Glycemic response and health—a systematic review and
meta-analysis: relations between dietary glycemic properties and
health outcomes1–5
Geoffrey Livesey, Richard Taylor, Toine Hulshof, and John Howlett
ABSTRACT
Background: Reduction of dietary glycemic response has been
proposed as a means of reducing the risk of diabetes and coronary
heart disease. The impact of glycemic response on markers of health
remains to be elucidated.
Objective: We assessed the evidence relating the glycemic impact
of foods to measures relevant for health maintenance and manage-
ment of disease.
Design: This was a systematic review and synthesis of interven-
tional evidence from literature reported on glycemic index and mark-
ers of health through the use of meta-analyses and meta-regression
models.
Results: Data from 45 relevant publications were found to January
2005. Lower glycemic index (GI) diets reduced both fasting blood
glucose and glycated proteins independently of variance in available
and unavailable carbohydrate intakes. Elevated unavailable carbo-
hydrate added to improvements in both blood glucose and glycated
protein control. These effects were greater in persons with poor
fasting blood glucose control. No effects were seen on fasting insulin
쏝100 pmol/L; above this, study numbers were few but consistent
with prevention of hyperinsulinemia in some but not all overweight
persons. Insulin sensitivity according to a variety of measurement
methods was improved by lower GI, higher unavailable carbohy-
drate interventions in persons with type 2 diabetes, in overweight and
obese persons, and in all studies combined. Fasting triacylglycerol in
addition to body weight reduction related more to glycemic load than
to GI. Glycemic load reduction by 쏜17 g glucose equivalents/d was
associated with reduced body weight.
Conclusions: Consumption of reduced glycemic response diets are
followed by favorable changes in the health markers examined. The
case for the use of such diets looks compelling. Unavailable carbo-
hydrate intake is equally important. Am J Clin Nutr 2008;
87(suppl):258S– 68S.
KEY WORDS Carbohydrate, glycemic response, glycemic in-
dex, glycemic load, meta-analysis, fasting blood glucose, glycated
proteins, fasting insulin, insulin sensitivity, fasting triacylglycerols
INTRODUCTION
Foods vary in their ability to provoke a postprandial glycemic
response in humans. This response has been quantified in various
ways, including the glycemic index (GI) (1), the relative GI (2),
the glycemic load (GL) (3), and others such as the glycemic
glucose equivalent (4) or the equivalent GL (5). For the present,
258S Am J Clin Nutr 2008;87(suppl):258S– 68S. Printed in USA. © 2008 American Society for Nutrition
information about GI and GL is available for a wide variety of
foods (6), and such information has formed the basis of 45 rel-
evant intervention studies in humans to ascertain whether foods
with a low impact on blood glucose also have a high impact on
disease risk reduction (7). Possibly central to all these approaches
is GL, which may be indexed to carbohydrate or to food weight
or to serving size for the purpose of describing the food and
Downloaded from www.ajcn.org by on January 2, 2009
estimating cumulative intakes (8).
Although information from intervention studies has been re-
viewed previously (7, 9 –12), none of those reviews focused on
the contribution of GL versus GI and the role of unavailable
carbohydrate in studies on reduced GI, or why GI appears not to
show dose-dependent effects on health risk factors, issues that
are addressed at present. Such information is critical to an un-
derstanding of whether interventions to reduce the glycemic im-
pact of the diet might be useful either as a treatment modality or
as a public health measure. It is already clear from epidemiologic
studies in the United States and Europe that the intake of both
unavailable and available carbohydrate of low GI or low GL each
may have a role in prevention of metabolic disease. For example,
type 2 diabetes (3, 13, 14), coronary heart disease (2), and health
risk markers for both diabetes and heart disease (15, 16) each
associate statistically with the glycemic impact of foods (GI or
GL). Although epidemiology can provide ideas about possible
public health strategy and treatment plans, interventional studies
are essential to proving whether such ideas work in practice and
to uncovering any unforeseen influences limiting utility of the
approach.
1
From Independent Nutrition Logic, Wymondham, Norfolk, United
Kingdom (GL and RT); Kellogg Europe, Den Bosch, The Netherlands (TH);
and Wembley Park, Middlesex, United Kingdom (JH).
2
Presented at an ILSI Europe workshop “Glycemic Response and
Health,” held in Nice, France, on 6-8 December 2006.
3
The review was commissioned by the Dietary Carbohydrates Task Force
of the European Branch of the International Life Sciences Institute (ILSI
Europe) and was funded by industry members Cerestar, Coca-Cola, Danisco,
Groupe Danone, Kellogg, Kraft Foods, National Starch, Nestlé, RHM Tech-
nology, Royal Cosun, Südzucker, Tate & Lyle Speciality Sweeteners, and
Unilever. The opinions expressed herein are those of the authors(s) and do not
necessarily represent the view of ILSI or ILSI Europe.
4
Address reprint requests to ILSI Europe. E-mail: publications
@ilsieurope.be.
5
Address correspondence to G Livesey, Independent Nutrition Logic,
Pealerswell House, 21 Bellrope Lane, Wymondham, Norfolk, NR18 0QX
United Kingdom. E-mail: glivesey@inlogic.co.uk.
 FOOD, GLYCEMIA, AND HEALTH RISK
A previous article showed that studies that lower GI as a
treatment plan may also reduce the intakes of available carbo-
hydrate and metabolizable energy significantly, though without
elevating fat intake (7). Low GI diets also may elevate the intake
of unavailable carbohydrate and protein to variable extents (7).
We therefore implement meta-regression as well as meta-
analysis as one of our investigative tools to assess what dietary
factors are important in affecting health risk markers.
From a user perspective, meta-analysis and meta-regression
call for somewhat different approaches. The latter requires more
data, but has the potential to assess the significance of covariates.
The former aims for datasets that are homogenous, ie without
variation in observations between different studies of the same
‘fixed’ effect (17, 18). However, meta-regression aims for a
heterogeneous dataset and attempts to explain the heterogeneity
(19, 20), which may otherwise appear to be a ‘random’ effect.
Doing so in the present context can provide a broader under-
standing of the factors influencing health risk. This would seem
especially important in nutrition in general where study out-
comes can often vary appreciably from one study to another. The
heterogeneous data set analyzed at present includes persons who
are healthy, glucose intolerant, type 1 or 2 diabetic, at primary
(due to family history) or secondary coronary heart disease risk,
and hyperlipidemic. Among these subject groups were either
normal weight or overweight or obese persons (7). In combining
data from such varied health types it was convenient to consider
health risk markers to represent a continuum from the healthy
state to the diseased state, and this seems to represent the situation
for both diabetes and coronary heart disease (21, 22). In addition
we considered study outcomes by categories: by health type, by
different types of food intake control and by different body
weight bands; normal, overweight, and obese (7).
MATERIALS AND METHODS
The construction of a database comprising data extracted from
45 controlled dietary intervention trials on GI reported in the
literature to January 2005 has been described previously (7).
Information is given there also about the literature search, the
inclusion and exclusion criteria, data extraction, and the calcu-
lation methods.
The database includes observations from 972 participants per
treatment arm of all ages (study group mean ages 10 to 63 y) with
both males (511) and females (461) represented. Participants
were either normal weight (16 studies), overweight (18 studies)
or obese (10 studies) or unclassified by weight (1 study). Similar
numbers of participants took part per treatment arm (770 on the
high and 793 on the low GI treatment). Study participants were
either healthy, ie no diagnosis of disease was evident (13 studies)
or had impaired glucose tolerance (2 studies, duration of impair-
ment unknown) or had type 1 diabetes (7 studies, mean duration
from diagnosis 3 to 16 y, one unknown duration) or had type 2
diabetes (17 studies, mean duration from diagnosis 앒0 to 12.5 y,
7 unknown duration) or were at risk of primary CHD (4 studies,
duration unknown) or secondary CHD (1 study, duration un-
known) or had hyperlipidemia (1 study combining Type II a,
Type II b, and Type III; mean duration from diagnosis 4.7 y).
Studies included participants on medication. Insulin dosage was
reported in all 7 studies with persons with type 1 diabetes, 2 of 17
studies with persons with type 2 diabetes, and in one of one study
with participants at risk of secondary CHD. In all but one study
259S
with persons with type 2 diabetes, subjects received non-insulin
medication for glycemic control (hypoglycemic agents).
Categorization of studies according health type was based on
what authors stated was the study group condition. The same
studies were cross categorized also according to body weight as
normal, overweight or obese, again according to what authors
stated or according to body mass index (bands cut at BMI 쏜 25
or 쏜 30 kg/m2). Interventions were by diet, with intention to
exchange the form of available carbohydrate (high versus low
GI). All studies were free-living, ie no subjects were hospital-
ized, housed in metabolic wards or centers of human nutrition.
Biochemical risk factors extracted were fasting blood glucose,
fasting insulin, glycated proteins (HbA1c and fructosamine and
these combined, glycated albumin and glycated protein), insulin
sensitivity, retrospectively calculated insulin sensitivity by ho-
meostatic model assessment (HOMA %S), calculated pancreatic
B-cell function (HOMA %B), cholesterol (total, LDL and HDL)
and fasting plasma triacylglycerols. Risk factors not reported on
here were either a) complex (HOMA model) and so receives only
brief comment or b) showed evidence of one or more factors
confounding results of simple meta-analyses (total and fractions
Downloaded from www.ajcn.org by on January 2, 2009
of cholesterol) or c) were too few to analyze separately (glycated
albumin and glycated protein combined). At this time these data
simply remain to be more fully assessed.
Body weight was the only constitutional risk factor extracted.
Dietary risk or nutritional factors extracted were metabolizable
energy, fat, available carbohydrate, unavailable carbohydrate,
and protein intake, together with potential risk factors GI and the
calculated GL. The last was calculated as GI multiplied by avail-
able carbohydrate intake and was expressed as g. glucose equiv-
alents. Duration of treatment was extracted either as a continuous
variable (weeks) or as a categorical variable by treatment dura-
tion 쏝 or 욷 12 wk.
Data were subjected to random effects meta-analyses and
meta-regressions using meta and metareg the latter using re-
stricted maximum likelihood (REML) in Stata 9.2 SE (Stata-
Corp, TX) according to Cochrane guidelines (17). Studies are
weighted by inverse variance. Analyses are reported as either
random effects or when variation between studies was zero they
are reported as fixed effects. Computation of the study effect and
dependent SE were as described previously (7) for inputting as
the mean effect (Ø) and the SEM effect (seØ) study-by-study (in
STATA terminology). For comparison of treatments we use
methods difference versus methods average (23). Discussion of
this and other approaches can be found elsewhere together with
developments in multivariate meta-analyses (20). While the be-
havior of univariate meta-regression allows valid inferences
given valid datasets, the validity and behavior of bivariate meta-
regression has only recently been investigated and recommended
over multiple uses of univariate analyses (19). Bivariate meta-
regression is reported to inflate heterogeneity without systematic
bias in the coefficients; this tends to widen the CIs and provide
conservative estimates of significant effects. It seems reasonable
for the present to make a similar assumption about multivariate
meta-regression models in general. We assume ‘measurement
error’ is not a cause of bias. This appears reasonable when in-
vestigating relations with dietary variables that are highly heter-
ogeneous, as in the present dataset (7). However, failure of this
assumption would lead to under strength regression coefficients
260S LIVESEY ET AL

