Articles (to February 11, 2017) were electronically searched and screened a from databases including PubMed, Embase, Cochrane Library, China National Knowledge Infrastructure database (CNKI), and Wanfang database (an affiliate of the Chinese Ministry of Science & Technology). Mesh terms with their corresponding synonyms were combined to form the search strategy: “Single nucleotide polymorphisms”, “Alzheimer’s disease”, “CD33” and “ABCA7”. Additional articles were also screened manually from the references in each eligible study.

Inclusion criteria were defined as:

• Studies must assess the association between polymorphism of CD33 rs3865444 or ABCA7 rs3764650 and susceptibility to AD;

• Case-control studies must be based on humans;

• Studies containing sufficient information to obtain the odds ratio (OR) and 95% confidence intervals (CI);

• AD diagnosis should meet clinical criteria set by the NINCDS-ADRDA Alzheimer’s Criteria (National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer’s Disease and Related Disorders Association);

• Genotype distribution of controls conform to the Hardy-Weinberg equilibrium (HWE).

Exclusion criteria were defined as follows:

• Abstracts, reviews, meta-analysis, case reports, comments, and editorials;

• Duplications, grey literatures, unpublished articles;

• Studies without sufficient genotyping information.

Allele, dominant, recessive, and additive models were employed for the meta-analysis. The CD33 rs3865444 polymorphism has two alleles, A and C, A is the minor allele. The ABCA7 rs3764650 polymorphism has two alleles, G and T, G is the minor allele. The four models were described as: CD33 rs3865444: allele model (A allele versus C allele), dominant model (AA + AC versus CC), recessive model (AA versus AC + CC), and additive model (AA versus CC); ABCA7 rs3764650: allele model (G allele versus T allele), dominant model (GG + GT versus TT), recessive model (GG versus GT + TT), and additive model (GG versus TT).

Cochran’s Q-statistic and I² test were used to evaluate statistical heterogeneity. A p value of Cochran’s Q test greater than 0.05 and I² statistic less than 50% (p ${\mathrm{}}_h>$ 0.05 and I² $<$ 50%), the fixed effect model was suitable for analysis due to lack of significant heterogeneity. When the $p$ value of the Cochran’s Q-test was less than or equal to 0.05 or I ${\mathrm{}}^2$ statistic was greater than or equal to 50% (p ${\mathrm{}}_h\le$ 0.05 or I² $\ge$ 50%), the random effects model was used for studies owing to the presence of significant heterogeneity [20]. Subgroup analysis by ethnicity (Asian, Caucasian) was performed for different heritability models that included allele, dominant, recessive, and additive models. Hardy-Weinberg equilibrium among the controls for each study was examined using Pearson’s chi-square test (p $<$ 0.05 was assumed to indicate deviation from Hardy-Weinberg equilibrium). Sensitivity analysis proceeded by individually omitting each study. Publication bias was assessed with Egger’s test, Begg’s test, and the inverted funnel plot [21]. All statistical analysis was performed with RevMan 5.3 and Stata 12.0 Software (StataCorp LP, College Station, Texas, USA).

The meta-analysis started with 173 articles. Initial review removed 46 duplicate articles to give 127 articles which were then subjected to abstract and keyword review. Of the 127 articles, 109 unrelated articles were discarded to give only 18 articles suited for full-text and data assessment, 6 articles lacked genotyping information, thus 12 eligible articles were finally available for data extraction [22-33]. The detailed information of each study is given in Table 1 and Table 2.

CD33 rs3865444. A total of 10 studies that included 22,262 patients with AD and 32,244 controls were suitable for the meta-analysis of CD33 rs3865444 polymorphism. Significant statistical heterogeneity was identified at the allelic level (I² = 71%, p ${\mathrm{}}_h<$ 0.00001), the overall OR was then calculated with the random effect model and no significant association was found between CD33 rs3865444 polymorphism and AD (p = 0.99, OR = 1.00, 95% CI, 0.92-1.09, minor allele = A) (Fig. 1/Table 3). In addition to the allele model, dominant, recessive, and additive models were used to evaluate association between the CD33 rs3865444 polymorphism and AD. Only seven of ten articles were selected for further analysis as three models required the exact number of original genotypes. There was significant heterogeneity among the selected studies for the dominant model (I² = 92%, p ${\mathrm{}}_h<$ 0.00001), the recessive model (I ${\mathrm{}}^2$ = 65%, p ${\mathrm{}}_h=$ 0.001) and the additive model (I² = 69%, p ${\mathrm{}}_h=$ 0.002). Furthermore, the results of the meta-analysis revealed no significant associations using the dominant model (p = 0.18, OR = 1.25, 95% CI, 0.90-1.73), recessive model (p = 0.15, OR = 1.21, 95% CI, 0.93-1.58), or additive model (p = 0.51, OR = 1.10, 95% CI, 0.83-1.44) (Table 3).

