Statistical overview

Quantile regression was used to examine the association between exposure and outcome, including adjusting for covariates identified by using a DAG (direct acyclic graph) (71).

4.4.3 Study III

A population-based cohort study including nulliparous women with spontaneous labor onset, Table 2 shows the study population in detail. The final study cohort comprised 13 379 women.

The main exposure was active first stage labor duration. The association between the exposure (active first stage labor duration) and the outcome was examined using multivariate and multivariable quantile regressions. Differences in the continuous outcome (second stage duration) was presented as both mean, median and at distinct quantiles with 95 % confidence intervals (CI). To check linearity of the associations restricted cubic splines were used to model the relationship between the continuous exposure and outcome. Multinomial logistic regression was used to quantify the odds of operative vaginal delivery (OVD) or cesarean delivery (CD) for each category, Q1-Q4 of active first stage labor duration. Categorization of the exposure Q1-Q4 was done based on the distribution of the exposure, i.e. quantiles for the study population. Maternal age, maternal height, early-pregnancy BMI, birthweight, gestational length, epidural analgesia before start of active first stage of labor, year of delivery and hospital of birth were considered potential confounders using a DAG framework, in the multivariable regression analyses. In contrast, obstetrical management indicators such as oxytocin and epidural, occurring after start of the exposure was considered potential mediators.

4.4.4 Study IV

A population-based cohort study including nulliparous women with spontaneous labor onset, Table 2 shows the study population in detail. The final study cohort comprised 46 040 women. In this study, single imputation was used to impute missing notations for the timepoint of 5 cm cervical dilation based on the target population data presented in Study I(72). Details for the imputation method used are found in the manuscript for Study IV and in Methodological considerations. Our exposure was duration of active first stage of labor, with onset of labor defined as the timepoint of 5 cm dilatation.

Outcome of interest were adverse neonatal outcomes. All identified outcomes were categorized hierarchal into two separate composite neonatal outcomes severe and moderate based on severity grading a) severe; conditions with a high risk of death and/or major neurodevelopmental impairments or b) moderate; conditions with a low risk of death and/or major impairments if adequate treatment is given quickly. Any neonates fulfilling the inclusion in more than one group, were hierarchically included in the severe group. Estimated distribution of the exposure was described at distinct

percentiles, the 5th, 10th, 25th, 50th, 75th, 90th, 95^th percentile in univariate analysis and stratified by outcome of interest. Modified Poisson regression was applied to investigate the association between increasing active first stage of labor duration and the relative risk of either severe or moderate adverse neonatal outcome.

Categorization of the exposure was done based on the its distribution, i.e.

percentiles for the study population, to quantify the relative risk of the outcome in each category (3-4) compared to the set reference group (active first stage of labor duration less than the 50^th percentile).

The approach was modelled on the probability of moderate or severe neonatal outcomes, conditioned on normal neonatal outcomes in the denominator for each category, presented as crude and adjusted relative risk with corresponding 95 % CI.

In multivariable analysis, early pregnancy BMI, maternal age and gestational week (categorized) were selected as covariates using a DAG framework. Birthweight was not adjusted for since gestational week is considered a proxy for birthweight.

Epidural use, oxytocin use and operative delivery were considered as mediators as they lie on the pathway between the exposure and outcomes. Mediation analysis was performed to decompose the total effect of active first stage of labor duration on adverse neonatal outcomes into an indirect effect, mediated by second stage.

Several sensitivity analyses were performed:

✓ To evaluate the robustness of the imputed information on start of the

exposure in the study population complete case analysis, specifically for the exposure variable, was performed.

✓ To verify the categorisation methods we ran the same analysis by four alternative groups with time-based cutoffs instead of quantiles (i.e. 6 hours or less, 6-8 hours, 8-12 hours, and 12 hours and more)

✓ To estimate the potential influence of classifying oxytocin as a mediator and not a confounder we performed additional adjusted analysis including oxytocin as a covariate

5 ETHICAL CONSIDERATIONS

In line with Ethical research principles to prevent from potential harm, all protection measures were taken when handling the data used in these studies (179).

Information in the database is retrieved from the medical record system Obstetrix.

