Study design
We conducted a registry-based cohort study of 1,543,653 women who were born in Sweden from 1960 to 1990 and resided in Sweden in 1990, according to the Swedish Population and Housing Census. Using the unique personal identification numbers assigned to every resident, we cross-linked all women to the nationwide Causes of Death Register and Migration Register. All women were followed from January 1, 2001, or their 15 birthday (as 96% Swedish women would have had menarche by age 15 [
20]), whichever occurred last, until death, emigration, bilateral oophorectomy, or hysterectomy, their 52nd birthday (the average age of menopause in Sweden [
21]), or December 31, 2012, whichever occurred first. We excluded 71,346 women due to bilateral oophorectomy or hysterectomy (
n=2698), emigration (
n=64,689), or death (
n=3,959) before cohort entry, leaving 1,472,379 women in the cohort (population analysis). The Swedish Multi-Generation Register consists of information on 97% of mothers and 95% of fathers of individuals born after 1932 and alive in 1961 [
22]. By linking to the register, we identified 638,910 women who were full sisters (i.e., sharing the same biological parents) for the sibling analysis.
Ascertainment of premenstrual disorders
Clinical indications of PMD were identified using any primary or secondary clinical diagnosis of PMD from the Patient Register (625E in International Classification of Diseases (ICD) 9th revision (ICD-9) and N943 in ICD-10). The Patient Register has nationwide coverage on inpatient care from 1987 onward and includes information on more than 80% of specialist-based outpatient visits from 2001 onward with high validity (positive predicted value of 85–95% across diseases) [
23]. All PMD diagnoses obtained from the Patient Register are considered provisional, as we could not confirm that all providers of PMD diagnoses ascertained from the Patient Register used two cycles of prospective daily symptom ratings for diagnosing [
3], although it is required in the Swedish healthcare guidelines for PMDD in many regions [
24]. The Patient Register does not cover diagnoses made in primary care, and 49.3% of mental health conditions are treated in primary care in Sweden [
9]. Therefore, we also identified PMD by searching the Prescribed Drug Register for any PMD diagnosis or clear indication of PMD treatment in prescriptions of anti-depressants (the Anatomical Therapeutic Chemical (ATC) code: N06AB, N06AX, N06AA) and oral contraceptives (ATC code: G03A, G02B). We included only prescriptions explicitly made for PMD treatment, as indicated by PMD diagnosis/prescription specific to PMD. Diagnostic codes and keywords indicated in prescriptions are described in supplementary Table S
1. The register collects information on drugs redeemed with a prescription from all pharmacies in Sweden from July 2005 onward, including drugs prescribed in primary care [
25]. We defined the date of PMD diagnosis as either the date of clinical diagnosis or the date that a prescription for PMD medication was filled, whichever came first.
PMD were treated as a time-varying exposure. Women who had not received a diagnosis of PMD during follow-up contributed person-time to the reference group. Women with PMD contributed person-time to the reference group from the start of follow-up until the date of diagnosis and contributed to the PMD group thereafter.
Ascertainment of injury
By linking to the Patient Register and the Causes of Death Register, we identified the first injury resulting in healthcare visit (either as the primary or secondary diagnosis) or death (either as underlying or contributory cause) experienced by participants during follow-up. We also identified first injury event due to suicidal behavior (i.e., completed suicide and suicide attempt), accidents, assaults, and undetermined injury, where available information was insufficient to make a distinction between the aforementioned three subtypes, separately
. Given our a priori hypotheses, we primarily focused on suicidal behavior and accidents in our main analysis. We further sub-grouped accidents into (1) falls, (2) transportation and accidents by other external factors, (3) accidents by natural forces or contact with animals or plants, (4) cutting or piercing, (5) poisoning, and (6) others (supplementary Table S
1). Injuries that occurred before the start of the follow-up were defined as “history of injury” and not treated as outcomes; only the first injury occurring during the follow-up was registered as an outcome event.
The Patient Register covers >90% of hospital discharges nationwide for injury since 1987 and has a high validity (94.8%) for diagnosing injury [
23]. Ascertainment of injury-related deaths in the Causes of Death Register is considered highly accurate [
26] and complete [
27].
Statistical analysis
First, we compared the distributions of demographic characteristics between the reference and PMD groups in both the population analysis and the sibling analysis. The characteristics of PMD patients identified via clinical diagnosis and through treatment indication were evaluated separately. Next, we calculated unadjusted incidence rates (number of events divided by accumulated person-years) of injury in both groups, separately.
In the population analysis, we used Cox proportional hazards regression (attained age as the underlying timescale) to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) of injury among women with PMD, compared to the reference group. We accounted for the relatedness of individuals (i.e., sisters) using robust sandwich estimator of variance in the population analysis [
28]. The proportional hazard (PH) assumption was tested by plotting the Schoenfeld residuals. Additionally, we conducted a sibling analysis using stratified Cox proportional hazards regression. The sibling analysis contrasts the rates within each set of full sister discordant on PMD diagnosis and inherently controls for unmeasured factors (e.g., shared genetic and familial environmental factors) shared between sisters [
29].
We performed separate analyses for the 4 types of injury: (1) suicidal behavior, (2) accidents, (3) assaults, and (4) undetermined injury. We further conducted an analysis of accident subtypes.
