Regression-Based Modeling of Environmental Factors Affecting

412 Chiang Mai J. Sci. 2011; 38(3)

Chiang Mai J. Sci. 2011; 38(3) : 412-429 http://it.science.cmu.ac.th/ejournal/ Contributed Paper

Regression-based Modeling of Environmental Factors Affecting Macrobenthic Fauna in Middle Songkhla Lake, Thailand Uraiwan Sampantarak*[a], Noodchanath Kongchouy**[b], and Saowapa Angsupanich [c] [a] Pattani Inland Fisheries Research and Development Center, Department of Fisheries, Pattani 94160, Thailand. [b] Department of Mathematics, Faculty of Science, Prince of Songkla University, Songkhla 90112, Thailand. [c] Department of Aquatic Science, Faculty of Natural Resources, Prince of Songkla University, Songkhla 90112, Thailand. Author for correspondence; e-mail: *[email protected], **[email protected]

Received: 23 March 2010 Accepted: 21 June 2011

ABSTRACT The main objective of this study was to consider multivariate multiple regression (MMR) as an alternative to canonical correspondence analysis (CCA) for examining associations between macrobenthic fauna assemblage densities and environmental characteristics. The data used for this comparison comprise macrobenthic fauna densities and water quality parameters including sediment characteristics obtained from nine sampling sites in Middle Songkhla Lake at bimonthly intervals from April 1998 to February 1999. The outcome variables were the densities of twenty four macrobenthic fauna families with greater than 35% occurrence, log- transformed to remove skewness. Factor analysis was used to deﬁ ne ﬁ ve predictors including three composite variables based on salinity, physical sediment characteristics, and physico- chemical properties of water and sediment, together with total suspended solids and dissolved oxygen as single variables. The results show that MMR can be used as an additional or alternative method and has some advantages over the CCA method.

Keywords: macrobenthic fauna, multivariate multiple regression model, factor analysis, canonical correspondence analysis.

1. INTRODUCTION Macrobenthic fauna are recognized them are unable to avoid adverse conditions as sensitive indicators of environmental brought about by natural stresses or human disturbance [1-3] partly because of the limits impacts. Moreover, their relative longevity, to their mobility. They can be very mobile but with many species having life cycles over only at a scale of mm or cm; pollution acts at two years, allows them to integrate responses a scale of 100s of metres or more. Many of to environmental processes over extended Chiang Mai J. Sci. 2011; 38(3) 413

time periods [4]. In addition, the observed method for assessing relationships between distribution of macrobenthic fauna may be species abundance and their environmental useful in diagnostic studies and environmental predictors, for the simple reason that it is monitoring [5]. It is desirable to fi nd a useful the natural extension of ordinary regression predictive model for assemblages to identify analysis involving a single outcome to any factors infl uencing their absence or presence, number of mutually correlated outcomes such and abundance, and which can be used in as species abundances. It is thus of interest management of environments. to compare this method with its biologically Clarke and Warwick [6] outlined the preferred counterpart using common sets of basic methods commonly used by biological biological data relating taxonomic abundances scientists for preliminary analysis of their data. to environmental determinants. For this For comparative studies to assess associations comparison we used data from a study between taxa abundance and environmental involving macrobenthic fauna abundances predictors, canonical correspondence analysis and various water and sediment characteristics [7] is used extensively in the biological collected at specifi ed locations in an estuarine literature [8]. Some important studies using lake over a period of one year, reported by this method include those reported by Angsupanich et al. [22]. Rakocinski et al. [9], Hawkins et al. [10], Joy and Death [11], Guerra-García et al. 2. MATERIALS AND METHODS [12], Hajisamae and Chou [13], Morrisey et al. [14], Ysebaert et al. [15], Quintino et al. 2.1 Materials [16], Anderson [17], Glockzin and Zettler Data on macrobenthic fauna abundances [18]. This non-parametric method gives and various water and sediment characteristics graphical displays in biplots but does not were collected bimonthly from April 1998 to provide numerical measures of strengths of February 1999 at nine specifi c locations in the associations and their standard errors. Such Middle Songkhla Lake in Southern Thailand measures are provided by regression models, (Figure 1). For more details on the study area provided data are transformed to reduce and data collection see Angsupanich et al. skewness. [22]. The macrobenthic fauna samples were Although exceptions exist, such collected using a Tamura’s grab (0.05 m2), as studies by Liang et al. [19] exploring sieved consecutively through the screens and relationships between benthic fauna and water fi xed in 10% Rose Bengal-formalin for later quality variables using structural equation identifi cation. modeling (SEM) and by Warton and Hudson The densities of macrobenthic fauna [20] for tests of multivariate abundance were recorded as the number of individuals data using multivariate analysis of variance per square meter (ind m-2) for each species. (MANOVA), multivariate multiple regression A total of 161 taxa of macrobenthic fauna (MMR) analysis is not commonly used in were found and classifi ed into 81 families. In the biological literature for analyzing species many cases the species could not be identifi ed abundance patterns, possibly because of exactly, so the outcomes were classifi ed by skewness in data sets (see, for example, family instead of species. With nine locations Warton [21]). However, with an appropriate and six bimonthly data study periods, we data transformation to remove skewness, this defi ned the occurrence for a specifi ed family method would appear to be an ideal statistical as the proportion of these 54 occasions on 414 Chiang Mai J. Sci. 2011; 38(3)

which at least one organism was found. For salinity (Sal), water pH (wpH), dissolved the purposes of our comparison study, we oxygen (DO), total suspended solids (TSS), selected the 24 families with greater than 35% with sediment pH (spH), total nitrogen occurrence, thus covering 93.2% of the total content (TN), organic carbon content (OC), macrobenthic fauna assemblages. and grain size (percentages of sand, silt, and Environmental variables comprised water clay). These data were measured with three depth (wDep), water temperature (wTemp), replications on each occasion.

