Deworming and adjuvant interventions for improving the developmental health and well-being of children in low- and middle-income countries: a systematic review and network meta-analysis

Calculating Effect Size

The effect size of weight and weight-for-age(WAZ) was analysed as standardised mean differences (SMD) of change from baseline, since this increased our sample of studies for exploring heterogeneity than if we had used weight (kg) alone or WAZ alone. We also did this for height (cm) and height for age (HAZ). This decision was based on discussion with the nutritionists and clinicians on the research team (ZB, SA, SK) and consultation with two external nutritionists. The SMD was back transformed to weight (kg) using the median standard deviation for studies which reported weight in kg, as described in the Cochrane Handbook in section 12.6.4.

Continuous outcomes of test scores and cognition tests were analysed as change scores using standardised mean differences as the scales used differed across studies. If studies reported baseline and end of study data, we calculated change scores and the standard deviation for change, using the formulae in the Cochrane Handbook. We used a correlation coefficient of 0.9 for weight, height and haemoglobin (based on before-after correlations used by Kristjanssson 2015 for school-feeding review), and 0.71 for cognition (based on correlation matrices in a study we included for sensitivity analyses (Sternberg 1997), and 0.71 was also used by Kristjansson 2015 for cognition). We consulted with a specialist in educational measures about whether to group similar tests together (EK). We decided to summarise three types of cognitive tests: 1) short-term attention tests (e.g. digit span, number recall), 2) general intelligence tests (e.g. Peabody Vocabulary Test and Raven's progressive matrices), and 3) development outcomes for young children (e.g. language and motor development). Where possible, we used unadjusted estimates. However, if these were not available, we used adjusted estimates.

Dichotomous outcomes were analysed as relative risks, using random effect methods. We used random effects models since we expected the underlying treatment effect would vary depending on the context, populations and setting.

We report analyses for each outcome separately, focusing on the timepoint closest to 12 months as the main analysis.

We did not conduct meta-analyses for the secondary outcomes.

Where the reported outcome data was not in the required format, (e.g. means, standard deviation and sample sizes), effect sizes were calculated using the appropriate formulae provided in the Cochrane Handbook (Higgins and Green 2011). If the data were adjusted, we used the adjustment included in the reported estimate. For regression studies, we planned to use the IDCG protocol and review guidelines, but we did not find any such studies.

Costs and resource use data were synthesized in tables, if provided. We did not conduct a cost-effectiveness analysis.

Unit of analysis issues

Where the unit of allocation was by groups (e.g. schools, communities, village, region), we used the standard deviation adjusted for clustering, if provided by the study. In the case of randomised studies, if the study had not adjusted for clustering, we adjusted the standard deviations using the variance inflation factor, as described in the Cochrane Handbook (Higgins and Green 2011). The variance inflation factor is calculated using the equation: (1 + (m-1) x ICC) where (m) is the cluster size and ICC is the intra cluster correlation. If cluster size is not reported, the number of participants in each analysis or total number of participants (where former is not available) was divided by the number of clusters to calculate cluster size. If ICC was not reported, we estimated ICC values for the corresponding outcome measure using published ICCs for similar outcome measures. The effect of ICC values was assessed using sensitivity analysis.

For quasi-randomised studies, where an adjusted estimate was determined but clustering was not taken into consideration, we would have derived procedures to accommodate for clustering in the modelling process, if possible, but no studies like this were found.

Effects of treatment externalities

Assessment of heterogeneity

Heterogeneity was assessed by visual inspection of forest plots, chi-squared test and I². I²was used to quantify heterogeneity across studies, as it describes the percentage of variability in effect estimates that is due to heterogeneity. We explored heterogeneity using subgroup and sensitivity analyses.

We explored heterogeneity using pre-planned subgroup and sensitivity analyses to assess the role of possible effect modifiers such as sanitation, poverty, under-nutrition, prevalence of different types of worms, intensity of infection, co-infection, concomitant interventions (e.g. micronutrients, hygiene) and risk of bias. We chose to assess the role of these factors based on prior theory or evidence, as shown in the logic model (Jamison, Breman et al. 2006; Bundy, Kremer et al. 2009; Kremer 2004). Subgroup analyses reported in the included studies were extracted and were compared to these subgroup and sensitivity analyses.

Subgroup Analyses

These subgroup analyses were conducted using Review Manager 5.3 using a test for interaction. To address possible concerns about not making use of all available data, we also conducted subgroup analyses for any mass deworming treatment vs. control. In studies with more than two arms, we selected the intervention arm that was most similar to mass deworming twice per year and that had common cointerventions in both arms.

We also compared our subgroup analyses with subgroup analyses across the same factors reported within the primary studies.

Sensitivity analysis

Sensitivity analyses were planned to assess the impact of outlier individual studies (e.g. very large studies, very large effects, very precise confidence intervals) on the overall effect size, and studies that may not fully fit inclusion criteria.

Dealing with missing data

We contacted authors of the studies if information reported was insufficient to calculate effect size and standard deviation. When authors did not reply, we listed these studies and/or outcomes as pending. If other measures of variation were available, such as exact p-values or standard error of the difference between means, we used the formulae in the Cochrane Handbook to calculate SD for each group. We did not impute missing values (e.g. missing variance or outcome data) for the primary studies, except for two studies where we received full datasets from the authors and we used information in those datasets to impute missing values (see below for details).

Study setting

Of the included 46 RCTs and five CBAs of mass deworming, 12 were conducted in India, five in Kenya, four in South Africa, three in Bangladesh, three in Tanzania, three in Vietnam, two in Haiti, two in Indonesia, two in Zaire, and one each in Benin, Botswana, Brazil, Cameroon, China, Guatemala, Malaysia, Papua New Guinea, Senegal, Sierra Leone, Sri Lanka, Uganda and one multi-country (China, Philippines, Kenya).

Of the 13 studies that described the socioeconomic or occupational characteristics of the households, eleven studies described all participants as belonging to low-income households. Forty-one studies did not provide detail on socioeconomic status.

Sanitation (e.g. water supply) and/or hygiene practices were described as poor, with high likelihood of reinfection in 17 studies. The setting of the studies was described as school for 22 studies, clinic for 13 studies, and 11 studies were described as home or community settings.

Participants

The median age was 6.5 years (range 0.8 to 12.9 years). Of 17 studies that reported the proportion underweight, the studies reported from 0.8 per cent to 73 per cent underweight, with a median of 42 per cent underweight. For studies which reported the sex distribution, the overall proportion was 51 per cent female.

Sick children were excluded from eight studies if they had severe anaemia or haemoglobin levels below cut-offs of 80 or 70 g/L (Beach 1999, Stoltzfus 1997, This Le Huong 2007, Nga 2009), infection (Kruger 1996, Monse 2013),or concurrent illness (Lai 1995, Monse 2013). Four studies (Garg 2002, Nga 2009, Olds 1999, Watkins 1996)excluded children who had received deworming medicine in the last six months to one year.

Interventions

The most common intervention was albendazole 400 mg twice per year compared to placebo or untreated control (21 studies). Co-administration of praziquantel for schistosomiasis with deworming for soil transmitted helminths was conducted by four studies in locations where schistosomiasis was endemic (Jinabhai 2001A, Jinabhai 2001B, Olds 1999, Miguel 2004). One RCT assessed mass deworming of children for schistosomiais only (Olds 1999). One long term study evaluated the effects of a schistosomiasis control programme in Nigeria (Makamu 2016).

Co-interventions were provided in 21 studies, which included iron (Bhoite 2012, Bobonis 2006, Dossa 2001, Ebenezer 2013, Gopaldas 1983, Huong 2007, Kruger 1996, Rohner 2010, Taylor 2001), vitamin A (Awasthi 2001, Awasthi 2008, Awasthi 2013, Hall 2006, Reddy 1986), multiple micronutrient fortified biscuits or soup (Jinabhai 2001B, Nga 2009, Solon 2003), some type of unfortified food or drink (Huong 2007, Kruger 1996, Nga 2009, Pust 1985, Solon 2003), anti-giardial (Goto 2009, Gupta 1982), and intermittent presumptive treatment for malaria (Rohner 2010). Sixteen studies reported compliance with deworming medication, ranging from 65 per cent to 98 per cent. Deworming was administered by field workers (eight studies), school staff (e.g. teachers) in nine studies, study staff in 11 studies, health workers in seven studies and at home in one study. The method of administration was not reported in the other 29 studies (Additional Table 7).

Controls

The most common comparator was placebo, used in 33 studies. For the other RCTs and CBAs, the comparator group received “no treatment” (Alderman 2006, Awasthi 2013, Bhoite 2012, Donnen 1998, Linnemayr 2011, Miguel 2004, Monse 2013, Ostwald 1984, Rozelle 2015, Stoltzfus 1997) or an intervention given to both treatment and control arms which included vitamin A 100,000 or 200,000 iu (Awasthi 2001, Awasthi 2008, Bobonis 2006) or unfortified noodles (Huong 2007, Kruger 1996), biscuits without fortification (Jinabhai 2001B, Nga 2009), unfortified beverage (Solon 2003). Since all of these comparators were either equal in both groups or considered a “do nothing” control, we refer hereon to the comparators as “placebo”.

