Changes in intake of dairy product subgroups and risk of type 2 diabetes: modelling specified food substitutions in the Danish Diet, Cancer and Health cohort

We investigated the association between an increased intake of one dairy product subgroup at the expense of another within a 5-year period and the subsequent 10-year risk of developing type 2 diabetes. Effect modification by the initial level of the dairy product intake was also examined. The cohort included 39,438 men and women with two measurements of diet assessed using food frequency questionnaires (FFQ) administered in 1993-1997 and 1999-2003. Dairy products were skimmed milk, semi-skimmed milk, whole-fat milk, buttermilk, low-fat yogurt, whole-fat yogurt, cheese and butter. Type 2 diabetes cases were ascertained from the Danish National Diabetes Register. The pseudo-observation method was used to calculate risk differences (RD) with 95% confidence intervals (CI). Among participants aged 56-59 years at hand-in of the follow-up FFQ, increased intake of whole-fat yogurt in place of skimmed, semi-skimmed or whole-fat milk was associated with a reduced risk (RD% [95% CI]: -0.8% [-1.3, -0.2]; -0.6% [-1,1, -0.1]; -0.7 [-1.2, -0.1]; per 50g/d, respectively). Among participants aged 60-64 and 65-72, substitution of skimmed milk for semi-skimmed milk was associated with an increased risk of type 2 diabetes (0.5% [0.2, 0.7]; 0.4% [0.1, 0.7]; per 50g/d, respectively). No effect modification by the initial intake of milk products was observed. Our results suggest that substitution of whole-fat yogurt for milk among those aged 56-59 decreases the risk of type 2 diabetes and substitution of skimmed milk for semi-skimmed milk was associated with an increased risk of type 2 diabetes among those aged 60-64 and 65-72, regardless of the initial intake level.

CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05. 28.20116459 doi: medRxiv preprint Introduction 54 The incidence of type 2 diabetes is increasing rapidly worldwide [1] and the prevention of type 2 55 diabetes is a top priority in public health [2]. Diet, together with other lifestyle factors, plays a key 56 role in prevention of type 2 diabetes. Indeed, risk of type 2 diabetes has been found to remain lower 57 at least 10 years after participation in lifestyle interventions that included dietary modifications 58 lasting two to six years [3]. 59 Current dietary recommendations to the general population [4,5] and for prevention of type 2 60 diabetes [6,7] recommend low-fat dairy products instead of whole-fat dairy products [8]. However, 61 meta-analyses of randomized interventions have found no clear effect of increasing the total intake 62 of dairy products on glucose metabolism or insulin sensitivity and no effect of increasing the intake 63 of low-fat or high-fat dairy products on cardio-metabolic risk markers, expect for a slightly 64 increased body weight [9][10][11]. This lack of an effect may be due to small sample sizes and short 65 durations of the included trials and to differences in types of dairy products provided and control 66 groups used, either consuming their habitual diet, a low-dairy or no-dairy diet, between the included 67 trials. Only few cohort studies have investigated the association between changes in intake of dairy 68 products on the risk of developing type 2 diabetes [12]. Of the few cohort studies that investigated 69 changes in intake of dairy products, higher consumption of yogurt, not milk products, was 70 associated with a lower risk of type 2 diabetes [13]. To be sure that assessed associations are 71 independent of changes in energy intake, it is important to adjust for energy intake in the statistical 72 model. Consequently, a food substitution is introduced where the model now reflects the influence 73 of increasing the intake of, for example, yogurt products, as in the example, at the expense of a mix 74 of other foods on the risk of type 2 diabetes; mimicking an isocaloric intervention study. As all 75 foods are either harmful, neutral or beneficial in the prevention of type 2 diabetes, the association 76 between intake of a specific dairy product subgroup and the risk of type 2 diabetes depend on the 77 substituted food. Unfortunately, this has rarely been specified, resulting in an unclear interpretation 78 of the results. 79 Another relevant question when investigating dietary changes is whether the initial intake 80 modifies the association between a dietary change and development of disease. In the context of 81 changes in dairy product subgroups, the association between substitution of, for instance, skimmed 82 milk for whole-fat milk and the risk of type 2 diabetes may be different, depending on whether the 83 initial intake of whole-fat milk was high or low. 84 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05.28.20116459 doi: medRxiv preprint 5   Specifying food substitutions when studying associations between dietary changes and long-term  85   disease risk may add important new knowledge on the public health impact of changing diet. Thus,  86 we investigated the association between an increased intake of one dairy product subgroup at the 87 expense of another within a 5-year period, and the subsequent 10-year risk of developing type 2 88 diabetes in middle-aged men and women. In addition, we examined whether the association was 89 modified by the initial intake of these products. 90 91

