Temperament as a predictor of eating behavior in middle childhood – A fixed effects approach

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.


Introduction
less because gut activity decreases in the presence of emotional arousal, normally suppressing 141 hunger and eating (Heatherton, Herman, & Polivy, 1991). Thus, although highly negative 142 affective children might be at risk for emotional overeating, they might be just as likely to 143 display more emotional undereating than less reactive children. We therefore hypothesize 144 that greater negative affectivity will also be prospectively associated with more emotional 145 undereating. Additionally, because fear, shyness and discomfort characterize negative 146 affectivity and fear makes humans more reluctant to try new foods (Pliner,Pelchat,& 147 Grabski, 1993) and possibly more likely to eat at a slower pace, we hypothesize that greater 148 negative affectivity will be prospectively associated with more food fussiness and slowness in 149 eating. As regards effortful control, which can be seen as a top-down self-regulatory-or 150 control process (Bridgett, Burt, Edwards, & Deater-Deckard, 2015), we hypothesize that 151 higher effortful control will predict lower food responsiveness, less emotional overeating, 152 higher satiety responsiveness and slowness in eating over time (i.e., better self-regulation of 153 eating). Put simply; in today's western 'obesogenic' environment where food is easily 154 accessible, we often have to decide actively whether, what and how much to eat -and those 155 who have well-developed self-regulation abilities (i.e., high levels of effortful control) are 156 probably more adept at regulating their intake according to their needs. The third responsiveness, emotional overeating and desire to drink), whereas children low in surgency 165 will become more 'food avoidant' (i.e., more food fussy, eating at a slower pace) over time.

167
Participants and Procedure 168 The 2003 and 2004 birth cohorts (N= 3,456) living in Trondheim, Norway, and their parents, 169 were invited to participate in the Trondheim Early Secure Study (TESS) (Steinsbekk & 170 Wichstrom, 2018), which the present study is built on. Because the primary aim of TESS was 171 to assess mental health, parents also received the Strengths and Difficulties Questionnaire 172 (SDQ) (Goodman, 1997) version 4-16, a brief measure of emotional and behavioral 173 problems, in addition to the invitation letter. Parents brought the completed SDQ when they 174 attended the well-child clinic for the routine health check at age 4 years, and the health nurse 175 obtained the parents' written consent to participate (5.2% of eligible parents were missed 176 being asked) (n = 2,475). Procedure and flow of participants are presented in Figure 1, and 177 additional details can be found in Steinsbekk & Wichstrøm (2018). Because almost all 178 children in the two cohorts appeared at the city's well-child clinic (97.2%) for the health 179 check-up (age 4), the sample is effectively a community sample. To increase sample 180 variability, children with higher SDQ scores (i.e., more problems) were oversampled. In 181 doing so, children were allocated to four strata according to their SDQ scores (cut-offs: 0-4, 182 5-8, 9-11, and 12-40), and the probability of selection increased with increasing SDQ scores 183 (0.37, 0.48, 0.70, and 0.89 in the four strata, respectively). To produce appropriate population 184 estimates, we accounted for this oversampling in the statistical analyses applied (see Results).

185
As can be seen in Figure 1 were higher level professionals, whereas 39% were lower level professionals; 26% were 193 formally skilled workers; 0.5% were farmers/fishermen and 3.1% were unskilled workers.

