You are on page 1of 8

Published May, 1982

Design of Experiments in Growth Chambers -- Uniformity Trials in


Ithe North Carolina State University Phytotron
2Chong-soon Lee and J. O. Rawlings

ABSTRACT experiment requires several chambers or at least several


Experimentsconductedin growthchambersfrequently are defi- runs over time in one chamber. The former will be
cient with respect to one or moreaspects of experimentaldesign, referred to as within-chamber experiments; the latter, as
primarily, properuse of randomization and replication. To provide between-chamber experiments. If both levels of experi-
informationwith whichto counter the usual argumentsassociated
withthe failure to use properdesigns andto improvethe efficiency mentation are involved in a single study, the arrangement
of growthchamber experiments,extensive uniformitytrials (using is the classical split-plot experiment with the individual
soybeans) were conductedin the NorthCarolinaState University chambers being the whole-plot experimental units and the
Phytrotronin three sizes of chambers.
Theresults confirmedthe existence of sizeable, patternedplant- individual plants, or sets of plants as the case may be,
to-plant variation within chambersand amongchambersof all within chambers being the sub-plot experimental units.
sizes. Patternsof plant growthwithin chambers tendedto be char- Two fundamental experimental design concepts, repli-
acteristic of chamber type and to persist over time. Bothchamber cation and randomization, are frequently ignored in phy-
and timeeffects werevery important.
Correlationsamongspatially related observationsconfirmedthe totron experiments. In some cases they are ignored
potential for seriousbias in estimatesof both treatmenteffects and because of a commonmisconception that the highly tech-
experimentalerror in experimentswhichignore either randomiza- nical control of the growth chamber environments insures
tion or properreplication. Properlydesignedexperimentsare as
importantin growth chambersas in greenhouseor field experi- uniform conditions, at least for all practical purposes, both
ments. Properblocking both within chambers(for within-chamber within and among chambers (with the same nominal set-
experiments) and with respect to time and/or chambers(for tings) and, therefore, placement of the plants and assign-
between-chamber experiments)appearsto be effective in control-
ling a majorportionof the variation. ment of treatments is not that critical. In other cases, the
problem arises because of misconception of what consti-
Additional index words:Experimentaldesign, Uniformitytrials, tutes replication in the experimental design context. Both
Randomization,Truereplication, Growthchamberstudies, Intra- proper replication and proper use of randomization
class correlations.
require a clear understanding of what constitutes an
experimental unit. Absence of replication means that the
W HILE controlled-environment growth chambers have
proven to be a valuable asset to plant research, there
experiment does not contain the information needed to
estimate the experimental error; absence of randomiza-
is a continuing tendency for some researchers to ignore
tion means that the estimate of experimental error, even if
good experimental design practices in growth chamber
there is proper replication, may be seriously biased.
experiments. Discussion of design of experiments within
growth chambers must recognize two levels of experimen- Several authors have issued warnings that growth
chamber environments are not spatially uniform.
tation in terms of what constitutes an experimental unit.
On the one hand, there is the class of treatments which Hruschka and Koch (1964-) reported "strong gradients"
may be applied to individual pots, or sets of contiguous in growth in all directions in a temperature and humidity
pots, within each chamber; e.g., levels of various controlled storage room and warned, "... heterogeneity
nutrients, moisture levels, cultivars, or plant densities. In is inevitable in environmentally controlled experimental
this case, the individual pot, or set of pots, is the experl- chambers .... " Collip and Acock (1967), using lettuce,
mental unit and each chamber will contain many such and Hammer and Langhans (1972), using maize, both
experimental units. On the other hand, there is the class reported large patterned within-chamber variation.
of treatments, e.g., irradiance levels, day/night tempera- Measures et al., (1973) presented results from cucumbers
ture regimes, or CO2levels, which by their nature require (Cucumis sativus L.), corn (Zea mays L.), oats (Avena sativa
that all plants within the chamber be exposed to the same L.), and wheat (Triticurn aestivum L.) discounting the con-
treatment. Hence, the chamber is the experimental unit cept of environmental homogeneity within growth cham-
for these treatments and even a single replicate of an bers. They indicated that heterogeneity within chambers
was so great that choice of position within the chambers
could critically bias the results. All of these authors rec-
’Paper No. 7044 of the Journal Series of the North Carolina ARS
Raleigh,N.C.This paper presents part of the results of a Ph.D.Diss. ommended reliance on randomization, replication, and
preparedby the senior author while enrolled at NorthCarolina State blocking to increase precision and eliminate bias. With
Univ.Theauthors wishto thankDr. R. J. Downs,director, NorthCar- respect to heterogeneity between chambers, Hammer et
olina State Univ. Phytotron,Dr. V. P. Bonaminio,extension assistant
professor of horticultural science, NorthCarolinaState Univ. and the al., (I 978) reported large experimental variation over time
phytotron staff for extensiveassistance andcooperationduringthe con- within laboratories and emphasized the necessity of rep-
duct of these studies. Received10 Aug.1981. licating growth chamber experiments over time to obtain
~Research assistant, presentlysenior researchofficer, MalaysianAgri-
cultural Researchand Development Institute, Serdang,Selangor, West a proper estimate of experimental error.
Malaysia, and professor of statistics, NorthCarolina State Univ., The lack of spatial uniformity within growth chambers
Raleigh, N.C. of the physical environment, primarily temperature and