Study population 93
The Diet, Cancer and Health cohort is a Danish cohort that recruited participants from 1993-1997. 94 Middle-aged men and women from the Aarhus and Copenhagen areas were invited to participate in 95 the study. At inclusion in 1993-1997 participants answered a validated food frequency 96 questionnaire (FFQ) mailed to the participants [14,15] . Furthermore, the participants underwent a 97 physical examination and answered a lifestyle questionnaire at one of two study centres [16]. In 98 1999-2003, participants from the original cohort answered mailed follow-up questionnaires on diet 99 and lifestyle similar to the questionnaires provided in 1993-1997. 100 For the purposes of the present study, participants with a diagnosis of diabetes, myocardial 101 infarction, stroke or cancer before the time of the hand-in of questionnaires in 1999-2003 were 102 excluded. Furthermore, participants for whom we did not have information on exposures or 103 covariates were also excluded. 104 105 Dairy product subgroups 106 The initial intake of dairy product subgroups (skimmed milk 0.1% fat, semi-skimmed milk 1.5%, 107 whole-fat milk 3.5%, buttermilk 0.5%, low-fat yogurt <1.5%, whole-fat yogurt 3.5%, cheese and 108 butter) was assessed in 1993-1997 using a validated FFQ designed for the Diet, Cancer and Health 109 cohort [14,15]. The FFQ inquired participants about their average intake of the dairy product 110 subgroups during the previous 12 months according to 12 possible categories ranging from 'never' 111 to 'eight times or more per day'. Low-fat yogurt and whole-fat yogurt included plain yogurt types 112 and yogurt with fruit. Participants handed in the FFQ at one of the two study centres and it was 113 checked by trained personnel that immediately addressed inconsistencies of the questionnaires with 114 the participants, if necessary. In 1999-2003, a FFQ similar to the 1993-1997 FFQ was mailed to the 115 participants and the returned FFQs were checked by trained personnel. Missing values were scored 116 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05.28.20116459 doi: medRxiv preprint 6 in terms of the type and relevance of the question (0, 0.5, 1 or 10 points). Participants with >9 117 points were contacted by phone. If participants had >55 points, the FFQ was mailed back to the 118 participants. The remaining missing values were coded as the lowest intake level "never/seldom". 119 Harmonization of the dietary data collected in 1993-1997 and 1999-2003 is described in the 120 Supplemental text. Food and nutrient intakes in g/day were calculated using FoodCalc [17] and 121 Danish food composition tables [18] version 1996. Changes in intake of dairy product subgroups 122 were defined as the difference in reported intake in 1993-1997 (initial intake) and reported intake in Registry (excluding females prescribed metformin alone). The algorithm has been validated with a 136 positive predictive value of 89% as compared to medical records from the general practitioner [19]. 137 The register does not distinguish between type 1 and type 2 diabetes. However, as all participants in 138 the Diet, Cancer and Health cohort were middle-aged, the vast majority of incident cases is 139 expected to be type 2 diabetic cases. The participants were followed up from the hand-in of 140 questionnaires in 1999-2003 until the date of the diagnosis of type 2 diabetes, death, emigration or 141 the date at which participants had been followed for 10 years after hand-in of questionnaires in 142 1999-2003, whichever came first. It is not known when the participants changed their dietary intake 143 during the 5 years between the two FFQs, but 10 years of subsequent follow-up was considered 144 reasonable based on evidence from clinical trials showing a lower risk of type 2 diabetes for up to 145 10 years after initiation of lifestyle interventions lasting two to six years [3]. 146 147 Assessment of covariates 148 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05.28.20116459 doi: medRxiv preprint Information on covariates was obtained from the lifestyle questionnaires and FFQs administered in 149 1993-1997 and 1999-2003. The analyses were adjusted for known risk factors of type 2 diabetes 150 selected a priori based on a review of the literature. From the 1993-1997 questionnaires, 151 information on education, history of hypertension, history of hypercholesterolemia, total energy 152 intake and food groups associated with type 2 diabetes was obtained. From the questionnaires in 153 1999-2003, information on alcohol intake, smoking status, physical activity, body weight and waist 154 circumference and updated information on history of hypertension, history of hypercholesterolemia, 155 total energy intake and food groups was obtained. In addition, participants self-reported their family 156 history of diabetes. Height was measured at the physical examination in 1993-1997. 157 158

