Mechanisms Associated with Type 2 Diabetes Occurrences: Structural Equation Confirmatory Model
Keywords: Diabetes production mechanism; Structural equations model; Confirmatory statistical model
normality.
profile.
- 01 Volume 2 Issue 1 - 1005 Diabetes production mechanism, Structural equations model, Confirmatory statistical model
Diabetes is the epidemic disease of the 21st century [1]. The WHO estimates that high blood glucose is the third most important risk factor for 6% of deaths worldwide [2]. According to the International Diabetes Federation (IDF), in 2015, diabetes prevalence was 8.8% and 12% of the world's health expenditure was applied to this disease. Over 415 million people suffer from it and approximately five million die from this cause, which is equivalent to one death every six seconds [3].
Currently, 80% of the disease burden is in developing countries [4–6]. By 2015, the prevalence was 12.9% in North America and the Caribbean region, the age-adjusted prevalence was 9.6% in Central and South America [3] and in Colombia, it varied between 7.1% and 8.5%. The average global cost attributed to the disease is $2.7 million and it is the fifth leading cause of death in Colombia, with a rate of 15 deaths per 100,000 people [7]. At such a high cost, the disease is a major challenge to health systems and an obstacle to economic equity and sustainable development [8,9].
Risk factors associated with this disease have traditionally been studied at an individual level and most of the methods used to confirm theoretical hypotheses based on the analysis of empirical data offer a simple structural model, which is a complex and multifactorial world, provides partial knowledge that is difficult to understand and apply in public health [10,11]. This research aimed to identify whether the causal mechanisms of the main theoretically established risk factors had an empirical correlation with the disease burden. The empirical adequacy of the theorization of the causal mechanism of type 2 diabetes was evaluated, through a structural equation model, with globally accepted risk factors, measured in a transversal way in people living in the city of Santiago de Cali, Colombia.
AII records from blood glucose tests measuring total cholesterol and lipid profiles were selected from a database containing information on risk measurement factors for no communicable diseases in the city of Santiago de Cali, during 2013. We excluded those with a prior diagnosis of diabetes, a height measurement of 1.35m or less, weight at 35 kg or less, systolic pressure measuring less than or equal to 50 or greater than 220 mm/Hg, diastolic pressure measuring less than or equal to 20 or greater than 180 mm/Hg, triglycerides lower than 80 or greater than 1500 mg/dL, and those with missing data on the variables of main interest in the modeling. In the end, information was obtained for 26 variables and 423 records.
The WHO (2)defined a theoretical framework focused on globally accepted factors with data available to estimate exposure in the population on which pathways or mechanisms for its reduction are known. This was the starting point for a specification model that could preserve classification factors (distal, intermediate, and direct) as well as a priori factorial structure identifying the person's socioeconomic status as a behavioral conditioner, reflected in diet and a sedentary lifestyle (Figure 1). (Lifestyles including alcohol and tobacco consumption, physical activity, and dietary changes, reported being associated with the occurrence of metabolic syndrome, influence increases in blood glucose levels as a predictor of diabetes. Pathways that would directly or indirectly affect blood glucose levels were posed and lifestyles were assumed usual pattern behaviors in the respondents [12–17].
The response variable, type 2 diabetes, was defined based on the blood glucose level, measured by the glycated hemoglobin test. Three categories were stipulated: a) norm glycemic: individuals with blood glucose levels lower than 5.6%, b) pre-diabetic, individuals with blood glucose levels between 5.7% and 6.4%, and e) diabetic, those with blood glucose levels higher than 6.5%.
Any factor that could confuse the relationship that was sought through the study was presented as a variable with direct effects on both the exposure of interest and the dependent variable; this was the case for independent variables such as age, sex, ethnicity, tobacco, and alcohol consumption. Measurement errors were assumed independent of each other, indicators as endogenous and factors as exogenous variables that varied and covaried freely. The standard reflexive measurement of the factors was used, assuming that there were positive intercorrelations between the indicators of the same factor, added to the fact that the exogenous variables contain measurement error terms that had to be manifested at the level of the indicator and not at the level of the latent compound. It is worth noting that each error term represented a unique variance, and it is a proxy indicator of all residual variation sources that were not explained by the model [18].