Figure 1. Songkhla Lake and sampling sites (labeled 1-9). Chiang Mai J. Sci. 2011; 38(3) 415

2.2 METHODS To satisfy statistical assumptions, promax rotation in preference to varimax, the response variable was defined as which requires the rotation to be orthogonal log(1cudensity), with the multiplier c chosen [26, 27]. to best approximate normality of error distributions. The predictors comprised 2.2.2 Multivariate Multiple Regression environmental components derived from a Multivariate multiple regression (MMR) factor analysis together with unique variables is used to evaluate the effects of multiple not accommodated by the factor model. predictor variables on multiple response The methods compared were canonical variables. The model [25] may be deﬁ ned in correspondence analysis (CCA) using matrix form as CANOCO Version 4.5 [23] and multivariate multiple regression (MMR) using R Version 8 n u p & nuq % qu p ( nu p (1) 2.10.0 [24].

In this formulation Y(nup) is an observed 2.2.1 Factor Analysis matrix of p response variables on each X Factor analysis was performed on the of n occasions, (n u q) is the matrix of q environmental variables with the aim of predictors (including a vector of 1s) in substantially reducing correlations between columns and n occasions in rows, B them that could mask their associations with (q u p) contains the regression coefficients E the outcome variables (see, for example, (including the intercept terms), and (n u p) is Chapter 9 of Mardia et al. [25]). Each factor a matrix of unobserved random errors with identifies correlated groups of variables. mean zero and common covariance matrix Ideally each group (which should contain at . Ordinary (univariate) multiple regression least two variables to contribute to the factor arises as the special case when p 1. If (k) analysis) contains variables having small q1 environmental predictors fi (k 1, correlations with variables in other groups. 2,…, q 1) are available, the prediction model To achieve this goal, any variable uncorrelated for outcome j on occasion i may be expressed with all other variables is omitted from as the factor analysis. Each factor comprises q1 k k weighted linear combinations of the variables yij P j ¦E j f i (2) and these factors are rotated to maximize the k 1 weights of variables within the factor group The model fi t may be assessed by plotting and minimize the weights of variables outside the residuals against normal quantiles [28], the group. The resulting weights are called and also by using the set of r-squared values “loadings”. Variables omitted from the factor for the response variables to see how much analysis due to low correlation with all other of the variation in each is accounted for by variables (high “uniqueness”) are treated as the model. separate predictors, so predictors include single variables as well as factors. The number of factors selected was based on obtaining an acceptable statistical fi t using the chi-squared test, and these factors were fi tted using maximum likelihood with 416 Chiang Mai J. Sci. 2011; 38(3)

The method also provides standard also produces coordinate scores and p-values errors for each of the p u q regression for the overall associations based on Monte coefficients thus providing p-values for Carlo permutation tests but not p-values for testing their statistical significance after local relationships. appropriate allowance for multiple hypothesis testing. The multivariate analysis of variance 3. RESULTS (MANOVA) decomposition is also used to 3.1 Environmental Parameters assess the overall association between each Figure 2 plots the water characteristics environmental predictor and the set of in Middle Songkhla Lake from April 1998 outcomes by the likelihood ratio [28-30]. to February 1999. The water depth was higher during the rainy season, varying with 2.2.3 Canonical Correspondence Analysis location from an average of less than 1 m at Assuming that the data structure stations four and nine to more than 2 m at comprises the Y and X matrices with rows station eight. The water temperature showed corresponding to measurements of outcomes decreased values in the rainy season, with and predictors taken on the same occasions, a range of 27-34°C. The salinity increased canonical correspondence analysis [7] from close to zero during the rainy season produces a two-dimensional biplot comprising (December to February) to an average close arrows with variable lengths and directions to 20 in other months. The pH of water was (gradients) emanating from a common origin also lowest in December. representing the predictor variables, together Figure 3 plots sediment characteristics with superimposed points denoting the measured on the same occasions as the water outcome variables. The relative lengths of characteristics. The total nitrogen content the arrows and the angles between them at each station was very low (0.02%) from are based on the correlation matrix of the October to February. The organic carbon predictor variables, and the coordinates of the content was relatively constant with respect to points are planar projections of the density month, but showed the highest value at station outcomes, computed in such a way that nine in every month except August. The lake their positions relative to the arrows portray bed at station six was mostly characterized their associations with the environmental by sand (mean 84.6%) and station 9 was predictors. Relatively greater distance along mostly characterized by clay (mean 53.2%), the arrow and relatively closer proximity also with high values of organic carbon. Note to the arrow line indicates likelihood of that the sand, silt, and clay percentages sum relatively stronger association. The method to 100%. Chiang Mai J. Sci. 2011; 38(3) 417

