Effect of pregnancy on HIV disease progression and survival among women in rural Uganda

Authors


Corresponding Author Lieve Van der Paal, MRC/UVRI Uganda Research Unit on AIDS, PO Box 49, Entebbe, Uganda. Tel.: +256 41 320042; Fax: +256 41 321137; E-mail: lieve.vanderpaal@mrcuganda.org

Summary

Objective  To investigate the effect of pregnancy on HIV disease progression and survival among HIV-infected women in rural Uganda, prior to the introduction of anti-retroviral therapy (ART).

Methods  From a clinical cohort established in 1990, we selected records from HIV-infected women of reproductive age. We conducted two analyses: (1) all HIV-infected cases contributing to analysis of CD4 decline, using a linear regression model with random intercepts and slopes; (b) incident cases with known date of seroconversion contributed to analyses of median time to CD4 <200 cells/μl, AIDS and death.

Results  A total of 139 women were included in the analysis of CD4 decline. Women who subsequently became pregnant had higher CD4 counts at enrolment and had a slower CD4 decline than those who did not become pregnant. In women who became pregnant, CD4 decline was faster after pregnancy than before (P < 0.0001). The survival analyses showed no significant differences between women who became pregnant and those who did not with respect to median time to CD4 count <200, AIDS or death.

Conclusions  The initial comparative immunological advantage possessed by fertile women before they become pregnant is subsequently lost as a result of their pregnancy. Women should be informed about the potential negative effect of pregnancy on their immunological status and should be offered contraception. In resource-limited settings, women determined to become pregnant should be given priority for ART if eligible.

Abstract

Objectif  Investiguer l'effet de la grossesse sur la progression de la maladie VIH et la survie chez les femmes infectées par le VIH en zone rurale en Ouganda avant l'introduction de la thérapie antirétrovirale.

Méthodes  Á partir d'une cohorte clinique établie en 1990, nous avons choisi des données sur des femmes infectées en âge de reproduction par le VIH. Nous avons effectué deux analyses: (1) tous les cas infectés par le VIH contribuant à l'analyse du déclin des CD4, en utilisant un modèle de régression linéaire avec des interceptions et des pentes aléatoires, (2) les cas d'incident avec date de la séroconversion connue ont contribués aux analyses du temps médian pour atteindre un taux de cellules CD4 <200/μl, le SIDA et la mort.

Résultats  Un total de 139 femmes ont été inclues dans l'analyse du déclin des CD4. Les femmes qui sont tombées enceintes par la suite avaient des taux de CD4 plus élevés au recrutement et ont eu un déclin des CD4 plus lent que celles qui ne sont pas tombées enceintes. Chez les femmes qui sont tombées enceintes, le déclin des CD4 était plus rapide après la grossesse qu'avant (P < 0,0001). Les analyses de survie n'ont montré aucune différence significative entre les femmes qui sont tombées enceintes et celles qui n'en sont pas, en ce qui concerne le temps médian pour atteindre un taux de CD4 <200/μl, le SIDA ou la mort.

Conclusions  L'avantage immunologique comparatif initial que possèdent les femmes fertiles avant qu'elles ne tombent enceintes est perdu plus tard à cause de leur grossesse. Les femmes devraient être informées au sujet de l'effet potentiellement négatif de la grossesse sur leur statut immunologique et devraient se voir proposées la contraception. Dans les endroits à ressources limitées, les femmes déterminées à tomber enceintes devraient recevoir en priorité la thérapie antirétrovirale si elles en sont éligibles.

Abstract

Objetivo  Investigar el efecto del embarazo sobre la progresión del VIH/SIDA y la supervivencia entre mujeres infectadas con VIH en Uganda rural, antes de la introducción de la terapia antirretroviral.

Métodos  Se seleccionaron historias clínicas de mujeres en edad reproductiva infectadas con VIH, de una cohorte clínica establecida en 1990. Se realizaron dos análisis: (1) todos los casos de seropositivas contribuían al análisis de una disminución de CD4, utilizando un modelo de regresión lineal con intercepciones y pendientes al azar; (2) los casos incidentes con una fecha de seroconversión conocida contribuyó al análisis del tiempo medio para alcanzar un conteo de CD4 <200 células/μl, SIDA y muerte.