and conservative estimates of the statistical significance of re-

sults. Abbreviations used in the results are provided in the foot-
notes to Table 1.

RESULTS

Fasting blood glucose concentrations

Thirty-six studies reported fasting blood glucose concentra-
tions in either venous plasma or capillary blood and achieved
reductions in GI ranging from 4 to 32, with accompanying re-
ductions in GL ranging from 6 to 134 g glucose equivalents.
Study durations ranged from 2 to 26 wk.
Subjects were reported as normal healthy (8 studies), glucose
intolerant (2 studies), type 1diabetic (4 studies), type 2 diabetic
(16 studies), type 1 and type 2 diabetic observations combined (1
study) and at risk of primary (4 studies) or secondary coronary FIGURE 1. Difference in fasting glucose concentration after treatment as
a function of difference in the average fasting glucose concentration after
heart disease (1 study). For analysis, fasting venous plasma glu- treatment with lower glycemic, variably higher unavailable carbohydrate
cose concentrations were converted to the equivalent capillary diets. Bubble sizes are relative 公n weights for 36 end of study results and 14
intermediate results, totaling 50.

TABLE 1

Downloaded from www.ajcn.org by on January 2, 2009

Difference in fasting blood glucose as a function of the average fasting blood glucose concentrations (fasting capillary blood glucose ҃
glucose at the end of treatment with lower glycemic index (GI) or – 0.61 ѿ 0.94 x fasting venous plasma glucose (24)).
glycemic load (GL), variably higher unavailable carbohydrate diets1 The totality of evidence on fasting blood glucose (Figure 1)
Fall in fasting
includes intermediate observations made at timed intervals up to
glucose2,3 (mmol/L the end of the treatment period in addition to the end of treatment
per mmol/L) observations. The figure is plotted according to the convention
for methods comparison (difference plotted against average).
Meanslope1 SEslope1 P ⬎ 兩kh-t兩 k Visual inspection shows no overall treatment difference in fast-
ing blood glucose concentration to be evident when the study
Before adjustment 2
Ҁ0.3 0.12 0.026 18
After adjustment for: population mean is approx. 5 mmol/L. When above 5 mmol/L
⌬ Carbohydrate (g/d)4 Ҁ0.34 0.13 0.021 16 (22 studies) the lower glycemic treatment outcome departed in-
⌬ Energy (kJ/d)4 Ҁ0.33 0.13 0.027 16 creasingly from that on the higher glycemic treatment, with the
⌬ GL (g eq/d)4 Ҁ0.34 0.12 0.017 16 difference being greatest for the highest average of treatment
⌬ Fat (g/d)4 Ҁ0.32 0.12 0.024 16 means. After omitting the intermediate observations to obtain
⌬ GI (%glucose)4 Ҁ0.30 0.12 0.029 18 independent information, meta-regression revealed a significant
⌬ Protein (g/d)4 Ҁ0.32 0.13 0.030 15 relation (slope Ҁ0.30 SE 0.10 ⌬mmol/L per mmol/L; P쏝 兩kh-t兩
⌬ Unavailable carbohydrate Ҁ0.26 0.14 0.085 15 ҃ 0.010, df 20, Tau 0.6 mmol/L). Further omitting of observa-
(g/d)4
tions so that all studies had fasting blood glucose 쏜 5 mmol/L and
1
Results for high minus low GI or GL diet treatments before and after all study participants were fed diets intended at energy mainte-
adjustment for treatment differences in potential covariates. All study groups nance resulted in 18 studies remaining for which meta-regression
were at near maintenance and had fasting blood glucose 쏜 5 mmol/L. FBG, again indicates a significant relation (Figure 2; slope Ҁ0.3 SE
fasting blood glucose; ⌬, difference, here in treatment means; x៮ , bar indicat- 0.12 ⌬mmol/L per mmol/L, P 쏝 兩kh-t兩 ҃ 0.026, df 16, Tau 0.5
ing a mean or average, here the average of treatment means; Meanslope n and
mmol/L).
Slopen, refer to the regression coefficients, and SEslope n the corresponding
It has been reported previously (7) that among these studies
Knapp & Hartung’s estimates of the SE; SEresidual, residual SE (ie, the SE of
the fitted values also known as the within-study SE); Tau, the between-study differences occur between treatments with respect to intakes of
SE; P ⬎ 兩X兩, the level of statistical significance of Tau based on the energy, protein, and available and unavailable carbohydrate, al-
likelihood-ratio test of Tau2 ҃ 0; P ⬎ 兩kh-t兩, level of statistical significance though not fat. When these dietary components, together with
based on the t test by using Knapp & Hartung’s estimate of SE; k, number of GI and GL, are used as covariates (Table 1), the regression
studies; df, error degrees of freedom. slope remains essentially unaltered at around – 0.3 ⌬mmol/L
2
REML random effects model weighted by inverse variance (df ⫽ k Ҁ per mmol/L (range – 0.26 to – 0.34). Such similar slopes
3): ⌬FBG (mmol/L) ҃ constant ѿ slope1 ҂ FBG (mmol/L) ѿ slope2 ҂ suggest that none of these dietary factors were sufficiently
covariate shown (units shown) 앐 Tau (mmol/L) 앐 SEresidual (mmol/L). The determinant, either uniquely by themselves or collinearly, to
univariate model (before adjustment) was the same as the bivariate model,
displace fasting blood glucose as an over-riding determinant
except for constraining slope2 to zero and df ҃ k ⫺ 2.
3 of the effect size in simple meta-regressions; that is regres-
The univariate model before adjustment was heterogeneous, with a
between-studies error (Tau) of 0.63 mmol/L and remained in the range of sions not including interactive terms.
from 0.58 to 0.61 after invoking one of the possible covariates listed in the It was found that many interventions that intended to lower the
bivariate model. Tau was significant in all analyses (P ⬎ 兩X兩 쏝 0.001). GI of a diet also resulted in increased intakes of unavailable
4
None of the covariates were statistically significant (P ⬎ 兩kh-t兩 쏜 0.2), carbohydrates and varied intake of available carbohydrate caus-
because of either an absence of effect or too narrow a range of values. ing usually a decreased GL. It is important to try to disentangle
 FOOD, GLYCEMIA, AND HEALTH RISK
FIGURE 2. Difference in fasting blood glucose after treatment versus the
average fasting blood glucose concentration after treatment with lower gly-
cemic, variably higher unavailable carbohydrate diets (k ҃ 18, univariate
model Table 1). Bubbles are weights by inverse variance and lines are the
meta-regression line, its SE (—), and 95% CI (- - -).
261S
GI diets and illustrates that the GL is as effective as the unavail-
able carbohydrate in lowering fasting blood glucose. Of the 15
studies in Figure 3 upper, no more than 3 reached 12 wk treatment
duration.
An assessment was made of whether, after accounting for
unavailable carbohydrate intake, the variation remaining in fast-
ing blood glucose reduction attributable to GL could be parti-
tioned between change in GI and change in available carbohy-
drate intake. Changes in GI in the range from Ҁ4 to Ҁ32 in these
studies were accompanied by a range for change in available
carbohydrates intake from ѿ 53 g/d to Ҁ35 g/d with a low
correlation between the 2 dietary measures (adjusted r2, Ҁ0.02);
this allowed a potentially good distinction between them. The
bivariate analysis of change in fasting blood glucose due to
change in GL showed no significant association for available
carbohydrate (P 쏜 兩kh-t兩 ҃ 0.54) while a strong association was
evident for GI (P 쏜 兩kh-t兩 ҃ 0.004). Thus it appears that GL is
better reduced by reduction in GI than by reduction in available
carbohydrate.
Further, observations (Table 2) indicate that the size of the
effects of GL or GI and unavailable carbohydrate is dependent on
the persons consuming the diet, that is their level of blood glucose

Downloaded from www.ajcn.org by on January 2, 2009

the effects of each of these. To understand the separate roles of
reduced GL of foods and increased unavailable carbohydrates
intake on fasting blood glucose, the interaction between these 2
variables and the severity (or impairment) of glycemic control
were studied. Severity (S) is quantified here by excessive fasting
blood glucose (S ҃ FBG – 5 mmol/L). In this more complex
model both interactions were significant (Table 2). Therefore
GL clearly has an effect independently of the unavailable carbo-
hydrate content of the diet. Likewise unavailable carbohydrates
appear to act independently of GL. Both act to control fasting
blood glucose as can be seen in Figure 3 (upper). This helps to
explain a reported variability in the effectiveness of the reduced
control (S, the severity of impairment); this is all important in
enabling any understanding of these studies.
Maximum control of fasting blood glucose was achieved by el-
evation of unavailable carbohydrate intake together with a lowering
of GI. Unavailable carbohydrate had stronger impact than GL on a
per g weight basis but remained approximately equally important
when taken together with the range of intakes (Figure 4 upper).
Glycated proteins
Twenty-eight studies reported on glycated protein concentra-
tions. Measures include glycated albumin, glycated plasma pro-
tein, fructosamine, and HbA1c. Subjects were categorized as

