Research question

What are the associations between modifiable risk and protective factors and medically attended injuries resulting from five common injury mechanisms in children under the age of 5 years?

Methods

Five multicentre case–control studies were undertaken (study A). Cases were children aged < 5 years attending secondary care with a fall (three types: fall from furniture, fall on one level or a stair fall), poisoning or a scald. Control subjects (controls) were matched to cases on age, sex and calendar time, and were recruited from the register of the cases’ general practice (or neighbouring general practice). Exposures (safety equipment use, safety behaviours and hazards) were measured using parent-completed questionnaires and were validated by home observations in a sample of cases and controls (study B). Odds ratios (ORs) were estimated using conditional logistic regression adjusted for confounding factors.

Results

Conclusions

Despite a small number of apparently counterintuitive findings (indicated with an asterisk), a range of modifiable risk factors were associated with falls from furniture, falls on stairs, poisonings and scalds in children aged 0–4 years. These results provide evidence on which to base safety advice and recommendations.

Work stream 1 consisted of five case–control studies (study A) quantifying associations between modifiable risk factors and falls from furniture, falls on one level, falls on steps or stairs, poisonings and scalds. Work stream 1 also included a study to validate self-reported exposures in the case–control studies (study B). The findings from work stream 1 informed:

• the decision analyses undertaken in work stream 5
• the development of an IPB for the prevention of fire-related injuries, falls, poisonings and scalds undertaken in work stream 6 (see Chapter 7).

Objectives

The primary objectives of study A were to estimate associations between modifiable risk and protective factors and medically attended injuries resulting from five injury mechanisms in children aged < 5 years:

1. falls from furniture
2. falls on one level
3. stair falls
4. poisoning
5. scalds.

Our secondary objectives were to explore whether or not associations between risk and protective factors and injuries varied by child age, sex, ethnicity, single parenthood, housing tenure and unemployment and injury severity.⁴⁹

Study design

We used five multicentre matched case–control studies [one for each of the injury mechanisms (a)–(e)].

Setting

We recruited participants from EDs, minor injury units (MIUs) and hospital wards from acute NHS trusts in Nottingham, Bristol, Newcastle upon Tyne, Norwich, Gateshead, Derby, Great Yarmouth and Lincoln, UK. Recruitment of cases commenced on 14 June 2010 for all studies and finished on (a) 15 November 2011 for the falls from furniture study, (b) 15 November 2011 for the falls on one level study, (c) 30 September 2012 for the stair falls study, (d) 18 January 2013 for the poisoning study and (e) 18 January 2013 for the scalds study. Recruitment of controls commenced with recruitment of the first case to each study and controls were recruited within 4 months of recruitment of cases.

Participants

Cases were children aged 0–4 years with:

1. a fall from furniture
2. a fall on one level
3. a stair fall
4. a poisoning or suspected poisoning from a medicinal or other household product or
5. a scald, resulting in hospital admission or ED or MIU attendance.

Injuries had to have occurred at the address at which the child was registered with a general practitioner (GP) (hereafter referred to as the child’s home). Intentional and fatal injuries were excluded, as were children living in residential care. Cases were eligible to be recruited only once to the study.

We used two sources of controls; community controls and hospital controls. For clarity and simplicity the findings relating to community controls (hereafter referred to as controls) are presented in the main text of the report. Findings relating to hospital controls are summarised in Appendix 1. Children living in residential care were excluded. Controls were children aged 0–4 years without a medically attended injury of the same mechanism as the case on the date of the case’s injury. Controls were eligible to be recruited as a case or as a further control if their second recruitment occurred at least 12 months after their first recruitment. They were not eligible to be recruited more than twice to the study. We aimed to recruit an average of four controls per case, individually matched on age (within 4 months of age of case), sex and calendar time (within 4 months of case injury). To increase the study power and make the most efficient use of controls, when we recruited more than four controls per case (or when cases were later excluded), the extra controls were eligible to be matched to other cases who did not have four matched controls. These were matched on age (within 4 months of age of case), sex and calendar time (within 4 months of the case injury) and study centre, and were eligible to be used only once as an extra matched control.

The eligibility of putative cases to take part in the study was assessed from medical records by clinical staff prior to study invitations being issued. Research staff also assessed eligibility on receipt of completed study questionnaires. Potentially eligible cases were approached by clinical staff face to face during their medical attendance or by telephone or post within 72 hours of their attendance. Controls were recruited by post by general practice or primary care trust (PCT) staff, from the practice register of the case’s GP or, when the case’s practice was unable to participate, from that of a neighbouring practice. To minimise age differences between cases and controls resulting from the time taken to recruit practices and then recruit controls, study invites were sent to children born up to 4 months before and 2 months after the case’s date of birth. Ten children were invited to participate for each case. One reminder was sent to case and control non-respondents 2 weeks after the original mailing.

Variables

Bias

We used several strategies to try to minimise bias. We aimed to minimise recall bias by inviting cases to participate in the study within 72 hours of the injury attendance and measured exposures over a short time period prior to the injury attendance (ranging from 24 hours to 1 week); for controls we measured exposures over the same time period prior to completing the study questionnaire. When possible, we validated the accuracy of self-reported exposures in cases and controls by home observations. To minimise non-response bias, we used methods shown in systematic reviews to increase response rates, including providing a small monetary incentive (£5) for the return of completed questionnaires, using personalised letters, sending mail by first class post, providing Freepost reply envelopes, using reminders including the provision of further questionnaires, keeping the questionnaire as short as possible and using university logos on study documentation.^69,70

Study size

Quantitative variables

All exposures were categorical variables. For confounders measured on a continuous scale, we assessed the linearity of their relationship with outcome measures by adding higher-order terms to regression models and tested significance using likelihood ratio tests with a p-value of < 0.05 taken as significant. When the relationship between age and the outcome of interest was non-linear we grouped age into the three age groups consistent with the age groups for which we had developed age-specific questionnaires (0–12 months, 13–36 months, ≥ 37 months). When other relationships were non-linear, we examined distributions of the confounders and grouped values based on cut-off points that separated the distribution into groups of similar values while ensuring sufficient numbers in each group for analysis. When standard groupings had been used in previous research, for example quintiles of deprivation scores, we grouped values similarly to allow comparisons with previous research. The cut-off points for groupings are given in the results tables.

We devised a score representing parents’ perceptions of their child’s ability to climb by combining responses across eight questions asking about perceptions of ability to climb or reach a range of hazards. Each question had a three-point Likert scale response from ‘not likely’ to ‘very likely’, with a ‘don’t know’ option. The score was created by categorising responses into (1) all not likely, (2) at least one quite likely but none very likely and (3) at least one very likely. Those with missing or ‘don’t know’ responses to individual items were categorised as missing an overall score unless respondents had at least one ‘very likely’ response. We also devised a composite categorical variable describing parents’ perceptions of their stair characteristics by combining responses across seven questions with a three-point Likert response from ‘agree’ to ‘disagree’ (stairs too steep, stairs too narrow, stairs poorly lit, steps in need of repair, banister/handrail in need of repair, stair covering in need of repair and stairs being safe to use). The categories of the composite variable were ‘unsafe’ (answered agree to any of the first six questions or disagree to last question), ‘moderately safe’ (answered combinations of agree, disagree and neither agree nor disagree) and ‘safe’ (answered disagree to all first six questions and agree to last question). Those with missing responses to individual items were categorised as missing a response on the composite variable unless they agreed with any of the first six questions or disagreed with the last question.

When the relationship between distance from hospital and the outcome of interest was non-linear, distance was grouped into quintiles. When the relationship between IMD and the outcome of interest was non-linear IMD was grouped into quintiles.

We used the treatment received as a proxy for injury severity. We created two categories: those who were seen and examined but who did not require any treatment and those requiring treatment in the ED, admitted to hospital or discharged with outpatient or primary care follow-up. We chose these groupings based on the number of cases in each group and combined hospital admissions, those treated in the ED and those discharged with outpatient or primary care follow-up as the numbers admitted to hospital and discharged with outpatient or primary care follow-up were small.