One RCT compared three different active deworming treatments, with no control group (Kaba 1978).

For the three long-termstudies that followed participants of randomised trials, the control group was a group that received deworming at the end of the study period. For one study, which followed participants of the Kenya Primary School Deworming Project (Miguel 2004), the control group received on average 2.41 years less deworming (Baird 2016). Ozier 2016 identified children who were not in school at the time of the Kenya PSDP study (children aged 8-14 years of age in 2009 and 2010; 11 and 12 years after the PSDP study), and compared children who were exposed to treated siblings before the age of one year to those exposed to treated siblings after one year of age (excluding children approximately age one). Ozier 2016 conducted two analyses: one for each of seven birth cohorts as randomized and one using age of exposure to a treated sibling to creat treatment groups. Croke 2014 identified children exposed tofrom 22 of 48 parishes in Alderman 2006, and compared children from 10 intervention parishes to children from 12 control parishes.

Worms: type and prevalence

Of the studies that reported STH worm prevalence, 36 were classified as high prevalence (>50 per centworm prevalence), 7 were moderate prevalence (20-50 per cent), and six were low prevalence (<20 per cent).

Eight of the included studies reported schistosomiasis infection (ranging from 1 per centto 80 per cent). Ascaris was reported in 47 studies, ranging from 0 per centto 93 per cent, hookworm was reported in 38 studies, ranging from 0.3 per cent to 91 per cent, and trichuris trichuria was reported in 35 studies ranging from 1 per cent to 97 per cent (Table 14).

Outcomes

For all outcomes, data were sought for the full sample at the time-point of interest (12 months).

Study duration

The median study duration for RCTs and CBAs was 12 months (range four to 60 months). Twenty-one studies had durations <12 months, 16 studies had a duration of 12 months, and 16 were longer than 12 months.

The long-termstudies had a follow-up of 10 years (Baird 2016, Croke 2014),eight years (Ozier 2016) and 10 years (Makamu 2016).

Unit of analysis issues

We included 16 cluster RCTs (Additional Table 11) and one cluster-allocated controlled before-after study (Pust 1985). We judged that 12 of these cluster RCTs had analysed data appropriately considering the cluster allocation, using methods described in Additional Table 11. We used the variance inflation factor method described in the Cochrane Handbook to adjust growth and haemoglobin outcomes in three cluster RCTs (Hall 2006, Kruger 1996, Bhoite 2012) and one controlled before-after study (Pust 1985). For Hall 2006, we calculated the ICC for weight and height from their raw dataset, which waskindly provided to us by Dr. Don Bundy (ICC for weight= 0.021, ICC for height= 0.015). For Kruger 1996, Bhoite 2012, and Pust 1985, we used the more conservative ICC of 0.17 for weight and 0.11 for height as reported by Awasthi 2008, and tested the influence of a less conservative ICC of 0.01 in sensitivity analyses. Awasthi 2013 used an ICC of 0.25 for calculating sample size for effects of deworming on mortality.

We calculated ICC for school attendance from teacher records using the full dataset kindly provided for the Ebenezer 2013 study. The ICC was 0.42 for the treatment arm and 0.30 for the control arm (using last observation carried forward to account for children with missing data in one or more of the five month of data provided. The ICC was 0.44 and 0.28 for the dataset with missing data excluded.

For dichotomous outcomes (mortality, adverse effects, proportion stunted) reported as number of events in cluster trials, we adjusted for the unit of analysis by dividing both the numerator and denominator by the design effect, using an ICC of 0.02 for stunting (based on ICC for weight from Hall 2006) and 0.01 for mortality.

For one study of school attendance, we received the full dataset of monthly attendance for six months prior to the intervention, and analysed the average attendance during the study period (November 2009 to March 2010), with consideration for cluster allocation (Ebenezer 2013), using the formulae from Ukoumunne 2009.

Data received from authors

We received data from 10 studies (Additional Table 4). We also received three replies that data had been archived or was not possible to find.

For Olds 1999, who kindly provided individual participant data for two sites of the five site study (371 children of 1,518 randomised in the original study), we imputed values for 76 children with missing data at the end of study for weight, height and haemoglobin using a single imputation technique, based on available information in the dataset on baseline characteristics of weight and height. We did a sensitivity analysis using the non-imputed values, and found no difference in the meta-analysis results (results not shown).

For Miguel 2004, we received from Dr. Joan Hamory Hicks the corrected dataset and analysis files from the replication study, led by Dr. Aiken. We first replicated the results tables using these files exactly. Then, using the identification numbers of schools which received albendazole and praziquantel combined (provided by Dr. Hicks), we ran the analysis for tables II, V and VIII from the original Miguel 2004 study and the Aiken 2015 pure replication for the schools which received albendazole alone and also for the schools which received both albendazole and praziquantel. This allowed us to analyse this study as a four arm trial for weight and height (which were only collected for groups 1 and 2 in the first year of the study): of 1) albendazole twice per year + hygiene promotion (19 schools); 2) control receiving nothing in areas with low schistosomiasis prevalence (15 schools), 3) albendazole + praziquantel +hygiene promotion (6 schools), 4) control (10 schools), which received nothing in areas of 30 per cent schistosomiasis prevalence. For school participation, data was available on school participation for each of three groups of schools after one year of treatment, and after two years of treatment. We replicated the original author analyses with restriction to schools which received only albendazole and then for schools which received albendazole and praziquantel. For the control group, (group 3 which eventually went on to receive deworming treatment), we used the treatment assignment which they received in the third year (after the end of the Miguel 2004 study).

STH prevalence

We assessed the impact of deworming interventions on worm burden at the end of each study. For the 23 studies that report impact on worms for mass deworming of children programmes, there is considerable variability in effectiveness at reducing worm prevalence at the end of study in these studies, ranging from a relative risk reduction of 9 per cent to 83 per cent (Additional Figure 4). Since some of the variation in relative risk reduction may have been due to differences in baseline prevalence, we also report the absolute risk reduction below, which ranged from 2 to 57 per cent (Table 1).

We explore the impact on worms and the worm prevalence as possible effect modifiers in subgroup, sensitivity and weighted least squares regression analyses, described below.

Haemoglobin

In summary, data from the included studies shows that mass deworming of children leads to little to no increase in haemoglobin, except for interventions where praziquantel was included in the mass deworming (moderate to high certainty evidence). These analyses do not represent the full spectrum of studies which have measured haemoglobin following mass deworming because we did not include studies if they did not report one of our primary outcomes.

Twenty-one studies reported haemoglobin in sufficient detail for meta-analysis. We did not meta-analyse haemoglobin since it was one of our secondary outcomes. Results are presented in Additional Figure 3. For all comparisons of deworming vs. placebo, the range of effect sizes in individual studies is -0.61 to 0.30 g/dL.

For six comparisons which included iron or praziquantel, statistically significant effect sizes of 0.27 to 0.93 g/dL were found in four studies (Olds 1999, Taylor 2001, Bhoite 2012, Nga 2009,). These four studies also measured anthropometric outcomes, and found no effect of deworming on anthropometric outcomes (Olds 1999, Taylor 2001, Bhoite 2012 and Nga 2009). Also, Nga 2009 assessed cognition using a general intelligence test, and found little to no effect of mass deworming on general intelligence (SMD 0.19, 95% CI: -0.07 to 0.44).

Weight or weight for age (WAZ) gain

In summary: based on our analyses, deworming probably leads to little or no difference in weight or weight gain compared to placebo (moderate certainty evidence). Based on our network meta-analysis, there is little to no difference between different types of deworming (or their frequency), their combinations with other deworming drugs or with micronutrients or food (moderate certainty evidence).

Pairwise meta-analysis: Thirty RCTs reported weight or weight of age with sufficient detail for meta-analysis. We conducted pairwise meta-analysis for each of the 26 nodes of our treatment network (Figure 8). All studies used standard doses of deworming drugs, with the exception of five studies which used Albendazole 600 mg instead of Albendazole 400 mg. Therefore, for our nodes, we combined all doses of each drug together. However, because frequency of treatment was considered a determinant of reinfection, we defined separate nodes for low frequency (1/year), standard frequency (twice per year) and high frequency (3-6 times per year). For albendazole 400 mg, twice per year, with 13 RCTs and 18, 957 participants, we found statistically significant heterogeneity with an I² of 93 per cent, which is considered very high. We considered it too high to allow pooling to estimate a single estimate of effect (Figure 6).