Resultados  En el análisis se incluyeron 139 mujeres con una disminución de CD4. Las mujeres que posteriormente quedaron embarazadas, tenían un mayor conteo de CD4 en el momento de ingresar al estudio y tuvieron una disminución de CD4 más lenta que aquellas que no quedaron embarazadas. En las mujeres que se quedaron embarazadas, la disminución de CD4 fue más rápida después del embarazo que antes del mismo (P < 0.0001). El análisis de supervivencia mostró que no había diferencias significativas entre las mujeres que se quedaban embarazadas y aquellas que no, con respecto al tiempo medio para alcanzar un conteo de CD4 <200 células/μl, SIDA y muerte.

Conclusiones  La ventaja comparativa inicial inmunológica de mujeres fértiles antes de quedarse embarazadas se pierde como resultado del embarazo. Las mujeres deberían ser informadas sobre el potencial efecto negativo del embarazo sobre su estatus inmunológico y debería ofrecérseles algún método de anticoncepción. En lugares con recursos limitados, a las mujeres que están decididas a quedarse embarazadas se les debería priorizar para tratamiento antirretroviral si son elegibles.

Introduction

There is conflicting evidence in the literature on the effect of pregnancy on disease progression and survival in HIV-infected women. A literature review and meta-analysis of seven prospective cohort studies found that in a developing country setting, HIV disease progressed more rapidly and death occurred earlier among pregnant women than among women who did not become pregnant (French & Brocklehurst 1998). A study in Haiti (Deschamps et al. 1993) showed a trend towards earlier manifestation of HIV-related symptoms among pregnant women, but no significant difference in rate of progression to AIDS or death. Studies done in industrialised countries did not confirm an effect of pregnancy on HIV progression. There was no adverse effect of pregnancy on immunological markers found in a cohort in Scotland (Brettle et al. 1995) and no association of pregnancy with accelerated disease progression in an American cohort (Bessinger et al. 1998). European cohort studies with known dates of seroconversion did not show an effect of pregnancy on CD4 decline (Van Benthem et al. 2002) or a faster progression from seroconversion to AIDS associated with pregnancy(Alliegro et al. 1997; Saada et al. 2000). In a Swiss cohort, the rate of any AIDS-defining event was higher in pregnant women, but a statistically significant difference only emerged for recurrent bacterial pneumonia (Weisser et al. 1998).

Data from sub-Saharan Africa on whether pregnancy accelerates disease progression are currently lacking. Cohorts of HIV-infected patients with known dates of seroconversion and with regular clinical and laboratory follow-up are rare in this setting. Information on the effect of pregnancy on HIV progression is important when counselling HIV-infected women about pregnancy, especially in sub-Saharan Africa where access to anti-retroviral therapy (ART) is, unfortunately, still limited.

Methodological difficulties pose a challenge to answering this very relevant research question. Pregnancy cannot be randomised, and women who become pregnant are different from women who do not: a ‘healthy pregnant woman’ effect has been described in Senegal (Ronsmans et al. 2001), Bangladesh (Khlat & Ronsmans 2000) and industrialised countries (Gissler et al. 1997; Jocums et al. 1998).

Methods

Study population

A clinical cohort was established in 1990 in rural S-W Uganda. Participants are adults aged 15 or more, identified from a large general population cohort with annual serosurveys (Mulder et al. 1994). The clinical cohort consists of a selection of HIV prevalent cases identified in 1990, HIV incident cases with known dates of seroconversion identified during subsequent survey rounds and age-and sex-stratified randomly selected negative controls. Date of seroconversion for the incident cases is estimated as mid-point date between first positive and last negative HIV antibody test, with a maximum of 4-year interval between test dates. The cohort has been described in detail elsewhere (Morgan et al. 1997). Briefly, participants give informed written consent before enrolment in the study. They are seen routinely every 3 months by one of two study physicians, who take a medical and sexual history and perform a clinical examination. Questions are also asked about date of last menses, reason for any amenorrhea, outcome of pregnancy; and the clinical examination includes assessment of fundus uteri. Blood, urine and stool specimens are collected for laboratory investigations at these routine visits. The laboratory tests do not include routine pregnancy testing. Participants are staged at each routine visit using the WHO staging system (World Health Organisation 1990). Study subjects are also seen by the clinicians at other times when ill. They receive treatment for opportunistic infections and secondary prophylaxis where indicated. Participants are encouraged to learn their HIV status from voluntary counselling and testing services provided in the villages where they live. However, few cohort participants (about 25% in 2003) know their status. During the study period, no primary prophylactic regimens or ART were given. ART for eligible patients was introduced in January 2004.