Statistical methods

We calculated kappa coefficients, sensitivities, specificities and predictive values (and 95% CIs) comparing each reported exposure with the observed exposure, with observations used as the ‘gold standard’.

Characteristics of cases and controls have been described using frequencies and percentages for categorical variables and means [standard deviations (SDs)] or medians [interquartile ranges (IQRs)] for continuous variables dependent on their distributions. Conditional logistic regression was used to estimate unadjusted ORs and AORs and 95% CIs for the matched analysis. Analyses were adjusted for area-level deprivation⁶⁵ and distance from hospital and for confounders identified from DAGs. We developed separate DAGs for each exposure–outcome analysis. All variables that we considered as potential confounders were included in the DAG, and we used Dagitty software [see www.dagitty.net/ (accessed 2 October 2016)] to create a causal diagram for each exposure–outcome analysis and to identify the minimum adjustment set of variables. The regression model for each analysis was adjusted for the variables belonging to the minimum adjustment set for that exposure–outcome analysis by entering them on one step into the model. They were retained in the model regardless of statistical significance or effect on the OR for the exposure. Potential differential effects by child age, sex, ethnicity, single parenthood, housing tenure and unemployment were assessed by adding interaction terms to models. Significance was assessed using likelihood ratio tests with a p-value of < 0.01 taken as significant. When significant interactions were found, the results are presented stratified by socioeconomic variables. Differential effects by injury severity were assessed by stratifying analyses into those who were seen and examined but who did not require any further treatment and those who received treatment or who were admitted to hospital. For the unmatched analyses using hospital controls, unconditional logistic regression was used to estimate unadjusted ORs and AORs and 95% CIs. Analyses were adjusted for age, sex, area-level deprivation⁶⁵ and distance from hospital, in addition to other confounders identified from DAGs. The population attributable fraction (PAF) per cent was calculated for exposures with statistically significantly raised AORs using a published formula.⁷³

We followed standard guidance on missing data for the PedsQL score and did not compute mean scale scores when > 50% of the individual scale score responses were missing. When > 50% of questions were answered, mean scale scores were generated with imputation of missing values using the mean of the answered questions.⁷⁴ For the HADS score we imputed single missing item values for each subscale using the mean of the remaining six items. When more than one item was missing, subscale scores were not computed.⁷⁵ The IBQ, ECBQ and CBQ allowed missing values and were scored as the total score divided by the number of questions answered.⁷⁶ We were unable to find guidance for dealing with missing values for the PDH scale so we used the same approach as for the HADS. The main analyses are complete-case analyses, excluding cases and controls with missing data for the exposure or confounding variables. Complete-case analysis gives unbiased estimates when people with missing data are a completely random subset of the individuals in a particular study; the missing data are then called missing completely at random. If, however, missingness is related to other observed or participant data, for example age or sex, this is called missing at random. We undertook multiple imputation, which assumes that missing data are missing at random, to create 20 imputed data sets. These were combined using Rubin’s rules.⁷⁷ The multiple imputation models included all sociodemographic characteristics, exposures and confounding variables considered in the analysis models, along with case–control status. The imputation models included interaction terms identified in the complete-case analyses when possible, but in some cases the imputation models would not converge when interaction terms were included so these were omitted. When exposures had > 5% of ‘not applicable’ responses, analyses were repeated coding these as a separate category.

For the validation of exposure measures, to assess whether differences between reported and observed practices may have arisen because of changes made to the home by families after completing questionnaires, we incorporated any changes made in the last 3 months as reported at the home visit to derive a modified value for each exposure. For any cell within the tables comparing reported and observed values, when the percentage of people reporting a change in the previous 3 months was > 20%, the numbers were adjusted to accommodate an assumed change from ‘yes’ to ‘no’ and vice versa, and positive predictive values (PPVs) and negative predictive values (NPVs) were recalculated.

Ethics

Approval for the case–control studies and the validation of exposures study was granted by Nottingham Research Ethics Committee 1 (reference number 09/H0407/14).

Validation of exposures study (study B)

The process of recruitment to study B is shown in Figure 2. In total, 113 cases and 119 controls were contacted by the research team, of whom 81 (72%) and 81 (68%), respectively, received a home visit. This represents 3% of cases and 1% of controls eligible for study A. The period of time between receipt of questionnaire and the visit being carried out varied between 1 and 92 days, with a median of 29 days.

FIGURE 2

Recruitment to the validation of exposures study (study B). a, Includes eight cases subsequently found not to be eligible for study A (study C, n = 7; study G, n = 1). These eight were not used to compare characteristics (more...)

Table 1 shows the characteristics of families participating in the home observations and those returning completed questionnaires who were eligible to participate in any of the five case–control studies but who did not have a home observation. For most characteristics, there was no significant difference between families who participated and those who did not. Families for whom the questionnaire respondent was female or the participating child was male were more likely to participate; this was also the case for single-parent families or households with more adults out of work.

TABLE 1

Characteristics of families observed at home and case–control study participants not observed at home

Table 2 shows the sensitivities, specificities and predictive values for exposures related to falls. NPVs were high (≥ 70%) for all 12 exposures relating to risk of falls. However, the PPV was ≥ 70% for only five of the fall exposures: safety gates at the top of the stairs; safety gates at the bottom of the stairs; use of safety gates elsewhere in the house; carpeted stairs; and the presence of a landing half-way up the stairs. The sensitivity for eight of the exposures was ≥ 70%. Exposures for which the sensitivity was < 70% were the presence of safety gates other than on the stairs and use of baby walkers, use of stationary play centres and use of travel cots as play pens. The specificity was ≥ 70% for 10 of the 12 exposures, being below this for banisters on stairs and handrails on stairs. Kappa coefficients ranged from 0.2 for use of baby walkers (slight agreement)⁷⁹ to 0.74 for carpeted stairs (substantial agreement).⁷⁹ There was no significant difference [t = 1.77, degrees of freedom (df) = 1.42, p = 0.08] in measured stair steepness (stair height-to-depth ratio) between those reporting that their stairs were too steep (n = 23; mean 0.87, SD 0.21) and those not reporting that their stairs were too steep (n = 121; mean 0.82, SD 0.09). Observed banister gaps were significantly larger than reported gaps (n = 55; Z = 3.12, p = 0.002). The median reported gap was 3.0 inches (IQR 2.0–4.0 inches) whereas the median observed gap was 3.8 inches (IQR 3.5–4.3 inches).

TABLE 2

Sensitivities, specificities, predictive values and kappa coefficients for agreement between the questionnaire and observations for exposures related to falls

Table 3 shows the sensitivities, specificities, predictive values and percentage agreement for 16 exposures relating to poisoning. All PPVs were low, the highest being 68% for all medicines having CRCs or blister packs. For 11 of the exposures, the NPV was > 70%, whereas sensitivity was > 70% for eight exposures. Kappa coefficients varied from –0.03 for all medicines stored in locked cupboard, cabinet, drawer or fridge (poor agreement)⁷⁹ to 0.54 for medicines kept in fridge (moderate agreement).⁷⁹

TABLE 3

Sensitivities, specificities, predictive values and kappa coefficients for agreement between the questionnaire and observations for exposures related to poisons

Table 4 shows the sensitivities, specificities, predictive values and percentage agreement for exposures relating to scalds. Sensitivity was > 70% for all three scald-related exposures. The PPV was high for two exposures (a kettle with a curly flex and kettle kept at the back of the kitchen surface), whereas the NPV was high for having a safety gate across the kitchen doorway and for having a kettle kept at the back of the kitchen surface. Kappa coefficients ranged from 0.13 (slight agreement) to 0.57 (moderate agreement).⁷⁹

TABLE 4

Sensitivities, specificities, predictive values and kappa coefficients for agreement between the questionnaire and observations for exposures related to scalds

Over-reporting of safety practices was more common than under-reporting. We were able to calculate predictive values for 30 safety practices and found that, for 24 of these, more families over-reported than under-reported (NPV exceeds PPV) and, for the remaining six practices, more families under-reported than over-reported (PPV exceeds NPV).

