Evaluating Coastal Douglas Fir Growth Responses to Nitrogen Application Using Tree Ring Chronologies

Abstract

Dendrochronology is a technique that can be applied as a retrospective monitoring (RM) approach to evaluate the performance of nutrient application in forest ecosystems. Applying the RM approach across operations lacks experimental controls, which may adversely affect accuracy and precision of estimates due to greater mismatches in stand conditions between treated and untreated plots. To test the rigor of the RM approach, we collected increment cores of coastal Douglas fir (Pseudotsuga menziesii var. menziesii) at eight sites of an experiment where stands were fertilized in 1971. First, we tested the approach under ideal conditions by sampling from treated and untreated plots of the experiment. Second, we tested the approach using newly established surrogate control (SC) plots, which differed in ecological site classification from those of the treated plots to understand how robust the approach was to mismatches in conditions between treated and untreated samples. We hypothesized that detrending ring width would mitigate error in responses to nutrient application resulting from mismatches in site classification. Within the experiment, the approach indicated an average increase of 15% (5% to 26%, p < 0.05) growth response to operational doses of urea. Different responses were found when the analysis relied on SC plots. Detrending low-frequency variation in ring widths eliminated differences in results arising, at least in part, to mismatches in site class. However, it also reduced the growth response using the experimental control plots to 10%. Dendrochronology with detrending shows promise in the ability to mitigate variation introduced by mismatches in ecological site classification that may occur in operational monitoring. Based on these results, we see potential to implement RM with operations to evaluate and optimize stand selection criteria.

1. Introduction

Nutrient management is valued as a cost-effective way to accelerate the development of forest stands; it can help generate more wood fibre over smaller harvest areas and sequester greenhouse gases in the process [1,2,3]. In British Columbia (B.C.), Canada, the aerial application of nitrogen (N) was conducted over 15,290 hectares (ha) annually from 2000 to 2018 (range: 923–29,898 ha) across public forests of B.C. [4]. The government-led program was guided by stand selection criteria that are designed to minimize the risk of adverse environmental impacts and maximize the response of stemwood growth. However, there are no operational-scale monitoring initiatives to evaluate how successfully the criteria maximize stemwood growth responses. In the absence of monitoring, previous experimental projects with N applications are the primary way to measure tree growth response and improve stand selection criteria. Although experiments have been effective in establishing average responses to N application for several important commercial tree species [5,6,7,8,9,10,11,12,13], additional information about how responses vary with site conditions, stand age, time after application, and repeated application remains unknown. The stand selection criteria, therefore, often remain general in nature.

Adopting data assimilation methods to improve the performance of stand selection criteria depends on the efficient collection of growth responses from a subset of operations. It is difficult to achieve the required sampling with fixed-area plots because the magnitude of area-based response to N application is low relative to among-plot variability in periodic estimates of stand growth derived from measurements of tree height and diameter at breast height (DBH). High spatial variability can be overcome by large sample size, yet that amount of sampling is prohibitive for large monitoring programs. This problem has led many to consider dendrochronology as a practical alternative source of response information [14,15,16].

Herein referred to as retrospective monitoring (RM), dendrochronology has been widely used to estimate tree growth responses to nutrient application [15,17,18,19,20]. The RM approach provides estimates of annual cambial growth of stemwood from measurements of ring width (w) that are produced by coring trees several years after fertilizer application. The strength of the RM approach comes from the ability to analyze observations of tree growth prior to fertilization, which is rarely considered in fixed-area plot experiments. Additional value is gained because growth of focal trees in the post-application period can be normalized by their previous growth in the pre-application period, which makes this approach best suited to assess responses after canopy closure [14].

Applying the RM approach in young lodgepole pine (Pinus contorta var. latifolia Engelm. ex S. Watson) stands, Brockley and Yole [16] described dissimilarities between treated and untreated plots as a complication that precluded reporting the significance of estimates. They called for improved stand selection criteria for operational fertilization as well as post-treatment monitoring.

In some cases, it may be easy to pair treated and untreated plots by ecological site classification (herein site series) using the well-established edatopic grid system for ecosystems of B.C., known as biogeoclimatic classification [21,22]. However, the prevalence of those opportunities from past operations remains in question, as some foresters familiar with operations (pers. comm.) have suggested that neighboring untreated areas were likely not treated for a reason, implying differences in site characteristics. This requires greater understanding of the robustness of the RM approach in instances where, at the stand level, it is not possible to locate homogenous areas with fertilized and unfertilized trees to sample.