Statistical analysis 159
We calculated medians and 80% central ranges for continuous variables, and proportions for 160 categorical variables among the cohort and cases. 161 We used the pseudo-observation method [20,21] with age as the underlying time-scale to 162 calculate risk differences (RDs) and their corresponding 95% confidence intervals (CIs). A central 163 assumption of the pseudo-observation method is that there must be marginally independent entry 164 and censoring. We investigated the validity of this assumption using a previously described 165 methodology stratifying the cohort based on age and date of inclusion, which was the date at hand-166 in of the questionnaires in 1999-2003 [22]. Because age was the underlying-time scale, a specific 167 age was set as the end of the follow-up period as required for the generation of the pseudo-values. 168 We used the median age at which participants were followed-up for 10 years after hand-in of 169 questionnaires in 1999-2003 as the end of the risk period. Risk periods were 56-67 years, 60-71 170 years and 65-76 years, in stratum 1, 2 and 3, respectively. Death during follow-up was handled as a 171 competing risk. The substitution models included all dairy product subgroups as change variables 172 and for all dairy product subgroups assessed in 1993-1997. The difference between two change-173 variables' beta-coefficients (the beta-coefficient for the product increased minus the beta-coefficient 174 for the product decreased) and their variances and covariance were used to estimate the RDs and 175 corresponding 95% CIs for a given substitution. Hence, the association of an increased intake of a 176 specific dairy product subgroup and a concomitant decreased intake of another specified dairy 177 product subgroup (i.e. a substitution) in the 5-year period from 1993-1997 to 1999-2003, with the 178 subsequent 10-year risk of developing type 2 diabetes, was estimated. In multivariable analyses we 179 included total energy intake (1993-1997 and 1999-2003, kJ/day; continuous) and sex (man, woman) 180 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05.28.20116459 doi: medRxiv preprint 8 in Model 1a. Total energy intake at both time-points was included to assess the association between 181 substitutions independent of total energy intake at both time-points and to reduce measurement 182 error of reported food intake [23]. Consequently, an unspecified residual substitution was 183 introduced, which is equal to the difference in energy provided by the substituted products. Model 184 1b was further adjusted for the duration of education (≤7, 8-10, >10 years), assessed in 1993-1997, 185 and alcohol intake (g/day; continuous, as a restricted cubic spline with three knots), smoking status 186 (never, former, current <15 g tobacco/day, current 15-25 g tobacco/day, current >25 g tobacco/day), 187 physical activity (participating in leisure-time sport activities or not), and family history of diabetes 188 considered as mediators rather than confounders. Linearity for the dairy product variables was 195 investigated using restricted cubic spline of each dairy product subgroup. We constructed a 196 correlation matrix of dairy product intake, assessed in 1993-1997, and changes in dairy product 197 intake in order to assess potential multicollinearity (i.e. one variable almost completely explains 198 another variable) as changes in dairy product variables may be highly correlated with the initial 199 intake of dairy product subgroups. 200 We investigated potential effect modification by the initial intake level of the replaced dairy 201 product group assessed in 1993-1997 in subgroup analyses. For instance, the association between 202 substitution of skimmed milk for whole-fat milk and risk of type 2 diabetes in subgroup analyses by 203 initial low or high whole-fat milk intake assessed in 1993-1997 (cut at the 50th percentile). We also 204 conducted sex-stratified analyses. 205 As a sensitivity analysis, we excluded participants who developed hypertension and/or 206 hypercholesterolemia between 1993-1997 and 1999-2003. This was done because these participants 207 are likely to have changed their dietary intake as a result of diagnosis and treatment. Furthermore, 208 non-participation at follow-up may result in selection bias. Thus, in another sensitivity analysis, we 209 adjusted for potential selection bias by using inverse probability weighting [24]. The probability 210 weights were based on sex, age, education, alcohol, smoking, physical activity, BMI, waist 211 circumference, history of hypertension, history of hypercholesterolemia, history of diabetes and 212 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05.28.20116459 doi: medRxiv preprint 9 alternate healthy eating index-2010 score [25] assessed in 1993-1997. Also, we compared 213 characteristics assessed with the 1993-1997 questionnaires between those who participated at both 214 time-points and those who only participated in 1993-1997. Changing two dietary components may 215 be followed by other dietary changes as well. We constructed spider plots to visually assess 