TABLE 2
Differences in fasting blood glucose and fructosamine as a function of the average fasting blood glucose in excess of 5 mmol/L (S), the difference in
unavailable carbohydrate intake (⌬UC), and the glycemic load from the diet (⌬GL)1
⌬Glycemic load or ⌬Glycemic index ҂
⌬Unavailable carbohydrate intake ҂ severity interaction
severity interaction [S 䡠 ⌬GL; U per (mmol/L 䡠 g/d)] or Fraction of variance
[S 䡠 ⌬UC; U per (mmol/L 䡠 g/d)]: ␤1 [S 䡠 ⌬GI; [U per (mmol/L 䡠 %)]: ␤2 due to between-
Between-study study variance:
k Meanslope SEslope P ⬎ 兩kh-t} Meanslope SEslope P ⬎ 兩kh-t} SE (U): Tau I2
⌬Fasting blood glucose (mmol/L)
Unavailable carbohydrate and glycemic load
15 Ҁ0.013 0.005 0.04 0.0034 0.0016 0.05 0.57 0.64
Unavailable carbohydrate and glycemic index
15 Ҁ0.012 0.005 0.03 0.0088 0.0033 0.02 0.46 0.59
⌬Fructosamine (%)
Unavailable carbohydrate and glycemic load
12 Ҁ0.169 0.078 0.055 0.033 0.013 0.026 9.3 0.81
Unavailable carbohydrate and glycemic index
12 Ҁ0.136 0.067 0.097 0.080 0.026 0.012 7.0 0.81
1
All studies (food intake categories) combined. REML random effects regression weighted by inverse variance (df ҃ k Ҁ 2): ⌬Fasting blood glucose
(mmol/L) or ⌬Fructosamine (%) ҃ noconstant ѿ slope ␤1 ⫻ S 䡠 ⌬UC (mmol/L 䡠 g/d) ѿ slope ␤2 ⫻ S 䡠 ⌬GL (mmol/L 䡠 g/d) or ѿ slope ␤1 ⫻ S 䡠 ⌬UC
(mmol/L 䡠 g/d) ѿ slope ␤2 ⫻ S 䡠 ⌬GI (mmol/L 䡠 %) 앐 Tau (mmol/L or %) 앐 SEresidual (mmol/L or %). The same model was used for ⌬fructosamine (%
difference) in place of ⌬fasting blood glucose (mmol/L). The model was constrained to have no effect (noconstant) at a fasting blood glucose concentration
of 5 mmol/L for fasting blood glucose (see Figure 2) and 3.5 mmol/L for fructosamine concentration (see Figure 5) to facilitate convergence consistent
with a consideration that little or no treatment effect is expected at or below this concentration. U, units for the analyte concentration (for change in glucose
this was mmol/L, for change in fructosamine this was % of treatment average). Other abbreviations are as in Table 1.
262S LIVESEY ET AL
FIGURE 3. Upper: Implications of the effect of unavailable carbohydrate
intake and glycemic load for fasting blood glucose concentrations. A REML
meta-regression model (Equation 1) result was obtained interrelating the
change (⌬) in fasting blood glucose to change in both unavailable carbohy-
drate intake (UC, g/d) and glycemic load [GL, g (glucose equivalents)/d] at
the different levels of severity of diabetes (Equation 1). This was recast to
assess the potential implications of unavailable carbohydrate intake and
glycemic load on the fasting blood glucose values (Equation 2). The relative
fasting blood glucose concentrations were estimated independently of sever-
ity S (Equation 3). Thus, estimates (Equation 2) were calculated for all 9 (␣)
combinations of GL and UC shown in the figure. All 9 estimates were
expressed as a fraction of the central values at 15 g UC/d and 150 g (glucose
equivalents)/d GL.
⌬Fasting plasma glucose ⫽ S ⫻ ⌬UC ⫻ ␤1 ⫹ S ⫻ ⌬GL ⫻ ␤2 (1)
Fasting plasma glucose ⫽ S ⫻ 关UC ⫻ ␤1 ⫹ GL ⫻ ␤2兴 (2)
Relative fasting plasma glucose
⫽ [(UC ⫻ ␤1 ⫹ GL ⫻ ␤2)UC҃␣,GL҃␣]/[(UC ⫻ ␤1 ⫹ GL ⫻ ␤2)UC҃15,GL҃150]
(3)
Lower: Implications of the effect of unavailable carbohydrate intake and
glycemic load for plasma fructosamine concentrations. The above procedure
(Equations 1 to 3) was followed by using fructosamine concentrations in
place of fasting blood glucose concentrations. Note that recasting the model
in relative values facilitates comparison of results between unavailable car-
bohydrate and glycemic load and between these for fasting blood glucose and
fructosamine.
healthy (2 studies), at risk of primary coronary heart disease (1
study), impaired glucose tolerance (1 study), type 1 diabetic (7
studies), type 2 diabetic (16 studies) and hyperlipidemic (1
study). Among these studies, some offered observations at inter-
mediate time points providing a combined total of 38 effect
estimates.
Downloaded from www.ajcn.org by on January 2, 2009
FIGURE 4. Upper: Implications of the effect of unavailable carbohydrate
intake and glycemic index for fasting blood glucose concentrations. Lower:
Implications of the effect of unavailable carbohydrate intake and glycemic
index for plasma fructosamine concentrations. In both the upper and lower
figures, the observations and models are the same as in Figure 3, with the use
of glycemic index in place of glycemic load. eq, glucose equivalents.
The totality of evidence shows treatment diets to lower the
blood glycated protein concentrations (Figure 5). The effect is
greater for some persons than others dependent on their level of
glycemic control. This figure combines information on fruc-
tosamine and glycated hemoglobin, observations from the end of
studies, and observations from intermediate time points. Choos-
ing only the independent observations (discarding intermediate
time points), statistical significance of this trend toward greater
effect in persons with poor blood glucose control is achieved both
FIGURE 5. Difference in glycated protein after treatment (⌬T) as a
percentage of the mean fasting blood glucose after treatment (MT) with lower
glycemic index or glycemic load, higher unavailable carbohydrate diets.
Bubble sizes are relative 公n weights for 28 studies and 10 repeats (totaling
38 bubbles). As the result of variation in measurement methods, the treatment
average fasting blood glucose was used in place of average fasting glycated
protein concentrations, and treatment difference is expressed as a percentage
of the treatment average.
 FOOD, GLYCEMIA, AND HEALTH RISK 263S
TABLE 3
Difference in glycated proteins as a function of average fasting blood
glucose at the end of the treatment with lower glycemic index or glycemic
load, variably higher unavailable carbohydrate diets1
Fall in glycated protein
(% per mmol/L)

Measurement Meanslope SEslope P ⬎ 兩kh-t兩 k

Unadjusted
⌬Fructosamine (%) Ҁ2.03 0.37 0.001 12
⌬HbA1c (%) Ҁ0.83 0.42 0.07 17
Combined Ҁ1.43 0.23 0.001 24
Adjusted for half-lives
⌬Fructosamine (%) Ҁ2.85 0.56 0.001 12
⌬HbA1c (%) Ҁ1.55 0.94 0.12 17
Combined Ҁ2.21 0.44 0.001 24
1
FIGURE 6. Difference in fasting insulin concentration after treatment as
All studies (food intake categories) combined. REML random effects a function of average fasting insulin after treatment with lower glycemic,
model weighted by inverse variance (df ⫽ k ⫺ 1): ⌬Glycated protein (%) ҃ variably higher unavailable carbohydrate diets. Bubble sizes are relative 公n
noconstant ѿ slope ҂ [FBG Ҁ 3.5] (mmol/L) 앐 Tau (mmol/L) 앐 SEresidual weight for 18 end of study and 5 intermediate study results, totaling 23.
(mmol/L). The model was constrained to have no effect (noconstant) at a
fasting blood glucose concentration of 3.5 mmol/L to facilitate convergence
consistent with a consideration that little or no treatment effect is expected at
Fasting insulin

