E. Douglass in Arizona". While it was Twining who first grasped the full significance and potential of crossdating, it was the work of Douglass that led to dendrochronology being recognized as a legitimate and continuing field of scientific investigation. Dendrochonology in its purest form, as implied by the word itself, deals with the use of tree rings to date particular events, such as otherwise unrecorded geomorphological phenomena, for example earthquakes, landslides and floods. Examples of this type are provided by Druce (1966), Helley and LaMarche (1973), and LaMarche and Wallace (1972). The incidence of forest fires has likewise been dated, e.g. Dieterich (1980). 2 ------- The greatest interest, perhaps, has been in climatic phenomena; in relation to the above it may be noted that it was the excessively cold winter of 1708-09, one that was beyond living memory, that destroyed Charles XII*s army in its invasion of Russia. One particular area of interest comes from the suggestion that the blocking of radiant energy from the sun by dust in the atmosphere, resulting from major volcanic erruptions could cause a reduction in temperature which would be reflected in tree ring widths of the period. While the so-called "year without a summer" ,1816, (Francis 1976) followed the erruption of Tambora in 1815 and appears to be reflected in the ring widths of white spruce growing near Great Whale River (Cri Lake), Quebec, Canada (Parker et al. 1981), there appears to be no universal association between major volcanic erruptions and tree growth. Thus, while LaMarche and Hirschboek (1984) report several instances of frost-ring events in conjunction with volcanic erruptions, they also note various instances of frost rings in years with no known volcanic activity, and instances of major erruptions for which no frost rings have been associated. In reporting his own study, Parker (1985) concluded that "No immediately-apparent relationship was detected on a continent- wide scale. ... A possible response is suggested for trees from sites near the tree line in northern Quebec and northern Manitoba". The phrase "near the tree line in northern Quebec and northern Manitoba" should be borne in mind in what follows concerning the relationship between stress and sensitivity. Dendoclimatology goes further than the dating of extreme climatic events, and much of the work in this area has been devoted to climatic reconstruction. For example, Schweingruber et al. (1978) have demonstrated that excellent reconstructions of past climate could be produced using series of ring-width and wood-density measurements from subalpine firs. Also Hughes et al. (1984) found that maximum latewood density and ring width measurements from the Scottish Highlands could be used to reconstruct July-August temperature for Edinburgh, the period studied being from 1721 to 1975. (See below for futher discussion on the role of tree-ring wood density in dendrochronology, dendroclimatology etc.) However not all such studies have been so successful. A thorough description of dendrochronology and dendro- climatology has been given by Fritts (1976) who himself has been a major contributor to the methodology and application. Much of the work of the past decade has been founded on this definitive text as well as the various papers of its author. As will be demonstrated below, this "standard" methodology has its limitations which have, perhaps, come more into focus with the current necessity to assess the effect of atmospheric pollution on forest growth. Accordingly, recent years have seen the 3 ------- introduction of some alternative methodological approaches, for example Warren (1980), Monserud (1986), Visser and Molenaar (1987), Van Deusen (1987), and Cook (1987b). There are several well defined and generally recognized steps in a dendrochronologically-based study directed to establishing an effect of atmospheric pollution, namely crossdating, standardization (the removal of the natural growth trend), identification and removal of the effects of disturbances unrelated to pollution, and estimation of the effects of climatic variation. Each of these steps will now be critically examined. Crossdating. It should go without saying that any dendrochronological study is dependent on the accurate dating of each ring. This may not seem like a problem if the date of felling, or of the collection of an increment core, of a tree is known; one should simply be able to count, ring-by-ring, back from the bark to identify the year in which each ring was laid down. This would be the case if it were not for the possible occurrence of missing and/or false rings. Tree rings are identified by the visible transition from the higher density wood laid down towards the end of the one growing season (the latewood) to the lower density wood laid down at the beginning of the subsequent growing season (the earlywood). There is, however, no sharp transition from earlywood to latewood. False rings tend to arise when, because of some late-season climatic phenomenon, there is a transition from latewood back to earlywood within the growing season; thus two (or even more) rings may appear within the one year. On the other hand, sometimes during a year of extreme climate, or as a result of some disturbance, the tree may not form a ring on all portions of the stem; the ring is then said to be partial, locally absent or missing along certain stem segments or radii. Again climatic conditions may be such that wood laid down in a growing season never attains sufficient density to be identified as latewood; in this case the ring is not so much missing as inseparable from the following ring. Fritts (1976) observes that "A surprising number of absent sets or false rings occur, even for sites that are relatively temperate". Further, Fritts writes "... ring widths can be crossdated only if one or more environmental factor becomes critically limiting, persists sufficiently long, and acts over a wide enough geographic area to cause ring widths or other features to vary the same way in many trees. The principle implies that the narrower rings provide more precise information on limiting climatic conditions than do wider rings. During years when rings 4 ------- are generally wide, factors may become limiting to different degrees in each tree, depending on its locale, ecological position in the site, and a great variety of nonclim^iic factors. As a result considerable variation in growth patterns may occur from tree to tree. If the growth of a tree is never limited by some climatic or environmental condition, there will be no information on climate in the widths of the rings and they will not crossdate". In a later section he states "If climate is highly limiting to growth, all replicated samples within and among trees will show approximately the same ringwidth variation and the rings will be easy to crossdate. ... If climate is not highly limiting, variations in factors of the site may cause marked differences in ring sizes between trees, and differences in growth on opposite sides of the same trees may result from variations in the structure of the forest stand, lean of the stem, and competition from neighboring trees. A greater amount of replication may be necessary to achieve a reliable chronology from such a site". In a later chapter it is observed that "No specimen is selected for climatic analysis if there is any question as to the dating, even if the problem involves only one ring". These statements imply that crossdating is best achieved with trees that are growing under somewhat, but sublethally, stressed conditions and that reliable crossdating would be difficult, if not impossible, for unstressed trees. This is not necessarily the case. Sequences that exhibit a wide variation in ring width from year to year are commonly referred to as "sensitive"; such are observed in trees from, for example, the dry interior of British Columbia or near the northern tree line in the Northest Territories. On the other hand, sequences that exhibit little variation in ring width are referred to as "complacent"; trees from the moist British Columbia coast, for example, generally fall into this category. From the traditional dendrochronological point of view sensitive series are the most useful since, if there are no missing or false rings, these are the most easily crossdated. Unfortunately, it is with senstive series that missing and false rings are the more likely to arise. Thus, complacent series, from trees for which the date of felling or sample extraction is known, present fewer problems in crossdating. Complacent