Inflection and stability points of diphasic logistic analysis of growth

Abstract

Growth functions with inflection points following a diphasic model, can be adjusted by two approaches using segmented regression or the sum of two functions. In both cases, there are two functions, one for each phase, with inflection and stability points. However, when they are summed, the result is a new function and the points of inflection and stability are different from those obtained from using each function individually. A method to determine these points in a diphasic logistics sum of functions is suggested and the results obtained from fitting the models to eucalyptus growth data showed a better fit of the logistic diphasic sum as compared with segmented regression and monophasic logistic models.

nonlinear models; segmented regression; sum of logistics

Introduction

Monophasic models of linear or non-linear growth are a description of the total growth cycle, from birth to adulthood of the individual, which may be greatly simplified. However, multiple growth cycles can be found in the literature (Seber and Wild, 1989) from the beginning of the 20^th century (Brody and Ragsdale, 1921). Since then, curve adjustment by multiphasic models has been employed in various areas, such as human height growth (Bock et al., 1973), growth in weight of rats (Koops et al., 1987;Kurnianto et al., 1999), chickens (Grossman and Koops, 1988), rice crop biomass (Sheehy et al., 2004) and cows (Mendes et al., 2008;Mendes et al., 2009).

The logistic function has been the one most frequently used in multiphasic models due to its properties of symmetry in the growth velocity curve (Koops, 1986;Kurnianto et al., 1999;Mendes et al., 2008). This model may be fitted either by segmented regression (Portz et al., 2000;Robbins et al., 2006) or by the sum of functions – the most commonly employed methods (Koops, 1986;Koops and Grossman, 1991;Kurnianto et al., 1999;Özkan, 2004;Nešetřilová, 2005;Mendes et al., 2008;Mendes et al., 2009;Fenner et al., 2013).

In the nonlinear multiphasic model with inflection points, adjusted by the sum of functions, the points are determined in the functions that correspond to each growth phase (Koops, 1986;Koops et al., 1987;Koops and Grossman, 1991;Kurnianto et al., 1999;Mendes et al., 2009;Fenner et al., 2013), but the inflection points of the sum function are not determined

In this study, a methodology for the determination of inflection and stability points of logistic diphasic models as adjusted by the sum of functions is presented, which are compared with the points obtained by segmented regression and the monophasic model.

Materials and Methods

Consider the following logistic models, where α > 0 and γ > 0, fitted to a data set (x_i; y_i) from observed data, i = 1, 2, ..., n, where errors are considered independent, normal and homoscedastic, ε~N (0, σ²_ε).

Model I - monophasic logistic model

Model II - diphasic logistic model segmented regression

θ = [α_kβ_k γ_kι]', k = 1, 2, and ι = abscissa of the intersection point of the two functions.

Model III - diphasic logistic sum of functions

Residual homoscedasticity is checked by the Breusch-Pagan test (Breusch and Pagan, 1979) and normality by the Shapiro-Wilk test. The residual autocorrelation is verified by the use of tables for testing randomness of grouping in a sequence of residual signs (Draper and Smith, 1998)

Adjustments to the models can be checked by the usual criteria for comparing models: the residual mean square of the model, RMS; the square of the correlation coefficient between observed and fitted values by the model, r²_y.ŷ (Schinckel and Craig, 2002), and the F criterion

where: SSR₁ is the sum of the squares of residuals of the model with a smaller number of parameters, SSR₂ is the sum of the squares of residuals of the model with a larger number of parameters, and df₁ and df₂ are degrees of freedom associated with SSR₁ and SSR₂, respectively;

the corrected Akaike Information Criterion:

where: SSR is the sum of the squares of residuals of the model, andp is the number of model parameters, (Narinc et al., 2010). The Akaike information criterion evaluate whether the model adequately describes the studied population: the lower the value, the better the model.

To check the fit of the models in the early growth stage the residual square mean ofm values, RSMm, was calculated as follows:

where: y_i is the i-th observed value, ŷ_i is the i-th value estimated by the model andm is the number of observations corresponding to the first phase of the fitted logistic by segmented regression (MII).