Downloaded from www.ajcn.org by on January 2, 2009

or below this concentration (see Figure 5). Tau minimized at a fasting glucose
concentration of 3.5 mmol/L when the SE between studies was 4.5% of the
prevailing glycated protein concentration (P ⬎ 兩X兩 ҃ 0.001). HbA1c, gly-
Eighteen relevant studies reported on fasting insulin concen-
cated hemoglobin. Other abbreviations are as in Table 1. trations. No treatment effects or meta-regression trends were
observed for concentrations 쏝 100 pmol/L in the totality of
evidence (Figure 6) or in end of study results whether for all
studies combined or by health type or body weight band (Table
4). Fasting insulin concentrations are difficult to interpret with-
for the 2 glycated protein types combined and for fructosamine out reference to corresponding fasting blood glucose concentra-
alone; for HbA1c alone, the effect is more probable than not tions (25) and so the absence of change in insulin concentration
(Table 3). The effect appears greater after adjustment for the does not necessarily mean an absence of effect on insulin pro-
half-life of fructosamine (2.5 wk) and HbA1c (4.5 wk). This is duction or effectiveness. Treatment effects on the insulin con-
because many of the studies are shorter in duration than the 3 centration may depend on hyperinsulinemia 쏜 100pmol/L (Fig-
half-lives necessary to avoid dilution of effect by the subjects ure 6, Table 4) which developed on the high GI/GL diets in a
pretreatment diets (ie 쏝12 wk for HbA1c). small number of studies in some overweight or obese persons but
Variability in the effect of the lower GI/GL higher unavailable not all.
carbohydrate diets on fructosamine, as with fasting blood glucose,
depended on the extent to which both the GI (and so GL) and the
unavailable carbohydrate intake varied (Table 2, Figure 4).
Again attempts were made to assess whether after accounting TABLE 4
Summary of fixed and random effects meta-analyses for differences in
for unavailable carbohydrate intake, the remaining variation at-
fasting insulin on treatment with lower glycemic, variably higher
tributable to change in GL could be partitioned between GI and unavailable carbohydrate diets1
available carbohydrate intake. For these studies, the change in GI
was in the range from Ҁ5 to Ҁ31 while the range of change in Combined
available carbohydrates intakes (g/d) was from ѿ 43 to Ҁ40. effect ⌬insulin
(pmol/L)
Although correlation between the 2 dietary measures was high
(r2-adjusted, 0.78) bivariate analysis resolved that any effect of Studies combined x៮ SE P ⬎ 兩z兩 k
available carbohydrate intake was non-significant (P 쏝 兩kh-t兩 ҃
0.54) while GI has significant effect (P 쏝 兩kh-t兩 쏜0.03), a result All with insulin 쏝 100 pmol/L 2.7 3 0.32 12
consistent with observations on fasting blood glucose. Hence All with insulin 쏜 100 pmol/L Ҁ12 18 0.49 6
All nondiabetes with insulin 쏜 100 Ҁ73 23 0.001 22
variation in fructosamine concentrations is contributed to in the
pmol/L2
greater part by variations in unavailable carbohydrate intake and All type 2 diabetes 5.9 6.8 0.38 5
GI (Figure 3 and 4 lower). All overweight or obese nondiabetes 4.6 3.9 0.24 5
Altogether, the observations made imply optimum reduction with insulin 쏝 100 pmol/L
in fasting blood glucose and fructosamine occur with intakes of All overweight or obese nondiabetes Ҁ73 23 0.001 22
unavailable carbohydrate at or above 25g/d (Figure 3 and 4), GL with insulin 쏜 100 pmol/L2
at or below 100g (glucose equivalents) per d (Figure 3) or GI 쏝 1
All meta-analyses converged on fixed effects except for insulin 쏜 100
45 (Figure 4). Of the 15 studies in Figure 3 and 4 on fasting blood pmol/L, which was marginally heterogeneous (Q-test P ⬎ 兩X兩 쏜 0.40). k,
glucose and the 12 studies in Figure 3 and 4 concerning fruc- number of studies.
2
tosamine, no more than 3 reached 12 wk treatment duration. Below the minimum of 3 studies normally required for meta-analysis.
264S LIVESEY ET AL

Insulin sensitivity TABLE 5

Summary of fixed and random effects meta-analysis for differences in
Eighteen relevant studies either reported on insulin sensitivity insulin sensitivity after treatment with lower glycemic index or load,
using a range of different methods, or provided postprandial variably higher unavailable carbohydrate diets1
information from which this could be assessed. Among these
studies, observations were from healthy subjects (5 studies), Combined
effect
glucose intolerant subjects (2 studies), persons with type 1 dia-
⌬sensitivity
betes (1 study), persons with type 2 diabetes (5 studies), and
(% increase):
subjects at risk of primary or secondary coronary heart disease (5
studies). The methods used included the euglycemic hyperinsu- Studies combined x៮ (95% CI) P ⬎ 兩z兩 k
linemic clamp technique, the insulin tolerance test, the frequent
sampling intravenous glucose tolerance test, erythrocyte insulin All studies 20 6 to 33 0.004 18
Health type
binding, and the inverse postprandial homeostatic model assess-
All nondiabetes (ostensibly healthy) 25 5 to 44 0.014 12
ment method (inverse HOMA PP). Some studies reported as- Healthy 22 Ҁ5 to 49 0.11 5
sessments using more than one method. Some provided addi- Impaired glucose tolerance 16 Ҁ4 to 36 0.11 22
tional information from which the inverse HOMA PP was Type 2 diabetes 12 2 to 22 0.014 53
calculable and others describe the sensitivity to insulin of glucose Type 1 diabetes 3 Ҁ28 to 33 0.87 22
consumption by human adipocytes in vitro. Fixed and random 1° and 2° CHD risk 29 Ҁ3 to 63 0.08 5
effects meta-analyses for differences in insulin sensitivity Body weight band4
achieved by intervention expressed as a percentage of average Normal 27 Ҁ15 to 69 0.21 4
insulin sensitivity are summarized by health type, by body Overweight 12 4 to 20 0.004 113
weight band, by method of assessment, and by study duration Obese 28 7 to 49 0.009 3
Overweight and obese together 14 7 to 21 0.001 143
(Table 5). For all 18 studies combined, the lower glycemic,

Downloaded from www.ajcn.org by on January 2, 2009

Methods of assessment
variably higher unavailable carbohydrate intervention results in Euglycemic hyperinsulinemic clamp 7 Ҁ5 to 30 0.54 3
a 20% improvement in insulin sensitivity (P 쏜 兩z兩 ҃ 0.004). HOMA S 13 Ҁ2 to 29 0.10 5
By subjects’ health type, improved insulin sensitivity was Insulin tolerance test 41 4 to 78 0.03 33
statistically significant for 12 studies of nondiabetics combined Frequent sample intravenous 20 Ҁ38 to 77 0.05 12
and for 5 studies on persons with type 2 diabetes combined. glucose tolerance
Results for other health types were consistently increased al- Inverse HOMA PP 30 Ҁ9 to 79 0.03 8
though did not reach statistical significance. Erythrocyte binding 3 Ҁ23 to 28 0.84 22
By body weight band, statistically significant increases in Adipocyte 35 Ҁ20 to 91 0.21 3
insulin sensitivity are apparent for overweight subjects com- Study duration
All studies 쏝 12 wk 20 3 to 35 0.015 145
bined (11 studies), obese subjects combined (3 studies) and both
All studies 욷 12 wk 16 Ҁ4 to 35 0.11 46
these groups combined (14 studies). The mean increase for all
normal weight subjects combined (4 studies) suggests a poten-
1
All meta-analyses were conducted with fixed or random effects (df ⫽
tially sensitive response but it does not achieve statistical signif- k Ҁ 1); data report random effects when the Der Simonian and Laird estimate
icance. of between-studies variance was 쏜0; otherwise, data report on fixed effects
(see footnote 3). Between-studies SEs (Tau) ranged from 8% to 41% of the
All methods of assessing insulin sensitivity yield combined
treatment average insulin sensitivity at the end of study and was statistically
means that are positive toward improved sensitivity, and statis- significant when 쏜15% (Q-test, P ⬎ 兩X兩 쏝 0.05). k, number of studies; CHD,
tically significant in some though not in others (Table 5). Com- coronary heart disease; HOMA S, insulin sensitivity by homeostatic model
parison between methods is hampered by the occurrence of het- assessment; HOMA B, pancreatic ␤-cell function by homeostatic model
erogeneity and the small number of studies. Most observations assessment.
were available for HOMA PP (8 studies) the combined mean for 2
Too few studies (쏝3) were undertaken to assess the between-studies
which indicates a 30% improvement in insulin sensitivity of variance so that random effects were assumed.
3
glucose disposal in the postprandial state. Fixed effects analysis.
4
For all studies and methods combined (Table 5) the mean Bands are based on author descriptions or BMI cutoff at 25 and 30
improvement in studies of 쏝 and 욷 12 wk treatment duration was (kg/m2).
5
Between-studies error 26% (Q-test, P ⬎ 兩X兩 쏝 0.001; significant).
similar at 20% and 16% respectively for a total of 14 and 4 studies 6
Between-studies error 11% (Q-test, P ⬎ 兩X兩 쏝 0.24; NS).
respectively, and with similar 95% CIs (앑0 to 35% each). The
larger number of studies in the short term indicate a significant
effect while over the longer term the effect appears more prob-
able than not.
included 45 effect estimates (Figure 7). There was no clear
evidence for a difference in fasting triacylglycerols following
Fasting plasma triacylglycerols treatment with the lower GI/GL intervention, although triacyl-
Thirty-two studies reported on fasting plasma triacylglycerol glycerols concentrations were reduced among groups with the
concentrations. The observations were for healthy subjects (7 highest concentrations.
studies), persons with type 1 diabetes (4 studies), persons with Meta-regression was used to search for an explanatory factor
type 2 diabetes (13 studies), subjects at risk of primary CHD (4 confounding a possible effect of GL, examining treatment dif-
studies), glucose intolerant subjects (1 study), and hyperlipid- ferences in the intake of protein, energy, available and unavail-
emic groups (1 study with 4 types). Six of the studies provide able carbohydrate, fat and GL as covariates in bivariate meta-
repeated observations at various time points. The total evidence regression (Table 6). Statistically significant effects lowering
 FOOD, GLYCEMIA, AND HEALTH RISK 265S

