School Feeding for Improving the Physical and Psychosocial Health of Disadvantaged Students

3.1 Continuous data. RCTs, CCTs, and CBAs

To perform meta-analyses of continuous data, we input data on means, standard deviations, and the number of participants for each outcome in the two groups. It is important to note that, in all cases, these means and standard deviations were unadjusted for confounders; however, they were adjusted for clustering when needed. We used mean and standard deviation for change in the meta-analyses in all but one case ( Agarwal 1989 ). In this case, we used end-of-study results for intelligence because we could not find an appropriate Rho (correlations between beginning and end-of-study in the control group).

In cases where standard deviations and/or standard deviations for change were not published, and the study was published < 30 years ago (after 1976), we wrote to the authors requesting this information. Susan Walker provided us with means and standard deviations for change, and with before and after correlations in the control group for the Powell 1998 study; data from the Powell 1983 study was unavailable. For Du 2004 , Zhang Qian provided change data for weight and height, corrected for clustering (Zhang 2006). We received no reply from Agarwal and Paige concerning their respective studies ( Agarwal 1989 , Paige 1976 ).

Change data

Data on change was either taken directly or calculated from other information presented in the papers for the following studies: Bailey ( Bailey 1962 ), Baker (Baker 1980 (boys)), Corry-Mann ( Corry-Mann 1926 ), and Devadas (Devadas 1979: 5-6). Susan Walker provided data on change and on correlations between the beginning and end-of study for the Powell 1998 study (Walker 2004). These correlations were used in calculating standard deviations for change for the Agarwal ( Agarwal 1989 ) and Powell 1983 studies, using standard deviation from baseline and end of study according to the methods described in section 8.5.2.10 of the Cochrane Handbook. In cases where data were not meta-analyzed, regression analyses, multilevel analyses, or Analyses of Variance were selected as providing the better estimate of effect, because: a) multilevel analyses accounted for clustering or because b) other ANOVAs and regressions provided results for change.

3.2. Interrupted time series

We used the average of before and after session presented in the two ITS studies ( Bro 1994 , Bro 1996 ).

We did not have any discrete outcomes in our meta-analyses.

3.4 Skewness

Except for Shemilt 2004's trail-making test ( Shemilt 2004 ), skewness was not indicated in any of the studies included in this review. For all papers except Corry-Mann 1926 , summary statistics were provided rather than the original raw data and we were thus unable to check the raw data for skewness. Furthermore, because change scores were used in most analyses, we were unable to do the rough check described in Section 8.5.2.11 of the Cochrane Handbook ( Higgins 2005 ).

3.5 Data synthesis

We conducted meta-analysis using a random effects model. Weighted mean differences were calculated unless otherwise stated. We conducted separate analyses for each outcome across: 1) Lower income countries vs. Higher-income country, and 2) Different study designs (i.e. ITS, RCT and CBA). A clinical psychologist (LJ) commented on the clinical relevance of the psychological outcomes.

3.6 Methods of analysis for studies with potential unit of analysis errors

Studies which allocated by school or class could have unit of analysis errors if they did not adjust for between-cluster correlation. Therefore, we reviewed each primary study to determine whether or not clustering was part of the study design. If it was, we then determined whether or not the analyses adjusted for clustering. If analyses were adjusted in the primary study, we used them.

3.6.1. Methods used to correct for design effect in clustered trials or CBAs that were not adjusted for clustering:

When the pooled estimate used was an SMD (because of varying metrics), the method outlined in section 8.11.2.4 of the Cochrane Handbook ( Higgins 2005 ) was used exactly as described to inflate the standard error. First, we calculated the unadjusted SMD and 95% confidence interval. The unadjusted SMD was entered as the effect estimate in the generic inverse variance method, then we inflated the standard error of the effect estimate by multiplying by the square root of the variance inflation factor, calculated as: (1+ (m-1) multiplied by (ICC), where m is the average cluster size. The standard error was calculated at the confidence interval divided by 3.92.

This approach was used because there is a real possibility that the cluster sizes differ between the treatment and control groups and therefore the VIF, which depends on cluster size, would be different. Therefore, the adjustment for variance inflation would differ between the treatment and control groups. As far as we know, the Cochrane Handbook does not provide for this eventuality. However it is necessary to take account of this difference in VIFs (see, for example, Agarwal in our review ( Agarwal 1989 )).

3.6.2 Calculating the Variance Inflation factor

First we calculated cluster size. When numbers of participants in each analysis were provided, these numbers were divided by the number of clusters to calculate cluster size. Otherwise, we used the number of participants provided in the methods sections of the primary studies and divided that by the number of clusters.

3.7 Subgroup analyses

3.7.1 Equity question. We had planned to perform subgroup analyses by socioeconomic status as well as baseline nutritional status. However, none of the included studies presented tables stratified by SES (although two did briefly describe results by SES). Two studies presented results stratified by baseline nutritional status (Powell 1998 (A), Agarwal 1989 ), this was used in a subgroup analysis as a proxy for SES. Meta-regression was used to test for an interaction between baseline nutritional status and height and weight.

3.7.2 Age

Because growth rates may vary by age, we conducted a subgroup analysis across age groups for weight and height. We also planned to do subgroup analyses by age for cognition, as it is possible that the effectiveness of school meals in changing cognitive outcomes may vary according to age. However, our data were not sufficient to do this.

3.7.3 Sex

We also conducted subgroup analyses by sex, and reviewed each study to determine whether or not they reported interactions between sex and outcomes.

3.8. Potential effect modifiers

We carefully examined several of the process elements listed above: high/low energy, compliance, substitution, and duration of the intervention. Study quality may also impact on findings; studies of lower quality often show higher effect sizes than those of higher quality. For example, biased outcome assessment is possible in situations where those who assess the outcome variables are not blinded to study group (i.e., intervention or control).

To better understand the influence of potential effect modifiers, we tabulated effects for each study sorting them by type of study, blinding versus unclear blinding, date of study, and high versus low energy ( Table 06 , Table 07 ).

3.9 Presentation of raw and absolute differences

Since absolute and relative differences convey different information, and both are needed to make an informed decision ( Hembroff 2004 ) we calculated both absolute and relative differences between intervention and control conditions. The absolute difference in change from baseline was calculated by the Review Manager analysis as the difference in change from baseline between the school meal and control groups. We calculated the relative % change as the difference in change divided by the baseline of the control group of the most representative study. Because height and weight both increase over time, we also calculated the % change relative to the rate of change as the difference in change scores (over the study periods) divided by the control group rate of change over the same time period (using a weighted average of the rate of change in all studies, using the same weights as the Review Manager analysis).

3.10 Heterogeneity

Where meta-analysis was conducted, we assessed heterogeneity using the steps below.

1)

Common sense (e.g. Are the interventions, participants or outcomes so different that they cannot be combined?)

3.11. Assessment of publication bias

There were too few studies in meta-analyses to evaluate the risk of publication bias by means of funnel plots ( Egger 1997 ). As the best protection from publication bias is a robust search strategy, we made strenuous efforts to locate unpublished and grey literature.