The RM approach relies on the comparison of growth for periods of time before and after N application; thus, a mismatch in site series between plots can introduce error if trees from sites of different quality may exhibit distinct average trajectories of stemwood growth as stands develop. While site effects on cambial growth are not well known, empirical studies commonly identify responses to terrain variables of ecological classification [23,24,25], which underpin the emphasis on polymorphic specification of tree growth models [26,27,28].

The use of tree rings for studying environmental change is extensive; however, the extraction of the desired signal from unwanted variation can be challenging. Tree ring series can be thought of as an aggregation of several signals that are either signal (information relevant to the study) or noise (information that is irrelevant to the study) depending on the study hypothesis [29]. Thus, standardization (i.e., detrending) of ring width is commonly performed in dendrochronology to remove variance explained by factors such as tree age, and/or endogenous (i.e., internal to stand) or exogenous (i.e., external to the stand) disturbances, resulting a stationary tree ring index (at the population level) with a relatively constant variance [29]. The signal of interest is a key consideration to determine what detrending function will be used, as the signal in one study may be the noise of another. For example, a climatologist is most interested in the tree ring variability that is caused by climate fluctuations, representing the signal of interest, while for a forester, this variability may be unwanted variation if it conceals the treatment effect of interest [30,31]. To our knowledge, no studies in B.C. have investigated whether detrending can help to overcome the potential hazard of mismatches in site series masking signals from N application using the RM approach. The motivation of this study is to identify methods that may help isolate the response of tree growth to N application from all other sources of variability, which are assumed to be more extensive under imperfect (non-experimental) conditions. This is relevant because the difficulty of assessing the efficacy of fertilization treatments is compounded by the lack of controls in the operational context. We hypothesize that by selecting non-experimental stands and applying tree ring width detrending, we should be able to isolate the short-term effect of N application by reducing variation introduced by site and age differences.

To build on previous investigation of operational monitoring methods, this study implemented the RM approach at sites of a long-term forest fertilization trial and on carefully selected non-experimental sites in the region. The purpose of the study was to demonstrate the effectiveness of dendrochronology to evaluate tree growth responses to N application, and to determine if non-experimental plots could be used with equal effectiveness as the experimental controls in evaluating growth response to nutrient application across operations of the nutrient management program.

2. Data and Methods

2.1. Study Site

This study was conducted at eight sites within the previously established experimental project (EP) 703 on the central coast of Vancouver Island, B.C., Canada (Figure 1). The project was established in 1971–1972 to investigate the effect of intensive forest management practices (i.e., fertilization and thinning) on immature Douglas fir (Pseudotsuga menziesii var. menziessii (Mirb.) Franco) and western hemlock (Tsuga heterophylla (Raf.) Sarg.) growth. All sites were located in the very dry (xm) Coastal Western Hemlock biogeoclimatic zone (CWHxm; sensu Meidinger and Pojar [22]). Soils were classified mainly as Duric Humo-ferric Podzol or Degraded Dystric Brunisol; medium nutrient and mesic moisture regimes were most common in the sites [32]. At establishment, stand age ranged between 19 and 43 years old. By basal area, tree species composition consisted of 90% Douglas fir and 10% western hemlock. The average stand density was 2847 trees ha⁻¹. The average site index (at base age 50) was 32 m. The most recent measurements from 2015 to 2020 indicated that the quadratic mean diameter and stand density of the plots ranged from 19.7 to 36.3 cm (mean: 25.2 cm) and 700 to 2840 trees ha⁻¹ (mean: 1569 trees ha⁻¹), respectively.

2.2. Field and Laboratory Methods

Eight Douglas fir sites were selected from the EP703 for field sampling. The sites were designed as combined treatments of multiple thinning (0 to 35% basal area removal) and N (urea) fertilization (0 to 675 kg N ha⁻¹) levels [32,33]. During December 2020, field sampling was conducted at unthinned experimental plots to exclude the thinning effect, consisting of two unfertilized (control; C) plots and two fertilized (F; 225 kg N ha⁻¹) plots per installation. In addition, four “surrogate control” (SC) plots per site were established from the untreated forest areas (i.e., not part of experimental design) adjacent to EP703 sites. This resulted in sampling at a total of 64 plots (8 sites × (2 C plots + 2 F plots + 4 SC plots)).

