Search strategy

A systematic search was conducted from a range of electronic databases, including The Cochrane Library, EMBASE, PsycINFO, OVID and PubMED and complemented by a manual search with bibliographic cross-referencing. Key words used to identify the studies were: magnetic resonance imaging, MRI, bipolar disorder, mania, mood disorders and schizophrenia. Studies were included if they presented original data and were published by March 2008, compared individuals with schizophrenia, bipolar disorder and/or healthy controls, reported volumetric measures of brain areas according to the international system of units (SI units) as means and standard deviations. If standard deviations were missing from the published articles, these were conservatively estimated from the largest standard deviation of other studies that measured the same structure in the same volumetric units. Studies that included participants with unipolar depression were included provided participants with bipolar disorder made up more than 79% of the sample. Researchers were contacted if this information was not readily available. Studies were excluded if data were subsumed in more recent larger studies. Information systematically extracted from the studies included diagnosis according to diagnostic criteria, volumetric measurements and number of participants essential to calculate effect sizes, but also a number of potentially critical confounding variables. These included demographics (age, gender), illness variables (age at onset, duration of illness, presence of euthymia, medication, chronicity of the condition), year of publication, magnetic field strength of the scanner and slice thickness. Conference abstracts and letters were included only if there were no other publications from the same study that had been published in full as peer reviewed articles. Where a single study was published in several journal articles, the article reporting the largest group size for that volume of interest was used. When multiple publications were identified, disagreement was resolved by consensus between the authors. Studies were excluded when there was a comorbid diagnosis of intellectual disability, chromosomal or genetic disorder. Studies were included irrespective of slice thickness, although these factors were recorded as potential sources of heterogeneity. Studies were not included when the control group were genetically related to affected probands.

Statistical analysis

Statistical analysis was conducted using STATA 8.0 for Windows (Stata Corp, College Station, Texas) supplemented by Metan software (www.stata.com/stb/stb44/sbe24/metan.hlp; downloadable from the Centre for Statistics in Medicine, Oxford, UK). Standardised mean differences were calculated using Cohen's d statistic.

Random effects analyses ^{Reference DerSimonian and Laird11} were used throughout to weight each study. The presence of heterogeneity was tested using the Q-test and its magnitude estimated using I ^{Reference Wright, Rabe-Hesketh, Woodruff, David, Murray and Bullmore2} and can be interpreted as the proportion of variance in effect size due to heterogeneity. ^{Reference Higgins, Thompson, Deeks and Altman12} When the Q-test was significant, we used a Galbraith plot to identify those studies contributing the greatest amount to that heterogeneity, in order to investigate potential causes. Publication bias, which describes the tendency of small studies to report large effect sizes, was examined using the Egger's test. ^{Reference Egger, Davey Smith, Schneider and Minder13} The significance level was set at P<0.05.

To further investigate causes for heterogeneity, meta-regression analyses were performed for the following variables: age at onset, duration of illness, presence of euthymia, chronicity, mean age, scanner strength, slice thickness, year of publication, and current medication, including mood stabilisers, antipsychotics and antidepressants. The STATA program ‘metareg.ado’ was used throughout and the REML (restricted maximum likelihood) method used to estimate the model parameters.