All normal wt (k = 11)*

All studies (k = 30)*

All over wt (k = 14)

Healthy (k = 7)

T2DM (k = 13)
Obese (k = 5)

T1DM (k = 4)

CHD (k = 4)
0

Fall in FTG (mmol/L) per unit

fall in glycemic load (g eq /d)
-0.1

-0.2

FIGURE 7. Difference in fasting triacylglycerol (FTG) concentrations

after treatment as a function of the average fasting triacylglycerol after
-0.3
treatment with lower glycemic, variably higher unavailable carbohydrate
diets. Bubble sizes show relative 公n weight for 32 end of study and 9
intermediate results, totaling 45.
-0.4
FIGURE 8. Fall in fasting triacylglycerols (FTG) per unit fall in glycemic
load after adjustment for fat intake, by body weight band, and by health type
fasting triacylglycerols were evident for lower GL (dependent on
after adjustment for fat intake. Bars show the meta-regression coefficients

Downloaded from www.ajcn.org by on January 2, 2009

GI) and higher fat intake (independent of GI).
After adjustment for variation in fat intake the effect of GL was
significant (P 쏜 兩kh-t兩 ҃ 0.014) among the 30 studies that in-
formed about both variables (fat and GL). This effect of GL was
significant in all studies combined and in all studies with subjects
in the normal range of body weight (Figure 8). Each health type
(healthy, type 1 and 2 diabetic and CHD risk) and body weight
band (normal, overweight, obese) contributed to varying extends
to the statistical significance found for all studies combined,
although each category alone was without statistically signifi-
cant effect. Over all a 10% fall in fasting triacylglycerols requires
achievement of a drop in GL by 30 to 100g eq./d for those
categories showing large or significant effects (Figure 8).
Body weight
Twenty-three relevant studies reported on body weight and
change from high to low GI diets; of these 19 informed about GI.
Among these studies, observations were on healthy subjects (4
studies), glucose intolerant subjects (1 study), persons with type
1 diabetes (1 study), persons with type 2 diabetes (9 studies), and
and standard errors. Heterogeneity was 11% to 21% of the average of treat-
ment means, although not identified for k 쏝 5. *Statistically significant effect
at P 쏜 兩kh-t兩 울 0.02.
subjects at risk of primary or secondary coronary heart disease (4
studies).
Meta-regression indicates body weight to fall with reduction
in GL and vice versa (Table 7). The trend was statistically
significant for all studies combined and occurred in 2 of the 3
food intake control categories: ad libitum and limited controlled
intake but not in the controlled food intake category. Combining
the ad libitum and limited controlled food intake categories,
reductions in body weight occur when GL is reduced by at
least 17 SE 12 g eq./d (intercept on x-axis for y ҃ 0) and most
consistently when the reduction is by 쏜 42 g. eq./d (95%CI)
(Figure 9).
Among a number of dietary variables, differences in body
weight associates univariately most closely with treatment dif-
ferences in GL, though body weight differences also associate

TABLE 6
Difference in fasting triacylglycerols after treatment as a function of differences in macronutrient intake and glycemic load during treatment with reduced
glycemic index diets1
Covariate effect for all
studies combined

Covariate Units for covariate effect Meanslope1 SE P ⬎ 兩kh-t兩 k

⌬ Protein (g/d) (% per g/d) Ҁ0.19 0.36 0.61 29

⌬ Energy (kJ/d) (% per kJ/d) Ҁ0.002 0.007 0.87 30
⌬ Unavailable carbohydrate (g/d) (% per g/d) Ҁ0.1 0.48 0.84 30
⌬ Available carbohydrate (g/d) (% per g/d) 0.21 0.13 0.13 30
⌬ Fat (g/d) (% per g/d) Ҁ1.08 0.43 0.019 30
⌬ Glycemic load (g eq/d) (% per g eq/d) 0.34 0.10 0.003 30
1
REML random effects model weighted by inverse variance (df ⫽ k Ҁ 3): ⌬FTG (% treatment average) ҃ constant ѿ slope1 ҂ covariate (units shown)
ѿ slope2 ҂ treatment average FTG (mmol/L) 앐 Tau (% treatment average FTG) 앐 SEresidual (% treatment average FTG). Between-studies errors (Tau) were
in the range of from 16% to 20% of the treatment average triacylglycerol (TG) concentration; all were significant (P ⬎ 兩X兩 쏝 0.001). Other abbreviations are
as in Table 1.
266S LIVESEY ET AL

TABLE 7 TABLE 8
Difference in body weight after treatment as a function of difference in Difference in body weight after treatment as a function of difference in
glycemic load during treatment with low glycemic, variably higher dietary attribute during treatment with lower glycemic, variably higher
unavailable carbohydrate diets1 unavailable carbohydrate diets1
Combined effect Combined effect
(g/wk per g eq/d) (g/wk per unit)

Food intake category Meanslope1 SE P ⬎ 兩kh-t兩 k Dietary attribute (units) MeanSlope1 SE P ⬎ 兩kh-t兩 k