216 differences in changes in intake of different food groups between participants that substituted one 217 for another dairy product subgroup, compared with participants who maintained a stable intake of 218 both dairy product subgroups. Substitutions were defined as increased intake of one dairy product 219 subgroup by ≥1 serving/day and decreased intake of another dairy product subgroup by ≥1 220 serving/day. To investigate the time-scale sensitivity of our results, we compared risk ratios from 221 the analyses with age as the underlying time-scale with risk ratios from analyses using time in the 222 study as the underlying time-scale. We compared risk ratios and not RDs as risk ratios are generally 223 less sensitive across studies than risk differences. Hence, differences between time-scales using risk 224 ratios would be of a greater concern than dissimilarities using risk differences. Lastly, we excluded 225 7179 participants with >9 missing values in their 1999-2003 FFQ coded as "no intake or less than 226 once per month" in the main analysis. 227 All analyses were performed in Stata version 15.1 (StataCorp, College Station, Texas, USA). 228 Forest plots were, however, coded in R version 3.5.0, R Core Team, Vienna, Austria, using the 229 ggepi package [26]. The assumption of independent entry within the entire cohort did not hold (hazard ratio [95% CI] 238 for type 2 diabetes comparing first and third tertile of age at inclusion: 1.26 [1.11, 1.42]). Based on 239 this result and a previous study [27], we analysed substitutions within three strata based on age at  (Table 1). There were no substantial differences in the median changes in 256 intake of dairy product subgroups between the cohort and those who developed type 2 diabetes. 257 However, the ranges of changes in intake (p10-p90) were different between the two (Table 1). 258 For the main substitution analyses we found different patterns across the age-strata (Figure 2  is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05.28.20116459 doi: medRxiv preprint Figure 3 shows the associations for substitutions between dairy product subgroups and risk of 276 type 2 diabetes stratified by the initial intake level of the replaced dairy product group. The patterns 277 of associations were largely similar between those with an initial low and high intake of the 278 substituted dairy product subgroup. When the analyses were stratified by sex, no associations 279 between substitutions of whole-fat yogurt for milk and risk of type 2 diabetes were observed among 280 women aged 56-59 whereas substitution of whole-fat yogurt for milk was associated with a reduced 281 risk in men aged 56-59. Substitution of skimmed milk for semi-skimmed milk was associated with 282 an increased risk among women aged 60-64 and 65-72 and only among men aged 60-64 283 (Supplemental Figure 2). 284 After excluding participants who developed hypertension or hypercholesterolemia between the 285 two dietary assessments, replacement of milk products with yogurt became more strongly 286 associated with a reduced risk of type 2 diabetes (Supplemental Figure 3). Otherwise, similar results 287 as in the main analyses were observed. Supplemental Figure 4 shows the flow of participants that 288 participated both in 1993-1997 and in 1999-2003 and those who only participated in 1993-1997. 289 Non-participants (i.e. those who only participated in 1993-1997) were more likely to be younger, 290 have a lower education, consume less alcohol, be smokers, be less physically active, have a higher 291 BMI and waist circumference and have a lower healthy eating index score in 1993-1997 compared 292 to those who answered both 1993-1997 and 1999-2003 questionnaires (Supplemental Table 1). 293 Adjusting for non-participation using inverse probability weighting provided similar results as the 294 main analyses (Supplemental Figure 5). Supplemental Figure 6 illustrates differences in intake of 295 other foods related to type 2 diabetes comparing participants that have increased their intake of ≥1 296 serving/day of, for example skimmed milk, and decreased their intake of ≥1 serving/day of, for 297 example semi-skimmed milk (i.e. substituted skimmed milk for semi-skimmed milk), with 298 individuals that did not change their intake of the dairy product subgroups. Subjects that substituted 299 low-fat dairy products, particularly skimmed milk or low-fat yogurt for other dairy products, 300 increased their intake of fruit and vegetables to a greater extent than those with a stable intake. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05. 28.20116459 doi: medRxiv preprint In this large cohort of middle-aged Danish men and women with repeated measures of diet, we 307 found different patterns of associations across age-strata for specified substitutions of dairy product 308 subgroups. We found that replacing milk products with whole-fat yogurt reduced the risk of type 2 309 diabetes among those aged 56-59 years. Replacing semi-skimmed milk with skimmed milk was 310 associated with an increased risk of type 2 diabetes among those aged 60-64 and 65-72 years. No 311 clear associations were observed for replacements with buttermilk, cheese or butter. 