SENSITIVITY ANALYSES Early exclusive / total enteral feeding versus gradual enteral feeding with parenteral support in preterm infants Generated 2026-09-02 - includes author-supplied additional arm data for Razzaghy 2024 and the corrected Razzaghy 2024 length-of-stay row (published mean +/- SD, Table 2) ================================================================================================================ S1. FIXED-EFFECT VERSUS RANDOM-EFFECTS MODEL ---------------------------------------------------------------------------------------------------------------- Necrotising enterocolitis fixed RR 0.80 [0.43, 1.49] random RR 0.84 [0.42, 1.71] I2=2% All-cause mortality before discharge fixed RR 1.00 [0.67, 1.49] random RR 1.03 [0.68, 1.54] I2=0% Culture-positive late-onset sepsis fixed RR 0.79 [0.58, 1.06] random RR 0.63 [0.35, 1.15] I2=50% Feed intolerance fixed RR 0.71 [0.57, 0.88] random RR 0.72 [0.54, 0.95] I2=28% Hypoglycaemia fixed RR 0.76 [0.50, 1.17] random RR 0.80 [0.37, 1.75] I2=45% Duration of hospital stay (days) fixed MD -1.80 [ -2.47, -1.14] random MD -3.94 [ -6.40, -1.49] I2=90% Time to attain full enteral feeds (days) fixed MD -1.36 [ -1.63, -1.10] random MD -2.30 [ -3.47, -1.14] I2=90% Time to regain birth weight (days) fixed MD -2.08 [ -2.57, -1.59] random MD -2.82 [ -5.12, -0.53] I2=94% Interpretation: for the five dichotomous outcomes the two models agree closely because between-trial variance is small. For the three continuous outcomes I2 is 90% or above, so the fixed-effect intervals are implausibly narrow; the pre-specified random-effects estimates are reported throughout. S2. EXCLUDING TRIALS AT HIGH OVERALL RISK OF BIAS ---------------------------------------------------------------------------------------------------------------- Trials retained ( 6 ): Nangia 2019; Razzaghy 2024; Sahu 2024; Alshaikh 2025; Ojha 2025 (FEED1); Mishra 2026 Trials excluded ( 6 ): Sanghvi 2013; Zecca 2014; Bora 2017; Jajoo 2022; Raman 2023; Nangia 2026 Necrotising enterocolitis RR 0.67 [0.28, 1.58], p=0.3605, k=5, I2=0% all trials: RR 0.84 [0.42, 1.71], p=0.6380, k=8, I2=2% All-cause mortality before discharge RR 1.17 [0.74, 1.83], p=0.5065, k=4, I2=0% all trials: RR 1.03 [0.68, 1.54], p=0.8958, k=6, I2=0% Culture-positive late-onset sepsis RR 0.82 [0.45, 1.52], p=0.5314, k=5, I2=54% all trials: RR 0.63 [0.35, 1.15], p=0.1353, k=8, I2=50% Feed intolerance RR 0.69 [0.51, 0.92], p=0.0131, k=5, I2=9% all trials: RR 0.72 [0.54, 0.95], p=0.0225, k=10, I2=28% Hypoglycaemia RR 1.14 [0.72, 1.81], p=0.5817, k=4, I2=0% all trials: RR 0.80 [0.37, 1.75], p=0.5741, k=6, I2=45% Duration of hospital stay (days) MD -3.90 [-10.05, 2.24], p=0.2133, k=3, I2=75% all trials: MD -3.94 [-6.40, -1.49], p=0.0016, k=7, I2=90% Time to attain full enteral feeds (days) MD -2.34 [-3.76, -0.91], p=0.0013, k=4, I2=92% all trials: MD -2.30 [-3.47, -1.14], p=0.0001, k=5, I2=90% Time to regain birth weight (days) MD -1.71 [-6.22, 2.81], p=0.4590, k=2, I2=93% all trials: MD -2.82 [-5.12, -0.53], p=0.0157, k=5, I2=94% Interpretation: excluding the six trials at high overall risk of bias leaves the direction of the two feeding-progression outcomes and feed intolerance unchanged, and feed intolerance remains statistically significant with heterogeneity falling to 9%. Length of stay and birth-weight regain lose significance, but only three and two trials respectively remain and the intervals widen greatly, so this reflects loss of precision rather than bias-driven effect. Hypoglycaemia reverses direction (RR 1.14) once the high-risk trials are removed; with wide intervals on either side of unity in both analyses this outcome is best read as no demonstrated difference, and the instability is carried into the GRADE judgement. S3. ALTERNATIVE METHODS FOR RARE DICHOTOMOUS EVENTS ---------------------------------------------------------------------------------------------------------------- Mantel-Haenszel risk ratios exclude trials with no events in either arm. Peto odds ratios, risk differences and inverse-variance risk ratios with a continuity correction retain them. Necrotising enterocolitis MH RR (primary) k= 8 RR 0.84 [0.42, 1.71] p=0.6380 Peto OR k= 8 OR 0.79 [0.42, 1.51] p=0.4765 Risk difference k=12 RD -0.002 [-0.008, 0.003] p=0.4274 IV RR (incr 0.5) k=12 RR 0.86 [0.44, 1.66] p=0.6532 All-cause mortality before discharge MH RR (primary) k= 6 RR 1.03 [0.68, 1.54] p=0.8958 Peto OR k= 6 OR 1.00 [0.61, 1.63] p=0.9889 Risk difference k= 7 RD -0.000 [-0.005, 0.005] p=0.9844 IV RR (incr 0.5) k= 7 RR 1.03 [0.69, 1.54] p=0.8964 Culture-positive late-onset sepsis MH RR (primary) k= 8 RR 0.63 [0.35, 1.15] p=0.1353 Peto OR k= 8 OR 0.76 [0.54, 1.07] p=0.1153 Risk difference k= 9 RD -0.036 [-0.076, 0.004] p=0.0806 IV RR (incr 0.5) k= 9 RR 0.66 [0.38, 1.16] p=0.1461 Feed intolerance MH RR (primary) k=10 RR 0.72 [0.54, 0.95] p=0.0225 Peto OR k=10 OR 0.63 [0.47, 0.84] p=0.0018 Risk difference k=11 RD -0.050 [-0.110, 0.010] p=0.1035 IV RR (incr 0.5) k=11 RR 0.72 [0.55, 0.94] p=0.0144 Hypoglycaemia MH RR (primary) k= 6 RR 0.80 [0.37, 1.75] p=0.5741 Peto OR k= 6 OR 0.71 [0.42, 1.20] p=0.2007 Risk difference k= 6 RD -0.045 [-0.137, 0.047] p=0.3373 IV RR (incr 0.5) k= 6 RR 0.85 [0.43, 1.68] p=0.6475 Interpretation: for enterocolitis, mortality and hypoglycaemia the four measures agree that no difference is demonstrated. For feed intolerance the relative measures are consistent and statistically significant while the risk difference is not (-5.0 percentage points, 95% CI -11.0 to +1.0), reflecting the wide spread of baseline risk across trials; the pre-specified risk ratio is retained as primary. Sepsis is not significant on any measure, with the risk difference closest to conventional significance. S4. ALTERNATIVE HETEROGENEITY ESTIMATOR AND SMALL-SAMPLE CONFIDENCE INTERVALS ---------------------------------------------------------------------------------------------------------------- DL = DerSimonian-Laird (pre-specified). REML = restricted maximum likelihood. HK = Hartung-Knapp-Sidik-Jonkman small-sample adjustment. Duration