Fig. 1.

Forest plot of CD33 rs3865444 polymorphism association with AD using the allele model. M-H, Mantel-Haenszel, random effect model, confidence interval (CI).

ABCA7 rs3764650. A total of eight studies including 52,214 patients with AD and 82,948 controls were included for the meta-analysis of ABCA7 rs3764650 polymorphism. There was no significant statistical heterogeneity among these studies using the allele model (I ${\mathrm{}}^2$ =49%, p ${\mathrm{}}_h=$ 0.001), overall OR was then calculated with the random effect model and no significant association was found between the ABCA7 rs3764650 polymorphism and AD (p = 0.06, OR = 1.06, 95% CI, 1.00-1.13, minor allele = G) (Fig. 2/Table 4). Dominant, recessive, and additive models were used to evaluate the association between ABCA7 rs3764650 and AD. Only five of eight articles were suited for further analysis due to three models requiring the exact number of original genotypes. There was no significant heterogeneity among selected studies for the dominant model (I ${\mathrm{}}^2$ = 30%, p ${\mathrm{}}_h$ = 0.20), the recessive model (I ${\mathrm{}}^2$ = 44%, p ${\mathrm{}}_h=$ 0.10) or additive model (I ${\mathrm{}}^2$ = 36%, p ${\mathrm{}}_h=$ 0.37). Furthermore, results of the meta-analysis revealed significant associations using the dominant model (p $<$ 0.0001 OR = 1.20, 95% CI, 1.10-1.31), recessive model (p = 0.01, OR = 1.59, 95% CI, 1.12-2.28), and additive model (p = 0.003, OR = 1.44, 95% CI, 1.13-1.83) (Table 4).

Fig. 2.

Forest plot of ABCA7 rs3764650 polymorphism association with AD using the allele model. M-H, Mantel-Haenszel, random effect model, confidence interval (CI).

CD33 rs3865444. The frequency of CD33 rs3865444 polymorphism was variable among two different subgroups (Asian and Caucasian). A total of 10 articles were suitable for further subgroup analysis of the allele model. There was a significant heterogeneity in Asians (I ${\mathrm{}}^2$ = 92%, p ${\mathrm{}}_h$ $<$ 0.00001), but no significant heterogeneity in Caucasians (I ${\mathrm{}}^2$ = 0%, p ${\mathrm{}}_h=$ 0.68), the random effect model was then used for further subgroup analysis. Subsequent meta-analysis revealed no significant association for either Asian (p = 0.71 OR = 1.10, 95% CI, 0.66-1.83) or Caucasian (p = 0.26, OR = 0.98, 95% CI, 0.93-1.02) populations (Fig. 3/Table 3). Furthermore, only eight of ten articles were selected for subgroup analysis using the dominant, recessive, and additive models. A subgroup analysis was then conducted for these models for Asians and Caucasians. There was significant heterogeneity in both Asians and Caucasians (I ${\mathrm{}}^2>$ 50%, p ${\mathrm{}}_h<$ 0.05) with the exception of the additive model (AA versus CC) for Caucasians (I ${\mathrm{}}^2$ = 0%, p ${\mathrm{}}_h$ = 0.49). The random effect model was ultimately selected for further subgroup analysis. There was no significant association with AD under these models for either Asians (dominant: p = 0.33, OR = 1.36, 95% CI, 0.74-2.51; recessive: p = 0.80, OR = 1.11, 95% CI, 0.49-2.50; additive p = 0.27, OR = 1.76, 95% CI, 0.65-4.81) or Caucasians (dominant: p = 0.32, OR = 1.22, 95% CI, 0.82-1.81; recessive: p = 0.45, OR = 1.09, 95% CI, 0.87-1.38; additive: p = 1.31, OR = 0.90, 95% CI, 0.78-1.03) (Table 3).