Figure 2. Water characteristics (wDep: water depth; wTemp: water temperature; Sal: salinity; wpH: water pH; DO: dissolved oxygen; and TSS: total suspended solids) in Middle Songkhla Lake from April 1998 to February 1999 by station (left panel) and month (right panel). 418 Chiang Mai J. Sci. 2011; 38(3)

Figure 3. Sediment characteristics (spH: sediment pH; TN: total nitrogen contents; OC: organic carbon contents; percentage of clay, silt, and sand) in Middle Songkhla Lake from April 1998 to February 1999 by station (left panel) and month (right panel). Chiang Mai J. Sci. 2011; 38(3) 419

3.2 Occurrence and Abundance of Macrobenthic Fauna Table 1 shows the taxa percentages Terebellidae). Crustacea was also represented of occurrence and density as individuals by nine families (Aoridae, Isaeidae, Melitidae, per square meter of the 24 families of Oedicerotidae, Apseudidae, Pseudotanaidae, macrobenthic fauna measured with the Anthuridae, Cirolanidae and Alpheidae). water characteristics. A total of 24 families Marginellidae, Retusidae, Skeneopsidae and were classified in three phyla of Annelida Stenothyridae were in the Gastropoda whilst (Polychaeta), Arthropoda (Crustacea) and the two remaining families (Tellinidae and Mollusca (Gastropoda and Bivalvia), which unidentiﬁ ed species were in the Bivalvia). comprised the most diverse group (35.2-98.2% Nereididae was the most commonly of occurrence). Polychaeta was represented observed family, found in 98.2% occurrence by nine families (Capitellidae, Goniadidae, of all station-month samples whereas Hesionidae, Nephtyidae, Nereididae, Terebellidae and Stenothyridae had the lowest Pilargiidae, Pholoidae, Spionidae and occurrence (35.2%). Apseudidae was the

Table 1. Taxa occurrence (%occ) and density in individual per square meter (ind m-2) of 24 families of macrobenthic fauna in Middle Songkhla Lake from April 1998 to February 1999.

Density (ind m-2) Phylum Class Order Family %occ (mean ± SE) Polychaeta Capitellida Capitellidae (Cap) 87.0 1,227.3 ± 227.0 Polychaeta Phyllodocida Goniadidae (Gon) 37.0 443.6 ± 149.7 Polychaeta Phyllodocida Hesionidae (Hes) 55.6 698.2 ± 177.9 Polychaeta Phyllodocida Nephtyidae (Nep) 87.0 2,218.2 ± 508.9 Annelida Polychaeta Phyllodocida Nereididae (Ner) 98.2 8,507.3 ± 1,579.5 Polychaeta Phyllodocida Pilargiidae (Pil) 70.4 1,625.5 ± 364.0 Polychaeta Phyllodocida Pholoidae (Pho) 59.3 658.2 ± 185.3 Polychaeta Spionida Spionidae (Spi) 92.6 5,056.4 ± 145.9 Polychaeta Terebellida Terebellidae (Ter) 35.2 1,136.4 ± 854.1 Crustacea Amphipoda Aoridae (Aor) 59.3 2,421.8 ± 968.5 Crustacea Amphipoda Isaeidae (Isa) 87.0 6,900.0 ± 2,014.4 Crustacea Amphipoda Melitidae (Mel) 94.4 4,438.2 ± 982.9 Crustacea Amphipoda Oedicerotidae (Oed) 66.7 667.3 ± 129.2 Arthropoda Crustacea Tanaidacea Apseudidae (Aps) 90.7 40,083.6 ± 8,553.7

Crustacea Tanaidacea Pseudotanaidae (Pse) 37.0 4,265.5 ± 2,172.3 Crustacea Isopoda Anthuridae (Ant) 75.9 3,816.4 ± 1,335.2 Crustacea Isopoda Cirolanidae (Cir) 37.0 427.3 ± 229.4 Crustacea Decapoda Alpheidae (Alp) 40.7 98.2 ± 23.9 Gastropoda Neogastropoda Marginellidae (Mar) 85.2 3,963.6 ± 940.9 Gastropoda Cephalaspidea Retusidae (Ret) 55.6 5,536.4 ± 2,509.2 Gastropoda Mesogastropoda Skeneopsidae (Ske) 38.9 956.4 ± 704.3 Mollusca Gastropoda Mesogastropoda Stenothyridae (Ste) 35.2 581.8 ± 386.9 Bivalvia Unidentiﬁ ed Unidentiﬁ ed (Uni) 44.4 338.2 ± 119.2 Bivalvia Veneroida Tellinidae (Tel) 81.5 17,134.5 ± 5,931.7 420 Chiang Mai J. Sci. 2011; 38(3)

Table 2. Factor analysis (with loadings below 0.2 omitted).