The database is kept at the Clinical epidemiology division at Karolinska Institutet, only accessed by assigned researchers involved in the particular project. For completeness of data on infant outcome, the Stockholm-Gotland Obstetric Cohort has been linked with the Swedish Neonatal Quality Register (SNQ) containing

specified information on infant diagnoses from neonatal care units. All these linkages have been approved by the ethical committee. Data used in all studies included in this thesis contains sensitive information from registers which may have legal unique concerns. Although registry-based research does not include any invasive

procedures or randomizations to a treatment there may a risk that the participants integrity could be violated. Hypothetically, that could be the case if any data is presented at such a detailed level that participants could be identified. There is a strict framework surrounding anonymization of participants before collected data are shared to researchers and how data should be stored (180). Permission for all studies in this thesis were obtained from Swedish Ethical Review Authority (In Swedish: Etikprövningsmyndigheten, Sweden, http://etikprovningsmyndigheten.se).

Swedish Ethical Review Authority was previously (at the time for permission) named Regional Ethical Review board in Stockholm, (In Swedish: Regionala

etikprövningsnämnden Stockholm). In accordance with their decision, we did not obtain informed consent from participants in the study and all research was

performed in accordance with relevant guidelines and regulations. Ethical permits for the studies in this thesis, 2009/275-31, 2012/365-32, 2013/792-32, 2014/177-32, 2014/962-32, 2019-02818.

6 METHODOLOGICAL CONSIDERATIONS

6.1 USAGE OF STATISTICS IN EPIDEMIOLOGICAL RESEARCH

In general, statistical analysis in epidemiology aims to describe the relationship between an exposure and an outcome and in observational research it is used to investigate the magnitude of an association between the two (181). Importantly, statistically significant associations or correlations do not necessarily imply a causal relationship, Hence, prior to describing and discussing the different pros and cons of the methodology for each study in this thesis, a brief introduction to the challenges and strategies used to navigate in the field of epidemiological research in general is essential.

Descriptive studies within epidemiology establish the empirical basis for

understanding populations and for tracking trends and outcomes both within and between different populations. Descriptive studies can be methodologically complex and depend largely on the collected data. Minimizing systematic errors like selection bias, misclassification and confounding bias is a challenge in both observational and descriptive studies. Historically, it has been believed that causal associations may only be possible to confirm in a world where the observation of counterfactual outcomes is possible, i.e. where one study subject is both non-exposed and exposed. This is of course an unrealistic scenario for observational studies in

general, unless in situations where a pragmatic trial concept like “target trials” can be applied (181). Most healthy women in high resource countries like Sweden

experience normal labor outcomes, limiting the numbers of adverse outcomes which further requires large data samples to draw conclusions on rare events.

Physiological processes can be difficult to model since many influential factors change as the process evolves over time. Some factors can be inaccessible or difficult to measure accurately. Some factors can interact with each other, for example the effect of BMI and age present at the same time may be much more than just the sum of their effects. Biological variation, given the same stimulus, can result in variation in response from person to person, which all needs to be

addressed when constructing models.

6.1.1 Exposure and outcome

difficult to classify the exposure status when conducting research, particularly when the exposure is continuous. When possible, it can be necessary to divide continuous exposures into categorical levels to investigate associations. The categorization as well as the continuous exposure can be complicated because the onset of the exposure needs to be decided upon by the researcher, i.e. prior knowledge is

necessary to accurately define the exposure of interest if there is no clear consensus within research on how to define it (182).

The considerations of outcome definitions in epidemiology research enumerates issues of bias that may arise, these are similar those described for the definition of the exposure. To improve validity of the study, it is important to have a clear and objective outcome definition that corresponds to the nature of the hypothesis investigated and the research question of interest but also to have validated the outcomes for the studied population of interest. In general, attention to how the collection of exposure and outcome data is done across previous studies and to use appropriate analytic methods suitable to address any varying definitions is essential to draw any conclusions. Observational studies of any exposure and outcome generally focus on clinical meaningful outcomes. In some situations, for example when addressing a clinical question and the diagnostic accuracy of commonly used ICD (International Classification of Disease) codes within the topic are inaccurate or when the biologic mechanism behind the diagnose is not well understood it could be challenging to adequately capture any exposure or outcome of interest (181, 182).