of hospital stay (days) DL MD -3.94 [-6.40, -1.49], p=0.0016, k=7, I2=90% REML MD -4.34 [-7.88, -0.81], p=0.0159 REML+HK MD -4.34 [-8.67, -0.02], p=0.0492 Time to attain full enteral feeds (days) DL MD -2.30 [-3.47, -1.14], p=0.0001, k=5, I2=90% REML MD -2.29 [-3.42, -1.17], p=0.0001 REML+HK MD -2.29 [-3.90, -0.69], p=0.0166 Time to regain birth weight (days) DL MD -2.82 [-5.12, -0.53], p=0.0157, k=5, I2=94% REML MD -2.83 [-5.39, -0.26], p=0.0306 REML+HK MD -2.83 [-6.48, 0.82], p=0.0979 Necrotising enterocolitis DL RR 0.84 [0.42, 1.71], p=0.6380, k=8, I2=2% REML RR 0.84 [0.42, 1.70], p=0.6329 REML+HK RR 0.84 [0.36, 1.98], p=0.6494 All-cause mortality before discharge DL RR 1.03 [0.68, 1.54], p=0.8958, k=6, I2=0% REML RR 1.00 [0.62, 1.60], p=0.9921 REML+HK RR 1.00 [0.57, 1.75], p=0.9918 Culture-positive late-onset sepsis DL RR 0.63 [0.35, 1.15], p=0.1353, k=8, I2=50% REML RR 0.59 [0.30, 1.17], p=0.1304 REML+HK RR 0.59 [0.26, 1.36], p=0.1814 Feed intolerance DL RR 0.72 [0.54, 0.95], p=0.0225, k=10, I2=28% REML RR 0.72 [0.54, 0.97], p=0.0279 REML+HK RR 0.72 [0.52, 1.01], p=0.0531 Hypoglycaemia DL RR 0.80 [0.37, 1.75], p=0.5741, k=6, I2=45% REML RR 0.92 [0.51, 1.65], p=0.7798 REML+HK RR 0.92 [0.39, 2.16], p=0.8118 Interpretation: REML gives point estimates almost identical to DerSimonian-Laird with somewhat wider intervals for the heterogeneous continuous outcomes. Under the Hartung-Knapp adjustment, which is preferred when between-trial variance is large and the number of trials small, the time-to-full-feeds benefit persists (MD -2.29 days, 95% CI -3.90 to -0.69) and length of stay is only marginally significant (p=0.049), while birth-weight regain (p=0.098) and feed intolerance (p=0.053) both cross unity. This fragility is carried directly into the GRADE imprecision and inconsistency judgements. S5. LEAVE-ONE-OUT INFLUENCE ANALYSIS (random-effects) ---------------------------------------------------------------------------------------------------------------- Necrotising enterocolitis omit Omitting Sanghvi 2013 RR 0.84 [0.42, 1.71] I2=2% omit Omitting Zecca 2014 RR 0.84 [0.42, 1.71] I2=2% omit Omitting Bora 2017 RR 0.70 [0.33, 1.47] I2=0% omit Omitting Nangia 2019 RR 1.01 [0.48, 2.12] I2=0% omit Omitting Jajoo 2022 RR 0.76 [0.33, 1.76] I2=10% omit Omitting Raman 2023 RR 0.84 [0.42, 1.71] I2=2% omit Omitting Razzaghy 2024 RR 0.83 [0.37, 1.87] I2=15% omit Omitting Sahu 2024 RR 0.90 [0.40, 2.02] I2=12% omit Omitting Alshaikh 2025 RR 0.84 [0.42, 1.71] I2=2% omit Omitting Ojha 2025 (FEED1) RR 0.92 [0.37, 2.30] I2=14% omit Omitting Nangia 2026 RR 0.94 [0.45, 1.93] I2=0% omit Omitting Mishra 2026 RR 0.74 [0.35, 1.55] I2=0% All-cause mortality before discharge omit Omitting Jajoo 2022 RR 0.96 [0.57, 1.60] I2=10% omit Omitting Raman 2023 RR 1.03 [0.68, 1.54] I2=0% omit Omitting Razzaghy 2024 RR 1.00 [0.66, 1.51] I2=0% omit Omitting Sahu 2024 RR 1.05 [0.69, 1.58] I2=0% omit Omitting Ojha 2025 (FEED1) RR 0.97 [0.56, 1.71] I2=13% omit Omitting Nangia 2026 RR 1.18 [0.76, 1.83] I2=0% omit Omitting