2.1 Study design

The 18 included studies comprised seven randomized controlled trials (RCTs), nine controlled before and after studies (CBAs), and two interrupted time series (ITS). Details of included studies are contained in the Characteristics of Included studies).

2.1.1. Allocation/clustering

Clustering was an issue in nine studies; these were all allocated by school or class. Of these nine, three ( Jacoby 1996 ; Neumann 2003 ; Shemilt 2004 ) adjusted for clustering in their analyses. We adjusted for clustering in five ( Agarwal 1989 - boys ; Bailey 1962 ; Du 2004 ; Orr 1928 - 13 yrs ; Powell 1983 ). We were unable to adjust for clustering in one study ( Chandler 1995 ) and therefore reported unadjusted Analysis of Variance results. Clustering was not an issue in the other nine studies (Baker 1980 (boys), Bro 1994 ; Bro 1996 ; Corry-Mann 1926 ; Devadas 1979: 5-6; Lieberman 1976 ; Paige 1976 ; Powell 1998 (A); Tisdall 1951 ). In five of the studies, students were allocated by individual and in four one experimental and one control school were allocated or we chose one experimental group to compare to another.

2.1.2. Inclusion in meta-analyses

Eight studies were not included in any of the meta-analyses. In five of these (Bro ( Bro 1994 ; Bro 1996 )); Lieberman (Lieberman 1976 ); Paige ( Paige 1976 ), and Tisdall ( Tisdall 1951 )), standard deviations were unavailable. We report Analysis of Variance results from Jacoby ( Jacoby 1996 ) and regression analyses from Shemilt ( Shemilt 2004 ) as these analyses corrected for clustering. The Powell 1998 study (Powell 1998 (A)) and the Kenya study ( Neumann 2003 ) were both included in meta-analyses for weight and height, but for psychosocial outcomes, we used results of multilevel regression. Similarly, the Orr study ( Orr 1928 - 13 yrs ) was included in meta-analyses for height, but lacked appropriate data for weight.

2.2 Location

Nine studies were from Lower Income Countries; five of the seven RCTs and four of the eight CBAs. Of the nine studies performed in higher-income countries; two of these were RCTs, six were CBAs, and two were Interrupted Time Series (see also Characteristics of Included Studies).

2.3 Participants

Children participating in studies included within this review ranged in age from 5 to 19 years; most studies included primary school children. Details on age, sex, and socio-economic characteristics of children in each study can be found in the ‘Characteristics of Included Studies’ table.

2.4 Description of the interventions

2.4.1 Lower income countries

Five programs provided meals: four provided breakfast and one provided lunch. The remaining four programs provided snacks/milk. Energy provided ranged from 195 to 730 kcal and protein provided ranged from 10-27 grams per day. Three studies ( Bailey 1962 ; Du 2004 ; Neumann 2003 ) provided less than <15% of RDA for energy, at 8-10% of the RDA for energy, 10% of the RDA for energy and 13% of the RDA for energy, respectively.

Five studies in lower-income countries provided no intervention to children in control groups; these children may have had meals at home, or had no meals. In four studies, children in the control group were given a very low energy snack (e.g. quarter of an orange), drink (e.g. syrup) or iron pill; the three Jamaican studies ( Chandler 1995 ; Powell 1983 ; Powell 1998 (A)), did this to control for the effects of benevolent attention. In one study ( Bailey 1962 ) boys in the control group were given iron tablets.

2.4.2. Higher-Income Countries

Five of the feeding programs involved meals: four breakfasts and one lunch, and four involved snacks/milk. Energy content of seven studies ranged from 126 to 705 kcal and the amount of protein provided ranged from 3 to 26 grams.

2.5. Outcomes

2.5.1 Physical outcomes included: weight gain in kg and height gain in cm (11 studies), change in weight and height-for-age z-score or percent height and weight-for-age (three studies), change in mid-upper arm muscle area (one study), change in mid upper arm fat area (one study), change in Bone Mineral Density, Bone Area and Bone Content (one study), and change in haemoglobin (two studies), hematocrit (one study), and in other biochemical indices (two studies).

2.5.2. Psychosocial outcomes included educational and cognitive test outcomes

Educational outcomes comprised: change in school attendance (five studies), change in achievement in arithmetic (four studies), reading (two studies), spelling (two studies). Change in cognitive test outcomes included: end-of-study general intelligence (one study), change in Raven's Progressive Matrices score (one study), change in performance on Trail Making Test (Part A) (one study), sparse reports on ‘mental tests’ or achievement (three studies), and short-term change in cognitive tasks (two studies). Although we had planned to include school enrolment, we could not find any studies on enrolment that met our criteria.

2.5.3. Behavioural outcomes

Change in playground activity was considered in one study, change in hyperactivity (one study), behavioural conduct (two studies), and in on-task behaviour (two studies).

Adverse outcomes were not reported by any studies in this review.

1.1. Criteria for RCTs

We considered allocation concealment, comparability of groups on physical and/or psychological measures at baseline, reliable primary outcome measures, protection against contamination, blinded assessment of primary outcomes, co-intervention, and loss to follow-up. For each of these criteria, a rating of adequate (criteria met), unclear, or inadequate (criteria not met) was given. We also used the Jadad Scale ( Jadad 1996 ). For the purposes of this review, we report on individual elements of the Jadad scale rather than on an overall score. For each of the Jadad criteria, a score of 1 (for met) or 0 (for not met) was given. One of the elements in the Jadad scale is double blinding, but it is important to note that double blinding is not always possible in a school meals situation.

1.2. Criteria for CBAs

In assessing methodological quality of the CBAs, we considered equivalence of baseline measurement of physical/psychological outcomes across groups, reliable primary outcome measures, protection against contamination, blinded assessment of primary outcomes, co-intervention, and loss to follow-up. For each of these criteria, a rating of adequate (criteria met), unclear, or inadequate (criteria not met) was given.

1.3 Criteria for ITS

In assessing methodological quality of the ITS designs, we considered protection against secular changes (including intervention independent of other changes, appropriate data analysis, and reason for number of pre and post points given), protection against detection bias (including intervention unlikely to affect data collection and blinded assessment), reliability of the outcome measures, and completeness of the data set. Each aspect of these criteria is described in more detail in Figure One. For each of these criteria, a rating of adequate (criteria ‘met’), ‘unclear’, or ‘inadequate’ (criteria not met) was given.

Detailed methods for rating

Two reviewers (EK and BS) independently rated each aspect of study quality (EK and VR rated Shemilt). Disagreements were resolved at a consensus meeting. A nutritionist (JK) and a registered dietician (AF) assessed the reliability of anthropometric measurements and equivalence of baseline anthropometric measures across experimental and control groups. A registered clinical psychologist (LJ) assessed the reliability of psychological measures and the equivalence of baseline psychological measures across groups. An internal medicine specialist (PT) assessed the quality of haemoglobin, hematocrit, and bone mineral density measurements. Table 03 and Table 04 provide details on baseline equivalence of physical and psychological outcomes.