series carry a weaker climatic signal and are thus of less value for climatic reconstruction or the dating of historic, or prehistoric, phenomena; the last-mentioned may well be impossible. Conversely it may be that, because of the weaker climatic signal, such samples may provide a clearer indication, or measure, of any effects of pollution. Most existing chronologies have, however, been built from samples selected on the basis of the traditional objectives of dendrochronology and dendroclimatology and, indeed, not surprisingly, these principles have maintained themselves in the 5 ------- collection of tree-ring data when the objective has been the isolation of possible pollution effects. In short, the analysis of tree-ring data collected under conventional dendro- chronological principles is unlikely to be efficient for the detection of a pollution signal. Further, if sampling is restricted to sensitive series, how representative of the population would such trees be? Is it possible that pollutant would have a greater effect on otherwise stressed trees than on unstressed trees, or indeed have little or no effect on the latter? Any restriction imposed on sample collection may well lead to biased estimates of the effects of pollution. With respect to the actual crossdating operation, Fritts (1976) observes that "Crossdating, as practiced in North America, is a time-consuming process in that it requires both tedious examination of the specimens by eye, an operation that can be facilitated by certain graphical techniques, and computer analysis". He also notes that "The crossdating obtained in most North American studies by professional dendrochronologists involves visual examination of every specimen. The Laboratory of Tree-Ring Research requires that a routine dating check be performed by a second person, and if serious discrepancies are found, the dates for the entire set must be verified by yet another worker". He concludes that "It is highly likely that some of the tedium involved in the visual crossdating operation will be increasingly facilitated by computer analysis. However, at present there is no fool-proof shortcut to the procedure of careful visual comparison". Many recent studies now employ the computer program COFECHA (Holmes, 1983; Holmes et al., 1986) which calculates the correlation coefficients between a single series and a master tree series. As with much modern-day analysis, there is a temptation to eliminate the tedium by total reliance on a piece of computer software. The consensus of experienced dendrochronologists is, however, that the proper use of COHECHA is for verification after visual crossdating; that is, the software is not a replacement for personal skill and experience. Indeed the linear correlation between two ring-width sequences, progressively shifted against each other, not uncommonly reveals more than one distinct, and roughly equal, maxima. In such circumstance an automated process could be quite erroneous. Standardization In its simplest form standardization involves the smooth tracking of the natural growth trend of ring width on age, followed by the division of the actual ring width by the smoothed estimates to form a sequence of relative ring width indices. Ideally such indices could be looked on as random variables distributed about unity with constant variance. The ratio of the 6 ------- actual and smoothed widths is taken rather than their difference since the raw ring-width series are, in general, heteroscedastic, i.e. the variance of the ring widths in a relatievly .small time period is a function of the mean ring width for the period. Indeed there tends to be a linear relationship between the mean and the standard deviation of ring width which, in turn, implies that the ratio will have stable variance. Early attempts at standardization relied on linear or modified negative exponential functions to model the natural growth trend, i.e Gt - P0 + Dit and Gt = k + a exp(-pt), respectively. Note that (Jj would generally be constrained to be < 0. Both these functions are monotonic and thus cannot accomodate the common pattern of increasing ring widths during the first several years of growth followed by the decreasing ring widths of subsequent years. In the fitting of these functions it is necessary, therefore, to remove from the analysis all data, prior to a subjectively chosen point. Since, from geometical argument, some form of exponential decline in ring width is, ultimately, to be expected, other more flexible exponential forms have been considered, e.g. Gt = atp exp(-ct), (Siren, 1961; Warren,1980) Gj = ata_,|f'xexp[-(t/P)B], (Yang et al., 1979) Gt = a[l + Pexp(*yt)]k, (Van Deusen, pers. com.) The last mentioned embodies the modified negative expontial (k = 1) and the logistic growth curve (k = -1) as special cases. In addition, the power function Gt = at~p has been employed by Kuusela and Kilkki (1963). These functions are either monotonic or possess a single maximum and, thus, are unable to track any departures from trend caused by endogenous or exogenous disturbances, including change in competitive status. Their applicability is thus generally limited to undisturbed trees in open canopy stands where inter- tree competition is not a factor. This has often led dendro- chronologists to avoid sampling closed-canopy stands; such action would, however, prohibit a study of the interaction between pollution effects and competition. 7 ------- The earliest, and simplest, method of avoiding this limitation was to represent the growth trend as a polynomial, G, = pQ + Pjt + P2t2 + . . . + pptp with the order of the polynomial unknown and dictated by the usual forward inclusion stepwise procedure of fitting. Whereas there may be some physical justification for the exponential-type functions, the same cannot be said of the polynomial; to describe the latter as a "model" is to stretch the meaning of the word. While the exponential-type models may be too simplistic the polynomial is too data dependent; it is also known to suffer from potentially severe end-fitting problems and poor local goodness of fit (Cook and Peters, 1981). A rather different approach, advocated by, in particular, Parker (1970), is the application a digitial filter. Specifically Gt = £ wi^t-i i i=-n rt 2 w, = 1, and, for i ~ 0,w, = w_r l=-n This type digital filter is simply a centrally weighted moving average of the actual ring widths. The number, 2n+l, and value of the weights, wt, depends on which frequencies of variation are to be retained and which are to be eliminated (filtered out). Thus, three weights in the proportion 1:2:1 will block out rapidly changing variations and pass slowly changing variations; such is refered to as a low-pass filter. Conversely one may have a high-pass filter, i.e. one that blocks frequencies with relatively long wave lengths (low frequencies) and transmits the high frequencies (short wave lengths). Both kinds may be used simultaneously. It is clear that, for application to tree-ring widths, there is little, if any, basis for selecting the weights and, thus, it would be difficult to justify the choice of any particular filter. Different filters will lead to different index values, although once a filter has been prescribed the procedure becomes totally data dependent; it can not therefore be properly described as model-based. A somewhat similar procedure is cubic-spline smoothing introduced by Cook and Peters (1981). In this, cubic curves are fitted to subsets of points and pieced together under the restriction that the first derivative be continuous. This is again a totally data-dependent, rather than model-based, technique. A rather different approach is that taken by Barefoot et al. (1974), namely exponential weighted smoothing, in which an estimate of each point in a series is a weighted sum of all 8 ------- observations preceding the point, with heaviest weight being given to the most recent observations. Specifically Gt = Rt + [(l-ot)/a]Rt where Rt = aRt + (1—0£) Rt_j and Rt = a(Rt - r,_,) +(i-a) Rt.j (Brown, 1959) The current average, Rt, is, thus, taken as a weighted average of the current observation, R(, and the previous average, Rt_j. <\> Likewise, the current trend, Rt/ is taken as a weighted average of the current trend, Rt - Rt_,, and the previous trend, Rt_j. The smoothed current value, Gt, is then estimated as the current average plus an adjustment for trend. The average and estimate for the first year may be taken as the observed value, and the trend as zero. [Note: the versions given in Barefoot (1974) and Cook (1987a) both contain