312 The primary exposures of this study were changes in intake of dairy product subgroups. To 313 investigate changes we used two measures of diet. The dietary intakes were self-reported and thus 314 measurement error is present. However, the FFQ was validated for nutrient intake and similar 315 FFQs, validated on a food level, tend to show a high correlation with food records for dairy 316 products (r = 0.53-0.90) [28][29][30]. We thoroughly adjusted for established risk factors of type 2 317 diabetes, but residual confounding cannot be excluded. Furthermore, after excluding those who 318 developed hypertension or hypercholesterolemia between 1993-1997 and 1999-2003, replacement 319 of milk with yogurt became more strongly associated with a reduced risk of type 2 diabetes. This 320 raise concerns about potential confounding by indication. However, the point in time when the 321 change in diet and development of hypertension or hypercholesterolemia was unknown. To our 322 surprise, we observed an increased risk of type 2 diabetes when replacing semi-skimmed milk with 323 skimmed milk among those aged 60-64 and 65-72 years. This is not consistent with the current 324 evidence, which mostly suggests no association between intake of milk, regardless of fat content, 325 and type 2 diabetes risk [12,31]. Other indicators of deteriorating health status (such as physical 326 disability or chronic inflammation) were not taken into account and thus, we cannot exclude the 327 possibility of confounding by indication here as well. All Danish men and women between ages 50-328 64 years, without any registered cancer diagnosis, living in the two largest cities in Denmark were 329 invited to participate in the Diet, Cancer and Health cohort. It has previously been shown that those 330 who chose to participate generally had a higher socioeconomic position [16], which limits the 331 generalizability of our results. We observed similar tendencies among participants eligible for this 332 study. Participants had longer duration of education than non-participants. Hence, non-participation 333 at follow-up of about 17% may have introduced selection bias. However, after using inverse 334 probability weights of participation in our analyses, we found similar results compared to the main 335 analyses, suggesting that selection bias is not a major concern. 336 Only few previous studies have investigated changes in intake of dairy product subgroups on risk 337 of developing type 2 diabetes. Interestingly, we found that replacing milk with yogurt products was 338 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05.28.20116459 doi: medRxiv preprint associated with a reduced risk of type 2 diabetes in those aged 56-59, but not those aged 60-72 339 years. As the effect of dietary intake on future risk of developing type 2 diabetes is likely 340 cumulative, changes among older populations may not have the same impact as dietary changes 341 among younger populations. For instance, although intake of fruit and vegetables have been 342 consistently associated with a lower risk of type 2 diabetes [32], these associations were not found 343 in a meta-analysis of cohorts of elderly individuals (>60 years) [33]. Alternatively, higher baseline 344 hazards among the elderly may also explain the findings from this meta-analysis and none of the 345 included studies investigated changes in intake of fruit and vegetables. Current dietary 346 recommendations emphasize substitution of high-fat dairy product with low-fat dairy products, 347 although few cohort studies have investigated the health impact of changes in intake of dairy 348 products. In a cohort of Spanish men and women aged 55-80 years, Diaz-Lopez et al. [13], reported 349 that an increased intake of low-fat and whole-fat yogurt from below the median intake to more than 350 or equal to the median intake was associated with a lower risk of type 2 diabetes compared with a 351 stable intake below the median, whereas no similar association was observed for low-fat or whole-352 fat milk. These findings support our findings among those aged 56-59. However, the authors did not 353 stratify according to different age groups and they did not specify relevant substitutions for the 354 changes in dairy intake. In a 9-week randomized intervention comparing commercially available 355 low-fat yogurt with a soy pudding, reduced levels of some, not all, measured biomarkers of chronic 356 inflammation and endotoxin exposure were observed in the low-fat yogurt group [34]. Overall, 357 these findings support a benefit of yogurt products on risk of development of type 2 diabetes 358 through mechanisms that may be mediated by changes in the gut microbiota, as suggested by others 359 [35]. 