Multi-country Tabulated Data or Reports

We identified multi-country indicator tabulations on length of stay from the OECD for 40 countries. We identified two survey programmes, namely MICS and CDC-RHS, that reported relevant data on 21 countries and one country, respectively. The DHS did not report on length of stay, but we used its electronically available datasets to conduct our own analyses for 30 countries [20,23–25]. When countries had multiple sources of estimates, we preferentially used the DHS because we could use the raw data to conduct our own analyses. If there were multiple surveys within a survey programme, we selected the most recent.

Organisation for Economic Co-operation and Development

The OECD is a forum for governments that asks governments to report on, and sets standards for, a wide range of indicators, and analyses and compares data to predict future trends. Its indicators include length of stay following vaginal singleton delivery [20,22]. We searched all OECD tabulations of this indicator and identified the 40 countries reporting data for at least one year since 2000. Where indicators were tabulated for more than one year, we selected the most recent.

Multiple Indicator Cluster Surveys

MICS are standardised, cross-sectional, nationally representative household surveys, covering an average of 11,000 households per country. Recent MICS country reports include the percentage of all facility births (typically in the 2 y before the survey) with a length of stay < 12 h. This is not stratified by mode of delivery or by singleton versus multiple (twin or triplet) status. Some countries also report on the percentage staying 12–23 h, totalling to the percentage staying <24 h. We reviewed all reports from the most recent MICS (2000 to 2015), and identified 21 countries with length-of-stay data that were not also captured via DHS [26]. Many MICS datasets are not in the public domain and so were not accessible to us.

Centers for Disease Control and Prevention Reproductive Health Surveys

The CDC-RHS consists of cross-sectional, nationally representative household surveys largely focused on Latin America and Eastern Europe. Since the late 1980s, questions included in the CDC-RHS are comparable with those in the DHS; however, the CDC-RHS is not provided in a standardised format, and many of its surveys are not in the public domain. Summaries, and in some cases datasets and reports, for all surveys are available on the Global Health Data Exchange (http://ghdx.healthdata.org/series/reproductive-health-survey-rhs). We reviewed the summaries and reports and identified one survey with appropriately tabulated data (although several others had also asked about length of stay).

Demographic and Health Surveys Data

The DHS are standardised, cross-sectional, nationally representative household surveys, usually covering 5,000 to 30,000 households [27]. Of relevance to our analyses, they measure length of stay, household and individual characteristics, child survival, and maternal health care use. We reviewed the most recent DHS (2000 to mid-2013) and identified 30 countries with length-of-stay data (S1 Table). If there were multiple surveys for a country, we selected the most recent.

Length of Stay

Women in MICS who delivered a live birth in a facility in the 2 y before the survey were asked, “You have said that you gave birth in [name or type of facility]. How long did you stay there after the delivery?” Responses were recorded in hours if less than 1 d, days if less than 1 wk, and weeks otherwise.

Women in the CDC-RHS who delivered a live birth in a facility in the 5 y before the survey were asked, “How many nights were you in that place after delivery?” Responses were recorded as number of nights for 0 to 84 nights, “85+” for 85 or more nights, or “88” for “Don’t know” or “Don’t remember”.

Women in the DHS who delivered a live birth in a facility in the 5 y (or 3 y) before the survey were asked, “How long after [name] was delivered did you stay there?” Responses were recorded in hours if less than 1 d, days if less than 1 wk, and weeks otherwise. We converted responses recorded in days or weeks into hours, assuming stays ended at the mid-point of the day or week, respectively (for example, a stay of 1 d could have been 24–47 h, and so was assumed to be 36 h. We saw many responses ranging from 1 to 23 h, which suggests that a response of 1 d was at least 24 h or longer, up to 47 h, after which we assumed women would have reported staying 2 d, although we cannot rule out reporting error. Similarly, a stay of 1 wk was assumed to be 10.5 d, or 252 h. We analysed the responses for the most recent birth in the survey recall period.

Length-of-stay data are frequently right-skewed, with extreme outliers [28]. Our data were right-skewed, and some outliers were so extreme that we judged them to be errors in the unit of reporting (number of days incorrectly reported as number of weeks). We were also concerned that including these values would make it difficult to show the main distribution in the descriptive data on a linear scale and that some extreme outliers would be influential in the linear regression. It is also possible that the factors associated with staying an extreme length of time were different from those associated with an increased or decreased length of stay within more normal limits. In the logistic regression, these extreme outliers were not relevant. Based on our knowledge of guidelines on length of stay, we decided that 3 wk (equivalent to 588 h in our conversion) was a reasonable cutoff for extreme values and excluded observations with lengths of stay greater than 3 wk (584 observations, 0.5% of the unweighted sample). The influence of any remaining skewness was assessed in a sensitivity analysis that log-transformed the data.

Conceptual Framework and Determinants

Conceptual frameworks lay out the key factors, constructs, or variables, and the relationships among them [29]. We developed our framework iteratively, using our technical knowledge and research background, a literature review of related theory and research, and concepts used previously to represent similar problems [30]. We classified and organised the concepts influencing length of stay to emphasise connections between them, connecting the main variables with each other, in a presumed relationship [29,30]. We used the framework to identify the determinants available for analysis within our DHS dataset and to group them into broad categories: need-related characteristics, facility/provider (health system) characteristics, women’s characteristics, and child-related characteristics. The main factors available to analyse, together with sub-categories used, are shown in the tables and S1 Text.