apparent typographical errors]. The quantity a is the weighting factor that determines the degree of smoothing; Barefoot et al. chose a = 0.2 which allowed the previous 10 - 15 years data to influence the current year's estimate. Thus, in contrast to the symmetrical digital filter, the exponential smoother is a one-sided filter that relies on only the current and prior values for its estimate; the estimates thus have the desirable feature of evolving in time in a manner that parallels the growth of the tree. The resultant series is nevertheless dependent on a, the choice of which is again omewhat arbitrary, and once a is selected, the procedure is likewise totally data dependent. Another method of trend removal is obtained from simply differencing the successive observations. If the ring-width series is taken as a random walk with a determininstic drift then Rt = Rt_i + et + 8 where et is a serially uncorrelated random variable and 6 is the deterministic drift of the process. Then the average of the VRt = Rt - Rt_, provides a measure of the drift. In practice a logarithmic transformation is usually applied, i.e. one works with the log(Rt) rather than the Rt. This is tantamount to assuming that Rt = P exp(-kt)(1 + €t) . A disadvantage of this procedure is that no direct estimate of the growth trend component, Gtf is obtained. However, as with other standardization processes, it might provide a stationary series of random variables to which autoregressive - 9 ------- moving average modeling (e.g. Box and Jenkins, 1970, 1976) can be applied [see the following section on chronology development]. On the other hand Monserud (1986) points out that a f-irst order autoregressive model with $ = 1.0 [defined in the following section] is assumed to be the underlying model if first differences are to produce a stationary series, and that such strong nonstationarity is unlikely to be found in tree growth, for the factors causing the changing growth trend over time are fairly well understood and can largely be explained by changes in tree size and stand density. Taking first differences thus seems an unnecessarily strong assumption for producing stationary series, and Monserud believes that fitting a growth trend and then standardizing seems more reasonable. Warren (1980) attempted to bridge the gap between overly simplistic models and totally data-driven smoothing by the introduction of a compound increment function. His approach has not received much attention, probably because of a perceived awkwardness in estimating the parameters, although Monserud (1986) notes, with respect to four chronologies for which he was forced to use a ninth degree polynomial for removal of the growth trend, that "the additive [i.e. compound] version of Warren's (1980) original model would likely perform very well in describing such complicated series". Before closing this section mention should be made of the very different method employed by the Russians Shiyatov and Mazepa (1987) . It is based on generating smooth curves from the maximal and mininal possible ring widths for each series as a function of time; this leads to a so-called "corridor". Currently it appears to be a purely graphical and subjective technique with undefined mathematical properties. Chronology Development On the completion of standardization for each tree from a particular site, the next step in the conventional process is to form a site chronology by averaging the index values on a year- by-year basis. Such a series should then contain a measure of the climatic component, Ct, plus any endogenous and exogenous disturbances that have not been removed along with the natural growth trend. If standardization has been accomplished by fitting (non-compound) exponential-type (or linear) functions, the effects of any endogenous or exogenous disturbances will also remain in the individual series. The effects of non-synchronous endogenous disturbances should, of course, be reduced by the averaging process. A more effective means of minimizing these effects seems likely achieved by the use of robust estimates such as the biweight mean 10 ------- (Mosteller and Tukey, 1977). Cook (1987a) gives some indication of the quantitative improvement that can be achieved in this manner and concludes that "the results indicate the high level of outlier contamination in closed-canopy forest tree-ring data that will corrupt the common climate signal if left unattended". With the exception of Cook (1987b) no attention appears to have been given to the forming of the principal components, yet this seems to be a situation in which principal components logically might be expected to be useful. Suppose that the majority of the standarized series (from a given site) carry a common signal (plus random noise) but a few series are aberrant in their pattern due to undetected or unallowed for endogenous disturbances. The first principal component should then tend to give equal weight to those series with the common signal but little, or inconsistent weight, to the remaining series. Indeed, if the remaining series also carried a common, albeit different, signal, this should show up in the second principal component. Thus, the computation of the principal components opens up the possibility of demarcating subgroups of trees that respond similarly within the subgroup but differently between subgroups. By definition, the first principal component should maximize the signal to noise ratio. Any need to exclude, subjectively, series perceived to contain a low common signal would be obviated. Whether the principal component would have any advantage over the biweight estimate is an open question since, presumably, the latter allocates its weights on a year-by-year basis. This could result in a series being heavily weighted in one year while receiving little weight in another. This aspect merits further investigation. Because of the likely existence of persistence effects, the application of autoregressive (AR) or autoregressive-moving average (ARMA) methodology would seem appropriate, and possibly mandatory in the case of standardization by the fitting of (non- compound) exponential-type functions or where the detrending has been done by the method of differencing. An autoregressive model for the index values may be written as = + $2*»-2 * * * $p*l"P + ef Thus the response at time t is dependent on the responses of the p previous years? the model is said to be of order p and written AR(p). There is thus a carry over effect of previous years plus a random perturbation. Complementary to this is a moving average model (this should not be confused with the moving average, or running mean, filter used to remove certain frequencies for curve smoothing? 11 ------- see above). It may be written and said to be of order q, i.e. MA(q). It represents a linear combination of the current and the q prior random perturbations. The two may be combined into an autogressive-moving average model of order p,q, i.e. ARMA(p,q). Differences in tree size or position within a stand can produce differences in either the speed of degree of response to an environmental "shock". Thus "to remove the effects of unwanted, disturbance-related transience on the common signal among trees, the tree-ring indices can be modelled and "prewhitened" by AR(p) and ARMA(p,q) processes before the mean value function is computed. The order of the process can be determined at the time of estimation using the Akiake Information Criterion (Akiake, 1974)" (Cook, 1987a). The biweight mean is recommended. The state-of-the-art means of accomplishing this is generally looked on as being the autoregressive standardization program, ARSTAN, of Cook (1985a) [see also Cook and Holmes, 1985; Holmes et al. 1986]. Monserud (1986) reports that, in 31 out of the 33 series he examined, a satisfactory representation was given by ARMA(1,1) models. With AR(p) models a satisfactory representation was obtained with p = 1, 2, 3 and 4 in 3, 15, 5 and 10 cases respectively. Cook (1987a) reports that he found AR(1) - AR(3) models satisfactory in most cases with his studies of eastern North American conifer and hardwood chronologies. From the ARMA(p,q) parmeters, $ , 6j, one may estimate the et and hence the (robust) means et, which may be taken as an initial estimate of the climate component, Cr Cook (1987) points out that this estimate is incomplete in that there still may exist a persistence component. An autorgressive model may again be applied and, hence, a final estimate of Ct generated. [There are complications in employing an ARMA model at this stage; these are