360 An advantage of specifying the food substitution is that it provides a clear interpretation of the 361 results. We specified the association for increasing the intake of one dairy product subgroup at the 362 expense of another. Our main model was Model 2, which also adjusted for other foods. However, it 363 could be argued, for public health recommendations, that Model 1b, not adjusted for other foods, 364 may be more relevant as this model did not constrain the underlying dietary pattern. For example, 365 substitution of low-fat yogurt for skimmed milk may be associated with changes in the intake of 366 other foods. We illustrated this using spider plots and generally found that those who increased their 367 intake of low-fat yogurt also increased their intake of fruit, vegetables and coffee compared to those 368 who did not change their intake of these dairy products. Indeed, after further adjustment for other 369 foods (Model 2) some of the observed associations (Model 1b) disappeared. Although the 370 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05.28.20116459 doi: medRxiv preprint magnitude of the RDs observed was rather small, around -0.9% to -0.5% for a substitution of 50 g/d 371 for whole-fat yogurt with 50 g/d for milk products, this is not surprising, as we only modelled a 372 single dietary change consisting of two different products. 373 We also investigated the modifying effect of the initial intake of the dairy product subgroup 374 being replaced. Overall, the pattern of associations was similar among those with an initial high or 375 low intake of the substituted dairy product. The low average initial intake of whole-fat dairy 376 products in this cohort may partly explain the lack of a modifying effect. In our sex-stratified 377 analyses we observed different patterns of associations between men and women. However, as there 378 is little evidence of biological reasons for such differences, we interpret this with caution and 379 encourage replication of this finding in other cohorts. 380 Future research should focus on randomized interventions and cohort studies with repeated 381 measures of diet to evaluate these suggested differences between yogurt and milk products on risk 382 of developing type 2 diabetes or risk markers hereof. 383 In conclusion, this study suggests a reduced risk of type 2 diabetes when replacing milk with 384 whole-fat yogurt among men and women aged 56-59 years and an increased risk of type 2 diabetes 385 when replacing skimmed milk for semi-skimmed milk among those aged 60-64 and 65-72, 386 regardless of initial intake level. 387 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05.28.20116459 doi: medRxiv preprint . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 29, 2020. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05.28.20116459 doi: medRxiv preprint Tables   Table 1 Characteristics at hand-in of follow-up questionnaires (1999)(2000)(2001)(2002)(2003) and changes in intake of dairy product subgroups of participants in the Diet, Cancer and Health cohort in three strata divided into tertiles of age 1 .  1 All medians and percentiles are reported as pseudo-medians and pseudo-percentiles (i.e. an average of the five surrounding values of each percentiles) in order to comply with European data protection regulations. Figure 1 Flowchart of the study sample in the Diet, Cancer and Health cohort. FFQ: food frequency questionnaire.

Figures
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2020. . https://doi.org/10.1101/2020.05.28.20116459 doi: medRxiv preprint Figure 2 Forest plot of substitutions between dairy product subgroups on 10 years risk of type 2 diabetes in the Diet, Cancer and Health cohort across three strata divided into tertiles of age at hand-in of questionnaires in 1999-2003. 10-year risk periods were 56-67 years, 60-71 years and 65-76 years, in strata 56-59, 60-64 and 65-72 years, respectively. Intakes were modelled per 50 g/day for all dairy products except for cheese (20 g/day) or butter (6 g/day). Adjusted for Model 2 covariates: total energy intake (assessed in 1993-1997 and 1999-2003), sex, duration of education (assessed in 1993-1997) and alcohol intake, smoking status, leisure-time sports activity and family history of diabetes, fruit, vegetables, red and processed meat, sugar-sweetened beverages, whole grains, potato chips and coffee (assessed in 1999-2003). The larger and darker the dots, the lower the p-value for a given association.

Figure 3
Forest plot of substitutions between dairy product subgroups on risk of type 2 diabetes, followed-up for subsequent 10 years, in the Diet, Cancer and Health cohort across three strata divided into tertiles of age at hand-in of questionnaires in 1999-2003 and low or high initial intake of the substituted dairy product, cut at the 50th percentile. 10-year risk periods were 56-67 years, 60-71 years and 65-76 years, in strata 56-59, 60-64 and 65-72 years, respectively. Intakes were modelled per 50 g/day for all dairy products except for cheese (20 g/day) or butter (6 g/day). Adjusted for Model 2 covariates: total energy intake (assessed in 1993-1997 and 1999-2003), sex, duration of education (assessed in 1993-1997) and alcohol intake, smoking status, leisure-time sports activity and family history of diabetes, fruit, vegetables, red and processed meat, sugar-sweetened beverages, whole grains, potato chips and coffee (assessed in 1999-2003). The larger and darker the dots, the lower the p-value for a given association.