Systematic search

Seventy-two reports ^{Reference Ahn, Breeze, Makris, Kennedy, Hodge and Herbert14–Reference Zipursky, Bury, Langevin, Wortzman and Katz85} from 180 studies identified met criteria for inclusion in the meta-analysis and are described in detail in online Table DS1. Sixty-five compared people with bipolar disorder with controls. Of these 65, 18 articles included a further comparison group of individuals diagnosed with schizophrenia, ^{Reference Altshuler, Bartzokis, Grieder, Curran and Mintz15,Reference Altshuler, Bartzokis, Grieder, Curran, Jimenez and Leight16,Reference Dasari, Friedman, Jesberger, Stuve, Findling and Swales36,Reference Harvey, Ron, Baker and Murray46,Reference Hirayasu, McCarley, Salisbury, Tanaka, Kwon and Frumin49,Reference Kasai, Shenton, Salisbury, Hirayasu, Lee and Ciszewski52–Reference Kasai, Shenton, Salisbury, Hirayasu, Onitsuka and Spencer54,Reference Lim, Rosenbloom, Faustman, Sullivan and Pfefferbaum56,Reference McDonald, Marshall, Sham, Bullmore, Schulze and Chapple58,Reference McIntosh, Forrester, Lawrie, Byrne, Harper and Kestelman59,Reference Pearlson, Barta, Powers, Menon, Richards and Aylward65,Reference Roy, Zipursky, Saint-Cyr, Bury, Langevin and Seeman68,Reference Schlaepfer, Harris, Tien, Peng, Lee and Federman75,Reference Strasser, Lilyestrom, Ashby, Honeycutt, Schretlen and Pulver80–Reference Swayze, Andreasen, Alliger, Yuh and Ehrhardt82,Reference Zipursky, Bury, Langevin, Wortzman and Katz85} one article added a third comparison group, of schizoaffective disorder ^{Reference Velakoulis, Wood, Wong, McGorry, Yung and Phillips83} and one compared with schizoaffective disorder but not with schizophrenia. ^{Reference Getz, DelBello, Fleck, Zimmerman, Schwiers and Strakowski45} Only two reports compared bipolar disorder with schizophrenia without the comparator of healthy controls. ^{Reference Frazier, Hodge, Breeze, Giuliano, Terry and Moore44,Reference Rossi, Stratta, Di Michele, Gallucci, Splendiani and de Cataldo66}

Thirty-six reports did not meet inclusion criteria mainly because authors used alternative or qualitative measurements of brain areas, reports were superseded by subsequent inclusive publications, volumes were not retrievable or there were fewer than three studies available for a given brain region. Years of publication ranged from 1990 to 2008. Studies were exclusively published in English. Studies used comparable diagnostic criteria and tested a total of 1823 participants with bipolar disorder, 670 with schizophrenia, 29 with schizoaffective disorder, 106 with unipolar depression (not included in the analysis) and 1940 healthy controls. This number includes duplicate publications investigating different brain regions. Same samples were considered only once in each individual analysis.

Studies generally considered individuals with recurrent episodes of illness who were treated with one or more mood stabilisers. Basic demographic characteristics were generally well reported but clinical details such as medication status, number of episodes, duration of illness, age at first presentation and illness subtype were not. Most studies included male and female participants but only a few offered separated analyses according to gender. There were 926 male participants with bipolar disorder, equivalent to 51% of the total sample. Age ranged from 10.6 to 58.8 years with a mean of 29.4 (s.d. = 11.8).

Bipolar disorder in comparison with healthy controls

Comparisons of regional brain volumes between people with bipolar disorder and healthy controls are described in detail in online Table DS2. There was a small but significant reduction in whole brain volume in bipolar disorder (n = 661) compared with healthy controls (n = 723) with an estimated standardised effect size of −0.15 (95% CI −0.27 to −0.02) and without significant heterogeneity (I ² = 0.23, P = 0.15) or publication bias (Egger's P = 0.9). Left and right lateral ventricles were significantly enlarged in bipolar disorder (n = 157, healthy controls n = 179) with an effect size estimate of 0.27 (95% CI 0.05–0.49) with no heterogeneity (I ² = 0, P = 0.7) or publication bias (Egger's P = 0.07). A larger number of studies (n = 11) measured lateral ventricles separately, suggesting a significant contribution of the left but not the right lateral ventricle, with no significant heterogeneity or publication bias. An analysis of five studies that measured the volume of the globus pallidus bilaterally (n = 135, healthy controls n = 106) showed a significantly increased volume in participants with bipolar disorder (estimate 0.57, 95% CI 0.03–1.11) with a significant level of heterogeneity (I ² = 0.74, P = 0.004) and publication bias (Egger's P = 0.02). This effect was not evident in the analysis of the three studies which measured the volume of the left and right globi pallidi separately (Table DS2).

