Comorbidities in women with polycystic ovary syndrome: a sibling study

Multiple births and women with a missing link to either biological parent were excluded (n=330, 378). We further excluded those who died or emigrated before the age of 17 or before the start of the follow-up period (n = 45,467), and those who were outside of the follow-up period (1997-2011) (n = 50,396). After excluding those with missing information on the explanatory variables and those without at least one full biological sister, two sub-samples were created. One consisted of women with at least one full sister (Sister sample: n = 333,999) and another with additional restriction to women with information on characteristics measured at birth, retrieved from the MBR, implying only Swedish-born women born between 1973-1980 are included (MBR Sister sample: n = 77,034). The flow chart in Fig. 1 illustrates how the analytical samples were generated.

The follow-up period starts on January 1, 1997, as International Classification of Disease (ICD) 10 codes started being used in the same year in Sweden, with the sample restricted to women observed between the ages of 17 and 45 years. The study population was followed continuously until whichever happened first: turning 45 years of age, death, emigration, or the end of our follow-up period on December 31, 2011. Figure 2 depicts the samples used in different parts of analysis and also illustrates the periods of availability of relevant data in Swedish administrative registers.

Using the NPR, we defined the following potential comorbidities, using the ICD-10 diagnostic codes: obesity (E66), depression (F32 – Depressive episode, and F33 – Recurrent depressive disorder), anxiety (F41 – Other anxiety disorder), and SSE disorders (F51 – Non-organic sleeping disorder, F52 – Sexual dysfunction not caused by organic dysfunction or disease, F50 – Eating disorders). Each condition was defined as a binary variable, indicating whether the individual had been diagnosed at any time during the follow-up period and not necessarily subsequent to the PCOS diagnosis.

PCOS was identified through the NPR [36] at any time during the follow-up period as the binary exposure, including both inpatient and outpatient diagnoses (E28), with the vast majority of the diagnoses coming from outpatient data (99.7%). The aggregated ICD-10 code E28 was used, due to limited access to the more detailed PCOS diagnosis code (ICD-10: E28.2). Women diagnosed with conditions that could cause similar symptoms to PCOS, including Turner syndrome (Q96), Malignant neoplasm of ovary (C56), Suprarenal tumor (C74), Adrenogenital syndrome (E25), Cushing disease (E24) and Pituitary hypersecretion (E22), have been excluded to ensure specificity, similar to previous literature [37]. Both main and contributing diagnoses were considered. In our main analysis, we only included women who were never diagnosed with any of the exclusion criteria, regardless of whether these happened before or after the PCOS diagnosis.

Statistical analyses were performed using STATA/MP 17.0 (StataCorp). The main analyses were restricted to individuals with non-missing information on all covariates included in the models.

In order to address the first aim of the study, we treated the sisters as unrelated individuals and thus used unconditional logistic regression to examine the association between PCOS and each comorbidity condition in the Sister sample, while controlling for previously outlined covariates. Covariates were added to the model group-wise, from controls for year of birth to family background characteristics and adult characteristics. Additionally, the same models were run on a sample that was not restricted to full sisters.

Addressing the second aim, comparing the associations between PCOS and examined comorbidity conditions among full sisters, a FE logistic regression was used (also known as conditional logistic regression). This accounts for unobserved, time-invariant factors coming from the shared family environmental, social or genetic factors that could affect both PCOS and the selected comorbidity.

A potential downside to this approach pertains to it requiring within-sister combination variation in the outcome. Thus, at least one (but not every) sister has to be diagnosed with the examined comorbidity, meaning that only discordant sisters are included in the analysis. This restriction reduced the sample for the examination of all outcomes; obesity (n = 7,994 families, n = 17,732 women), depression (n = 15,232 families, n = 33,709 women), anxiety (n = 12,597 families, n = 27,881 women), SSE disorders (n = 5,129 families, n = 11,351 women).

For each selected comorbidity, we compared the results from unconditional logistic regression models and sister FE models on the Sister and MBR Sister sample, respectively. The FE approach adjusts for unmeasured factors that are shared between the sisters, such as genes which may cause a predisposition to disease or other factors linked to sharing an environment during the upbringing, such as parenting style, and attitudes towards exercise and diet. While we are unable to quantify the relative importance of each of the aforementioned factors, the method allows for obtaining the association after adjusting for shared factors between sisters.

To investigate our third aim, to what extent the association between PCOS and the selected comorbidities changes when controlling for early-life factors such as birthweight, gestational age and one-minute Apgar score, we restricted our sample to full sisters born in Sweden between 1973-1980 and with information on their selected birth characteristics available (MBR Sister sample).

