Studies conducted at the Department of Nutrition, Exercise and Sports, Faculty of Science, University of Copenhagen, Denmark, from 1992 to 2006 were available for inclusion in this study. Studies were included if they contained: 1) data on at least 24-h EE measurements in whole-body respiration chambers; 2) baseline measurements and a minimum of three measurements of appetite sensations after a dinner meal using visual analog scales (VAS); and 3) detailed descriptions of energy intake (EI). To obtain individual energy balance during the chamber stay, EI provided was based on equations estimating EE. In studies where active components (e.g. medication, enriched foods, etc.) were tested, only the control measurements done without these active components were included in this IPD meta-analysis. If more than one measurement was carried out on the same participant, and if both measurements matched the criteria above, only the last measurement was included. Finally, all included participants had to be healthy, non-smoking, and non-elite athletes (less than 10 h of exercise a week).

EE was measured by indirect whole-body calorimetry in a 14.7 m³ open-circuit respiration chamber at the Faculty of Science (University of Copenhagen, Denmark). The design of the chambers has been described in detail elsewhere (10). In the included studies, oxygen and carbon dioxide exchange including urinary nitrogen measurements were used to calculate EE. Spontaneous physical activity (SPA) in the respiration chambers was assessed using two microwave radar devices. The within-subject variation for 24-h EE measured in the chambers is 2.3% and the within-subject variation for 1-h measurements of basal EE is 5% (10). Resting metabolic rate (RMR) was measured in a resting period of 30 min just before the dinner meal was served.

VAS were used for measuring the subjective appetite sensations of satiety, hunger, fullness, and prospective food intake. The scale consists of a horizontal line (100 mm in length) with the most positive and most negative sensations at opposite ends of the line. Participants mark the line at a point corresponding to their perceived appetite at a given time. For satiety, the question was ‘How satisfied do you feel?’, and the text anchors were ‘I am completely empty’ and ‘I cannot eat another bite’. For fullness, hunger and prospective food intake the questions were ‘How full do you feel?’ (text anchors: ‘Not at all full’ and ‘Totally full’), ‘How hungry do you feel?’ (text anchors: ‘I am not hungry at all’ and ‘I have never been hungrier’) and ‘How much do you think you could eat?’ (text anchors: ‘Nothing at all’ and ‘A lot’), respectively.

The reproducibility and validity of VAS have previously been examined by Flint et al. (11), who concluded that VAS provide a reliable tool for a quantitative measurement of subjective sensations such as appetite in single meal-test trials. Finally, Composite Appetite Score (CAS) was included in the study. This measure reflects the four VAS questions and was included in the study as a summary measure of appetite. CAS was calculated using the formula inspired by Anderson et al. (12): CAS=((100 − satiety)+(100 − fullness) + hunger + prospective food intake)/4.

Data were acquired by contacting the principal investigators, searching old records, and contacting the research department where the studies were conducted. The data of all of the included participants were checked for duplicates and were checked twice (by a second reviewer) before being included in the IPD meta-analysis. Data were extracted from previously collected data and were treated as confidential.

Data on appetite (VAS scores on satiety and CAS [including satiety, fullness, hunger, and prospective food intake]) just before ingestion of the dinner meal and until 180 min after the meal were extracted from each of the included studies. Similarly, data on EE measurements 30 min before the dinner and until 180 min postprandially were also extracted. Furthermore, data on sex, age, height, body weight, EI, fat-free mass, fat mass, body fat percentage, and tea/coffee ingested after dinner (yes/no) were also collected. VAS was calculated as the incremental area under or over the curve (iAUC and iAOC, respectively) using VAS measurements filled out immediately before the dinner as baseline. The examined appetite measures were Satiety_iAUC and CAS.

TEF was calculated in three different ways. The primary TEF measure was calculated as the incremental area under the curve for resting EE after the dinner meal with RMR used as the baseline measure (TEF_iAUC) and expressed in kJ/3 h. For the purpose of sensitivity analysis, we also scrutinized other ways to interpret TEF: (1) TEF as the increase in EE above baseline expressed as a percentage of EI in dinner meal (TEF%_EI); and (2) TEF as the increase in EE above baseline expressed as a percentage of baseline (TEF%) as secondary TEF measures. These TEF measures were included in posthoc analyses of satiety and CAS. We preferred the dinner meal as this normally included 30–40% of the total EI during the stay.