55~
552 CROPSCIENCE, VOL. 22, MAY-JUNE
I982

radiance, has been reported by many authors; e.g., and experimental error must be recognized. Of even more
Measures et al., (1973), Tibbitts et al., (1976), Hammer potential seriousness in between-chamber experiments is
and Langhans (1972). Knievel (1973) and Tibbitts et the very strong desire, for obvious economic reasons, to
(1977) reported on other environmental factors which use only one chamber (one experimental unit) for each
might be contributing to environmental variation within treatment and to regard plants within chambers as proper
and between chambers. replication for the estimation of experimental error. This
Hammer and Langhans (1978) reiterate the impor- is analogous to the previous situation of using plants
tance of using proper replication and experimental designs within trucks as an estimate of experimental error and
in an introductory manual on the use of controlled envi- provides the potential for gross underestimation of the.
ronment growth chambers, and Hammer and Urquhart experimental error.
(1979) present a more complete discussion of the applica- Of course, one does not expect complete uniformity of
tions of experimental design to growth chamber experi- plant development even under highly controlled environ-
ments. In spite of the many references to lack of mental conditions. The critical point, however, in the
uniformity within and between chambers and in spite of a event randomization has not been used is whether the
general acceptance of the need for proper experimental variation is random or patterned in some way. If the var-
design in both field and greenhouse experiments, there iation is nonrandom, then the magnitude of the variation
continue to be rather frequent occurrences of failure to use will affect the biases in the estimation of treatment effects
proper experimental design in controlled environment and experimental error. The discussion in Cochran and
experiments. Cox (1957) of randomization as an insurance "against
In the context of the within-chamber experiment, the disturbances that may or may not occur and that may or
experimental unit implicitly includes the space within the may not be serious if they do occur" is relevant but is, in
chamber assigned to the pot or set of pots. It is not the authors’ opinions, an understatement of the serious-
uncommon, for example, to allocate one treatment to all ness of lack of randomization.
the plants on one truck, another treatment to all plants on To quantify the seriousness of the failure to make proper
a second truck, and to place the trucks wherever space use of randomization and true replication in growth
permits within the given chamber. In the event environ- chamber studies and to obtain basic information relevant
mental conditions are not uniform, it has long been rec- to the design of experiments in growth chambers, a series
ognized that such procedures introduce unknown biases of uniformity studies was run in the North Carolina State
in the estimation of treatment effects and an underesti- University Phytotron (Lee, 1977). The objective was
mation of experimental error, if variation among plants on assess the magnitude and nature of the within-chamber
a truck is used as the error. Even if the growth chamber and between-chamber variation under uniform condi-
provides a perfectly uniform environment, the mere tions in the three sizes of growth chambers in the NCSU
grouping together of all plants for a given treatment and Phytotron. With this information, the magnitude of the
the inescapable tendency to handle the group as a unit, bias from misspecification of experimental error and fail-
such as in thinning, watering, and data collection will tend ur.e to use randomization can be ascertained and efficient
to introduce positive correlations among the plants within experimental designs identified.
a treatment with the net result that variation amongplants
on a truck will underestimate experimental error. The MATERIALS AND METHODS
truck or group of plants, in effect, becomes the experi- Since the objective was to detect what, a priori, were expected
mental unit so that the experiment has no true replica- to be relatively minor environmental variations, it was decided
tion. Variation among plants within treatments is really that a relatively sensitive plant species would be used in a tem-
sampling variation within experimental units. perature regime that would favor detection of environmental
The same problem occurs, and is probably much more variation. Soybeans (Glycine max.), ’Bragg’, was chosen as the
serious in nature, with between-chamber experiments. test species and 35 of the 47 uniformity studies were conducted
Because of the nature of the treatment, all plants within a with a day/night temperature regime of 22/18 C. To check the
effect of temperature, the other 12 studies were conductedat 26/
chamber must be subject to the same treatment. It is com-
22 C. The distribution of the studies over the three sizes of
mon practice with between-chamber experiments for a chambers (A, B, and C) in the North Carolina State University
series of standard environments involving day/night tem- Phytotron, the dates and the temperatures at which they were
perature and light regimes to be provided on a continuous done are shown in Table 1.
basis by designated chambers, with individual researchers All seeds were pregerminated. The seedlings were selected for
using the chambers providing the treatments of interest uniformity and transplanted three to each pot. At 7 days of
when space permits. Operationally, this is much more growth the two extreme plants were eliminated leaving one plant
convenient than allowing random assignment of treat- per pot. All studies were terminated at 20 days and the following
ments to chambers and it is convenient to leave unchanged five measurementstaken on each plant:
1. plant height (ram) from the cotyledonary node to the tip
over relatively long periods the assignment of standard
the apical bud,
environments to chambers. In such case, if replication in 2. total leaf area (era~) of all fully opehedleaves,
time is used, it means repeated use of the same growth 3. petiole length (ram) of the first trifoliate leaf,
chamber for a given treatment. As before, the failure to 4. fresh weight (rag) and,
use randomization in the assignment of treatments to 5. dry weight (rag) of the plant above the cotyledonary node.
experimental units, the chambers, and the associated Each A chamber (3.66 × 2.44 m) held 24 movable trucks in
potential for biases in the estimation of treatment effects 6 × 4 configuration; each truck held 16 plants in a 4 × 4 config-
LEE & RAWLINGS:DESIGNOF EXPERIMENTS
IN GROWTH
CHAMBERS 553