The cohort study has received ethical approval from the Ethics Committee of the Uganda Virus Research Institute and the Ugandan National Council for Science and Technology.

Laboratory methods

The HIV testing is done using two ELISA antibody tests confirmed with Western Blot in case of discordant or first time positive ELISAs (Nunn et al. 1993). Since 1996, CD4 lymphocyte counts have been performed every 6 months using a FACScount system (Becton Dickinson, San Jose, CA, USA). Until 1996, samples were tested by flow-cytometry in an external laboratory in Kampala.

Statistical methods

Study participants for the analysis of the effect of pregnancy on HIV progression were HIV-infected women of reproductive age (15 to 49 years old) enrolled in the cohort between October 1990 and December 2003. Data were collected on standardised forms by the clinicians, double entered and checked for consistency using Epi-Info version 6.0 (Centers for Disease Control, Atlanta) until 1998, and Fox-Pro for Windows version 2.6 (Microsoft Corporation, USA) since 1998. The exposure of interest for both the linear regression analysis and the survival analysis was pregnancy, whether ending with the delivery of a live baby, a stillbirth or a miscarriage.

Linear regression analysis of CD4 decline

We graphed CD4 decline over time by a regression predicting CD4 count by time since first CD4 count, in order to describe the variability of CD4 decline between individual women. HIV incident and prevalent cases with at least one CD4 measurement contributed to the analysis of the effect of pregnancy on CD4 decline. Although they contributed only little information, the six women with only one measurement were included in the analysis. This follows accepted multilevel modelling practice (Crystal & Sambamoorthi 1996; Snijders & Bosker 1999). Because there were no CD4 count measurements in this cohort before August 1992, the study period for the analysis of CD4 decline was restricted to the period from August 1992 to December 2003. We conducted a multilevel analysis to account for multiple CD4 lymphocyte measurements in each woman using MlwiN version 2.1 (Center for Multilevel Modelling, Institute of Education, University of London, London, UK). This was a linear regression analysis of CD4 count against time since first CD4 count and against time since first delivery. We used time since first CD4 count, rather than time since seroconversion, to enable us to include prevalent cases for whom we did not have a date of seroconversion. We accounted for variability between individual women by allowing the time variable to have random intercepts and slopes. The random intercepts and slopes also allowed us to account for variation in disease stage because it incorporated a correlation between first CD4 count (intercept) and CD4 decline (slope). We additionally controlled for age at first CD4 count, whether or not the woman was currently pregnant, whether or not the woman was within 3 months postpartum, and whether or not the woman had more than one pregnancy during the follow-up period. We did not control for frequency of sexual activity or use of contraception. We used a square root transformation of CD4 measurements in order to normalise their distribution.

We developed different regression models to describe CD4 decline in the study population (Appendix 1). These regression models were built up stepwise to investigate different situations with respect to the occurrence of pregnancy. The first model describes CD4 count decline over time since the first CD4 count. The second model incorporates the interaction between whether the woman was ever pregnant during follow-up and time since first CD4 count in order to explore whether women who became pregnant had a CD4 decline before pregnancy that differed from that among women who were never pregnant. The third model incorporated the effect of the pregnancy on CD4 decline by including time since first delivery.

The results of the linear regression model were then used to graphically present the predicted CD4 decline for the three different scenarios encountered in this cohort. These were: (1) the period before pregnancy for women who became pregnant, (2) the post-pregnancy period for women who became pregnant and (3) the entire period of follow-up for women who never became pregnant. To facilitate graphical presentation of the results, we arbitrarily assumed a woman who started with a CD4 of 900 cells/μl for each of the three scenarios.