The inflection points of the logistic function, determined by equating its second order derivative to zero, are obtained by the usual formula (-β/γ; α/2) for model I (-β_k/γ_k; α_k/2), k = 1, 2, for model II. In model III, however, the function (3) is a sum and its second order derivative equated to zero:

is an equation without explicit solution, which makes the problem more complex (Beyene and Ramakrishnan, 2013), though the solution can be approximated by iterative techniques. In this paper, the Newton-Raphson method is used,in the SASproc model (Statistical Analysis System, version 9.2).

The abscissas of the inflection points of the diphasic logistic sum of functions, therefore, cannot be determined by common formulas for each plot of (7), since there is no fitted function for each phase, but a sum of two functions in the same phase. The solutions of equation (7), denoted by υ, are contained in the interval τ₁ < x < τ_2,where τ_k = -β_k/γ_k, k = 1, 2; this is shown below:

Consider the logistic functions defined for α_k > 0, β_k < 0 and γ_k> 0,

The functions are continuous, differentiable, positive and increasing in the interval (-∞, ∞) and have an inflection point where x equals τ_k = –β_k/γ_k. The first order derivatives of y_k are continuous, positive and have maxima at x = τ_k, whose values are the roots of the second order derivative functions:

that are continuous, positive in the interval -∞ < x < τ_k and negative in τ_k < x < ∞. Then the sum function

is also continuous, differentiable, positive and increasing at (- ∞, ∞). Assuming τ₁ < τ₂ we have in the interval -∞ < x < τ_1,by (8), X₁ > 1 and X₂ > 1 and by (9), y”₁ > 0 and y”₂ > 0. Therefore y” = y”₁ + y”₂ > 0 in this interval. At the point where x = τ₁, X₁ = 1, X₂ > 1, y”₁ = 0, y”₂> 0 and y” > 0.

Then,

in -∞ < x ≤ τ₁, y” > 0; there are no inflections iny in this interval. (11)

In the interval τ₂ < x < ∞, by (8), 0 < X₁ < 1 and 0 < X₂ < 1 and by (9), y”₁ < 0 and y”₂ < 0. Therefore y” < 0 in this interval. At the value where x = τ₂, X₁ < 1, X₂ = 1, y”₁ < 0, y”₂ = 0 and y” < 0. Then,

where τ₂ ≤ x < ∞, y” < 0; there are no inflections iny in this interval. (12)

By (11) and (12), as y’’ is continuous on (- ∞, ∞) it follows that there is at least one value ofx in the interval (τ₁, τ₂) where y’’ is equal to zero. This value where x = υ is the abscissa of the inflection point of y in (10).

The stability points of the logistic function, mono or diphasic can be determined by various methods (Passos et al., 2012). The method that equates the fourth-order derivative of the functiony to zero is employed here (Mischan et al., 2011):

which gives the points

in model I and

in model II.

In model III, where there is a sum of functions, the equation

has no explicit solution. The solution method is the same used for the determination of inflection points.

The models were adjusted using the proceduremodel,method=marquardt, from SAS (Statistical Analysis System, version 9.2). The optionsbreusch ‘pagan’ andnormal in theproc model were employed to verify the homoscedasticity and normality of the residuals. All tests were verified at the significance level of α = 0.05.

Observational growth data (volume-age) of the trunks ofEucaliptus grandis L. of a reforestation zone in Jacareí, in the state of São Paulo (23^o22’27’’ S, 46^o1’34’’ W) were used to illustrate the methodology. The reforestation zone was available for research and consisted of 150 plants, arranged in three rows of approximately 50 trees each, with a spacing of 3.0 × 2.5 m between plants. One row of 50 plants selected at random and 29 trees with measurements taken at all times were considered. Individual settings for each tree were marked by 11 observations made from 8 to 50 months with the following values of x = {8, 10, 12, 15, 19, 21, 25, 27, 30, 36, 50 months}. The trunk volume data (m³) were calculated from the diameter at breast height (m) and tree total height (m).