Table 1. Dates, chambers,and temperatures of the uniformity Table 2. Means,standarddeviations {within studies), and coef-
studies with soybeans in the NCSUPhytotron. ficients of variationaveragedover all studies for a given cham-
ber type.
Trial Date{1975} Chamber
Variable
A Chambers
1 7/8 -7/30 A8 A9 A10 Plant Leaf Petiole Fresh Dry
2 7/29-8/20 A4 A9 A10 ht. area length wt. wt.
3 8/19-9/10 A4 A8 A9 mm cm* mm mg mg
4 9/9 -10/1 A4 A8 A9
Achambers
Ill studiesat 22/18CI~
B Chambers~f
1 3/14-4/3 B2 B3 B8 Mean 77 133 27 318 49
4/26-5/16 B1 SD 12 24 5.5 53 8.5
3 B2~: B85 C.V. 16 18 20 17 17
4 5/17-6/6 B2 B3 B8
5 6/7 -6/27 B1 B3 B8 B chambers
(12studiesat 22/18C)
6 6/28-7/18 B1 B2 B3§ Mean 119 155 39 386 54
C Chambers SD 27 33 10.3 86 10.9
C12 C15 C.V. 23, 22 27 22 20
1 5/31-6/20 C3:]: C8:~ C9
2 6/21-7/11 C3 C8 C9:~ C12~: C15~/ Cchambers
(10 studiesat 22/18CI
3 7/12-8/1 C3~t C8 C9 C12~t C15 Mean 99 110 24 254 37
4 8/2 -8/22 C3 C8:~ cg~t C12 C15~ SD 17 23 4.6 46 6.2
C.V. 17 21 19 18 17
Trial 2 wasdiscardeddueto atypicalplant growthin all three chambers.
Studies conductedat 26/22C. All otherstudieswereconducted at 22/18 (10studies
C chambers at26/22 C)
C. All pooledresultsincludeonlytrials conducted
at 22/18C. Mean 295 257 76 565 75
A faulty temperature controlrequiredmovement of the plants to cham- SD 40 32 8.3 74 9.8
ber A15for 3 days.Datafromthis trial are not includedin the combined C.V. 14 12 11 13 13
analyses.
Malfunction
of the light controlin chamber
A8in trial 3 wasrecorded
for
threeconsecutivedays.Datafromthis trial areexcluded.
uration giving a total of 384 plants per A chamber, 16 columns
and 24 rows of plants. Each B chamber (1.22 × 2.44 m) held Alternatively, large intraclass correlations identify classifica- "
trucks in a 2 × 4 configuration, 16 plants each, giving a total of tions that would constitute effective blocks for experimental
128 plants per B chamber, 16 columnsand eight rows of plants. design purposes, assuming treatments are randomly assigned to
Each C chamber (0.91 × 1.22 m) held 15 racks in a 3 × 5 con- individual plants within the classification. Therelative efficiency
figuration, four plants per rack, giving 60 plants per C chamber, of the blocked design to a completely randomdesign is (approx-
10 columnsand six rows of plants. For a detailed description of imately) 1/(1-~), ignoring adjustments for loss of degrees
the North Carolina State University Phytotron facility, see freedom, where~ is the appropriate intraclass correlation.
Downs and Bonaminio (1976). As a second test for randomnessof variation within chambers
All analyses of the data were done by chamber type, and and to characterize the pattern of the variation, polynomial
within temperature levels for the C chambers. Nested analyses response surfaces were fit to the data for each plant variable in
of variance were used to test and quantify the importance of dif- each study using row and columncodes as the coordinates for the
ferences amongstudies (= chamber-trial combinations), posi- response surface. Similarities of the response surfaces over