When variables like socioeconomic status, present a priority effect on other variables, such as diet and physical activity, the latter was considered as endogenous variables that have a disturbance (residual error) designated as "D" to represent all the unmeasured effects of the factor and not only those derived from the measurement error. In turn, physical activity and diet would, directly and indirectly, affect blood sugar levels, mediated through a factor called metabolic syndrome. It is worth noting that priority effects between latent variables and the variations of the disturbances were calculated by controlling the measurement error in the observed variables.
In this research, several alternative models were considered that played an important role in the phase of re-specification and detection bias due to the presence of equivalent models.
The analysis was conducted in four stages: model specification, identification, estimation, and assessment. A specified and over-identified model was available. The method of robust weighted least squares was used to estimate the SEM, to salve the inconvenience of violating the multivariate normality assumption [19]. STATA/SE V.11.1 and descriptive statistics were used for database debugging, and then the Lavaan package of the R software version 3.3.3 was used for model estimation [20]. Inherent weights to the sampling design were not used in the adjustment of the estimates since the calculation of the results was based on the analysis of a subsample formed from the filters of the inclusion-exclusion criteria, so that these weights were not inherited.
The model fit was assessed in terms of the magnitude and significance of the estimated parameters, and the variance explained by the independent variables of interest. Several statistical measures were used to derive the correspondence of the model with the data: the goodness of absolute fit (Chi-square), the relative fit (CFI), and the parsimonious fit (RMSEA). Since none of these procedures provides the information needed to assess the model, they were used simultaneously. Modification of the indices was used to evaluate and select the specific routes or mechanisms identified by the model with the best fit, a p-value<0.05 was considered statistically significant.
Among the 423 participants, 86 (20.33%) were classified as norm glycemic, 141 (33.33%) as pre-diabetic, and 196 (46.33%) as diabetic. 65.01% were female, age range 12-98 years, the highest proportion of people with type 2 diabetes, 27.04%, were between 40 and 50 years of age. 62.7% of the population earned less than the current legal minimum wage and 51.3% fell within a range of 6-11 years of schooling.
When comparing the group of diabetics and pre-diabetics with norm glycemic individuals, at a significance level of 5%, it is observed that the former were more likely to have high diastolic blood pressure, 7.09% and 16.84% compared to 16.28%, p=0.008, have high LDL cholesterol levels 6.63% and 2.13% vs. 0.0%, p=0.001, consume alcoholic beverage 10.64% and 5.51% vs. 17.44%, p=0.008 and have high total cholesterol levels 20.57% and 18.37% vs. 12.79%, p= 0.006 (Table 1).
The estimation process of the first over-identified recursive structural model, with a measurement of 4 factors and 23 indicators, converged to an admissible solution. However, the model failed the Chi-square test whose value was equal to 474.94, with 269 degrees of freedom and a p-value of 0.00. Besides, it presented poor convergent validity, the low explanatory power of the factor on the variability of its constitutive indicators, and an AVE (Average Variance Extracted) less than 0.5 for most of the constructs, which suggests that more factors are required because the variables involved do not measure a common domain. RMSEA (90% CI) equal to 0.045 (0.030-0.052), p-value for RMSEA equal to 0.87.
The modeling consisted of the following steps to locate the source
Source: Compiled by author.