We conducted all of our pre-planned subgroup and sensitivity analyses for this comparison to assess whether any of them explained the high heterogeneity. None of these analyses explained this level of heterogeneity (results not shown). Therefore, we conducted influence analysis by removing single studies. The heterogeneity did not improve upon removal of any single study, with the exception of removing both Koroma 1996 and Stephenson 1989. Removal of each study yielded an I² of 89 per cent. Removal of both studies yielded an I² of 61 per cent.

To explore reasons for these outlier studies, we assessed all studies for baseline imbalance, using standardized differences in baseline scores on prognostically important variables of worm prevalence, intensity, nutritional status (weight and height) and age (Austin 2009). The two outlier studies had baseline imbalance of > 0.5 standardized difference on one or more of these variables and both were rated unclear for allocation concealment. Koroma 1996 had a clinically important baseline imbalance in weight for age of one standard deviation lower in the intervention group (e.g. -2.18 WAZ in the treatment group vs. -1.07 WAZ in the control group at baseline in the urban population). In Stephenson 1989, the albendazole group had higher initial prevalence of hookworm (95% vs. 79% for the control group) and higher intensity of hookworm infection (geometric mean egg counts of 1,183 vs. 384 epg for the control group). The pooled random effects effect size for albendazole 2/year vs. control was 0.05 SMD (95%CI: -0.02,0.11), I²=61% without these two studies, compared to 0.25 SMD (95%Ci: 0.11 to 0.39), I²=93% with these studies. If converting to kg using the median standard deviation for weight in kg from all included RCTs, this equates to a difference between 0.07 kg without these studies compared to 0.35 kg with these studies. We decided to run the pairwise and network meta-analyses without these two studies, and conducted sensitivity analyses to explore the impact of this decision on the results.

Network meta-analysis: The network meta-analysis of weight and weight for age included 29 trialswith a total of 61,857 participants (Alderman 2006; Awasthi 2000; Awasthi 2001; Awasthi 2008; Bhoite 2012; Donnen 1998; Dossa 2001; Garg 2002; Gateff 1972; Goto 2009; Greenberg 1991; Gupta 1982; Hadju 1997;Hall 2006; Kruger 1996; Ndibazza 2012; Olds 1999; Ostwald 1984; Jinabhai 2001A; Miguel 2004; Nga 2008; Rousham 1994; Rozelle 2015; Stephenson 1993; Stoltzfus 1997; Sur 2005; Watkins 1996; Willett 1979; Wiria 2013). The study by Miguel 2004 was analysed as two separate studies because we obtained data from the authors on which schools received praziquantel due to >30 per cent schistosomiasis prevalence. Because we had data on intervention schools and control schools with >30 per cent schistosomiasis prevalence, we analysed this study as two separate comparisons with two separate control groups: 1) Albendazole vs. control for schools where no praziquantel was given, and 2) Albendazole+praziquantel vs. control for schools where praziquantel was given.

We assessed 26 interventions (defined by drug type, frequency of administration, and types of co-interventions) in 20 two-arm studies, five studies with three arms and five studies with four arms. The nodes with the most studies were placebo (28 studies), Albendazole two per year (11 studies), Mebendazole three-four per year (four studies) and Albendazole three to four per year (four studies) (see network geometryFigure 8). The network analysis was consistent, as assessed by the consistency plot, and model diagnostics (Additional Table 12), with a total residual deviance 55.67, DIC -31.69.

Compared to placebo: For interventions compared to placebo, we found no statistically significant treatment effects in the pairwise analyses, with effect sizes ranging from -0.14 to 0.29 SMD, with three exceptions. The pairwise comparison of Albendazole twice per year+vitamin A vs. vitamin A was statistically significant with a random effects SMD of 0.06 (95%CI: 0.01 to 0.11) and mebendazole twice/year vs. placebo was statistically significant with a random effects SMD of 0.10 (95%CI: 0.01, 0.18). These one-year effect sizes equate to 330 grams and 110 grams, respectively. Piperazine >2/year + metronizadole (anti-giardial) vs. placebo was statistically significant in one four arm study with an SMD 0.46 (95% CI:0.02, 0.91), equating to 0.35 kg (95%CI: 0.02 to 0.68 kg); this difference was not statistically significant in the network analysis. The comparison of the network to pairwise results were consistent with each other, and found no statistically significant effects for any treatment or treatment combination on weight or weight for age (Table 2). The full comparison of all pairwise and network meta-analyses and interventions is in Appendix 2.

Two studies reported only end of study difference between arms. Bobonis 2006 reported a non-statistically significant difference at end of study (without providing results for each arm of the study) between albendazole+iron+vitamin A vs. control in preschool children of 0.50 kg (95%CI: -0.09, 1.09). Linnemayr 2011 found a difference between a comprehensive nutrition package which included deworming, growth promotion, vitamin A, iron, bednets for a fee, cooking workshops and breast feeding support vs. control of 0.26 kg (95% CI: 0.06, 0.46), which was robust to a number of sensitivity analyses (results not provided for each arm of the study).

Compared to active interventions, in the network analysis, we found no statistically significant or clinically important (using SMD of 0.3 to assess clinical importance) differences between any interventions. We report selected head to head treatments below(Table 3).

Power analysis: We assessed the power to detect a difference of 0.2 kg (postulated by Croke et al 2016 to be clinically important if only a portion of the population benefits) for albendazole twice per year vs. placebo, using the methods described by Hedges and Pigott 2001. Using these methods, we had 99.98 per cent power to detect a difference of 200 grams in our analysis of albendazole twice per year vs. placebo, with an alpha of 0.05 (two tailed test).

Controlled before after studies: The results from the controlled before after study which reported weight (Pust 1987) agree with these, and the network meta-analysis including Pust 1987 converged, was consistent (as assessed with consistency plots, residual deviance and deviance information criteria) and agreed with these results (results not shown).

Publication bias: The funnel plot for albendazole twice per year vs. placebo for weight or weight for age does not suggest publication bias(Figure 35).

Studies not included in meta-analysis

Five additional studies reported no statistically significant difference between deworming and control for weight or weight for age, without sufficient detail for meta-analysis (Kloetzel 1982, Taylor 2001, Beach 1999, Shah 1975, Solon 2003).

Height gain or Height for age (HAZ)

In summary: Deworming probably leads to little or no difference in height gain compared to placebo (moderate to high certainty evidence). There isprobably little or no difference between different types of deworming, frequencies of deworming or combinations with other deworming drugs, micronutrients or food (moderate to high certainty evidence).

Meta-analysis: In pairwise analyses, we found high heterogeneity (I² of 86%) with two studies described above: Koroma 1996 and Stephenson 1989 because of baseline imbalance. For congruence with the weight analysis, we excluded these studies from the base case analysis, and conducted sensitivity analyses with and without them (Appendix 4 for forest plot). With these two studies, the pairwise results for albendazole vs. placebo had an SMD of 0.18 (95%Ci: 0.03, 0.33) vs. an SMD of 0.03 (95%CI: -0.02, 0.09) without these two studies (I² of 11%). Using a typical standard deviation for height in the included studies of 3.4 cm, SMD of 0.18 is equivalent to 0.44 cm and SMD of 0.03 is equivalent to 0.07 cm.

Network meta-analysis structure: The network meta-analysis for height or height gain had 25 studies (Awasthi 2000; Awasthi 2001; Awasthi 2008; Jinabhai 2001A; Ndibazza 2012; Watkins 1996; Garg 2002; Ostwald 1984; Hall 2006; Kruger 1996, Miguel 2004; Wiria 2013; Rozelle 2015; Goto 2009; Bhoite 2012; Donnen 1998; Stoltzfus 1997; Stephenson 1993; Nga 2011; Dossa 2001; Gupta 1982; Olds 1999; Hadju 1997) (Figure 9). As with weight, Miguel 2004 was entered as two separate studies because we received data from authors to identify which schools had prevalence of schistosomiasis >30 per cent and thus received praziquantel, thus it was entered as Albendazole twice per year vs. control and Albendazole+Praziquantel vs. control. There were 24 different treatment combinations, in 16 two arm studies, five three-arm studies and five four-arm studies. The network was consistent. Total residual deviance was 43.36 (sd 7.783) and the deviance information criterion was -39.794 (Additional Table 13).

Compared to placebo: For treatments compared to placebo (Table 4), there were no statistically significant differences in pairwise analyses or the network, with the exception of the pairwise analyses of height for age in the Miguel 2004 study, where there was a statistically significant difference in height for age between children who received one year of treatment of 0.13 Z-score (95% CI: 0.08, 0.18)with albendazole alone and hygiene promotion vs.children in control schools and of 0.12 Z score units (95% CI: 0.04 to 0.20) for albendazole+praziquantel and hygiene promotion vs. control, where praziquantel was given once per year only in regions with schistosomiasis >30 per cent [data provided by the authors]. These comparisons were not statistically significant in the network analysis with a 0.12 SMD (95%CI: -0.01,0.25) and 0.11 SMD (95%CI: -0.03,0.25), respectively. Noeffect sizes were greater than and SMD of 0.15, with the exception of combinations of deworming with anti-giardials, where the effect size was 0.2 SMD for deworming and secnizadole vs. placebo (Goto 2009) and 0.41 and 0.42 for piperazine and metronizadole and metronizadole alone vs. placebo, respectively (Gupta 1982). See Appendix 13.5 for comparison of all network and pairwise results.