194
Differences in rates of occupations between the present sample and the Norwegian parent 195 population were negligible, and never exceeded 3.6% (Statistics Norway). The sample was 196 also comparable with the Norwegian parent population with regard to the parents' level of 197 education (Statistics Norway, 2012) and children's BMI (Juliusson et al., 2013). All  Food Responsiveness (range of internal consistency for age 6 to10: α = .65-.71; 5 items, e.g., 204 "Even if my child is full, she/he finds room to eat her/his favorite food"); Enjoyment of Food 205 (α = .81-.83; 4 items, e.g., "My child enjoys eating"); Emotional Overeating (α = .75-.77; 4 206 items, e.g., "My child eats more when worried"); Emotional undereating (α = .75-.78; 5 207 items, e.g., "My child eats less when upset"); Satiety Responsiveness (α = .70-.74; 5 items, 208 e.g., "My child gets full easily"); Food Fussiness (α = .89-.90; 6 items, e.g., "My child is 209 difficult to please with meals"); Slowness in Eating (α = .70-.72; 4 items, e.g., "My child eats 210 slowly"); and Desire for Drinks (α = .65-.71; 3 items, e.g., "My child is always asking for a 211 drink"). The CEBQ has been validated using objective measures of eating behavior (Carnell & Wardle, 2007), and it has been shown to have good test-retest reliability (Wardle, Guthrie, 213 Sanderson, & Rapoport, 2001).  specifying the relationship between model parameters to arrive at a best-fitting model, while 231 effectively handling missing data. Figure 2 illustrates the fixed effects model tested (details of 232 the model fitting procedure is displayed in supplemental material). Due to the high number of 233 parameters to be estimated relative to the number of children, not all eating behaviors could 234 be analyzed in one model. Separate models for each of the eight eating behaviors were 235 therefore created. In each model, eating behavior (e.g., Food Responsiveness) at ages 8 and 236 10 was regressed on temperament (i.e., negative affectivity, effortful control and surgency) at age 6, whereas eating behavior at age 6 was regressed on temperament at age 4. To include 238 unmeasured time-invariant effects and thus adjust for them, a fixed effects part was added to 239 each model by constructing a latent variable loading on the eating behavior in question. This 240 latent time-invariant variable was allowed to correlate with temperament at age 6, whereas 241 the correlations with temperament at age 4 were set to zero (because these must be 242 considered exogenous variables given that eating behavior (i.e., outcome variable) was 243 measured from age 6 onwards). Temperament variables at all time points were allowed to 244 correlate and age-6 temperament was allowed to correlate with concurrent eating behavior. In 245 addition, because we hypothesized that the influence of temperament on eating behavior 246 would increase with age, Satorra-Bentler qhi-square tests (Satorra & Bentler, 2001) were 247 used to examine such age differences by comparing the paths from temperament at age 4 to 248 eating behavior at age 6 with the corresponding age 6 to 8 paths.

249
When modeling the hypothesized paths from temperament to eating behavior, we examined 250 whether random or fixed effects fit the data best. Because of their exclusive reliance on 251 within-person variance, fixed effects models have limited statistical power. In contrast, a 252 random effects model utilizes between-person variance as well and is thus statistically more 253 powerful but presupposes that the predictors are uncorrelated with the latent time-invariant 254 factor -which may not necessarily be true. We therefore compared the random effects 255 models to the fixed effects models, testing differences in χ 2 . However, because differences in 256 χ 2 do not follow a χ 2 distribution when a robust maximum likelihood estimator is applied, 257 Satorra-Bentler's scaled χ 2 was used (Satorra & Bentler, 2001); which thus becomes a 258 functional equivalent to the Hausmann test (Allison, 2009). Furthermore, hybrid models (i.e., Parental socioeconomic status was neither associated with temperament nor eating behavior 268 and was therefore not included as a confounder in the analysis.

270
Descriptives are displayed in Tables 1 and 2, whereas bivariate correlations between all study 271 variables are presented in supplemental material (Table S1). The results of the model fitting

277
The parameter estimates from temperament to eating behaviors in each of the 278 preferred models are shown in Table 4 (food approach behaviors) and Table 5 (food avoidant   279 behaviors). At all time points examined, negative affectivity significantly predicted higher 280 levels of food responsiveness, emotional overeating, emotional undereating, satiety 281 responsiveness, food fussiness, slowness in eating and desire to drink, even when all 282 unmeasured time-invariant confounders were accounted for. Enjoyment of food was the only 283 eating behavior prospectively unrelated to negative affectivity, but this eating behavior was 284 significantly predicted by higher levels of effortful control, as was slowness in eating (ages 6 285 to 8 and 8 to 10 years). Lower effortful control, on the other hand, predicted more food fussiness at all time points, as well as greater food responsiveness from ages 4 to 6, emotional 287 overeating and desire to drink from ages 6 to 8. Children higher on surgency at age 6 were 288 more likely to enjoy food more and be more food responsive but displaying less satiety 289 responsiveness and less food fussiness at age 8. The diminished satiety responsiveness and 290 food fussiness were also still evident at age 10 (Table 5). Surgent children also displayed 291 more rapid eating over time, apart from the age 6 to 8 years lag (Table 5).