explored in Guiot (1987)]. Disturbance Effects Once the climate component, which is assumed to be common to all trees in a site, has been estimated, it is possible to estimate the effects of any endogenous disturbances, that were not taken out in the standardization process, on the individual trees. [Any effects of exogenous disturbances that remain are common to the trees of a site and should be embodied in the estimate of the climate component]. Removal of the common 12 ------- component from an individual tree series should leave the effect of the disturbance(s) plus random noise. In the terms of Cook (1987a) a disturbance pulse~can be thought of as an autocorrelated form of Et in the sense that Dlt represents a persistent departure from the common signal. However Dt differs from Et in that, while Et is always present to some extent, the random shock leading to the creation of Dlt has a defined arrival time and an impact on radial growth that eventually disappears from the record (cf. Warren, 1980). These properties suggest that an endogenous disturbance may be modelled by means of intervention analysis (Box and Tiao, 1975), cf. Warren (1983). This assumes that the times of the interventions are known and, unfortunately, for trees within forest stands, this is rarely, if ever, the case. The problem can, however, be addressed by "intervention detection" (e.g. Reilly, 1984) which is essentially an a posteriori search performed by undertaking intervenion analysis at successive points in the sequence until an identifiable effect is located. However, from the viewpoint of detecting a pollution-related signal, it is the removal, rather than the estimation, of the effects of endogenous effects that is important. Intervention analysis and intervention detection may also be appropriate tools for separating the effect of exogenous disturbances, D2t, from the climate component, Ct. The date of exogenous disturbances, such as fire, fertilization, or epidemics of defoliating insects, will often be known, or much more readily inferred from the data than the effects of endogenous disturbances. It is also necessary that this component be removed before one can attempt to relate the climate effect to meteorological measurements. Effect of Climate To this point the objective has been to obtain an estimate of the climate component, Ct, free from the natural growth trend and any endogenous or exogenous disturbances, including atmospheric pollution. If this component can be accurately described in terms of meteorological measurements, after allowing for the effects of age and other disturbances, it may be possible to predict ring widths of trees suspected of being affected by atmospheric pollution. Consistent overestimation (or under- estimation) would then reasonably suggest a negative (or positive) effect of pollution and, indeed, provide a measure of that effect. Most attempts at relating the estimated climatic signal to meteorological variables have followed Fritts et al. (1971), or Ashby and Fritts (1972) (also Fritts, 1976). The dependent 13 ------- variable is the site chronology obtained from the indexed series, or, of course, the estimated Ct. The weather variables are taken as the monthly mean temperatures and total precipitation for January through September of the associated growth year and for May through September of the previous year, that is 14 temperatures and 14 precipitations for a total of 28 potential predictor variables. Mean ring-width indices from l to 3 years previously have sometimes been included, primarily when the persistence effect has not otherwise been accomodated. There may thus be up to 31 potential predictor variables. To avoid the problems of multicollinearity the principal components of the weather variables then replace the actual monthly values. The number of potential predictors is sometimes then reduced by omitting those components which account for a negligible amount of the variability amongst the predictors. In this way Ashby and Fritts (1972) reduced the number of weather- related variables to 20. A forward-selection stepwise-regression procedure is then employed, inclusion being based on the sequential F statistic being > 1. This seems a very liberal criterion since, under the null hypothesis, the expected value of the, F statistic is d/(d-2) where d is the number of degrees of freedom of the denominator. More stringent criteria (e.g. F > 2) have sometimes been used. The resulting multiple coefficient of determination, R2, is usually computed and its significance judged by reference to standard statistical tables. A drawback here, that seems not to be generally recognized, is that the probability values obtained from standard tables are not applicable in the case of a "steered" regression, of which forward selection is a particular form. In this, one is not picking a potential predictor, or set of predictors, at random, for which the standard tables would be applicable, but the predictor, or subset, with the maximum R2. Thus the critical values of a test of the null hypothesis, at any specified level, must exceed those given in the standard tables. In the limited number of cases where appropriate values have been determined, the difference between the proper value and that given by standard tables is, in general, substantial. (See for example, Draper et al., 1971; Pope and Webster, 1972; Rencher and Pun, 1980). Unfortunately the problem is not well documented in the statistical literature, possibly because of the inherent theoretical difficulties. The few explicit results that are available have been, for the most part, generated by Monte Carlo methods. There are vast number of possible combinations of the number of data points, the number of potential predictors and the number of these selected (let alone the correlation structure of the predictor variables, in the general case). Also the time for 14 ------- computation becomes substantial so it is not surprising that only a small segment of the possible space has been covered. Reliance on conventionally computed probabilities or tabled critical values seems to have led to undue confidence in the resulting regressions to predict the effect of weather variables. McLenahen and Dochinger (1985), for example, present, inter alia, the number of principal components entered under a forward- selection F > 1 criterion and the resulting values of the multiple R2, for all combinations of 5 sites and two periods (1901-1930 and 1931-1978); all 28 principal components were, however, considered. Interpolation and modest extrapolation of the Rencher and Pun (1980) tables show that, while all ten values of R2 exceed their expected values under the null hypothesis, only one exceeds the 5%-level critical value, and it only marginally so (see Appendix). There is thus serious doubt whether the derived regressions have any real predictability. Indeed, in relation to their expected values, the pattern of results for the two periods is remarkably similar, seemingly invalidating the authors' conclusion that tree growth in the more recent period had come under strong influence of factors unrelated to climate. Ashby and Fritts (1972) included 12 out of 2 0 orthogonal (principal-component) predictors. These accounted for 59% of the variability in the ring-width index chronology. The series length was 55 years. This combination of values is too far removed from the Rencher and Pun tables for extrapolation to be reliable (Rechner and Pun provide a formula for approximate interpolation but state that extrapolation with it may be risky and needs further investigation). It would appear, however, that significance at the 5% level would be, at best, marginally attained. Cropper (1982) took an actual 32-year chronology and regressed it against regional climatic data. He did the same with a computer-generated independent chronology (i.e one generated from random numbers without regard for the climatic data, so that only chance associations with a climatic data set could be expected). He reported that 88% of the variance of the actual chronology was accounted for exclusively by climate and 15 climatic variables were deemed significant at the 5% level. Slightly less variance was accounted for in the case of the artificial chronology but 16 variables were included as significant. The analyses were run in exactly the same way following Fritts et al. (1971). Cropper concluded that "there is the potential for the response-function techniques to be misleading by attributing significance to non-significant variables". 