One cluster RCT (Bobonis 2006) reported no difference in height gain for albendazole + iron + vitamin A in preschool vs. control -0.75 cm (95%CI: -2.40, 0.90 cm).

Active comparators: For head to head comparisons of deworming compared to other active treatments, there were no statistically significant, nor clinically important differences in either the pairwise or network meta-analysis, with the exception of combinations with antigiardials (secnizadole or metronizadole), where there were effect sizes of 0.2 to 0.4 SMD, which was equivalent to a difference of 0.4 to 0.8 cm in the original studies (Table 5).

Studies not included in meta-analysis: Three studies were not included in meta-analysis due to missing data. All of these studies reported no statistically significant effect of deworming vs. control on height (Taylor 2001, Beach 1999, Solon 2003).

Controlled before after studies: We did not find any controlled before after studies with data for height in a usable format.

Publication bias: A funnel plot of albendazole twice per year reporting height or height for age did not suggest publication bias (Additional Figure 2).

Subgroup and sensitivity analyses: As described below, these main effects on height were robust to four subgroup analyses and our pre-planned sensitivity analyses.

Weight for height (WHZ)

In summary: Mass deworming results in little to no difference in weight for height compared to placebo (high certainty evidence). There is probably little to no difference between different drug types, frequencies or combinations with food or micronutrients (moderate to high certainty evidence).

Pairwise meta-analysis: Twelve studies were included in pairwise analysis for weight for height (Awasthi 2001; Dossa 2001; Hadju 1997; Nga 2011; Stephenson 1993; Kruger 1996; Ndibazza 2012; Watkins 1996; Garg 2002; Ostwald 1984; Greenberg 1981;Goto 2009).

Network meta-analysis structure: The network meta-analysis comprised 15 different treatment comparisons, assessed in seven two-arm studies, two three-arm studies, and three four-arm studies (Figure 10). One study with five arms was coded as a four arm study for programming reasons (we dropped the arm with pyrantel once per year, considered not applicable to current deworming programmes) (Hadju 1997). The network was consistent; the total residual deviance was 33.54, DIC -7.733 (Additional Table 13).

Compared to placebo: No comparisons of treatments vs. placebo were statistically significant in the pairwise or the network analysis for weight for height. The effect sizes in pairwise and network comparison vs. placebo were <0.2 SMD, with the exception of albendazole once/year vs. placebo, which was based on two studies (Hadju 1997; Stephenson 1993)with an SMD of 0.44 (95% CI -0.58 to 1.46) and I² of 95% (low certainty evidence, downgraded for less than optimal information size) (Table 6). For full comparison of network to pairwise results, see Appendix 2.

Compared to active interventions: There were no statistically significant differences between any active comparators. All effect sizes were less than 0.4 SMD (Table 7).

Controlled before-after studies: No controlled before after studies were found that reported WHZ.

Proportion stunted

In summary: Mass deworming with albendazole twice per year leads to little to no difference in the proportion of stunted children compared to placebo (high certainty evidence). There was little to no difference between different drug types, combinations or frequency (low to moderate certainty evidence).

Pairwise meta-analyses: We included seven RCTs which reported proportion of children stunted at the end of study. The pairwise meta-analyses were not statistically significant and ranged from 0.63 to 0.98 relative risk (Appendix 6 for forest plots).

Network meta-analysis structure: Because there were three studies with multiple arms (Stoltzfus 2001, Bhoite 2012 and Thi Le Huong 2007), we judged that network meta-analysis was worthwhile. However, since there were no studies with vitamin A alone, the network failed to converge if we considered Awasthi 2001 as Albendazole twice per year+vitamin A vs. vitamin A. As we considered there to be a low risk of synergistic effects of vitamin A (and vitamin A was in both arms), we coded this study as Albendazole twice/year vs. placebo (Figure 11). The network was consistent (residual deviance 18.58 vs. 19 data points, deviance information criteria 133.008).

Compared to placebo, mass deworming with albendazole twice per year leads to little to no effect on proportion of stunted children, with relative risk (RR) of 0.97(95%CI: 0.78,1.18)[high certainty evidence] (Table 10). Compared to placebo, deworming with mebendazole, albendazole+praziquantel, albendazole+iron, and mebendazole+iron probably leads to little to no effect on proportion of stunted children (low to moderate certainty evidence for mebendazole, albendazole combined with praziquantel, and mebendazole combined with iron).

For comparison of active treatments, there may be little to no difference in effect on proportion stunted between albendazole, albendazole +praziquantel, albendazole +iron and mebendazole (with or without iron)(relative risks range from 0.74 to 1.10) (low to moderate certainty evidence) (Table 8). Mebendazole + iron compared to iron alone may increase the proportion stunted with a RR of 1.49 (95%CI: 0.88 to 2.48) (low certainty evidence).

Controlled before after studies: We did not find any CBAs that reported proportion of stunted children as an outcome.

Studies not included in meta-analysis: One study could not be included in the meta-analyses (Bobonis 2006) and reported no difference in height for age between albendazole+iron+vitamin A vs. control in preschool children five months after treatment.

Other measures of anthropometry: Body mass index, mid-upper arm circumference, skinfold, proportion wasted, proportion underweight

Because these outcomes were reported in a small number of studies (less than 10 studies, as described above), and all studies reporting these outcomes also reported either weight or height or both, we did not conduct analysis for these outcomes.

One CBA reported BMI but not weight or height (Monse 2013). It found little to no effect on BMI between a class treated with mass deworming and an external concurrent untreated control school, with an effect size of 0.02 BMI units (95%CI: -0.33 to 0.37) (very low certainty evidence).

Cognitive processing and development

In summary: Based on our analyses, mass deworming results in little or no difference in short-term cognitive tasks (such as attention), little to no difference in general intelligence and little to no difference in childhood development compared to placebo (high certainty evidence).

Pairwise meta-analysis: We conducted three meta-analyses: 1) cognitive outcomes that are sensitive to attention and short-term memory (e.g. digit span, number recall); 2) general intelligence measures (e.g. Peabody Vocabulary scale and Raven's progressive matrices); and 3) childhood development measures (e.g. language development).

For short-term attention, three studies reported outcomes of tasks which measure short-term attention: 1) Ebenezer 2013 assessed a coding task; 2) Rozelle 2015 reported the WISC IV working memory index and processing speed index (in Chinese), and 3) Ndibazza 2012 measured an attention task. We used the WISC IV working memory index from Rozelle 2015, and tested whether choosing the Processing speed index would change results (it did not). These studies show little to no difference in short-term attention-focused tasks (high certainty evidence). Miguel also found no difference in short-term attention tasks (insufficient information for meta-analysis).

For general intelligence, we selected the outcomes common to more than one study (Peabody Vocabulary scale reported by Ndibazza 2012 and Watkins 1996). We also felt that Raven's progressive coloured matrices from Nga 2009 could be combined with this. These results found little to no difference in general intelligence measures for mass deworming compared to control (Figure 13).

Ndibazza et al. 2012 also measured other cognition outcomes including working memory, cognitive flexibility, inhibition, planning, fine motor function, gross motor function, and reported no statistically significant differences for any of these outcomes in children at age five years who had been treated with albendazole 400 mg twice per year from 15 months of age.

For early childhood development, we decided to analyse language development reported in Stoltzfus 2001 separately from cognitive processing measures since the scale measured development in language over one year. This was done by assessing items such as “child can say three words” and “child uses the words I, me and you” and was considered to capture a different concept than the above cognitive processing outcomes. We also analysed the motor development scale separately (consisting of 20 items, Kohnbachs alpha=0.9492 (for children ages 12-36 months), with items such as “child can kick a ball forward, child can walk on tiptoe, child can stand for a moment on their own”). For language development, Stoltzfus 2001 used a scale of language development modified for use in Kiswahili (range 0-18). The main effect of mebendazole four times per year was not statistically significant (0.3 units out of 18 point scale (95%CI: -0.3 to 0.9) and the interaction term of mebendazole and iron was not statistically significant either. The main effect of iron vs. no iron was a statistically significant difference of 0.8 units on an 18 point scale (95% CI: 0.2 to 1.4). Stoltzfus 2001 also found no difference in motor development between mebendazole vs. no mebendazole, reported as a difference of 0.44 units on a 20 unit scale (95% CI: -0.22 to 1.10).

Awasthi 2000 assessed the revised pre-screening Denver questionnaire (also a measure of early child development), and foundno difference in the outcome of questionable Denver questionnaire in children of 1.01 (95% CI: 0.82 to 1.23).

Network meta-analysis: We did not perform network meta-analysis due to the paucity of studies.