Controlled intake Ҁ0.4 1.4 0.8 12 2

⌬ Glycemic load (g eq/d) 2.1 0.7 0.01 19
Limited controlled intake 1.9 0.9 0.052 143 ⌬ Available carbohydrate (g/d) 2.4 1 0.04 20
Ad libitum intake 2.8 1.3 0.11 53 ⌬ Glycemic index (% glucose) 6.9 3.2 0.05 19
Ad libitum ѿ limited controlled 2.1 0.7 0.011 193 ⌬ Metabolizable energy (kJ/d) 0.8 0.4 0.05 20
intakes ⌬ Fat (g/d) 4 4 0.36 20
All categories combined 1.7 0.7 0.017 313 ⌬ Unavailable carbohydrate (g/d) 7 35 0.85 18
⌬ Protein (g/d) 4 4 0.33 19
1
REML random effects model weighted by inverse variance (df ⫽ k Ҁ
2): ⌬Body weight (g/wk) ҃ constant ѿ slope1 ҂ ⌬ glycemic load (g eq/d) 앐 1
All studies were combined from the ad libitum and limited controlled
Tau (g/wk) 앐 SEwithin studies (g/wk). Abbreviations are as in Table 1. food intake categories (ie, studies with controlled food intake were ex-
2
The model yielded fixed effects results because the between-studies cluded). REML random effects model weighted by inverse variance (df ⫽ k
error was zero. Ҁ 2): ⌬Body weight (g/wk) ҃ constant ѿ slope1 ҂ ⌬ dietary attribute (units
3
The between-studies error (Tau) ranged from 52 to 71 g/wk. shown) 앐 Tau (g/wk) 앐 SEwithin studies (g/wk). Abbreviations are as in Table
1. The between-studies error (Tau) ranged from 71 to 110 g/wk.
with treatment differences in available carbohydrate intake, GI