Statistical Analysis

We used Stata (version 14.0; Stata Corporation, College Station, Texas, US) to conduct the analyses. The complex, multi-level cluster survey design used in the DHS, which often oversamples certain areas, was accounted for in the analysis by applying individual women’s survey weights that allowed for the survey design and non-response. These survey weights “weight down” women who are from areas that have been oversampled and “weight up” women from areas that have been undersampled. Similarly, when we present proportion estimates for the entire DHS sample (with countries pooled), the countries that have been oversampled are “weighted down”, and those undersampled “weighted up”, to get estimates that represent the proportion of women across the 30 DHS countries surveyed.

We used violin plots [31] and Kaplan-Meier graphs to explore the distribution of length of stay, and we estimated the proportions of women in each category pooled across countries, weighting according to country populations in 2008. Within each category, we estimated the mean and median length of stay.

Linear regression

We performed a linear regression for each factor in turn using length of stay as the outcome, adjusting only for country. To understand how the rate of cesarean section influenced each of these odds ratios, we repeated the analyses adjusting for mode of delivery. Next, we performed a multivariable linear regression including all factors or potential confounders. Finally, we ran analyses stratified by mode of delivery. Results are expressed as the expected increase (or decrease) in length of stay in hours between the category and the baseline reference category, along with its 95% confidence interval.

Logistic regression

We ran two sets of logistic regression analyses to identify factors associated with lengths of stay that were “too short”, defined as <24 h for vaginal and <72 h for cesarean-section births. The 24-h cutoff was based on a WHO recommendation for uncomplicated vaginal deliveries, but there was less basis for choosing a cesarean-section cutoff. We chose 72 h because the United States deemed stays of 48–72 h for cesarean section problematic and legislated that women be allowed to stay at least 96 h [11,32]. The model-building process was the same for the logistic regression as for the linear regression, and results are expressed as odds ratios comparing each category to the baseline reference category. We could not assess the proportion of women who stayed “too long”, as to do so would have required much better data on need-related characteristics (in terms of information on complications in the mother or newborn) than were available to us.

Each of the variables included in the data analysis was checked for missing observations, and the number of missing values recorded. For the regression analyses, a complete case analysis was performed, excluding all observations that had either the outcome or any of the exposure variables missing. However, if any individual variables had greater than 1% missing data or the total proportion of excluded observations exceeded 5%, then a sensitivity analysis was performed to identify whether excluding these records influenced the estimates obtained.

Syntheses of Results

Across the DHS analyses and models, many of the associations were consistent, and, with few exceptions, the factors associated with longer lengths of stay in the linear regression were associated with smaller proportions of stays that were too short in the logistic regression for both modes of delivery (S2 Table). Cesarean-section mode of delivery was independently and consistently associated with longer lengths of stay, and had the largest effect size. Deliveries with other need-related factors (twins or triplets, lower birthweight, or an index child who died before or on the discharge day) were independently and consistently associated with longer lengths of stay in the linear regression, or with a lower proportion staying too short in the logistic analyses stratified by mode of delivery. Deliveries where infants died after discharge were associated with shorter average stays and more stays that were too short in the crude means and proportions, but the pattern reversed, such that death after discharge was associated with longer stays (although not always statistically significantly), once country was adjusted for. This reversal of the direction of association was because, on average, countries with higher infant mortality—and hence higher post-discharge mortality—had shorter lengths of stay.

Facility/Provider-Related Characteristics

Doctor-attended deliveries were associated with longer lengths of stay and a lower proportion staying too short in all analyses. The subset of women whose infants were delivered by cesarean section are expected to have been attended largely, if not exclusively, by doctors, so we excluded birth attendant from this model. Women attended by non-skilled staff consistently had shorter lengths of stay than those attended by nurse-midwives. In the crude analyses, it appeared that women who delivered in private-sector facilities stayed longer than those who delivered in the public sector did. However, once we adjusted for cesarean mode of delivery or stratified by it, women in the private sector stayed shorter. The apparent association with longer stay in the crude analyses was due to a greater proportion of cesarean sections in the private sector. In the stratified analyses, the association of length of stay with sector was not significant among vaginal births, while among cesarean-section births, private-sector births were 50% more likely to be too short (<72 h) than public-sector births.

Women’s Characteristics

Women’s characteristics independently and consistently associated with longer lengths of stay, or with a lower proportion staying too short, included older age and not being currently married. In the crude analyses of mean length of stay and proportion of women staying too short, the older the women were, the more likely they were to have a shorter or too short stay. Once country was adjusted for, this trend reversed, so older age became associated with longer stays, because countries with more births to older mothers usually had shorter lengths of stay. The same reversal in the direction of the association also occurred for women who were not currently married, although the association for not currently married women was not statistically significantly different from that of currently married women in the cesarean-section model. Having higher education was independently and consistently associated with a lower proportion staying too short in all analyses except those that accounted for cesarean-section delivery via adjustment (linear regression) or stratification (logistic regression). Urban residence also increased length of stay, but this crude association operated largely through higher cesarean-section rates in urban areas. Once mode of delivery was adjusted for, residence ceased to have a statistically significant association with length of stay. In the analysis stratified by mode of delivery, women with a cesarean-section delivery who lived in rural areas were less likely to stay too short than women who lived in urban areas.