6.1.1.1 Implications regarding exposure and outcome used in this thesis

For studies evaluating labor duration and progression, the definition of start of active phase has been under re-evaluation and, as previously discussed, no clear

consensus exists of the definition of phases or stages. For the main exposure of this thesis, active first stage labor duration, we used four slightly different definitions of start of the exposure and the same definition for end of the exposure. When the active phase of labor duration was the exposure the defined starting point was 3, 4 or 5 cm cervical dilatation, whatever timepoint came first for each individual. In Study III this was slightly changed and IV this starting point was changed to 5, since study I informed us that 5 cm is more likely the cervical dilatation at which labor progress accelerates more rapidly.

In Study II in the thesis, active first stage of labor duration was the outcome, here we used the previously described definitions to identify the start and end of this period of time.

Total active labor duration was also examined as a continuous exposure of interest, the onset was defined as the time that active labor began, same start as for active first stage of labor and end of the exposure defined as the time of birth in Study II.

The starting point of the continuous outcome in Study III, second stage labor duration was defined as the first timepoint for full cervical dilation for each individual and end was time of birth.

When investigating the effects of labor duration, it is important to consider that some of the outcomes may not be manifest until after a long period of time, such as long-term consequences regarding birth experience, cesarean delivery and neonatal outcomes. Therefore, in studies investigating the effects of labor duration, the study period and length of follow-up has to be adequately specified. In Study IV the

outcome was adverse neonatal outcomes. The selection of outcomes to include and categorization of them involved considerations by the expected frequency of the outcome, sample size and prior research and clinical knowledge of expected long-term consequences. The hypothesis studied was translated into two specific composite neonatal outcomes with clear definitions, described in detail in the overview for Study IV and in detail both in the manuscript for Study IV and under Methodological considerations.

6.1.2 Confounders

Confounders are factors that have the potential to cause spurious results to scientific studies if not properly accounted for. In short, a confounding variable must influence both the independent variable (exposure) and the dependent variable (outcome).

Confounding variables can result in both over- or underestimations of the association between the exposure and the outcome and obscure the true effect. Controlling for confounders can be done by adjusting in multivariable regression analysis,

restriction, matching or stratification. Tests of statistical significance to detect confounding have led to a trend of over-adjustments in observational studies (183, 184). Residual confounding challenges the internal validity of all observational studies, since the confounding factor is unknown or unmeasured and therefore cannot be accounted for (185). In all four studies in this thesis we have restricted our study population to women with spontaneous onsets, term gestations and vertex presentations and used DAGs as a concept for choosing which confounders to adjust for.

6.1.3 Directed Acyclic Graphs

One approach that can be applied for estimating any effects of labor duration on

identified variables, measured and unmeasured and arrows. Its main feature is to identify an optimal analytic approach and inform the choice of which variables to include in a statistical model. Besides representing causal relations, it could also be used to specify statistical associations and help to account for confounding and prevent introduction of spurious association due to over-adjustments. To enumerate for the complex continuous process of labor duration, DAGs became the core to model our research questions for this thesis. In contrast to most seminal

approaches, where each stage of labor has been separately evaluated, the focus in this thesis was to clinically and scientifically incorporate the continuous process of labor.

Constructing DAGs for this process was necessary to differentiate confounders from mediators, since controlling for mediators in regression models (i.e. treating

mediators like confounders) will consequently change the magnitude of the effect size of the relationship between exposures and outcomes (181, 183, 186). Figure 10 illustrates the basic concepts of a DAG. All variables of interest are linked by

different pathways (arrows) to describe their direction and effect. The association of interest is the relationship between X and Y. Prior to estimating the effect of X on Y, it is important to decide which effect that is of interest, the total effect or the direct effect. Confounders (U) are common causes of the exposure and outcome,

mediators (V) are factors that are in the causal pathway, and colliders (Z) are factors that can arise from both exposure and outcome. In the figure, both arrows are

pointing at the collider (Z), i.e. this “back-door path” is closed. Back-door pathways for X on Y via confounders and mediators are open. Statistical analysis for direct effect needs to account for all confounders and mediators, in contrast to analysis for the total effect where this is not necessary. To avoid collider stratification bias, it is of importance to not control for Z.

As an example, in Figure 1, X could indicate maternal BMI, Y could indicate

cesarean delivery, U could then indicate maternal age, a variable confounding this association and V could indicate labor duration, a potential mediator of an increased risk of cesarean delivery in obese women, while Z could indicate postpartum

hemorrhage, i.e. a variable of which there is an increased risk with increasing BMI and cesarean delivery. In this example adjusting for maternal age would be correct while adjusting for postpartum hemorrhage would introduce collider stratification bias. Importantly, not taking labor duration (V) into account will mean that

interpretation will be limited to a “total effect” and not the direct effect of BMI on cesarean delivery.