Environmental variables Factor 1 Factor 2 Factor 3 Organic carbon (OC) - - 0.47 Total nitrogen (TN) 0.34 - 0.58 Sediment pH (spH) 0.39 - î 0.57 Water depth (wDep) î 0.53 - - Water pH (wpH) 0.70 - - Salinity (Sal) 0.99 - - Water temperature (wTemp) 0.42 - 0.54 Sand - 0.94 - Clay - î 0.95 - % Total variance 24.6 20.7 13.7 % Cumulative variance 24.6 45.3 59.0

most abundant family with average density of represents the effect of sediment characteristics 40,083.6 ± 8,553.7 ind m-2; on the other hand, in the lake bed (sand-clay habitat), consisting the Alpheidae was least abundant with average of a positive loading for sand and a similar density 98.2 ± 23.9 ind m-2. negative loading for clay. Factor 3 characterizes physical and chemical compositions in the 3.3 Factor Analysis lake, comprising positive loadings for TN, DO and TSS were omitted from the OC, and wTemp, and a negative loading for factor model due to high uniquenesses spH. Thus Factor 1 was defi ned as 0.53 × (0.975 and 0.848, respectively). The model wDep 0.70 × wpH 0.99 × Sal, Factor 2 provided an adequate fi t using three factors as 0.94 × Sand 0.95 × Clay, and Factor 3 (chi-squared 20.23, 12 df, p-value 0.063). as 0.47 × OC 0.58 × TN 0.57 × spH Table 2 shows the loadings, with values less 0.54 × wTemp. The three factors respectively than 0.20 in magnitude suppressed. If only accounted for 24.6%, 20.7%, and 13.7% of loadings greater in magnitude than 0.45 are the variance in the environmental data, a total considered, the three factors do not contain of 59.0%. Thus three factors were included any overlapping variables. in the regression model as predictors together Factor 1 encompasses salinity, containing with the two singleton variables omitted from positive loadings for Sal and wpH, and a the factor analysis, with each of these fi ve negative loading for wDep as expected, with predictor variables scaled to have mean 0 and deeper water during the rainy season. Factor 2 standard deviation 1. Chiang Mai J. Sci. 2011; 38(3) 421 2 Spi = Spionidae; ; Ant = Anthuridae; ) 0.15 ed; Tel = Tellinidae. ed; Tel 0.21 (0.24) 0.16 ( (0.21) (0.14) 0.19 0.17 (0.27) 0.14 (0.24) (0.24) 0.32 0.17 (0.21) 0.35 tifi ns ns 0.42 (0.20) 0.08 0.66 1.08 0.45 0.75 0.85 0.73 î î predictors (left panel) environmental ve (0.23) 0.47 (0.22) 0.28 (0.23) 0.42 (0.14) 0.27 0.65 (0.25) 0.79 Factor 1Factor 2 Factor r 0.41 0.69 î î 2 cients with p-values > 0.05 in both models are omitted; cients with p-values 0.86 (0.36) 0.15 - - 0.07 î ) - 0.26 - ) - 0.45 - 0.53 ( 0.41 ( (0.58) - 0.22 - - 0.05 1.11 1.56 0.92 î tting multivariate multiple regression multiple models with all fi tting multivariate ) ) - - 0.25 ) - - 0.25 0.52 (0.23) - 0.17 0.28 0.19 0.30 cant in the MANOVA (right panel). Coeffi cant in the MANOVA ( ( ( 5 predictors 2 predictors 0.76 (0.31) 1.08 (0.45) - 0.31 0.61 (0.25) - 0.11 0.60 0.42 0.65 î cant are shown in italics and those with p-values < 0.01 are shown in bold. Note that TSS and DO referred < 0.01 are shown to in italics and those with p-values cant are shown ) (0.24) - - - 0.12 - 0.14 (0.22) - -(0.23) -(0.22) - - 0.33 - - - - 0.16 0.16 - - 0.47 (0.23) 0.09 ns (0.24) (0.26) - - - - 0.19 - ( (0.20) (0.24) - - - 0.38 ns ns ns 0.68 (0.25) - - - 0.17 - 0.16 1.04 0.81 0.30 0.69 0.87 î î (0.24) (0.15) (0.25) 0.51 (0.21) (0.18) 0.67 (0.29) - - - 0.58 (0.22) - - 0.30 - 0.24 0.42 (0.18) - 0.11 (0.28) - - - - 0.23 0.67 (0.25) - 0.18 (0.26) - ns ns ns ns 0.64 (0.28) 0.61 (0.25) 0.43 Factor 1Factor 2 Factor 3 Factor –TSS DO r 1.01 0.76 î î cients and standard errors (in parenthesis) from fi Coeffi Pil - Isa - 0.46 Ste 0.35 Cir 0.42 Tel - - - - - 0.10 - - 0.08 Spi 0.44 (0.21) - - - - 0.13 0.43 (0.19) - 0.09 Ret - - - Ter - Pse Ske - - - - - 0.02 - - 0.01 Alp - - - - - 0.09 - - 0.04 Uni 0.71 (0.27) - - - - 0.16 0.53 (0.25) - 0.08 Ant - Mel 0.20 Aor - Ner 0.27 Hes Cap - - - - Aps - - - - - 0.05 - - 0.04 Mar - 0.42 Pho Oed - - - - - 0.08 - - 0.04 Nep - Gon Family total suspended solids and dissolved oxygen, respectively. oxygen, total suspended solids and dissolved those adjudged not honestly statistically signifi and with only the two predictors found statistically signifi and with only the two Table 3. Table Cap = Capitellidae; Gon Goniadidae; Hes Hesionidae; Nep = Nephtyidae; Ner = Nereididae; Pil Pilargiidae; Pho Pholoidae; Ter = Terebellidae; Aor = Aoridae; Isa Isaeidae; Mel Melitidae; Oed Oedicerotidae; Aps Apseudidae; Pse Pseudotanaidae = Terebellidae; Ter Uni = Uniden Ste = Stenothyridae; = Skeneopsidae; Ske = Retusidae; Cir = Cirolanidae; Alp Alpheidae; Mar Marginellidae; Ret 422 Chiang Mai J. Sci. 2011; 38(3)

Table 4. MANOVA decomposition for multivariate multiple regression model with ﬁ ve predictors.

