Introduction

Associations between cognitive function and anthropometric measures in childhood and adulthood, such as body height and head circumference, are well documented. Several studies have reported positive associations of cognitive ability with body length in childhood [1] and in adulthood [2,3], as well as with childhood and adult head circumference [2,4–6]. Further, both adult height and head circumference have been shown to be negatively associated with late life cognitive decline [6–8] and dementia [9–12].

Further, it has repeatedly been shown that adults with higher cognitive ability, higher education, and higher occupational achievements seem to be somewhat protected from developing age-related cognitive impairments such as dementia [13–16]. To account for this finding, the concept of cognitive reserve has been proposed, a concept that has gained much attention in regard to age-related cognitive decline [13]. The cognitive reserve theory proposes that differences in complex mental activity and ability throughout adulthood (cognitive reserve) as well as in brain structure and size (brain reserve) affect resilience against age-related decline [13,17]. This theory has been extended to childhood cognitive ability, with children scoring low on ability tests having an increased risk for dementia late in life [18]. Recently, a study found evidence that education not only has an enhancing effect on cognitive reserve but also contributes to brain reserve through increased cerebral volume [19]. These findings suggest that cognitive reserve is at least partly established early in life. Another possible explanation could be that other factors contribute to higher education and higher cognitive ability (i.e., cognitive reserve) as well as lower risk for age-related cognitive decline. Such factors could be influences very early in life, such as fetal growth patterns, gestational age at birth, or genetic influences.

Finally, birth characteristics, such as preterm birth, low birth weight, short birth length, small head circumference, and indicators of reduced fetal growth (i.e., small for gestational age), are independently associated with lower academic achievement, cognitive function, and intellectual performance in childhood [5,20–22] and young adulthood [23–25], even after controlling for familial factors and socioeconomic status. To the best of our knowledge, thus far only a few small studies [6,26,27] have explored associations between different birth characteristics and cognition in old age, reporting mixed findings. While some research reported significant associations for head circumference at birth, but not birth weight or length [26], other studies found no significant association between late life cognition or cognitive decline and the birth characteristics included, i.e., head circumference [6] or birth weight or length [27].

Together, the well-documented relationships between size at birth, gestational age, and cognition, as well as adult cognition and dementia, and the associations between adult anthropometric measures and dementia imply potentially adverse effects of reduced fetal growth for age-related cognitive decline and dementia. At this time, when population aging is one of the most important demographic phenomena in the world [28], and given the large burden of cognitive decline in older individuals [29,30], with currently no available cure for cognitive decline and dementia, research towards determining early predictors of cognitive impairment and dementia is warranted.

A link between birth characteristics and age-related cognitive dysfunction could in part explain the reported associations between adult anthropometric measurements, cognitive ability, and age-related cognitive dysfunction. Understanding these relationships is essential for the development of prevention strategies and to identify individuals who may benefit from early intervention. The present study explores potential associations between birth characteristics—i.e., weight, length, head circumference, and gestational age, as absolute measures as well as relative to gestational age—and all-cause dementia, using a large Swedish population-based study of more than 35,000 individuals. In addition, as registry-based dementia diagnoses may underestimate true numbers of cases of dementia in the population [31], associations between birth characteristics and age-related cognitive impairment based on comprehensive computer-assisted telephone screening were investigated in a subsample of 4,000 individuals.