Downloaded from www.ajcn.org by on January 2, 2009

and metabolizable energy intake (Table 8).
DISCUSSION
Low glycemic response diets are proposed as a means to fa-
vorably influence physiologic parameters implicated as markers
for conditions including overweight and obesity, diabetes mel-
litus and risk of coronary heart disease. The present meta-
analyses provide evidence that supports the view that interven-
tion to reduce the diets glycemic impact will favorably affect
several health risk markers.
For individuals with fasting blood glucose concentrations in
excess of 5 mmol/L, there is evidence that fasting blood glucose
is reduced by the consumption of lower GI or GL. Higher un-
available carbohydrate has an effect that is additive to that of
lower GI (and resulting GL), ie both together have optimum
FIGURE 9. Difference in the average rate of change in body weight
during treatment as a function of the difference in glycemic load during the
treatment with low glycemic, variably higher unavailable carbohydrate diets.
Data are combined from the ad libitum and limited controlled food intake
groups. Lines are the regression coefficient and 95% CIs. Labels are first
author and year of publication. eq, glucose equivalents. (For citations and
table of study characteristics, see reference 7.)
effect. Further, the evidence indicates that the effect of consum-
ing lower GI and higher UC diets is greater in absolute units in
persons with poorer control of blood glucose, including persons
with both type 1 and type 2 diabetes.
Similarly, for individuals with fasting blood glucose concen-
trations in excess of 5 mmol/L, there is evidence that both lower
GI (and so lower GL) and higher UC in diets reduce the levels of
glycated proteins. The evidence is stronger for fructosamine than
for glycated hemoglobin. Again, the size of the effect is greater
in persons with poorer control over blood glucose.
Although the data suggest that the consumption of lower GI/
GL, higher unavailable carbohydrate diets lead to a reduction in
fasting insulin concentrations in overweight or obese individuals
who had fasting concentrations 쏜 100 pmol/L, the evidence is
weak due to few studies. In individuals with fasting insulin levels
쏝 100 pmol/L, meta-analysis provides evidence of no significant
effect.
Overall, this review provides evidence that insulin sensitivity
is improved by consumption of lower GI/GL, higher unavailable
carbohydrate diets. Although the increase is evident in all body
weight bands combined, and in nondiabetics, the increase is not
shown to be statistically significant in subjects in the normal
body weight range.
While a simple meta-analysis of interventions with low GI/
GL, variably higher unavailable carbohydrate diets did not pro-
vide a combined mean effect on fasting triacylglycerol concen-
trations, a more sophisticated analysis suggests an apparent
absence of effect is due to a confounding factor. After adjustment
for fat intake unrelated to GI or load (7), the evidence shows
reductions in GL to reduce fasting triacylglycerols.
Over all studies relevant for the purpose taken together, it is
evident that intervention with low GI/GL, high unavailable car-
bohydrate diets do associate significantly with reductions in
body weight. Such evidence is consistent with evidence on re-
ductions in food energy intake (7). The present analysis suggests
a minimal reduction in GL or index is necessary for this effect to
occur (Figure 9 and Table 8 in ref. 7). This could arise either
because a threshold needs to be surpassed to achieve an effect or
 FOOD, GLYCEMIA, AND HEALTH RISK
because dietary advice is imprecisely implemented by the re-
searchers, health professionals, and consumers involved.
Although the evidence generally supports the view that inter-
vention with a low GI/GL and higher unavailable carbohydrate
diet is associated with favorable changes in a number of health
risk markers relevant to persons who are overweight, obese,
diabetic or at risk of coronary heart disease, the evidence also
indicates for some markers that not all subjects respond equally.
For reductions in fasting blood glucose concentrations and gly-
cated proteins, there is evidence that the effects are greatest
among those with poorest glycemic control. For reduction in
fasting blood glucose, the threshold for this effect is at about 5
mmol/L. For glycated proteins the threshold is at about 3.5 glu-
cose mmol/L. Unpublished meta-analysis of the effect of the low
GI sugar fructose indicates a threshold of effect on glycated
hemoglobin at about 4% HbA1c (Livesey and Taylor, unpub-
lished observations, 2007). Hence absence of effect in those with
good glycemic control is not evident. Also, when fasting glucose
is low (쏝5 mmol/L) and study precision is accounted for, then
below median normal fasting blood glucose is elevated to a small
but statistically significant extent by such diets (in which case the
‘threshold’ mentioned above is a pivot point). The evidence for
a small rise in fasting glucose below 5mmol/L comes with evi-
dence of publication bias consistent with hesitance to report such
normalization. While the present data taken together with influ-
ence on insulin sensitivity points toward improved control
among nondiabetics (in addition to persons with diabetes), a role
in disease prevention remains to be established.
The studies reviewed had the intention to treat by reducing the
GI of available carbohydrate ingested. Compliant treatments
show variable elevations in unavailable carbohydrate and protein
intake and reductions in available carbohydrate, metabolizable
energy, and GL when food intake was not firmly controlled (7).
Even when intake was controlled the reductions in GI were ac-
companied by significant reductions in available carbohydrate
intake and so in GL (7). However, among the health markers
examined 3 factors were clearly important, glycaemic index,
available carbohydrate intake and unavailable carbohydrate in-
take. Altogether, improvement in the control of health markers
cannot be said to be clearly due to GI alone.
Overall, considering all the markers of health examined here
together, it is difficult to establish whether GI is stronger than GL,
in part due to their co-linearity (7) and in part because some
markers appear affected by available carbohydrate intake while
others are not. While there is no affirmative evidence that vari-
ation in available carbohydrate intake (앐 50g/d) influences fast-
ing blood glucose or glycated proteins in these studies, there is
evidence that reductions in available carbohydrate intake do ac-
company or play a role alongside GI in the beneficial effects of
lower GL on fasting triacylglycerols (herein), body weight
(herein) and food intake (7). On the other hand, it must also be
considered that limiting harm due to glycaemic load by attempts
to lower available carbohydrate intake will risk elevating harm
from higher total or saturated fat intake. Benefits of reduction in