Source of Df Variation Df Pillai approx F (num) Df (denom) Prob (!F)

Intercept 1 0.990 104.092 24 25 0.0001 Factor 1 1 0.793 3.968 24 25 0.0005 Factor 2 1 0.778 3.641 24 25 0.0010 Factor 3 1 0.596 1.535 24 25 0.1468 TSS 1 0.503 1.055 24 25 0.4466 DO 1 0.339 0.691 24 25 0.8157 Residuals 48

3.4 Regression Analysis statistically signifi cant at 5% and 1% (in bold). The choice c 100 in the transformation Since there are 120 regression coeffi cients log (1 c u density) gave residuals satisfying in all and 5% of these would be expected the normality assumption. The left panel of to have p-values less than 0.05 even if all Table 3 shows the corresponding individual their corresponding population parameters regression coeffi cients and standard errors and were zero, the largest six p-values less than r-squared values for each family after fi tting 0.05 are italicized to indicate failure to the MMR model with all fi ve environmental achieve “honest” signifi cance. The additional predictors included. The right panel shows coeffi cients (labeled ns) were not statistically the corresponding results for a reduced model signifi cant in their fi tted model, but achieved containing only the two predictors that were signifi cance in the other model. Note that statistically significant in the MANOVA the coeffi cients for TSS are reversed in sign (Table 4). because most were negative, so this predictor The coefficients listed are the ones is labeled TSS. Chiang Mai J. Sci. 2011; 38(3) 423

Figure 4. Upper panel: biplot of fi rst two axes of CCA ordination diagram with 24 macrobenthic fauna families against fi ve environmental predictors. Lower panel: similar biplot omitting the three environmental variables not statistically signifi cant in the MANOVA, with families showing highly signifi cant coeffi cients in the multivariate multiple regression model (p-values < 0.01) connected to the corresponding arrows representing the predictors. (Cap = Capitellidae; Gon = Goniadidae; Hes = Hesionidae; Nep = Nephtyidae; Ner = Nereididae; Pil = Pilargiidae; Pho = Pholoidae; Spi = Spionidae; Ter = Terebellidae; Aor = Aoridae; Isa = Isaeidae; Mel = Melitidae; Oed = Oedicerotidae; Aps = Apseudidae; Pse = Pseudotanaidae; Ant = Anthuridae; Cir = Cirolanidae; Alp = Alpheidae; Mar = Marginellidae; Ret = Retusidae; Ske = Skeneopsidae; Ste = Stenothyridae; Uni = Unidentifi ed; Tel = Tellinidae).

3.5 Canonical Correspondence Analysis Figure 4 shows biplots based on the CCA matching the two MMR analyses. The families are represented by dots whereas the environmental predictors are represented by arrows. Each arrow determines an axis in the plots, obtained by extending the arrows in both directions. 424 Chiang Mai J. Sci. 2011; 38(3)

4. DISCUSSION Factor 1 and TSS have identical directions, but the correlation between these variables (0.30) 4.1 Comparison of Methods is not high. Multivariate multiple regression (MMR) and canonical correspondence analysis (CCA) 4.2 Biological Conclusions were used to examine the relations between The results, both by MMR and CCA, the macrobenthic fauna family densities and clearly indicate that the salinity factor was the reduced set of environmental predictors. positively associated with the densities of The biplots from using CCA method result the Goniadidae, Hesionidae, and Spionidae in locations relative to arrows that suggest and the unidentifi ed families in the Bivalvia, possible strong relationships between each of and negatively associated with the densities the fi ve predictors for at least one family, but of Pholoidae and Pseudotanaidae. This is in without p-values this cannot be confi dently contrast with analysis by Angsupanich et al. assumed. The MMR method provided [22] using BIOENV based on the same data, p-values that enabled the assumptions to which showed no evidence of salinity effect be made with confidence. In applying on benthos density. In general, salinity is an MANOVA decomposition only two factors important factor affecting the distribution were found to be statistically significant and structure composition of macrobenthic overall. The biplot produced by the CCA was fauna in brackish water of coastal habitats seen to be more informative when only these [34-36]. Although Middle Songkhla Lake two predictors were included. is not connected to the sea directly, this The MMR model containing all five zone receives the effect of salinity from the predictors gave seven associations between saltwater inflow through the Lower Lake a family density and an environmental which is open to the Gulf of Thailand. Salinity determinant that are highly statistically is often regarded as a primary descriptor in signifi cant (p-value 0.01), and a further nine estuarine ecosystems [37, 38]. with p-value between 0.01 and 0.05, however A sedimentary habitat contains ten families showed no evidence of an information mirroring the functional association with any of the fi ve determinants. biodiversity and activity patterns Most of these associations can also be seen in of macrobenthic fauna [39]. The main the CCA biplot. characteristics at the bottom of Middle The most noticeable difference between Songkhla Lake are clay and silt [22]) except the results of the two methods is that the for station six, which is mainly sand (84.6%). family Spionidae was found to be associated We found that sand/clay excess was positively with the salinity factor in the MMR model but associated with the densities of Terebellidae, this association was not seen in the biplots Aoridae, Pseudotanaidae, and Anthuridae, from CCA. Since Spionidae could be regarded while a negative association was found with as a marine taxon [32] with typical dominant Pilargiidae. A typical genus Sigambra within species Pseudopolydora kempi and Prionospio Pilargiidae [22] was found to be negatively cirrifera [22], there is evidence supporting the related with sand-clay excess, a finding MMR result in this case. On the other hand, supported by a study in the south-eastern there is evidence of occupation of estuaries Gulf of California reporting that the genus by spionidae [33]. In the biplot containing all Sigambra was dominant in the areas of sand fi ve environmental predictors the arrows for percentage of 1% or mud of 60-70% [40]. Chiang Mai J. Sci. 2011; 38(3) 425