Studies not included in meta-analysis: Three studies had insufficient information for meta-analysis and reported no effect of deworming vs. control on cognitive processing. Miguel 2004 reported no statistically significant effects of deworming vs. control on a battery of cognitive tests given in 2000 (the third year of the study) to all groups, which included both attention/short-term memory tasks as well as general intelligence measures of “picture search, Raven's matrix, verbal fluency, digit span … and a dynamic test using syllogisms”(Miguel 2004). Solon 2003 reported no differences on a measure of general intelligence, the primary mental abilities test for Filipino children, which measures verbal and nonverbal quantitative abilities between children treated with albendazole twice per year or those receiving a placebo (Solon 2003). Jinabhai 2001B reported no differences in Raven's coloured progressive matrices, Rey Auditory Verbal Learning Test or Symbol Digit Modalities Test between children treated with albendazole and praziquantel compared to those with no deworming.

Controlled before after studies: We did not include any CBA with cognition outcomes.

Publication bias: There was an insufficient number of studies (<10) to rely on funnel plots.

School attendance or participation

In summary: Based on our analyses, mass deworming results in little or no difference in school participation vs. placebo (moderate certainty evidence).

Pairwise meta-analysis: Seven studies reported school attendance, using either school records or unannounced site visits throughout the year. The interventions in the analysis included: Albendazole twice/year (Rozelle 2015), Albendazole four times per year (Watkins 1996), Albendazole twice/year +iron and vitamin A (Bobonis 2006), mebendazole+iron once in six month period (Ebenezer 2012), Albendazole 2/year +hygiene promotion with or without praziquantel once per year (Miguel 2004), Albendazole 4/year (Kruger 1996), and thiabendazole four times per year (Gateff 1972). We used generic inverse variance method because we only had difference scores for two studies (Gateff 1972 and Bobonis 2006). We converted all measures to percentage of school attended over the duration of the study (e.g. by expressing days missed by the number of possible school days in the term).

Network meta-analysis: We decided to attempt network analysis because we had ≥ five studies. However, two studies could not be included because they did not report attendance for each arm; they only reported the difference between groups (Gateff 1972, Bobonis 2006). The network contained five trials, with six nodes (Figure 14). The network converged, however, the credible intervals were wide (from -100 to +100 per cent attendance differences), and we could not assess the assumption of consistency because we had no closed loops (i.e. no studies with three or four arms). Therefore, we felt these results were not robust and do not report them here.

Compared to placebo, using pairwise analyses, the pooled result across all mass deworming interventions was a difference of 1 per cent attendance (95%CI: -1% to 3%), with moderate heterogeneity indicated by I² of 67 per cent and p=0.004 for the chi² of 21.05 with seven degrees of freedom (Figure 15). We rated down the certainty by one level (to moderate) because of heterogeneity(see below for sensitivity analyses).

We explored reasons for heterogeneity with pre-planned subgroup and sensitivity analyses (described in detail below). The pooled result was consistent for all sensitivity analyses and there were no subgroup differences, with one exception, described below.

The only subgroup or sensitivity analysis which explained heterogeneity was confounded by two factors: 1) risk of bias and 2) method of measuring school attendance (Figure 16). Two studies which measured school attendance with on-site visits also had high risk of bias for lack of blinding both personnel and participants (in contrast to the other studies which used school records to monitor attendance that had low or unclear risk of bias for blinding).

Power analysis: These seven studies provided school attendance data on approximately 20,000 children (5,142 from six studies, and the remaining from the Miguel 2004 study). Using the Hedges and Pigott (2001) method, assuming 8 independent studies, we had 99.97 per cent power to detect a difference of 7.5 per cent in school attendance with an alpha of 0.05 (two-tailed test). We used 7.5% as a minimum important difference since this was described as an important difference in Miguel and Kremer (2004).

Controlled before after studies: There were no CBAs with school attendance outcomes.

Studies not included in meta-analysis: Two long-term studies report attendance. These are discussed below in a section on long-term outcomes.

School performance (math and reading)

In summary: based on our analyses, mass deworming results in little or no difference in school performance in math or language compared to placebo (high certainty evidence).

We conducted separate analyses for math and language/reading tests (e.g. English, Vietnamese, and French). We decided that network meta-analysis was not worthwhile since there were no closed loops, and five or fewer studies for each outcome.

Pairwise meta-analysis, math: For math, five studies could be combined using generic inverse variance. We used standardised mean differences in math scores at end of study since two studies reported math scores standardised to a mean of zero and a standard deviation of one (Miguel 2004 and Rozelle 2015), while the others reported the scores out of 100. There was no materially important difference between deworming and control, with a pooled difference of -0.03 SMD(95%CI: -0.07, 0.01) (Figure 17). This SMD difference is equivalent to 0.7 per cent when converted to percentage out of 100 scores.

Pairwise meta-analysis, language: Tests of language or reading were reported in sufficient detail for meta-analysis by three studies (Ebenezer 2013, Hall 2006, Watkins 1996). We used standardised mean difference because two reported percentages and one reported the mean on the Interamerican vocabulary test (Watkins 1996). The pooled effect of deworming vs. control showed no materially important effect with an SMD of-0.02(95% CI -0.08, 0.04). This SMD difference is equivalent to 1.1 per cent when converted to a score out of 100 (Figure 18).

These results on school performance in language were robust to sensitivity analyses, as described below.

Controlled before after studies: There were no CBAs with math or language outcomes.

Studies not included in meta-analysis: Two other studies reported no statistically significant difference between deworming and placebo for test scores. Nga 2011 reported no statistically significant differences between any groups (Albendazole, fortified biscuits, placebo or fortified biscuits+albendazole) for math or Vietnamese tests (size of effect not given). Jinabhai 2001B found no difference in math scores between deworming and control groups.

Mortality

In summary: based on our analyses, mass deworming results in little or no difference in mortality compared to placebo (high certainty evidence).

Pairwise meta-analysis: Six RCTs reported mortality. The DEVTA study (Awasthi 2013) was designed to assess the primary outcome of mortality. Deaths were also reported in five other studies. The pooled risk ratio showed no materially important effect of deworming vs. control with risk ratio of 0.95 (95% CI:0.89 to 1.02), Heterogeneity I²=0 per cent (Figure 19).

Network meta-analysis: We did not perform network meta-analysis.

Controlled before after studies: There were no CBAs with mortality as an outcome.

Adverse effects

In summary: Based on our analyses, deworming with albendazole results in little or no difference in adverse effectsvs. placebo (moderate certainty evidence). Deworming with other drugs (thiabendazole, praziquantel, and tetrachloroethylene) may increase nausea, headache, vomiting and abdominal pain (low certainty evidence¹).

Adverse effects were reported by Fox 2005, Gateff 1972, Michaelsen 1995, Olds 1999, Sur 2005 and Wiria 2013. Fox 2005 and Wiria 2013 both studied albendazole and reported no difference in fever, headache, myalgia or cough with albendazole vs. placebo (or for diethyl carbamazine in Fox 2005 vs. placebo or albendazole). Gateff 1972 reported statistically significant greater headache with thiabendazole 3/year vs. placebo. Michaelsen 1995 reported nausea, drowsiness and sleepiness in 17 per cent of the children treated with tetrachloroethylene 0.1 ml/kg vs. cough syrup (given as placebo). Olds 1999 studies albendazole alone, praziquantel alone and albendazole + praizquantel vs. placebo, and reported statistically significantly more nausea reporting an adjusted odds ratio of 2.11 (95% CI 1.49, 2.98) with praziquantel and praziquantel + albendazole vs. placebo or albendazole alone, as well as similar magnitude increase in vomiting, abdominal pain, headache, and diarrhea.

Controlled before after studies: No CBAs reported adverse effects.

Long term outcomes

In summary: mass deworming for soil-transmitted helminths may slightly increase hours worked per week and school enrolment and have little to no effet on self-reported health and height (low certainty evidence) but it is uncertain if this is due to deworming alone or deworming and hygiene education. It is uncertain whether mass deworming leads to improved math achievement, language achievement or reduces work days missed due to sickness because the certainty of evidence is very low. It is uncertain whether early life (less than one year of age) exposure to dewormed siblings or communities leads to improved height, reduced stunting, or improved cognitive processing because the certainty of evidence is very low.