We explored whether or not differences between reported and observed safety practices could be accounted for by families changing safety practices between completing the questionnaire and the home observation. This did not appear to explain the differences between reported and observed practices, as the findings were similar to those from the main analysis when using the adjusted figures.⁷⁸ The results are available from the authors on request.

Associations between observations and self-reports differed significantly between cases and controls for only one exposure, which was storage of household products in containers that were different from the ones in which they came (χ² = 4.91, p = 0.03). The results are available from the authors on request.

Case–control study of risk and protective factors for falls from furniture (study A)

A total of 672 cases and 2648 controls participated in the study. The process of recruitment to the study is shown in Figure 3. In total, 35% of cases and 33% of controls agreed to participate. The age and sex of participants and non-participants in the falls from furniture study were similar, as shown in Table 5.

FIGURE 3

Flow of cases and controls through the falls from furniture study: (a) recruitment of cases; and (b) recruitment of controls. a, Assumed to be 10 times the number of cases as practices were asked to invite 10 controls for each case; b, controls for cases (more...)

TABLE 5

Age and sex of participants and non-participants

The mean number of controls per case was 3.94. The median time from date of injury to date of questionnaire completion for cases was 10 days (IQR 6–20 days).

Most cases sustained single injuries (86%), most commonly a bang on the head (59%), cuts/grazes not requiring stitches (19%) and fractures (14%). Most cases (60%) were seen and examined but did not require treatment, with 29% treated in the ED, 4% were admitted to hospital and 7% treated and discharged with a follow-up appointment.

Table 6 shows the sociodemographic characteristics of cases and controls. As expected, because controls were recruited after the matched cases, they were slightly older than cases (median age 1.91 vs. 1.74 years). Cases were slightly more likely than controls to live in a household with no adults in paid work (17.7% vs. 12.6%), in a household receiving state benefits (43.0% vs. 35.9%) and in non-owner-occupied housing (39.5% vs. 32.2%).

TABLE 6

Sociodemographic characteristics of cases and controls

Table 7 shows the frequency of exposures among cases and controls and unadjusted ORs, and Table 8 shows ORs adjusted for a range of confounding variables. Adjusting for confounders had relatively little impact on most ORs; only one out of 13 ORs changed by > 10% after adjustment (had things child could climb on to reach high surfaces – OR 0.85, AOR 0.94). Compared with parents of controls, in the adjusted analyses parents of cases were more likely not to use a safety gate (AOR 1.65, 95% CI 1.29 to 2.12; PAF 15%), more likely to leave children on raised surfaces (AOR 1.66, 95% CI 1.34 to 2.06, PAF 23%) and more likely not to have taught their children rules about climbing on objects in the kitchen (AOR 1.58, 95% CI 1.16 to 2.15, PAF 16%) and their children were less likely to climb or play on garden furniture (AOR 0.74, 95% CI 0.56 to 0.97). Most of the ORs for the remaining nine exposures were close to 1, with seven being > 1 (ranging from 1.01 to 1.35) and two being < 1 (0.77 and 0.94). All had CIs indicating that associations could have occurred by chance.

TABLE 7

Frequency of exposures and unadjusted ORs comparing cases with controls

TABLE 8

Adjusted ORs comparing cases with controls

There were significant interactions with child age. ORs for each age group are shown in Table 9. Comparing cases to controls, for children aged 0–12 months, parents of cases were more likely to leave children on raised surfaces (AOR 5.62, 95% CI 3.62 to 8.72; PAF 50%), change nappies on raised surfaces (AOR 1.89, 95% CI 1.24 to 2.88; PAF 34%) and put children in car/bouncing seats on raised surfaces (AOR 2.05, 95% CI 1.29 to 3.27; PAF 12%) than parents of controls. In the 13–36 month age group, parents of cases were less likely to put car or bouncing seats on raised surfaces than controls (AOR 0.22, 95% CI 0.05 to 0.94). The effect for children climbing or playing on furniture was only significant in children aged 3 years and older, with cases being more likely to climb or play on furniture (AOR 9.25, 95% CI 1.22 to 70.07; PAF 88%) than controls.

TABLE 9

Significant interactions in adjusted analyses between exposures and age comparing cases with controls

Tables showing the results of the following sensitivity analyses are available from the authors on request. None of the AORs using the multiply imputed data differed by > 10% from those using the complete-case data for the main analysis. Five of the AORs for the interaction analyses differed by > 10% between the analysis using the multiply imputed data and the analysis using the complete-case data. All six exposures with a ‘not applicable’ response option had > 5% of ‘not applicable’ responses. Analyses were undertaken incorporating a separate category for ‘not applicable’ responses. None of the AORs for the six exposures differed by > 10% between the analyses with and the analyses without the ‘not applicable’ category.

Stratifying analyses by the treatment received (as a proxy for injury severity) resulted in AORs with fairly broad and overlapping 95% CIs for those seen and examined in the ED and those admitted/treated in the ED/discharged with follow-up for all exposures. The results from these analyses are available from the authors on request.

Case–control study of risk and protective factors for falls on one level (study A)

A total of 582 cases and 2460 controls participated in the falls on one level study. The process of recruitment to the study is shown in Figure 4. In total, 24% of cases and 24% of controls agreed to participate and were included in the analysis. Table 10 shows that participants and non-participants were similar in terms of sex and age.

FIGURE 4

Flow of cases and controls through the falls on one level study: (a) recruitment of cases; and (b) recruitment of controls. a, Assumed to be 10 times the number of cases as practices were asked to invite 10 controls for each case; b, controls for cases (more...)

TABLE 10

Age and sex of participants and non-participants

The mean number of controls per case was 4.23. The median time from date of injury to date of questionnaire completion for cases was 10 days (IQR 6–20 days).

The majority of cases had sustained single injuries (80%), most commonly a bang on the head (52%), cuts/grazes not requiring stitches (29%) and cuts needing stitches (17%). In total, 47% of cases were seen and examined but did not require treatment, 46% were treated in the ED, 3% were admitted to hospital and 4% were treated and discharged with a follow-up appointment.

Table 11 shows the sociodemographic characteristics of cases and controls. Controls were slightly older than cases (median age 2.16 vs. 2.08 years). Cases were slightly more likely than controls to have a mother who was aged ≤ 19 years when she had her first child (16.5% vs. 10.8%), to live in a household with no adults in paid work (19.1% vs. 12.4%), to live in a household receiving state benefits (44.3% vs. 37.0%) and to live in non-owner-occupied housing (42.5% vs. 32.7%).

TABLE 11

Sociodemographic characteristics of cases and controls

Table 12 shows the frequency of exposures among cases and controls and unadjusted ORs, and Table 13 shows ORs adjusted for a range of confounding variables. Adjusting for confounders had relatively little impact on most ORs. Four out of 13 ORs changed by > 10% on adjustment (having tripping hazards on floors: OR 0.88, AOR 1.07; not using safety gate to prevent access to garden: OR 0.58, AOR 1.01; unsupervised playing in garden: OR 0.76, AOR 0.89; not using a playpen/travel cot: OR 0.76, AOR 0.90). Compared with parents of controls, in the adjusted analyses parents of cases were less likely not to use furniture corner covers (AOR 0.72, 95% CI 0.54 to 0.94) and less likely to have rugs/carpets not firmly fixed to the floor (AOR 0.77, 95% CI 0.59 to 0.99). ORs for most of the remaining 11 exposures were close to 1, with five being > 1 (ranging from 1.01 to 1.37) and six being < 1 (ranging from 0.73 to 0.97). All had CIs indicating that associations could have occurred by chance.