Whereas all original C and F plots were located in zonal sites of the CWHxm subzone, SC plots were intentionally located in two adjacent stands with drier (SC_Dry) and wetter (SC_Wet) site classification (site series 03 and 05), respectively. (See Klassen et al. [21] for a description of site classification in B.C.) The SC samples were intended to test the effectiveness of the RM approach where experimental controls are unavailable. The results can provide insight into the feasibility of broader implementation of the RM approach under operational conditions where a mismatch in site series between the treatment area and control plots may be common.

From all plots (i.e., EP703 sites and SC plots), a single core was collected at breast height (1.3 m above the ground) from eight healthy dominant/co-dominant Douglas fir trees, except at site 5. At site 5, 16 trees were sampled per treatment, resulting in a total of 544 sampled trees (8 trees × 64 plots + 32 additional trees at site 5). The DBH and height of sampled trees were recorded in 2020 at the time of coring. The mean DBH and height of the sampled trees across the treatments were 37.6 cm and 30.3 m, respectively (

Table 1. Mean diameter at breast height (DBH) and height of sampled trees in this study.

Site	DBH (cm)				Height (m)
	F †	C	SCDry	SCWet	F	C	SCDry	SCWet
3	45.4	45.6	37.2	52.8	34.3	35.6	30.4	34.5
4	37.2	35.2	42.0	48.2	33.8	30.5	31.5	37.8
5	29.1	30.5	30.0	29.1	25.6	27.0	23.2	42.1
10	33.4	31.9	32.8	59.1	27.6	25.9	24.2	34.4
11	32.8	35.1	40.4	44.6	28.7	30.3	26.4	37.7
12	28.3	29.2	36.2	51.0	29.1	27.7	20.9	38.5
16	36.4	37.1	31.2	46.5	32.6	31.7	23.3	35.6
71	34.1	30.4	34.5	50.5	28.7	25.0	26.0	36.8
Mean	34.0	34.0	35.5	47.7	29.6	29.0	25.7	37.2

^† Notation: F: experimental fertilized; C: experimental control; SC_Dry: surrogate control (dry); SC_Wet: surrogate control (wet).

).

In the laboratory, increment cores were air-dried and mounted and then sanded with increasingly finer grades of sandpaper (100 to 330 grit) to reveal the annual ring structure (Stokes and Smiley, 1968). Tree cores were scanned with an Epson Perfection V850 Pro-flatbed scanner, and ring widths were measured to the nearest 0.01 mm using CooRecorder and visually cross-dated in C-Dendro [34,35]. Cross-dating was verified statistically using the software program COFECHA [36]. For cross-dating purposes, all trees sampled per site (C, F, and SC) were cross-dated together as a population. Of the original 544 tree cores, 462 were successfully cross-dated (85%) with an average inter-series correlation of 0.58 (

Table 2. Site conditions and cross-dating statistics of sampled trees. Number of dated series reflect the final number of trees included in analysis after removal of cores that could not be cross-dated with confidence. Climate variables express the 1971–2000 mean for the nearest 1-hectare grid cell in raster coverages from ClimateNA [37].

Site	Mean Annual Air Temperature (°C)	Mean Annual Precipitation Depth (mm yr−1)	Number of Dated Series	Average Series Length (Years)	Timespan, Calendar Years	Mean Inter-Series Correlation	Mean First-Order Autocorrelation
3	7.4	1759	45	79.9	1919–2020	0.50	0.84
4	10.2	1105	49	62.2	1937–2020	0.57	0.82
5	4.5	2505	72	62.5	1949–2020	0.61	0.82
10	6.2	2153	62	65.7	1947–2020	0.58	0.83
11	8.1	1530	57	61.4	1948–2020	0.61	0.83
12	7.9	1571	58	57.4	1955–2020	0.63	0.82
16	7.7	1474	59	57.5	1957–2020	0.60	0.81
71	8.3	1518	60	67.0	1949–2020	0.61	0.77

). The remaining 15% were excluded from subsequent analysis.