available carbohydrate intake would be expected only in a con-
text of no increase in total or saturated fat intake. It is evident in
the studies reviewed here that lower available carbohydrate in-
take following marked reductions in glycaemic index were not
accompanied by elevation in fat intake (7). In this context, the
balance favors reduced harm.
267S
The meta-analyses confirm that GI or load has a significantly
stronger relation with glycemic control than does available car-
bohydrate in the present context in which change in fat intake is
minimal. Attaining an optimum diet for health therefore requires
consideration of glycemic impact of foods eaten in preference to
consideration of carbohydrate content alone, each in the context
of a balanced diet meeting nutrient requirements. The extent to
which reduction in GL using ingredient carbohydrates has sim-
ilar effects to those in the studies reviewed requires to be more
fully evaluated. Fructose appears at least equally effective for
HbA1c control (Livesey and Taylor, unpublished observations,
2007) and added unavailable carbohydrate can be effective in
glycemic control (8, 26, 27).
The evidence shows also that, independently of increasing
unavailable carbohydrate intake, reductions in GI (and so GL) do
improve glycemic control. For optimal control over fasting blood
glucose concentrations in persons with diabetes, the evidence
points toward the need for foods that are of low impact on gly-
cemia (independently of fat intake) and foods that are high in
unavailable carbohydrate. Diets that achieve one without the
other would be suboptimal for diabetes control.
Downloaded from www.ajcn.org by on January 2, 2009
The authors thank TH for considerable help with structuring a report
before a workshop in Nice, which provided comments on the preliminary
results, and to G Janecek, University of East Anglia, Norwich, United King-
dom for preliminary discussion of the statistical approach.
The contributions of the authors were as follows—GL and RT: data col-
lection; GL: analysis; JH and GL: writing; and TH: comment. GL and RT had
no financial or personal conflicts of interest. JH is currently advising an
industry group comprising food manufacturers and retailers who are prepar-
ing a submission to the authorities in Europe supporting the case for the use
of glycemic index in the labeling of food products. TH works for Kellogg.
REFERENCES
1. Jenkins DJ, Wolever TM, Taylor RH, et al. Glycemic index of foods: a
physiological basis for carbohydrate exchange. Am J Clin Nutr 1981;
34:362– 6.
2. Brouns F, Bjorck I, Frayn KN, et al. Glycaemic index methodology. Nutr
Res Rev 2005;18:145–71.
3. Liu S, Willett WC, Stampfer MJ, et al. A prospective study of dietary
glycemic load, carbohydrate intake, and risk of coronary heart disease in
US women. Am J Clin Nutr 2000;71:1455– 61.
4. Monro JA. Glycaemic glucose equivalent: combining carbohydrate con-
tent, quantity and glycaemic index of foods for precision in glycaemia
management. Asia Pac J Clin Nutr 2002;11:217–24.
5. Wolever TM, Gibbs AL, Spolar M, Hitchner EV, Heimowitz C. Equiv-
alent glycemic load (EGL): a method for quantifying the glycemic re-
sponses elicited by low carbohydrate foods. Nutr Metab (Lond) 2006;
3:33.
6. Foster-Powell K, Holt SH, Brand-Miller JC. International table of gly-
cemic index and glycemic load values: 2002. Am J Clin Nutr 2002;76:
5–56.
7. Livesey G, Taylor R, Hulshof T, Howlett J. Glycemic response and
health—a systematic review and meta-analysis: the database, study
characteristics, and macronutrient intakes. Am J Clin Nutr 2008;
87(suppl):223S–36S.
8. Livesey G. Low-glycaemic diets and health—implications for obesity.
Proc Nutr Soc 2005;161:105–13.
9. Brand-Miller JC, Hayne S, Petocz P, Colagiuri S. Low-glycemic index
diets in the management of diabetes: a meta-analysis of randomized
controlled trials. Diabetes Care 2003;26:2261–7.
10. Opperman AM, Venter CS, Oosthuizen W, Thompson RL, Vorster HH.
Meta-analysis of the health effects of using the glycaemic index in
meal-planning. Br J Nutr 2004;92:367– 81.
11. Anderson JW, Randles KM, Kendall CW, Jenkins DJ. Carbohydrate and
fiber recommendations for individuals with diabetes: a quantitative as-
sessment and meta-analysis of the evidence. J Am Coll Nutr 2004;23:
5–17.
268S LIVESEY ET AL
12. Kelly S, Frost G, Whittaker V, Summerbell C. Low glycaemic index
diets for coronary heart disease. Cochrane Database Syst Rev 2004:
CD004467.
13. Salmeron J, Ascherio A, Rimm EB, et al. Dietary fiber, glycemic load,
and risk of NIDDM in men. Diabetes Care 1997;20:545–50.
14. Salmeron J, Manson JE, Stampfer MJ, Colditz GA, Wing AL, Willett
WC. Dietary fiber, glycemic load, and risk of non-insulin-dependent
diabetes mellitus in women. JAMA 1997;277:472–7.
15. Frost G, Leeds AA, Dore CJ, Madeiros S, Brading S, Dornhorst A.
Glycaemic index as a determinant of serum HDL-cholesterol concen-
tration. Lancet 1999;353:1045– 8.
16. Liese AD, Gilliard T, Schulz M, D’Agostino RB Jr, Wolever TM. Car-
bohydrate nutrition, glycaemic load, and plasma lipids: the Insulin Re-
sistance Atherosclerosis Study. Eur Heart J 2007;28:80 –7.
17. Deeks J, Higgins JDA. Section 8: analysis and presentation of results.
Cochrane Reviewers’ Handbook 4.2.2. Chichester, United Kingdom:
Wiley and Sons Ltd, 2002.
18. Leandro G. Meta-analysis in medical research. Oxford, United King-
dom: BMJ Books, Blackwell Publishing, 2005.
19. Riley RD, Abrams KR, Sutton AJ, Lambert PC, Thompson JR. Bivariate
random-effects meta-analysis and the estimation of between-study cor-
relation. BMC Med Res Method 2007;7:3doi:10.1186/1471–2288-7–3.
20. Arends LR. Multivariate meta-analysis: modelling the heterogeneity.
PhD thesis. Rotterdam, Netherlands: Erasmus University, 2006. [ISBN
90-9020786-4.]
21. Khaw KT, Wareham N, Luben R, et al. Glycated haemoglobin, diabetes,
and mortality in men in Norfolk cohort of European prospective inves-
tigation of cancer and nutrition (EPIC-Norfolk). BMJ 2001;322:15– 8.
22. Stratton IM, Adler AI, Neil HA, et al. Association of glycaemia with
macrovascular and microvascular complications of type 2 diabetes (UK-
PDS 35): prospective observational study. BMJ 2000;321:405–12.
23. Bland JM, Altman DG. Applying the right statistics: analyses of mea-
surement studies. Ultrasound Obstet Gynecol 2003;22:85–93.
24. Colagiuri S, Sandbaek A, Carstensen B, et al. Comparability of venous
and capillary glucose measurements in blood. Diabet Med 2003;20:
953– 6.
25. Wallace TM, Levy JC, Matthews DR. Use and abuse of HOMA mod-
eling. Diabetes Care 2004;27:1487–95.
26. Livesey G. Glycaemic responses and toleration. In: Mitchell H, ed.
Sweeteners and sugar alternatives in food technology. Oxford, United
Kingdom: Blackwell, 2006:1–18.
27. Robertson MD, Bickerton AS, Dennis AL, Vidal H, Frayn KN. Insulin-
sensitizing effects of dietary resistant starch and effects on skeletal
muscle and adipose tissue metabolism. Am J Clin Nutr 2005;82:559 – 67.
Downloaded from www.ajcn.org by on January 2, 2009