TABLE 12

Frequency of exposures and unadjusted ORs comparing cases with controls

TABLE 13

Adjusted ORs comparing cases with controls

There was a significant interaction between the number of adults living with the child and rugs/carpets not being firmly fixed to the floor. ORs for each category are shown in Table 14. In households where the child lived with one parent, rugs/carpets were more likely not to be fixed firmly to the floor in cases than in controls (AOR 2.54, 95% CI 1.16 to 5.54, PAF 18%), whereas in households where the child lived with more than one adult they were less likely not to be fixed firmly to the floor in cases than in controls (AOR 0.69, 95% CI 0.52 to 0.90).

TABLE 14

Significant interactions in adjusted analyses between rugs/carpets firmly fixed to the floor and number of adults living with the child, comparing cases with controls

Tables showing the results of the following sensitivity analyses are available from the authors on request. Two of the AORs using the multiply imputed data differed by > 10% from those using the complete-case data for the main analysis (did not use safety gate to prevent access to garden: multiply imputed AOR 0.78, 95% CI 0.50 to 1.21; complete case AOR 1.01, 95% CI 0.58 to 1.74; not taught rules about running in the house: multiply imputed AOR 0.82, 95% CI 0.64 to 1.06; complete case AOR 0.73, 95% CI 0.54 to 1.00). One of the AORs for the interaction analyses differed by > 10% between analyses using the multiply imputed data and analyses using the complete-case data.

Three exposures with a ‘not applicable’ response option had > 5% of ‘not applicable’ responses. Analyses were undertaken incorporating a separate category for ‘not applicable’ responses. One of the AORs for the three exposures differed by > 10% between the analyses with and the analyses without the ‘not applicable’ category (did not use safety gate to prevent access to the garden: AOR with ‘not applicable’ category 1.01, 95% CI 0.58 to 1.74; AOR without ‘not applicable’ category 0.89, 95% CI 0.54 to 1.48).

Stratifying analyses by the treatment received (as a proxy for injury severity) resulted in AORs with fairly broad and overlapping 95% CIs for those seen and examined in the ED and those admitted/treated in the ED/discharged with follow-up for all exposures. The results from these analyses are available from the authors on request.

Case–control study of risk and protective factors for stair falls (study A)

A total of 610 cases and 2658 controls participated in the stair falls study. The process of recruitment to the study is shown in Figure 5. In total, 33% of cases and 29% of controls agreed to participate and were included in the analysis. Table 15 shows that participants and non-participants were similar in terms of sex, but a higher proportion of participants than non-participants were aged 0–12 months (19.0% vs. 12.3%).

FIGURE 5

Flow of cases and controls through the stair falls study: (a) recruitment of cases; and (b) recruitment of controls. a, Assumed to be 10 times the number of cases as practices were asked to invite 10 controls for each case; b, controls for cases from (more...)

TABLE 15

Age and sex of participants and non-participants

The mean number of controls per case was 4.36. The median time from date of injury to date of questionnaire completion for cases was 11 days (IQR 7–21 days).

The majority of cases had sustained single injuries (85%), most commonly a bang on the head (66%), cuts/grazes not requiring stitches (14%) and broken bones (12%). Most cases (64%) were seen and examined but did not require treatment, 25% were treated in the ED, 5% were admitted to hospital and 6% were treated and discharged with a follow-up appointment.

Table 16 shows the sociodemographic characteristics of cases and controls. Compared with controls, cases were less likely to live in a household with more than one adult in paid work (50.0% vs. 59.0%). They also lived in areas with higher levels of deprivation (median IMD score 18.7 vs. 15.2), were more likely to have a mother who had had her first child aged ≤ 19 years (18.5% vs. 9.1%) and were more likely to live in a single adult household (14.6% vs. 10.5%), a household in receipt of state benefits (40.9% vs. 32.4%), in non-owner-occupied housing (40.4% vs. 32.2%) or in a household without a car (14.7% vs. 9.7%).

TABLE 16

Sociodemographic characteristics of cases and controls

Table 17 shows the frequency of exposures among cases and controls and unadjusted ORs and Table 18 shows ORs adjusted for a range of confounding variables. Ten of the 24 ORs changed by > 10% on adjustment, with all reducing in magnitude (no carpet on stairs: OR 1.91, AOR 1.52; stairs too steep: OR 1.35, AOR 1.21; stairs too narrow: OR 1.35, AOR 1.28; stairs need repair: OR 1.97, AOR 1.71; stair covering needs repair: OR 1.74, AOR 1.41; stairs not safe: OR 1.71, AOR 1.46; no banisters: OR 1.44, AOR 1.27; not taught child rules about going downstairs: OR 1.60, AOR 1.36; not taught child rules about leaving things on stairs: OR 1.00, AOR 0.85; banister width 2.5–3.75 inches: OR 0.84, AOR 0.75).

TABLE 17

Frequency of exposures and unadjusted ORs comparing cases with controls

TABLE 18

Adjusted ORs comparing cases with controls

Compared with parents of controls, in the adjusted analysis parents of cases were more likely not to use safety gates on their stairs (AOR 2.50, 95% CI 1.90 to 3.29; PAF 21%) or to leave the gates open (AOR 3.09, 95% CI 2.39 to 4.00; PAF 24%), not to have carpeted stairs (AOR 1.52, 95% CI 1.09 to 2.10; PAF 5%) and not to have a landing part-way up their stairs (AOR 1.34, 95% CI 1.08 to 1.65; PAF 18%). They were also more likely to consider their stairs not safe to use (AOR 1.46, 95% CI 1.07 to 1.99; PAF 5%) or their steps in need of repair (AOR 1.71, 95% CI 1.16 to 2.50; PAF 5%). Case households were less likely than control households to have tripping hazards on their stairs (AOR 0.77, 95% CI 0.62 to 0.97) or to not have handrails on all stairs (AOR 0.69, 95% CI 0.56 to 0.86). ORs for most of the remaining 16 exposures were close to 1, with 11 being > 1 (ranging from 1.07 to 1.41) and five being < 1 (ranging from 0.69 to 0.97). All had CIs indicating that associations could have occurred by chance.

Table 19 shows significant interactions in the adjusted analysis. Comparing cases with controls, parents of cases aged 0–12 months and 13–36 months were more likely to have left safety gates on stairs open than closed, with a particularly high OR in parents with infants (AOR 0–12 months 8.64, 95% CI 3.99 to 18.68, PAF 46%; AOR 13–36 months 2.64, 95% CI 1.92 to 3.64, PAF 24%). Parents of cases in all age groups were more likely not to have a safety gate on stairs than to have a closed gate and the OR was again higher in parents with infants (AOR 0–12 months 3.27, 95% CI 1.48 to 7.20, PAF 18%; AOR 13–36 months 2.33, 95% CI 1.60 to 3.39, PAF 15%; AOR ≥ 37 months 2.08, 95% CI 1.23 to 3.51, PAF 32%). The association between safety gates and stair falls varied by baby walker use, with a particularly high OR for leaving a safety gate open compared with having a closed gate in walker users (AOR walker users 7.37, 95% CI 4.36 to 12.45, PAF 44%). Parents of cases in households in which no adults were in paid work were less likely not to have taught children rules about carrying things downstairs (AOR 0.44, 95% CI 0.20 to 0.96) or not to have taught children rules about leaving things on stairs (AOR 0.27, 95% CI 0.12 to 0.60). Parents of cases in single adult households were more likely not to have carpeted stairs (AOR 11.07, 95% CI 3.89 to 31.53, PAF 26%) and were less likely not to have taught children rules about leaving things on stairs (AOR 0.33, 95% CI 0.15 to 0.75).