302
This study aimed to establish whether temperament is involved in the etiology of eating 303 behavior in middle childhood, by studying a sample of Norwegian 4-year olds followed up at 304 ages 6, 8 and 10, and applying a statistical approach that accounts for all unmeasured time-305 invariant confounders (e.g., genetics). We found that higher negative affectivity predicted 306 higher levels of food responsiveness, emotional overeating, emotional undereating, satiety 307 responsiveness, food fussiness, slowness in eating and desire to drink. Lower effortful control 308 predicted more food fussiness, food responsiveness, emotional overeating and desire to drink, 309 whereas higher effortful control predicted more enjoyment of food and slowness in eating, although not consistently through all time-points. Higher levels of surgency was 311 prospectively associated with more enjoyment of food and food responsiveness, as well as 312 lower satiety responsiveness, food fussiness and slowness in eating, but again, not 313 consistently through all time-points.

314
Negative affectivity. The results indicated that among the three temperamental 315 dimensions examined, negative affectivity was the one most consistently related to eating 316 behavior, which accords with a previous cross-sectional study of pre-schoolers capturing 317 several temperamental dimensions (Haycraft et al., 2011). As hypothesized, over time, 318 negative affectivity predicted more emotional over-and undereating, food fussiness, slowness 319 of eating and desire to drink. Although emotional distress may trigger eating (e.g., for those 320 who have learned that eating soothes negative emotions (Kaplan & Kaplan, 1957)), the most 321 natural response to distress is to eat less because gut activity decreases in the presence of 322 emotional arousal, normally suppressing hunger and eating (Heatherton et al., 1991;Van 323 Strien & Ouwens, 2007), possibly explaining why negative affectivity forecast both 324 emotional over,-and undereating. Research does show that emotions can both increase and 325 decrease food intake, but less is known about which emotional or individual characteristics 326 predict more or less eating (Macht, 2008). It might be, for example, that highly negative 327 reactive children eat more under positive circumstances and less during negative ones, being 328 especially malleable to environmental influences, for better or worse, as suggested by the 329 differential susceptibility hypothesis (Belsky & Pluess, 2009).

330
The fact that fear makes humans more reluctant to try new foods (Pliner et al., 1993) 331 and that negative affectivity is characterized by fear and related constructs such as shyness 332 and discomfort may explain why highly negative affective children become more food fussy 333 over time. Interestingly, negative affectivity also predicted more food responsiveness and 334 higher satiety responsiveness, the latter association possibly being due to high satiety sensitivity indicating a poorer or smaller overall appetite. This also fits with the finding that 336 negative affectivity predicted more slowness in eating (i.e., eating slower if reduced appetite), 337 which has also been found in a former study of young children (Haycraft et al., 2011).

338
Although further studies are needed before conclusions can be drawn, the same physiological 339 mechanism as described above might therefore explain the relationship between negative 340 affectivity and satiety responsiveness and slowness in eating finding (i.e., emotional arousal -341 decreased gut activity -reduced appetite).

342
Effortful control. Children with lower levels of effortful control were more food 343 responsive (from ages 4 to 6), displayed more emotional overeating (from ages 6 to 8) and control, on the other hand, predicted more enjoyment of food and a slower eating pace (from 346 ages 6 to 8 and 8 to 10); in line with this finding, a link has previously been reported between 347 behavioural inhibition (i.e., the ability to inhibit behavior) and slowness in eating 348 (Vandeweghe, Vervoort, Verbeken, Moens, & Braet, 2016). The relationship between 349 effortful control and enjoyment of food might seem surprising though, given that enjoyment 350 of food is also considered to be a food-approach behavior. Although they are positively 351 associated, greater 'food responsiveness', in contrast to 'enjoyment of food', is indicative of 352 less self-regulated eating. Children high on temperamental effortful control may indeed enjoy 353 food, but still be better at self-regulating their food intake because they have the ability to 354 withhold impulses (i.e., inhibition) and re-direct behavior (Rothbart & Bates, 2006), and thus 355 display lower food responsiveness.