2.3. Estimating the Response to N Application

To evaluate growth response to N application, we first analyzed the raw (i.e., not detrended) tree ring widths (mm) at the C and F plots of the original experiment. As these plots were selected prior to N application with the intention of controlling for differences in initial stand conditions, this analysis reflects the idealized performance of the RM methodology through the pairing of control and treated plots.

Following Ballard and Majid [14], ratios that normalize the growth of fertilized ($f$) and unfertilized ($u$) trees were calculated as:

$$r\left(f\right)=\frac{{\overline{w}}_{After}\left(f\right)}{{\overline{w}}_{Before}\left(f\right)}$$

(1)

and

$$r\left(u\right)=\frac{{\overline{w}}_{After}\left(u\right)}{{\overline{w}}_{Before}\left(u\right)}$$

(2)

where w_After(f) and w_Before(f) are the mean ring width of fertilized trees for a defined period after and before application, and w_After(u) and w_Before(u) are the mean ring width of unfertilized trees for a period after and before application. Ballard and Majid [14] defined a standardized index in absolute terms, denoted here as follows:

$$I_a=r\left(f\right)-r\left(u\right)$$

(3)

In more general terms, the index can be expressed as:

$$I_{a,t}=\frac{w_t\left(f\right)}{{\overline{w}}_{t_{ref}}\left(f\right)}-\frac{w_t\left(u\right)}{{\overline{w}}_{t_{ref}}\left(u\right)}$$

(4)

where w_t(f) and w_t(u) indicate ring width of fertilized and unfertilized trees in calendar year t, and ${\overline{w}}_{tref}$ (f) and ${\overline{w}}_{tref}$(u) indicate mean ring width of fertilized and unfertilized trees during a specified reference period, respectively. As the units and magnitude of I_a,t are uncommon and unrelatable, we expressed the index as a percent difference:

$$I_{r,t}=100·I_{a,t}/\left(\frac{w_t\left(u\right)}{{\overline{w}}_{t_{ref}}\left(u\right)}\right)$$

(5)

Previous studies provide little guidance on how the reference period should be selected. Here, we defined w_tref(f) and w_tref(u) by the mean ring width over 5 years preceding application (1966–1970). The semi-continuous annual time series of the I_r,t was plotted to visualize the temporal response of ring width to N application.

To further summarize the findings, we reported the mean response over the first ten years after N application following Ballard and Majid [14] and later adopted by Brockley [15]:

$${\overline{\mathsf{\Delta }w}}_{1:10}={\overline{w}}_{1:10}\left(f\right)-{\overline{w}}_{t_{ref}}\left(f\right)·\left(\frac{{\overline{w}}_{1:10}\left(u\right)}{{\overline{w}}_{t_{ref}}\left(u\right)}\right)$$

(6)

where $\overline{w}$_1:10(f) and $\overline{w}$_1:10(u) are the mean ring width of fertilized and unfertilized trees for the ten years after application (1972–1981). Again, because the units of $\overline{∆w}$_1:10 were unrelatable, we also took advantage of the ability to apply the general form of Equation (5) to estimate the ten-year response as a percent (I_{r, 1:10}). Following the analysis of response from the original experimental plots, Equations (5) and (6) were also calculated using raw ring-width measurements from surrogate control (SC_Dry or SC_Wet) plots.

The evaluation process for the experimental controls was repeated using detrended tree ring series. We adopted a widely applied method that tested fits to each tree series with a linear or a negative exponential function of ring age using the R version 4.0.5 [38] package dplR version 1.7.6 [39]. It is important that detrending preserves any potential effect of N application in the treated plots. As the age series of each tree included at least 15 years prior to N application and 49 years after application, we assumed that the potential effects of N application on ring width in the treated plots would have negligible influence on the age function fit. Subsequent steps of calculating the fertilization response indices from the detrended residual chronologies were identical to the previous analysis.

In most experiments, there is greater reliance on quantifying responses from a comparison of growth measurements for periods after application (i.e., w_After). Sampling must, therefore, overcome high variance in growth among trees and plots, Var(w_After). The RM approach is predicated on the assumption that variance in r(u) is lower than that of w_After, which should improve the signal-to-noise ratio. The hypothesis that Var(r(u)) < Var(w_After) was tested to give insight into the additional value of pre-application sampling attained by the RM approach. The value Var(r(u)) was approximated by the coefficient of variation in r(u) (the last term of Equation (6)). The value Var(w_After) was approximated by randomly selecting 10 pairs of trees from experimental C plots at each installation, calculating the percent difference in mean ring width (of trees in each pair) for the ten-year period following application, and then calculating the mean absolute value of the percent differences in randomly paired trees from all eight sites.