TABLE 19

Significant interactions in adjusted analyses comparing cases with controls

Tables showing the results of the following sensitivity analyses are available from the authors on request. In the analysis using community controls, the AORs for four exposures using the multiply imputed data differed by > 10% from those using the complete-case data for the main analysis. These were not having carpeted stairs (multiply imputed AOR 1.68, 95% CI 1.25 to 2.25; complete case AOR 1.52, 95% CI 1.09 to 2.10), having a stair carpet in need of repair (multiply imputed AOR 1.57, 95% CI 1.13 to 2.18; complete case AOR 1.41, 95% CI 0.99 to 2.03), not using a stationary activity centre (multiply imputed AOR 0.96, 95% CI 0.75 to 1.24; complete case AOR 1.08, 95% CI 0.80 to 1.46) and banister width (2.5–3.75 inches: multiply imputed AOR 0.85, 95% CI 0.61 to 1.20; complete case AOR 0.83, 95% CI 0.53 to 1.29; > 3.75 inches: multiply imputed AOR 0.90, 95% CI 0.65 to 1.25; complete case AOR 0.75, 95% CI 0.48 to 1.18). Four of the interactions from the complete-case analysis remained significant at the 1% level in the multiply imputed analysis and two [interactions between (1) teaching children rules about leaving things on stairs and single parent households and (2) teaching children rules about carrying things on stairs and parental unemployment] had p-values of 0.011 in the multiply imputed analysis. Patterns of risk were similar for the multiply imputed and complete-case analyses, but AORs differed by > 10% between the complete-case analyses and the multiply imputed analyses for 12 out of 20 AORs. Stratifying analyses by the treatment received (as a proxy for injury severity) resulted in AORs with fairly broad and overlapping 95% CIs for the seen and examined in the ED and the admitted/treated in the ED/discharged with follow-up groups for all exposures. The results from these analyses are available from the authors on request.

Case–control study of risk and protective factors for poisonings (study A)

A total of 567 cases and 2320 controls participated in the poisonings study. The process of recruitment to the study is shown in Figure 6. In total, 28% of cases and 28% of controls agreed to participate and were included in the analysis. The age and sex of participants and non-participants in the poisonings study were similar, as shown in Table 20.

FIGURE 6

Flow of cases and controls through the poisonings study: (a) recruitment of cases; and (b) recruitment of controls. a, The study the participant had previously been recruited to was not recorded; b, includes eight participants previously recruited to (more...)

TABLE 20

Age and sex of participants and non-participants

The mean number of controls per case was 4.09. The median time from date of injury to date of questionnaire completion for cases was 12 days (IQR 6–22 days).

The majority of cases had sustained single injuries. Parents of seven cases reported injuries in addition to poisonings. Most cases (84%) were seen and examined but did not require treatment, 6% were treated in the ED, 8% were admitted to hospital and 2% were treated and discharged with a follow-up appointment.

Table 21 shows the sociodemographic characteristics of cases and controls.

TABLE 21

Sociodemographic characteristics of cases and controls

Table 22 shows the frequency of exposures among cases and controls and unadjusted ORs, and Table 23 shows ORs adjusted for a range of confounding variables. Four of the eighteen ORs changed by > 10% on adjustment (medicines transferred into different container: OR 1.15, AOR 0.96; household products not stored at adult eye level or above: OR 0.84, AOR 0.95; household products transferred into different containers: OR 1.74, AOR 1.20; no safety gate to stop access to kitchen: OR 0.91, AOR 1.05).

TABLE 22

Frequency of exposures and unadjusted ORs comparing cases with controls

TABLE 23

Adjusted ORs comparing cases with controls

Compared with parents of controls, in the adjusted analyses parents of cases were more likely not to store all medicines at adult eye level or above (AOR 1.59, 95% CI 1.21 to 2.09; PAF 15%) and were more likely not to store all medicines safely (locked away or at adult eye level or above) (AOR 1.83, 95% CI 1.38 to 2.42; PAF 16%). Parents of cases were less likely than parents of controls not to store all household products safely (locked away or at adult eye level or above) (AOR 0.77, 95% CI 0.59 to 0.99). They were also more likely not to put medicines away immediately after use (AOR 2.11, 95% CI 1.54 to 2.90; PAF 20%) or to put household products away immediately (AOR 1.79, 95% CI 1.29 to 2.48;, PAF 11%) and less likely not to have taught their children what to do or not do if medicines were left on the worktop (AOR 0.66, 95% CI 0.45 to 0.96). ORs for the remaining 12 exposures were close to 1, with four being > 1 (ranging from 1.05 to 1.25), and eight being < 1 (ranging from 0.81 to 0.96). All had CIs indicating that associations could have occurred by chance.

Table 24 shows significant interactions in the adjusted analysis comparing cases with controls. Parents of cases living in single adult households were more likely not to store all household products at adult eye level or above (AOR 2.43, 95% CI 1.09 to 5.43; PAF 50%). Parents of female cases were less likely not to keep all medicines locked away (AOR 0.59, 95% CI 0.37 to 0.94).

TABLE 24

Significant interactions in adjusted analyses comparing cases with controls

Tables showing the results of the following sensitivity analyses are available from the authors on request. In the analysis using community controls, four of the AORs using the multiply imputed data differed by > 10% from those using the complete-case data for the main analysis. These were medicines not stored safely (multiply imputed AOR 1.62, 95% CI 1.29 to 2.05; complete case AOR 1.83, 95% CI 1.38 to 2.42), household/cleaning products not stored safety (multiply imputed AOR 0.91, 95% CI 0.73 to 1.13; complete case AOR 0.77, 95% CI 0.59 to 0.99), transferring cleaning products to different containers (multiply imputed AOR 1.47, 95% CI 0.80 to 2.69; complete case AOR 1.20, 95% CI 0.54 to 2.65) and not putting household/cleaning products away immediately after use (multiply imputed AOR 1.61, 95% CI 1.26 to 2.05; complete case AOR 1.79, 95% CI 1.29 to 2.48). Two of the AORs differed by > 10% between the multiply imputed and the complete-case interaction analyses.

One exposure with a ‘not applicable’ response option had > 5% of ‘not applicable’ responses. Analyses were undertaken incorporating a separate category for ‘not applicable’ responses. The AORs did not differ by > 10% between analyses with and analyses without the ‘not applicable’ category. The results from these analyses are available from the authors on request. Stratifying analyses by the treatment received (as a proxy for injury severity) resulted in AORs with fairly broad and overlapping 95% CIs for those seen and examined in the ED and those admitted/treated in the ED/discharged with follow-up for all exposures. The results from these analyses are available from the authors on request.

Case–control study of risk and protective factors for scalds (study A)

A total of 338 cases and 1438 controls participated in the scalds study. The process of recruitment to the study is shown in Figure 7. In total, 32% of cases and 29% of controls agreed to participate and were included in the analysis. The age and sex of participants and non-participants in the scalds study were similar, as shown in Table 25.

FIGURE 7

Flow of cases and controls through the scalds study: (a) recruitment of cases; and (b) recruitment of controls. a, Assumed to be ten times the number of cases as practices were asked to invite 10 controls for each case; b, controls for cases from the (more...)

TABLE 25

Age and sex of participants and non-participants

The mean number of controls per case was 4.25. The median time from date of injury to date of questionnaire completion for cases was 11 days (IQR 6–21 days).

All cases had sustained a scald and no other injury in addition to the scald. In total, 31% received treatment at the ED, 24% were seen and examined but did not require treatment, 18% were admitted to hospital and 27% were treated and discharged with a follow-up appointment.

Table 26 shows the sociodemographic characteristics of cases and controls. Controls were slightly older than cases (median age 1.56 vs. 1.47 years). Cases were less likely to be of white ethnic origin (81.8% vs. 91.3%) and more likely to have only one child under the age of 5 years (67.5% vs. 62.3%), live in an overcrowded household (15.2% vs. 8.6%) or live in a rented home (49.6% vs. 37.1%). Case households were more likely to receive state benefits than control households (46.0% vs. 35.0%). Cases lived in areas of higher social deprivation (median IMD score 20.6 vs. 15.7) and typically spent less time being cared for outside the home (median number of hours per week 5.5 vs. 12). Parental perception of their child’s ability to climb was lower for cases: 78.6% of case parents thought that it was very likely that their child could climb to at least one of the eight places described on the questionnaire compared with 82.8% of control parents.