≤ $589.500b 53 61.63 85 > $589.500 33 38.37 56 Body-mass index Underweight (≤18.5) 0 0.00 3 Normal (18.5 - 24.9) 39 45.35 64 Overweight (25.0 - 33 38.37 46 29.9) Obesity (≥ 30) 14 16.28 28 Waist-to-hip ratio Men ≤1 10 34.48 21 > 1 19 65.52 31 Women ≤0.8 1 1.75 4 >0.8 56 98.25 85 Triglycerides (mg/dL) ≤150 46 53.49 70 151-199 14 16.28 37 200 – 499 26 30.23 34 Blood pressure (mm/Hg) Systolic pressure ≤ 130 78 90.70 130 >130 8 9.30 11 Diastolic pressure ≤ 85 72 83.72 131 >85 14 16.28 10 Cholesterol HDL (mg/dL) <40 35 40.70 72 40 – 60 39 45.35 40 >60 12 13.95 29 Cholesterol LDL (mg/dL) <100 46 53.49 52 100 – 129 21 24.42 38 130 – 159 17 19.77 39 160 – 189 2 2.33 9 > 190 0 0.00 3 Total Cholesterol (mg/dL) <200 63 73.26 78 200 – 239 12 13.95 34 ≥ 240 11 12.79 29 Smokes Si 6 6.98 10 No 80 93.02 131 Alcohol Consumption a 0.394 0.769 a 0.638 0.490 0.08 0.025 0.073 0.001 0.006 0.985 0.008 Si 15 17.44 15 No 71 82.56 126 Exercise Si 23 26.74 45 No 63 73.26 96 a b Fisher’s exact test, legal minimum wage (2013) of poor fit: first, the measurement model was proposed (Figure 2) and once an acceptable model was reached, the structural part was evaluated. The factorial model was reformulated eliminating those indicators that did not have a significant factorial load and that presented notable measurement problems, configuring them into potential factors of mismatch in the model, such as the waist-tohip ratio, salt intake, sugar intake, intense physical activity at work and level of economic income.
The first outlined model (Figure 2) presented a convergent and admissible solution, and it fitted the data well with an x value of 67.87 and an associated p-value of 0.255. The values of other adjustment statistics are favorable (RMSEA=0.017, CFI= 0.98). More convergent validity is evident in this specification and correlations between factors ranging from -0.162 to 0.312, i.e. moderate correlations indicating discriminant validity. The re-specification required examination of modification indices that suggested error covariance between weight and size indicators belonging to the obesity factor, and LDL vs. triglycerides of the lipid profile factor. Under the mentioned arguments and considering the pertinent modifications, a second structural model was configured (Figure 3).
Variables related to diet and socioeconomic status were excluded from the model because of their poor convergent validity, low reliability of the constituent indicators, more than 90% of the variability explained by measurement error, and increased overall model mismatch. Covariates such as sex, age, tobacco, and alcohol consumption were also tested. However, they did not present statistical significance. A substantial improvement in adjustment indicators was evidenced compared to the first model. The χ statistic value was 44.58, with an associated p-value of 0.447, so the exact fit hypothesis is not rejected (Table 2).
0.118 The RMSEA = 0.006 with a confidence interval that has not a 5% value, indicates that the hypothesis of a close fit is not rejected, this conclusion is corroborated by the p-value=0.009. The relative adjustment index CFI, went from 0.729 in Model 1, to 0.999 in Model II, a value very close to 1.0, which indicates the best adjustment. Estimates of factorial loads, variances, and covariance’s of Model II are shown in Table 3. The standardized loads for the lipid profile and blood pressure constructs are high, with a range of variation of 0.530 - 0.928 and 0.608 - 0.800 respectively.
Although the physical activity construct presents very low factorial loads in its indicators, AVE = 12.6%, the panorama is different for the lipid profile and blood pressure constructs, which present high factorial loads, so that they can explain, on average, 50.2% and 50.4% of the variability of their constituent indicators, respectively. AII error variances are significant, except the one associated with the arterial index (0.238 which rejects the hypothesis that this index is measured with little error. AII error covariance’s are significant and no significance was found in the association between physical activity and obesity. This correlation is not very large, but its presence helps to "clean" some local adjustment problems in some parts of the measurement model (Table 3).
The lipid profile effect on glycosylated hemoglobin is significant, with a value of 0.608. An increase of one point in the lipid factor, consisting of LDL, triglycerides, and arterial index, predicts an increase in the blood glucose level of 0.608 units in its original metric. Statistical significance was also found in the trajectory coefficient directed from the physical activity towards the lipid profile; so that an increase in physical activity generates a decrease of 0.318 in the lipid profile. Similarly, an increase in the lipid profile causes an increase of 0.768 units in the blood pressure factor, a relationship found to be statistically significant (Table 4).