Methods

Multiple imputation

As some of the variables (see below) had missing values, resulting in reduced Ns in those models corrected for increasing numbers of covariates, the multiple imputation (MI) procedure in StataIC 14 [49] was used to impute missing data. MI is a simulation-based approach that uses all known variables in the model to replace missing values with multiple sets of simulated values, rather than a single value, to complete the data [50]. That way, MI preserves the distributions and relationships in the data while still reflecting the uncertainty in the predictions of the missing values. All variables forming part of our analyses were also included in our imputation models (including the outcome and fully observed variables). The variables with missing data included in the imputation model were HC (continuous), BL (continuous), BWGA (continuous or dichotomous, i.e., SGA), HCGA (continuous or dichotomous), BLGA (continuous or dichotomous), GA (continuous or dichotomous), parity (continuous), age of mother (continuous), birth SES (ordinal), and education level (dichotomous). In addition, the fully observed variables BW, YOB, sex, age, zygosity, LBW, and either dementia diagnosis and the cumulative hazard function or cognitive impairment were also included to preserve their relationships with the imputed variables. Note that imputations for each model were run twice, first with the continuous GA variables (i.e., BWGA, HCGA, BLGA, and GA) and then a second time with the dichotomized GA variables (i.e., SGA, SHCGA, SBLGA, and preterm). Further, mother’s age only had missing values in the dementia analyses while it was included as a fully observed variable for the cognitive impairment analyses. To calculate how many imputations should be regarded as sufficient for the pattern of missing data in the 2 samples, a 2-step procedure using a quadratic rule recommended by von Hippel [51] was used. Accordingly, 58 and 100 imputed datasets were created in each run for the dementia dataset and the cognitive impairment subset, respectively, using chained equations (using the STATA command mi impute chained), a MI method using a sequence of univariate imputations with fully conditional specification of prediction equations [49]. The augmented regression option was used as there was perfect prediction. The resulting estimates were combined using Rubin’s rule to derive 1 final MI result. See Table 1 for the distribution of missing/imputed values for each variable and set of analyses, i.e., dementia and cognitive impairment. The subsequent analyses (described below) were conducted using the imputed datasets. For comparison and to check whether the imputation process altered our findings, all analyses were then repeated in the original dataset with missing values and in a reduced dataset including only individuals with complete data. The results of these additional analyses are reported in S3–S6 Tables.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Patterns of missingness in the birth variables and covariates before imputation.

https://doi.org/10.1371/journal.pmed.1002609.t001

Results

Descriptive statistics

The final dataset for the birth characteristics–dementia analyses contained 35,191 twin individuals aged 55 to 89 years (mean = 69.21, SD = 8.83) at follow-up, with information on at least 1 of the birth characteristics and dementia diagnosis (Table 2). BW ranged between 1,010 and 4,700 g (mean = 2,664), HC ranged between 230 and 430 mm (mean = 329.5), and BL ranged between 34 and 59 cm (mean = 47.7)—expected birth characteristics in a twin sample [55]. Of this sample, 2.6% were diagnosed with dementia (an incidence rate of 5.9% per 100,000 person-years), with a mean age at diagnosis of 74.31 years (SD = 7.38, range = 55–88). Of the 4,000 individuals in the subsample included in the cognitive impairment analyses (mean age at screening = 68.36 years, SD = 2.55, range = 65–74), 14.2% were identified as cognitively impaired based on the telephone screening. Descriptive statistics for the birth characteristics and cognitive variables in both samples are presented for sexes separately and combined in Table 2. Additional testing for a main effect of birth order on the 2 measures of cognitive dysfunction showed that although the second-born twin was on average somewhat smaller than the first-born, birth order was far from significant (p > 0.5) for dementia and cognitive impairment, suggesting no association between birth order and cognitive dysfunction.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Descriptive statistics for the birth and cognitive variables (unimputed).

https://doi.org/10.1371/journal.pmed.1002609.t002

Birth characteristics–dementia analyses

Table 3 shows the associations between indices of BW, HC, BL, and GA and dementia diagnosis from analyses both not adjusted and subsequently adjusted for covariates based on the imputed dataset. Regardless of the covariates included, lower BW significantly increased the risk for dementia, with an approximately 2% reduction in dementia risk with each additional 100 g BW and a 19%–23% increase of dementia risk for infants with LBW. Similarly, each additional standard deviation BWGA resulted in an 8%–9% significantly reduced risk for dementia, and SGA infants showed a 19% higher dementia risk, though the latter association was not significant.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Hazard ratios for dementia diagnosis based on survival analyses in relation to birth characteristics.