Fig. 3.

Forest plot of CD33 rs3865444 polymorphism association with AD using the allele model for different ethnicities. M-H, Mantel-Haenszel, rondom effct model, confidence interval (CI).

ABCA7 rs3764650. The frequency of ABCA7 rs3764650 polymorphism was variable among different subgroups (Asian and Caucasian). A total of eight articles were suited for further subgroup analysis for the allele model. There was no significant heterogeneity for Asians (I ${\mathrm{}}^2$ = 0%, p ${\mathrm{}}_h$ = 0.92) but there was significant heterogeneity for Caucasians (I ${\mathrm{}}^2$ = 59%, p ${\mathrm{}}_h$ = 0.003). The random effect model was selected to conduct a further subgroup analysis. The subsequent meta-analysis revealed no significant association for Asian (p = 0.33, OR = 1.06, 95% CI, 0.94-1.19) or Caucasian (p = 0.1, OR = 1.06, 95% CI, 0.99-1.14) populations (Table 4). Furthermore, only five of eight articles were selected for analysis using the dominant, recessive, or additive models. A subgroup analysis was then conducted for these models for Asians and Caucasians. There was no significant heterogeneity in Caucasians or Asians (p $>$ 0.05 and I ${\mathrm{}}^2<$ 50%) under the dominant and additive models, thus the fixed effect model was selected for subgroup analysis under these two genetic models in Caucasians and Asians. A significant association was found under these models in Caucasians (dominant: p $<$ 0.00001, OR = 1.28, 95% CI, 1.16-1.41) (Fig. 4/Table 4); (additive: p = 0.001, OR = 1.96, 95% CI, 1.30-2.94) (Table 4). However, a significant heterogeneity for Asians was found (I ${\mathrm{}}^2$ = 70%,p ${\mathrm{}}_h$ = 0.07) for the recessive model, thus, the random effect model was ultimately selected. A significant association was found under the recessive model in Caucasians (p = 0.002, OR = 1.96, 95% CI, 1.27-3.04), while there was no association between ABCA7 rs3764650 polymorphism and Alzheimer’s disease in Asians subgroups (p = 0.28, OR = 1.34, 95% CI, 0.79-2.26) (Table 4).

Fig. 4.

plot of ABCA7 rs3764650 polymorphism association with AD using the dominant model for different ethnicities. M-H, Mantel-Haenszel, random effect model, confidence interval (CI).

Sensitivity analysis was conducted by excluding one study at a time to assess whether any single study had a strong influence on the pooled OR. For CD33 rs3865444, sensitivity analysis indicate that no single study significantly influenced the pooled OR (data not shown). However, for ABCA7 rs3764650, by excluding the study from Harold et al. [32]. (UK/Ireland) using the additive model, G versus T with OR = 1.44, 95% CI, 1.13-1.83, p = 0.003 changed to OR = 1.29, 95% CI, 0.99-1.68, p = 0.06. Furthermore, Begg’s funnel plot and Egger’s linear regression test were performed to assess publication bias. For CD33 rs3865444, the shapes of the funnel plots show no evidence of publication bias under either the allele model (Begg’s test, p = 0.972; Egger’s test, t = 0.04, p = 0.972) (Fig. 5), dominant model (Begg’s test, p = 0.828; Egger’s test, t = $-$ 0.22, p = 0.828), recessive model (Begg’s test, p = 0.862; Egger’s test, t = 0.18, p = 0.862), or additive model (Begg’s test, p = 0.333; Egger’s test, t = 1.02, p = 0.333). For ABCA7 rs3764650, there was also no significant publication bias for the allele model (Begg’s test, p = 0.118; Egger’s test, t = 1.66, p = 0.118) (Fig. 5), dominant model (Begg’s test, p = 0.674; Egger’s test, t = -0.46, p = 0.674), recessive model (Begg’s test, p = 0.606; Egger’s test, t = 0.55, p = 0.606), or additive model (Begg’s test, p = 0.459; Egger’s test, t = 0.80, p = 0.459).

Fig. 5.

Publication bias for CD33 rs3865444 and ABCA7 rs3764650 under the allele model detected by Egger’s publication bias plot analysis.

SE:Standard error of mean.