tions in studies, and plants in positions, wherepositions were chambersand trials were investigated.
defined, in turn, by carts, rows, or columnswithin chambers.If
the variation in plant growth is random,all meansquares will be RESULTS AND DISCUSSION
estimating the same variance and should be approximately
equal. Equivalently, all intraclass correlations for different clas- The means, standard deviations (within study), and
sifications of spatially (or temporarily) related plants should coefficients of variation, based on the pooled within-study
zero. The results of the analysis of variance are summarizedin individual plant variation, averaged over all studies in a
terms of different types of intraclass correlations on the bases of given chamber type arc shown in Table 2. Plant-to-plant
within-chamber variation, relevant to within-chamber experi- variation within chambers was large in all experiments.
ments, and on the basis of total variation over studies, relevant Average coefficients of variation at 22/18 C ranged from
to between-chamberexperiments. The intraclass correlations is 16 to 27 %. In terms of extremes in plant performance, the
the average pairwise covariance between observations within
classes relative to the total variance in the reference base. Sig- largest measurement for any variable in a study was two
to three times the smallest. The average coefficients of
nificance of intraclass correlations (different from zero)
inferred from variance ratio tests in the appropriate analyses of variation for the higher temperature wcrc smaller but still
variance. For simplicity of presentation, and because of the sim- ranged from 11 to 14%. Inspection of the variance showed
ilarity of results for all plant variables, all intraclass correlations significant heterogeneity of variances associated with the
are presented as averages over all plant variables and all studies mean level of performance. Amongseveral transforma-
within chamber type. tions tried, the square root transformation was the most
The intraclass correlations are useful for two purpsoes. The consistent in stabilizing the variance and all subsequent
impact of misusing a within experimental unit variance as analyses arc in terms of square root transformed varia-
experimental error, rather than variance among experimental bles.
units, is quantified by the "error bias ratio" = 1 + (n-1)~/
The analyses of variance, using trucks to identify posi-
(1-~), where ~ is an estimate of the appropriate intraclass cor-
relation and n is the numberof plants per experimental unit. The tions in chambers (Table 3) shows a very significantly
error bias ratio estimates the size of the F-ratio for the test of increasing hierarchy of mean squares for every variable in
treatment effects under the null hypothesis of no treatment effects every chamber type. Clearly, all components of variance
whenthe incorrect within experimental unit variance is used as are greater than zero (or the intraclass correlations appro-
experimental error. priate to this nested system are greater than zero).
554 CROPSCIENCE, VOL. 22, MAY-JUNEI982