https://doi.org/10.1371/journal.pmed.1002609.t003

While there was not sufficient evidence in support of an association between the 2 continuous HC variables, i.e., HC (in mm) and HCGA, and dementia, infants with SHCGA had an about 65% higher risk for dementia, independent of the covariates included. Though the estimates were in the expected direction for all 3 BL variables, with an up to 40% increased risk for dementia for those born with SBLGA, none of the estimates were significant. Similarly, there was little evidence in support of an association between GA or preterm and dementia. Of the covariates, parity and birth decade were the only significant risk factors for dementia, with increased parity (p < 0.05) and later birth decade (with the exception of the most recent birth decade, 1946 to 1955) somewhat increasing the risk for dementia, i.e., those born between 1936 and 1945 showed a somewhat increased risk (non-significant), while those born between 1946 and 1955 had an about 40% higher risk compared to those born 1926 to 1935 (p < 0.05). There was little evidence in support of an association between birth SES or education and dementia (even when included without any of the birth-related variables). Note that overall results based on the full unimputed sample as well as the reduced sample (only participants with all covariates) were very similar in terms of both effect sizes and significance levels, although the latter varied slightly between analyses depending on the sample size (see S3 and S5 Tables). Further, adjustment for birth order had no effect on any of the results (see S7 Table).

Birth characteristics–cognitive impairment analyses

Table 4 shows associations between indices of BW, HC, BL, and GA and cognitive impairment from analyses both uncorrected and subsequently corrected for covariates in the imputed subsample. While there was little evidence in support of an association between BW, LBW, or BWGA and cognitive impairment, ORs for SGA ranged between 1.69 and 1.76 depending on the covariates included, suggesting a somewhat increased risk of cognitive impairment for infants born SGA.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Odds ratios for cognitive impairment based on logistic regression in the subsample in relation to birth characteristics.

https://doi.org/10.1371/journal.pmed.1002609.t004

All 3 measures of HC had a significant influence on cognitive impairment regardless of the covariates included, with a 1% reduction in odds with each additional millimeter HC, about 15% lower odds for each additional standard deviation HCGA, and a more than 2-fold greater odds for individuals with SHCGA at birth. Which covariates were included had little influence on effect sizes and significance levels. Again, there was little evidence in support of an association between any BL or GA variables and cognitive impairment, although infants born with SBLGA showed a 51%–61% increased odds for cognitive impairment (non-significant). Of the covariates, age, parity, and education were significant, with 7% higher odds of being diagnosed with cognitive impairment with each additional year of age (p < 0.01) and parity number (p < 0.01), as well as a 35% lower odds for those with higher education (p < 0.001). Again, results based on the full unimputed sample as well as the reduced sample (only participants with all covariates) were very similar in terms of both effect sizes and significance levels, although the latter varied slightly between analyses depending on the sample size (see S4 and S6 Tables). Additional adjustment for birth order had no effect on any of the results (see S8 Table).

Within-pair analyses

Within-pair analyses in twins were performed for BW and HC (both converted to z-scores to compare results with the between-pair analyses of the BWGA and HCGA variables, which also were standardized) to explore whether the significant associations observed in the above analyses were confounded by underlying shared liability (Table 5). In total, there were 182 MZ pairs who were discordant for dementia diagnosis (days between dementia diagnosis of 1 twin and diagnosis or censoring of the co-twin: mean = 2,215, SD = 2281, range = 32–13,364) and 62 who were discordant for cognitive impairment. HRs and ORs in the within-pair analyses showed a similar effect of birth characteristics compared to the GA-adjusted results for BW and HC from the between-pair analyses (i.e., BWGA and HCGA), though they were not significant. When including both MZ and DZ twins in the analyses, effect sizes were still not significantly different from 1, and also the interaction effect was non-significant, suggesting that there was no significant difference in effect size between MZ and DZ twins. Adjusting for birth order did not change results significantly (see S9 Table). However, while effect sizes remained comparable for cognitive impairment, the point estimates for dementia risk changed somewhat, with wide confidence intervals reflecting reduced power. Considering the risk for misclassification in registry data (the source of the dementia diagnoses) and that such misclassification has a bigger impact in co-twin control analyses as compared to cohort analyses [56], it is unlikely that the change in effect size reflects an effect of birth order on the associations with dementia.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Results of within-pair analyses in twin pairs discordant for birth characteristics and dementia or cognitive impairment, adjusted for familial factors shared within twin pairs.