Survival analysis

All HIV incident cases contributed to analyses of median time to CD4 < 200 cells/μl, to AIDS (defined as WHO clinical stage 4) and to death. We applied the Cox Proportional Hazards method using Stata version 8 (Stata Corporation, College Station, TX, USA). The semi-parametric Cox regression is often used when comparing hazards of survival from HIV seroconversion to death, among two or more groups (Brahmbhatt et al. 2006; Lavreys et al. 2006). Hazard ratios were adjusted for age at seroconversion because older age is associated with decreased fertility and with faster disease progression. Women were grouped according to whether during follow-up they had zero pregnancies since HIV seroconversion, one pregnancy, or more than one pregnancy.

Results

Between October 1990 and December 2003, we followed 109 HIV-infected women of reproductive age with known date of seroconversion and who had at least one CD4 lymphocyte measurement (seroincident cases). Additionally, there were 30 HIV-infected women with an unknown date of seroconversion (but before 1990). Characteristics of both HIV sero-incident and -prevalent women are shown in Table 1. Median age at seroconversion was 23.7 years, and the median interval between seroconversion and enrolment in the cohort was 11.6 months. Median age at enrolment for the prevalent cases was 28.2 years.

Table 1.   Characteristics of study participants
 Incident casesPrevalent cases
Age at first follow-up
 15–2455 (50.5%)8 (26.7%)
 25–3435 (32.1%)13 (43.3%)
 35–4414 (12.8%)7 (23.3%)
 45–495 (4.6%)2 (6.7%)
 Total10930
Median age at seroconversion (IQR)24 (19, 32)Not known
Median age at enrolment (IQR)24 (21, 33)28 (21, 37)
Median number CD4 counts performed (IQR)8 (3, 16)9 (3, 15.25)
Median first available CD4 (IQR)616 (453, 919)615 (380, 800)
Pregnancies during study period [N women (%)]
 No pregnancy58 (53.2%)17 (56.7%)
 One pregnancy22 (20.2%)3 (10.0%)
 Two pregnancies15 (13.8%)7 (23.3%)
 Three or more   pregnancies14 (12.8%)3 (10.0%)
Number of pregnancies10128
 Live birth78 (77.2%)24 (85.7%)
 Stillbirth5 (5.0%)0 (0.0%)
 Miscarriage18 (17.8%)4 (14.3%)

More than half of the women did not become pregnant during follow-up. On the contrary, 28% of women had more than one pregnancy during follow-up.

CD4 decline over time

There was substantial variability in CD4 decline between individual women. This is shown in Figure 1. The bold line represents the CD4 count as obtained by the regression analysis of individual data, while the light grey lines are the observed CD4 counts over time for each of the 139 women who contributed to this analysis. The variability across women in both intercepts and slopes of CD4 decline necessitated the multilevel modelling approach in analysis that we took in assessing the impact of pregnancy on CD4 decline.

Figure 1.

 CD4 decline over time actual (grey lines) and regression predicted (bold line).

The regression analyses of the decline in CD4 counts for the period before pregnancy (line A), for the post-pregnancy period (line B), and for women who never became pregnant (line C) are shown in Figure 2. Women who were able to become pregnant had slower CD4 decline before their pregnancy than women who never got pregnant (line A vs. line C; P = 0.047, log-likelihood deviance test). Women who became pregnant had a transient decline in CD4 count during pregnancy, and a steeper CD4 decline after the end of pregnancy than before (line B vs. line A; P < 0.0001, log-likelihood deviance test). The CD4 decline for women who had been pregnant approached that of women who never became pregnant (lines B and C), indicating that the initial comparative immunological advantage of women who were in sufficiently good health to become pregnant was lost as a result of pregnancy.

Figure 2.

 CD4 decline women with vs. without pregnancy. To facilitate graphical presentation of results of the linear regression model, we arbitrarily assumed a woman who started with a CD4 count of 900 cells/μl for each of the three scenarios.

Survival analyses

No significant differences were seen in median time to CD4 count <200 cells/μl, AIDS or death between women with zero, one or more than one pregnancy during follow-up (Table 2). However, for all three endpoints, women with more than one pregnancy appeared to have a slower clinical progression than the other groups.