356
In contrast to what we expected, satiety responsiveness was unaffected by children's 357 effortful control. Satiety responsiveness, or 'fullness' sensitivity  358 has a strong genetic basis (Carnell, Haworth, Plomin, & Wardle, 2008;Llewellyn, 359 Trzaskowski, van Jaarsveld, Plomin, & Wardle, 2014; Llewellyn, van Jaarsveld, Johnson, Carnell, & Wardle, 2010) and reflects the homeostatic appetite system; this controls hunger 361 and satiety according to energy needs, primarily via the melanocortin pathway, which is 362 regulated by hormones that signal shorter-and longer-term energy balance (e.g., gut 363 hormones released periodically in response to energy intake, and adiponectins produced by 364 adipose tissue) (Anderson et al., 2016). The biological basis of satiety sensitivity may make 365 it less amenable to modification by psychological processes such as effortful control. Food 366 approach behaviors such as food responsiveness, on the other hand, are regulated by the 367 hedonic appetite system (i.e., 'eating for pleasure'), which involve the neuropsychological 368 processes of wanting and liking, regulated by the dopamine pathways, and the opioid and 369 endocannabinoid systems (Zheng & Berthoud, 2008). Food responsiveness may thus be more 370 likely to be affected by psychological factors such as effortful control. In summary, our study 371 extends the existing cross-sectional research that has shown effortful control (or 372 corresponding concepts/phenomenon such as executive function and self-regulation) to 373 correlate with 'food approach' behavior (Godefroy et al., 2016;Leung et al., 2014). One may 374 argue that common underlying neurobiological functions (i.e., the genetic basis of executive 375 functions) might influence both, but our findings indicate that effortful control also predicts 376 'food approach' behaviours independently of such time invariant factors.

Surgency.
Our results further revealed that higher surgency may promote more 'food 378 approach' ('Food responsiveness', 'Enjoyment of food'; from age 6 to 8 years; 'Desire to 379 drink'; from age 6 to 10 years) and less 'food avoidant' behavior (`Food fussiness', 'Satiety 380 responsiveness'; from ages 6 to 8 and 6 to 10 years; 'Slowness in eating': from ages 4 to 6 381 ang 8 to 10 years), as we hypothesized. No former longitudinal studies of surgency and 'food 382 approach' behavior exist, but our finding corresponds to earlier research reporting cross-383 sectional associations between surgency and 'food approach' behaviors (e.g., food 384 responsiveness) (Leung et al., 2016). Even though replications are needed, it might be that the outgoing, explorative style of surgent children, akin to 'openness to experience' in adult 386 personality, do cause them to be more open towards novel food experiences as well (i.e., less 387 food fussiness) and to enjoy food more, which might also cause them to be more prone to eat 388 in response to external food cues, and eat at a faster pace. Highly surgent children whose 389 focus is on the outside world might also be less sensitive to inner signals, such as those of 390 fullness, and therefore display lower levels of satiety responsiveness, compared to less 391 surgent children. 392 We hypothesized that the prospective relationships between temperament and eating 393 behaviors would strengthen with age, which was confirmed with regards to the association 394 between negative affectivity and food responsiveness and emotional overeating, respectively.

395
Other age-related increases in associations were also observed; surgency was a stronger 396 predictor of food responsiveness from age 6 to 8 years as compared to the years from age 4 to 397 6, and the magnitude of the association between effortful control and slowness in eating also 398 increased with age. However, one exception was revealed -the associated between surgency 399 and slowness in eating weakened by age. Our findings may indicate that as children take 400 more responsibility for their own eating as they mature (i.e., less parental control), their inner 401 dispositions such as temperament are able to play a greater role in shaping their own eating 402 behavior.

403
Unmeasured time-invariant factors, such as changes in parenting over time may also 404 affect both temperament and eating behavior, and thus produce spurious associations between 405 them. For example, parental sensitivity is associated with fussiness in children (Steinsbekk,406 Bonneville-Roussy, Fildes, , a parent characteristic that may 407 vary over time (Dallaire & Weinraub, 2005) and which is also linked to the development of 408 temperament (Parade, Armstrong, Dickstein, & Seifer, 2018). Furthermore, parental stress 409 can vary over time and may undermine the development of effortful control (Gartstein, Bridgett, Young, Panksepp, & Power, 2013), and stress is also associated with higher levels 411 of food responsiveness in children (Boswell, Byrne, & Davies, 2018) and might thus have 412 contributed to the associations between temperament and eating behavior found here. We 413 have previously shown that negative affectivity predicts emotional feeding and emotional 414 eating in children, the latter two being reciprocally related (Steinsbekk, Barker, et al., 2017).