https://doi.org/10.1371/journal.pmed.1002609.t005

Discussion

The present study explored associations between birth characteristics (BW, BL, HC, and GA) and dementia and cognitive impairment later in life. Overall, lower BW and smaller HC showed associations in the expected direction, resulting in increased risks for age-related cognitive dysfunction, i.e., dementia or cognitive impairment. For dementia, lower BW (independent of GA) was a significant risk factor, with a 2% higher risk with each 100 g lower BW and about 22% higher risk if weighing less than 2.5 kg at birth. Similar associations were observed when BW was adjusted for GA, with a 9% risk reduction for dementia for each additional standard deviation BWGA and a 17% risk increase for individuals born SGA, though the latter was not significant. HC was a significant risk factor only when the head was SGA (65% risk increase), suggesting harmful effects of reduced fetal growth for dementia. Similarly, detrimental effects were observed for cognitive impairment, with individuals born SGA and with SHCGA showing approximately 73% and 124% higher odds, respectively. In addition, larger HC was, independent of GA as well as adjusted for GA, a significant protective factor for cognitive impairment (but not for dementia), with each additional millimeter resulting in a 1% lower odds for cognitive impairment and each additional SD HCGA resulting in a 15% lower odds. BL and GA (independent of other birth characteristics) showed no significant association with late life cognitive dysfunction. Overall, these findings suggest negative long-term effects of small birth size on risk of cognitive dysfunction late in life, with additional protective effects of each extra 100 g of BW for dementia and each additional millimeter of HC for cognitive impairment (independent of GA).

Past research has shown birth characteristics adjusted and unadjusted for GA to be independently associated with academic achievement, cognitive function, and intellectual performance early in life [5,20–25], traits known to be protective against cognitive dysfunction late in life and frequently referred to as cognitive reserve [14]. The exact mechanisms underlying such long-term effects of birth characteristics are unclear, although some research has suggested that fetal growth restriction can result in structural brain differences, such as reduced grey matter [57,58], i.e., lower brain reserve. In view of the previous literature, the present findings suggest that the reported protective value of cognitive reserve (including brain reserve) against cognitive dysfunction may at least to some degree be explained by factors influencing fetal growth and development.

Contrary to prior literature and the findings of another study using a partly overlapping sample in the STR [59,60], education level had little impact on dementia diagnoses (even when included as the sole predictor). The same was observed for birth SES. There are several possible explanations for the missing protective effect of these potential indicators of early and later cognitive stimulation on dementia. First, in the present sample only individuals who had birth information could be included, restricting the maximum age to 89 years. As a result, we may have an overrepresentation of individuals diagnosed towards the young end of the typical age of onset of dementia, and, as one effect of cognitive reserve is to delay dementia, we may miss those protected by cognitive reserve. Second, registry-based dementia diagnoses have been shown to capture only about 63% of the true number of dementia cases in the population, despite showing a near perfect specificity [31,61]. Consequently, individuals may have been misclassified as healthy, resulting in a bias towards null. In that case, it is even more remarkable that we still observe such strong associations between birth characteristics and dementia, suggesting that such associations may be stronger than birth SES, education, and possibly cognitive reserve. Third, it could be that individuals with higher education and higher SES are more likely to seek medical advice when they or their family members notice cognitive symptoms and, as such, are more likely to be diagnosed with dementia. This would result in lower sensitivity of dementia diagnosis among lower educated individuals, which in turn could cover up the protective effect of education. Further, though the effect of education on dementia has been shown to be quite consistent [62], previous research has suggested that it varies somewhat depending on the study population [30,62] and is most consistent when years of education reflect cognitive capacity [63,64]. Therefore, another possible explanation could be that the 2 measures used here do not very well reflect true cognitive capacity and as such do not affect cognitive reserve, especially given that the education measure included was coarse, with only 2 categories. However, a significant association of education, but not childhood SES, was observed with the cognitive impairment measure. The association between education and cognitive impairment was mostly independent from the birth characteristic effects, suggesting that the protective effects of education on age-related cognitive impairment risk cannot fully be explained by prenatal factors. A possible explanation as to why a significant protective effect of education on risk of cognitive impairment but not dementia was observed could be that our cognitive impairment measure taps into learned information, which would be strongly related to education.