Table 2.   Survival analysis for time to death, AIDS (WHO stage 4) and CD4 < 200 cells/μl
 Not pregnantOne pregnancy>1 Pregnancies
  1. Pyo, person-years-of-observation.

  2. †Hazard ratio adjusted for age at seroconversion.

CD4 < 200 cells/μl
 Rate8.4/100 pyo12.2/100 pyo7.0/100 pyo
 Median time to   CD4 < 200 cells/μl6.0 years4.9 years8.9 years
 Hazard ratio† (95% CI)11.67 (0.77–3.63)0.63 (0.30–1.33)
 P-value 0.1940.222
AIDS
 Rate6.2/100 pyo6.7/100 pyo3.8/100 pyo
 Median time to AIDS8.2 years10.4 years11.7 years
 Hazard ratio† (95% CI)11.16 (0.51–2.65)0.71 (0.30–1.68)
 P-value 0.7200.439
Death
 Mortality rate7.0/100 pyo9.7/100 pyo3.6/100 pyo
 Median time to death6.8 years7.8 years11.8 years
 Hazard ratio† (95% CI)11.78 (0.84–3.78)0.51 (0.22–1.18)
 P-value 0.1350.114

Discussion

Through our multilevel model analyses, we were able to examine the effect of pregnancy among HIV-infected women in different ways. In our cohort, we were able to follow women who became pregnant five or more years after enrolment. We therefore were able to compare disease progression before pregnancy with disease progression after pregnancy in the same women and also to compare those women who became pregnant with women who never became pregnant.

These analyses show that pregnancy in HIV infected women is not a random event: women who get pregnant are different from those who do not get pregnant. Specifically, women who subsequently became pregnant had slower CD4 decline than those who never became pregnant. Additionally, in those women who became pregnant CD4 decline was faster after pregnancy than before, indicating that pregnancy has a detrimental effect on immune status.

We conducted the same analysis of the effect of pregnancy on CD4 decline, in which we only included cases with more than one CD4 measurement (data not shown). The results were essentially the same. When the same model was fitted against time since seroconversion, only including the incident cases with known date of seroconversion, we observed similar trends, but they were not statistically significant resulting from the smaller number of women and therefore reduced power (data not shown). We chose to use CD4 decline as a surrogate measure for HIV disease progression, as this has been shown in different contexts to be a reliable predictor of how fast HIV-infected individuals progress to AIDS (Gardner et al. 1992; Alcabes et al. 1994; Wannamethee et al. 1998).

Other studies have documented a decrease in CD4 lymphocytes during pregnancy in both HIV-infected and -uninfected women (Rich et al. 1995) and we found the same in our study participants, although there seemed to be a temporary increase in CD4 lymphocytes in the postpartum period. This may be explained by haemodilution associated with pregnancy.

This analysis of CD4 decline helps explain why the results of our survival analyses in Table 2 seem to be ambiguous. While women who became pregnant were initially healthier than women who did not become pregnant (as indicated by a less steep CD4 decline), they had a faster CD4 decline after pregnancy. These two phenomena work in opposite directions, leaving no overall effect on clinical progression. Similar unrevealing findings on survival analysis have been found in other studies (Bessinger et al. 1998). A ‘healthy pregnant woman effect’ has been described in a study in Senegal (Ronsmans et al. 2001), where it was demonstrated that pregnancy does not confer additional risks to women between 20 and 44 years of age; after excluding direct obstetric deaths, women between 20 and 39 in this study, who did become pregnant, had a lower mortality risk than women who did not. A study from the United States, comparing immunological and clinical progression in HIV-positive women after one index pregnancy with women who had a second pregnancy after the index pregnancy, found that there was almost no difference in CD4 decline between both groups, but women with a second pregnancy tended to survive longer (Minkoff et al. 2003). The women who became pregnant in our study, and in particular, those women who had more than one pregnancy after seroconversion were most likely healthier at the beginning of the study and had a slower disease progression; they were thus more likely to be sexually active and to become pregnant.