In addition, the genus Marginella within structures, and the construction of a deep sea Marginellidae was also listed as being present port [46]. The analytic methods we have used in Middle Songkhla Lake [22] thus showing a are designed to gain a better understanding positive association with sand. This fi nding of the environmental factors associated with agrees with a study of invertebrate species macrobenthic fauna and can be used as an identified in Fresh Creek, Bahamas where additional or alternative method for analysis Marginella was listed as most commonly having of the relationships between environmental the habitat type of sandfl at [41]. variables and the abundance of benthic or Ten families (Nereididae, Stenothyridae, other aquatic organisms . This knowledge is Nephtyidae, Isaeidae, Marginellidae, useful for the natural resource management Alpheidae, Oedicerotidae, Apseudidae, of estuarine environments. Skeneopsidae, and Tellinidae) showed The major fi nding from this study was that no evidence of association with any of MMR was more informative when describing the environmental variables. Although relationships between macrobenthic fauna Alpheidae was found to have the lowest densities and environmental factors than was density among the families included in our the CCA method. MMR’s advantages over study, it is commonly found in the stomach conventional methods are that it separates contents of the dominant bottom feeding the effects of the environment on organisms, fi sh (Osteogeneiosus militaris and Arius maculates) gives statistics such as standard errors for in Middle Songkhla Lake. Angsupanich et al. these estimated effects and p-values and [42] implied that these catfi sh species feed also provides a predictive model for the opportunistically on a variety of prey in outcomes. The family was used in this study their environment coupled with preferential to investigate the appropriateness of the feeding. method, but other studies using the MMR Nereididae are the most important method could use species taxa if the species polychaetes due to their diversity and can be easily distinguished and if they occur abundance and found not only in marine in suffi cient numbers. environments [43] but also in brackish water such as occurs in Middle Songkhla Lake. 4.3 Limitation of Study Fourteen species of Nereididae were reported A serious limitation of our study was in a former study [22] and it seems that that the MMR method was compared with Nereididae is widespread in Middle Songkhla only one alternative method. Although Lake where it had the highest species richness. CCA has been the method most widely Some species, such as Ceratonereis hircinicola, used by ecologists for determining were widely spread in the high salinity areas associations between taxa abundances and [44], whereas Namalycastis indica has been their environmental determinants [8,47] found to inhabit fresh to slightly brackish other methods are also popular. These water in cisterns, pools and lagoons [45]. include non-parametric methods (MDS, Others are euryhaline. ANOSIM, BIOENV) [6] designed to avoid Songkhla Lake nowadays suffers from the the need to transform the outcome variables use of coastal land and water resources for to satisfy statistical assumptions. In recent uncontrolled shrimp farming, the destruction years permutation multivariate analysis of both mangrove areas and peat swamp of variance (PERMANOVA) [48-50] has forest, construction of intake and outfall become increasingly popular. This method 426 Chiang Mai J. Sci. 2011; 38(3)