Table 3. Meansquares in the nested analyses of variance over Table 5. Coefficients of determination, averaged over all
experimentsfor each chambertype using trucks to define posi- studies, of polynomial response surfaces fit to uniformity
tions within chambers. Analyses are on square root trans- data.
formeddata.
Chamber
Degree of Plant Leaf Petiole Fresh Dry
Variable type polynomial ht. area length wt. wt.
Plant Leaf Petiole Fresh Dry A 6th 0.51 0.46 0.57 0.42 0.46
Source~ df ht. area length wt. wt. B 4th 0.70 0.52 0.69 0.53 0.51
C 4th 0.64 0.48 0.53 0.46 0.42
A chambers
Amongexperiments 10 93.7 236.0 48.9 384.2 43.8
Amongtrucks in
experiments 253 3.6 6.9 2.1 11.9 2.2 the A chambers. (All plant variables showed similar pat-
Amongplants in terns.) Note in particular the pattern characteristic to each
trucks 3960 0.3 0.7 0.2 1.6 0.2
chamber. A4 tends to have a pattern of its own with peaks
B chambers (P) near the left rear and right front of the chamber.
Amongexperiments 11 153.7 204.3 45.6 525.8 43.3
Amongtrucks in has a similar pattern to A4 but the valley displaced closer
experiments 84 14.4 14.9 6.7 36.6 4.0 to the right front. A9 and A10 patterns appear to be
Amongplants in almost mirror images of A8 and A4. Similar consistency
trucks 1440 0.7 1.2 0.3 2.8 0.4
of patterns over time is exhibited by the other variables.
C chambers
Among experiments Figure 2 shows the response surfaces for plant height for
in temperatures 18 33.2 33.1 14.2 96.5 9.6 the B chambers. In this case the major features of the con-
Among racks in tour surface seem to be characteristic of all chambers.
experiments 280 2.7 2.4 0.5 4.6 0.5
Amongplants in Closer examinaton suggests some chamber-to-chamber
racks 900 0.6 0.7 0.1 1.7 0.2 differences. Particular attention should be paid to B2 and
In the absenceof experiment andtruckeffects andwithindependent B8 in trial 3, where the pattern holds up very well at the
errors, the threemeansquareswithineachanalysishavethe sameex- 26/22 C temperature regime, and to B3 in trial 6, where
pectation.In a correlatederrorsmodel,withQ, beingthe correlationbe- the pattern is quite different, possibly disrupted by the
tweenplants on the sametruckandQ2beingthe correlationbetween
plantsondifferenttrucksbut in the samechamber, the meansquareex- movement of all plants to A15 for 3 days when the tem-
pectationsarea~[l+ (n-I)~i + n{k-I)~i],
a~[l+ {n-I)~i], ando~{I ~) perature control in B3 malfunctioned.
respectively, where n = number ofplants pertruck andk = number of
trucks perchamber. The response surfaces in the C chambers, (not shown)
fail to show the same high degree of pattern consistency
Table4. Estimatedintraclass correlations for various classifica- over time. There is suggestion of a tendency for the peak
tions within chambers,averagedover five plant variables. to be near the left-rear of the chamber but the patterns
shift considerably both with chambers and time (and with
Chamber
type
temperature).
Classification A B C These two methods of summarization clearly show that
Trucks(or racks) 0.37 0.47’ 0.36 the variation within chambers is not random; that as much
Rows 0.02~ 0.05 0.08 as 50 to 70% of the plant-to-plant variation within a
Columns 0.17 0.43 0.21
chamber can be ascribed to position differences within
Two of the five correlations
averagedto give this valueweresignificant-
ly differentfromzero,p = 0.05. In all othercases,all intraclasscorre- chambers. These results are surprising only in their mag-
lationsweresignificantlydifferentfromzero. nitude. They verify that the basic principle in field exper-
iments, that experimental units adjacent in time or space
Considering first the within chamber variation, the tend to be positively correlated, extends to the highly con-
large intraclass correlations (Table 4) for trucks (racks trolled environments in growth chambers. Hence, failure
the C chambers) and for columns clearly indicate nonran- to utilize proper design techniques for within-chamber
domness of the variation within chambers. The smaller experiments will lead to biased estimation of treatment
intraclass correlations for rows simply reflect the fact that effects and underestimation of the true error.
the individual rows span most of the variation within each The seriousness of failing to use proper design princi-
chamber, since we know from the other correlations that ples is demonstrated by noting that the error bias ratio
there are strong patterns to the variation within cham- resulting from using the within truck variance as an esti-
bers. mate of error, when the truck, due to lack of randomiza-
The polynomial response surfaces fit to each plant var- tion, is the experimental unit, is 10, 14, and 3 for the A,
iable in each study confirm that there is strong patterning B, and C chambers, respectively (computed from Table
in the plant-to-plant variation within chambers. In all 4). It is also clear that blocking by trucks or columns
cases, many terms of the polynomials were significant. A would be very efficient, relative efficiency, 1/(1-~), var-
sixth degree polynomial seemed to adequately character- ies from 1.56 to 1.89 for trucks and from 1.20 to 1.75 for
ize the surfaces in the A chambers, a fourth degree poly- columns as blocks. Randomization is imperative and
nomial in the B and C chambers. The coefficients of blocking within chambers would be beneficial. In addi-
determination (R2’s) for these models, averaged over tion, to the extent that the response surface in a chamber
studies, varied from 0.42 to 0.70 (Table 5). is consistent over time, prior information on the response
Inspection of the response surfaces reveals a fairly high surface would be a potentially useful covariate for control
consistency of chamber patterns over time. Figure 1 shows of within-chamber variation.
contours of the fitted response surface for plant height in The preceding has been concerned primarily with the
LEE & RAWLINGS: DESIGN OF EXPERIMENTS IN GROWTH CHAMBERS ,~,~.~