Results and discussion

The estimates of parameters α, β, γ, in model I are denoted bya,b andc, respectively; in models II and III, the parameters α_k, β_k, γ_k bya_k,b_k,c_k, k = 1, 2, respectively; and in model II, the parameter ι is denoted byri.Table 1 shows the parameter estimates of models I, II and III. The criteria used to check the fit to the data of eucalyptus are presented inTable 2.

Table 1
– Mean (standard deviation) of parameters estimates of models I - monophasic, II - diphasic segmented and III - diphasic sum. (n = number of trees).

Models (n)	a1	b1	c1	a2	b2	c2	ri
Model I (29)	278.96	-5.17	0.1544	..	..	..	..
	(60.374)	(0.425)	(0.0093)	..	..	..	..
Model II (26)	101.64	-7.67	0.3542	287.88	-5.08	0.1478	25.59
	(54.832)	(2.307)	(0.1266)	(66.376)	(0.833)	(0.0260)	(3.749)
Model III (26)	65.49	-8.00	0.3822	200.28	-8.32	0.2316	..
	(33.958)	(3.359)	(0.1791)	(61.590)	(2.245)	(0.0634)	..

Table 2
– Mean (standard deviation) of criteria for comparison of models, I - monophasic, II - diphasic segmented and III - diphasic sum. Residual mean square, RMS, number of parameters,p, corrected Akaike Information Criterion, AICc, square of correlation coefficient, r2y.ŷ, and residual squares mean, RSMm. (n = number of trees).

Model (n)	RMS	p	AICc	r2y.ŷ	RSMm
Model I (29)	46.05(32.669)	3	45.52(7.784)	0.9953(0.0027)	35.29(25.439)
Model II (26)	11.39(14.374)	6	45.08(12.822)	0.9993(0.0006)	4.01(4.620)
Model III (26)	8.38(6.498)	6	44.98(7.878)	0.9994(0.0004)	3.30(2.406)

• m is the number of observations in interval x = 0 to x = ri, the interception value of the two logistics in model II.

The analysis of the residuals showed that the null hypothesis of homoscedasticity cannot be rejected, for the three models in all fits. The normality hypothesis was not rejected in 100 % of the cases for model I, 84 % for model II and 88 % for model III. The residuals can be considered independent for 100 % of the adjustments of model I, 88 % for model II and 96 % for model III.

In model III all parameter estimates were in accordance with the constraints α_k > 0, β_k< 0 and γ_k > 0, k = 1, 2, characterizing positive and increasing functions. The fitted values of asymptote in phase 2 of model II, with mean a₂ = 287.8842, and the values of the sum of the two asymptotes in the model III, where mean a₁ + a₂ = 265.7759, are near to the asymptote estimated in model I where mean a = 278.9652.

Compared with the monophasic model, the diphasic models fit better, not only throughout the measuring interval of trees, as shown by the values of residual mean square, RMS, as well as during the initial phase, as shown by the residual squares mean ofminitial values, RSMm. Model III is more efficient, both in the initial period fit as for the whole. On average, the reduction in RMS values compared to model I were 75 % for model II and 82 % for model III. The AICc and r²_y.ŷ criteria are similar in the three models, with slight improvements in the diphasic models.

Table 3 shows the values of the abscissas of inflection and stability points in adjusted models andFigure 1 illustrates the adjustments.

Table 3
– Mean (standard deviation) of the x values that are roots of the second order derivative functions (abscissas of the inflection points) and fourth order (abscissas of the stability points) of models I - monophasic, II - diphasic segmented and III - diphasic sum. (n = number of trees).

	Abscissas of inflection points			Abscissas of stability points
Models (n)	phase 1	inter-phases	phase 2	phase 1	phase 2
Model I (29)	..	..	33.46(1.902)	..	48.36(2.088)
Model II (26)	22.07(2.104)	..	34.47(1.560)	29.07(3.583)	50.46(3.594)
Model III (26)	21.26(1.392)	26.49(2.295)	36.60(0.950)	27.16(2.314)	47.41(2.828)
	[20.90(1.271)]	..	[36.80(1.062)]	[27.48(2.626)]	[47.44(2.843)]

• In brackets, abscissas determined by the functions at each phase of model III.