TABLE 26

Sociodemographic characteristics of cases and controls

Table 27 shows the frequency of exposures among cases and controls and unadjusted ORs and Table 28 shows ORs adjusted for a range of confounding variables. Fifteen of the 28 ORs changed by > 10% after adjustment (no safety gate: OR 1.79, AOR 1.46; things that could be climbed on to reach high surfaces: OR 1.07, AOR 1.24; kettle not at back of worktop/cooker: OR 1.46, AOR 1.20; water temperature too hot: OR 0.67, AOR 0.96; water temperature not known or > 54°C: OR 1.29, AOR 0.99; played/climbed on furniture: OR 0.54, AOR 0.62; held by someone holding hot drink: OR 0.95, AOR 0.83; hot drinks in reach of child: OR 1.99, AOR 2.33; pan handles not turned to back of cooker: OR 1.26, AOR 0.91; child left alone in bathroom: OR 0.53, AOR 0.70; bath run by older child: OR 1.13, AOR 0.74; older child supervised child in bath: OR 0.82, AOR 1.10; not taught rules about bathtubs: OR 2.16, AOR 1.42; not using playpen/travel cot: OR 1.16, AOR 1.33; not using stationary activity centre: OR 1.62, AOR 1.22).

TABLE 27

Frequency of exposures and unadjusted ORs comparing cases with controls

TABLE 28

Adjusted ORs comparing cases with controls

Compared with parents of controls, in the adjusted analysis parents of cases were more likely to have not taught their child rules about things not to climb on in the kitchen (AOR 1.66, 95% CI 1.12 to 2.47; PAF 20%), what to do or not do when parents are cooking using the top of the cooker (AOR 1.95, 95% CI 1.33 to 2.85; PAF 26%) and about hot things in the kitchen (AOR 1.89, 95% CI 1.30 to 2.75; PAF 26%). They were also more likely than control parents to have left hot drinks within reach of their child (AOR 2.33, 95% CI 1.63 to 3.31; PAF 31%). Cases were less likely than controls to have played or climbed on furniture (AOR 0.62, 95% CI 0.40 to 0.96) or to have been left alone in the bath (AOR 0.47, 95% CI 0.30 to 0.75). ORs for most of the remaining 22 exposures were close to 1, with 10 being > 1 (ranging from 1.10 to 1.46) and 12 being ≤ 1 (ranging from 0.70 to 1.00). All had CIs indicating that associations could have occurred by chance.

As shown in Table 29, there were three exposures for which there was a significant interaction with one of the sociodemographic variables. Comparing cases with controls, parents in households with two or more adults in paid work were more likely not to have taught their child rules about what to do or not do when in the bathtub (AOR 2.81, 95% CI 1.43 to 5.53; PAF 33%). Compared with parents of controls, in single adult households parents of cases were less likely to have a hot water temperature of ≥ 54 °C (or not know the water temperature) (AOR 0.42, 95% CI 0.07 to 2.72), whereas, in households with two or more adults, parents of cases were more likely to have a hot water temperature of ≥ 54 °C (or not know the water temperature) (AOR 1.47, 95% CI 0.85 to 2.56). Compared with control parents, case parents living in rented accommodation were more likely never to check their child’s bathwater temperature using a thermometer or other gadget (AOR 1.84, 95% CI 1.03 to 3.28, PAF 36%), whereas parents of cases in owner-occupied housing were less likely to never check the bathwater temperature using a thermometer or other gadget (AOR 0.65, 95% CI 0.42 to 1.03).

TABLE 29

Significant interactions in adjusted analyses comparing cases with controls

Tables showing the complete set of results of the following sensitivity analyses are available from the authors on request. Seventeen of the AORs using the multiply imputed data differed by > 10% from those using the complete-case data for the main analysis and these are shown in Table 30. Three of the AORs differed by > 10% between the multiply imputed and the complete-case interaction analyses.

TABLE 30

Comparison of the results from the complete-case and multiple imputation analyses for those exposures for which there was a > 10% difference

For five exposures, the proportion of ‘not applicable’ responses was > 5%. Analyses were undertaken incorporating a separate category for ‘not applicable’ responses. The AOR differed by > 10% in the analyses comparing cases with controls with and without a ‘not applicable’ category for the bath being run by an older child (AOR with ‘not applicable’ category 0.62, 95% CI 0.28 to 1.34; AOR without ‘not applicable’ category 0.74, 95% CI 0.31 to 1.82) and for older children looking after a younger child in the bath (AOR with ‘not applicable’ category 0.95, 95% CI 0.56 to 1.00; AOR without ‘not applicable’ category 1.10, 95% CI 0.63 to 1.93). The results from these analyses are available from the authors on request.

Stratifying analyses by the treatment received (as a proxy for injury severity) resulted in AORs with fairly broad and overlapping 95% CIs for the seen and examined in the ED and the admitted/treated in the ED/discharged with follow-up groups for almost all exposures. However, the AORs differed between those admitted/treated in the ED/followed up post discharge and those seen and examined in the ED for putting hot drinks or liquids on tables with tablecloths on. Parents of cases who were seen and examined in the ED were less likely than parents of controls to put hot drinks on tables with cloths (AOR 0.11, 95% CI 0.02 to 0.63), but parents of cases admitted to hospital/treated in the ED/followed up post discharge were more likely to put hot drinks on tables with tablecloths than parents of controls (AOR 2.02, 95% CI 1.22 to 3.36). The results from these analyses are available from the authors on request.

Main findings

We found that a range of modifiable risk factors were associated with secondary care-attended falls from furniture, falls on stairs or steps, poisonings and scalds in children aged 0–4 years. Only two modifiable risk factors were associated with secondary care-attended falls on one level in children aged 0–4 years.

Compared with control parents, parents of children who had a fall from furniture were more likely not to use safety gates anywhere in the home (PAF 15%) and, for those with children aged 0–12 months, were more likely to have left them on raised surfaces (PAF 50%), changed nappies on raised surfaces (PAF 34%) and put them in car/bouncing seats on raised surfaces (PAF 12%). Parents of children who had fallen from furniture were less likely to put children aged 13–36 months in car/bouncing seats on raised surfaces (78% reduction in odds). They were more likely not to have taught children rules about things they should not climb on in the kitchen (PAF 16%) and, for those with children aged ≥ 37 months, their children played or climbed on furniture more often (PAF 88%) than control children. Case children who had had a fall from furniture played or climbed on garden furniture less often than control children (26% reduction in odds).

Compared with parents of controls, parents of children who had had a fall on one level were less likely not to use furniture corner covers (28% reduction in odds) and less likely not to have rugs/carpets firmly fixed to the floor (23% reduction in odds). The association with rugs/carpets firmly fixed to the floor varied with the number of adults in the household. Households with one adult in which children had had a fall on one level were more likely not to have rugs/carpets fixed to floors than control households (PAF 18%), whereas households with two adults in which children had had a fall on one level were less likely not to have rugs/carpets fixed to floors than control households (31% reduction in odds).