Bipolar disorder in comparison with schizophrenia

Table DS2 shows that in comparison with schizophrenia, people with bipolar disorder showed an increased right amygdala volume (n = 115 v. n = 200), effect size estimate 0.47, 95% CI 0.21–0.73, I ² = 0, Egger's P = 0.59. Lateral ventricles in bipolar disorder appeared bilaterally smaller than in schizophrenia (n = 126 v. n = 158). The effect size estimate for the left lateral ventricle was: −0.35, 95% CI −0.59 to −0.11, I ² = 0.007, Egger's P = 0.11; for right lateral ventricle it was: −0.26, 95% CI −0.49 to −0.02, I ² = 0; Egger's P = 0.06. This effect was not present in the analysis of the three studies that measured the cumulative volume of left and right lateral ventricles (Table DS2).

Heterogeneity and publication bias

Heterogeneity and publication bias were not detected in structures that showed significant volumetric differences, except for whole brain grey matter and globus pallidus in the comparison of bipolar participants with healthy controls. Heterogeneity was, however, detected in a larger number of structures (as shown in Table DS2) and it is likely that methodological differences and clinical sample variation (including the effect of medication) are accountable for such an effect. With the limitation of selective reporting of relevant variables across the included studies we systematically investigated causes of heterogeneity in multiple meta-regression analyses. The results are displayed in Tables 1 and 2 and the main findings described below.

Table 1

Meta-regression analyses of the effect of illness characteristics and demographic variables on volumetric differences of brain structures that showed statistically significant levels of heterogeneity.^a Results are expressed as regression coefficient (C), Z-statistic and P

	Mean age, years			Duration of illness, years			Euthymia			Age at onset, years			Chronicity			Male participants
Structure	C	Z	P	C	Z	P	C	Z	P	C	Z	P	C	Z	P	C	Z	P
Bipolar disorder in comparison with healthy controls
Whole brain (grey matter)	0.0045	0.21	0.83	0.09	2.38	0.02	0.66	1.08	0.28	-0.23	-1.29	0.2	NA			-1.19	-1.49	0.14
Amygdalae	0.033	1.42	0.16	0.05	1.04	0.30	0.17	0.20	0.84	0.006	1.13	0.26	NA			2.16	1.64	0.1
Left amygdala	0.032	1.91	0.056	0.08	2.21	0.03	-0.01	-0.02	0.98	0.022	0.74	0.46	0.04	0.07	0.94	0.08	0.03	1
Right amygdala	0.025	1.49	0.14	0.06	1.07	0.30	-0.16	-0.30	0.76	0.02	0.38	0.7	0.11	0.21	0.83	-1.32	-0.53	0.6
Hippocampi	-0.017	-1.53	0.13	0.024	0.75	0.45	0.66	1.53	0.13	0.063	2.96	0.003	0.12	0.32	0.75	0.63	0.65	0.51
Globi pallidi	0.022	0.64	0.52	0.15	0.91	0.36	-0.76	-0.80	0.42	0.02	0.67	0.5	NA			6.23	1.55	0.12
Right globus pallidus	-0.05	-1.16	0.25	-0.08	-2.39	0.02	-2.07	-2.76	0.006	NA			NA			6.85	1.08	0.28
Left anterior cingulate cortex	0.13	1.58	0.12	0.028	0.83	0.41	-8.82	-1.66	0.1	NA			NA			-0.04	-1.23	0.22
Thalami	0.004	0.34	0.73	0.28	1.98	0.047	-0.13	-0.26	0.8	NA			NA			2.0	1.21	0.228
Temporal lobes	0.03	2.92	0.004	0.054	2.87	0.004	0.89	0.75	0.45	NA			NA			0.54	0.48	0.63
Left temporal lobe	0.012	0.70	0.48	0.07	1.09	0.3	-1.34	-1.46	0.14	-0.0071	-0.09	0.92	NA			0.74	0.97	0.33
Right temporal lobe	-0.012	-0.66	0.51	0.076	2.22	0.03	0.83	0.50	0.62	0.091	1.04	0.3	NA			0.90	1.12	0.3
Pituitary gland	-0.014	-1.58	0.11	NA			-1.09	-1.65	0.1	NA			NA			-5.80	-2.37	0.02
Bipolar disorder in comparison with schizophrenia
Intracranial volume	0.072	0.97	0.33	NA			NA			NA			NA			2.66	1.32	0.19
Whole brain (white matter)	0.15	1.64	0.1	NA			NA			NA			NA			1.94	2.55	0.011
Left amygdala	0.016	0.58	0.6	NA			0.98	2.10	0.035	NA			NA			2.30	1.41	0.16
Left hippocampus	0.043	4.29	<0.001	NA			1.62	3.43	0.001	0.043	1.28	0.2	NA			1.3	0.97	0.33
Right hippocampus	0.039	4.04	<0.001	NA			1.73	3.45	0.001	0.04	1.19	0.23	NA			1.58	1.25	0.21