Mishra 2026 RR 0.75 [0.36, 1.57] I2=0% Culture-positive late-onset sepsis omit Omitting Sanghvi 2013 RR 0.63 [0.35, 1.15] I2=50% omit Omitting Nangia 2019 RR 0.66 [0.34, 1.28] I2=53% omit Omitting Jajoo 2022 RR 0.68 [0.37, 1.24] I2=51% omit Omitting Raman 2023 RR 0.66 [0.35, 1.22] I2=54% omit Omitting Razzaghy 2024 RR 0.57 [0.30, 1.07] I2=54% omit Omitting Sahu 2024 RR 0.78 [0.47, 1.31] I2=36% omit Omitting Ojha 2025 (FEED1) RR 0.46 [0.20, 1.01] I2=45% omit Omitting Nangia 2026 RR 0.72 [0.39, 1.32] I2=47% omit Omitting Mishra 2026 RR 0.47 [0.19, 1.20] I2=57% Feed intolerance omit Omitting Sanghvi 2013 RR 0.73 [0.54, 0.99] I2=35% omit Omitting Zecca 2014 RR 0.72 [0.54, 0.95] I2=28% omit Omitting Bora 2017 RR 0.66 [0.52, 0.83] I2=0% omit Omitting Nangia 2019 RR 0.76 [0.56, 1.05] I2=29% omit Omitting Jajoo 2022 RR 0.72 [0.53, 0.98] I2=35% omit Omitting Raman 2023 RR 0.71 [0.53, 0.95] I2=32% omit Omitting Razzaghy 2024 RR 0.68 [0.51, 0.92] I2=26% omit Omitting Sahu 2024 RR 0.76 [0.57, 1.02] I2=24% omit Omitting Alshaikh 2025 RR 0.72 [0.53, 0.98] I2=36% omit Omitting Nangia 2026 RR 0.78 [0.57, 1.05] I2=22% omit Omitting Mishra 2026 RR 0.72 [0.51, 1.01] I2=34% Hypoglycaemia omit Omitting Zecca 2014 RR 1.14 [0.72, 1.80] I2=0% omit Omitting Raman 2023 RR 0.75 [0.31, 1.82] I2=56% omit Omitting Razzaghy 2024 RR 0.61 [0.23, 1.60] I2=34% omit Omitting Sahu 2024 RR 0.82 [0.31, 2.16] I2=52% omit Omitting Alshaikh 2025 RR 0.66 [0.24, 1.80] I2=56% omit Omitting Mishra 2026 RR 0.77 [0.29, 2.04] I2=56% Duration of hospital stay (days) omit Omitting Sanghvi 2013 MD -1.97 [ -3.53, -0.41] I2=73% omit Omitting Zecca 2014 MD -4.47 [ -7.45, -1.50] I2=92% omit Omitting Bora 2017 MD -4.74 [ -8.00, -1.49] I2=92% omit Omitting Raman 2023 MD -4.34 [ -7.41, -1.27] I2=91% omit Omitting Razzaghy 2024 MD -3.73 [ -6.27, -1.19] I2=91% omit Omitting Alshaikh 2025 MD -3.64 [ -6.17, -1.10] I2=91% omit Omitting Ojha 2025 (FEED1) MD -4.94 [ -7.75, -2.14] I2=88% Time to attain full enteral feeds (days) omit Omitting Nangia 2019 MD -1.81 [ -2.76, -0.86] I2=80% omit Omitting Sahu 2024 MD -2.06 [ -3.25, -0.86] I2=92% omit Omitting Alshaikh 2025 MD -2.57 [ -4.17, -0.96] I2=93% omit Omitting Ojha 2025 (FEED1) MD -2.70 [ -3.81, -1.60] I2=76% omit Omitting Nangia 2026 MD -2.34 [ -3.76, -0.91] I2=92% Time to regain birth weight (days) omit Omitting Sanghvi 2013 MD -1.73 [ -3.21, -0.26] I2=82% omit Omitting Zecca 2014 MD -2.94 [ -6.01, 0.13] I2=95% omit Omitting Bora 2017 MD -3.26 [ -6.33, -0.19] I2=94% omit Omitting Sahu 2024 MD -2.54 [ -5.25, 0.16] I2=95% omit Omitting Alshaikh 2025 MD -3.64 [ -6.24, -1.05] I2=95% Interpretation: no single trial drives any conclusion, but several outcomes are fragile. For feed intolerance every leave-one-out estimate stays at or below 0.78 and the pooled effect loses conventional significance only when Nangia 2019, Nangia 2026, Mishra 2026 or Sahu 2024 is omitted (upper limits 1.01-1.05), so significance rests on no one trial but is not robust to removing any of the larger contributors. For sepsis the