Compared with parents who kept safety gates closed, parents of children who had fallen down stairs or steps were more likely not to use safety gates on stairs (PAF 21%) or to leave safety gates on stairs open (PAF 24%) than parents of controls. Compared with control households, the odds of not using a safety gate on stairs and of leaving the gate open appeared to be particularly high in families with children aged 0–12 months (PAF 18% and 46%, respectively) and the odds of leaving safety gates open appeared to be high in families who used baby walkers (PAF 44%). Families with children who had fallen down stairs or steps were more likely than controls not to have carpeted stairs (PAF 5%) or not to have a landing part-way up their stairs (PAF 18%). They were also more likely to consider their stairs not safe to use (PAF 5%) or in need of repair (PAF 5%). They were less likely than controls to have tripping hazards on their stairs (23% reduction in odds) or not to have handrails on all stairs (31% reduction in odds). Compared with controls, families with no adults in paid work whose children had fallen down stairs or steps were more likely to have taught children rules about carrying things down stairs and leaving things on stairs (56% and 73% reduction in odds of not teaching rules, respectively). Compared with controls, single adult families whose children had fallen down stairs or steps were more likely to have taught children rules about leaving things on stairs (67% reduction in odds of not teaching rules) and to not have carpeted stairs (PAF 26%).

Compared with parents of controls, parents of children who had had a poisoning were more likely not to store medicines at adult eye level or above (PAF 15%), not to store medicines safely (locked away or at adult eye level or above) (PAF 16%) and not to put medicines (PAF 20%) or household products (PAF 11%) away immediately after use. They were also less likely not to store household products safely (23% reduction in odds) and not to have taught children rules about what to do if medicines are left on the worktop (34% reduction in odds). Parents of children who had had a poisoning in single adult households were more likely than control parents not to store household products at adult eye level or above (PAF 50%). Parents of girls who had had a poisoning were less likely than parents of controls not to lock medicines away (41% reduction in odds).

Compared with parents of controls, parents of children who had had a scald were more likely to have left hot drinks in reach of children (PAF 31%) and more likely not to have taught children rules about climbing in the kitchen (PAF 20%), about what to do or not to do when adults are using the top of the cooker (PAF 26%) or about hot things in the kitchen (PAF 26%). Parents of children who had had a scald were less likely than control parents to have left a child alone in the bath (53% reduction in odds) and their children climbed or played on furniture less often (38% reduction in odds). Compared with controls, families with at least two adults in paid work whose children had had a scald were more likely not to have taught children rules about what to do or not to do when in the bathtub (PAF 33%). Single adult families with children who had had a scald were less likely than controls to report an unsafe or unknown hot water temperature (58% reduction in odds). Compared with parents of controls, parents of children with a scald living in rented accommodation were more likely not to check the bathwater temperature with a thermometer (PAF 36%).

Most exposures were study specific. Not using safety gates (other than on stairs) was an exposure in all five case–control studies and ORs were > 1 for all studies (ranging from 1.05, 95% CI 0.80 to 1.37 for poisonings to 1.65, 95% CI 1.29 to 2.12 for falls from furniture). All case–control studies included exposures related to teaching children safety rules, but only teaching rules about climbing in the kitchen was measured in more than one study, with both studies finding ORs > 1 (falls from furniture OR 1.58, 95% CI 1.16 to 2.15; scalds OR 1.66, 95% CI 1.12 to 2.47). Across all studies, nine of the 13 ORs related to teaching safety rules were > 1 (ranging from 1.13, 95% CI 0.83 to 1.52 for rules about slippery floors and falls on one level to 1.95, 95% CI 1.33 to 2.85 for rules about what to do or not do when parents are using the top of the cooker and scalds).

Strengths and limitations

We report the largest case–control studies to date examining associations between a range of modifiable risk factors for falls, poisoning and scalds. These studies were conducted in NHS hospitals across England, including urban and rural areas. Adjustment was made for a wide range of potential confounding factors using DAGs. For four of our five studies, the majority of ORs using multiply imputed data for the main analysis did not differ by > 10% from those in the complete-case analysis. However, in most studies, a larger number of ORs in the interaction analyses differed by > 10% between the multiple imputation and the complete-case analyses.

Case–control studies have limitations arising from their observational nature. These include, but are not limited to, measurement error, a range of different types of bias and confounding. Each of these are discussed below. We validated measures for exposures when possible and found high (> 70%) sensitivities and specificities for six out of 12 falls exposures, for two out of 15 poisoning exposures and for two out of three scalds exposures. We used home observations as the ‘gold standard’, but it is possible that families may have made changes to their homes either as a result of the injury or in anticipation of the home observation. In addition, there may have been some social desirability bias, with parents reporting responses that they judged to be ‘more acceptable’. These may potentially explain some of the lack of agreement that we found between parent-reported and observed exposures. It is likely that some misclassification of exposures occurred. Misclassification is likely to be lowest when sensitivity and specificity are both high. Sensitivity and specificity were both > 70% for having safety gates on stairs or across kitchen doorways, having carpeted stairs and landings part-way up stairs, having playpens, furniture corner covers and cordless or curly-flexed kettles and storing household products at adult eye level or above. For other exposures, when misclassification is higher it is more likely that ORs are biased towards the null (although this does not always occur⁸⁶) and this may partly explain our failure to find associations between some exposures and falls, poisonings and scalds. We did explore whether or not the differences between self-reported and observed exposures varied between cases and controls, and we found similar levels of under- and over-reporting for both groups for most exposures. This suggests that differential misclassification is unlikely to have occurred to a large extent. Our analyses did not take misclassification of exposures into account, and for this reason our findings should be interpreted with caution. Future studies should consider adjusting for exposure misclassification.

The participation rates for cases and controls were similar, but were low for all studies, ranging from 24% to 35% for cases and from 24% to 33% for controls across the five studies. If reasons for participation are associated with the exposure or outcome of interest, selection bias may have occurred. Our participation rates do not show large differences by case/control status, age and sex, but we were not able to assess the prevalence of exposures in participants and non-participants, and caution should be exercised in interpreting our findings. In addition to social desirability bias discussed above, recall bias may have occurred. These two types of bias could potentially impact on our ORs in different directions. As most injuries were relatively minor, it is possible that a small number of controls may have had similar injuries to cases and on the same date as the case injury but did not seek medical attention for those injuries. This could lead to misclassification of cases and controls. However, this is likely to apply to only a small number of controls; hence, we would expect this to have at most only a minor impact on our results. If seeking medical attention is associated with exposures of interest, this may lead to overestimation of ORs, but the extent to which this may have introduced bias is difficult to assess.

Most ORs for not teaching children various safety rules were > 1 across the five case–control studies and were significantly raised for rules about climbing in the kitchen for falls from furniture and scalds and for rules about what to do or not do when parents were using the cooker top and about not touching hot things for scalds. It is possible that parents who use safety rules supervise their children differently from parents who do not use rules and, as we were not able to adjust for this, residual confounding may partly explain these findings.

In all five case–control studies, cases were more disadvantaged than controls. Socioeconomic disadvantage is likely to be associated with some of the exposures of interest in our studies and, although we adjusted for a range of confounders, it is possible that some residual confounding remained. It is also possible that families belonging to black and minority ethnic (BME) groups were under-represented in our studies and that for some studies (scalds in particular) a higher proportion of cases than controls belonged to a BME group. The proportion of the population reported as white in the 2011 UK census was 86%,⁸⁷ whereas the proportion of cases and controls from a BME group ranged from 8% to 18% and from 7% to 9%, respectively, across our five studies. The generalisability of our findings to BME groups may therefore be limited.

Overall, there were few significant interactions found, but our interaction analyses should be interpreted with caution for several reasons. First, a large number of tests for interactions were performed and, although we used a significance level of 1% for these tests, one in every 100 tests will be significant by chance alone. Second, subgroup numbers were small for many comparisons, leading to insufficient power to detect anything other than large differences in associations. Third, differences in estimates of associations between complete-case and multiple imputation analyses suggest that the findings of some of our interaction analyses were not robust to missing data. Consequently, these analyses should be considered as generating rather than testing hypotheses. The significant interactions that were found need confirmation from further research.