Details of long-term studies

Baird 2016 used the Kenya Life Panel Survey (2007-2009), which followed a sample of 7500 children in primary school in grades 2 to 7 at the time of the Kenya Primary School Deworming Project (PDSP)(Miguel 2004). The Kenya PDSP assigned 75 schools with 31,445 children to mass deworming using a stepped wedge design. Baird et al 2016 followed arepresentative sample of 7,500 children (randomly selected) from the Miguel 2004 study who were in grades 2 to 7 at the time of the PDSP study. Since all schools eventually received the mass deworming, the average difference in deworming between treatment and control children was 2.41 years in the 10 year follow-up study. The Kenya Life Panel Survey found86.7 per cent of this pre-specified group of children(Baird 2016). The children were an average age of 11.9 years in 1998. The sample of found children included 5,569 respondents (3,686 treatment and 1,883 control), and there was no difference in the tracking rate across treatment or control children. Long-term outcomes were collected for 5,084 of these identified children (16 per cent of the children randomized to the original study). We rate the risk of bias for this follow-up study based on the original study and the follow-up study, as high risk of bias for allocation concealment, blinding and outcome reporting (Table 2). We downgrade the GRADE certainty by two levels for study limitations for all outcomes. While the original study did assess the potential for the control group to take up deworming on their own as very low since less than 5% of people were buying deworming medicines in a nearby area in Kenya, and the attitudes of parents to deworming were not positive. However, there is still the potential for the knowledge of the treatment schools to affect the participant behaviour. This study conducted a number of sensitivity analyses that increase the confidence in their findings and estimation strategy. Also, since the difference in time of treatment between the treatment and control arms was only 2.41 years, any differences observed are likely to be smaller than would be observed between a group which received no mass deworming at all. According to their results, mass deworming and hygiene education may improve hours worked per week and school enrolment, and may have little to no effect on height and self-reported health. This study also reported differences in the size of effect for men compared to women for hours worked (3.49 hours for men, 95 per cent confidence interval: 0.71-6.27 and 0.32 hours for women (95%CI: -2.35 to 2.99), which was hypothesized to be related to differences in effects of educational investment for women and men in this setting. This was not reported as a pre-planned analysis, and when we conducted a test for interaction to assess subgroup differences, this subgroup difference was not statistically significant (results not shown), thus we assessed this subgroup analysis as very low certainty evidence.

Ozier 2016 collected data on 21,309 children in Kenya from the areas of the Primary School Deworming Project cluster randomized trial conducted in Kenya in 1998 (Miguel 2004). Children were sampled if they were aged 8-14 years in 2009 and 9-15 years in 2010, according to self-reported age by the children (who often could only report age to the nearest year). A random sample of eligible non-migrants were selected for cognitive testing, using a Stata random number generator, which resulted in a sample of 923 from group 1 schools, 934 from group 2 schools and 514 from group 3 schools (control). It is unclear why the sampling strategy selected almost twice as many children from intervention schools as from control schools; this was described as an error in the Stata code (Ozier 2011). This study was rated at unclear risk of bias due to sampling strategy. As described above, Ozier 2016 conducted two analyses: 1) comparing 21 within birth cohort two-arm tests to contrast intervention and control schools for cognition measured with Raven's matrices; and 2) to compare each of the primary outcomes for children exposed to a treated sibling before age 1 to children exposed toa treated sibling after age 1. Tests of falsification and robustness were conducted to support the hypothesis of the article. However, the tests as randomized were described as “low power”, and consisted of about 353 children per comparison. Exposure to mass deworming and hygiene promotion of school-age children in the community before age one year may leadto long-term improvements in Raven's test scores but not height or height for age, but it is uncertain whether this effect is due to mass deworming or to hygiene promotion (Table 12) [very low certainty evidence].

Croke 2014 followed the sample from the Alderman 2006 RCT conducted in church parishes in Uganda, where there was a baseline prevalence of 60 per cent of STH (mostly hookworm). Children in intervention groups were offered albendazole 400 mg at the health clinic days. The authors estimated that 34 per cent of the control children had accessed deworming medicine through an external source in the original RCT by the end of the three year study. Croke 2014 used data collected by Uwezo in 2010, which surveyed 22 out of the 48 parishes in the deworming study (10 treatment and 12 control), and identified 763 children who were aged 1-7 years in the study period of 2000-2003 and were eligible for the Alderman 2006 study (2.7 per cent of the original sample of 27,995 children). This follow-up study was rated at high risk of samplingbias, since it is unknown why the 22 parishes were sampled and whether they were a representative sample of the original 48 parishes randomized. Data on math and English scores were available for 710 children, standardised to a mean of zero and a standard deviation of 1. The unadjusted effect of deworming was 0.30 standard deviations [95% CI: -0.00, 0.60] for math and 0.16 standard deviations [95% CI -0.17, 0.50] for English. When adjusted for age, gender, survey year and interactions of these variables, the effect for math was 0.36 standard deviations (95% CI: 0.11, 0.62) and for English was 0.25 standard deviations (95% CI: -0.01, 0.50). Croke 2014 found these results were robust to a range of sensitivity analyses, and the effect was larger in poorest income quintiles and girls. Using the correlations from Ozier 2016, these effect sizes equate to 0.5 to 0.8 years of education. Despite the importance of this magnitude of effect, it is uncertain whether mass deworming leads to long-term improvements in English or math scores because of very low certainty evidence².

Makamu 2016 compared four states in Nigeria which received mass deworming with praziquantel for schistosomiasis control to 33 states in Nigeria which did not, and assessed education outcomes. The findings of 0.642 more years of education (7.9 per cent more than the average years of education of 8.13 years in control states) are considered very low certainty evidence because of the observational design and the moderate risk of bias. The study conducted two analyses: a difference in differences analysis with 1990 as the pre-intervention year and 2008 and 2013 as post-intervention, and a cohort study. Baseline characteristics and comparability across groups in 1990 were not shown (Table 12).

Costs or resource use

Five studies reported the cost per single dose of deworming or the cost per year for deworming (Alderman 2006, Awasthi 2000, Garg 2002, Gateff 1972, Nga 2011). These estimates ranged from less than $0.05 per dose of albendazole (Nga 2009) to $0.40/dose of albendazole (Alderman 2006). These do not consider delivery or implementation costs.

Eight studies conducted some type of economic evaluation within their studies or using study data. We did not appraise the quality of these economic evaluations. Three studies reported the cost to deliver deworming drugs (considering the administration and services required when incorporated into health or school systems). Alderman 2006 estimated $ 1-1.33 per child to deliver deworming once to twice a year when incorporated into child health days in Uganda; Awasthi 2013 estimated $D 0.10/child incorporated into child health management by Anganwadi workers in India; Bobonis 2006 estimated $1.70 per child per year for preschool delivery in India. Two studies compared costs of deworming to cost of providing school meals. Gupta 1977 estimated deworming drugs for one year cost less than 4 per cent of a year's ration of food for a child. Gateff 1972 calculated that deworming cost seven to 18 times less than food programmes. Four studies calculated the incremental cost-effectiveness ratio per unit of outcome. Alderman 2006 reported $0.42 per 10 per cent increase in weight, Awasthi 2000 reported 543 Indian rupees per case of stunting, Miguel 2004 reported $D 3.5 per additional year of school participation, and Stoltzfus 1997 reported $D3.57 per case of moderate to severe anaemia prevented.

Miguel 2007 found that a cost-recovery programme reduced the take-up of deworming medicine by 80 per cent.

Physical fitness

In summary: It is uncertain whether mass deworming leads to differences in physical fitness (very low certainty evidence³).

Three studies reported physical fitness; two reported no effect of deworming vs. control (Solon 2003, Bhoite 2012). We decided not to statistically pool the two studies with sufficient data for meta-analysis (Bhoite 2012 and Stephenson 1993) due to high heterogeneity (I² of 93%), with results going in opposite directions.

In one cluster randomised trial (Bhoite 2012) of three schools, which we adjusted for unit of analysis errors, there was a small difference of3.40 steps on the Harvard Step Test [0.05, 6.75] (very low certainty evidence).

In the other study of 53 boys (Stephenson 1993), there was an increase in the Harvard step test score of five units with single dose of deworming vs. 0 units in the placebo group (95%CI: 3.34 to 6.66), when assessed four months after deworming (very low certainty evidence). The Harvard step test score is calculated as 300*100 divided by sum of heart rates per minutes at one, two and three minutes after test completion. Scores of <55 are considered poor fitness, and scores >90 are considered excellent. The difference of five units between intervention and control is thus approximately equivalent to 5 per cent of the whole scale.

Malaria, HIV, tuberculosis outcomes

These outcomes may be subject to selective reporting since 48 studies were at unclear or high risk of bias for selective reporting. HIV and TB outcomes were not reported by any of the included studies.

Malaria incidence was reported by three studies. Ndibazza found reduced malaria incidence in children who received albendazole from age six months to five years compared to placebo (hazard ratio 0.85, 95%CI 0.73, 0.98). Wiria 2013 found a significant increase in malaria parasitemia (P=0.0064) in children treated with albendazole every three months for 21 months, which the authors attribute to the transient increase in malarial parasitemia observed at six months post treatment (odds ratio 4.16(95% CI: 1.35, 12.80) in children treated with albendazole. In Stoltzfus 2004, malaria outcomes were similar across four groups of mebendazole, mebendazole+iron, iron alone and placebo (results not shown).