Although the analysis of CD4 decline presents strong evidence that one pregnancy is detrimental, the survival analyses suggest that more than one pregnancy is protective (P-value non-significant). The most likely explanation for this apparent paradox is the temporal explanation: a woman who lives longer will have more time and opportunity to become pregnant more than once.

Contraceptive use was very limited among these women, because the cultural pressure to have children is very high in rural Uganda as in many parts of sub-Saharan Africa. Only between 10% and 20% of women in this study were aware of their HIV status, and other studies in Uganda and neighbouring countries have shown that knowledge of HIV status does not affect contraceptive behaviour (Ryder et al. 1991; Allen et al. 1993) or pregnancy rates (Gray et al. 1998).

We did not adjust the analysis for potential confounders like parity at enrolment and breastfeeding. Breastfeeding after a successful pregnancy was universal among these women living in a rural area, both for cultural and economical reasons. Since the introduction of a prevention of mother-to-child transmission programme in the study area in 2002, only very few HIV-positive women (less than 5%) in the study area have opted to not breastfeed their newborn baby.

As no pregnancy testing was done during routine visits, we are likely to have missed most early miscarriages, and could consequently have underestimated the number of pregnancies among the study participants.

We have shown that pregnancy on average has a detrimental effect on HIV progression. Pregnancy requires unprotected sexual intercourse; therefore, for HIV-infected women it is associated with other risks, including the possibility of HIV transmission to the sexual partner and vertical transmission to the baby. For these reasons, women with HIV infection should be advised to avoid unprotected sexual intercourse and pregnancy. However, some women want to become pregnant even after careful counselling about all related risks. Such women should be informed that it is likely to be safer to have a pregnancy in the early clinical stage of HIV infection, when CD4 counts are still relatively high, so that the steeper CD4 decline after pregnancy is less likely to result in progression to clinically manifest disease and death.

In resource poor settings with restricted availability of ART, women who are pregnant and eligible for ART should be offered such treatment as a priority group, as they are at high risk for fast progression and as ART will at the same time protect the baby from HIV infection. It might be advisable for women on ART with low CD4 counts and who are determined to have children, to postpone a pregnancy until CD4 counts have increased.

Appendices

Appendix 1: Multilevel models for CD4 count decline

Table A1.   Comparison of CD4 decline among women who: (a) never get pregnant, (b) before pregnancy, and (c) after pregnancy. Outcome: square root (CD4) (1312 CD4 count measurements on 139 women)
Fixed effectCoefficient (SE)
Model 11Model 22Model 33
  1. 1Base model, indicating that CD4 generally declines over time for all women. This seemingly obvious model is required in order to obtain the log-likelihood deviance test for model 2.

  2. 2This model shows that women who become pregnant have a slower CD4 decline before pregnancy than women who never become pregnant.

  3. 3This model shows that women who become pregnant have a steeper CD4 decline after pregnancy than before pregnancy.

  4. 4In order to account for the temporary postpartum increase in CD4, this is time since 3 months after the first delivery.

  5. 5Variance component.

  6. 6Var(U0j)… Var(U3j) represent the intercept and slope variances. Including these variances in the model allow the decline in CD4 to vary across women.

  7. 7Cov(U0j,U1j)… Cov(U2j,U3j) represent the covariances of intercepts and slopes across women. For example, Cov(U0j,U1j) is the covariance between the intercept and the decline in CD4 since first CD4. This covariance is positive and statistically significant, meaning that there is a positive correlation between initial CD4 and CD4 decline; that is, the higher the initial CD4, the slower the CD4 decline over time.