Our findings in relation to age are, in general, in keeping with what would be expected based on child development. For example, the increased odds of falls from furniture only in 0- to 12-month-olds who were left on, had nappies changed on or who were put in car/bouncing seats on raised surfaces is to be expected given that these are activities likely to be undertaken by parents whose children are not independently mobile. We found only a small number of exposures with significant interactions with sex, with the exposure more strongly associated with an injury in boys than in girls. The majority of these were in the poisonings case–control study. Boys have higher mortality rates for poisoning than girls⁸⁸ and have been found to have higher hospital admission rates for poisonings than girls,^55,89 although some studies including poisonings not requiring hospital admissions have failed to find significant differences by child sex.^90,91 As our cases mainly included minor injuries, this would be consistent with finding only a few exposures more strongly associated with poisonings in boys than in girls.

There were significant interactions for several exposures related to teaching safety rules with either the number of adults in the household or the number in paid work in the stair falls and scalds studies. Previous research suggests that parents start teaching safety rules between the ages of 2 and 4 years and at this point they move from mainly using strategies based on supervision and changing the home environment to teaching- and rule-based strategies to prevent injury.^92,93 Previous research also suggests that teaching safety rules can increase the risk of injury^93,94 and that teaching needs to increase children’s understanding of the safety issue to reduce the extent to which they interact with hazards.⁹⁴ Our findings suggest that cases in single adult households and in households without adults in paid work are more likely to have been taught safety rules than controls. It is possible that families in which supervision may be more challenging (e.g. single adult households) or with fewer material resources (e.g. those without adults in paid work) rely more on teaching safety rules than on other injury prevention strategies, but this may not be an effective strategy in these circumstances. Further work is required to explore these hypotheses.

The many exposures in our studies resulted in multiple significance testing; hence, some associations may have been significant by chance alone. Our estimates of associations for some exposures were imprecise because of the low prevalence of some exposures such as use of playpens or stationary activity centres, use of safety gates to prevent access to gardens, transferring household products or medicines to other containers or baths being run by older children. Some exposures were not measured in our studies because they were known to be rare, and our studies were underpowered to detect anything other than implausibly large associations. For example, bunk bed falls account for only 10% of falls from beds and have an annual incidence rate of 0.3 per 1000 children-years.^45,95–98

Comparisons with existing literature

Comparing our findings with those of previously published studies has been limited by differences in the exposures measured, similar exposures measured in different ways or inadequately detailed descriptions of exposure measures. We have not compared our findings to studies in which cases represent a wide range of injury mechanisms⁹⁹ because of the difficulty in interpreting findings that are not specific to single injury mechanisms. We found only one Australian case–control study of infants with head or face trauma¹⁰⁰ matched on age to controls with which to compare the findings from our falls studies. The findings of the Australian study were consistent with our findings for changing nappies on high surfaces (OR 1.77, 95% CI 1.07 to 2.92) and use of high chairs without harnesses (OR 1.47, 95% CI 0.73 to 2.98). We found slightly raised odds of a fall from furniture for children who had not used a baby walker (OR 1.22, 95% CI 0.90 to 1.65) but slightly reduced odds of a fall on one level (0.83, 95% CI 0.58 to 1.10) and for a fall on stairs or steps (AOR 0.83, 95% CI 0.62 to 1.09), which was consistent with the results from the Australian study (OR for ever using a baby walker 0.83, 95% CI 0.50 to 1.38). However, this was inconsistent with the increased odds of a head injury in those using a baby walker most days (OR 2.47, 95% CI 0.97 to 6.48) found by the same study.

There are several case–control studies that we can compare the findings from our poisonings study with. A study from Greece including children predominantly aged 2–4 years attending hospital emergency clinics following a poisoning and non-injured age-, sex- and hospital-matched controls attending other outpatient clinics found no significant association between mother’s use of safely packaged products, mainly for detergents, and poisoning attendance (OR not presented).¹⁰¹ An Australian study of 1- to 3-year-olds attending an ED following a poisoning and three different types of controls (community, ED attenders with another type of injury and ED non-injured attenders) found that a 1% increase in the percentage of medicinal substances stored in accessible locations in bathrooms increased the odds of poisoning by 3% (OR 1.03, 95% CI 1.002 to 1.080).¹⁰²

A case–control study from Thailand of children aged < 5 years attending hospitals following a poisoning matched with controls on age, sex and area of residence found no association between toxic substances in the home, storage practices, frequency of use, packaging or disposal practices and poisoning (ORs not presented).¹⁰³ A case–control study from Brazil of children aged < 5 years treated in hospital, matched with children seeking emergency care for other reasons on age, sex, hospital and presence in the home of the same toxic substances found in the case home, found an increased odds of poisoning in families storing toxic substances in boxes/cabinets (OR 3.80, 95% CI 1.15 to 12.49) and an increased odds of poisoning in families storing toxic substances < 150 cm from the floor (OR 16.59, 95% CI 2.86 to 96.20).¹⁰⁴

Comparing the findings from these studies with our control findings shows some consistency in terms of our higher odds of a poisoning in families not storing medicines at adult eye level or above and in families not storing poisons safely (at adult eye level or above or locked). The previous studies have not measured associations between putting medicines or household products away immediately and poisoning, and so we cannot compare our findings in this respect.

There are several case–control studies that we can compare our scalds study with. A Greek study of children predominantly aged 0–4 years attending an ED following a burn injury, of which 61% were scalds, matched on age and sex with non-injured ED attenders, found that a 1-unit increase in a burn avoidance index was associated with a 40% reduction in the odds of a burn (OR 0.6, 95% CI 0.5 to 0.8).¹⁰⁵ The burn avoidance index was a composite measure consisting of direction of handles of cooking utensils on the cooker while cooking, use of front/rear hotplates during cooking, keeping hot objects, foods and liquids in places inaccessible to children and avoidance of tablecloths on kitchen tables.

A study from Iraq of children aged 0–5 years admitted to a burns centre following a burn occurring at home, of which 79% were scalds, matched on age and sex to children admitted to hospital for other reasons, found that a 1-unit increase in a burns hazard score increased the odds of a burn by 32% (OR 1.32, 95% CI 1.02 to 1.71).¹⁰⁶ The burns hazard score included use of kerosene cookers, kerosene heaters, samovars for tea, home generators and non-electric heaters for bathwater, not knowing the boiler temperature, storing petrol at home, having no fire extinguisher and having no smoke alarm. A Dutch study of children aged 0–4 years attending an ED with burn injuries, 62% of which were scalds, matched on age with controls, found that storage of hot drinks in their original containers instead of in vacuum flasks increased the odds of a burn (OR 2.0, 90% CI 1.2 to 3.1).¹⁰⁷ A study in Bangladesh of children aged 0–12 years admitted to a burns unit and controls matched on age, sex and area of residence found that significantly more case households than control households had cooking equipment within reach of children (p < 0.001; OR not reported).¹⁰⁸

All of these case–control studies that we can compare our scalds study findings to studied children with burns and, although most of the burns were caused by scalds in these studies, some of the differences between these studies and our findings may reflect differing case definitions. In addition, there were few common exposures between these studies and our study, which may in part reflect differences between heating and cooking practices between countries. Our findings that families who had hot drinks in reach of children, drank hot drinks while holding children or passed hot drinks over children’s heads had an increased odds of a scald are in keeping with the findings from the Greek case–control study.¹⁰⁵ However, as that study used a composite measure of exposure, it is difficult to know the contribution made by keeping hot foods and liquids in inaccessible places to the odds of a scald. Similarly, we did not find significant associations between cooking practices such as use of the front rings of the cooker and turning pan handles towards the back of the cooker and, although these were included in the composite exposure measure in the Greek study, the contribution of these items to the odds of a scald in that study is unknown.¹⁰⁵

How these findings inform other research within the Keeping Children Safe programme

The findings from study A have been used to inform the decision analyses (study K) undertaken in work stream 5 assessing the cost-effectiveness of interventions to prevent falls and poisoning. They have also been used to inform recommendations on scalds, falls and poison prevention practices in the IPB for the prevention of fire-related injuries, scalds, falls and poisoning (study M).