estimate strengthens to RR 0.46 (0.20-1.01) when FEED1, the only high-income trial, is omitted, consistent with the setting subgroup effect. Enterocolitis, mortality and hypoglycaemia remain non-significant under every omission; hypoglycaemia flips above unity only when Zecca 2014 is dropped, and Bora 2017 accounts for the entire residual heterogeneity in feed intolerance. For length of stay, omitting Sanghvi 2013 halves the effect from -3.94 to -1.97 days and reduces I2 from 90% to 73%, so that early trial contributes much of both the magnitude and the heterogeneity. Every length-of-stay and time-to-full-feeds omission remains significant; birth-weight regain loses significance when Zecca 2014 or Sahu 2024 is omitted. S6. SMALL-STUDY EFFECTS AND FUNNEL PLOT ASYMMETRY ---------------------------------------------------------------------------------------------------------------- Cochrane guidance advises against formal asymmetry testing with fewer than 10 trials. Only feed intolerance reaches that threshold, so all tests below are exploratory and reported for completeness. Necrotising enterocolitis k= 8 Egger p=0.8614 Begg-Mazumdar p=0.6207 All-cause mortality before discharge k= 6 Egger p=0.7831 Begg-Mazumdar p=0.8510 Culture-positive late-onset sepsis k= 8 Egger p=0.0759 Begg-Mazumdar p=0.4579 Feed intolerance k=10 Egger p=0.3663 Begg-Mazumdar p=0.2449 Hypoglycaemia k= 6 Egger p=0.1088 Begg-Mazumdar p=0.1885 Duration of hospital stay (days) k= 7 Egger p=0.0841 Begg-Mazumdar p=0.1765 Time to attain full enteral feeds (days) k= 5 Egger p=0.0664 Begg-Mazumdar p=0.1416 Time to regain birth weight (days) k= 5 Egger p=0.4190 Begg-Mazumdar p=0.6242 Trim-and-fill (exploratory, feed intolerance only - the single outcome with k>=10): observed RR 0.72 [0.54, 0.95], k=10 trim-and-fill RR 0.65 [0.47, 0.90], k=12 (2 imputed) Interpretation: no test reaches conventional significance. The borderline Egger p-values for sepsis (0.076), length of stay (0.084) and time to full feeds (0.066) are computed on 5-8 trials, where the test has almost no power and is strongly influenced by the true heterogeneity already documented; they cannot distinguish publication bias from genuine between-trial variation. For feed intolerance, the only adequately powered outcome, both tests are clearly non-significant, but trim-and-fill nonetheless imputes two hypothetical unpublished trials and moves the estimate away from the null (RR 0.65, 95% CI 0.47 to 0.90) rather than towards it, so the observed RR of 0.72 is not an artefact of missing small negative trials on this analysis. Visual inspection of the funnel plots shows the small trials from lower-income settings clustered towards larger benefit for the feeding-progression outcomes, which is the same signal the setting subgroup analysis captures; for length of stay Sanghvi 2013 sits far outside the funnel on the benefit side. Publication bias cannot be excluded for any outcome given that most trials are small and single-centre, and this is reflected in the GRADE judgements.