2.1. Jadad criteria

2.1.1. DESCRIBED AS RANDOMIZED. All seven RCTs were described as randomized.

2.1.2 DOUBLE BLINDING. Double-blinding was not done in any of the studies.

2.1.3 DESCRIPTION OF WITHDRAWALS AND DROPOUTS. Five RCTs provided descriptions of withdrawals and dropouts, two did not provide this description.

2.1.4. DESCRIPTION OF METHOD OF RANDOMIZATION. None of the studies described their method of randomization.

2.1.5. DESCRIPTION OF METHOD OF DOUBLE-BLINDING. This was not given for any of the studies.

2.2. EPOC criteria

2.2.1. RELIABLE OUTCOME MEASUREMENTS. Five studies were rated as adequate for all, seven studies were rated as unclear for all, five studies were rated as mixed adequate and unclear, and one study was rated as inadequate.

2.2.2. EQUIVALENCE OF BASELINE MEASURES. Seven studies were rated as adequate for all, five were rated as unclear for all, two studies were rated as mixed adequate and unclear, one was rated mixed inadequate and adequate, and three were not applicable.

2.2.3. PROTECTION AGAINST CONTAMINATION. Nine ratings of adequate were given, five were unclear, two were inadequate, and two were not applicable.

2.2.4. BLINDING OF OUTCOME ASSESSMENT. Blinding of outcome assessment was generally poor; six were rated adequate, nine were unclear, two were inadequate, and one was mixed inadequate and unclear.

2.2.5. CO-INTERVENTION. Co-intervention was almost always unclear; we gave fifteen ratings of unclear, one of adequate, and two were rated not applicable.

2.2.6. LOSS TO FOLLOW-UP. Loss to follow-up was rated adequate for seven of the studies, unclear for seven, inadequate for two, and not applicable for two.

2.2.7. ALLOCATION CONCEALMENT (RCTs only). Allocation concealment was rated as unclear for all 7 of the RCTs.

A1. Weight gain

A1.1 Lower income countries: RCTs. Three RCTs ( Du 2004 ; Neumann 2003 ; Powell 1998 (A)) were included in a meta-analysis for weight. There was some heterogeneity (Chi-square = 3.49, df = 2, p = 0.18, I-squared = 41.1%). Children who were fed at school gained an average of 0.39 kg more than those who were not supplemented (95% confidence interval 0.11 to 0.67); this is significant. Sensitivity analyses with ICCs at 0.01, 0.05, and 0.10 made little difference. This absolute difference of 0.39 kg represented a 1.2% difference relative to baseline and 12% relative to the control rate of weight gain. Study duration ranged from 7 to 24 months. Taking 19 months as the average study duration, this works out to a gain of 0.25 kg a year.

A1.2. Lower Income Countries: CBAs. Three CBAs in lower income countries ( Agarwal 1989 - boys , Bailey 1962 ; Devadas1979: 5-6) were included in meta-analyses for weight. These CBAs were heterogeneous (Chi-square= 15.8, df = 5, p = 0.007, I-square = 68.3%). A statistically significant difference was found between supplemented and non-supplemented groups; on average, intervention groups gained 0.71 kg more than the control groups (95% confidence interval 0.48 to 0.95). Sensitivity analyses with ICCs of 0.01, 0.05, and 0.10 made little difference. This absolute difference in weight of 0.71 kg represented a 3.9% difference relative to baseline weight and a 44% difference relative to the control rate of weight gain. Study durations ranged from 10 to 12 months, for an average of 11.3 months. This works out to an average gain of 0.75 kilograms a year.

A1.3. Higher Income Countries: RCTs. Only one RCT in higher income countries assessed weight gain (Baker 1980 (boys)). Our analyses in REVMAN show no statistically significant differences in weight gain between children who received school milk and those who did not (WMD (kg) = 0.13 (95% confidence interval, -0.23 to 0.49).

A1.4. Higher Income Countries: CBAs. Four CBAs in higher-income countries assessed changes in weight. In the year long Corry-Mann study ( Corry-Mann 1926 ), boys who received milk gained an average of 1.42 kg more than controls (95% C.I = 1.19, 1.65) over 12 months. For the 7 month long Orr (Orr 1928) study, compared to controls, 5-year-old, 8 year olds, and 13 year olds who received milk gained from 0.18 to 0.44 more, eight-year-old children gained from 0.01 kg to 0.54 kg and 13-year-old children gained from 0.28 kg to 0.38 kg more. In Paige 1976 , the average weight increment for children who received the fortified drink was 0.5 kg greater than that of the controls (P value < 0.001); according to the graphs, the gain was much greater in younger children (below 7 and 7 to 8 years) than in the older children (8 and 9 year old) age groups (significance not given). In the Tisdall study, children in the school lunch group gained an average of 0.32 kg more than the controls (n.s.) over the 27 month period of the study ( Tisdall 1951 ).

A2 Height gain in cm

A2.1. Lower Income Countries: RCTs. Three RCTs ( Du 2004 ; Neumann 2003 ; Powell 1998 (A)) were included in a meta-analysis; these studies were heterogeneous (Chi-square = 7.9, df = 2, p value = 0.02; I-square = 74.8%). Overall, there was a small, non- statistically difference in change in height between experimental and control groups (WMD = 0.38, 95% confidence interval -0.32 to 1.08). The height gain of 0.38 represented a 0.2% difference relative to baseline height (134.5 cm) and a 5% difference relative to the control group (Additional Table 01 ). The duration of these studies ranged from 7 months (Powell 1998) to 24 months ( Du 2004 ).

A2.2. Lower Income Countries: CBAs. Three CBAs with six independent samples ( Agarwal 1989 - boys ; Bailey 1962 ; Devadas1979: 5-6) were included in the metaanalysis. There was severe heterogeneity (Chi-square = 81.8, df = 5, P < .0001, I-square = 96%). Significant gains in height were seen; children in the intervention groups gained an average of 1.43 cm more than children in the control groups (95% confidence interval 0.46 to 2.41). Sensitivity analyses made little difference. This height difference represented a difference of 1.2% relative to the baseline height (of 116.6 cm), and a 33% difference relative to control group rate of height gain.

A2.3 Higher-Income Countries: RCTs. Only one RCT (Baker 1980 (boys)) studied height. Our analyses in RevMan indicate that children who were given school milk gained a small amount more than controls over two years (WMD = 0.28, 95% confidence interval: -0.01 to 0.56).

A2.4 Higher-Income Countries: CBAs. Height gain was studied in four CBAs. In a meta-analysis of the Corry-Mann ( Corry-Mann 1926 ) and 1928 Orr ( Orr 1928 - 13 yrs ) studies, high heterogeneity was observed (Chi-Square = 29.66, df = 3, P < 0.0001, I-squared = 89.9%). Children who received milk gained an average of 0.92 cm more in height than children in the control group (95% C.I.: 0.16 to 1.69). In the 7-month long Paige study, the average height gain in the experimental group was greater than that of controls by 0.5 cm in the 6 year old group (n.s.), 1.0 cm in the 7 year old group, (p < 0.01), 0.6 cm in the 8 year olds (p < 0.01) and 0.2 cm in the 9 year olds (n.s.) ( Paige 1976 ). In the Tisdall study, the intervention group grew 0.25 cm more than the controls over two years; this was non-significant ( Tisdall 1951 ).