Trial I

^8

,I. "",,.’’"’1’... ~"~’:~~"’~


Fig. 1. Gontours of individual experiments, arranged by trial, of a sixth degree response surface for square root of plant height. Each contour
covers an interval t~ units. Chamber identification
of 0.5 mm is shown at the left. Trial identification is shown at the top. Chamber A8 in
trial ~ with a defective light control is shown with ?. The bold-face letters P and V on each diagram show locations of peaks and valleys in
the contour surface.

effects of nonrandom within-chamber variations on ments was partitioned into that due to chambersand trials
within-chamber experimental designs. It is clear from M1 (or time) differences and the "amongtrucks" variation
the nested anMysesof variance, see Table 3, that the posi- was partitioned into average position effects and interac-
tive intraclass correlations of observations related spa- tions of positions with chambersand with trims (to show
tially (and temporarily) extend across chambersand trims; the consistencyor lack of consistency of the position effects
the among experiments mean squares are much larger over chamber and trims).
than the within experiments mean squares. First, with respect to the anMysisof position effects, the
The intraclass correlations (Table 6) are far from zero following conclusions might be made:
at MIlevels and, consequently, the use of any within-clas- 1. In all chambertypes the position maineffects (trucks
sification (within-chamber, within-truck) sampling vari- or racks) were highly significant indicating at least
ance as the experimental error grossly underestimates the some consistency of the pattern over trials and
true experimentMerror. For example, the average intra- chambers.
chambercorrelations were 0.31, 0.42, and 0.38 in the A, B, 9. In the A chambers, the chamber × truck interaction
and C chambers, respectively. Using these averages, and meansquares were nearly as large as the truck main
assuming full chambers for each treatment in an inter- effects mean squares indicating that the position
chamber experiment, the error bias ratios from inappro- effects are quite different from chamberto chamber
priately using a within chambervariance as experimentM but, were reasonably consistent over time for a given.
error would be 175, 91, and 37, respectively (Table 7). chamber. That is, the within-chamber pattern of
The other lines in Table 7 showthe bias resulting from variation for the A chambers seems to be relatively
using only part of each chamber.For example, if one truck stable over the duration of these studies but the pat-
(or rack) within a chamberis used for each treatment, the tern in one chamber is not transferable to another
error bias ratios are 90, 32, and 5, for the A, B, and C without testing.
chambers, respectively. 3. In the B chambers the truck × chamber mean
Each of the sources of variation in the anMysisshownin square is again the next largest meansquare but this
Table 3 was partitioned further; variation amongexperi- time muchsmMler, relatively, than in the A chain-
556 CROP SCIENCE, VOL. 92, MAY-JUNE ,982