In the crude analyses of mean length of stay and proportion staying too short, the wealthier the quintile, the longer women appeared to stay. However, once the linear regression was adjusted for cesarean section or stratified by mode of delivery, the pattern reversed and wealthier women stayed shorter. In the stratified analysis of women with cesarean-section deliveries, the odds of staying too short were 27% higher among the richest compared to the poorest quintile of women.

Child-Related Characteristics

Women delivering infants of lower birth order or male infants stayed longer and were less likely to stay too short compared to women delivering infants of higher birth order or female infants. The association of birth order with length of stay was not significant in either of the stratified logistic analyses. The association of giving birth to a male child with length of stay was small and not significant in either stratified logistic regression. Women whose newborns were wanted stayed longer (and were less likely to stay too short) than those whose pregnancies were mistimed or unwanted in crude analyses. In the adjusted linear regressions, women with mistimed or unwanted pregnancies stayed slightly longer. Wantedness was not a significant factor associated with the proportion staying too short in the adjusted logistic regressions stratified by mode of delivery.

Limitations

Using existing tabulated data meant we relied on definitions used by others; further, in our secondary data analyses, we were unable to analyse all the potential determinants elaborated in the conceptual framework, particularly those related to medical need (of mother or newborn), facility features, providers, or the health system. The strengths of our DHS analyses were that they used nationally representative, comparable data for many countries in which, to our knowledge, length of stay after childbirth had not been studied previously. However, our DHS analyses also have limitations. Responses relied on women’s self-reports of their most recent live birth, recalled for up to 5 y, and stillbirths were excluded. Similarly, women who died, albeit a small percentage, were also excluded. The response option grouping of lengths of stay above 6 d into weeks resulted in heaping, because, for example, women staying between 7 to 13 d would have all been reported as staying 1 wk; our analysis considered them as staying 10.5 d (the mid-point of 7 to 13 d). This approach may have exaggerated the mean length of stay for cesarean sections or for countries at the longer end of the distribution. Also, some women identified non-doctor/non-clinical-officer individuals as their birth attendant for a cesarean section, an anomaly noted previously [33]. We are uncertain whether women misclassified their delivery provider or their mode of delivery, although the latter is thought to be well understood by women [34].

Until now, the main data available on mean length of stay were for vaginal singleton births compiled for 40 middle- and high-income countries by the OECD [20]. We conducted new analyses with 30 DHS datasets and compiled reports on a further 22 countries from MICS (21 countries) and the CDC-RHS (one country). Between-country variability was extremely large; for example, women with vaginal singleton deliveries stayed 12 times longer on average in the Ukraine than in Egypt. Country remained strongly associated with length of stay even after adjusting for many other variables, suggesting that national norms and health system features are important, beyond variations in case mix and consequent need. Norms that stipulate a minimum stay, for example, the 7 d required in the former Soviet Union [35], may explain the longer lengths of stays.

The more worrying finding was that in some countries many women left almost immediately, well before the 24 h recommended (the proportion staying <6 h ranged from 0.2% in Ukraine to 43.1% in Pakistan). Very short lengths of stay partially obviate the purpose of facility delivery, and seem perverse when so many initiatives seek to encourage postnatal checking of newborns, including by having health workers visit women’s homes up to three times in the first 10 d [36]. While some countries might mitigate the potential adverse consequences of early facility discharge through routine postnatal home visits by midwives after discharge [37,38], many do not. Countries with substantial proportions staying too short should seek to understand whether women leave so soon because of overcrowding, insufficient nursing support, or poor physical environments. Irrespective of reason, wide variations in length of stay are suggestive of potentially poor-quality care, possibly reinforced by a weak scientific evidence base on the optimal length of stay.

Nearly all the variables we explored were associated with length of stay in the expected direction, and the indicators of need or of complication had the largest associations, with cesarean-section mode of delivery having the single largest association. Women with a cesarean-section delivery residing in urban areas were more likely to stay too short compared to women residing in rural areas, possibly because those in remote areas received cesarean section for more severe indications and needed longer to recover. We were interested by the finding that the richest women stayed shorter than the poorest women did, because we expected women with greater financial resources to have fewer resource constraints and more options for attaining higher-quality care. The reasons wealthier women stayed shorter might include that they were healthier, operated with more self-efficacy, or were treated more efficiently by facility staff. For example, in some settings, women are detained until they pay facility fees [39]. Women with greater economic resources may have found arranging transport back home easier. Alternatively, they may have delivered in larger, higher-volume facilities that discharged women quickly to reduce overcrowding. Finally, our conceptual framework suggested that the balance of attractiveness between the home and facility environments may be a driver, and for the richest, the balance may favour the home environment [40].