Intercept26.4075 (2.5893)26.1660 (2.5863)26.1178 (2.5559)
Time since first CD4−0.2148 (0.0550)−0.3115 (0.0725)−0.2408 (0.0746)
 Square(time since first CD4)0.00121 (0.00045)0.0012 (0.0004)−0.0007 (0.0004)
 Cube(time since first CD4)−0.000014 (0.000003)−0.000013 (0.000003)−0.0000006 (0.000004)
Age at first follow up−0.0718 (0.0795)−0.0674 (0.0793)−0.0749 (0.0784)
(ageff) × (time since first CD4)−0.0003 (0.0018)0.0016 (0.0020)0.0014 (0.0020)
Currently pregnant (yes/no)−1.5743 (0.2593)−1.5419 (0.2598)−1.6832 (0.2583)
Three months postpartum pregnancy (yes/no)0.5010 (0.3904)0.5061 (0.3905)0.3986 (0.3912)
More than one pregnancy (yes/no)2.0407 (1.8161)1.9195 (1.8009)2.6135 (1.7806)
Ever pregnant (yes/no)0.4171 (1.6524)0.7707 (1.6508)0.7018 (1.6307)
(ever pregnant) × (time since 1st CD4) 0.0682 (0.0335)0.0866 (0.0432)
Time since first delivery4  −0.1229 (0.0537)
 Square(time since delivery)  0.0015 (0.0008)
 Cube(time since delivery)  −0.000026 (0.000006)
Random effectVar comp5. (SE)Var comp. (SE)Var comp. (SE)
Level-two random effects
 Var(U0j) – intercept648.4002 (6.2045)48.1395 (6.1739)46.4620 (5.9725)
 Var(U1j) – time since first CD460.0205 (0.0045)0.0207 (0.0045)0.0186 (0.0054)
 Var(U2j) – square(time since CD4)60.000003 (0.000001)0.000003 (0.000001)0.000035 (0.000009)
 Var(U3j) – time since first delivery6  0.0516 (0.0173)
 Cov(U0j,U1j)70.1886 (0.1279)0.2174 (0.1281)0.3610 (0.1401)
 Cov(U0j,U2j)70.0019 (0.0018)0.0019 (0.0018)0.0010 (0.0019)
 Cov(U0j,U3j)7  −0.1406 (0.2333)
 Cov(U1j,U2j)7−0.00010 (0.00005)−0.00011 (0.00005)−0.00008 (0.00006)
 Cov(U1j,U3j)7  −0.00746 (0.00724)
 Cov(U2j,U3j)7  −0.00027 (0.00009)
Level-one variance
 Var(Rij) – level-1 residual5.9706 (0.2652)5.9787 (0.2656)5.6973 (0.2550)
 Deviance (−2 × log-likelihood)6978.6076974.6496936.971
 Log-likelihood deviance test P = 0.047P < 0.0001

Model interpretation

The interpretation of the coefficients in all three models depends on whether the coefficient describes a time variable or not. The situation when the coefficient represents a time variable is described below. If the coefficient does not represent a time variable, e.g. the coefficient for ‘currently pregnant’ or the coefficient for ‘3 months post-partum’, then a positive coefficient means a higher CD4 count, while a negative coefficient means a lower CD4 count. For example, the coefficient for ‘currently pregnant’ in the final model is −1.6832. This negative coefficient indicates that, while pregnant, the CD4 count drops, but it comes back up after pregnancy. This particular effect is statistically significant at P < 0.05. The statistical significance of all effects in the models can be approximated by the Wald test; that is, if the absolute value of the coefficient divided by its standard error (SE) is greater than 2, then the effect is statistically significant at P < 0.05. Note that in a multilevel regression model, unlike a standard (ordinary least squares) linear regression, the Wald test is only approximate. This is why, to test the two main hypotheses of this paper, we developed three models and employed the log-likelihood deviance tests. A log-likelihood deviance test compares one model to another. To assess the impact on CD4 decline of ever becoming pregnant, for example, one subtracts the deviance of model 2 from the deviance of model 1 and this is distributed as a chi-square with 1 degree of freedom.

There are two groups of time variables – the time since first CD4 count group, and the time since first delivery group. In general, the negative coefficient for time since first CD4 count means that CD4 declines over time. The negative coefficient for time since first delivery indicates that, among women who ever become pregnant, their CD4 decline is steeper after delivery than before delivery. The square and cubic components of these time variables, however, make interpretation of a single time coefficient conceptually difficult. This is why we graphically depicted the effect of time on CD4 decline in Figure 1. The square and cubic components of the time variables were included in order to allow the CD4 decline to be non-linear. This follows the typical time course of CD4 cell depletion that was found in the Multicenter AIDS Cohort Study (MACS) (Kaslow et al. 1987).

Ancillary