A3. Change in Weight-for-age z-scores

A3.1 Lower Income Countries: RCTs. Powell 1998 (A) assessed change in weight-for-age z-scores in Jamaica using the U.S. National Center for Health Statistics references. Analyses showed a statistically significant effect of treatment on weight-for-age; the change in z-score was 0.07 higher in the experimental group than in the control group (95% confidence interval 0.04 to 0.10).

A3.2. Lower-Income Countries: CBAs. In the 1983 Powell study ( Powell 1983 ) there was no difference in final percentage of expected weight-for-age between children who were given breakfast and children in two control groups; one given a low energy syrup and a no treatment control group (WMD = 0.75, 95% C.I. = -0.61, 2.40). Sensitivity analyses did not change the outcome.

A4. Change in Weight-for-height z-scores

A4.1 Lower-Income Countries: RCTs. Grillenberger reported that in the Kenya study the decline in z-score was 0.20 less for the children who received meat as a snack than for the control group; this difference was not statistically significant (95% confidence interval, - 0.24 to 0.64) ( Grillenberger 2003 ; Neumann 2003 ).

A5. Change in Height-for-age z-scores

A5.1 Lower-Income Countries: RCTs. In a meta-analysis of two RCTs ( Neumann 2003 ; Powell 1998 (A)) there was no heterogeneity (Chi-square = 0.49, df = 1, p = 0.49; I-square = 0%). A small, significant effect of school feeding on height-for-age (z-score diff = 0.04, 95 % confidence interval 0.02 to 0.06) was seen.

A6. Percentage of Height-for-age

A6.1. Lower-Income Countries: CBAs. Our analyses for Powell's study ( Powell 1983 ) show a small, significant difference in percentage height that favors control children (WMD = -.75, 95% C.I. = -0.75, 95% C.I. = -1.19, -0.32).

A7. Change in mid-arm muscle area, mid-arm fat area, and mid-upper arm circumference

A7.1 Lower-Income Countries: RCT. Authors of the Kenya study presented data on change in mid-arm muscle area, mid-arm fat area, and MUAC ( Neumann 2003 ). They reported that children who were given meat gained significantly more mid-upper arm muscle area than controls (WMD = 68.22 mm-squared (95% confidence interval 39.57 to 96.87). No differences were found for change in mid-upper arm fat area between the experimental children who were given meat and the controls (WMD = -0.31 mm-squared; 95% confidence interval = -26.12 to 25.50). There was a significant increase in MUAC in the meat group compared to controls (WMD = 0.31 cm (95% confidence interval 0.14 to 0.48)).

A8. Change in bone mass and body composition

A8.1 Lower-Income Countries: RCT. Du and his colleagues ( Du 2004 ) studied the effects of milk fortified with calcium. These results were initially not adjusted for clustering, but in a later letter to the editor reporting on an updated analysis controlling for clustering, results showed no significant differences in percentage change in Bone Mineral Content (Exp - Control = 1.7, P = 0.12) and bone area (Exp - control = - 1.2, P =0 .22). However, for Bone Mineral Density, the group who received milk with calcium experienced a 3.1% greater gain than the control group (P = 0.03).

A9. Biochemical outcomes

A9.1. Lower-income countries: RCT. In the Kenya study ( Neumann 2003 ), a number of micronutrient status indicators including hemoglobin, plasma ferritin, serum iron, serum zinc, serum copper, plasma vitamin B-12, folate and retinol, and erythrocyte riboflavin were assessed. After a year of intervention, the only significant difference was in plasma vitamin B-12. While plasma Vitamin B-12 concentration increased by 47 pmol/L (s.d. = 66) in children who received a meat supplement, it decreased by 13 (s.d. = 65) in the control group (p < 0.0001).

A9.2. Higher income countries. Tisdall ( Tisdall 1951 ) compared ‘good attenders’, ‘poor attenders’ and controls on serum ascorbic acid, serum carotene, and serum Vitamin A. Statistically significant differences were found favouring the school lunch children.

A10. Hemoglobin and hematocrit

A10.1 Lower-Income countries. RCT. One paper produced from the Kenya study (an RCT in a lower-income country) found no differences in hemoglobin increase between the experimental (meat) and the control group; the mean increase in g/L was 7.5 (s.d = 15.1) in the experimental group and 11.5 (s.d. = 19.8) in the control group ( Siekmann 2003 ; Neumann 2003 ).

A10.2. Lower-Income Countries: CBAs. Devadas (Devadas 1979: 5-6) reported a greater increase in haemoglobin in the experimental groups of 0.38 g/dl. Significance was not reported. Bailey ( Bailey 1962 ) reported a decrease of 0.83 percent in the experimental group and no change in the iron-supplemented control group. However, iron supplementation is not an appropriate control group for this outcome.

A10.3 Higher income countries: CBAs. Tisdall ( Tisdall 1951 ) reported no significant difference in increase in hemoglobin between ‘good attenders’, ‘poor attenders’, and controls (statistics not given). Paige ( Paige 1976 ) reported a larger increase in % hematocrit in the supplemented group (35.9% to 39.4%) than in the control group (increase from 35.8% to 38.2%; p < 0.001). Paige reported that the end-of-study difference in percent of children with levels below 33.9% was significant (p value < 0.01).

B1. School Attendance

B1.1 Lower-Income Countries: RCTs. Jacoby and colleagues ( Jacoby 1996 ) and Powell and colleagues (Powell 1998 (A)) assessed school attendance. In the Powell study, multilevel analysis (school, class and pupil as fixed parameters) and with initial score, sex, grade, and nutrition group as covariates revealed a significant effect of breakfast on attendance (b (unstandardized regression coefficient) = 2.32, s.e. = 0.78, p < .05). This means that children in the experimental group attended school 2.3% more days in the study period than children in the control group. Jacoby and his colleagues ( Jacoby 1996 ) performed descriptive analyses on attendance, and reported significant effects of school feeding on attendance (note that they were unable to provide us with standard deviations). In the experimental group, attendance increased by 0.58% while it decreased by 2.92% in the control group. In the first month of the intervention, the attendance of the experimental group was 5.1% higher than that of the control group. The difference in change between the experimental and control groups is 3.4 %.

B1.2 Lower-Income Countries: CBAs. Our meta-analyses adjusting for clustering in the control groups showed that change in attendance was not significantly different between the breakfast group and the pooled control groups (WMD = 4.95, 95% C.I. = -3.95, 13.46). Sensitivity analyses at 0.10 and 0.20 did not change the outcome. Agarwal ( Agarwal 1989 ) provided a very sparse report on attendance; in his discussion, he noted that there was ‘significant improvement in attendance’ (p.172). It is unclear whether this was based on statistical analyses or was merely impressionistic.