415
In sum, a range of factors may interact and change over time, to influence eating behavior.

417
The present study has many strengths; a large community sample, longitudinal data, and the 418 use of an analytical technique that allowed us to discount the influence of all unmeasured 419 time-invariant confounders. Nevertheless, there were some limitations. Parents reported on 420 both their child's temperament and eating behavior, which could have inflated associations 421 between temperament and eating behavior due to common rater bias. Notably though, rater 422 bias contains both transient/time-varying (e.g., mood-of-the-day effects) and more stable 423 aspects (e.g., social desirability or acquiescence) (Moum, 1988) and because the latter part is 424 partly time-invariant, this time-invariant aspect was accounted for in our hybrid fixed-effects 425 approach. Furthermore, temperament was measured at ages 4 and 6, whereas eating behavior 426 was measured at ages 6, 8, and 10. We could not therefore account for baseline levels of 427 eating behavior when examining the associations between temperament and eating from age 428 4 to 6 and eating behavior at age 10 was predicted by temperament at age 6. However, both 429 temperament and eating behavior are considered biologically based/dispositional 430 characteristics displaying modest to moderate stability (Ashcroft,Semmler,Carnell,van 431 Jaarsveld, Roberts & DelVecchio, 2000). Even so, prospective associations 432 tend to decrease with increasing time span between predictor and outcome. Thus, the age 6 433 temperament to age 10 eating behavior paths may have been attenuated compared to the 434 association obtained if we measured temperament at age 8. Furthermore, child temperament has its own origins, and merits separate studies that could complement the present one to 436 provide a fuller picture of the temperament-eating association. Finally, although the influence 437 of time-invariant factors (e.g., genetics) was ruled out, uncontrolled time-varying factors such 438 as unstable aspects of parenting (e.g., changes due to the child's development, family 439 situation) or negative life-events may affect both temperament and eating, and thus influence 440 the results.

442
Following a community sample of 4-year-olds with biennial assessments until age 10, we 443 found that negative affectivity was prospectively associated with a range of eating behaviors, 444 whereas low effortful control may be involved in the development of 'food approach' 445 behavior specifically. Surgency negatively predicted 'food avoidant' behavior and was 446 inconsistently related to 'food approach' behavior. We add to existing research by using a The hybrid fixed/random effects model: Cross-lagged part (normal font) and time-invariant factor part (in bold) Note: Presentation of the analytical model tested. T1: Age 4; T2: Age 6; T3: Age 8; T4: Age 10. Note that the model is abbreviated for illustrative purposes. Due to the high number of parameters to be estimated relative to the number of children, a model for each of the eating behaviors in question was created (i.e., 6 models). Each model consists of 1 time-invariant factor, 1 eating behavior (measured at T2, T3, T4) and 3 temperamental traits (Measured at T1,T2) (Results: see Table 2). The "latent factor" is a time-invariant factor that loads on the respective factor, e.g., on 'Food responsiveness'. In random effects models, the correlations between temperament (i.e., predictors) and the time-invariant factor are fixed to zero, whereas in fixed models these correlations are freely estimated. In a hybrid model, the temperamental dimensions shown to be uncorrelated with the time-invariant factor are fixed, whereas those who are associated with the latent factor are freely estimated. Time-invariant factor part (a) and fixed (b)/random; (c) cross-lagged paths. In all models, temperamental factors (i.e., negative affectivity, surgency, and effortful control) are allowed to correlate with each other and with eating behavior (not shown).   Note. All models are nested and compared with the next model (i.e., random models are compared with fixed models, fixed models are compared with hybrid models); ∆χ 2 is corrected according to Satorra-Bentler's procedure; preferred model in bold. a The baseline model is an unstructured model (null model/null hypothesis) assuming zero covariation between the observed variables; b Root mean square error of approximation; c Standardized root mean square residual; d Comparative fit index; e Tucker Lewis index.  Note. For 'Emotional undereating' and 'Slowness in eating', results from the preferred hybrid model (M4) (Table S1) are displayed, whereas for 'Satiety responsiveness' and 'Food fussiness', the results of the preferred random model (M2) is presented. B=unstandardized beta coefficients; β=standardized beta coefficients.