also does not require data transformation Langesundfjord and in a mesocosm (unless biologically desirable) and has the experiment, Mar. Ecol. Prog. Ser., 1988; further advantage that it is carry out on a 46: 151-165. complete data matrix. Thus it would be of [5] Warwick R.M., A new method for interest to undertake comparisons of all three detecting pollution effects on marine methods in future studies. macrobenthic communities, Mar. Biol., 1986; 92: 557-562. ACKNOWLEDGEMENTS The data were collected using funds [6] Clarke K.R. and Warwick R.M., Change provided by the Biodiversity Research and in Marine Communities: an Approach to Training Program (BRT), under the National Statistical Analysis and Interpretation, Science and Technology Development Natural Environment Research Council, Agency (NSTDA), together with the Thailand Plymouth, UK, 1994. Research Fund (TRF) with the grant number [7] Ter Braak, C.J.F., Canonical corres- 142016. The authors are grateful to Emeritus pondence analysis: a new eigenvector Prof. Dr. Don McNeil and Asst. Prof. Dr. technique for multivariate direct gradient Sukree Hajisamae for their advice on the analysis, Ecology, 1986; 67: 1,167-1,179. statistical analysis and to the anonymous [8] von Wehrden H., Hanspach J., referees for invaluable suggestions. Bruelheide H. and Wesche K., Pluralism and diversity-trends in the use and REFERENCES application of ordination methods 1990- [1] Weisberg S.B., Dauer D.M., Schaffner 2007, J. Veg. Sci., 2009; 20: 695-705. L.C., and Frithsen J.B., An estuarine benthic index of biotic integrity (B-BI) [9] Rakocinski C.F., Brown S.S., Gaston for Chesapeake Bay, Estuaries, 1997; 20: G.R., Heard R.W., Walker W.W. and 149-158. Summers J.K., Macrobenthic responses to natural and contaminant-related [2] Borja A., Franco J., and Pérez V., A gradients in Northern Gulf of Mexico marine biotic index to establish the estuaries, Ecol. Appl., 1997; 7: 1,278-1,298. ecological quality of soft-bottom benthos within European estuarine and [10] Hawkins C.P., Norris R.H., Hogue J.N. coastal environments, Mar. Pollut. Bull., and Feminella J.W., Development and 2000; 40: 1,100-1,114. evaluation of predictive models for measuring the biological integrity of [3] Ranasinghe J.A., Thompson B., Smith streams, Ecol. Appl., 2000; 10: 1,456-1,477. R.W., Lowe S., and Schiff K.C., Evaluation of benthic assessment [11] Joy M.K. and Death R.G., Development methodology in Southern California and application of a predictive model of Bays and San Francisco Bay, Technical riverine ﬁ sh community assemblages in Report, Southern California Coastal the Taranaki region of the North Island, Water Research Project, 2004; 432. New Zealand, N.Z. J. Mar. Freshwat. Res., 2000; 34: 241-252. [4] Gray J.S., Aschan M., Carr M.R., Clarke K.R., Rosenberg R. and Warwick R.M., [12] Guerra-García J.M., González-Villa F.J. Analysis of community attributes of and García-Gómez J.C., Aliphatic hydro- the benthic macrofauna of Frierfjord/ carbon pollution and macrobenthic Chiang Mai J. Sci. 2011; 38(3) 427

assemblages in Ceuta harbour: a variables of water, benthic invertebrates, multivariate approach, Mar. Ecol. Prog. and feeding activity of waterbirds in the Ser., 2003; 263: 127-138. Sitsao wetlands of Southern Taiwan, Zool. Stud., 2002; 41: 441-451. [13] Hajisamae S. and Chou L.M., Do shallow water habitats of an impacted coastal [20] Warton D.I. and Hudson H.M., A strait serve as nursery grounds for fi sh?, MANOVA statistic is just as powerful as Estuar. Coast. Shelf Sci., 2003; 56: 281- distance-based statistics, for multivariate 290. abundances, Ecology, 2004; 85: 858-874. [14] Morrisey D.J., Turner S.J., Mills G.N., [21] Warton D.I., Many zeros does not mean Williamson R.B. and Wise B.E., Factors zero infl ation: comparing the goodness- affecting the distribution of benthic of-fi t of parametric models to multivariate, macrofauna in estuaries contaminated Environmetrics, 2005; 16: 275-289. by urban runoff, Mar. Environ. Res., 2003; [22] Angsupanich S., Siripech A. and 55: 113-136. Charoenpornthip M., Macrobenthic [15] Ysebaert T., Herman P.M.J., Meire P., fauna community in the Middle Songkhla Craeymeersch J., Verbeek H. and Heip Lake, Southern Thailand, Songklanakarin C.H.R., Large-scale spatial patterns J. Sci. Technol., 2005(a); 27: 365-390. in estuaries: estuarine macrobenthic [23] Ter Braak C.J.F. and Šmilauer P., communities in the Schelde estuary, NW CANOCO Reference manual and Europe, Estuar. Coast. Shelf Sci., 2003; CanoDraw for Windows User’s guide: 57: 335-355. Software for Canonical Community [16] Quintino V., Elliott M. and Rodrigues Ordination (version 4.5), Microcomputer A.M., The derivation, performance Power, Ithaca, NY, USA, 2002. and role of univariate and multivariate [24] R Development Core Team, R: A indicators of benthic change: Case language and environment for statistical studies at differing spatial scales, J. Exp. computing, R Foundation for Statistical Mar. Biol. Ecol., 2006; 330: 368-382. Computing, Vienna, Austria, ISBN [17] Anderson M.J., Animal-sediment 3-900051-07-0, URL http://www.R- relationships re-visited: Characterising project.org, 2009. species distributions along an environ- [25] Mardia K.V., Kent J.T. and Bibby J.M., mental gradient using canonical analysis Multivariate Analysis, Academic Press, and quantile regression splines, J. Exp. USA, 1979. Mar. Biol. Ecol., 2008; 366: 16-27. [26] Browne M.W., An overview analytic [18] Glockzin M. and Zettler M.L., Spatial rotation in exploratory factor analysis, macrozoobenthic distribution patterns in Multivariate. Behav. Res., 2001; 36: 111- relation to major environmental factors- 150. A case study from the Pomeranian Bay (southern Baltic Sea), J. Sea. Res., 2008; [27] Abdi H., Factor rotations in factor 59: 144-161. analyses. In: Lewis-Beck, M., A. Bryman, T. Futing (Eds), Encyclopedia of Social [19] Liang S., Shieh B. and Fu. Y., A structural Sciences Research Methods, Thousand Oaks, equation model for physiochemical CA, Sage, 2003: 1-8. 428 Chiang Mai J. Sci. 2011; 38(3)