15 ------- Notwithstanding the significance, real or otherwise, of the resulting regression, validation of the chosen regression is to be recommended. The general approach here is to divide the total series into two parts, one from which the regression parameters are estimated, the other to which the resulting regression is applied and its performance evaluated. It is assumed that the "verification" set is not affected by some undetected disturbance, such as pollution; thus the time period up to, say 1940 may be used for estimation, 1940-1960 for verification, and the years since 1960 omitted. The difficulty lies in choosing a meaningful measure of the regression's applicability in the verification period. The so called "reduction of error" (RE) statistic has been employed (Kutzbach and Guetter, 1980; Fritts et al., 1979). Specifically RE = 1 - i(yt - y^VKy, - yc)2 where the summation is over the verification set, the y( and £ are the actual and estimated values, while y^ is the mean value from the estimation (or calibration) set. Although the properties of this statistic do not appear to have been studied, positive values are regarded as as indicitive of satisfactory performance. Unfortunately, it is possible to obtain, from the estimation sets, a collection of multiple R2 values that appear to be in good agreement with the null distribution (after allowance for forward selection), and thus imply little or no predictability, along with a preponderance of positive values of RE from the verification sets. This seems pardoxical, but it should be noted that RE may be positive because £(y - y()2 is small, which would be the case if the verification data were closely tracked by the regression estimates, and/or because £(Y, " Yc)2 is large. The latter may result from the climate variables in the verification period being atypical in relation to those experienced in the calibration period. This is an aspect that deserves further study. There is another problem. The majority of trees for which dendrochronological records are available are those which have survived to the present, or at least recent, time. Thus trees that may have succumbed to some environmental factor will be excluded from the record. A reccurrence of that factor may then show up as a decline simply because it cannot been fairly represented in the regressions. Finally there is the choice of potential predictors. Calendar months are an artificial division of the year which are not, in general, in synchronization with biophysical developments. It would seem to make very little difference to a plant if 2 inches of rain fell on 31 March or on 1 April, but this could have a noticeable effect on regressions based on 16 ------- monthly values. The forming of principal components, which is first and foremost and arithmetical artifact, is unlikely to remedy this. Next, from the physiological point of view, it is difficult to see how additive linear models of precipitation and temperature can give anything other than, at best, a very crude predictor. The use of the Palmer drought index, as a measure of moisture availability, as advocated by some authors, would seem to be a step in the right direction, but not on a calendar month basis. Further, there is no doubt a differential temperature response at any given moisture level, provided the latter is adequate, but what constitutes an adequate moisture level is, itself, likely a function of temperature. Further, if the moisture level is adequate, why should additional moisture result in any response? May there not be a threshhold temperature below which growth will be impeded whatever the moisture availability? A more sophisticated physiological approach to relating growth to climate seems necessary. Newer Concepts The most recent development in the analysis of tree-ring data has been the introduction of the Kalman filter. The method stems from Kalman (1960) who presented a system for updating and predicting that is based on the space-state formulation of a linear dynamic model. It encompasses ARMA models (Box and Jenkins 1970, 1976), standard multiple regression, and regression models with time-varying parameters (Harvey, 1981). The methodology has application in other areas but it relevance to dendrochronology seems not to have been recognized prior to Visser and Molenaar (1986, 1987) and Van Deusen (1987). It provides a means of simul- taneously reducing a number of series to a single chronology and generating climate-based predictions. The climate parameters can be allowed to vary over time to provide a test of the hypothesis that conditions in the past are similar to conditions at present. The essence of the Kalman filter is in two equations, the "observation" or "measurement" equation, + Xt and the "transition" or "system" equation, a = G.a w . in which the state variables a( evolve over time according to a first-order Markov process. In a dendrochronological application y is an nt x 1 vector of tree-ring data at time t, F, is an nt x p matrix of known quantities (e.g. the values of climate variables) , at is a p x 1 vector of 17 ------- underlying state parameters, Gt is a fixed p x p matrix, v and wt are, respectively, nt x 1 and p x 1 vectors of residuals with zero expectation of covariance matrices Vt and Wt. Van Deusen(1987) and Visser and Molenaar (1987) provide examples. This seem to provide the closest thing to a unified mathematical framework for analyzing tree-ring series; indeed, Visser and Molenaar (1987) note the possibility of incorporating the estimation of growth curves and the effect of weather variables in the one model. Nevertheless the approach is not without limitations. Firstly there is the need to specify, explicitly or implicitly, a form for the natural growth function; Van Deusen's (1987) standarization by differencing the logarithms of successive ring widths is a case of implicit definition. Secondly, the first-order Markov assumption for the transition equations is rather strong; put simply, it means that the situation in the current year is determined, apart from a random perturbation, solely by the situation in the immediately preceding year. Finally, the covariance matrices, Vt and Wt, have to be known in advance. In general this is not the case and their "estimation" will involve some element of subjectivity. It may well be that departures from these assumptions have little impact on the inferences that result; this possibility merits investigation. In addition, there is nothing in this approach that would resolve the criticism concerning the choice of climate variables; in short, monthly temperatures and precipitations are no less inappropriate under this formulation as under any other. If these and other such difficulties prove inconsequential, or can be satisfactorily resolved, the amalgamation of intervention analysis and Kalman filtering appears to have considerable potential as an aid to answering the current concerns about the existence and magnitude of any effect of atmospheric pollution on tree growth. This potential is highlighted by the apparent success of this mix elsewhere. Thus, the abstract of the paper "Using the Kalman filter to include intervention analysis in starmax models" presented by D. K. Blough, University of Arizona, at the 1987 joint meetings of the American Statistical Association reads as follows: "The Kalman-filter approach has been usd in fitting ARIMA 18 ------- models. This paper uses the multivariate extension of this approach to model spatial time series which include periodic interventions whose effects damp out exponentially. Included in the development of these models will be previously studied techniques for the inclusion of missing values, aggregate values, non-linear transformation of the data, covariates and spatial relationships. An example will be presented in which these models are applied to the study of boll worm moth counts in cotton fields over time. The amount of irrigation water in the field is the covariate and the periodic application of insecticide is the intervention. Maximum likelihood estimates of the model . parameters are obtained using a quasi-Newton routine in the GAUSS programming language on a personal computer." Wood Density The statement that the history of a tree's response to its environment is embodied in its sequence of tree rings is not meant to apply solely to ring widths. To determine the amount of woody material actually produced one must consider also the density of the wood, or the ring mass, the latter being defined as the product of ring density and ring width. As noted above, the density is not uniformly distributed across the