B2.1.3 Higher-Income Countries: CBAs. Three authors ( Paige 1976 ; Tisdall 1951 ; Lieberman 1976 ) all reported that children who received school meals attended school more days per year than children in control groups. Paige found that children who received school meals attended, on average, 2.5 more days per year than children in the control group (significance not reported). Lieberman reported that in the group of children who received school breakfast, attendance increased from 158 to 161 days per year while in the control group, attendance decreased from 158 to 156 days (nonsignificant). Tisdall found that students who received school lunch attended 1.4 days more than controls (non-significant); it is not clear whether this was over the two-year period of the study, or per year. Tisdall also reported that the rate of absence for medical causes per 1000 pupil days was 36.0 for school lunch children and 39.8 for control children; the rate was only 31.8 per 1000 pupil days among the group of children who regularly attended the school lunch program.

B2. Math performance

B2.1. Lower-Income Countries: RCTs Two RCTs (Powell 1998 (A)); and the Kenya study ( Whaley 2003 ; Neumann 2003 )) reported data measuring change in math performance. Powell (Powell 1998 (A)) used the Wide Range Achievement Test (WRAT), and Whaley used an adapted arithmetic test from the Wechsler Intelligence Scale for Children (WISC). In the 1998 Powell study, multilevel analyses were performed; predictors included initial score, sex, grade, nutrition group, treatment group, school, class, pupil, and a treatment by grade interaction (Powell 1998 (A)). Powell reported a significant effect of breakfast on arithmetic achievement over the seven month period of the study; children in the experimental group gained significantly more than those in the control group; (b = 0.71, s.e. = .0.31, P value < 0.05). Multi-level regression analyses in the Kenya study performed by some of the authors of the Kenyan study showed that children who were given meat gained 0.17 points per year, (s.e = 0.10, P value < 0.05) in math knowledge than children in the control group ( Whaley 2003 ; Neumann 2003 ).

B2.2. Lower income Countries: CBAs. Two CBAs (Powell 1983 ; Agarwal 1989 ) assessed math achievement. Results of our adjusted meta-analysis (with ICC = .15) show no heterogeneity (Chi-Square = 0.02, df = 1, P = 0.86, I-squared = 0). Change in math achievement was significantly greater for children who had received school meals than those who did not (SMD = 0.31, 95C.l. = 0.09 to 0.53). We also performed an analysis in which Agarwal was broken down into four nutritional subgroups; results were very similar (SMD = 0.44, 95% C.I. = 0.22 to 0.67) ( Agarwal 1989 ). Sensitivity analyses for ICCs of 0.10 and 0.20 made little difference.

B3. Spelling achievement

B3.1 Lower Income Countries: RCTs. Multilevel regression analyses performed by Powell (Powell 1998) showed no difference in change in spelling achievement as measured by an adapted version of the WRAT for the experimental group than for the controls (b = -0.5, s.e. = 0.27, ns.).

B3.2 Lower Income Countries: CBAs. In one CBA ( Powell 1983 ), our adjusted metaanalysis (ICC = 0.15) shows a small, significant difference between children in the experimental and two control groups combined (SMD = 0.24, 95% C.I. = 0.01, 0.47). Sensitivity analysis with an ICC of 0.10 showed much the same results, however, the sensitivity analysis with an ICC of 0.20 was non-significant.

B4. Reading achievement

B4.1 Lower-Income countries: RCTs. Powell and colleagues (Powell 1998 (A)) assessed change in reading performance on an adapted version of the WRAT. A multi-level regression analysis demonstrated that there was no difference in reading performance between experimental and control groups after adjusting for baseline performance, age, and sex (b = 0.12, s.e. = .29, n.s.).

B4.2 Powell and colleagues ( Powell 1983 ) assessed the effect of school breakfast on reading performance using the adapted version of the WRAT. Our adjusted meta-analyses demonstrated no difference between children in the experimental groups and those in the control groups combined (SMD = 0.09, 95% C.I. = -0.11, 0.29). Sensitivity analyses made little difference.

B5. Intelligence-type test scores

B5.1 Lower Income Countries: RCTs. Some trialists ( Whaley 2003 ) involved with the Kenya study ( Neumann 2003 ) studied effects of a school snack on fluid intelligence using Raven's Progressive Matrices. They found a significant effect of treatment: children in the group who received meat gained an average of 0.34 points per year more than the control group who were not fed (P value < 0.05).

B5.2. Lower Income Countries: CBAs. In the Agarwal study ( Agarwal 1989 - boys ), our adjusted analyses (ICC = 0.15) show moderate positive, but non-significant effects of school feeding on Full Scale, Performance, and Verbal IQ. Children who were given school lunches had an end-of-study Full-Scale IQ that was 3.9 points higher than those who were not given school lunch (WMD = 3.9, 95% C.I. = -2.88, 10.68). The end-of study Performance IQ was 5 points higher for those children who were given lunch than for the controls (WMD = 5.0, 95% C.I. = -2.60 to 12.6). Differences in Verbal IQ were smaller (WMD = 3.10, 95% C.I. = -2.99, 9.19). Sensitivity analyses made very little difference to either the point estimate or the significance.

In another analysis, we used subgroups from Agarwal ( Agarwal 1989 ). Here, our adjusted analyses (ICC = 0.15) show significant effects of school feeding on Full Scale and performance IQ. Children who were given school lunches had an end-of-study Full-Scale IQ that was 3.8 points higher than those who were not given school lunch (WMD = 3.80, 95% C.I. = 0.51, 7.10). Sensitivity analyses with ICCs at 0.10 and 0.20 were still significant. Larger differences were seen in performance IQ; the end-of study Performance IQ was nearly 6 points higher for those children who were given lunch than for the controls (WMD = 5.74, 95% C.I. = 1.75, 9.74). Sensitivity analyses with ICCs at 0.10 and 0.20 were both significant. Differences in Verbal IQ were smaller and non-significant (WMD = 3.32, 95% C.I. = -0.21, 6.92). Sensitivity analyses made very little difference to either the point estimates or the significance. Another CBA (Devadas 1979: 5-6) provided very sparse reports on mental ability; she found no differences between experimental and control groups in end-of-study mental ability.

B5.3. Higher Income Countries: CBAs. One CBA in a higher income country (Lieberman 1976) reported no difference in magnitude of improvement on Raven's Progressive Matrices and other psychological tests (figure copying, making x's, listening attention, and memory for numbers). Another CBA (Tisdall 1951) reported no significant differences in change in intelligence test scores, school marks, and reading and math tests.

B6. Short-term effects on cognition

B6.1 Lower-income countries. RCTs. Chandler and her colleagues (Chandler 1995) performed a randomized cross-over study of the short-term effects of providing breakfast on four cognitive tasks. They report significant treatment effects for verbal fluency (P value < 0.02) and a significant treatment by nutritional status interaction (P value < 0.05). Those children who were undernourished (defined as 1 or more than 1 standard deviation below National Centre for Health Statistics references) had scores (unadjusted for clustering) that were 1.5 points higher after receiving breakfast than when they received the placebo (P value < 0.01). However, breakfast made no difference to the children who were classified as adequately nourished (above - 1 s.d. for weight- for-age). No significant effects of school feeding were found on information processing, visual search or digit span for either well-nourished or undernourished children.