Model I (Figure 2) 0.045 (0.039 – 0.052)
χ 474.944 Obesity's indirect effect on blood glucose levels, mediated by the lipid profile, is equal to 0.466. The rationale for this derivation is as follows: even though obesity does not seem to have a direct effect on blood glucose level, it has a direct effect on lipid profile (0.768), and part of this effect (0.608) is transmitted to blood glucose levels. The 0.466 result indicates that it is expected for sugar Model II (Figure 3) 44.588 269 44 0.000 0.447 0.006 (0.000 – 0.035) 0.872 0.999 0.729 0.999 blood levels to increase in 0.466 units due to obesity increase in one point, mediated through the lipid profile. The total effect of physical activity on blood glucose level is -0.344, i.e. a decrease of 0.344 units is observed in blood glucose level due to physical activity increase in one unit, through all causal relationships direct and indirect, between these two variables (Table 4).
A model was constructed under the hypothesis proposed in previous studies demonstrating causal associations between sociodemographic, behavioral, anthropometric, and metabolic risk factors, considered responsible for the occurrence of type 2 diabetes [21–25]. The fundamental hypothesis of this model is that the effect of the socioeconomic status on the level of glucose in the blood is indirect and other mediators like diet and physical activity, which in turn affect it directly and indirectly through another factor named metabolic syndrome [26–28], transmit it.
The results of the first stated model showed a poor representation of the constructs for the constitutive indicators in the model of measurement, which represent latent variables that may account, in average, for less than 20% of the variability of its indicators, lack of convergent validity and most of the explained variance is charged to the measurement error variances. In the assessment of the model goodness of fit to data, some discrepancies were found between the predicted and the estimated values (χ = 474.944, M CFL = 0.729, RMSEA = 0.045) which are higher than the expected by chance, for this reason, this model was rejected as the explanation of the chain of causalities of the type 2 diabetes, and the model re-specified.
The resulting model did not take into consideration the socioeco-
Factor Loads Variances and covariances Factor Variances Parameter Not standardized Physical Activity Time physical activity 1.000 Time cycling or walking 1.452 Attend gymnasium 1.760 Lipid profile LDL 1.000 Triglycerides 1.452 Arterial I. 1.760 Obesity Weight 1.000 Size 0.182 Abdominal perimeter 1.191 Blood pressure Systolic pressure 1.000 Diastolic pressure 0.652 Time physical activity 987.510 Time cycling or walking 1504.679 Attend gymnasium 0.016 LDL 873.743 Triglycerides 2704.319 Arterial I. 250.877 Weight 82.561 Size 78.253 Abdominal perimeter 55.096 Weight Size 41.230 -404.201 LDL Triglycerides Time walking- cycling Diabetes 183.510 Physical activity Obesity 5.041 Diabetes 2822.662 Physical activity 55.299 Lipid profile 446.507 Obesity 87.834 Blood pressure 2287.571 Standard error P-value Standardized - - 0.230 0.816 0.052 0.291 0.004 0.007 0.492 - - 0.604 0.226 0.000 0.530 0.275 0.000 0.928 - - 0.718 0.063 0.004 0.190 0.206 0.000 0.833 - - 0.800 0.194 0.001 0.608 462.612 0.033 0.947 557.233 0.007 0.915 0.005 0.002 0.758 99.639 0.000 0.635 321.380 0.000 0.719 212.427 0.238 0.139 16.747 0.000 0.485 6.127 0.000 0.964 22.567 0.015 0.307 6.766 0.000 0.513 141.124 0.004 -0.263 75.740 0.015 0.089 7.344 0.492 0.072 410.113 0.000 0.934 37.386 0.139 1.000 97.665 0.000 0.890 17.686 0.000 1.000 750.757 0.002 0.909
nomic status or the diet as factors explaining the relationship with blood glucose level, directly or indirectly. Unlike other models [27,29] indicating significant causal associations between these factors and diabetes, the absence of contribution is possible due to the omission of more important predictive variables or to measurement errors of indicators in the definition of factors of diet and socioeconomic status. This is based on the poor convergent validity evidenced, the indicators low reliability, and the high proportion of variability explained by the measurement error, about 90%. This lack of consistency does not mean that direct effect between them does not exist, what it does point out is that its constitutive indicators are not well defined or measured, which causes confusing causal inferences on how to be useful for clinical or public health purposes.