ring but normally increases steadily from the low density earlywood to the high density latewood. The ability to examine the intra-ring density profile essentially began with the pioneering work of Polge (1963) in the application of X-ray densitometry to tree-ring research. Prior to this some use had been made of beta radiation (Cameron et al. 1959) but it is now generally conceded that, for most purposes, the X-ray method is superior. Soft gamma rays have also been employed (Woods and Lawhon, 1974) as well as photogrammetric methods (Green, 1964, 1965; Elliott and Brook, 1967). Such techniques permit the measurement not only of the ring width but also the complete intra-ring density profile. From this profile one can derive such summary statistics as the proportion of latewood (and, of course, the widths of the earlywood and latewood), the average and maximum density of the latewood, the average and minimum density of the earlywood and, of course, the whole ring average density. The location within the ring of the maximum and minimum density also can be recorded. Further any deviations from the normal profile can be identified. Since the wood cells are laid down sequentially during the growing season, such deviations must be in response to some more or less synchronous environmental phenomena. Indeed, Fritts (Chapter 2 1976) writes "Considerable attention has been given in this chapter to the variations that 19 ------- occur in cell structure within the annual ring. It has been stated that these variations can be attributed to particular characteristics of the growing season and the associated climates. While such small-scale variations in cell size have little net effect on ring widths, they are an important source of information that can be measured by densitometric analyses. The close association between cell size, wood density, and factors of the environment point to a tremendous amount of information that must be present in the rings and which should be extractable by measuring density variations along ring widths. If these new data on density can be successfully related to and calibrated with environmental data, the simultaneous analysis of density and ring width may provide significantly more information on past climate than is possible using only the widths of rings". The addition of density to ring width can aid in the troublesome operation of crossdating. Parker (1967, 1970) reports the dating of high latitude spruce and larch samples that may not have been possible from ring widths alone. Nevertheless, even during the past two decades, most dendro- chronological studies have been based solely on ring widths. This, perhaps, reflects the forester's preoccupation with volume rather than, and sometimes at the expense of, wood quality. Several facilities for X-ray densitometric measurement of increment cores exist in North America, principally the Laboratory for Tree-Ring Reseach at the University of Arizona, the Lamont-Doherty Geological Observatory, Palisades NY, and Forintek Canada Corp., Vancouver BC, but only at the last mentioned does it appear than any extensive densitometric chronologies have been constructed. Other North American facilities include the Oak Ridge National Laboratory, Oregon State University, the University of Kentucky, and Laval University, Quebec; these, however, have to date made little or no contibution to the subject. Outside North America the principal facilities appear to be at the Centre National de Recherches Forestieres, Champenoux, France, the Swiss Federal Forestry Research Institute, Birmensdorf, the Oxford Forestry Institute, England, and the Forestry Research Institute, Rotorua, New Zealand. Notable conributions to the application of densitometry to tree rings have been made by workers at all these four installations. The influence of environemntal factors can be reflected in one of more of ring width, and the various density measures mentioned above. Conkey (1984a,b) reported that several stands of high elevation red spruce in Maine sampled in 1977 showed no real decline in ring-width values in recent decades relative to the 200-to-300-year record of growth, but the series of maximum wood densities from the same trees displayed a flattening of high- 20 ------- frequency variation and a lower mean value since 1962. This led her to suggest that wood density may provide an "earl^ warning signal" of growth decline (Conkey, 1987). Bodner (pers. comm.) examined the density profiles of trees in the neighborhood of a point source of pollution that became active in 1979. He found that, beginning with that date, there was a clear decrease in maximum density coupled with an increase in minimum density. An almost identical pattern was observed by Parker (pers. comm.) with material obtained under similar circumstances. Any one of the above density measures can be analyzed in a manner parallel to ring width, but it would seem appropriate to handle all variables simultaneously, i.e. construct multivariate analogs of all or any of the above approaches and, in particular, intervention analysis and Kalman filtering. There is, accordingly, a substantial, and seemingly relevant, area for research that has not, as yet, been adequately exploited. Other Concerns For practical reasons most tree-ring analysis has been performed on increment cores obtained at breast height. It seems possible, however, that environmental factors would be better refected in cores taken close to the base of the live crown. Sample collection of such would, however, be relatively expensive. Further, not only is the base of the live crown sometimes difficult to define but also its position advances up the bole as the tree ages. Some studies based on cores, or discs, take at various positions along the bole, nevertheless, have been, carried out, e.g. Parker et al. (1976). Most samples have come from individual trees, mainly dominant or codominant if not open-grown, without regard to their location relative to neighboring trees. Thus, while one may be able to relate individual tree growth to environmental factors, the data provide no information on stand behavior. For example, one may conjecture that, because of genetic variability, some trees could be more adversely affected by pollution than others. The more affected ultimately would provide less competition to any less affected neighbors which, in turn, should respond positively to the reduced competition. This is a possible explanation for the behavior of a set of cores presented by Schweingruber (1987). Of these, some show an abrupt and roughly simultaneous decrease in ring width, others show no change while in others there is a suggestion of a synchronous increase. Not surprisingly, therefore, Schweingruber appears to advocate looking at trees individually rather that automatically generating site chronologies. It also highlights the need for 21 ------- direct evidence of the magnitude of competitive effects on tree ring measures. While the temporal variation of ring-width series has received a great deal of attention, there has been little formal investigation of their spatial variation. By this is meant, not necessarily the relative location of sample trees within a site but, more particularly, the inter-site variation. Ord and Derr (1987) give an illustration of how the some of the relatively simple techniques given in, for example Cliff and Ord (1981) may be utilized, informatively, in this context. Note that Blough, quoted above, included a spatial component in his study. Conclusion Notwithstanding the problems, tree rings provide the best hope for determining if there is an effect of atmospheric pollution and, if so, how its magnitude relates to levels of deposition. The traditional methods of dendrochronology and dendroclimatology are, however, not well suited to these new objectives. In this respect, while great credit must be given to Fritts and his coworkers for their development of much of the conventional methodology, the uncritical adoption of that methodology may have done more harm than good. In some instances it has clearly led to a confidence in interpretations that, in actuality, cannot be justified. Alternative methodological ideas have recently found their way into this arena. Hands-on experience with these techniques in this context is to date very limited, and it is too early to determine whether they will provide the hoped-for panacea. Further, intra-ring density must also assume a role at least equal to that of ring width. 