Jacoby and his colleagues ( Jacoby 1996 ) studied the short-term effects on performance in a battery of psycho educational tests: 1) the coding subtest from the Weschler Intelligence Scale for Children, 2) a test of reading comprehension from the Inter-American Series, 3) a test of vocabulary from the Inter-American Series, and 4) a test of functions and abilities related to the math curriculum. An Analysis of Covariance controlling for sex, height-for-age z-score, weight - residual (weight regressed on age and height), SES, home language, repetition of any grade, and age on entry to school; current enrolment in grade, school nested in treatment group, and interactions between height-for-age and treatment and weight-for-age and height was used. Significant effects were seen for vocabulary only; the main effect of breakfast overall was not significant, but there was a significant interaction between baseline weight-for-age and height and treatment (parameter estimate = 0.37, F = 4.97, P value < 0.05). This interaction meant that the heavier children benefited most from school breakfast. The authors noted that the combination of smaller than average stature and normal weight for height are common among poor Peruvians, and that ‘this phenomenon reflects a protracted deficit of critical nutrients due to poor diet and infection’ (p. 62) ( Jacoby 1996 ). Thus he hypothesized that those who were heavier than normal were likely to be of poorer health and less well nourished than the other children.

B6.2. Higher Income Countries: RCTs. One cluster RCT in a high income country ( Shemilt 2004 ) assessed concentration with the Trail Making Test three months after breakfast clubs were initiated in a number of schools where each school decided on meal content. For children in school years 2-11, the time taken to complete the Trail Making Test was significantly shorter in the intervention group at 3 months; the time taken by the intervention group was 3.70 seconds logged and the time taken by the control group was 3.71 seconds. The ratio of adjusted geometric means was 0.90, 95% confidence interval 0.81 -1.00). It is important to note that contamination had occurred; by the time of first follow-up, 2 schools randomized to the control group were running breakfast clubs. Furthermore, not all of the 17 schools randomized to the experimental conditions ran breakfast clubs.

D1. Effects by socio-economic status

D1.1. Lower-Income countries: RCT. None of the included studies displayed data on outcome by SES. However, Grillenberger and her colleagues (part of Neumann 2003 ) reported that the effect of school meals on weight was greater for children of lower SES ( Grillenberger 2003 ).

D2. Effects by baseline nutritional status

D2.1. Lower-Income Countries

D.2.1.1. Weight. A subgroup analysis for change in weight from the Powell study (Powell 1998 (A)) showed no interaction between treatment and baseline nutritional status (fixed and mixed effects: Q/Z = 0.01, P-value = 0.97).

D.2.1.2. Height. Subgroup analysis of the Powell RCT (Powell 1998 (A) in RevMan demonstrate a significant gain of 0.30 cm in the undernourished group and a nonsignificant gain of 0.17 in the well-nourished group. However, formal analyses showed a non-significant interaction between baseline nutritional status and height gain (fixed and mixed effects Q/Z = 69.7, P = 0.404). Grillenberger ( Grillenberger 2003 , Neumann 2003 ) presented a graph showing results by baseline height-for-age z-score. They reported that height gain was not significantly different between the meat and control group for either the low HAZ (< - 1.4 standard deviations below the mean) or the high HAZ (> -1.4) groups.

D2.1.3. As previously described, Jacoby ( Jacoby 1996 ) reported an interaction between weight-residual (based on regressing weight-on-height and age) and treatment; those children in the experimental group with higher initial weights showed significantly improved vocabulary scores. Chandler ( Chandler 1995 ) found that breakfast significantly improved verbal fluency only for children who were underweight (<= -1 s.d. below reference values in weight-for-age). Another way of looking at effectiveness by need is to compare lower-income countries and higher income countries. The assumption here is that if school meals are more effective for those in greatest need, they should show larger effects in lower-income countries where the need is presumably greatest. As Table 06 and Table 07 show, effect sizes for weight and height are generally higher in lower-income countries when comparing studies in lower-income countries to studies in higher income countries from the latter half of this century (Paige is the exception to this ( Paige 1976 )). However, if one examines studies done in the 1920s in higher income countries, effect sizes for weight and height are very similar (or higher) than those for later studies in lower-income countries. The study by Paige also resulted in significant effects on weight and height ( Paige 1976 ).

D3. Age

D.3.1. Lower Income countries: weight. In a meta-regression of the three RCTs ( Du 2004 ; Neumann 2003 ; Powell 1998 (A)) with a total N of 1462, no significant treatment by age interaction for weight was found (Z (Q) Fixed and mixed effects = -4.58, P = 0.67).

In a meta-regression of the 3 CBAs ( Agarwal 1989 ; Bailey 1962 ; Devadas1979: 5-6) with a total N of 1022, there was a significant age-by-treatment interaction. The greatest benefit of school meals was shown for 5-6 year olds and 9 to 10 year olds (.95 and .89 kg respectively). The effect for 6 to 8 year old children was also large at .67 kg, but there was no effect in the Bailey study which included children up to age 13 in the analyses.

D.3.2. Higher Income Countries: weight. As previously shown, in the Paige study ( Paige 1976 ), weight gain was highest in the youngest children and lower and nonsignificant for older children.

D.3.3. Lower Income Countries: height. Meta-regression analysis for height in three RCTS ( Du 2004 ; Neumann 2003 ; Powell 1998 (A)) with an overall N of 1462 showed a significant interaction between age and treatment (Z(Q) Fixed and Mixed (moment) = -2.284, P = 0.022). An examination of the subgroup data shows that children who were 9 to 10 years old grew more relative to controls (0.40 cm., 95% CI = 0.03, 0.77) than did children in a group aged 6 to 16 (-0.37 cm, 95% C.I.= -1.30, 0.56) Meta-regression of the CBAs also showed a significant interaction with age ((Z(Q) Fixed = -9.608, p = .000: Mixed = -2.137, P = 0.033). In the three CBAs ( Agarwal 1989 ; Bailey 1962 ; Devadas 1980) with an overall N of 1022, significant effects of feeding were seen only for the youngest children (aged 5-6 and 6 to 8 years). These children gained 2.26 cm and 1.25 cm, respectively, more than children in the control groups. In the 9-10 and 6 -13 year old groups children who were fed did not gain significantly more in height than children in the control groups.

The relative differences in weight and height also show a gradient in benefit across age, with a relative benefit of 98%, 41%, 45%, 0.9% for the ages 5-6, 6–8, 9–10 and mixed ages respectively.

D. 3.4. Higher income countries: height. As mentioned previously, the average height gain was larger in the middle two age groups (7 and 8 year olds) in the Paige study (Paige 1976).

For cognition, Powell (Powell 1998 (A)) reported a significant grade by treatment interaction for math (b = -0.85, s.e. = 0.44, p = 0.05) indicating that the effect of school breakfast mainly occurred for younger children.