Figure 10. Example of a basic directed acyclic graph (DAG). X represents the exposure, Y the outcome, the black arrow indicates the direct effect of exposure on outcome. U represents a confounder, pink arrows the alternative pathways for confounders. V represents a mediator and blue arrows the mediated pathway. Z is a collider, indicated by the green arrowed pathway.

6.1.4 Effect modification

Effect modification is present if the strength of an association between an exposure and outcome differs according to the level of another variable and should not be confused with confounding. Effect modification is relative to the exposure and outcome being studied and not for a specific exposure or a specific outcome.

Depending on the research question the same variable can be both a confounder and an effect modifier, or effect modifier and not a confounder and vice versa. A helpful tool to identify potential effect modifiers is to performed stratified analysis, importantly this must not occur before the variable has been evaluated using the

6.1.5 Systematic and random errors overview

Epidemiological studies are sensitivity to both systematic error and random errors, which both can affect the internal validity of a study. Systematic errors, such as selection bias, information bias/misclassification and confounding are not influenced in its magnitude by sample size. Impact of random errors on the other hand

decreases with increasing sample size, p-values and confidence interval are standard methods for describing levels of uncertainty caused by random errors. 95

% of the confidence intervals will contain the true value if the study conducted was repeated and recalculated. Associations cannot rely on the p-value alone, it needs to be interpreted in conjunction with the magnitude of found associations and the

quality of the study design.

6.1.5.1 External validity

External validity refers to the generalizability of the study results and needs to be interpreted in the light of the studied question. For studies on labor duration this question is hard to answer, given that settings and populations differ worldwide. In short, among populations and settings similar to the region of Stockholm-Gotland the external validity for the studies in this thesis is high. When comparing our study data to the publications from prospectively collected randomized controlled trial in Norway (187) it confirms that the data have a high external validity in a Nordic setting.

6.1.5.2 Internal validity

Internal validity is highly dependent on the representativeness of the study

population and high internal validity implies that the studied association is similar to the true association for the target population (sometimes referred to as source population). High internal validity is obtained if findings are unlikely to be explained by systematic or random errors.

6.1.5.3 Selection bias

The consequence of selection bias is that the association between exposure and outcome among those included in the analysis differs from the association among those eligible. Although data from observational cohort studies is generally

prospectively collected, selection bias may occur as a result of a poorly defined study population. For example, including planned cesarean deliveries in cohorts evaluating labor duration will consequently introduce selection bias since these women will have short labor durations and 100% risk of cesarean delivery. Also, since women with long labors/labor dystocia are at increased risk of cesarean delivery at any stage of labor, excluding women with a cesarean delivery during the active phase of labor has been proposed to bias the results towards faster labors.

However, such implications are highly dependent on the frequency of cesarean deliveries in the target population along with sample size and will not always substantially affect the estimates when large data samples are being used. This issue is discussed in several publications on labor progression and duration (11, 29, 72, 79, 188, 189). DAGs can help to identify selection bias on beforehand and

provide knowledge on how to best select a valid study population (181, 183, 184, 186, 190).

6.1.5.4 Information bias

Information bias can occur if the classification of exposures or outcomes is

inaccurately measured or defined and can be either differential or non-differential.

Differential misclassification occurs if the information on the exposure or outcome is systematically misclassified between the comparison groups and this can introduce bias. Equally distributed, non-differential misclassification can lead to

underestimation or overestimation of associations. Spontaneous labor is a

continuous process where the onset most commonly occurs at home. This means that women provide data on the timepoint for start of contractions and rupture of membranes themselves which increases the risk of misclassification. It could be discussed if this is differential or non-differential. It is also possible that the exact transition of phases or cervical change has occurred both before or after the cervical examination recorded in a woman’s chart. For example, below both timepoint a (3-4 cm), b (4-5 cm) and c (4-5 cm) will be recorded as the timepoint for 4 cm.

Figure 11. Illustrative overview over cervical dilation, interpooled measurement could occur any time on the scale

6.2 METHODOLOGICAL CONSIDERATIONS FOR EACH STUDY