All included studies were randomized controlled trials with either parallel (13–16) or crossover (17) design. Four of the studies used either a single- (15, 17) or double- blinded design (13, 14). Two of the studies were placebo-controlled trials testing potential drug candidates (13, 14). The remaining three studies all tested different diets against each other (15–17).

The raw data included from each study were analyzed separately to acquire summary statistics in the form of Spearman's r-coefficient and the corresponding standard error (SE), reflecting the level of statistical significance in the individual tests for association in the different eligible trials. The SE for the non-parametric Spearman's r-coefficient (SE[r]) was estimated from the (two-sided) P-value converted into a standard normal score (Z), which enabled an estimate of SE[r] using Wald-test methodology (Z[r]= r/SE[r]). All results are reported with 95% confidence intervals (95% CI), that were calculated as the estimate for coefficient r±1.96*SE. For the purpose of sensitivity analysis, these Wald-test-converted SEs were re-confirmed when estimated using the approach proposed by Thompson et al. (18) using the Fisher-transformed correlation (and SE) derived from the reported P-value and the sample size.

The summary analyses were computed using homogeneity statistics to evaluate the agreement of the individual trial results with a fixed-effect meta-analytic summary (19). However, for the overall inference we used standard random-effects meta-analysis (20) as a default option; the fixed-effect analysis would apply as a sensitivity analysis in the case of inconsistency. We estimated inconsistency by calculating the I ² statistic (21), which describes the percentage of total variation across trials that is attributable to heterogeneity rather than to chance (22).

We performed a number of pre-specified sensitivity analyses based on a statistically more advanced hierarchical model. At level one in the hierarchical model, participants were compared with others from the same study (i.e. trial numbers were applied as clusters), enabling the entire dataset to be analyzed as if it originated from a single study (23). A random coefficient model was applied to assess the five different possible linear associations simultaneously (24). Random coefficient models emerge as natural mixed-model extensions of simple linear regression models in a hierarchical (nested) data setup. As we treated each study as a random sample, it was natural to incorporate this in the model by assuming the subject effects (trial intercepts and slopes) to be random (25). These models enabled inclusion of potentially confounding factors included in Table 1; factors and covariates that could influence the overall summary association across included studies (see Supplementary file). These hierarchical models derived from maximum likelihood estimates were performed using the SAS software (version 9.2).

As illustrated in Fig. 2, the IPD meta-analysis found no overall association between satiety and TEF_iAUC, with the Spearman's r-coefficients varying from −0.36 to +0.33, all being statistically non-significant (P>0.18). Pooling the data from these five individual trials produced a combined Spearman's r-coefficient of 0.06 (95% CI: −0.16 to 0.28), providing no support to the hypothesis of an association between satiety and TEF (P=0.58). This result was based on a small amount of heterogeneity, with a negligible degree of inconsistency (I ²= 15.8%). For sensitivity, we compare this with the more naïve approach – combining all participant data into one correlation analysis (i.e. ignoring the study structure), also resulting in an overall statistically non-significant Spearman's r-coefficient (r=0.079, P=0.42, N=109 observations). Inclusion of covariates did not affect the association between satiety and TEF_iAUC (see Supplementary file).

There was also no association between CAS and TEF_iAUC (P=0.45); the pooled Spearman's r-coefficient from these five trials was r=0.08 (95% CI: −0.12 to 0.28). As presented in Fig. 2, the r-values ranged from −0.20 to +0.31, all being statistically non-significant (P>0.20), and apparently with no inconsistency (I ²=0%). For sensitivity, combining all participant data into one correlation analysis also resulted in an overall statistically non-significant Spearman's r-coefficient (r=0.051, P=0.60, N=107 observations). Inclusion of covariates did not affect the association between CAS and TEF_iAUC (see Supplementary file).