D4. Sex

D4.1. Lower Income countries: RCTs. Powell (Powell 1998) found no significant sex by treatment interaction for height and weight. However, Grillenberger ( Neumann 2003 ) reported that school feeding (meat versus control) had a greater effect on weight and MUAC for boys than for girls. Furthermore, while feeding with meat had a significant positive effect on WHZ for boys, there was no significant effect for girls (for boys, WHZ declined less in the experimental group than in the control group).

D4.2 Lower Income: CBAs In the Agarwal study ( Agarwal 1989 ), effects of feeding on weight and height were very similar for boys and girls. For girls, weight gain of the experimental group relative to the control group was 0.70 (significant) while for boys, the weight gain relative to controls was 0.80. There was no significant relationship between feeding and height for either boys or girls.

D4.3. Higher Income countries: RCTs. In the Baker study, no differences were evident between boys and girls in weight gain of the experimental group relative to the controls ( Baker total 1980 ). For height, girls in the experimental group grew 0.45 of a cm more than controls (n.s.); results boys in the experimental group grew only 0.11 of a cm (n.s.).

None of the studies tested a treatment by sex interaction for cognition.

E1. Study quality

E1.1. Comparison of RCTs and CBAs. Our meta-analyses of height and weight gain described above show there was a large difference between RCTs and CBAs in height gain (0.35 cm (n.s) in RCTs and 1.45 cm (sig) in the CBAs). For weight, differences between RCTs and CBAs in meta-analyses were smaller (0.39 kg in RCTS and 0.71 kg in CBAs); both showed significant overall effects. For most of the cognitive functions, comparisons of effect sizes are difficult, as every study used different scales and different methods of reporting. For arithmetic, two RCTs and two combined CBAs showed significant effects.

In higher income countries, both weight and height gains were non-significant and very small in one RCT, but higher in the CBAs. However, it must be noted that the % RDA for energy provided in that RCT was very low. No comparison of cognitive functions is possible, as this RCT did not assess cognitive function.

E1.2. Comparison of blinding to unclear blinding

Findings for this comparison were very similar to findings from comparisons of study types. However, for attendance, one of two RCTs in lower-income countries that assessed attendance had unclear blinding. This study had a very slightly larger effect size than the one that did blind outcome assessment.

E2. Substitution

Only three included studies ( Agarwal 1989 ; Jacoby 1996 ; Murphy 2003 in Neumann 2003 ) assessed this issue, and all report evidence of substitution. They all found that the net increase in energy intake was far lower than the amount of energy provided by the meal. For example, Agarwal reported that the mid-day meal program provided 450-500 kcal and 25% of the RDA for energy, but that the average daily increase in the supplemented group was only 200 kcal while in the Peruvian breakfast study ( Jacoby 1996 ), the meal contained an average of 600 kcal but the net increase in intake was 288 kcal. In the Kenyan study, the net increase in the meat group was only 140 calories when 239 were given; the milk and energy groups actually showed decreases in net intake, and the control group showed an increase of 196 calories a day.

E3. Compliance

Compliance was very poor in some studies from higher income countries. For example, in the Lieberman study, only 10% of children attended breakfast 90% of the time, and less than half attended 55% or more of the time ( Lieberman 1976 ). In this study, non-significant results were found on cognitive tests. The Tisdall study ( Tisdall 1951 ) sheds important light on the issue as they compared ‘good attenders ‘to ‘poor attenders’ (nearly 2/3 of the school lunch groups) in many analyses. In these analyses, the ‘good attenders’ did better than the ‘poor attenders’.

Effects on Biochemical outcomes

It is difficult to draw any firm conclusions about the effects of school feeding on biochemical outcomes, as only a few studies reported on them. In those that did, results for haemoglobin and hematocrit were mixed, and except for Vitamin B12 and Vitamin A, other biochemical outcomes were unaffected by feeding. In the case of Vitamin A, Tisdall noted that differences were not clinically meaningful ( Tisdall 1951 ). However, Seikmann ( Neumann 2003 ) noted that the high prevalence of disease and infections among the children in their study could have accounted for their null biochemical findings. Furthermore, studies in this review focused on feeding rather than micronutrient supplementation.

Effects on School Attendance

One of the most commonly cited benefits of school feeding is in improved attendance. The mechanisms for enhancing attendance may include not only the attraction of a free meal, but also possible effects on immune function that reduce illness and the effect on concentration that may make school more enjoyable.

The results of this review are consistent with other findings showing that school meals can improve school attendance in lower income countries. However, effects were small. Taking estimates from the two RCTs together, in a school year of 172 days a year (i.e. Peru), this would mean an increase of 4 to 6 days per year. The improvement in attendance noted in three CBAs in the United States was lower (and nonsignificant). This difference may be due to the fact that children in higher-income countries had more regular attendance at baseline than those from lower-income countries. It may also be due to the fact that families in lower-income countries were more motivated by the prospects of receiving food. This fits well with the high non-compliance rate we observed in many of the North American studies. These studies suggest that the impact of free school meals on school attendance may be greatest in areas of greatest poverty.

Improved attendance could mean greater opportunities for learning and mental stimulation and consequently, improved academic performance, more opportunities for social interaction with adults and peers, and possibly, a better attitude towards school. However, it is difficult to say whether the projected 4 to 6 days of increased attendance per year in lower and lower-middle income countries is sufficient to result in such changes.

Behaviour

Playground activity. Results from one study in our review indicate that playground activity levels, particularly pro-social activity, are higher for those who received school meals; although absolute differences are small, relative differences are large.

Evidence from three of our included studies shows that school feeding may have positive effects on classroom behaviour in both high and low income countries. However, as discussed below, effects may depend on the quality of the educational environment.

D1. Substitution

As mentioned in the introduction, one important concern in school feeding studies is substitution. In poor families, to spread limited resources, the home diet may be reduced for children who are receiving food at school. In two studies that assessed substitution ( Agarwal 1989 ; Jacoby 1996 ), the net increase in energy realized by the children was less than half of that provided by school feeding (breakfast and lunch). In the Kenyan study, the net increase was 140 out of 239 calories for the meat group (our experimental group). Children who were given the ‘milk’ and ‘energy’ supplements actually showed net decreases of more than 100 calories. This may reduce the effectiveness of school feeding programs on children's growth and cognitive performance. Therefore, experts recommend that the amount and composition of food should be sufficient to overcome this problem. It may be that a mid-morning meal could help prevent substitution since it would be seen as a snack, not a meal (personal communication, Galloway 2006 ). Our data do not allow us to explore this hypothesis since none of these three studies with data on substitution used mid-morning snacks.

Interestingly, the results from the Kenyan Study ( Murphy 2003 ; Neumann 2003 ) suggest that parents of control children may compensate for the fact that their children don't get fed at school. Over the 24 month period of the study, energy increase for the children in the control group increased by nearly 200 calories. This finding may be peculiar to this particular study, but it is quite interesting.