3. Results

The age responses of ring width at the original control and treatment plots of the EP703 experiment were similar, although average ring widths tended to be slightly greater at experimental control plots over ages ranging from 5 to 20 years old (Figure 2a,b). Ring width was comparable between the original plots of the experiment and dry SC plots (Figure 2c), yet observations at dry SC sites showed signs of a slightly different pattern, with a lower average ring width from ages 5 to 10 years and a greater average ring width over ages from 18 to 40 years. The average ring width at wet SC sites was similar to that of other site types during the first five years of development, but then showed substantially less decline with age (Figure 2d). Visual assessment indicated that the negative exponential age response functions successfully captured age–size variability among plot types. While negative exponential functions tended not to fully capture the initial period of rapid increase in cambial growth, analysis of ages indicated that 100% of the sampled ring width observations (in the 10 years following N application) corresponded with ages greater than 17 years old.

Using unfertilized trees across all experimental sites, the mean r(u) was 0.81, and the standard deviation was 0.23. The coefficient of variation was, therefore, 28% (i.e., Var(r(u)) ≈ 28%). With a sample size of 462 trees, the standard error was ≈0.01. As defined by randomly pairing trees at unfertilized plots, we found that the mean absolute percent difference in growth (w_After) was 45%. With a difference in variation between the two variables of 17% (45% vs. 28%), the analysis supports the underlying assumption of the RM approach, that explanatory power can be gained by applying the RM approach because Var(r(u)) < Var(w_After).

The response of ring width to N application was visible from the time series of the relative response index calculated from Equation (5) (Figure 3). The response was greatest in years 2 and 3 after application, peaking at 34% three years after application. The index then rapidly decreased, consistent with modelled responses to N application [10]. The ten-year average value of I_r from this analysis was 16.0%. The time series indicated a small positive relative response extending well beyond the 10-year test period, except for multi-year episodes in 1988–1992, 2001–2003 and 2018–2020, when the difference was far less and sometimes negative.

Adopting Equation (6) outlined by Ballard and Majid [14], estimates of Δw_t1:10 were positive at seven of the eight sites (Figure 4). The mean value for all sites was Δw_t1:10 = 0.20 mm yr⁻¹. Individual sites showed substantial variability, with values of Δw_t1:10 ranging between −0.17 mm yr⁻¹ at site 4 and 0.39 mm yr⁻¹ at site 10.

Analysis of the response ratios showed that it was normal for w_After < w_Before irrespective of N application, as indicated by values of r(f) and r(u) less than 1.0 (Figure 5a), a consequence of the universal decline in ring width with age for all plot types (Figure 2). There were two exceptions at the plot level, including fertilized trees at site 3 and unfertilized trees at site 4. The ten-year response in the relative response index showed a significant positive average response to fertilization (Figure 5b), consistent with the previous analysis of the time series and values of Δw_t1:10. The variation in relative responses among sites differed slightly from that of Δw_t1:10 (in absolute units of the variable). Significant relative responses were detected at seven of eight sites; no response was found for site 4 (additionally see Figure 4). While pooling the time series of I_r yielded a ten-year response of 16.0% (Figure 3), stratifying the comparisons by site and then taking the average led to a mean ten-year response of 15.0% (“All” in Figure 5b), suggesting that the response estimates were slightly sensitive to differences in the number of cross-dated samples per site (Table 2).

The use of surrogate control plots to test the feasibility of this approach for operational fertilization applications had important effects on the time series of the response index (Figure 6a). Low-frequency variability differed substantially between zonal (experimental controls) sites and wet (SC_Wet) sites, as growth at wet sites showed a substantial decline after 1976. The response index using dry (SC_Dry) plots was more consistent with the experimental control but still deviated beyond a standard deviation of the response index for extended periods (e.g., 1979–1985). Consistent with the evaluation of age responses (Figure 2), detrending effectively removed differences in age–size variation in the response indices among types of controls (Figure 6b). A similar response to N application is clearly visible from the analysis of all control types.