We conducted a series of sensitivity analyses to assess the robustness of the PMD classification and assess the potential impact of misclassification. To assess potential differences in outcomes among PMD patients by diagnosis type, we differentiated between PMD identified by (1) clinical diagnosis, (2) treatment indication, and (3) clinical diagnosis and treatment indication. Although all diagnoses in the Patient Register are made by specialists with high validity [
23], the clinical diagnosis of PMD has not been validated. To further test the validity of PMD diagnoses, we restricted the PMD group to patients with at least two consecutive PMD diagnoses appearing at least 28 days apart in the Patient Register. We also conducted a simulation analysis and calculated hazard ratios for injury in testing conditions where 10–80% of PMD diagnoses were false positives [
30]. Finally, because use of antidepressants or hormonal contraceptives may be associated with different risk of suicidal behavior [
31,
32], a sensitivity analysis was conducted by restricting our primary analysis to untreated PMD.
In all analyses, attained age was used as the underlying timescale. We adjusted for calendar year of birth, educational level, region of residence, and history of injury in Model 1. Psychiatric comorbidities were additionally adjusted for in a time-varying manner in Model 2. However, we only employed Model 2 for the analyses subsequent to the primary analysis, whenever applicable. All covariates were categorized (Table
1).
Table 1
Characteristics of women with and without premenstrual disorders (PMD)
Total number | 1,472,310 | 18,628 | 631,236 | 7,674 |
Year of birth |
1960–1964 | 239,959 (16.3) | 4346 (23.3) | 80,813 (12.8) | 1471 (19.2) |
1965–1970 | 255,554 (17.4) | 5080 (27.3) | 111,933 (17.7) | 2142 (27.9) |
1970–1974 | 243,458 (16.5) | 4097 (22.0) | 112,549 (17.8) | 1837 (23.9) |
1975–1979 | 218,052 (14.8) | 2428 (13.0) | 103,626 (16.4) | 1097 (14.3) |
1980–1984 | 216,518 (14.7) | 1628 (8.7) | 106,187 (16.8) | 753 (9.8) |
1985–1990 | 298,769 (20.3) | 1049 (5.6) | 116,128 (18.4) | 374 (4.9) |
Educational level |
Primary | 226,819 (15.4) | 2304 (12.4) | 104,799 (16.6) | 957 (12.5) |
High school | 604,802 (41.1) | 9435 (50.6) | 262,378 (41.6) | 3910 (51.0) |
College and beyond | 379,324 (25.8) | 6005 (32.2) | 166,447 (26.4) | 2508 (32.7) |
Unknown | 261,365 (17.8) | 884 (4.7) | 97,612 (15.5) | 299 (3.9) |
Region of residency |
South | 342,159 (23.2) | 3602 (19.3) | 150,946 (23.9) | 1514 (19.7) |
Middle | 825,202 (56.0) | 11,293 (60.6) | 347,545 (55.1) | 4597 (59.9) |
North | 304,949 (20.7) | 3733 (20.0) | 132,745 (21.0) | 1563 (20.4) |
History of injury |
No | 1,327,241 (90.1) | 14,169 (76.1) | 571,236 (90.5) | 5870 (76.5) |
Yes | 145,069 (9.9) | 4459 (23.9) | 60,000 (9.5) | 1804 (23.5) |
| PYs (%) | PYs (%) | PYs (%) | PYs (%) |
Psychiatric comorbiditiesb |
No | 13,820,488 (95.6) | 83,418 (89.3) | 6,050,657 (95.9) | 34,195 (89.7) |
Yes | 636,709 (4.4) | 10,005 (10.7) | 261,787 (4.1) | 3932 (10.3) |
Because PMD has been associated with multiple psychiatric comorbidities [
33], we repeated our primary analysis by separately adjusting for the number of psychiatric comorbidities and individual type of psychiatric comorbidities (e.g., substance abuse, schizophasia) in a time-varying way. Finally, as previous suicidal behavior is a strong predictor of suicidal behavior [
34], we additionally adjusted for history of suicidal behavior when assessing the association between PMD and suicidal behavior.
To provide insights into the temporal pattern, for each PMD patient, we randomly selected 5 women from the study base who had not yet been diagnosed with PMD by the patient’s index date (i.e., the matching date); reference women were individually matched to PMD patients by birth year (within a 5-year range) and region of residency. We then plotted the cumulative incidence rates and CIs of suicidal behavior and accidents, from the diagnosis/matching date, using survci package in STATA. We further estimated HRs of suicidal behavior and accidents across different time windows, including within 1 year, >1 to 2 years, and >2 years after PMD diagnosis, separately.
A diagnosis of PMD may be delayed for years from symptom onset [
35]. We therefore conducted an additional analysis to estimate the risk of injury preceding diagnosis among PMD patients (i.e., from cohort entry to the date of diagnosis), compared to women without PMD. Due to risk of a delay in PMD diagnosis, we also conducted a sensitivity analysis by excluding person-time preceding the diagnosis of PMD (i.e., from cohort entry to the date of diagnosis).
To shed light on the impact of psychiatric comorbidities, we performed an analysis stratified by psychiatric comorbidities by estimating HRs in the presence and absence of psychiatric comorbidities. To identify potential risk modifiers, we also performed stratified analyses for any injury by calendar year of birth, educational level, region of residency, and history of injury.
Data were prepared in SAS statistical software version 9.4 (SAS Institute, Cary, NC) and analyzed in Stata 15.1 (STATA, College Station, TX). The statistical significance was set at the nominal two-sided 5% level.