The second model specified indicates that physical activity, blood pressure, and lipid profile have a direct effect on blood glucose levels. So physical activity causes variations on blood pressure, obesity, and lipid profile that are factors directly responsible for the modification of blood glucose.
The model showed a good overall adjustment (χ = 44.588, CFI = M 0.999, RMSEA = 0.006) its measurement component presented a mixed convergent validity; on the one hand, an explanation of important variability in the lipid profile, blood pressure, and obesity constructs (0.502, 0.504 y 0.415) though it is still a small one for the physical activity factor that explains only 12.6% of the average variability of the constituent indicators.
Physical activity protective effect on blood glucose levels is not direct but is mediated by the lipid profile (-0.193). The lipid profile constituted by the variables LDL, triglycerides, and the arterial index has a positive effect constituted on the levels of sugar in the blood (0.608), as well as it has it on the factor arterial pressure (0.768). No significant associations were found between obesity and physical activity; not between obesity and the lipid profile factor and blood pressure with blood glucose levels.
Blood pressure direct effect on blood glucose levels was confused by the lipid profile and, in part, by obesity represented by abdominal girth, weight, and height. This finding is supported by a previous study that found an increase in insulin resistance affecting glucose metabolism in patients with hypertension, regardless of the level of obesity compared to their controls with normal blood pressure [30].
Physical activity plays an important role in controlling factors associated with diabetes occurrence. increases in the level of physical activity were associated with decreases in dyslipidemia (-0.318) and high blood pressure (-0.244), results consistent with those obtained by Bardenheier et al. [27] in the modeling of diabetes risk factors using SEM.
No relationship was found between socioeconomic status and diet constructs, with factors such as alcohol intake, tobacco consumption, and race/ethnicity recognized in the disease production chain. This could be due to the omission of variables necessary for construct construction, to the lack of precision in measurement instruments, to ambiguity in question formulation resulting in imprecise responses, to the presence of memory and complacency bias, especially concerning behaviors subject to social control.
Among this study limitations, we have: the reliability and consistency of the measurement instrument used, although the instrument was validated in similar contexts; the error or bias in the measurement of the variables; observer errors, in this case of interviewer errors on the field, of individuals who accepted to participate in the interview, for this reason, a generalization of the empirical findings to new environments, contexts or populations is difficult (external validity), and to the set of variables that were not collected and that can affect the construct measurement [31]. Similarly, if the occurrence of diabetes were largely due to the accumulation of the deleterious effect of complex social processes throughout a lifetime, these would be better understood if studied longitudinally. The unique cross-sectional nature of this study, which was not repeated, imposes certain limitations on the causal inference, affecting the external validity of the study.
The main strength of this study lies in the formulation of a model that allows for the simultaneous evaluation of both the formation and the measurement of non-observable variables (theoretical constructs) from the interweaving of causal relationships for the testing of theoretical hypotheses based on the analysis of empirical data, which allows simulating the way these interrelations are presented in a complex, multifactorial and multilevel world [6].
The Institutional Committee of Human Ethics (CIREH) of the Health Faculty of the Universidad Del Valle approved the protocol, in September 2016.
Thanks to the Center for the Development and Evaluation of Public Health Policy and Technology (CEDETES) of the Universidad del Valle, for allowing the use of the database resulting from the survey on Risk Factors for Noncommunicable Diseases in Cali, 2013. Special thanks to Vanessa Bergonzoli for reviewing the English grammar.