Early treatment of depression has been proven to be very effective. However, diagnosing depression remains a challenge with current diagnostic tools. Research has begun exploring digital and mobile modalities for depression screening. In our research, we focus on screening for moderate depression with the typed and transcribed responses in the StudentSADD (Student Suicidal Ideation and Depression Detection) dataset. Our modeling approach involves comparing the performance of five BERT (Bidirectional Encoder Representations from Transformers) variants with and without fine-tuning layers. Our results indicate that typed responses are more useful than transcribed responses when screening for depression. In particular, the less computationally expensive BERT variants with both fine-tuning layers performed best for the typed responses. The highest classifier balanced accuracy was 0.63. Our research can help inform the future development of essential digital mental illness screening models.