D2. Energy intensity of meal/snack

It seems logical that the energy intensity provided by school meals would impact on outcome. Our review provides some support for this hypothesis. Three ( Bailey 1962 ; Baker 1980; Neumann 2003 ) that provided low energy levels (9, 5, and 13 percent, respectively) showed either small or null effects on weight and height, while two (Du 2003; Paige 1976 ) showed effectiveness for both, and another one (Leiberman 1976) did not report changes in weight relative to controls. However, in the two low energy studies that showed effectiveness (Du 2003; Paige 1976 ) fortified milk or drink was provided; this may have produced positive results in the absence of high energy. Furthermore, students in the Du study that did not drink the milk on > 4 days were excluded from the analyses, indicating a high compliance rate for this study ( Du 2004 ). Therefore, although the energy content was higher in some studies, children may have been ingesting more calories and other nutrients due to higher compliance.

D3. Pupil / student compliance

Another factor that is almost certain to impact on effectiveness is compliance. While compliance rates were very high in the Kenya study ( Neumann 2003 ), two studies from higher income countries ( Lieberman 1976 ; Tisdall 1951 ) reported very low compliance rates. It is therefore, not surprising that these studies showed little effect. Results from Tisdall ( Tisdall 1951 ) substantiate the idea that children who attend regularly are more likely to benefit. In analyses comparing ‘good attenders’ to ‘poor attenders’ and controls, the ‘good attenders’ did better than the poor attenders. In addition, Du 2004 eliminated non-compliant children (those not taking the milk supplement on >4 days for any reason) from analyses and had large effects for both weight and height gain. Thus, it seems as if compliance can indeed affect outcomes of school feeding. Most studies in the review did not report on compliance. We suggest that in the future, compliance be carefully monitored and reported.

The lack of compliance shown in several studies in high income countries suggests that school feeding programs in higher income countries should take measures to increase compliance. Timing of the meal may be important in high income countries; it seems logical that students would be less likely to attend school breakfast (which means leaving home early) than they would be to attend lunch or a snack session.

The degree to which those who implement the intervention comply with the protocol (e.g. continuous provision of food, complying with group assignment) is very likely to impact on results. For example, in the Shemilt study, some experimental schools implemented breakfast clubs before baseline measurements could be taken; others (4 out of 17) did not have breakfast clubs at the time of first follow-up ( Shemilt 2004 ). Also, two of the control groups provided breakfast at the time of the first follow-up. This type of contamination could have influenced results, although the extent of this influence is not clear.

D4. Short-term reduction in hunger

Evidence from this and other reviews ( Pollitt 1995 ) shows that short-term relief of hunger through breakfast can improve performance on cognitive tasks carried out on the same morning. Thus, in long-term studies, it is quite important to control for this by giving a meal to all children on the day of testing. In that way, researchers can determine whether results are due to the long-term effects of providing food rather than simply to immediate improvements as a result of morning feeding. However, only one of the longer-term studies in this review ( Powell 1983 ) did this. We would suggest that in the future, all studies assessing the impact of feeding on outcomes such as intelligence, neuropsychological functioning, and academic achievement should consider controlling for the effects of hunger by feeding children in both the intervention and control groups.

D5. Study duration

For growth, one would expect to see effects on weight with shorter study durations, and effects on height only with longer durations. In this review, studies that assessed growth were generally of long duration; the shortest was seven months. In reviewing the results, it is difficult to discern any clear pattern by duration of the study. The Du study ( Du 2004 ) was the longest at 24 months and showed large changes in height, but others with similar durations of 23 months (e.g. Grillenberger ( Neumann 2003 ) had smaller effects. It is possible that the very large changes shown in the Du study were due to factors other than study duration (e.g. compliance and addition of calcium).

For cognition, the importance of study length depends on the outcome measure. It is reasonable to expect to see improved attention and concentration following short-term interventions, while increased performance in other cognitive domains may be demonstrated only following longer feeding intervals. Two studies in this review have suggested that when performance is evaluated using a mechanical, stimulus-response format (for example, “put a ring round all the letter As on this page”), short-term feeding has little impact on performance, whereas if the test involves a greater degree of creativity and motivation and the expenditure of cognitive effort (for example, “name as many animals as you can in time limit Y”), feeding has a significant effect. For example, the Chandler study ( Chandler 1995 ) showed a small impact on verbal fluency for children who were at least one standard deviation below normal weight-for-age. However, this improvement was not seen in other tasks that are more rote such as digit span or information processing.

Within the longer-term interventions, it is difficult to discern whether or not a pattern exists according to length of the study. This is largely due to the fact that outcomes were so different, and few could be combined. For math, effects were strongest for the Agarwal study ( Agarwal 1989 ), which lasted longest 24 months. The Kenya study ( Neumann 2003 )) lasted 23 months, however, and effects were quite small.

The interval between feeding and cognitive testing may also prove to be an important variable. It is expected that feeding would have immediate, but short-lived effects on cognitive processes such as processing speed and attention. Measurement of these functions should occur within minutes or hours of the feeding in order to quantify these changes. In contrast, feeding programs of a longer duration, such as months to years would be required to effect improvements on language, learning or academic tests and the interval between feeding and testing is less important when measuring these functions.

D6. Benevolent attention

A mechanism of action, implicit in some later studies, is that provision of school breakfast makes children feel valued and increases the general attention given to them. According to Powell ( Powell 1983 ), benevolent attention may be possible confounding factor in school feeding studies. Thus it is important to ensure that children in the control group receive the same amount of attention as those in the experimental group. However, in the Powell 1983 study, benevolent attention didn't seem to make a difference. Despite this, we conclude that control groups in school-based feeding studies should receive benevolent attention in order to control for this potential confounder.

D7. Quality of the educational environment

The quality of the educational environment plays an important role in learning. For example, Levinger ( Levinger 1996 ) noted that quality of instruction, teacher quality, and quality of the learning materials can play an important role in improving children's capacity to learn. There is some evidence for this in a sub-study of the larger Chandler study. Here, improvements in behaviour were only seen in one school that was ‘adequately equipped and organized’. Chang speculated that when classroom conditions were conducive to learning, the extra energy provided by breakfast could result in improved behaviour. However, when the atmosphere for learning was poor, this extra energy could not be channelled appropriately.

E1 Age

Evidence from this review shows that younger children realize greater height benefits than older children. However, it is important to note that few studies on growth involved older children and that some of our studies did not separate out age groups. Few studies in our review assessed the interaction between age and treatment effectiveness on cognition, although one study did show that younger children benefited more from feeding.

E4. Sex

We found that there was generally little evidence for difference in effectiveness by sex.

In terms of cognition, developmental studies have shown that sex differences in cognition in children are evident as early as the age of five. For example, girls are more proficient than boys in verbal memory, ( Maccoby 1974 ; Kramer 1997 ) verbal fluency and spelling ( Gaddes 1975 ; Barr 2003 ) and one study suggested that in adolescence, females surpass males in information processing speed and working memory ( Barr 2003 ). Thus, we might expect to find sex differences in cognitive responsiveness to school feeding. However, none of the studies presented data on this topic. We also found no discussion of potential cultural and social differences (e.g. in food intake at home) according to sex.

Considering that the World Food Program strategic plan includes reduction of sex disparities as a goal of school-feeding, the relationship between sex and the effectiveness of school meals clearly warrants further investigation.