Trial Trial

B1

B2

B3

88

Fig. 2. Contours of individual experiments, arranged by chambers, of a fourth degree response surface for square root of plant height. Each
contour shows a range of 0.5 mm ~ units. Chamber identification is shown at the left. Trial identification is shown at the top. Chambers B2
and B8 in trial 3 conducted at 26/22 C temperature are shown with $. Chamber B3 in trial 6 with a defective temperature control is shown
with $. The bold-face letters P and V on each diagram show locations of peaks and valleys in the contour surface.

Table 6. Estimatedintraclass correlations for between-chamber Table 7. Error bias ratios~ resulting fromu~ing various within
experiments, averagedover variables and studies, where the experimental unit plant-to-plant variances as experimental
class, or experimentalunit, is the indicated portion of the error for testing between-chamber treatments. Theresults are
chamber. computedfrom the appropriate observed intraclass correla-
tions averagedover all variables and studies within a chamber
Chamber type type.
Class(experimental
unit) A B C Chamber type
Full chamber 0.31~ 0.42 0.38 Experimental unit A B C
One truck~chamber 0.56 0.67 0.60
One row/chamber 0.33 0.45 0.44 Full chamber 175 91 37
One column/chamber 0.42 0.67 0.52 One truck/chamber 20 32 5
One row/chamber 8 13 8
~fAll intraclass correlations involved in all averages are significantly One column/chamber 18 15 6
greater than zero (p = 0.05).
~fTheerrorbiasratiois computedas 1 + (n-l)(~ll-~)wheren is
bers. Thus, the large position effects are showing a numberof plantswithineachexperimental
unitand ~ is the appropriate
intraclass
correlation.
great deal of consistency over chambers as well as
time. This is a confirmation of the general appear-
ance of greater consistency in the response surface
contours for the B chambers shown earlier. may change with the plant species used and with the con-
4. In the C chambers, none of the interactions seem to ditions under which studies are conducted, it is evident
dominate although most are signifiant. Thus, the that use of proper within-chamber experimental designs
position effects show some consistency over cham- should increase the efficiency of experimentation. These
bers and time but the changes in pattern that do data suggest that simple blocking by trucks (or racks), for
occur are, more or less, unpredictable. example, in a similar experiment involving the entire
To summarize, there are distinct patterns to the within- chamber would reduce the error sum of squares by an
chamber variation in these studies that, to a sizeable average of 36% in A chambers, 43% in the B chambers
degree, are characteristic of the chamber type, i.e., and 52% in the C chambers. It is equally clear that ran-
repeatable over time and space. At the same time, there domization must be used to insure unbiased estimation of
are significant components of these patterns which "are experimental error.
unique to given chambers, particularly in the A cham- Shifting now to the analyses of the chamber-trial differ-
bers, and which change with time. While it must be rec- ences, the following conclusions are relevant:
ognized that the specific patterns of variation observed 1. A chambers: The large variation among experi-
LEE & RAWLINOS: DESIGN OF EXPERIMENTS IN GROWTH CHAMBERS 557