Of the other covariates, only higher parity and later birth decade were significant for dementia, and increased age and parity for cognitive impairment. The finding that higher parity increases the risk for cognitive dysfunction later in life is in line with previous findings showing an increased risk for low intellectual performance with increased parity [46]. Further, the increased dementia risk with higher birth decade is likely a reflection of an increased sensitivity of the different medical registers over time (the likelihood for dementia to be captured and diagnosed) rather than being due to an increase in dementia incidence. For example, the outpatient register only became part of the NPR in 2001. Further, those born in the most recent birth decade (i.e., 1946 to 1955) would not yet have reached the peak age of onset of dementia and therefore do not show a higher dementia risk. Birth order had no significant effect on the cognitive outcome measures or any of the analyses.

Effect sizes of the within-pair analyses, which are adjusted for shared influences, were of similar magnitude (though not significant) to those from between-pair analyses, suggesting that the increased risk for cognitive dysfunction for individuals who were of smaller birth size is likely not explained by underlying shared etiology, such as genetic factors or shared environmental influences (e.g., prenatal family environment or maternal factors including family SES). These findings, though hampered by a small sample size and low power, could suggest that the smaller twin was exposed to a less optimal intrauterine environment, which manifested itself in the less ideal birth characteristics as well as a worse trajectory throughout the lifetime in terms of cognitive development and cognitive aging. These results are in line with past reports of effects of poor fetal growth (as indicated by adverse birth characteristics) on cognitive development in early childhood based on co-twin control analyses [65]. However, as the HRs and ORs were not significant in the within-pair analyses and confidence intervals were wide, likely due to the small sample sizes, the results should be interpreted with caution.

In light of the past literature, our findings show that small birth size and low fetal growth may increase the risks of age-related cognitive dysfunction late in life even when controlling for familial factors and childhood SES as well as education in adulthood. Further, each additional 100 g BW and additional millimeter HC significantly reduced the risk of dementia and cognitive impairment, respectively, independent of as well as adjusted for GA. Although our findings suggest that the frequently reported protective value of cognitive reserve (as measured by cognitive ability) on cognitive dysfunction may at least to some degree be explained by factors established as early as during fetal development, the increased risk of age-related cognitive dysfunction associated with small birth size and the indicators of fetal growth measured here was also evident in models corrected for childhood SES and education—predictor and proxy of cognitive ability, respectively. It is known that a multitude of factors influence age-related cognitive decline (as well as cognitive reserve), including genetic and neurological factors, immune response, and various environmental factors such as nutrition, physical activity, family environment, physical and mental trauma, and occupation, which are likely to interact with each other throughout the lifetime. It is probable that exposure to a less optimal intrauterine environment will act negatively on several of the factors named above, including general health as well as brain and cognitive reserve, although it is yet to be determined whether the associations of small birth size and low fetal growth with cognitive decline are mediated by any of those factors. Our findings highlight the importance of preventing prenatal growth restriction and closely monitoring the cognitive development of growth-restricted infants. It has been shown that early life catchup growth may be beneficial for the development of those born with adverse birth characteristics [23], and growing evidence about protective factors for dementia emphasizes a life course model with a key role for early life factors [30]. Various reviews converge on similar suggested interventions to increase cognitive reserve: active treatment of hypertension, more childhood education, physical exercise, social engagement, reducing smoking, and preventing and treating depression (e.g., [30,62]).