[28] Venables W.N. and Ripley B.D., Modern macrobenthic invertebrates in a brackish Applied Statistics with S, 4th Edn., Springer lake in relation to environmental Science + Business Media, New York, variables, Estuar. Coast. Shelf Sci., 2005; USA, 2002. 63: 167-176. [29] Olson C.L., On choosing a test statistic [37] Gaston G.R., Lee D.A.L. and Nasci J.C., in multivariate analysis of variance, Estuarine macrobenthos in Calcasieu Psychol. Bull., 1976; 83: 579-586. Lake, Louisiana: Community and trophic structure, Estuaries and Coasts, 1988; 11: [30] Johnson R.A. and Wichern D.W., Applied 192-200. Multivariate Statistical Analysis, 4th Edn., Prentice-Hall, USA, 1998. [38] Lamptey, E., and Armah, A.L., Factors affecting macrobenthic fauna in a [31] Clarke K.R., Somerfi eld P.J. and tropical hypersaline coastal lagoon in Chapman M.G., On resemblance Ghana, West Africa, Estuaries and Coasts, measures for ecological studies, 2008; 31: 1,006-1,019. including taxonomic dissimilarities and a zero-adjusted Bray-Curtis coeffi cient for [39] Rosenberg R., Magnusson M. and denuded assemblages, J. Exp. Mar. Biol. Nilsson H.C., Temporal and spatial Ecol., 2006; 330: 55-80. changes in marine benthic habitats in relation to the EU Water Framework [32] Day R.L. and Blake J.A., Reproduction Directive: The use of sediment profile and larval development of Polydora imagery, Mar. Pollut. Bull., 2009 ; 58: 565- Giardi Mesnil (Polycheata: Spionidae), 572. Biol. Bull., 1979; 156: 20-30. [40] Méndez N., Relationships between [33] Walker L.M., Polydora and Dipolydora deep-water polychaete fauna and (Polychaeta: Spionidae) of estuaries and environmental factors in the south- bays of subtropical eastern Australia: A eastern Gulf of California, Mexico, Sci. review and morphometric investigation Mar. (Barc.), 2007; 71: 605-622. of their taxonomy and distribution, MSc thesis, Southern Cross University, [41] Layman C.A. and Silliman B.R., Lismore, NSW, 2009. Preliminary survey and diet analysis of juvenile fi shes of an estuarine creek on [34] Mannino A. and Montagna P.A., Andros Island, Bahamas, Bull. Mar. Sci., Small-scale spatial variation of 2002; 70: 199-210. macrobenthic community structure, Estuaries, 1997; 20: 159-173. [42] Angsupanich S., Somsak S. and Phrommoon J., Stomach contents [35] Ogunwenmo C.A. and Osuala I.A., of the catfi shes Osteogeneiosus militaris Physico-chemical parameters and (Linnaeus, 1758) and Arius maculatus macrobenthos of an estuarine creek (Thunberg, 1792) in the Songkhla Lake, and an artifi cial pond in Lagos, South- Songklanakarin J. Sci. Technol., 2005(b); 27: Western Nigeria, Acta SATECH, 2004; 391-402. 1: 128-132. [43] Gonzalez-Escalante L.E. and Salazar- [36] Nanami A., Saito H., Akita T., Vallejo S.I., A new estuarine species, Motomatsu K. and Kuwahara H., Spatial Nereis garwoodi (Polychaeta: Nereididae), distribution and assemblage structure of Chiang Mai J. Sci. 2011; 38(3) 429

from Bahia Chetumal, Macican [47] Kongchouy, K., Methods for modeling Caribbean coast, Rev. Biol. Trop., 2003; overdispered-outcomes with application 51: 155-164. to disease incidence rates and macrobenthic fauna densities, Ph.D Thesis in Research [44] Angsupanich S. and Kuwabara R., Methodology, Prince of Songkla University, Macrobenthic fauna in Thale Sap Thailand, 2010. Songkla, a brackish lake in southern Thailand, Lakes Reservoirs: Res. Manage, [48] Anderson M.J., A new method for 1995; 1: 115-125. non-parametric multivariate analysis of variance, Austral Ecology, 2001; 26: 32-46. [45] Glasby C.H., The Namanereidinae (Polychaeta: Nereididae), Part 1, Taxonomy [49] McArdle B.H. and Anderson M.J., and phylogeny, Records of the Australian Fitting multivariate models to community Museum, 1999: 129. data: a comment on distance-based redundancy analysis. Ecology, 2001; 82: [46] Chufamanee P., Boromthanarat S., and 290-297. Lønholdt J., A case experience on integrated environment and water management [50] Tracey S.R., Lyle J.M. and Duhamel G., towards people’s livelihoods, Case story 1: Application of elliptical Fourier analysis Partnership policy in Songkhla Lake Basin, of otolith form as a tool for stock Workshop report. Songkhla workshop identiﬁ cation, Fisheries Research, 2006; 77: on linking management of catchments 138-147. and coastal ecosystems, Royal Thai Ministry of Natural Resources and Environment, 2003: 15-32.