Table 2

Meta-regression analyses of the effect of medication, scanner strength and year of publication on volumetric differences of brain structures that showed statistically significant levels of heterogeneity.^a Results are expressed as regression coefficient (C), Z-statistic and P

	Mood stabilisers			Antipsychotics			Antidepressants			Scanner strength, T			Slice thickness, mm			Year of publication, years
Structure	C	Z	P	C	Z	P	C	Z	P	C	Z	P	C	Z	P	C	Z	P
Bipolar disorder in comparison with healthy controls
Whole brain (grey matter)	-0.40	-0.41	0.68	-1.13	-2.46	0.014	-0.93	-1.28	0.2	0.22	0.98	0.32	0.006	0.05	1	0.005	0.13	0.9
Amygdalae	2.28	2.33	0.02	-0.8	-0.48	0.63	-2.02	-1.64	0.1	-0.51	-0.83	0.41	-0.16	-0.31	0.75	-0.29	-3.31	0.001
Left amygdala	0.19	0.14	0.89	-0.63	-1.26	0.21	-1.2	-1.83	0.067	-0.4	-0.98	0.33	-0.04	-0.43	0.70	-0.03	-0.71	0.50
Right amygdala	-0.19	-0.23	0.82	-1.28	-2.38	0.017	-1.67	-2.08	0.038	-0.50	-1.49	0.13	0.01	0.13	0.90	-0.05	-1.31	0.20
Hippocampi	0.98	1.31	0.19	-1.05	-1.85	0.065	0.18	0.20	0.84	0.21	0.53	0.60	0.03	0.11	0.91	-0.06	-0.72	0.50
Globi pallidi	2.32	0.70	0.48	0.67	0.58	0.56	-0.46	-0.44	0.66	NA			-0.16	-0.72	0.47	0.022	0.64	0.52
Right globus pallidus	3.22	2.72	0.007	1.48	1.30	0.19	2.3	0.40	0.69	NA			-1.29	-0.78	0.43	-0.006	-0.03	1.0
Left anterior cingulate cortex	-7.91	-0.93	0.35	0.25	0.04	0.97	-2.59	-0.69	0.49	-3.28	-0.47	0.64	NA			0.33	0.42	0.70
Thalami	1.27	1.08	0.28	-0.86	-1.82	0.069	-0.77	-1.23	0.22	-0.07	-0.21	0.83	-0.46	-2.38	0.02	-0.15	-2.92	0.003
Temporal lobes	2.23	2.69	0.007	-1.62	-1.06	0.3	-1.48	-0.73	0.46	-0.29	-0.61	0.54	-0.07	-0.29	0.8	-0.11	-1.59	0.11
Left temporal lobe	-1.76	-0.37	0.71	NA			-0.32	-0.37	0.71	0.09	0.29	0.77	0.012	0.25	0.81	-0.044	-1.71	0.09
Right temporal lobe	-2.15	-0.15	0.88	NA			-0.39	-0.15	0.88	0.24	0.74	0.5	-0.053	-1.03	0.3	-0.044	-1.55	0.12
Pituitary gland	-1.58	-0.98	0.32	NA				NA		NA			-3.65	-0.97	0.33	0.18	1.39	0.17
Bipolar disorder in comparison with schizophrenia
Intracranial volume	NA			NA			NA			NA			NA			0.16	0.54	0.6
Whole brain (white matter)	NA			NA			NA			NA			NA			0.016	0.58	0.56
Left amygdala	NA			NA			NA			NA			NA			0.0076	0.009	0.93
Left hippocampus	NA			NA			NA			0.44	0.69	0.49	-0.34	-1.63	0.1	-0.024	-0.58	0.56
Right hippocampus	NA			NA			NA			0.44	0.73	0.5	-0.062	-1	0.31	-0.02	-0.45	0.65

