Written informed consent was obtained from all subjects included. Recruitment for the Caribbean Hispanic Study was approved by the Institutional Review Board of the Columbia University Medical Center. Recruitment for the NIALOAD Study was approved by the relevant institutional review boards of the participating centers (ie. the IRBs of Boston University, Columbia University, Duke University, Indiana University, Massachusetts General Hospital, Mayo Clinic, Mount Sinai School of Medicine, Oregon Health & Science University, Rush University Medical Center, University of Alabama at Birmingham, University of California Los Angeles; University of Kentucky; University of Pennsylvania; University of Pittsburgh; University of Southern California; University of Texas Southwestern; University of Washington; Washington University Medical Center; University of Miami; Northwestern University; Emory University).The study was conducted according to the principles expressed in the Declaration of Helsinki.

The two datasets included a) 162 Caucasian cases and 670 controls from the NIA-LOAD study [13] and b) 549 cases and 544 controls from a Caribbean Hispanic dataset that have been described in detail elsewhere [14]. The clinical characteristics of these datasets are summarized in Table 1. The diagnoses of ‘probable’ or ‘possible’ AD were defined according to the National Institute of Neurological and Communication Disorders and Stroke–Alzheimer's Disease and Related Disorders Association (NINCDS-ADRDA) diagnosis criteria at clinics specializing in memory disorders or in clinical investigations. Persons were classified as “controls” when they were without cognitive impairment or dementia at last visit. Informed consent was obtained from all participants using procedures approved by institutional review boards at each of the clinical research centers collecting human subjects.

For both studies, all participants underwent a standardized neuropsychological test battery that examined multiple domains [15]. In the Caribbean Hispanic Study, orientation was evaluated using parts of the modified Mini-Mental State Examination [16]. Language was assessed using the Boston Naming Test [17], the Controlled Word Association Test [18], category naming, and the Complex Ideational Material and Phrase Repetition subtests from the Boston Diagnostic Aphasia Evaluation [19]. Abstract Reasoning was evaluated using WAIS-R Similarities subtest [20], and the non-verbal Identities and Oddities subtest of the Mattis Dementia Rating Scale [21]. Visuospatial ability was examined using the Rosen Drawing Test [22], and a matching version of the Benton Visual Retention Test [23]. Memory was evaluated using the multiple choice version of the Benton Visual Retention Test [23] and the seven subtests of the Selective Reminding Test [24]: total recall, long-term recall, long-term storage, continuous long-term storage, words recalled on last trial, delayed recall, and delayed recognition. This neuropsychological test battery has established norms for the same community [25]. In the NIA-LOAD Study, cognition was measured with a battery of 7 brief tests [26]. Working memory was assessed with Digit Span Forward [27], Digit Span Backward [27], and Digit Ordering [28]. Two measures of episodic memory were included: immediate and delayed recall of story A from the Wechsler Memory Scale-Revised [27]. Semantic memory was assessed by asking persons to name members of two semantic categories (Animals, Vegetables) in separate 1-min trials [26], [28], [29]. While all subjects recruited into the NIA-LOAD study underwent standard neuropsychological assessment that contributed to the diagnosis of AD, the standardized neuropsychological test battery especially designed for the NIA-LOAD study was integrated in the study at a later stage (year 2004). Therefore only 832 subjects of the originally recruited case-control sample (n = 1877) have standardized neuropsychological data and contributed to the final analytic NIA-LOAD sample included in the present analysis.

Both study sites provided the results from genotyping of SORCS1 SNPs that were part of its genome-wide studies described previously [13], [14]. For the NIA-LOAD study, SNPs were genotyped using the Illumina Human610Quadv1_B BeadChips (Illumina, San Diego, CA, USA). For the Caribbean Hispanic study, SNPs were genotyped using the Illumina HumanHap 650Y chip. Genotyping of APOE polymorphisms (based on SNPs rs7412 and rs429358) for all samples was performed at PreventionGenetics.

First, for both datasets a memory savings score was calculated by dividing Delayed Free Recall by Trial 6 Recall (Caribbean Hispanic Study) or by Story (NIA-LOAD Study) multiplied by 100 and expressed as a percent. Using the means and standard deviations from the control samples, we then transformed the resulting savings scores into z-scores. Then, we restricted the genotyping data to the SNPs that were overlapping in both datasets (110 overlappimg SNPs spanning 590 kb). SNP marker data were assessed for deviations from Hardy-Weinberg equilibrium (HWE) in controls. Independently for each dataset, multivariate linear regression analyses were used to assess genotypic and allelic associations with the memory savings scores, adjusting for Population stratification, sex, APOE-ε4 and age-at-onset or age-at-examination. The False Discovery Rate (FDR) [30], which controls the expected proportion of incorrectly rejected null hypotheses (type I errors), was used to account for the error in multiple comparisons. PLINK (http://pngu.mgh.harvard.edu/~purcell/plink/) was used to perform a meta-analysis of both datasets.

We used Haploview (http://www.broad.mit.edu/mpg/haploview/index.php) to assess linkage disequilibrium (LD). Haplotype blocks were defined using the confidence intervals algorithm. The default settings were used in these analyses, which create 95% confidence bounds on D′ to define SNP pairs in strong LD. Analyses assessing associations between haplotypes and the memory savings score, were carried out using a window of three contiguous SNPs using PLINK v1.07 (http://pngu.mgh.harvard.edu/~purcell/plink/) for case-control data. We performed all analyses first for cases and controls combined, and then for the case and control groups separately.

We also performed a meta-analysis of both datasets for single marker and haplotype analyses. To determine the strength of associations between the individual SORCS1 SNPs (or haplotypes) and the memory savings score, we calculated a pooled OR for each marker/haplotype using fixed and random effects models using PLINK. We first performed meta-analyses of unadjusted results from the individual datasets, and then repeated the meta-analyses using the results from the individual datasets adjusting for Population stratification, sex, APOE-ε4 and age-at-onset or age-at-examination. The p-values for each SNP/haplotype were corrected for multiple testing using the False Discovery Rate (FDR). Between-dataset heterogeneity was quantified using the I² metric for inconsistency and its statistical significance was tested with the chi-square distributed Q statistic. I² is provided by the ratio of (Q−df)/Q, where df = the number of degrees of freedom (one less than the number of combined datasets); it is considered large for values above 50% and Q is considered statistically significant for p = 0.10.