22 ------- References Akiake, H. 1974. A new look at statistical model identification. IEEE Trans, on Automatic Control, AC-19(6):716-723. Ashby, W.C. and Fritts, H.C. 1972. Tree growth, air pollution and climate near LaPorte, Ind., Bull. Amer. Meteorol. Soc. 53:246-251. Barefoot, A.C., Woodhouse, L.B., Hafley, W.L. and Wilson, E.H., 1974. Developing a dendrochronology for Winchester, England. Jour. Inst. Wood. Sci. 6(5):34-40. Box, G.E.P. and Jenkins, G.M. 1970, 1976 (2nd ed.) Time Series Analysis: Forecasting and Control. Holden Day, San Francisco. Box, G.E.P. and Tiao, G.C. 1975. Intervention analysis with applications to environmental and economic problems. Jour. Amer. Statist. Assoc. 70:70-79. Brown, R.G. 1959. Loss risk in inventory estimates. Harvard Business Review, 37(4):104-116. Cameron, J.F., Berry, P.F. and Phillips, E.W.J. 1959. The determination of wood density using beta rays. Holzforsch. 13(3):78-84. Cliff, A.D. and Ord, J.K. 1981. Spatial Processes: Models and Applications. Pion, London, (Methuen NY). Conkey, L.E. 1984a. Dendrochronology and forest productivity: Red spruce wood density and ring width in Maine. USDA Forest Service General Technical Report NE 90, p.69-75. Conkey, L.E. 1984b. X-ray densitometry: Wood density as a measure of forest productivity and disturbance. In: Air Pollution and the Productivity of the Forest, Ed. D.D. Davis, Izaac Walton League of America, Washington DC, pp. 287-296. Conkey, L.E. 1987. Wood density: An early indicator of forest decline? In: Tree Rings and Forest Mensuration: How Can They Document Trends in Forest Health and Productivity. NCASI Technical Bulletin (in press). Cook, E.R. 1985a. A time series approach to tree-ring standard- ization. Unpubl. Ph.D. dissertion, University of Arizona. 23 ------- Cook, E.R. 1985b. The use and limitations of dendrochronology in studying the effects of air pollution on forests. Proceedings NATO Advanced Workshop on Effects of Acid Deposition on Forests, Wetlands and Agricultural Systems, Ed. T. Hutchinson, Toronto, Canada. Cook, E.R. 1987a. On the disaggregation of tree-ring series for environmental studies. Proceedings of the International Symposium on Ecological Aspects of Tree-Ring Analysis, Palisades, NY, August 17-21, 1986. pp. 522-542. Cook, E.R. 1987b. A tree-ring analysis of red spruce in the southern Appalachian Mountains. USDA Forest Service Report (in press). Cook E.R. and Holmes, R.L. 1985. User's manual for program ARSTAN. Laboratory of Tree-Ring Research, Univ. of Arizona, Tuscon AZ. 29pp. Cook, E.R. and Peters, K. The smoothing spline: a new approach to standardizing forest interior tree-ring width series for dendroclimatic studies. Tree-Ring Bull. 41:45-53. Cropper, J.P. 1982. Comment on H.C. Fritts: The climate-growth response. In: Climate from Tree Rings, Ed M.K. Hughes, P.M. Kelly, J.R. Pilcher and V.C. LaMarche Jr., Cambridge University Press, pp. 47-50. Dean, J.S. 1978. Tree-ring dating in archeology. University of Utah Anthropology Paper 99:129-163. Dieterich, J.H. 1980. Chimney Spring Forest fire history. USDA Forest Service Research Paper RM-220. Draper, N.R., Guttman, I. and Kanemasu, H. 1971. The distribution of certain regression statistics. Biometrika 58:295-298. Druce, A.P. 1966. Tree-ring dating of recent volcanic ash and lapilli. New Zealand Jour. Bot. 4(1):3-41. Elliott, G.K. and Brook, S.E.G. 1967. Microphotometric technique for growth-ring analysis. Jour. Inst. Wood Sci. 18:24-43. Francis, P. 1976. Volcanoes. Pelican Books, Harmondsworth. Fritts H.C. 1976. Tree Rings and Climate. Academic Press NY. 24 ------- Fritts, H.C., Biasing, T.J., Hayden, B.P. and Kutzbach, J.E. 1971. Multivariate techniques for specifying tree-growth and climate relationships and for reconstructing anoltralies in paleoclimate. Jour. Appl. Meteorol. 10:845-864. Fritts, H.C., Gordon, G.A. and Lofgren, G.R. 1979. Variations in climate since 1602 as reconstructed from tree rings. Quat. Res. 12:18-46. Graybill, D.A. 1982. Chronology development and analysis. In: Climate from Tree Rings, Ed. M.K. Hughes, P.M. Kelly, J.R. Pilcher and V.C. LaMarche Jr., Cambridge University Press pp. 21-28. Green, H.V. 1964. Supplementary details of construction of the stage and drive assembly of the scanning microphotometer. Pulp and Paper Research Institute of Canada, Res. Note 41. Green, H.V. 1965. Wood characteristics IV: The study of wood characteristics by means of a photometric technique. Pulp and Paper Research Institute of Canada, Tech. Pub. 419. Guiot, J. 1987. Standardization and selection of the chronologies by the ARMA process. In: Methods in Dendrochronology: East- West Approaches. Ed. L. Kairiukstis. IIASA/Polish Academy of Sciences (in press). Harvey, A.C. 1981. Time Series Models. Phillip Allan Publishers. Helley, E.J. and LaMarche, V.C. Jr. 1973. Historic flood information for Northern California streams from geological and botanical evidence. U.S. Geol. Surv. Prof. Pap. 485-E: 1-16. Holmes, R.L. 1983. Computer-assisted quality control in tree- ring dating and measurement. Tree-Ring Bull. 43:69-75. Holmes, R.L., Adams, R.K. and Fritts, H.C. 1986. Tree-ring chronologies of Western North America: California, Eastern Oregon and the Northern Great Basin, with procedures used in the chronology development work including users manuals for computer programs COFECHA and ARSTAN. Chronology Series VI, Laboratory of Tree-Ring Research, Univ. of Arizona, Tuscon. Hughes, M.K., Schweingruber, F.H., Cartwright, D. and Kelly, P.M. 1984. July-August temperature at Edinburgh between 1721 and 1975 from tree-ring density and width data. Nature 308:341- 344. 25 ------- Kutzbach, J.E. and Guetter, P.J. 198 0. On the design of paleo- environmental data networks for estimating large scale patterns of climate. Quat. Res. 14:169-187. Kuusela, J. and Kilkki, P. 1963. Multiple regression of increment percentages on other characteristics of Scots-pine stands. Finnish Soc. For. 3 5 pp. LaMarche, V.C. Jr. and Hirschboek, K.K. 1984. Frost rings as records of major volcanic srruptions. Nature 307:121-126. LaMarche, V.C. Jr. and Wallace, R.E. 1972. Evaluation of effects on trees of past movements on the San Andreas fault, northern California. Geol. Soc. Amer. Bull. 83:2665-2676. McClenahen, J.R. and Dochinger, L.S. 1985. Tree ring response of white oak to climate and air pollution near the Ohio River Valley. Jour. Environ. Qual. 14:274-280. Monserud, R.A. 1986. Time-series analysis of tree-ring chrono- logies. Forest Sci. 32:349-372. Mosteller, F. and Tukey, J.W. 1977. Data Analysis and Regression. Addison-Wesley, Reading MA. Ord, J.K. and Derr, J.A. 1987. The utility of time series models and spatial analysis of forecast residuals in detecting recent changes in ring width of red spruce (Picea rubens, Sarg.) in the Great Smoky Mountains. USDA Forest Service Report (in press). Parker, M.L. 1967. Dendrochronology of point of pines. M.A. thesis, University of Arizona, Tuscon, 168 pp. Parker, M.L. 1970. Dendrochronological techniques used by the Geological Survey of Canada. In: Tree-Ring Analysis with Special Reference to Northwest America. Ed. J.H.G. Smith and J. Worrall. Univ. of British Columbia, Faculty of Forestry Bull. No. 7:55-66. Parker, M.L 1985. Investigating the possibility of a relationship between volcanic erruptions and tree growth in Canada (1800- 1899). In: Climate Change in Canada 5: Critical Periods in the Quaternary Climatic History of Northern North America. Ed. C.R. Harrington. Syllogeus 55:249-264. Parker, M.L, Hunt, K., Warren, W.G. and Kennedy, R.W. 1976. Effect of thinning and fertilization on intra-ring characteristics and kraft pulp yield of Douglas fir. Applied Polymer Symposium No. 28. Wiley NY. pp. 1075-1086. 26 ------- Parker, M.L., Jozsa, L.A., Johnson, S.G. and Bramhall, P.A. 1981. Dendrochronological studies on the coasts of James Bay and Hudson Bay. In: Climate Change in Canada 2. "Ed. C.R. Harrington. Syllogeus 33:129-188. Polge, H. 1963. L"analyse densitometrique de cliches radio- graphiques: une nouvelle methode de detremination de la texture de bois. Ann. Ec. Natl. Eaux Forets St. Rech. Exper. 