Due to discordancy in the outcome variable and also the fact that each PCOS diagnosed woman needed to have a sister, there is a risk that the sample is selected and thereby yielding results with limited external validity in a larger representative sample. We address this concern by reporting results from models estimated on the sibling FE sample but without aforementioned FE. In addition, we compare these results to the corresponding results of a sample without restrictions to sisters. We argue that this provides a good approximation of the degree to which the sibling FE sample can be used to understand associations in a population of outcome-concordant siblings. Additionally, we estimate random intercept models on the Sister sample and Sister MBR sample. Random intercept models are less restrictive in terms of both study sample and the ability to obtain parameter estimates for independent variables that do not display any variation within sibling combinations. The disadvantage, however, is that the appropriateness of the method requires that there is no correlation between the random effects and the independent variables. We therefore used the Hausman (1978) [39] specification test to detect violations of the random effects approach assumptions.

The diagnostic codes for SSE disorders were combined, due to the generally low prevalence of these conditions (0.64%, 0.40% and 0.64% in the total sample, respectively), and were treated as one binary variable in the main analyses. Therefore, a separate analysis for each of the disorders was conducted.

Descriptive characteristics of the study population are presented in Table 1 for two samples of women that we used in our analyses: one sample with women born in Sweden in 1962-1980 with at least one full sister (Sister sample, n = 333,999), and another sample of women born in Sweden in 1973-1980 with at least one full sister (MBR Sister sample, n = 77,034). Descriptive characteristics of additional sub-samples such as: without restriction to sisters (n = 857,757), concordant sisters without PCOS (n = 326,422), concordant sisters with PCOS (n = 137), and discordant sisters (n = 7,440), are shown in Appendix Table A.1.

From a total of 333,999 women aged 17 to 45 years and followed from 1997-2011, 3,570 were diagnosed with PCOS, of which 14% were diagnosed with obesity, 8% with depression, 7% with anxiety and 4% with SSE disorders during follow-up. The prevalence of the comorbidities among cases of PCOS remained very similar in the MBR Sister sample (Fig. 3).

Estimates of ORs calculated for SSE disorders separately were similar to those with the combined SSE variables (Appendix. Table A.4). When comparing the FE and random intercept models, we see that the random intercept model estimates consistently are slightly smaller in size but yields qualitatively similar results (Appendix. Table A.3.). The results of the Hausman test, which compared the fitting of the fixed and random effects models, rejected the null hypothesis in the Sister sample (Appendix. Table A.3.), confirming that the FE estimators were better fitting our data.

The large, nationally representative Swedish register data allows for analyses with substantial statistical power even for relatively low-prevalence diseases in the registers such as PCOS. The high quality of the Swedish national registers with prospectively collected data [43] and a nationwide coverage also reduce the risk of selection, recall and information bias. However, this type of data source is prone to its own limitations, such as that they do not cover information on health risks or health-seeking behaviors.

Since data used for research purposes are further limited in detail in order to protect personal integrity, the research data extract do not contain the full information from the patient registers. We only had access to the more aggregated three-digit ICD-10 codes from the Swedish NPR, meaning that the E28 Ovarian dysfunction code was used throughout the analysis instead of the more detailed E28.2 Polycystic ovary syndrome code. As previously reported [12], the prevalence of PCOS is lower in the Swedish NPR than it could be expected on a population level. This could have led to measurement error. Due to information sharing between sisters, this may however be less of a concern in families where at least one sister has been diagnosed, which are the ones that are included in our FE analysis. As previously described by March et al. [3], the prevalence of PCOS is dependent on the diagnostic criteria used, finding that it can range from 8.7% up to 17.8%, depending on whether the National Institute of Health or the Rotterdam criteria were used for diagnosis. Since outpatient data is constrained by the inherent limitation of uncertainty in the diagnostic criteria used, many patients may only get diagnosed if a comorbidity condition such as obesity or sub-fertility is also present which they would primarily seek help for. This could also be the reason for the lower prevalence of PCOS in the Swedish NPR.

There was a paradigm shift in PCOS diagnosis during the follow-up period with the introduction of the Rotterdam criteria in 2003 [14] meaning that some of the index women and their sisters had been diagnosed before the shift while others were diagnosed subsequently. Additionally, outpatient specialist care was added to the Swedish NPR from 2001, and before that only inpatient care diagnoses were recorded. This could have caused a left censoring in the analysis as well as a higher concentration of more severe PCOS cases before 2001.