Without detrending, the surrogate control plots indicated lower responses to N application, with 8.5% and 9.1% growth response in the SC_wet and SC_dry, respectively (Figure 7a). In addition to substantial decreases in the best estimates of the growth response, the use of the surrogate controls failed to establish significance in response to N application at the 95% confidence level (Figure 7a). With detrending, the responses to N application using the experimental controls decreased from 15.0% to 9.6% (Figure 7b). Differences in estimates among the control types were also eliminated following detrending, and the response for all control types was significant (Figure 7b).

4. Discussion

This study of cored trees at sites of the EP703 experiment followed several characteristics of response to N application identified by previous studies. The average stemwood response of coastal Douglas fir was significantly positive, consistent with previous analysis of the EP703 experiment [33] and other sites [8,12,13,40,41]. However, there was one exception: one of eight sites (site 4) showed a weak negative response (Figure 5). This finding is similar to studies that report infrequent negative responders and non-responders [11,42,43,44]. After the peak response two years after application, annual growth response decreased gradually in a non-linear fashion (Figure 3), consistent with the previous observation and modelled responses to N application [10].

Testing responses from operations is important because operational conditions can differ from experiments in how fertilizer is applied, in weather conditions during application, and in the species composition and stand conditions of the treatment area. While the RM approach yielded reliable information about the growth response of dominant trees, the time and costs associated with plot establishment and plot pairing were avoided by applying the RM approach, making it an attractive option for broad application across operations. Yet, the time and cost need to be weighed against the quality of the information produced, including the variables that are observed and the precision and accuracy of the estimates.

The RM approach only describes the response of cambial growth for dominant trees, while paired-plot experiments are based on repeated censuses of vital status, DBH and height, and can, therefore, observe variables that are more directly relevant to performance measurement, including the growth of survivors, recruitment and mortality at the stand level. Nevertheless, the ability to verify significant positive responses in the cambial growth of dominant trees, like screening trials [45], can play a role in confirming the presence, if not magnitude, of significant positive responses under operational conditions. While the relative response indices offered by the RM approach are not representative of relative responses in net volume (or biomass) increment at the stand level, benchmarks could be established based on a standardized application of dendrochronology methods. Any future reliance on dendrochronology could benefit from additional knowledge about the relationship between the responses of DBH and height. For instance, if the RM approach indicates a significant response in cambial growth, is it safe to assume that the height increment also responded similarly, or is there evidence of common divergence in the response of cambial and apical growth? Likewise, the RM approach could benefit from a better understanding of the relationship between the responses of dominant tree growth and stand mortality. That is, if one finds a significant response in the cambial growth of dominant trees, does that correspond with a predictable response in mortality?

Since this study focused on management goals, we intentionally tried to control for factors other than N application. The implications for ecology may, therefore, be limited. However, the dendrochronological approach used in this study offers insight into methodologies to detect the responses to ecological events which are difficult to establish formal controls, such as wildfire and drought. Our hope is that it encourages others in forest ecology to think about the interesting and complex trade-offs in the information content of individual-tree data (with strong temporal sampling both before and after treatment) vs. paired fixed-plot data (with strong spatial sampling).

With regard to precision and accuracy, our study revisited the subject of trade-offs in the explanatory power of the RM approach vs. paired-plot experiments. The RM approach is predicated on the idea that the inclusion of sampling pre-application growth increases the explanatory power of the analysis. The present study provides quantitative support for the premise (when applied using plots that were carefully paired during establishment of the original experiment). Paired-plot experimental designs aim to estimate the response to N application by comparing w_After(f) with w_After(u), and must, therefore, overcome high spatial (i.e., among-plot) variation in tree growth. Values of r(f) in the RM approach, however, cannot be taken as a reliable indicator of response to N application due to inter-annual variation caused by weather and biotic agents, and potential lower-frequency variability due to changes in the age, ontogeny and competitive status of trees [14,16]. The RM approach may, therefore, be equally dependent on effective pairing with control plots. Again, desiring to balance quality of information with time and cost, more studies could be designed to understand the trade-off in quantitative terms, and specifically how much explanatory power gained from the inclusion of pre-application observations compensates for the potential increase in among-plot variation that might arise from the post facto selection of control plots when applying the RM approach across operations.