Figure 1
– Logistic models fitted to data from aEucalyptus grandis plant. (A) Monophasic logistic, (B) diphasic segmented logistic, (C) diphasic logistic sum and its components in phases 1 and 2; yo = observed values, ye = fitted values, a = asymptote, ri = abscissa of the intersection point between the phases in model II, pi = inflection points and pe = stability points. In model III, pi-y1 and pi-y2 = the inflection points in phases 1 and 2 curves, respectively; pi1, pi2 and pi3 are inflection points, pe1 and pe2 are stability points in the logistic sum curve. Indices 1 and 2 refer to phases 1 and 2.

Considering the average values of the inflection points obtained to each plant,Table 3, it is seen that the monophasic model has an inflection point with abscissa x = 33.46 months . In model II, diphasic segmented regression, the averages are x = 22.07 in the first phase and x = 34.47 months in the second, the latter being comparable to the average of the monophasic model. For model III – the diphasic sum of two logistics - 20 plants had three inflection points and six plants only one; the means were x = 21.26 months which corresponds to the inflection point in the first phase, x = 26.49, which is the abscissa of the point that separates the two phases and x = 36.60 months, which corresponds to the inflection point in the second phase. The abscissas values of the latter inflection point in the model III are higher in all plants, comparing to the values determined by the monophasic model. In this model III, the inflection points of the double logistic for all plants, are within the interval (t₁, t₂), wheret₁ is the abscissa of the inflection point of the first logistic andt₂, of the second, as demonstrated in this work.Table 3 shows 20.90 < 21.26 < 26.94 < 36.60 < 36.80 months.

Figure 2 shows the graphs of the derivatives of the first and second orders, with the location of the abscissas of the inflection points for two plants. In the interval (-∞; ∞), in (B) the derivative of second order y” = y”₁ + y”₂intercepts the x-axis three times, and in (D) just once. Thex value that defines the interphase, 26.49 months in the model III, is quite close to the estimate of the abscissa of the intersection point between the two logistics in model II, 25.59 months. These are points that can define the separation between the two phases of growth of the organism.

Figure 2
– Inflection points in the diphasic logistic sum in an example of two plants ofEucaliptus grandis. In (A) first order derivatives (y’) and in (B) second order (y”) for plant 1: a1 = 58.057, b1 = -6.506, c1 = 0.2898, a2 = 141.104, b2 = -7.632, c2 = 0.1959, t1 = 22.45, t2 = 38.95, v1 = 23.81, v2 = 28.61, v3 = 38.62; in (C) and (D) for plant 2: a1 = 52.603, b1 = -7.096, c1 = 0.3333, a2 = 222.031, b2 = -6.952, c2 = 0.1906, t1 = 21.29, t2 = 36.48, v = 36.28 months. The abscissas t1 and t2 are estimates of τ1 and τ2; v1, v2, v3, of υ1, υ2, υ3.

In determining the points that are the roots of the fourth order derivative in model III, the diphasic sum, there were five solutions for (16), considering as stability points the third solution, xe₁ = the abscissa of the stability point ofy in phase 1, and the fifth solution xe₂ = the abscissa of the stability point ofy in growth phase 2. These solutions are different from those obtained when determining stability points considering a function for each phase.Table 3 presents the averages of the abscissas of the stability points: in model I they are the third solution of the equation y⁽⁴⁾ = 0; in model II, the third solution for phase 1 and the sixth for phase 2; in model III, the third solution for phase 1 and the fifth for phase 2.

The abscissas of the stability points in phase two of models II and III are quite similar to the values found for model I, for all plants. The abscissas determined by the logistics at each stage in model III (Table 3 in brackets) are very similar, but not identical to those that are the roots of the fourth order derivative of the sum function; this can be observed in all plants.

Conclusions

The use of a diphasic logistic to represent eucalyptus growth up to 50 months was more effective than a monophasic logistic. The critical points of inflection and stability are determined in the diphasic segmented model by the known formulas for determining these points in monophasic models, but in the diphasic model sum, the points that are the roots of second and fourth orders derivatives cannot be determined explicitly, and their values are different from those determined for the individual logistics components of this model.