As touched upon above, there are some limitations to the use of registry-based data. Among these is the lack of sensitivity of dementia diagnoses, as well as the possibility of mistakes in the birth records (e.g., mix-ups of the twin pairs). Given this, it is even more remarkable that such strong associations between birth characteristics and dementia were observed in the registry data and that the findings could be replicated for cognitive impairment, which was not based on medical records but on a cross-sectional cognitive screening study. As a result, though, the cognitive impairment data had no time-to-event information (i.e., date of onset), which would have been desirable, and had to be modeled as ORs rather than HRs. Another potential caveat of registry-based dementia diagnosis is the potential for misdiagnosis, which has been shown to be especially high in those below the age of 65 years [66]. However, as mentioned above, recent findings on dementia diagnoses in the Swedish patient registries have shown a near perfect specificity [61], suggesting very low rates of misdiagnosis. In addition, the mean age of dementia diagnosis in the present study was 74.31 years, and we found a lower risk for dementia in those in the youngest age group (i.e., age 55–65 years), suggesting that the associations we observe are unlikely to be driven by misdiagnoses in young individuals. Further, the present study only explored associations between birth characteristics and all-cause dementia. The observed associations could be driven by a specific dementia subtype, warranting future research to investigate whether associations vary between dementia types. Measures of birth SES and education were included that may not have been ideal as proxies for baseline cognitive ability (or cognitive reserve). A measure of cognitive ability early in life would have been desirable, but such measures were not available in the samples used. However, education is generally considered a reliable proxy of cognitive ability [67], and the fact that the inclusion of education had no effect on the observed associations suggests that there may be little mediation from cognitive ability.

In addition, the dating (exact GA) of pregnancies would likely have been less accurate at the time the birth data were recorded compared to today, which could potentially explain the non-significant associations observed for GA and some GA-corrected variables. Further, perinatal complications, such as asphyxia, have been associated with developmental deficits in both premature and SGA infants [68], which in turn have been shown to predict worse cognitive development. Both asphyxia and LBW are more common in the second-born than in the first-born twin [69–71]. Therefore, additional analyses adjusting for birth order were conducted: these analyses showed little effect of birth order on cognitive dysfunction and little change to the observed associations. This, in combination with the fact that associations in the expected direction were observed across BW, HC, and BL, including HC independent of GA, and that GA (prematurity) had no main effect on the cognitive dysfunction measures, favors the hypothesis that factors associated with fetal growth explain the associations.

Not unexpectedly given the age of the records, there was information missing in the available birth data. In order to obtain less biased and more precise estimations of the parameters of interest, MI was conducted, which allowed us to include those individuals who would have been omitted in a complete case analysis. However, the overall rate of missingness was only 6.2% in the full sample and 9.7% in the subsample, which is generally regarded as inconsequential [72]. We also conducted a complete case analysis (only individuals who had no items missing) as well as analyses with all available data for each model (reported in S1–S9 Tables for comparison). Comparing the models indicates that overall, as expected with this small amount of missingness, the imputation procedure did not have a large effect on our main findings, with overall very similar effect sizes and p-values. An assumption of MI methods is that the data are missing at random. The greatest extent of missingness was found for birth SES. Should the data not be missing at random, it is likely that those missing would be at the lower end of the distribution, potentially resulting in residual confounding from birth SES of the effect sizes. Nevertheless, as mentioned above, the distributions of the birth characteristics and covariates in the present sample were in line with those previously reported for twin samples [55]. Finally, there are some potential limitations inherent to twin data, especially in regards to birth characteristics. Twins may generally be more growth restricted in utero compared to singletons (for a detailed review and discussion, see [73]), raising the concern of the generalizability of findings in twins to the general population. However, twins tend not to differ from the general population in morbidity and all-cause mortality across the life span [74,75], suggesting that twinning does not affect long-term health outcomes. Further, research has repeatedly shown that also in other domains such as cognitive ability, personality, and risk for mental disease, twins are similar to the general population (e.g., [76–78]). As such, it is likely that the present findings are also applicable to the general population. The within-pair analyses were unfortunately hampered by small sample size, which is not easily overcome, given the already very large overall sample size.

In conclusion, in this study, we found that smaller birth size (i.e., LBW and small HC adjusted and unadjusted for GA) is associated with a significant risk for late life cognitive dysfunction, largely independent from education level and childhood SES—factors related to cognitive ability. As such, our findings highlight the importance of preventing prenatal growth restriction, closely monitoring the cognitive development of growth-restricted infants, and exploring the potential of reserve-enhancing interventions to possibly protect individuals at higher risk for late life cognitive dysfunction.

Supporting information

Acknowledgments

We would like to thank the Swedish twins for their participation.