Age

In summary: For age, we found no subgroup differences across three categories of age (<2 years, 2-5 years or >5 years) for weight, height or school attendance. These agreed with within study analyses for weight and height. However, for school attendance, two studies showed that children aged 4-7 years had greater benefits than younger children or older children for attendance.

Pairwise meta-analysis, for weight and height: In the pairwise analysis, there was no statistically significant difference in effect across three age groupsfor albendazole 400 mg twice per year vs. control and albendazole >2/year vs. control (Error! Reference source not found. and Additional Table 21).

Pairwise meta-analysis for school attendance: The test for interaction for subgroup effects across age was not statistically significant for any deworming vs. control for school attendance (Additional Table 21).

Overall pairwise meta-analysis for weight and height: When we conducted subgroup analysis for any deworming vs. control, the test for interaction for subgroup effects was not statistically significant across age for weight gain (p=0.59)or height gain (p=0.93)(Additional Table 23).

For height and weight within studies: Three studies found no differences across age: Awasthi 2008 found no differencesin weight or height amongst one year age groups from 1-5 years of age), Awasthi 2013 (no difference between 1-3 and 3-6 years in weight, height and haemoglobin), Bobonis 2006 (no difference between 2-3 year olds and 4-6 year olds for weight). Stoltzfus 1997 found a statistically significant interaction between age and treatment, with larger weight and height gains for children younger than 10 years vs. those older than 10 years.

For wasting, stunting and anaemia within studies: Stoltzfus 2004 conducted a post-hoc analysis for children < 30 months vs. children greater than 30 months because they found age differences in effect of deworming on wasting, and chose the cut-off of 30 months because it differentiated the pre-intervention relationship between parasitic infection and iron status. Their analyses showed a statistically significant difference in effect of deworming on wasting malnutrition for children <30 months (relative risk 0.30 [0.11, 0.79]) vs. children >30 months (relative risk 1.18 (0.73 to 1.90). The effect of deworming on stunting was not statistically different between <30 months vs. >30 months, and there was no effect on anaemia in either age group.

For haemoglobin, within studies: Olds 1999 found no difference in effect on haemoglobin across age. Thi Le Huong 2007 found younger children had greater increase in haemoglobin than older children.

For attendance within studies: Bobonis 2006 found effect on school participation was larger for 4-6 year olds than 2-3 year olds and Miguel 2004 found the effect on school participation of albendazole vs. control was larger for preschool to grade 2 [10 per cent] than grade 3-5 [7 per cent], and grade 6-8 [5.9 per cent]).

Nutritional status (<30 per centstunted or =>30 per cent of population stunted)

In summary: For baseline nutritional status, we found no differences in height, weight or school attendance effects for populations more stunted at baseline than less stunted.

Pairwise meta-analysis for weight and height: We found little to no difference across baseline nutritional status in effect on weight gain in the pairwise meta-analysis for albendazole twice per year vs. placebo, with effect of 0.02 SMD (95%CI: -0.05 to 0.09) for populations with <30 per cent stunted vs. 0.05 SMD (95%CI: -0.62 to 0.72) for populations with >30 per cent of children stunted at baseline, test for interaction for subgroup effects not statistically significant (p=0.61) (Error! Reference source not found.).

Overall pairwise meta-analysis for weight and height: There was also little to nodifference in the effect on weight gain or height gain when all studies of deworming were pooled together, test for interaction for subgroup differences was p=0.92 for weight and p=0.80 for height (Additional Table 23). These results were robust to using a different cut-off of population proportion stunted (varying the cut-off from 20 per centto 60 per cent) (Appendix 13.8)

Pairwise meta-analysis for attendance: We did not have sufficient information to conduct a subgroup analysis because we could not ascertain the proportion stunted in three studies (Ebenezer 2013, Kruger 1996, Watkins 1996). The other studies all had less than 30 per cent of the population stunted at baseline, and the result was similar to the overall effect for school attendance with a difference of5 per cent [95%CI: -1% to 11%]).

Within studies, weight: Four out of five primary studies that assessed whether effects of deworming were different for stunted children found no difference in effects on weight. Awasthi 2008 and Dossa 2001 found no difference in effect on weight for children <-2HAZ vs. children greater than -2 HAZ. Ebenezer 2013 reported no interaction between baseline stunting and treatment effect on haemoglobin and attendance outcomes. Stoltzfus 1997 found a statistically significant interaction between HAZ and programme group for both weight gain (p<0.047) and height gain (p=0.008) for children <10 years of age, with greater weight and height gain in the children <10 years who were not stunted. The weight gain compared to control was 0.24 kg for normal HAZ vs. 0.20 kg for stunted children for mebendazole twice yearly, and 0.36 kg for normal HAZ vs. 0.1 kg for stunted children for mebendazole 3/year). Height gain compared to control for the twice-yearly mebendazole was 0.36 cm (0.38 cm for HAZ=0 vs. 0.02cm for HAZ=-2). A similar interaction was found in the older age group. The cut-off of age 10 years was a post-hoc decision.

Worm prevalence

In summary: There was no difference in effect of mass deworming on weight, height or school attendance depending on baseline prevalence of worms.

Pairwise meta-analysis for weight and height: The test for interaction for subgroup effects was not statistically significant for effects on weight or height foralbendazole twice per year vs. placebo or for albendazole >2/year vs. placebobetween categories of prevalence of <20 per cent, 20-50 per cent or >50 per cent (using WHO cut-offs for low, moderate and high prevalence) (Additional Table 21).

Overall pairwise any deworming vs. control, weight and height: There was no difference in the effect for weight or height gain between these three prevalence categories for any deworming vs. control (p=0.12 for weight, p=0.69 for height) (Additional Table 23). We also assessed the influence of the prevalencecut-offs by assessing the effect on weight gain for studies above the threshold from 10 per cent prevalence to 90 per cent prevalence, and there was no difference in effect of deworming at any of these cut-offs (Appendix 13.8).

Pairwise school attendance: The test for subgroup differences for studies <50 per cent prevalence or >=50 per cent prevalence was not statistically significant (p=0.21) (Additional Table 22). We also assessed the influence of prevalence cut-off by assessing the effect on attendance for studies above the threshold from 10 per cent prevalence to 90 per cent prevalence, and there was no difference in effect of deworming at any of these cut-offs(Appendix 13.8).

Sex

In summary: There is probably no difference in effect of deworming between boys and girls for weight or height gain. For school attendance, two studies at high risk of bias found greater effects on attendance for girls than boys.

Meta-analysis: We did not have sufficient data to conduct subgroup analyses across sex in meta-analysis for any outcome.

Within study, weight and height: Eleven studies assessed whether the treatment effect for weight or height was different for boys compared with girls. Seven found no difference in effects of deworming across sex. Of the other four studies, two found a larger treatment effect for girls (Donnen 1998 for weight gain of 1.9 vs. 1.6 kg, respectively and Bobonis 2006 for WHZ), and two found a larger treatment effect for boys (Stoltzfus 1997 for height, Awasthi 2001 for weight gain). Awasthi 2001 reported greater weight gain with albendazole+ vitamin A vs. vitamin A alone for males compared to females (3.31 kg in males vs. 3.11 kg in females).

Within study, attendance: Miguel 2004 and Bobonis 2006 found larger effect of deworming vs. control for girls than boys (in Miguel 2004, (0.067 difference for girls vs. 0.043 for boys in the second year; and 0.098 for girls vs. 0.041 for boys in first year of treatment). Gateff 1972 reported no significant difference between sexes for attendance.

Subgroup analysis for method of measuring school attendance: unannounced visits or school records

In summary: This analysis was confounded because the two studies which measured attendance with on-site visits were also at high risk of bias for lack of blinding personnel and participants (Bobonis 2006 and Miguel 2004). On-site enumerators were likely aware of the treatment status of the schools which may have affected their behaviour. However, the on-site visits were not announced thus decreasing the chance that students or teachers would influence the school attendance data collected.

There was a larger effect of deworming vs. control on school attendance in two cluster randomised trials which measured school attendance using unannounced site visits, with a pooled effect of 8 per cent (95%CI: 3 to 13%), low certainty evidence (Bobonis 2006, Miguel 2004) compared to an effect of deworming in five studies that used school records to measure attendance of -1%, 95%CI: -3 per cent to 1 per cent, high certainty evidence. As indicated in Figure 24, the studies with unannounced visits were at high risk of bias for blinding of personnel and participants.

Sensitivity including Koroma 1996 and Stephenson 1989

As described above, we excluded two studies from our base case network meta-analysis because they contributed significantly to the heterogeneity for the most common intervention of albendazole twice per year vs. control (Koroma 1996, Stephenson 1989). The pooled random effects effect size for weight for albendazole twice per year vs. control was 0.05 SMD (95%CI: -0.02,0.11), I²=61 per cent without these two studies, compared to 0.25 SMD (95%CI: 0.11 to 0.39), I²=93 per cent with these studies. Using the median standard deviation of all studies which measured weight in kg, this equates to a difference of 0.4 kg with the two studies or 0.08 kg without them.