References

• Beyene, S.; Ramakrishnan, V. 2013. A numerical method for estimating the variance of age at maximum growth rate in growth models. Communications in Statistics: Theory and Methods 42: 1464-1475.
• Bock, R.D.; Wainer, H.; Petersen, A.; Thissen, D.; Murray, J.; Roche, A. 1973. A parameterization for individual human growth curves. Human Biology 45: 63-80.
• Breusch, T.S.; Pagan, A.R. 1979. A simple test for heteroscedasticity and random coefficient variation. Econometrica 47: 1287-1294.
• Brody, S.; Ragsdale, A.C. 1921. The rate of growth of the dairy cow. The Journal of General Physiology 3: 623-633.
• Draper, N.R.; Smith, H. 1998. Applied Regression Analysis. 3ed. John Wiley, New York, NY, USA.
• Fenner, T.; Levene, M.; Loizou, G. 2013. A bi-logistic growth model for conference registration with an early bird deadline. Central European Journal of Physics 11: 904-909.
• Grossman, M.; Koops, W.J. 1988. Multiphasic analysis of growth curves in chickens. Poultry Science 67: 33-42.
• Koops, W.J. 1986. Multiphasic growth analysis. Growth 50: 169-177.
• Koops, W.J.; Grossman, M.; Michalska, E. 1987. Multiphasic growth curve analysis in mice. Growth 51: 372-382.
• Koops, W.J.; Grossman, M. 1991. Applications of a multiphasic growth function to body composition in pigs. Journal of Animal Science 69: 3265-3273.
• Kurnianto, E.; Shinjo, A.; Suga, D. 1999. Multiphasic analysis of growth curve body weight in mice. Asian-Australasian Journal of Animal Sciences 12: 331-335.
• Mendes, P.N.; Muniz, J.A.; Silva, F.F.; Mazzini, A.R.A. 2008. Difasics logistic model in the study of the growth of Hereford breed females. Ciência Rural 38: 1984-1990 (in Portuguese, with abstract in English).
• Mendes, P.N.; Muniz, J.A.; Silva, F.F.; Mazzini, A.R.A.; Silva, N.A.M. 2009. Analysis of the difasics growth curve of Hereford females by the Gompertz non-linear function. Ciência Animal Brasileira 10: 454-461 (in Portuguese, with abstract in English).
• Mischan, M.M.; Pinho, S.Z.; Carvalho, L.R. 2011. Determination of a point sufficiently close to the asymptote in nonlinear growth functions. Scientia Agricola 68: 109-114.
• Narinc, D.; Karaman, E.; Firat, M.Z.; Aksoy, T. 2010. Comparison of non-linear growth models to describe the growth in Japanese Quail. Journal of Animal and Veterinary Advances 9: 1961-1966.
• Nešetřilová, H. 2005. Multiphasic growth models for cattle. Czech Journal of Animal Science 50: 347-354.
• Özkan, M. 2004. Diphasic analysis of growth in Japanese quail. Asian-Australasian Journal of Animal Sciences 17: 1281-1285.
• Passos, J.R.S.; Pinho, S.Z.; Carvalho, L.R.; Mischan, M.M. 2012. Critical points in logistic growth curves and treatment comparisons. Scientia Agricola 69: 308-312.
• Portz, L.; Dias, C.T.S.; Cyrino, J.E.P. 2000. A broken-line model to fit fish nutrition requirements. Scientia Agricola 57: 601-607.
• Robbins, K.R.; Saxton, A.M.; Southern, L.L. 2006. Estimation of nutrient requirements using broken-line regression analysis. Journal of Animal Science 84: E155-E165.
• Schinckel, A.P.; Craig, B.A. 2002. Evaluation of alternative nonlinear mixed effects models of swine growth. The Professional Animal Scientist 18: 219-226.
• Seber, G.A.F.; Wild, C.J.1989. Nonlinear Regression. John Wiley, New York, NY, USA.
• Sheehy, J.E.; Mitchell, P.L.; Ferrer, A.B. 2004. Bi-phasic growth patterns in rice. Annals of Botany 94: 811-817.

Publication Dates

• Publication in this collection
  June 2015

History

• Received
  11 July 2014
• Accepted
  29 Nov 2014