A multivariate logistic regression equation to predict ART Adherence in Zimbabwe: Development and Validation

Introduction: Epidemiologic methods need to fit to contextual clinical practice. The aim was to develop and validate an empirical equation to predict optimal adherence to ART in Zimbabwe.

Methods: A 12 month prospective cohort study of treatment-experienced patients with measurement every three months using a thirty-day recall adherence method validated by a seven-day recall and CD4 count testing every three months and by the Pill Count method at twelve months. Descriptive statistics tested data for normality; bivariate statistics tested for causal associations; Pearson’s correlation coefficient tested for Collinearity and multivariate logistic regression modelling tested for associations of independent variables whilst controlling for the effects of other variables.  The Mantel Haensel Chi² tested for homogeneity within associations. Adjustments for potential confounders; interactions were assessed through the ‘collin’ function in STATA. Finally, the linktest and the Hosmer-Lemeshow goodness-of-fit tests were used in final model diagnostics.

 

Results: Stratification of age distribution by gender revealed normal distribution. Pearson’s Correlation Coefficients (r) for the final set of variables revealed strong positive correlations between “WHO clinical stages 3&4” and “Cotrimoxazole Prophylaxis” (r = 0.764 and 0.659 respectively); “Clinic” and “Time-on-treatment >24months” (r = 0.837); “Treatment Buddie” and “Confidence” (r = 0.743); and “Treatment Buddie” and “HIVDR awareness” (r = 0.749). Zero tolerance (0.0000) and high variance inflation factor (VIF) (8.58e⁺¹²) between “Cotrimoxazole” and “WHO staging” resulted in choosing “WHO staging” for the final model since “Cotrimoxazole prophylaxis” depended on WHO staging in the program theory. The final model contained 13 variables, 3 of which (age, clinic and gender) were forced into the final model to ensure public health significance:

 Optimal Adherence =

P = 1/(1 - e(-x)), where x = 1.703 + [0.46 (age in 5 yr categories) + 0.74 (if urban) + 0.80 (if unemployed) + 0.36 (if female) + 0.64 (if cannot tell how to take medication) + 0.34 (if unaware of HIVDR) + 1.24 (if Confident) + 0.24 (Difficulty taking medications) + 0.65 (if Alcohol dependency) + 0.63 (if Never Condom use) + 0.28 (No support group) + 0.47 (if No treatment buddy) + 0.33 (if Stavudine-based regimen) + 0.56 (If treatment duration >24months)]

The Linktest (Pseudo R=0.1415) and Hosmer and Lemeshow's (p=0.3406) goodness-of-fit test statistic confirm that the model’s estimates fit the data at an acceptable level with sensitivity = 16.09%; Specificity = 95.74%; PPV = 63.64%; and NPV = 71.15%. The model correctly classified 70.55% of patients.

 

Conclusion: This multivariate logistic equation improves on current recommendations of predicting optimal ART adherence in resource – limited settings where self reports can be easily implemented as inexpensive monitoring tools.