ments (Table 3) is not ascribable to either chambers ronment used in the broadest sense. Several other studies
or trials; that is, no consistent chamber effect nor (see the References) have shown similar results. It is
time effect is responsible for the large differences unlikely that such environmental variations would not
among experiments. The implication is that blocking impact to some extent on the growth of other species under
either by chamber or time would be largely ineffec- other conditions.
tive in controlling the experiment-to-experiment
variation. SUMMARY AND CONCLUSIONS
2. B chambers: The major contributor to the variance
among experiments is trial effects; that is, changes The uniformity data using soybeans in the NCSU Phy-
associated with time. It might be noted that the B totron vividly demonstrate that all of the conventional
chamber studies spanned a greater length of time experimental design principles must be considered in
than the studies in the other two types of chambers. experiments conducted in growth chambers. In spite of
3. C Chambers: There was again some indication that highly technical control of the environments and stand-
trial effects (time) was of some importance in ardization of daily operating procedures, the variation in
accounting for the experiment-to-experiment varia- plant growth both within chambers and between trials
tion but the chamber effects were nearly as impor- (chambers and/or time) was neither trivial nor random.
tant. Failure to use randomization and proper replication can
Thus, it is not clear whether blocking by chamber or lead to biased treatment effects and gross underestimation
time would be more effective for between-chamber stud- of experimental error.
ies. It is clear, however, that the chamber and the time Experiments involving treatments which require an
(trial) components of variance are sizable and need to be entire chamber for a given level must use chambers as
taken into account in the between-chamber experimental experimental units and, consequently, replication must
designs. The same nominal chamber setting is not pro- involve the use of several chambers (possibly over time).
ducing the same plant response over time over chambers. Experiments which use both between-chamber and
This suggests that the practice of using standard cham- within-chamber treatments belong to the classical split-
bers, i.e., certain chambers maintained continuously at plot experiments and, as in nearly all split-plot field
prescribed settings, may be yielding biased estimates of experiments, the sub-plot error (the within-chamber
treatment effects and, therefore, is an undesirable prac- error) is much smaller than the whole plot error and totally
tice. inappropriate for testing between-chamber treatments;
The final point to be emphasized is that, because of the the whole-plot error must be estimated via replication of
large components of variance for chambers and trials and the between-chamber treatments. Such replication is very
the very important spatial correlations among plants expensive, a fact which only increases the importance of
within chambers, it is an absolute requirement that exper- using proper and efficient experimental designs.
imental error for between-chamber experiments be esti- Even in experiments which involve only within-cham-
mated via proper replication of the experiment. The ber treatments, the nonrandom nature of the within-
chamber-trial combination is, in such cases, the experi- chamber variation dictates that proper randomization
mental unit and each treatment must be replicated by techniques be used to avoid bias in the estimation of
using several chambers and/or repeating the experiment experimental error and treatment effects and that the
over time. Experimental error estimated from within- experimental unit (a plant, a truck, etc.) be properly
chamber "replication" is inappropriate for testing identified so that true replication will be used. In addi-
between-chamber differences. The error bias ratios (Table tion, the magnitude of the patterned within-chamber var-
7) show that within-chamber estimates of error, com- iation shows that efficiency of experimentation can be
pared to proper between-chamber estimates, are too small increased by paying proper attention to use of blocking
by a factor (on the average in these studies) of from 6 to techniques and covariates.
175. Recalling that an F-ratio of 2.5 to 3.0 is declared sig- While within-chamber experiments can be completed,
nificant (a — 0.05) for any reasonable number of degrees including within-chamber replication, in a single trial in
of freedom, it is clear that even the smallest of error bias one chamber, it should be borne in mind that some uni-
ratios completely invalidates any tests of significance. dentified environmental and/or experimenter variables are
While these results were obtained from an extensive set changing with time and chamber even in the same phy-
of 39 uniformity trials, over 13 different growth cham- totron, as evidenced by the large chamber and trial effects.
bers, only the soybean cv. Bragg was grown and, for the Thus, it may be desirable, in order to broaden the base of
most part, in one standardized environmental regime. inference, to replicate even the within-chamber experi-
Consequently, one does not expect the quantitative results ments over chambers and/or time (or possibly over phy-
such as estimates of intraclass correlations, error bias totrons).
ratios, and response surface patterns to extrapolate with-
out modification to other conditions in this or other labo-
ratories. However, it is the mere presence of nonrandom
variation that is critical and the results have clearly shown
nonrandom variation in plant growth within and between
chambers in the NCSU Phytotron which, presumably, is
the result of nonrandom environmental variation, envi-
558 CROP SCIENCE, VOL. 22, MAY-JUNE 1982

You might also like