Our study provides further clarity into how mismatches in stand conditions can affect the estimation of responses to nutrient application. Mismatches in ecological site classification between the treated and untreated plots affected the results because the growth of dominant and co-dominant trees growing on wetter sites exhibited a distinct trajectory in ring width as forest stands developed (Figure 2). The smaller difference between the ring widths of zonal and SC_Dry plots may have reflected the greater fertility of the drier control plots (site series 03) found within the vicinity of the EP703 sites when compared to zonal sites (site series 01). That is, slightly greater fertility may have compensated for the lower moisture of the SC_Dry plots.

A consideration of site conditions in the analysis of tree growth is rare, yet the identification of differences in the relationship between age and ring width between zonal and wet sites in this study demonstrated value in using ecological site classification systems to represent the effects of local soil fertility and moisture on tree growth. In addition to potential differences in the trajectory of ring width, differences in site factors between fertilized and unfertilized plots may amplify variance in ratios, Var(r(u)), if interactions exist between climate anomalies and site factors [14]. Testing the latter effects requires meticulous efforts to differentiate variability in initial stand conditions (i.e., stem density) from site nutrient balance and water balance. While differences in the conditions of plots were known for the time of coring, the study design did not permit measurements of stand attributes of the surrogate control plots at the time of N application.

Detrending low-frequency variability in ring width from each tree showed promise in mitigating the undesirable effects of mismatches in site classification between treated and untreated samples. However, detrending affected the magnitude of the relative response tested under ideal conditions (i.e., experimental control and treatment plots), decreasing it from 15% to 10% (Figure 7). The negative exponential function was only correcting for variation explained by age and size. The residual time series should have preserved variability associated with climate, the presence of biotic factors (e.g., root rot), and N application. Without a better understanding of how commonly detrending affects response estimates based on the careful selection of experimental controls, we argue that operational monitoring with dendrochronology prioritizes application in instances where site conditions (site series and initial density and stand age) for control and treatment plots are matched. However, the fact that detrending the surrogate controls led to greater coherence and statistical significance (Figure 7b) implies that detrending may be valuable if our understanding of the reasons for alteration in average response can improve. Overall, we did not anticipate a difference in growth response between raw and detrended time series, and thus, exploring other detrending functions (e.g., cubic smoothing splines) warrants consideration.

While the RM approach cannot be applied in instances where fertilization occurred close to the time of stand establishment, the average stand age at the time of application in B.C. is 35 years, similar to the present study. The RM approach could, therefore, be applied broadly to most treatment areas. However, the capacity to perform the detrending in this study was aided by the fact that sampling was conducted well (49 years) after N application. We propose that the effectiveness of detrending to remove background variability might be compromised if applied to circumstances where sampling was performed shortly (e.g., 10 years) after N application. More generally, the approach likely deteriorates with a decreasing range of variability in age. This emphasizes the need for the development of statistical models that simultaneously consider the effects of tree age, N application dose, and time since N application as explanatory factors (e.g., Jang et al. [10]).

Paired-plot experiments have established that the average response of some major commercial tree species in B.C. is significantly positive. Directly evaluating responses from operational applications can build confidence that experiments are representative of operations. There are additional needs that may be fulfilled by the RM approach, including evaluating responses to repeated N applications, the optimal stand age at the time of application, interactions between application and disturbances (e.g., drought, non-lethal fire), and evaluation of responses in pilot test conditions outside the boundaries of traditional stand selection criteria.

The use of dendrochronology is an effective strategy for these applications when comparable control plots are present, while the present study presents insights into how the method can break down in the absence of effective controls or extensive efforts (through detrending or modelling) to control for among-plot variability in tree growth. If the sole purpose is to evaluate the efficacy of treatment, time and cost may be reduced by developing procedures that could eliminate reliance on extensive lab work (e.g., cross-dating). Indeed, dendrochronology was successfully applied based on a visual inspection of cores, just coring deep enough to span the time of N application (B.C. Forest Investment Program pers. comm.). However, coring to the pith and collecting measurements of DBH and height at the time of coring permit the reconstruction of tree size, which may provide more accurate responses and may promote the application of dendrochronology to serve multiple research purposes in forest management beyond just responses to N application.