Bipolar disorder in comparison with healthy controls

In the comparison between people with bipolar disorder and healthy controls significant heterogeneity was found for several regions. The effect size for whole grey matter volume correlated with duration of illness (C = 0.09, Z = 2.38, P = 0.02) and use of antipsychotic medication (C = −1.13, Z = −2.46, P = 0.014), suggesting larger bipolar v. control differences in people with longer durations of illness and larger reductions with use of antipsychotics, perhaps a proxy for a more severe illness. The differences in the right globus pallidus correlated with length of illness (C = −0.08, Z = −2.39, P = 0.02), percentage of participants with euthymia (C = −2.07, Z = −2.76, P = 0.006) and mood stabilisers (C = 3.22, Z = 2.72, P = 0.007), suggesting smaller differences with increasing illness duration and euthymia and increased differences with the use of mood stabilisers. In the amygdalae, volumetric differences tended to increase with the use of mood stabilisers (C = 2.28, Z = 2.33, P = 0.02) and to decrease with increasing year of publication (C = −0.29, Z = −3.31, P = 0.001). In the left amygdala, volumes increased with duration of illness (C = 0.08, Z = 2.21, P = 0.03). In the right amygdala, volumes decreased in relation to use of antipsychotic and antidepressant medication (C = −1.28, Z = −2.38, P = 0.017 and C = −1.67, Z = −2.08, P = 0.038 respectively). The hippocampi differences between participants with bipolar disorder and controls correlated positively with age at onset (C = 0.063, Z = 2.96, P = 0.003), suggesting larger bipolar v. control hippocampal differences in individuals with a later age at onset. The effect size with respect to the thalami correlated negatively with year of publication (C = −0.15, Z = −2.92, P = 0.003) and slice thickness (C = −0.46, Z = −2.38, P = 0.02), and positively with duration of illness (C = 0.28, Z = 1.98, P = 0.047). These findings suggest that the thalami may have been smaller in individuals with bipolar disorder in more recently published studies and in studies where larger voxel sizes were used. However, the observed non-significant reductions in participants with bipolar disorder may be less evident in those with longer durations of illness. Pituitary volumes were negatively associated with percentage of male participants (C = −5.80, Z = −2.37, P = 0.02). Volumetric differences in the temporal lobes correlated positively with mean age (C = 0.03, Z = 2.92, P = 0.004), duration of illness (C = 0.054, Z = 2.87, P = 0.004) and use of mood stabilisers (C = 2.23, Z = 2.69, P = 0.007). These results suggest smaller volumetric differences in this region in bipolar disorder with increasing age, duration of illness and the use of mood stabilisers. Duration of illness was positively associated with bipolar–control differences in right temporal lobe volume (C = 0.076, Z = 2.22, P = 0.03).

Bipolar disorder in comparison with schizophrenia

In the bipolar disorder v. schizophrenia comparison, whole brain volume was positively associated with percentage of male participants (C = 1.94, Z = 2.55, P = 0.011). Left amygdala volume was positively associated with number of participants with euthymia (C = 0.98, Z = 2.10, P = 0.035). Volume differences in the left and right hippocampus were positively associated with age (C = 0.043, Z = 4.29, P<0.001 and C = 0.039, Z = 4.04, P<0.001 respectively) and number of participants with euthymia (C = 1.62, Z = 3.43, P = 0.001 and C = 1.73, Z = 3.45, P = 0.001 respectively).