For height, with these two studies, the pairwise effect sizefor albendazole twice per year vs. placebo was an SMD of 0.18 [95% CI: 0.03, 0.33] vs. an SMD of 0.03 [95% CI: -0.02, 0.09] without these two studies (I² of 11%). This equates to a difference of 0.43 cm with the two studies vs. 0.07 cm without.

Choice of attendance effect size from Kenya PSDP (Miguel 2004)

Our analysis was robust to which measure of attendance from the Miguel 2004 study. As described above, for Miguel 2004, we chose to use the measure of deworming effect on school participation for intended recipients (girls <13 years and all boys) in the first year of treatment which was 0.093 (SE 0.031) for the whole study (in both Miguel 2004 and the pure replication Aiken 2015a). We chose this estimate because it was also used in the Cochrane review on deworming (Taylor-Robinson 2015), and also it is one of the largest estimates of the direct effect on treated children. The analysis of Miguel 2004 and Aiken 2015 that combined the first year of treatment for group 1 schools and first year of treatment of group 2 schools, yielded a more precise estimate of 0.062 (SE 0.015) [or 0.60, se 0.015 in Aiken 2015a) for school participation. We tested the pooled effect on school participation with this estimate, as well as with the overall treatment effect of 0.039 (SE 0.031) from Aiken 2015a and the overall treatment effect estimated by Hicks 2015 a in response to Aiken 2015a of 0.085 (SE 0.017). Our meta-analysis results remained non-statistically significant with less than 2 per cent difference in attendance between deworming and control for all of these analyses.

For Ebenezer 2013, there was no difference in the results for attendance if we used the dataset with missing children or the dataset where we imputed missing values using last observation carried forward.

Cluster randomised trials

In summary: Our analyses for weight, height and school attendance were robust to restricting to cluster randomised controlled trials.

Network meta-analysis: When the network meta-analysis for weight gain was restricted to cluster-randomised trials, the network was consistent (total residual deviance 11.72, deviance information criteria -36.6). Deworming compared to placebo in cluster randomised trials probably leads to little to no difference in weight gain, with effect sizes of between -0.12 to 0.17 (moderate certainty evidence) (Additional Table 24).

In the pairwise analysis for weight: The pairwise meta-analysis restricted to cluster trials only confirms the base case analysis of little to no effect of deworming on weight, with no statistically significant effects, and no effect size greater than SMD of 0.22 (equivalent to 300 grams).

In the pairwise analysis, for height in cluster randomised trials, deworming probably has little to no effect on height gain compared to placebo (moderate certainty evidence).

In pairwise meta-analysis, for school attendance: When restricted to cluster randomised trials, results were robust, and there was little to no effect on attendance 1 per cent (95%CI: -1% to 3%).

In pairwise meta-analysis, for worm prevalence: When restricted to cluster randomised trials, there is an important reduction in worm burden in the mass deworming compared to control at end of study (results not shown).

Studies with >50 per cent reduction in ascaris

In summary: Our results for weight, height and attendance are robust to restricting to studies with >50 per cent relative reduction in ascaris prevalence.

Network and pairwise meta-analysis, weight: We restricted the network meta-analysis for weight gain to studies with greater than 50 per cent relative risk reduction in burden of worms, when compared to the control group. The network was consistent, total residual deviance 12.43, DIC -10.625. The results are consistent with the base case analysis, with no statistically significant or clinically important (all less than 0.15 standard deviations) effects of deworming vs. placebo or any head-to-head comparisons, with two exceptions. The effect of albendazole twice per year (0.70 SMD (95% CI: 0.02,1.41)] and albendazole once per year (0.73(95%CI: 0.03,1.43)) were statistically significant, with a clinically important effect size. This is driven by a network with a small number of studies, and only one trial of Albendazole of these doses had >50 per cent impact on worms: Stephenson 1993 (Additional Table 25).

Weight, sensitivity for threshold impact on worms: We tested the threshold for impact on worms by comparing deworming vs. placebo for threshold cutoffs of high impact on worm burden from 10-90 per cent in 10 per cent increments, and found that the effect on weight with deworming vs. control was not statistically significant or clinically important at any threshold for impact on worms.

Network and pairwise meta-analysis for height: The network meta-analysis was consistent. There were no clinically or statistically significant differences in height for studies with >50 per cent impact on worm prevalence. As above, we tested the threshold for the cutoff from 10 to 90 per cent, and found no statistically or clinically important differences at any threshold for impact on worms.

Pairwise meta-analysis, attendance: Results for school attendance were robust to restricting to studies with >50 per cent impact on ascaris, with effect on attendance for deworming vs. control of 3 per cent (95% CI: -4% to 10%).

Studies with >50 per cent hookworm infection

In summary: Results for weight, height and attendance are robust to restricting to studies with >50 per cent hookworm infection.

Pairwise meta-analyses for weight: When restricted to studies with >50 per cent hookworm at baseline, effect on weight was less than 0.21 SMD for all comparisons, with none statistically important, with the exception of albendazole once per year (Stephenson 1993). This study found an effect of 0.73 SMD [95% CI: 0.44, 1.03] on weight for deworming vs. control (low certainty evidence)⁵.

Pairwise meta-analyses for height: The effects on height gain were robust to restricting to studies with >50 per cent hookworm infection, with no effect size greater than SMD 0.15 (equivalent to 0.36 cm). One exception is that the Miguel 2004 study found a small difference in height for age at the end of one year of treatment that was statistically significant (0.11 SMD [95% CI: 0.04, 0.18]).

Pairwise meta-analysis for attendance: Pairwise meta-analysis for attendance was robust to restricting to studies with >50 per cent hookworm (Gateff 1972, Miguel 2004), with mean difference in attendance of 5 per cent (95%CI: -4% to 13%, I²=74%).

Low risk of bias for allocation concealment

In summary: results of little to no effect on weight, height and attendance were robust to restricting to studies at low risk of bias.

Weight, network and pairwise meta-analyses: The network meta-analysis of studies at low risk of bias for allocation concealment was consistent, residual deviance 6.125, DIC -10.952. The results are congruentwith the base case analysis, showing no statistical or clinically important differences between deworming and placebo or any head to head comparisons (effect sizes were between -0.09 and 0.13 SMD, equating to -125 grams to 181 grams). Pairwise results were in alignment with these results.

Height, network and pairwise meta-analyses: The network meta-analysis was consistent, residual deviance 5.552, DIC -9.401. Results were congruent with the main analyses, showing little to no effect on height.

Pairwise meta-analyses, school attendance: When restricted to studies at low risk of bias for allocation concealment (Rozelle 2015, Ebenezer 2013), the effect on school attendance was small and not statistically significant:-2% (95%CI: -6% to 3%).

Sensitivity for eligibility criteria: including studies which screened for infection and treated infected children

We conducted a sensitivity analysis to assess the effect of including studies where children were screened for infection, and infected children were randomised to deworming or control.

Intensity of infection, >30 per cent with heavy infection

In summary: Results of little to no effect of mass deworming vs. control were robust to restricting to studies with >30 per cent heavy infection.

Network and pairwise meta-analysis, weight: For weight, network meta-analysis of 11 studies (8 two-arm, one 3-arm, and two 4-arm studies) converged and was consistent (total residence deviance 16.22, DIC -12.714). Compared to placebo, deworming had little to no effect for all comparisons (less than SMD 0.20), with one exception. Albendazole once per year vs. placebo may increase weight gain compared to placebo (SMD 0.48, 95% CI: -0.51 to 1.48). This was a weight gain of 1.1 kg in one year in Stephenson 1993. Pairwise meta-analyses agreed with these results (results not shown).

Network and pairwise meta-analysis, height: For height, network meta-analysis of 10 studies (7 two-arm, one 3-arm, two 4-arm studies) converged and was consistent (total residence deviance 13.75, DIC -24.283). Compared to placebo, deworming had little to no effect for all comparisons (less than SMD 0.13). There was agreement with pairwise meta-analyses (results not shown).

Pairwise meta-analysis for attendance: For attendance, pairwise analyses of studies where >30 per cent of infected children had moderate to heavy infections (Miguel 2004, Watkins 1996) showed no effect on attendance (-1%, 95%CI: -4 to 2%).

Lower ICC for weight and height (0.01 instead of 0.17 and 0.11)

Using a lower ICC had no effect on the pooled estimates for weight and height for any deworming treatment.

Higher ICC value for cognition of 0.15 instead of 0.07

We did not need to adjust for clustering in our base case analysis. In the sensitivity analysis for eligibility criteria, we adjusted Hadidjaja 1998. Using a higher ICC of 0.15 (as used by Kristjansson 2015 for cognition outcomes of schoolfeeding) does not change the results.

Effect of unpublished studies

We included two unpublished studies (Hall 2006, Rozelle 2015). The effects for weight, height, cognition and attendance were not changed when these studies were removed (results not shown).