20:530-581. Pope, P.T. and Webster, J.T. 1972. The use of an F-statistic in stepwise regression procedure. Technometrics 14:327-340. Reilly, D.P. 1984. Automatic intervention detection system. Proc. Business & Economic Statistics Section, American Statist. Assoc. pp.539-542. Rencher, A.C. and Pun, F.C. 1980. Inflation of R2 in best subset regression. Technometrics 22:4 9-53. Schweingruber, F.H. 1987. Dendrochronological studies in localised and regional areas with forest damage. IUFRO Workshop on Woody Plant Growth in a Changing Physical and Chemical Environment, Vancouver BC, July 27-31 (in press). Schweingruber, F.H., Fritts, H.C., Braeker, O.U., Drew, L.G. and Schaer, E. 1978. X-ray technique as applied to dendro- climatology. Tree-Ring Bull. 38:61-91. Shiyatov, S.G. and Mazepa, V.S. 1987. Some new approaches in the construction of more reliable dendroclimatological series and the analysis of cycle components. In: Methods of Dendrochronology; East-West Approaches. Ed. L. Kairiukstis. IIASA/Polish Academy of Sciences (in press). Siren, G. 1961. Skogsgranstallen som indikator for klimatfluk- tuationerna i norre fennoskandien under historik tid. Comm. Inst. For. Fenniae 54.2, Helsingfors, 66pp. Studhalter, R.A. 1956. Early histroy of crossdating. Tree-Ring Bull. 21:31-35. Sundberg, R. 1974. On the estimation of pollution-caused growth reduction in forest trees. In: Statistical and Mathematical Aspects of Pollution Problems. Ed. J.W. Pratt, Marcel Dekker, NY pp. 167-175. 27 ------- Van Deusen, P.C. 1987. Some application of the Kalman filter to tree-ring analysis. Proceedings of the Intenational Symposium on Ecological Aspects of Tree-Ring Analysis, Palisades NY, Aug. 17-21, 1986. pp. 566-578. Visser, H. and Molenaar, J. 1986. Time-dependent responses of trees to weather variations: an application of the Kalman filter. KEMA Report No. 50385-MOA 86-3041, The Netherlands. Visser, H. and Molenaar,J. 1987. Time dependent responses of trees to weather variations: an application of the Kalman filter. Procedings of the International Symposium on Ecological Aspects of Tree-Ring Analysis, Palisades NY, Aug 17-21, 1986, pp. 579-590. Warren, W.G. 1980. On removing the growth trend from dendro- chronological data. Tree-Ring Bull. 40:35-44. Warren W.G. 1983. The effect of industrial pollution on tree growth: a case study. In: Renewable Resource Inventories for Monitoring Changes and Trends. Ed. J.F. Bell and T. Atterbury. College of Forestry, Oregon State University, pp. 490-492. Yang, R.C., Kozak, A. and Smith, J.H.G. 1978. The potential of Weibull-type functions as flexible growth curves. Can. Jour. For. Res. 8:424-431. 28 ------- Appendix Tree Ring Response of White Oak to Climate and Air Pollution near the Ohio Valley: A Comment. William G. Warren In the study of the response of tree rings of white oak to climate and air pollution near the Ohio Valley by McClenahen and Dochinger (1985) the conclusions appear to be very much dependent on the values of the coefficient of determination, R2, obtained on fitting response functions in the manner of Fritts (1976). Specifically, chronologies of standardized ring-width indices were regressed on the principal components of monthly precipitation and mean temperature for January through September of the growth year and May through September of the preceding year - a total of 14 months or 28 potential predictor variables. The authors also considered additional variables, namely the growth of each of the preceding three years; regressions were performed with these variables included and excluded. A forward selection procedure was used with variables included as long as the F ratio for the next component to enter was greater than unity. Response functions were obtained for the years 1901-1930, 1931-1978 and 1900-1978. When prior growth was not included the number of principal components that entered ranged from 8 to 23 and was most commonly from 12 to 16. Our focus will be on these cases. Although McClenahen and Dochinger did not make any claims of formal statistical significance, it is clear that any inferences must be dependent on the results departing from what could be reasonably expected under the null hypothesis, namely the hypothesis that there is no association between the growth (as measured by the standardized ring-width indices) and the climatic factors, or the principal components thereof. An often unappreciated fact is that, in a forward selection procedure, the null distribution of R2 (of F) is not the standard distribution as widely tabled. One is not then dealing with a random value of R2 but with the maximum value of a set of random, but not mutually independent, values of R2, so that the expected value and the critical values for an alpha level test (alpha small, commonly 5%) of the null hypothesis are, in general, substantially greater than the tabled values. (See e.g. Draper et al. 1971, Pope and Webster 1972). The same applies to the multiple R2 obtained by a forward selection procedure. Because of the intrinsic theoretical difficulty little work on this topic has appeared. Rencher and Pun (1980) nevertheless provide Monte Carlo estimates of the expected value and 95th percentile of R2 for a limited number of combinations of sample size, n, number of potential predictors, l ------- k, and the number of variables selected, p. They assume that the regression variables are uncorrelated as is the case when the principal components of the set of potential predictors are used. Their results are based on 500 to 2000 realizations. Interpolation and modest extrapolation from the Rencher and Pun table gives Table 1. 2 4 P 6 8 10 30: E(R2) = .296 .461 .570 .651 .716 R. 95 = .430 .606 .716 .784 .839 48: R(R2) = .188 .295 .370 .426 .471 R. 95 = .275 .405 .493 .568 .620 For the regressions with prior growth excluded McClenahen and Dochinger give Table 2 1 2 Site 3 4 5 1901-30: n = 30: p = 23 15 19 21 19 R2 = .972 869 .965 .985 .960 1931-78: n = 48: p = 8 10 13 15 12 H CN « .529 484 .639 .576 .609 The values of Tables 1 and 2 are plotted in Figs. 1 and 2. Although the Rencher and Pun Tables do not extend beyond p = 10 it is clear in Fig. 2 that, while all the R2 values obtained by McClenahen and Dochinger fall above their expected values (i.e. their theoretical means) they all fall below the 95th percentile, i.e. are less than the critical value for a 5% level test and would commonly be declared "not significant". In Fig. 1 there is virtually an additional point since when p = ]c = n-1, R2 = 1.0 (assuming that an intercept is fitted). The dotted lines between p = 10 and p = 28 are conjectural but one can be confident that while all the McClenahen and Dochinger values of R2 again fall above their expected values, with one possible exception, they again all fall below the 95th percentile. If, indeed, the R2 for Site 4 exceeds the 95th percentile it does not do so by much. ii ------- A similar pattern arises for 1900-1978 but is not presented because of the uncertainty in extrapolating the Rencher and Pun table to n = 79. Thus, for both periods, 1901-30 and 1931-78, the pattern of results does not seem too far removed from what would be expected under the null hypothesis. In only one out of ten cases is an R2 possibly significant at the 5% level and there is about a one in three chance of such spurious significance arising in ten trials. Thus, notwithstanding the fact that all R2 values are somewhat greater their expected values, there is little that can be construed as definitive evidence of a relationship between ring width and the climatic factors as measured. More importantly, perhaps, is the fact that the pattern of the R2 values in relation to what would be expected appears to differ negligibly between the two periods. The choice of F > 1.0 as a selection criterion in forward inclusion seems rather foolhardy. Since the expected value of an F statistic is d/(d-2), where d is the number of degrees of freedom of the denominator, one must expect that a goodly proportion of variables with no predictive ability would be included. In summary, therefore, it seems legitimately questionable whether the climatic variables in the McClenahen and Dochinger study have any predictive ability and, thus, whether inferences based on the resulting response functions have any validity. References Draper, N.L., Guttman, I. and Kanemasu, H. 1971. The distribution of certain regression statistics. Biometrika 58:295-298. Fritts, H.C. 1976. Tree Rings and Climate. Academic Press, New York. McClenahen, J.R. and Dochinger, L.S. 1985. Tree ring response of white oak to climate and air pollution near the Ohio River Valley. J. Environ. Qual. 14:274-280. Pope, P.T. and Webster, J.T. 1972. The use of an F-statistic in stepwise regression procedure. Technometrics 14:327-340. Rencher, A.C. and Pun, F.C. 1980. Inflation of R2 in best subset regression. Technometrics 22:49-53. iii ------- Fig. 1 1901-1930 No Previous Growth 1.1 1 0.9 - 95th percentile 0:8 0.7 0.6 Expected 0.5 0.4 0.3 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 ------- Fig. 2 1931—1978 No Previous Growth 0.8 0.7 95th percentile 0.6 0.5 0.4 Expected 0.3 0.2 0.1 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 ------- |